build: rename the module to sourcedock.dev/petrbalvin/gasm-sdk

Assisted-by: GLM 5.3 Flash

.gitea/workflows/race.yml

@@ -0,0 +1,37 @@
+# Race, Go. Dispatched by hand, and never a gate on a push or a tag: the release tag is
+# cut only after `just gates` has already raced the tree, so this workflow is the
+# explicit second opinion, not a step of the release.
+#
+# The race detector roughly doubles both time and memory, which the shared runner box
+# cannot afford on every push. Locally it belongs to `just gates`, which runs it once per
+# task; here it is a decision rather than a routine.
+#
+# Every step is one command, so the step that fails is the gate that failed.
+name: Race
+
+on:
+  workflow_dispatch:
+
+env:
+  # One core: parallelism buys no speed here and costs memory the box does not have.
+  GOFLAGS: -p=1
+  GOMAXPROCS: "2"
+
+jobs:
+  race:
+    runs-on: fedora
+    timeout-minutes: 20
+    steps:
+      - uses: actions/checkout@v7
+
+      - uses: actions/setup-go@v6
+        with:
+          go-version-file: go.mod
+          cache: true
+
+      - name: Install gcc
+        # The race detector needs cgo and the runner image carries no C compiler.
+        run: dnf install -y gcc
+
+      - name: Race
+        run: go test -race -count=1 -timeout 10m ./...

.gitea/workflows/release.yml

@@ -1,73 +1,222 @@
-# Release — gasm binaries. Runs on version tags (v0.28.0) pushed to main.
+# Release, Go binaries. Runs on version tags (v1.2.3) pushed to main.
+#
+# The module sits at the repository root: the toolchain records a version only for a root
+# module, measured on go1.27.1, so a build of a module in a subdirectory reports (devel)
+# even at its own <module>/vX.Y.Z tag and this workflow's smoke test can never pass for
+# it. A Go repository is one module at the root.
+#
+# The version contract these steps implement: nothing is injected.  The toolchain records
+# the tag into the binary's build information, so the build simply has to happen at the
+# tag, which the trigger guarantees.
+#
+# The gates run in their own job, once, before the matrix, minus the race detector: race
+# never runs on a push path or a tag, and the local gate raced this tree before the tag
+# was cut. Putting the gates inside the matrix would run the whole suite once per target
+# on the box that also hosts the forge. Each job validates the tag for itself rather than
+# passing a value between jobs, so no workflow feature has to be trusted for the version
+# to reach the file name.
 name: Release
 
 on:
   push:
     tags: ["v*"]
 
+env:
+  # The box is shared with the forge, so parallelism is bounded on purpose. The gates job
+  # needs it most; the build jobs inherit it for their parallel compilation.
+  GOFLAGS: -p=1
+  GOMAXPROCS: "2"
+
 jobs:
+  gates:
+    runs-on: fedora
+    timeout-minutes: 10
+    steps:
+      - uses: actions/checkout@v7
+
+      - uses: actions/setup-go@v6
+        with:
+          go-version-file: go.mod
+          cache: true
+
+      - name: Install Perl
+        # Perl for the steps below. The install is a no-op where the package
+        # is already present.
+        run: dnf install -y perl
+
+      - name: Validate the tag
+        env:
+          VERSION: ${{ gitea.ref_name }}
+        run: |
+          perl -e '
+            my $v = $ENV{VERSION} // q{};
+            $v =~ m{^v[0-9]+(\.[0-9]+){0,2}([-+].*)?$}
+              or die qq{ERROR: expected a semver tag like v1.2.3, got: $v\n};
+            print qq{tag $v\n};
+          '
+
+      - name: Security policy names this release
+        # The supported-versions table is the one part of SECURITY.md that
+        # carries a version, so it goes stale the moment a tag is cut.  Fail
+        # here rather than publish a policy naming the previous release.
+        env:
+          VERSION: ${{ gitea.ref_name }}
+        run: |
+          perl -e '
+            my $v = $ENV{VERSION} // q{};
+            (my $nv = $v) =~ s/^v//;
+            open(my $f, q{<}, q{SECURITY.md}) or die qq{SECURITY.md: $!\n};
+            local $/;
+            my $t = <$f>;
+            close $f;
+            $t =~ m{^\|\s*\Q$nv\E\s*\|\s*yes\s*\|}m
+              or die qq{ERROR: SECURITY.md does not name $nv as supported; update the table before releasing.\n};
+            print qq{SECURITY.md names $nv\n};
+          '
+
+      - name: Build
+        run: go build ./...
+
+      - name: Format
+        run: |
+          perl -e '
+            open(my $g, q{-|}, q{gofmt}, q{-l}, q{.}) or die qq{gofmt: $!};
+            my @bad = <$g>;
+            close($g);
+            print @bad;
+            exit(@bad ? 1 : 0);
+          '
+
+      - name: Vet
+        run: go vet ./...
+
+      - name: Modernise
+        run: go fix -diff ./...
+
+      - name: Tests
+        # The same command as in test.yml, so the floor is the same number everywhere.
+        run: go test -count=1 -timeout 10m -coverprofile=coverage.out ./arch/... ./asm/... ./ast/... ./disasm/... ./format/... ./lexer/... ./lint/... ./lsp/... ./parser/... ./token/... ./verify/...
+
+      - name: Tests outside the coverage set
+        # The same command as in test.yml: the CLI's exit codes and manual-page guard,
+        # and the debugger's architecture-neutral units, run outside the floor.
+        run: go test -count=1 -timeout 10m ./cmd/... ./debug/...
+
+      - name: Coverage floor
+        run: |
+          perl -e '
+            open(my $c, q{-|}, q{go}, q{tool}, q{cover}, q{-func=coverage.out}) or die qq{cover: $!};
+            my $total;
+            while (my $l = <$c>) { $total = $1 if $l =~ m{^total:\s+\S+\s+([0-9.]+)%} }
+            close($c);
+            die qq{no total line in coverage.out\n} unless defined $total;
+            printf qq{Total coverage: %s%%\n}, $total;
+            exit($total < 80 ? 1 : 0);
+          '
+
   build:
     runs-on: fedora
+    timeout-minutes: 25
+    needs: gates
     strategy:
       fail-fast: false
       matrix:
+        # Portable targets: amd64, arm64, loong64 and riscv64 on Linux, at the toolchain
+        # default level. No 32-bit, no wasm, no macOS, no Windows. FreeBSD stays out until
+        # verify/jit.go ports off syscall.Mprotect: the Go syscall package defines no
+        # Mprotect for freebsd, and verify/jit.go:50 calls it to drop the write bit from
+        # the JIT mapping, so every freebsd target fails to build with "undefined:
+        # syscall.Mprotect" (verified for amd64, arm64 and riscv64 on go1.27.1).
         include:
           - goos: linux
             goarch: amd64
           - goos: linux
             goarch: arm64
-          - goos: linux
-            goarch: riscv64
           - goos: linux
             goarch: loong64
+          - goos: linux
+            goarch: riscv64
     steps:
       - uses: actions/checkout@v7
 
       - uses: actions/setup-go@v6
         with:
-          go-version: "1.27"
+          go-version-file: go.mod
+          cache: true
 
-      - name: Download dependencies
-        run: go mod download
+      - name: Install Perl
+        run: dnf install -y perl
 
-      - name: Validate tag and build
-        id: build
+      - name: Validate the tag
+        id: version
         env:
           VERSION: ${{ gitea.ref_name }}
         run: |
-          set -euo pipefail
+          perl -e '
+            my $v = $ENV{VERSION} // q{};
+            $v =~ m{^v[0-9]+(\.[0-9]+){0,2}([-+].*)?$}
+              or die qq{ERROR: expected a semver tag like v1.2.3, got: $v\n};
+            (my $nv = $v) =~ s{^v}{};
+            open(my $o, q{>>}, $ENV{GITEA_OUTPUT}) or die qq{GITEA_OUTPUT: $!};
+            print $o qq{version_no_v=$nv\n};
+            close($o);
+            print qq{version $nv\n};
+          '
 
-          if ! echo "$VERSION" | grep -qE '^v[0-9]+(\.[0-9]+){0,2}([-+].*)?$'; then
-            echo "ERROR: expected a semver tag like v1.2.3, got: '$VERSION'"
-            exit 1
-          fi
-
-          VERSION_NO_V="${VERSION#v}"
-          echo "version_no_v=${VERSION_NO_V}" >> "$GITEA_OUTPUT"
-
-          mkdir -p bin
-          GOOS=${{ matrix.goos }} GOARCH=${{ matrix.goarch }} CGO_ENABLED=0 \
-            go build -ldflags "-s -w -X main.version=${VERSION_NO_V}" \
-              -o "bin/gasm-${VERSION_NO_V}-${{ matrix.goos }}-${{ matrix.goarch }}" \
-              ./cmd/gasm
+      - name: Build
+        env:
+          VERSION_NO_V: ${{ steps.version.outputs.version_no_v }}
+          GOOS: ${{ matrix.goos }}
+          GOARCH: ${{ matrix.goarch }}
+          CGO_ENABLED: "0"
+        run: |
+          # Nothing is injected. The toolchain records the tag into the binary's build
+          # information, so the version is right because this build happens at the tag, and
+          # there is no path for anyone to get wrong. -s -w only strips symbols.
+          go build -ldflags "-s -w" -o "bin/gasm-${VERSION_NO_V}-${GOOS}-${GOARCH}" ./cmd/gasm
 
+      # Artifacts stay on v3: v4 and later detect Gitea as GHES and abort.
       - name: Upload artifact
         uses: actions/upload-artifact@v3
         with:
           name: gasm-${{ matrix.goos }}-${{ matrix.goarch }}
-          path: bin/gasm-${{ steps.build.outputs.version_no_v }}-${{ matrix.goos }}-${{ matrix.goarch }}
+          path: bin/gasm-${{ steps.version.outputs.version_no_v }}-${{ matrix.goos }}-${{ matrix.goarch }}
           if-no-files-found: error
 
       - name: Smoke test
+        # Only a binary matching the runner can be run here. The check is not that --version
+        # exits cleanly but that it reports the tag and nothing more: a build outside version
+        # control reports (devel), and a build whose tree was dirty reports +dirty, and both
+        # would otherwise be published.
         if: matrix.goos == 'linux' && matrix.goarch == 'amd64'
+        env:
+          TAG: ${{ gitea.ref_name }}
+          BIN: bin/gasm-${{ steps.version.outputs.version_no_v }}-${{ matrix.goos }}-${{ matrix.goarch }}
         run: |
-          chmod +x bin/gasm-${{ steps.build.outputs.version_no_v }}-${{ matrix.goos }}-${{ matrix.goarch }}
-          ./bin/gasm-${{ steps.build.outputs.version_no_v }}-${{ matrix.goos }}-${{ matrix.goarch }} --version
+          perl -e '
+            my $want = $ENV{TAG} // die qq{ERROR: no tag\n};
+            open(my $bin, q{-|}, $ENV{BIN}, q{--version}) or die qq{$ENV{BIN}: $!};
+            my $got = <$bin>;
+            close($bin);
+            $got = defined $got ? $got : q{};
+            chomp $got;
+            index($got, $want) >= 0
+              or die qq{ERROR: the binary printed "$got", which does not contain $want. Version control was disabled, so there is no recorded version.\n};
+            index($got, q{+dirty}) < 0
+              or die qq{ERROR: the binary printed "$got". The tree was dirty at build time, which means the checkout was not the tag, or the build artefacts are not ignored.\n};
+            print qq{$ENV{BIN} reports $got\n};
+          '
 
   release:
     runs-on: fedora
+    timeout-minutes: 15
     needs: build
     permissions:
+      # contents: read is required for the checkout: a job that declares any
+      # permissions gets a token scoped to exactly those, and releases: write
+      # alone leaves the fetch with no read access, which Gitea answers with
+      # a 404 "Repository not found". Verified on the instance 2026-09-16.
+      contents: read
       releases: write
     steps:
       - uses: actions/checkout@v7
@@ -77,81 +226,165 @@ jobs:
         with:
           path: dist
 
-      - name: Extract CHANGELOG section
+      - name: Install Perl
+        run: dnf install -y perl
+
+      - name: Extract the CHANGELOG section
         env:
           VERSION: ${{ gitea.ref_name }}
         run: |
-          set -euo pipefail
-          VERSION_NO_V="${VERSION#v}"
+          # Each step derives what it needs from the tag, so no value has to travel between
+          # jobs.
+          perl -e '
+            my $v = $ENV{VERSION} // q{};
+            $v =~ s{^v}{};
+            open(my $vout, q{>}, q{version-no-v.txt}) or die qq{version-no-v.txt: $!};
+            print $vout $v;
+            close($vout);
+            open(my $in, q{<}, q{CHANGELOG.md}) or die qq{CHANGELOG.md: $!};
+            my @lines = <$in>;
+            close($in);
+            my ($start, $end) = (-1, scalar @lines);
+            for my $i (0 .. $#lines) {
+              if ($start < 0) { $start = $i if $lines[$i] =~ m{^##\s+\[\Q$v\E\]} }
+              elsif ($lines[$i] =~ m{^##\s+\[}) { $end = $i; last }
+            }
+            $start >= 0 or die qq{ERROR: no CHANGELOG section for $v, expected a heading like: ## [$v] - YYYY-MM-DD\n};
+            my @body = grep { m{\S} } @lines[$start + 1 .. $end - 1];
+            @body or die qq{ERROR: the CHANGELOG section for $v is empty\n};
+            open(my $out, q{>}, q{release-body.md}) or die qq{release-body.md: $!};
+            print $out @body;
+            close($out);
+            printf qq{notes for %s: %d lines\n}, $v, scalar @body;
+          '
 
-          sed -n "/^## \[${VERSION_NO_V}\] /,/^## \[/p" CHANGELOG.md \
-            | sed '$d' \
-            | tail -n +2 \
-            > release-body.md
+      - name: Build the release request
+        run: |
+          perl -e '
+            open(my $vin, q{<}, q{version-no-v.txt}) or die qq{version-no-v.txt: $!};
+            my $v = <$vin>;
+            close($vin);
+            chomp $v;
+            open(my $in, q{<:raw}, q{release-body.md}) or die qq{release-body.md: $!};
+            my $body = do { local $/; <$in> };
+            close($in);
+            # Byte-oriented escaping: JSON is UTF-8, so non-ASCII passes through and only the
+            # characters JSON forbids are rewritten.
+            $body =~ s/([\\"])/\\$1/g;
+            $body =~ s/\t/\\t/g;
+            $body =~ s/\r//g;
+            $body =~ s/\n/\\n/g;
+            $body =~ s/([\x00-\x08\x0b\x0c\x0e-\x1f])/sprintf(q{\u%04x}, ord($1))/ge;
+            my $json = sprintf(qq{{"tag_name":"v%s","name":"v%s","body":"%s","draft":false,"prerelease":false}}, $v, $v, $body);
+            open(my $out, q{>}, q{release.json}) or die qq{release.json: $!};
+            print $out $json;
+            close($out);
+            print qq{release.json written for v$v\n};
+          '
 
-          if [ ! -s release-body.md ]; then
-            echo "ERROR: no CHANGELOG section found for ${VERSION_NO_V}"
-            echo "Expected a heading like: ## [${VERSION_NO_V}] — YYYY-MM-DD"
-            exit 1
-          fi
-
-      - name: Create release
+      - name: Create the release
         env:
           GITEA_TOKEN: ${{ secrets.GITEA_TOKEN }}
           GITEA_SERVER_URL: ${{ gitea.server_url }}
           GITEA_REPOSITORY: ${{ gitea.repository }}
-          GITEA_REF_NAME: ${{ gitea.ref_name }}
         run: |
-          set -euo pipefail
-
-          BODY=$(sed -e 's/\\/\\\\/g' -e 's/"/\\"/g' -e 's/\t/\\t/g' -e 's/\r//g' release-body.md | sed ':a;N;$!ba;s/\n/\\n/g')
-          BODY="\"${BODY}\""
-
-          response=$(curl -sS -w '\n%{http_code}' \
-            -H "Authorization: token ${GITEA_TOKEN}" \
-            -H "Content-Type: application/json" \
-            -X POST \
-            "${GITEA_SERVER_URL}/api/v1/repos/${GITEA_REPOSITORY}/releases" \
-            -d "{\"tag_name\":\"${GITEA_REF_NAME}\",\"name\":\"${GITEA_REF_NAME}\",\"body\":${BODY},\"draft\":false,\"prerelease\":false}")
-
-          http_code=$(echo "$response" | tail -1)
-          payload=$(echo "$response" | sed '$d')
-
-          echo "HTTP ${http_code}"
-          if [ "$http_code" != "201" ]; then
-            echo "Failed to create release: ${payload}"
-            exit 1
-          fi
-
-          RELEASE_ID=$(echo "$payload" | grep -oE '"id"[[:space:]]*:[[:space:]]*[0-9]+' | head -1 | grep -oE '[0-9]+')
-          echo "Created release ID=${RELEASE_ID}"
-          printf '%s' "${RELEASE_ID}" > release-id.txt
+          perl -e '
+            my @cmd = (q{curl}, q{-sS}, q{-o}, q{response.json}, q{-w}, q{%{http_code}},
+              q{-H}, qq{Authorization: token $ENV{GITEA_TOKEN}},
+              q{-H}, q{Content-Type: application/json},
+              q{-X}, q{POST},
+              qq{$ENV{GITEA_SERVER_URL}/api/v1/repos/$ENV{GITEA_REPOSITORY}/releases},
+              q{--data-binary}, q{@release.json});
+            open(my $curl, q{-|}, @cmd) or die qq{curl: $!};
+            my $code = <$curl>;
+            my $ok = close($curl);
+            my $exit = $? >> 8;
+            $code = defined $code ? $code : q{};
+            $ok or die qq{ERROR: curl failed (exit $exit) calling $ENV{GITEA_SERVER_URL}\n};
+            open(my $r, q{<:raw}, q{response.json}) or die qq{response.json: $!};
+            my $body = do { local $/; <$r> };
+            close($r);
+            $code eq q{201} or die qq{ERROR: the release was not created, HTTP $code: $body\n};
+            $body =~ m{"id"\s*:\s*([0-9]+)} or die qq{ERROR: no release id in the response: $body\n};
+            open(my $o, q{>}, q{release-id.txt}) or die qq{release-id.txt: $!};
+            print $o $1;
+            close($o);
+            print qq{release id $1\n};
+          '
 
       - name: Upload assets
         env:
           GITEA_TOKEN: ${{ secrets.GITEA_TOKEN }}
           GITEA_SERVER_URL: ${{ gitea.server_url }}
           GITEA_REPOSITORY: ${{ gitea.repository }}
-          GITEA_REF_NAME: ${{ gitea.ref_name }}
         run: |
-          set -euo pipefail
-          RELEASE_ID=$(cat release-id.txt)
-
-          for binary in dist/gasm-*/gasm-*; do
-            [ -f "$binary" ] || continue
-            fname=$(basename "$binary")
-            echo "Uploading ${fname}..."
-            http_code=$(curl -sS -o /dev/null -w '%{http_code}' \
-              -H "Authorization: token ${GITEA_TOKEN}" \
-              -H "Content-Type: application/octet-stream" \
-              -X POST \
-              --data-binary "@${binary}" \
-              "${GITEA_SERVER_URL}/api/v1/repos/${GITEA_REPOSITORY}/releases/${RELEASE_ID}/assets?name=${fname}")
-            echo "  HTTP ${http_code}"
-            if [ "$http_code" != "201" ]; then
-              echo "Failed to upload ${fname}"
-              exit 1
-            fi
-          done
-
-          echo "Release ${GITEA_REF_NAME} is live."
+          perl -e '
+            open(my $f, q{<}, q{release-id.txt}) or die qq{release-id.txt: $!};
+            my $id = <$f>;
+            close($f);
+            chomp $id;
+            my @files = grep { -f $_ } glob(q{dist/*/*});
+            @files or die qq{ERROR: no assets under dist/\n};
+            # A file that arrived empty from the artifact step would be uploaded as an
+            # empty attachment, every status would still be 201, and the run would go
+            # green over a release nobody can install.  Refuse it here, before the
+            # upload, and verify what was stored afterwards.
+            my %size;
+            for my $path (@files) {
+              my $n = -s $path // 0;
+              (my $name = $path) =~ s{.*/}{};
+              $n > 0 or die qq{ERROR: $path is empty, so there is nothing to upload\n};
+              $size{$name} = $n;
+            }
+            my $bad = 0;
+            for my $path (@files) {
+              (my $name = $path) =~ s{.*/}{};
+              my @cmd = (q{curl}, q{-sS}, q{-o}, q{/dev/null}, q{-w}, q{%{http_code}},
+                q{-H}, qq{Authorization: token $ENV{GITEA_TOKEN}},
+                q{-H}, q{Content-Type: application/octet-stream},
+                # The @ must not sit inside a qq{} string: there it starts an
+                # array interpolation and the upload body collapses to empty,
+                # which Gitea stores as a 201-created zero-byte attachment.
+                q{-X}, q{POST}, q{--data-binary}, q{@} . $path,
+                qq{$ENV{GITEA_SERVER_URL}/api/v1/repos/$ENV{GITEA_REPOSITORY}/releases/$id/assets?name=$name});
+              open(my $curl, q{-|}, @cmd) or die qq{curl: $!};
+              my $code = <$curl>;
+              my $ok = close($curl);
+              my $exit = $? >> 8;
+              $code = defined $code ? $code : q{};
+              unless ($ok) {
+                printf qq{%s: curl failed (exit %d)\n}, $name, $exit;
+                $bad = 1;
+                next;
+              }
+              printf qq{%s: HTTP %s\n}, $name, $code;
+              $bad = 1 if $code ne q{201};
+            }
+            # Read every asset back through the release download route and require the
+            # served length to be the file that was sent: stored but empty is a broken
+            # release however green the run looks.
+            open(my $v, q{<}, q{version-no-v.txt}) or die qq{version-no-v.txt: $!};
+            my $v = <$v>;
+            close($v);
+            chomp $v;
+            for my $name (sort keys %size) {
+              my $url = qq{$ENV{GITEA_SERVER_URL}/$ENV{GITEA_REPOSITORY}/releases/download/v$v/$name};
+              my @head = (q{curl}, q{-sS}, q{-I}, q{-H}, qq{Authorization: token $ENV{GITEA_TOKEN}}, $url);
+              open(my $h, q{-|}, @head) or die qq{curl: $!};
+              my $len;
+              my $status;
+              while (my $l = <$h>) {
+                $status = $1 if $l =~ m{^HTTP/\S+\s+(\d+)};
+                $len = $1 if $l =~ m{^content-length:\s*(\d+)}i;
+              }
+              my $ok = close($h);
+              $len = defined $len ? $len : 0;
+              if (!$ok || $status != 200 || $len != $size{$name}) {
+                printf qq{ERROR: %s serves %s bytes, expected %d\n}, $name, $len, $size{$name};
+                $bad = 1;
+                next;
+              }
+              printf qq{%s: serves %d bytes\n}, $name, $len;
+            }
+            exit($bad ? 1 : 0);
+          '

.gitea/workflows/test.yml

@@ -1,4 +1,17 @@
-# Test — gasm-devkit. Runs on push and pull request to development.
+# Test, Go. Push and pull request to development. Never on main.
+#
+# The gates are the ones the justfile's `gates` recipe runs, minus race: the shared
+# runner box cannot afford the race detector on every push, so it lives in race.yml.
+# The box is one core and 2 GB beside Gitea, so parallelism is bounded on purpose and
+# everything runs in one job. Extra jobs would duplicate the checkout, the Go setup and
+# the dependency download three times without buying any parallelism.
+#
+# Every step is one command, so the step that fails is the gate that failed, and no shell
+# option has to be trusted for the run to stop. The scripted steps are Perl, not shell and
+# not Python: Perl behaves the same on both runner images, there is no bashism to trip over
+# on ash, and it is one language instead of two. The Perl uses builtins only, because
+# Fedora packages the Perl modules separately and nothing beyond `perl` itself may be
+# assumed present.
 name: Test
 
 on:
@@ -7,90 +20,113 @@ on:
   pull_request:
     branches: [development]
 
+env:
+  # One core: parallelism buys no speed here and costs memory the box does not have.
+  GOFLAGS: -p=1
+  GOMAXPROCS: "2"
+
+# A superseded run of the same ref is cancelled instead of queueing behind one that
+# no longer matters. Verified on Gitea 1.27.1 on 2026-09-17: a queued run whose ref
+# moved on is cancelled before it ever reaches the runner, while a run already
+# dispatched there runs to completion.
+concurrency:
+  group: ${{ gitea.workflow }}-${{ gitea.ref }}
+  cancel-in-progress: true
+
 jobs:
-  vet:
-    runs-on: fedora
-    steps:
-      - uses: actions/checkout@v7
-
-      - uses: actions/setup-go@v6
-        with:
-          go-version: "1.27"
-
-      - name: Download dependencies
-        run: go mod download
-
-      - name: gofmt
-        run: |
-          set -euo pipefail
-          unformatted=$(gofmt -l .)
-          if [ -n "$unformatted" ]; then
-            echo "These files need gofmt:"
-            echo "$unformatted"
-            exit 1
-          fi
-
-      - name: go vet
-        run: go vet ./...
-
   test:
     runs-on: fedora
-    needs: vet
+    timeout-minutes: 10
     steps:
       - uses: actions/checkout@v7
 
       - uses: actions/setup-go@v6
         with:
-          go-version: "1.27"
+          # The module is the source of truth for the version, so it cannot drift.
+          go-version-file: go.mod
+          cache: true
 
-      - name: Download dependencies
-        run: go mod download
-
-      - name: Install gcc
-        run: dnf install -y gcc
-
-      - name: go test -race
-        run: go test -race -count=1 ./...
-
-      - name: Coverage gate — 80 % minimum
-        run: |
-          set -euo pipefail
-          # Exclude packages inherently untestable without hardware:
-          # debug — interactive ptrace, requires a live process
-          # cmd/gasm — CLI glue, covered by integration tests
-          go test -coverprofile=coverage.out \
-            sourcedock.dev/petrbalvin/gasm-devkit/arch \
-            sourcedock.dev/petrbalvin/gasm-devkit/asm \
-            sourcedock.dev/petrbalvin/gasm-devkit/ast \
-            sourcedock.dev/petrbalvin/gasm-devkit/format \
-            sourcedock.dev/petrbalvin/gasm-devkit/lexer \
-            sourcedock.dev/petrbalvin/gasm-devkit/lint \
-            sourcedock.dev/petrbalvin/gasm-devkit/lsp \
-            sourcedock.dev/petrbalvin/gasm-devkit/parser \
-            sourcedock.dev/petrbalvin/gasm-devkit/token \
-            sourcedock.dev/petrbalvin/gasm-devkit/verify
-          coverage=$(go tool cover -func=coverage.out | awk '/^total:/ { gsub("%", "", $3); print $3 }')
-          echo "Total coverage: ${coverage}%"
-          if awk -v c="$coverage" 'BEGIN { exit !(c+0 < 80) }'; then
-            echo "ERROR: coverage ${coverage}% is below the 80% threshold"
-            exit 1
-          fi
-
-  build:
-    runs-on: fedora
-    needs: test
-    steps:
-      - uses: actions/checkout@v7
-
-      - uses: actions/setup-go@v6
-        with:
-          go-version: "1.27"
-
-      - name: Download dependencies
-        run: go mod download
+      - name: Install Perl
+        # The runner images are minimal and Perl is not guaranteed. The install is a
+        # no-op where it is already present; drop this step once verified on the box.
+        run: dnf install -y perl
 
+      # The steps follow the `gates` order of the justfile contract: build, format,
+      # vet, test. The vet gate is go vet and go fix -diff, two steps here.
       - name: Build
-        run: go build -ldflags="-s -w" -o bin/gasm ./cmd/gasm
+        run: go build ./...
 
-      - name: Smoke test
-        run: ./bin/gasm --version
+      - name: FreeBSD build (amd64)
+        # The debugger's ptrace surface and the JIT substrate are the two
+        # FreeBSD-portable layers the tree carries; the forge has no FreeBSD
+        # runner, so a push can only compile-gate them. Running the ptrace
+        # suite needs real FreeBSD hardware.
+        env:
+          GOOS: freebsd
+          GOARCH: amd64
+        run: go build ./...
+
+      - name: FreeBSD build (arm64)
+        env:
+          GOOS: freebsd
+          GOARCH: arm64
+        run: go build ./...
+
+      - name: FreeBSD build (riscv64)
+        env:
+          GOOS: freebsd
+          GOARCH: riscv64
+        run: go build ./...
+
+      - name: Format
+        run: |
+          perl -e '
+            open(my $g, q{-|}, q{gofmt}, q{-l}, q{.}) or die qq{gofmt: $!};
+            my @bad = <$g>;
+            close($g);
+            print @bad;
+            exit(@bad ? 1 : 0);
+          '
+
+      - name: Vet
+        run: go vet ./...
+
+      - name: Modernise
+        # Exits non-zero when it has something to rewrite, so it needs no output capture.
+        run: go fix -diff ./...
+
+      - name: Tests
+        # The suite must be fast: a push pipeline that cannot finish in a few minutes moves
+        # its heavy part behind a dispatch. The inner timeout matches the job's, so a
+        # hanging test reports its own goroutine dump rather than a silent job kill.
+        # The pattern is `packages` in the project's justfile: the logic packages, since a
+        # thin cmd/ would drag the total under the floor. release.yml runs the same
+        # command, so the floor is the same number everywhere. ./verify/... carries the
+        # live oracle-parity comparison against `go tool asm` (the TestGroundTruth
+        # suites); the runner's Go setup provides both the tool and GOROOT.
+        run: go test -count=1 -timeout 10m -coverprofile=coverage.out ./arch/... ./asm/... ./ast/... ./disasm/... ./format/... ./lexer/... ./lint/... ./lsp/... ./parser/... ./token/... ./verify/...
+
+      - name: Tests outside the coverage set
+        # The CLI and the debugger sit outside `packages` because a thin main and a
+        # ptrace-bound package pull the total under the floor, but their tests guard
+        # shipped surfaces: the command exit codes, the manual pages against the
+        # binary's own help, and the debugger's architecture-neutral units.  They run
+        # here so the floor stays a product measure and nothing is left untested.
+        run: go test -count=1 -timeout 10m ./cmd/... ./debug/...
+
+      - name: Oracle parity
+        # Re-run the live go-tool-asm comparison as its own step so that a parity
+        # regression names the gate that failed instead of hiding inside the suite.
+        run: go test -count=1 -timeout 10m -run 'TestGroundTruth' ./verify/...
+
+      - name: Coverage floor
+        run: |
+          perl -e '
+            open(my $c, q{-|}, q{go}, q{tool}, q{cover}, q{-func=coverage.out}) or die qq{cover: $!};
+            my $total;
+            while (my $l = <$c>) { $total = $1 if $l =~ m{^total:\s+\S+\s+([0-9.]+)%} }
+            close($c);
+            die qq{no total line in coverage.out\n} unless defined $total;
+            printf qq{Total coverage: %s%%\n}, $total;
+            exit($total < 80 ? 1 : 0);
+          '

.gitignore

@@ -1,14 +1,13 @@
-# Binaries
-/gasm
-/bin/
-*.exe
+.idea/
+.zcode/
 
-# Test and coverage artefacts
+# Build output
+/bin/
+/gasm
 coverage.out
 *.test
 
-# Scratch / temporary work
-_scratch/
-
-# ZCode workspace
-.zcode
+# Crash dumps from the emulator runs
+core
+core.*
+*.core

AGENTS.md

@@ -1,119 +0,0 @@
-# AGENTS.md — gasm-devkit
-
-Repository rules for AI agents and contributors. Read before modifying any
-code in this repository.
-
-## AI Contribution Policy
-
-AI agents may assist with code, documentation, tests, and review in this
-repository. All AI-assisted changes must:
-
-- Follow the code style and conventions in this file.
-- Include the trailer `Assisted-by: <model-name>` in every commit message.
-- Not commit directly to `main` — work on `development`.
-- Pass the full Definition of Done before any commit.
-
-## Workflow
-
-- **Branching.** `development` is the working branch. `main` is
-  release-only: merge from `development`, then tag. Never commit directly
-  to `main`.
-- **Release procedure.**
-  1. Bump `version` in `justfile` and `cmd/gasm/main.go`.
-  2. Update `CHANGELOG.md` with a new `## [X.Y.Z] — YYYY-MM-DD` section.
-  3. Update `README.md` and `docs/ARCHITECTURE.md` if user-visible
-     behaviour changed.
-  4. Run the Definition of Done (below).
-  5. Commit on `development`.
-  6. `git checkout main && git merge --ff-only development`.
-  7. `git tag vX.Y.Z`.
-  8. `git checkout development`.
-  9. `GOBIN=~/.local/bin just install-bin`.
-
-## Commit Messages
-
-Conventional Commits, subject line only, imperative mood, lowercase after
-the colon:
-
-```
-feat(asm): add EVEX gather and scatter with VSIB addressing
-```
-
-Allowed types: `feat`, `fix`, `docs`, `style`, `refactor`, `perf`, `test`,
-`chore`, `ci`, `build`, `revert`.
-
-Every commit ends with exactly one trailer, using the model that
-assisted with the change:
-
-```
-Assisted-by: <model-name>
-```
-
-Replace `<model-name>` with the actual model (e.g. `DeepSeek V4 Pro`).
-
-No body, no footers, no trailing period on the subject.
-
-## Code Style
-
-Language: Go 1.27 (`toolchain go1.27.0`).
-
-### Formatter
-
-`gofmt` — zero diff. Run `just fmt` before committing.
-
-### Linter
-
-`go vet` — zero warnings. Run `just build` before committing.
-
-### Tests
-
-`go test -race -count=1 ./...` — all green, coverage ≥ 80 % (hard gate,
-enforced by `just test`).
-
-### Dependencies
-
-- **Production code:** standard library only. No third-party imports in
-  shipped code.
-- **Test code:** `golang.org/x/arch` is the sole test dependency (decode
-  oracle for round-trip validation). It is never linked into the binary.
-- **No cgo, no C, no external toolchains, no JavaScript.**
-
-### Error Handling
-
-Explicit `if err != nil`. Wrap with `fmt.Errorf("context: %w", err)`.
-No panics outside `main`. The one exception: the JIT trampoline's
-`recover`-guarded decoder hot path, which converts bounds panics to
-sentinel errors.
-
-### Assembly
-
-Plan 9 syntax (Go's assembler dialect). Hand-written — no code generators
-except `_gen/gen.go` for instruction tables (which parses the Go
-toolchain source). Every instruction table is committed; no runtime
-dependency on the Go toolchain.
-
-### File Naming
-
-- `_amd64.s`, `_arm64.s`, `_riscv64.s`, `_loong64.s` for
-  architecture-specific assembly.
-- `_linux_amd64.go` for platform-specific Go files.
-- `_test.go` suffix for test files.
-
-## Definition of Done
-
-A task is not complete until all of these pass:
-
-1. `just build` — `go vet` + `gofmt` check, zero errors, zero warnings.
-2. `just test` — full suite with `-race`, coverage ≥ 80 %.
-3. `just fmt` — produces no diff.
-4. Diagnostics — zero warnings across the project.
-5. Non-trivial changes reviewed.
-
-## Licence
-
-BSD-3-Clause. Every source file carries the SPDX header:
-
-```
-// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
-// SPDX-License-Identifier: BSD-3-Clause
-```

CONTRIBUTING.md

@@ -1,101 +1,134 @@
-# Contributing to gasm-devkit
+# Contributing
 
-## Prerequisites
+Contributions to **gasm-sdk** are governed by the Contributor terms
+below; submitting one means you accept them.
 
-- Go 1.27 or later (`toolchain go1.27.0`)
-- `just` command runner
-- A Linux host on amd64, arm64, riscv64 or loong64
+## Contributor terms
 
-## Development Setup
+1. This project belongs to its owner alone. The owner decides what is
+   accepted, in what form and when; the decision is final and needs no
+   justification.
+2. By submitting a contribution you assign to Petr Balvín
+   <opensource@petrbalvin.org> all present and future copyright and
+   related rights in it, worldwide, for the full term of the rights,
+   with the right to relicense and sublicense without restriction,
+   including under proprietary terms.
+3. Where that assignment is not effective, it counts as a perpetual,
+   irrevocable, royalty-free licence with the same scope.
+4. To the fullest extent permitted by law, you waive any right of
+   attribution and integrity in the contribution. The project names no
+   contributors and keeps no credits list.
+5. By submitting you represent that the work is yours and that you
+   hold the rights to assign it as above.
+
+## Development setup
+
+Requirements: Go 1.27.1, the exact version the `go` directive in `go.mod`
+declares, [just](https://github.com/casey/just) for the recipes, and a C
+compiler (gcc), because `just gates` includes `just race` and the race
+detector needs cgo.
 
 ```sh
-git clone https://sourcedock.dev/petrbalvin/gasm-devkit.git
-cd gasm-devkit
-just install        # download module dependencies
-just build          # go vet + gofmt check
-just test           # full test suite with race detector
+git clone https://sourcedock.dev/petrbalvin/gasm-sdk.git
+cd gasm-sdk
+just build
+just gates
 ```
 
-## Commands
+## Workflow
 
-Every just recipe:
+1. Branch from `development`. Never commit directly to `main`, which is release-only.
+2. Commit in [Conventional Commits](https://www.conventionalcommits.org/) form:
+   `type(scope): description`, subject line only, imperative mood, lowercase after the
+   colon, no trailing full stop. Allowed types: `feat`, `fix`, `docs`, `style`,
+   `refactor`, `perf`, `test`, `chore`, `ci`, `build`, `revert`.
+3. One logical change per commit. A refactor, a behaviour change and a formatting pass
+   are three commits, never one.
+4. Record every user-visible change in `CHANGELOG.md` under `## [development]`.
+5. Add or update tests. Coverage stays at 80 percent or more; it is a hard gate.
+6. Update the documentation when the public API, the configuration or the behaviour
+   changes.
+7. Open a pull request against `development`.
 
-| Recipe | What it does |
-|--------|-------------|
-| `just` | List all recipes |
-| `just install` | `go mod download` |
-| `just build` | `go vet ./...` + `gofmt -l .` check — zero errors required |
-| `just test` | `go test -race -count=1 -coverprofile=coverage.out ./...` + 80 % coverage gate |
-| `just fmt` | `gofmt -w .` |
-| `just run -- lint file.s` | Run the CLI with `go run` (args after `--`) |
-| `just install-bin` | Install `gasm` into `$GOBIN` with the release version stamped |
-| `just gen` | Regenerate `arch/*_gen.go` instruction tables from the Go toolchain |
-| `just uninstall` | Remove build artefacts (`coverage.out`, `gasm`, `*.test`) |
+Releases are cut by merging `development` into `main` and tagging `vX.Y.Z`. The release
+workflow builds the assets and publishes the release and its notes.
 
-## Running a Single Test
+## Code style
 
-```sh
-go test -run TestVexGroundTruth ./asm/
-go test -run TestDifferentialLZ4Fuzz ./verify/
-```
+`gofmt` and `go vet` run through `just fmt` and `just vet`, with zero diff and zero
+warnings tolerated. `just vet` is two gates, `go vet ./...` and `go fix -diff ./...`,
+so the modernisation rewrites are enforced too. `just gates` is the definition of done in
+one command, and the recipe file names what it contains. Errors are checked explicitly,
+wrapped as `fmt.Errorf("context: %w", err)`, and nothing panics outside `main`. The
+recipe file holds the commands, and the language and standard-library surface is the one
+the `go` directive in `go.mod` pins.
 
-## Testing the Debugger
+- `golang.org/x/arch` is the one module dependency, and it is linked into the binary:
+  `gasm dis` and the debugger's listings decode through it. Everything else is the
+  standard library.
+- No cgo and no C. The standalone encoder paths (`gasm asm --format raw` and `--format
+  elf`) need no Go installation; `gasm verify --ground-truth`, `gasm verify --fuzz`,
+  `gasm audit-instructions` and `gasm asm --format goobj` resolve through the installed
+  Go toolchain.
+- The parser, lexer and formatter are hand-written; the `arch` instruction tables are
+  generated only by `_gen/gen.go` (`just gen`) and never edited by hand.
+- Assembly committed to the repository goes through `gasm fmt` and `gasm lint`, so a
+  `.s` file that `gasm fmt -l .` lists is unfinished.
 
-The interactive debugger (`gasm debug`) requires a compiled binary —
-`go run` does not work for the child process. Install first:
+New source files open with the project's two-line licence header, whose SPDX
+identifier matches `LICENSE`. Configuration files, workflows and dotfiles do not carry
+it.
 
-```sh
-just install-bin
-gasm debug --func add testdata/verify/basic_amd64.s
-```
+## AI contribution policy
 
-## Code Style
+AI tools are welcome as productivity aids and are a normal part of modern software
+development. What matters is that the contribution stays understandable, reviewable and
+genuinely useful.
 
-See [AGENTS.md](AGENTS.md) for the full style guide. Key points:
+- **Disclose the assistance.** If AI helped draft any part of a commit, issue, pull
+  request or review, say so.
+- **Commit messages carry exactly one trailer**, as a git trailer on the line after a
+  blank line that closes the subject:
 
-- `gofmt` — zero diff.
-- `go vet` — zero warnings.
-- Standard library only in production code; `golang.org/x/arch` in tests.
-- No cgo, no C, no JavaScript.
-- Hand-written Plan 9 assembly; tables generated only via `_gen/gen.go`.
+  ```
+  Assisted-by: MODEL
+  ```
 
-## Branches and Releases
+  Name the model that did the work, spelled the way its maker spells it, for example
+  `GLM 5.3`, `DeepSeek V4.1 Flash` or `Qwen 3.8 Flash`. No `Co-Authored-By`, no `Signed-off-by`,
+  no other trailers, and no prose: the trailer is the disclosure.
+- **Issues and pull requests** attribute the assistance in a comment, for example
+  `_Assisted-by: GLM 5.3_`. It does not belong in the pull request description.
+- **Take responsibility.** You are accountable for the accuracy, completeness and
+  intent of everything you submit, whether or not AI produced it.
+- **Review before marking ready.** Read the diff carefully, run it locally, and add the
+  tests it needs. Do not mark a pull request ready until you can defend every change in
+  it.
+- **Quality over quantity.** Contributions that look like un-reviewed output, or whose
+  author cannot engage substantively during review, may be closed.
+- **Preferred models.** Prefer open-weight models with transparent training data and
+  minimal output filtering.
 
-- `development` is the working branch.
-- `main` is release-only: `git merge --ff-only development`, then `git tag vX.Y.Z`.
-- Conventional Commits: `feat(asm): add EVEX gather and scatter`.
-- Every commit ends with `Assisted-by: <model-name>`.
+AI assists. It does not replace judgement.
 
-## CI
+## Continuous integration
 
-CI runs on every push to `development` and on pull requests:
+Workflows live in `.gitea/workflows/` and run on the project's own runners:
 
-- **Test** (`test.yml`) — `gofmt` check, `go vet`, `go test -race` and the
-  80 % coverage gate.
-- **Release** (`release.yml`) — cross-compiles release binaries for
-  linux/{amd64,arm64,riscv64,loong64} on version tags and publishes them.
+| Workflow | Trigger | What it does |
+|---|---|---|
+| Test | push or pull request to `development` | build, format check, vet, modernisation, the test suite with the coverage floor, the CLI and debugger tests outside the profile, then the oracle-parity rerun against `go tool asm` |
+| Release | a `v*` tag | the same gates as Test minus the oracle-parity step, then the matrix build, the version smoke test and the release itself; the race detector runs locally in `just gates` before the tag is cut |
 
-The Definition of Done (`just build` + `just test` + `just fmt`) must still
-pass locally before pushing.
+The local equivalent is `just gates`, which is the same set plus the race detector. The
+race detector also has its own workflow, dispatched by hand; it never runs on a push or a
+tag, where it would double the time and the memory a shared runner cannot spare.
 
-## AI-Assisted Contributions
+## Reporting bugs
 
-AI agents may assist with code, documentation, tests, and review. All
-AI-assisted changes must:
+Open an issue at `https://sourcedock.dev/petrbalvin/gasm-sdk/issues` with the
+version, the operating system and architecture, the exact command, the full output,
+and the expected against the actual behaviour.
 
-- Include the trailer `Assisted-by: <model-name>` in the commit message
-  (e.g. `Assisted-by: DeepSeek V4 Pro`).
-- Follow the [AGENTS.md](AGENTS.md) rules.
-- Pass the Definition of Done before committing.
-
-Attribute agent authorship in issues and pull requests on one trailing
-line:
-
-```
-_Assisted-by: Qwen 3.8 Max_
-```
-
-## Questions
-
-Open an issue at
-[sourcedock.dev/petrbalvin/gasm-devkit](https://sourcedock.dev/petrbalvin/gasm-devkit/issues).
+**Security issues do not go in the issue tracker.** Report them as
+[SECURITY.md](SECURITY.md) describes, to **opensource@petrbalvin.org**.

README.md

@@ -1,144 +1,338 @@
-# gasm-devkit
+# GAsm: Software Development Kit for Plan 9 Assembly
 
-Developer tooling for **GAsm** — Go's built-in Plan 9 assembler.
+> **Warning: this is an experiment.** gasm-sdk is under active
+> development and is not stable. The version is 0.x.x: commands, flags,
+> output formats and behaviour can change without warning at any time.
+> A 1.0.0 release is light years away. Nothing in this document is a
+> stability promise. For all of that, this is not a paper project: gasm
+> is already in active use and is tested on real assembly work. Only
+> amd64 is validated on real hardware; the other three architectures run
+> under emulation ([Validation status](#validation-status)).
 
-Go ships an assembler but no tooling for it.  There is no syntax highlighting,
-no autocomplete, no linter, no static analyser, no formatter, no standalone
-assembler and no debugger for `.s` files.  Developers write assembly blind,
-validate it by benchmark, and debug it by print statement.
+**GAsm** is Go's Plan 9 assembler, and Go ships it without tooling:
+there is no formatter, no linter and no debugger for `.s` files, and no
+assembler that works without a Go installation. Developers write
+assembly blind, validate it by benchmark, and debug it by print
+statement. gasm-sdk is the missing toolkit: a single, self-contained
+binary, `gasm`, that serves both purposes.
 
-gasm-devkit is the missing toolkit.  It is a single, self-contained binary —
-`gasm` — that brings proper developer tooling to Plan 9 assembly:
+- **Help develop Plan 9 assembly.** Formatting, linting, disassembly,
+  dynamic verification, a source-level debugger and a language server,
+  for `.s` files in Go programs.
+- **Use Plan 9 assembly outside the Go toolchain.** `gasm asm` encodes
+  on its own and writes raw images or linkable ELF objects with DWARF5
+  debug sections, with no Go installation in the loop; the Go
+  toolchain's own GOOBJ format, which `go build` consumes in place of
+  the toolchain's output, needs the installed toolchain.
 
-```
-gasm tokens   dump the lexical token stream
-gasm parse    parse and report syntax errors
-gasm fmt      canonicalise formatting (gofmt for assembly)
-gasm lint     static checks
-gasm lsp      language server (completion, hover, symbols, diagnostics, highlighting)
-gasm asm      standalone assembler
-gasm verify   dynamic analysis & verification
-gasm debug    source-level debugger
-gasm diff     compare machine code of two .s files
-gasm profile  show basic-block structure of functions
+## Why Plan 9 assembly
+
+Plan 9 assembly is the quiet triumph of the field. One syntax across
+every architecture Go builds for: the same source-first operand order,
+the same four pseudo-registers, the same frame convention, whether the
+target is x86, ARM, RISC-V or LoongArch. Learn it once and you can
+read a kernel on any of them.
+
+Compare the alternatives. Intel syntax and AT&T syntax disagree on the
+one question every instruction answers, which operand is the source
+and which is the destination, so half the world writes it one way,
+half the other, and every assembly programmer carries both in their
+head forever. GNU as settles the argument with directives that switch
+dialects mid-file (`.intel_syntax noprefix`), a percent sign on every
+register and a dollar on every immediate: punctuation that carries
+nothing the operand order did not already say. And the x86 family
+fragments again underneath: NASM is not MASM is not GAS, each with its
+own directive zoo and macro language, so every project picks a dialect
+and every reader learns a different one by accident.
+
+Plan 9 assembly has none of it. Registers are bare names. Memory is
+one notation, `offset(base)`, extended by an index and a scale when
+the instruction needs it. Arguments arrive named and offset-checked:
+`x+0(FP)` is the argument x, on every architecture, and `go vet`
+polices the offsets against the Go prototype.
+
+```text
+AT&T (GNU as):   movq %rax, -16(%rbp)
+Plan 9 (Go):     MOVQ AX, total-16(SP)
 ```
 
-## Architecture support
+The same lines, but only one of them tells you what the number is for.
+The syntax is uppercase, regular and boring, which is the highest
+compliment a language for machine code can earn. gasm-sdk exists
+to give that syntax the tooling it deserves.
 
-gasm-devkit targets every architecture Go's assembler speaks.  The instruction
-tables are **generated from the Go toolchain's own assembler source**
-(`cmd/internal/obj/<arch>`), so gasm-devkit recognises *every* mnemonic the
-real assembler accepts — not a hand-maintained subset that drifts and rots.
+## Features
 
-| Architecture | GOARCH      | File suffix    | Instructions recognised            |
-|--------------|-------------|----------------|------------------------------------|
-| AMD64        | `amd64`     | `_amd64.s`     | 1600 + common opcodes + traditional aliases |
-| ARM64        | `arm64`     | `_arm64.s`     | 538 + common opcodes               |
-| RISC-V       | `riscv64`   | `_riscv64.s`   | 961 + common opcodes               |
-| LoongArch    | `loong64`   | `_loong64.s`   | 799 + common opcodes               |
+- **Front end.**  A hand-written lexer and an error-tolerant parser produce a
+  typed AST with source positions; `gasm tokens` and `gasm parse` expose them
+  directly.
+- **Formatter.**  `gasm fmt` canonicalises indentation, operand spacing,
+  per-function mnemonic alignment and blank-line layout: `gofmt` for assembly,
+  operating recursively on directories the way `go fmt` does.  `-l` lists
+  files whose formatting differs and `-d` prints a unified diff.
+- **Linter.**  `gasm lint` runs 18 conservative static checks, among them
+  `undefined-label`, `abi-argsize` (declared argument area vs the `// func`
+  signature), `register-clobber` (Go ABI register liveness over the
+  control-flow graph), `stack-imbalance`, `abi0-register-args` and
+  `unencodable-instruction`.
+- **Standalone assembler.**  `gasm asm` encodes all four architectures without
+  the Go toolchain and writes raw images or linkable ELF objects (with DWARF5
+  debug sections) with no Go installation needed, or the Go toolchain's own
+  GOOBJ format, which needs the installed toolchain and which `go build`
+  consumes in place of the toolchain's output.  Framed functions get the
+  stack-split guard and the morestack block, byte-identical to the
+  toolchain's, so split functions link too.  The assembler preprocesses
+  like the toolchain (`#define`, `#include` with `-I`, `#ifdef`), generates
+  `go_asm.h` from the package's Go files, and carries `PCALIGN`, the
+  `LOCK`/`REP` prefixes and the literal-data pseudo-ops.
+- **Disassembler.**  `gasm dis` lists a `.s` file's functions at their real
+  offsets after assembling, or disassembles raw bytes from a file or stdin.
+- **Dynamic verification.**  `gasm verify` JIT-loads assembled functions into
+  executable memory: smoke calls, ABI checks (sentinel registers, red-zone
+  canary), differential fuzzing against the `go tool asm` build, and
+  byte-for-byte ground-truth comparison of the machine code.
+- **Debugger.**  `gasm debug` is a source-level ptrace debugger with
+  breakpoints (optionally conditional), hardware watchpoints, register and
+  memory inspection, and headless script runs that report instruction and
+  label coverage; it runs on Linux (all four architectures) and FreeBSD
+  (amd64, arm64, riscv64).
+- **Language server.**  `gasm lsp` serves completion, hover, document symbols,
+  push and pull diagnostics, semantic-token highlighting, go-to-definition,
+  find references, rename, formatting, inlay hints, code actions, signature
+  help, document highlights, workspace symbol search, #include document
+  links and folding ranges over stdio; definition, references and rename
+  work across every open document and the indexed workspace files beyond
+  them, and the quick fixes add a missing textflag.h include and set the
+  argument area from the // func signature.
+- **Comparators and audits.**  `gasm diff` compares the machine code of two
+  assembly files byte-for-byte, `gasm profile` shows basic-block structure,
+  `gasm audit-instructions` diffs the encoder against the installed toolchain,
+  and `gasm scaffold` generates a differential test skeleton for a kernel.
+
+### Architecture support
+
+Four architectures, the four that matter in practice:
+
+| Architecture | GOARCH      | File suffix  | Instructions recognised                     |
+|--------------|-------------|--------------|---------------------------------------------|
+| AMD64        | `amd64`     | `_amd64.s`   | 1600 + common opcodes + traditional aliases |
+| ARM64        | `arm64`     | `_arm64.s`   | 645 + common opcodes                        |
+| RISC-V       | `riscv64`   | `_riscv64.s` | 992 + common opcodes                        |
+| LoongArch    | `loong64`   | `_loong64.s` | 808 + common opcodes                        |
 
 "Common opcodes" are the instructions shared by every architecture (`RET`,
-`JMP`, `NOP`, `CALL`, `TEXT`, `FUNCDATA`, `PCDATA`, …).  AMD64 additionally
+`JMP`, `NOP`, `CALL`, `TEXT`, `FUNCDATA`, `PCDATA`, ...).  AMD64 additionally
 carries the traditional conditional-jump spellings (`JZ`, `JNZ`, `JA`, `JC`,
-…) that the assembler accepts as aliases.  Regenerating the tables is one
-command — `just gen` — and requires only a Go installation; the committed
-output has no runtime dependency on the toolchain.
+...) that the assembler accepts as aliases.  The tables are generated from
+the Go toolchain's own assembler source (`just gen` refreshes them), so
+every mnemonic the real assembler accepts is recognised; what the encoder
+can emit today is narrower, and a recognised but unencodable instruction is
+reported as an explicit error, never as a wrong byte.
 
-## Supported Platforms
+The same measurement runs over GOROOT's whole assembly corpus:
+`gasm audit-instructions --corpus` reports every real-code GOROOT assembly
+file (the tree without testdata) assembling for every target its build
+admits: 250 of 250, 100 %. Over the whole tree including testdata the
+measure is 271 of 322 attemptable (84.2 %); files named for other Go ports
+are counted but never attempted, and `//go:build` constraints decide which
+targets attempt a file at all, exactly as the build does. The number moves
+with every release.
 
-The toolkit runs on Linux.  All four Linux architectures are supported as
-hosts — amd64, arm64, riscv64 and loong64 — and the release matrix
-cross-compiles the same four targets.
+### Validation status
 
-**FreeBSD support is planned for a future release.**
+**Only amd64 is validated on real hardware.** The other three
+architectures are validated under qemu-user emulation, because the
+project owns no arm64, riscv64 or loong64 machine, and emulation is the
+only substitute available for the hardware. The distinction matters and
+is stated rather than implied: everything below is a claim about what has
+actually been executed.
 
-## Principles
+| Layer | amd64 | arm64, riscv64, loong64 |
+|---|---|---|
+| Encoding: byte-for-byte against `go tool asm` | native hardware | native hardware (the toolchain cross-assembles any GOARCH on any host) |
+| Execution: JIT calls, ABI checks, differential fuzzing | native hardware | qemu-user emulation |
+| Debugger: ptrace tracing, breakpoints, watchpoints, coverage | native hardware | emulation cannot run ptrace; the layer compiles and its architecture-neutral units run under `go test ./...`, nothing more. FreeBSD (amd64, arm64, riscv64) is in the same position: the port compiles behind the cross-build gate and its integration test is ready, but no FreeBSD machine has executed it |
 
-- **Pure Go and GAsm only.**  No C, no cgo, no external toolchains, no native
-  binaries, no JavaScript runtimes.  The parser is hand-written; there is no
-  parser generator.
-- **Self-contained.**  The toolkit's production code depends only on the
-  standard library; one binary, no runtime data files.  The single module
-  dependency, `golang.org/x/arch`, is used **only in tests** to validate the
-  instruction encoder by round-trip decoding — it is never linked into the
-  `gasm` binary.
-- **Linux-only.**  Runs natively on amd64, arm64, riscv64 and loong64 Linux
-  hosts; the release matrix cross-compiles the same four targets.  Latest
-  stable Go only.
-- **No vendor lock-in.**  The integration surface is the Language Server
-  Protocol and a command-line interface — both open standards.  No cloud
-  service, no proprietary API, no dependence on any one editor's internals.
-- **Complete and verifiable.**  Instruction coverage is generated from the
-  assembler's own source and regenerated on demand, so it cannot silently fall
-  behind the toolchain.
+Consequences, stated plainly. An emulator is a model of a CPU, not the
+CPU: instruction semantics are implemented in software and can differ
+from silicon in ways a test suite does not reveal. A kernel that passes
+under qemu-user is therefore not proven correct on real hardware, and a
+discrepancy found on real hardware is a defect in gasm, reported like any
+other. Encoding parity is the exception: the byte comparison against the
+toolchain runs on the host for every architecture, so no emulator stands
+between the claim and the evidence. The debugger is the weakest case: on
+the three emulated architectures its per-architecture ptrace code has
+been compiled and read, never executed. Its architecture-neutral units
+run under `go test ./...`, which the race workflow and a manual run
+perform; the default `just test` gate does not sweep `./debug/...`.
 
-## Components
+## The documentation goal
 
-| Package | Purpose |
-|---------|---------|
-| `token` | Lexical token kinds and source positions. |
-| `lexer` | Hand-written scanner for Plan 9 assembly. |
-| `ast` | The abstract syntax tree. |
-| `parser` | Line-oriented, error-tolerant parser producing the AST. |
-| `arch` | amd64, arm64, riscv64 and loong64 register files and instruction tables. |
-| `lint` | Conservative static checks. |
-| `format` | A canonical formatter — `gofmt` for assembly. |
-| `asm` | The standalone assembler: amd64, RISC-V and LoongArch encoders, linker, object-file emitters (ELF, GOOBJ). |
-| `verify` | JIT execution substrate for dynamic analysis, combined ABI+fuzz differential testing. |
-| `debug` | Interactive ptrace debugger with GPR/YMM register display and named buffer allocation. |
-| `lsp` | Language Server Protocol server. |
-| `cmd/gasm` | The `gasm` binary tying it all together. |
-| `_gen` | The generator that rebuilds the instruction tables from the Go toolchain. |
+The toolkit is the primary goal. The secondary one is documentation: a
+specification of the Plan 9 assembly language and of the GOOBJ object
+format that is 100 % complete, detailed enough to implement against,
+and written to a professional standard. These are the two subjects this
+project works with every day, and they are the two for which no usable
+documentation exists.
 
-See [`docs/ARCHITECTURE.md`](docs/ARCHITECTURE.md) for the design rationale and
-data flow, and [`docs/DECISIONS.md`](docs/DECISIONS.md) for design decisions
-deliberately postponed (with the analysis needed to pick them up again).
+Go documents the language on a single page, "A Quick Guide to Go's
+Assembler", which carries no section for loong64, one of the four
+architectures gasm supports, and covers a fraction of what each
+assembler accepts. What exists beyond it lives as comments inside the
+toolchain's internal source: per-architecture reference manuals for
+arm64, ppc64, riscv64 and loong64, written for the toolchain's own
+maintainers rather than for an outside reader, and none at all for
+amd64. GOOBJ fares worst of all. The format that `go build` consumes
+has no specification anywhere: it is described by a comment in an
+internal package, it is not a stable interface, and it can change with
+any toolchain release.
+
+The gap is therefore filled the only way it can be filled: by reverse
+engineering the toolchain itself, the same work the encoders already
+perform. Most of the documentation can come from nowhere else, and it
+is written as that knowledge is produced during development. It is
+verified the way the code is verified: an encoding documented here is
+one that differential tests against `go tool asm` confirm
+byte-for-byte, and a format field documented here is one the linker
+demonstrably reads. The work has begun: [docs/GOOBJ.md](docs/GOOBJ.md)
+specifies the object file format completely, and
+[docs/asm/README.md](docs/asm/README.md) opens the language reference
+with its common core. The per-architecture pages follow.
+
+## Direction
+
+The plan, in the order it is being worked:
+
+- **Extended instruction support.**  Two layers. First, encoding
+  coverage for every mnemonic the Go toolchain itself accepts, closed in
+  order of how often real code needs each instruction;
+  `gasm audit-instructions` measures the gap. Second, the larger work:
+  an extended instruction set the toolchain does not know at all. The
+  toolchain-derived tables stay generated and untouched; only the
+  extended instructions are hand-maintained, with their own spellings
+  and encoders, verified by execution (on real hardware for amd64, under
+  emulation for the rest, per the validation status above) because the
+  toolchain offers no ground truth to compare against. The gaps exist
+  on every architecture, amd64 included.
+- **Full GOOBJ and ELF compilation.**  The destination is a complete,
+  standalone compilation path: linkable ELF objects for consumers outside
+  Go, and GOOBJ objects that `go build` links directly.  Through GOOBJ, a
+  Go program will be able to use machine instructions that the Go
+  toolchain itself does not support; through ELF, Plan 9 assembly becomes
+  usable outside Go entirely.
+- **Platforms: Linux and FreeBSD.**  Linux is supported today on all four
+  architectures and is where the binary builds.  FreeBSD follows on amd64,
+  arm64 and riscv64: the JIT's executable-memory mapping and the ptrace
+  debugger layer are ported (the debugger's live validation awaits a
+  FreeBSD machine, as the validation status states).  Other unix systems
+  may follow those two.
+- **Four architectures, no more.**  amd64, arm64, riscv64 and loong64.
+  No others are planned.
+
+## Install
+
+Prebuilt binaries for linux/amd64, linux/arm64, linux/riscv64 and
+linux/loong64 are on the
+[releases page](https://sourcedock.dev/petrbalvin/gasm-sdk/releases).
+From source (Go 1.27.1):
+
+```sh
+go install sourcedock.dev/petrbalvin/gasm-sdk/cmd/gasm@latest
+```
+
+Or from a repository checkout:
+
+```sh
+just install
+```
+
+The installed binary reports the version the toolchain recorded: the tag
+on a tagged checkout, a pseudo-version naming the commit below one.
 
 ## Quick start
 
 ```sh
-just install     # download dependencies (there are none)
-just build       # go vet + gofmt check — zero errors, zero warnings
-just test        # full suite, race detector, 80 % coverage gate
-just fmt         # gofmt the tree
-just gen         # regenerate the instruction tables from the Go toolchain
+cat > hello_amd64.s <<'EOF'
+#include "textflag.h"
+
+// func add(a, b int) int
+TEXT ·add(SB), NOSPLIT, $0-24
+	MOVQ a+0(FP), AX
+	ADDQ b+8(FP), AX
+	MOVQ AX, ret+16(FP)
+	RET
+EOF
+
+gasm lint hello_amd64.s                         # static checks
+gasm asm -o hello.bin hello_amd64.s             # assemble to a raw image
+gasm verify --call add --args a=2,b=3 hello_amd64.s   # JIT-call it with arguments
 ```
 
-Install the binary and use it:
+## Usage
 
 ```sh
-just install-bin                    # installs gasm into $GOBIN
-
-gasm --help                         # overview of commands and flags
-gasm tokens kernel_amd64.s          # dump the token stream
-gasm parse  kernel_amd64.s          # parse, report syntax errors
-gasm fmt    -w kernel_amd64.s       # canonicalise in place
-gasm fmt                            # reformat every .s below here, like go fmt
-gasm lint   *.s                     # static checks
-gasm asm --format elf -o k.o k.s    # assemble to a linkable ELF object
-gasm verify kernel_amd64.s           # JIT-load and report functions
-gasm verify --ground-truth k.s      # byte-for-byte vs go tool asm
-gasm verify --call decodeBlockAVX2 --buf src:64:hex...,dst:256:zero k.s
-gasm debug --func name k.s           # interactive debugger
-gasm diff a.s b.s                    # compare machine code byte-for-byte
+gasm fmt                                # reformat every .s below here, like go fmt
+gasm fmt    -w kernel_amd64.s           # canonicalise one file in place
+gasm fmt    -l *.s                      # list files whose formatting differs
+gasm fmt    -d kernel_amd64.s           # print a unified diff instead
+gasm lint   *.s                         # static checks
+gasm asm --format elf -o k.o k.s        # assemble to a linkable ELF object
+gasm asm --format goobj -p pkg/path -o k.o k.s   # Go object, consumed by go build
+gasm dis k.s                            # assemble, then list each function
+gasm dis -a amd64 - < dump.bin          # disassemble raw bytes from stdin
+gasm verify --ground-truth k.s          # byte-for-byte vs go tool asm
+gasm verify --fuzz k.s                  # differential fuzz vs the go tool asm build
+gasm debug --func name k.s              # interactive debugger
+gasm debug --func name --script cmds.txt --timeout 30s k.s   # headless run
+gasm debug --func name --cover k.s      # instruction and label coverage
+gasm diff a.s b.s                       # compare machine code byte-for-byte
 gasm diff --map wideCopyAVX2=wideCopyAVX512 avx2.s avx512.s
-gasm profile k.s                     # show basic-block structure
+gasm profile k.s                        # show basic-block structure
+gasm audit-instructions                 # encoder vs go tool asm name diff
+gasm scaffold differential k.s          # generate a differential test skeleton
 ```
 
-See [CONTRIBUTING.md](CONTRIBUTING.md) for the full development workflow,
-[docs/CLI.md](docs/CLI.md) for the command reference, and
-[docs/DEVELOPMENT.md](docs/DEVELOPMENT.md) for setup and recipes.
+Run `gasm --help` for the command overview and `gasm <command> -h` for a
+command's flags.  [docs/CLI.md](docs/CLI.md) is the full reference.
 
-## Editor integration
+### Editor integration
 
 `gasm lsp` speaks the Language Server Protocol over standard input/output, so
-any LSP-capable editor can use it — point your editor's LSP client at the
+any LSP-capable editor can use it: point your editor's LSP client at the
 binary and associate it with `.s` files.  Syntax highlighting is delivered as
-**LSP semantic tokens**, so no editor-specific grammar is required.  The server
+LSP semantic tokens, so no editor-specific grammar is required.  The server
 infers the target architecture from the file-name suffix
 (`_amd64.s` / `_arm64.s` / `_riscv64.s` / `_loong64.s`).
 
-## License
+## Development
+
+```sh
+just build        # compile, zero errors and zero warnings
+just test         # the suite, no cache, the 80 % coverage floor
+just gates        # build, fmt-check, vet, test, race: the definition of done
+just fmt          # gofmt the tree
+just gen          # regenerate the instruction tables from the Go toolchain
+```
+
+See [CONTRIBUTING.md](CONTRIBUTING.md) for the development workflow and
+[docs/DEVELOPMENT.md](docs/DEVELOPMENT.md) for setup details and every
+recipe.
+
+## Documentation
+
+- [docs/CLI.md](docs/CLI.md): full command reference
+- man pages: `just install-man` installs gasm(1) and one page per command
+  except `version`, which is documented inside gasm(1) instead, into
+  ~/.local/share/man (MANDIR overrides); `just uninstall-man` removes
+  them
+- [docs/ARCHITECTURE.md](docs/ARCHITECTURE.md): components and data flow
+- [docs/GOOBJ.md](docs/GOOBJ.md): the GOOBJ object file format specification
+- [docs/asm/](docs/asm/README.md): the Plan 9 assembly language reference
+- [docs/DEVELOPMENT.md](docs/DEVELOPMENT.md): development setup and recipes
+- [CHANGELOG.md](CHANGELOG.md): release history
+
+## Licence
+
+BSD-3-Clause; see [LICENSE](LICENSE).
 
-BSD-3-Clause — see [LICENSE](LICENSE).  
 Copyright © 2026 [Petr Balvín](https://petrbalvin.org)

SECURITY.md

@@ -0,0 +1,41 @@
+# Security policy
+
+## Supported versions
+
+Security fixes go to the newest release and to the `development` branch. Older
+releases do not receive them.
+
+| Version | Supported |
+|---|---|
+| 0.35.0 | yes |
+| older releases | no |
+
+## Reporting a vulnerability
+
+**Do not open a public issue for a security problem.** A public report tells everyone
+about the flaw before there is a fix. Report it privately to
+**opensource@petrbalvin.org**.
+
+Include:
+
+- the version or commit you tested, and the platform
+- what the problem is, and what an attacker gains from it
+- the smallest reproducer you have, ideally a test or a single command
+- a suggested fix, if you have one
+
+## What to expect
+
+- A human reads the report, and you get an acknowledgement.
+- You are kept informed while the fix is being made, and told when it ships.
+- The fix is released before the details are published, and the timing is agreed with
+  you.
+- The fix ships without naming you: the project keeps no credits list, so the release
+  notes, the changelog and the commits name no reporter.
+
+## Out of scope
+
+- Findings that require the attacker to already run code as the user, or to have local
+  access.
+- Missing hardening with no demonstrated impact.
+- Flaws in a third-party dependency: report them to that project, and to this one only
+  when this project's use of it makes them reachable.

_gen/gen.go

@@ -5,9 +5,13 @@
 // toolchain's own assembler source.  Go's Plan 9 assembler defines the exact,
 // complete set of mnemonics it accepts for each architecture in
 // $GOROOT/src/cmd/internal/obj/<arch>/anames.go; this tool extracts those
-// names so gasm-devkit supports every instruction the real assembler does,
+// names so gasm-sdk supports every instruction the real assembler does,
 // with no hand-maintained (and therefore inevitably incomplete) lists.
 //
+// The same data feeds the generated instruction appendices of the assembly
+// language reference, docs/asm/INSTRUCTIONS-<ARCH>.md, so that the reference
+// cannot drift from the tables it documents.
+//
 // Usage (via the justfile):
 //
 //	just gen
@@ -26,9 +30,12 @@ import (
 	"path/filepath"
 	"sort"
 	"strings"
+
+	"sourcedock.dev/petrbalvin/gasm-sdk/arch"
+	"sourcedock.dev/petrbalvin/gasm-sdk/asm"
 )
 
-// archDirs maps a gasm-devkit architecture name to its obj sub-directory.
+// archDirs maps a gasm-sdk architecture name to its obj sub-directory.
 var archDirs = []struct {
 	arch string
 	sub  string
@@ -39,11 +46,30 @@ var archDirs = []struct {
 	{"loong64", "loong64"},
 }
 
+// docPages maps an architecture to its generated appendix in the language
+// reference.  The amd64 page carries a per-mnemonic encodability column,
+// decided by asm.Encodable, which mirrors the encoder's own dispatch; the
+// other targets have no single cheap predicate, so their pages carry the
+// inventory and point at the live measurement instead.
+var docPages = []struct {
+	arch      arch.Arch
+	title     string
+	file      string
+	anames    string
+	encodable bool
+}{
+	{arch.AMD64, "AMD64", "INSTRUCTIONS-AMD64.md", "cmd/internal/obj/x86/anames.go", true},
+	{arch.ARM64, "ARM64", "INSTRUCTIONS-ARM64.md", "cmd/internal/obj/arm64/anames.go", false},
+	{arch.RISCV, "RISC-V 64", "INSTRUCTIONS-RISCV64.md", "cmd/internal/obj/riscv/anames.go", false},
+	{arch.LOONG64, "LoongArch 64", "INSTRUCTIONS-LOONG64.md", "cmd/internal/obj/loong64/anames.go", false},
+}
+
 func main() {
 	goroot := strings.TrimSpace(runGoEnvGOROOT())
 	if goroot == "" {
 		fatal("could not determine GOROOT")
 	}
+	version := strings.TrimSpace(runGoEnv("GOVERSION"))
 	// The common opcodes shared by every architecture (RET, JMP, NOP, CALL,
 	// TEXT, FUNCDATA, …) live in cmd/internal/obj/util.go.
 	commonPath := filepath.Join(goroot, "src", "cmd", "internal", "obj", "util.go")
@@ -57,16 +83,24 @@ func main() {
 	}
 	fmt.Printf("%-8s %4d instructions -> arch/common_gen.go\n", "common", len(common))
 
+	names := map[string][]string{}
 	for _, a := range archDirs {
 		path := filepath.Join(goroot, "src", "cmd", "internal", "obj", a.sub, "anames.go")
-		names, err := extractInstrs(path)
+		names[a.arch], err = extractInstrs(path)
 		if err != nil {
 			fatal("extract %s: %v", a.arch, err)
 		}
-		if err := writeGen(a.arch, a.sub, names); err != nil {
+		if err := writeGen(a.arch, a.sub, names[a.arch]); err != nil {
 			fatal("write %s: %v", a.arch, err)
 		}
-		fmt.Printf("%-8s %4d instructions -> arch/%s_gen.go\n", a.arch, len(names), a.arch)
+		fmt.Printf("%-8s %4d instructions -> arch/%s_gen.go\n", a.arch, len(names[a.arch]), a.arch)
+	}
+
+	for _, p := range docPages {
+		if err := writeDocPage(p.arch, p.title, p.file, p.anames, version, p.encodable); err != nil {
+			fatal("write %s: %v", p.file, err)
+		}
+		fmt.Printf("%-8s -> docs/asm/%s\n", p.arch, p.file)
 	}
 }
 
@@ -85,7 +119,7 @@ func filterCommon(names []string) []string {
 // writeCommon emits arch/common_gen.go.
 func writeCommon(names []string) error {
 	var b strings.Builder
-	b.WriteString("// Code generated by gasm-devkit _gen; DO NOT EDIT.\n")
+	b.WriteString("// Code generated by gasm-sdk _gen; DO NOT EDIT.\n")
 	b.WriteString("// Source: cmd/internal/obj/util.go from the Go toolchain.\n")
 	b.WriteString("//\n")
 	b.WriteString("// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)\n")
@@ -156,7 +190,7 @@ func stringLit(elt ast.Expr) string {
 // writeGen emits arch/<arch>_gen.go.
 func writeGen(arch, sub string, names []string) error {
 	var b strings.Builder
-	b.WriteString("// Code generated by gasm-devkit _gen; DO NOT EDIT.\n")
+	b.WriteString("// Code generated by gasm-sdk _gen; DO NOT EDIT.\n")
 	b.WriteString("// Source: cmd/internal/obj/" + sub + "/anames.go from the Go toolchain.\n")
 	b.WriteString("//\n")
 	b.WriteString("// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)\n")
@@ -172,6 +206,61 @@ func writeGen(arch, sub string, names []string) error {
 	return os.WriteFile(filepath.Join("arch", arch+"_gen.go"), []byte(b.String()), 0o644)
 }
 
+// writeDocPage emits docs/asm/<file>, the generated instruction appendix of
+// the language reference for one architecture: every mnemonic the toolchain
+// accepts, with the curated summary where the architecture table carries one
+// and, on amd64, a per-mnemonic encodability column.
+func writeDocPage(a arch.Arch, title, file, anames, version string, encodable bool) error {
+	table := arch.ForArch(a)
+	instrs := table.Instructions()
+
+	var b strings.Builder
+	b.WriteString("# " + title + ": instruction inventory\n\n")
+	b.WriteString("Generated by gasm-sdk's `_gen` from the Go toolchain's instruction table\n")
+	b.WriteString("(`" + anames + "`, " + version + "); DO NOT EDIT. This page lists every mnemonic\n")
+	b.WriteString("`go tool asm` accepts on this target, which is the upper bound of the\n")
+	b.WriteString("language on it: a name absent here is not an instruction of the target,\n")
+	b.WriteString("and a name present here may still be one gasm's encoder cannot emit yet.\n\n")
+
+	encodableCount := 0
+	if encodable {
+		b.WriteString("The `gasm encodes` column reports whether gasm's encoder can emit the\n")
+		b.WriteString("mnemonic today; the gap is the encoder backlog, measured live by\n")
+		b.WriteString("`gasm audit-instructions`.\n\n")
+		b.WriteString("| Mnemonic | gasm encodes | Notes |\n")
+		b.WriteString("|---|---|---|\n")
+		for _, in := range instrs {
+			ok := asm.Encodable(in.Name)
+			if ok {
+				encodableCount++
+			}
+			b.WriteString("| `" + in.Name + "` | " + yesNo(ok) + " | " + in.Summary + " |\n")
+		}
+		b.WriteString("\n")
+		fmt.Fprintf(&b, "Recognised: %d mnemonics. gasm encodes: %d.\n", len(instrs), encodableCount)
+	} else {
+		b.WriteString("The inventory carries no per-mnemonic encoder column: on this target\n")
+		b.WriteString("encodability is decided per operand shape, and the live measured\n")
+		b.WriteString("coverage is reported by `gasm audit-instructions`.\n\n")
+		b.WriteString("| Mnemonic | Notes |\n")
+		b.WriteString("|---|---|\n")
+		for _, in := range instrs {
+			b.WriteString("| `" + in.Name + "` | " + in.Summary + " |\n")
+		}
+		b.WriteString("\n")
+		fmt.Fprintf(&b, "Recognised: %d mnemonics.\n", len(instrs))
+	}
+	return os.WriteFile(filepath.Join("docs", "asm", file), []byte(b.String()), 0o644)
+}
+
+// yesNo renders a boolean as the word the appendix tables use.
+func yesNo(v bool) string {
+	if v {
+		return "yes"
+	}
+	return "no"
+}
+
 func runGoEnvGOROOT() string {
 	out, err := exec.Command("go", "env", "GOROOT").Output()
 	if err != nil {
@@ -180,6 +269,15 @@ func runGoEnvGOROOT() string {
 	return string(out)
 }
 
+// runGoEnv runs `go env` for a single variable.
+func runGoEnv(name string) string {
+	out, err := exec.Command("go", "env", name).Output()
+	if err != nil {
+		return ""
+	}
+	return string(out)
+}
+
 func fatal(format string, args ...any) {
 	fmt.Fprintf(os.Stderr, "gen: "+format+"\n", args...)
 	os.Exit(1)

arch/amd64.go

@@ -70,6 +70,10 @@ func amd64Registers() []Register {
 	for i := 0; i <= 7; i++ {
 		add(fmt.Sprintf("K%d", i), Mask, "AVX-512 mask register")
 	}
+	// x87 stack registers (FMOVD and the other x87 moves).
+	for i := 0; i <= 7; i++ {
+		add(fmt.Sprintf("F%d", i), Float, "x87 stack register")
+	}
 	return regs
 }
 
@@ -270,6 +274,8 @@ func amd64Curated() []Instr {
 		"VMINPD", "VMINPS", "VMINSD", "VMINSS", "VMAXPD", "VMAXPS", "VMAXSD", "VMAXSS",
 		"VXORPD", "VXORPS", "VANDPD", "VANDPS", "VANDNPD", "VANDNPS", "VORPD", "VORPS",
 		"VUNPCKHPD", "VUNPCKLPD", "VUNPCKHPS", "VUNPCKLPS",
+		"PSHUFD", "PSHUFHW", "PSHUFLW", "SHUFPS", "SHUFPD",
+		"UNPCKLPS", "UNPCKHPS", "UNPCKLPD", "UNPCKHPD",
 		"VSQRTPD", "VSQRTPS", "VSQRTSD", "VSQRTSS", "VRSQRTPS", "VRCPPS",
 		"VCMPPD", "VCMPPS", "VCMPSD", "VCMPSS",
 	} {

arch/amd64_gen.go

@@ -1,4 +1,4 @@
-// Code generated by gasm-devkit _gen; DO NOT EDIT.
+// Code generated by gasm-sdk _gen; DO NOT EDIT.
 // Source: cmd/internal/obj/x86/anames.go from the Go toolchain.
 //
 // Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)

arch/arch.go

@@ -55,6 +55,7 @@ const (
 	Mask                    // AVX-512 mask register (K)
 	Float                   // arm64 floating-point register (F)
 	VecARM                  // arm64 SIMD/vector register (V)
+	VecSIMD                 // architecture-neutral SIMD/vector register (LoongArch LSX/LASX)
 	Special                 // architecture-special register
 )
 
@@ -73,6 +74,8 @@ func (c RegClass) String() string {
 		return "float"
 	case VecARM:
 		return "vector (arm64)"
+	case VecSIMD:
+		return "vector"
 	case Special:
 		return "special"
 	default:

arch/arm64.go

@@ -29,10 +29,11 @@ func arm64Registers() []Register {
 		regs = append(regs, Register{Name: name, Class: class, Desc: desc})
 	}
 
-	// General-purpose integer registers R0–R30.
+	// General-purpose integer registers R0-R30.
 	for i := 0; i <= 30; i++ {
 		add(fmt.Sprintf("R%d", i), GPR, "64-bit general-purpose register")
 	}
+	add("R18_PLATFORM", GPR, "R18 under its toolchain-reserved Windows name (an alias of R18)")
 	add("ZR", Special, "zero register (reads as 0)")
 	add("SP", Special, "stack pointer")
 	add("LR", Special, "link register (alias of R30)")
@@ -145,11 +146,38 @@ func arm64Curated() []Instr {
 	for _, op := range []string{
 		"LDAXR", "LDAXRB", "LDAXRH", "LDAXRW", "STXR", "STXRB", "STXRH", "STXRW",
 		"LDAR", "LDARB", "LDARH", "LDARW", "STLR", "STLRB", "STLRH", "STLRW",
-		"LDADD", "LDCLR", "LDEOR", "LDSET", "SWP", "CAS", "CASAL", "CASL", "CASAL",
+		"LDADD", "LDCLR", "LDEOR", "LDSET", "SWP", "CAS", "CASAL", "CASL",
 	} {
 		t = append(t, i(op, "Atomic memory operation"))
 	}
 
+	// Register-pair loads and stores.
+	for _, op := range []string{"LDP", "STP", "LDPW", "STPW", "FLDPD", "FSTPD"} {
+		t = append(t, ic(op, "Register-pair load or store", 2, 2))
+	}
+
+	// Cache maintenance and prefetch.
+	t = append(t, i("DC", "Data cache maintenance"))
+	t = append(t, i("PRFM", "Memory prefetch"))
+	for _, op := range []string{"LDADDAL", "LDCLRAL", "LDORAL", "SWPAL"} {
+		t = append(t, i(op, "Atomic memory operation with acquire and release semantics"))
+	}
+
+	// Cryptographic extensions.
+	for _, op := range []string{"AESE", "AESD", "AESMC", "AESIMC"} {
+		t = append(t, i(op, "AES round"))
+	}
+	for _, op := range []string{
+		"SHA1C", "SHA1P", "SHA1M", "SHA1H", "SHA1SU0", "SHA1SU1",
+		"SHA256H", "SHA256H2", "SHA256SU0", "SHA256SU1",
+		"SHA512H", "SHA512H2", "SHA512SU0", "SHA512SU1",
+	} {
+		t = append(t, i(op, "SHA round"))
+	}
+	for _, op := range []string{"VEOR3", "VBCAX", "VXAR", "VRAX1"} {
+		t = append(t, i(op, "Three-way XOR / rotate crypto vector operation"))
+	}
+
 	// Floating-point scalar.
 	for _, op := range []string{
 		"FADD", "FSUB", "FMUL", "FDIV", "FNEG", "FABS", "FSQRT", "FMIN", "FMAX",

arch/arm64_gen.go

@@ -1,4 +1,4 @@
-// Code generated by gasm-devkit _gen; DO NOT EDIT.
+// Code generated by gasm-sdk _gen; DO NOT EDIT.
 // Source: cmd/internal/obj/arm64/anames.go from the Go toolchain.
 //
 // Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
@@ -364,6 +364,8 @@ var arm64GeneratedInstrs = []string{
 	"REVW",
 	"ROR",
 	"RORW",
+	"RPRFM",
+	"SB",
 	"SBC",
 	"SBCS",
 	"SBCSW",
@@ -477,23 +479,68 @@ var arm64GeneratedInstrs = []string{
 	"UXTH",
 	"UXTHW",
 	"UXTW",
+	"VABS",
 	"VADD",
 	"VADDP",
 	"VADDV",
 	"VAND",
 	"VBCAX",
+	"VBIC",
 	"VBIF",
 	"VBIT",
 	"VBSL",
+	"VCLS",
+	"VCLZ",
 	"VCMEQ",
+	"VCMGE",
+	"VCMGT",
+	"VCMHI",
+	"VCMHS",
+	"VCMLE",
+	"VCMLT",
 	"VCMTST",
 	"VCNT",
 	"VDUP",
 	"VEOR",
 	"VEOR3",
 	"VEXT",
+	"VFABS",
+	"VFADD",
+	"VFADDP",
+	"VFCMEQ",
+	"VFCMGE",
+	"VFCMGT",
+	"VFCMLE",
+	"VFCMLT",
+	"VFCVTL",
+	"VFCVTL2",
+	"VFCVTN",
+	"VFCVTN2",
+	"VFCVTZS",
+	"VFCVTZU",
+	"VFDIV",
+	"VFMAX",
+	"VFMAXNM",
+	"VFMAXNMP",
+	"VFMAXNMV",
+	"VFMAXP",
+	"VFMAXV",
+	"VFMIN",
+	"VFMINNM",
+	"VFMINNMP",
+	"VFMINNMV",
+	"VFMINP",
+	"VFMINV",
 	"VFMLA",
 	"VFMLS",
+	"VFMUL",
+	"VFNEG",
+	"VFRINTM",
+	"VFRINTN",
+	"VFRINTP",
+	"VFRINTZ",
+	"VFSQRT",
+	"VFSUB",
 	"VLD1",
 	"VLD1R",
 	"VLD2",
@@ -502,11 +549,17 @@ var arm64GeneratedInstrs = []string{
 	"VLD3R",
 	"VLD4",
 	"VLD4R",
+	"VMLA",
+	"VMLS",
 	"VMOV",
 	"VMOVD",
 	"VMOVI",
 	"VMOVQ",
 	"VMOVS",
+	"VMUL",
+	"VNEG",
+	"VNOT",
+	"VORN",
 	"VORR",
 	"VPMULL",
 	"VPMULL2",
@@ -515,14 +568,47 @@ var arm64GeneratedInstrs = []string{
 	"VREV16",
 	"VREV32",
 	"VREV64",
+	"VSCVTF",
+	"VSHADD",
 	"VSHL",
+	"VSHRN",
+	"VSHRN2",
 	"VSLI",
+	"VSMAX",
+	"VSMAXP",
+	"VSMAXV",
+	"VSMIN",
+	"VSMINP",
+	"VSMINV",
+	"VSMLAL",
+	"VSMLAL2",
+	"VSMLSL",
+	"VSMLSL2",
+	"VSMULL",
+	"VSMULL2",
+	"VSQABS",
+	"VSQADD",
+	"VSQNEG",
+	"VSQSHL",
+	"VSQSUB",
+	"VSQXTN",
+	"VSQXTN2",
+	"VSQXTUN",
+	"VSQXTUN2",
+	"VSRHADD",
 	"VSRI",
+	"VSRSHR",
+	"VSSHL",
+	"VSSHLL",
+	"VSSHLL2",
+	"VSSHR",
 	"VST1",
 	"VST2",
 	"VST3",
 	"VST4",
 	"VSUB",
+	"VSXTL",
+	"VSXTL2",
 	"VTBL",
 	"VTBX",
 	"VTRN1",
@@ -530,8 +616,27 @@ var arm64GeneratedInstrs = []string{
 	"VUADDLV",
 	"VUADDW",
 	"VUADDW2",
+	"VUCVTF",
+	"VUHADD",
 	"VUMAX",
+	"VUMAXP",
+	"VUMAXV",
 	"VUMIN",
+	"VUMINP",
+	"VUMINV",
+	"VUMLAL",
+	"VUMLAL2",
+	"VUMLSL",
+	"VUMLSL2",
+	"VUMULL",
+	"VUMULL2",
+	"VUQADD",
+	"VUQSHL",
+	"VUQSUB",
+	"VUQXTN",
+	"VUQXTN2",
+	"VURHADD",
+	"VUSHL",
 	"VUSHLL",
 	"VUSHLL2",
 	"VUSHR",
@@ -541,6 +646,8 @@ var arm64GeneratedInstrs = []string{
 	"VUZP1",
 	"VUZP2",
 	"VXAR",
+	"VXTN",
+	"VXTN2",
 	"VZIP1",
 	"VZIP2",
 	"WFE",

arch/common_gen.go

@@ -1,4 +1,4 @@
-// Code generated by gasm-devkit _gen; DO NOT EDIT.
+// Code generated by gasm-sdk _gen; DO NOT EDIT.
 // Source: cmd/internal/obj/util.go from the Go toolchain.
 //
 // Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)

arch/loong64.go

@@ -31,10 +31,10 @@ func loong64Registers() []Register {
 		add(fmt.Sprintf("F%d", i), Float, "floating-point register")
 	}
 	for i := 0; i <= 31; i++ {
-		add(fmt.Sprintf("V%d", i), VecARM, "LSX 128-bit vector register")
+		add(fmt.Sprintf("V%d", i), VecSIMD, "LSX 128-bit vector register")
 	}
 	for i := 0; i <= 31; i++ {
-		add(fmt.Sprintf("X%d", i), VecARM, "LASX 256-bit vector register")
+		add(fmt.Sprintf("X%d", i), VecSIMD, "LASX 256-bit vector register")
 	}
 	return regs
 }

arch/loong64_gen.go

@@ -1,4 +1,4 @@
-// Code generated by gasm-devkit _gen; DO NOT EDIT.
+// Code generated by gasm-sdk _gen; DO NOT EDIT.
 // Source: cmd/internal/obj/loong64/anames.go from the Go toolchain.
 //
 // Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
@@ -152,6 +152,8 @@ var loong64GeneratedInstrs = []string{
 	"FNMADDF",
 	"FNMSUBD",
 	"FNMSUBF",
+	"FRINTD",
+	"FRINTF",
 	"FSCALEBD",
 	"FSCALEBF",
 	"FSEL",
@@ -177,7 +179,10 @@ var loong64GeneratedInstrs = []string{
 	"FTINTWF",
 	"JIRL",
 	"LL",
+	"LLACQV",
+	"LLACQW",
 	"LLV",
+	"LLW",
 	"LU12IW",
 	"LU32ID",
 	"LU52ID",
@@ -248,7 +253,11 @@ var loong64GeneratedInstrs = []string{
 	"ROTR",
 	"ROTRV",
 	"SC",
+	"SCQ",
+	"SCRELV",
+	"SCRELW",
 	"SCV",
+	"SCW",
 	"SGT",
 	"SGTU",
 	"SLL",

arch/riscv_gen.go

@@ -1,4 +1,4 @@
-// Code generated by gasm-devkit _gen; DO NOT EDIT.
+// Code generated by gasm-sdk _gen; DO NOT EDIT.
 // Source: cmd/internal/obj/riscv/anames.go from the Go toolchain.
 //
 // Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
@@ -81,6 +81,9 @@ var riscvGeneratedInstrs = []string{
 	"CLD",
 	"CLDSP",
 	"CLI",
+	"CLMUL",
+	"CLMULH",
+	"CLMULR",
 	"CLUI",
 	"CLW",
 	"CLWSP",
@@ -95,13 +98,20 @@ var riscvGeneratedInstrs = []string{
 	"CSDSP",
 	"CSLLI",
 	"CSRAI",
+	"CSRC",
+	"CSRCI",
 	"CSRLI",
+	"CSRR",
 	"CSRRC",
 	"CSRRCI",
 	"CSRRS",
 	"CSRRSI",
 	"CSRRW",
 	"CSRRWI",
+	"CSRS",
+	"CSRSI",
+	"CSRW",
+	"CSRWI",
 	"CSUB",
 	"CSUBW",
 	"CSW",
@@ -259,6 +269,7 @@ var riscvGeneratedInstrs = []string{
 	"ORCB",
 	"ORI",
 	"ORN",
+	"PAUSE",
 	"RDCYCLE",
 	"RDINSTRET",
 	"RDTIME",
@@ -322,6 +333,8 @@ var riscvGeneratedInstrs = []string{
 	"VADDVI",
 	"VADDVV",
 	"VADDVX",
+	"VANDNVV",
+	"VANDNVX",
 	"VANDVI",
 	"VANDVV",
 	"VANDVX",
@@ -329,8 +342,17 @@ var riscvGeneratedInstrs = []string{
 	"VASUBUVX",
 	"VASUBVV",
 	"VASUBVX",
+	"VBREV8V",
+	"VBREVV",
+	"VCLMULHVV",
+	"VCLMULHVX",
+	"VCLMULVV",
+	"VCLMULVX",
+	"VCLZV",
 	"VCOMPRESSVM",
 	"VCPOPM",
+	"VCPOPV",
+	"VCTZV",
 	"VDIVUVV",
 	"VDIVUVX",
 	"VDIVVV",
@@ -743,10 +765,16 @@ var riscvGeneratedInstrs = []string{
 	"VREMUVX",
 	"VREMVV",
 	"VREMVX",
+	"VREV8V",
 	"VRGATHEREI16VV",
 	"VRGATHERVI",
 	"VRGATHERVV",
 	"VRGATHERVX",
+	"VROLVV",
+	"VROLVX",
+	"VRORVI",
+	"VRORVV",
+	"VRORVX",
 	"VRSUBVI",
 	"VRSUBVX",
 	"VS1RV",
@@ -950,6 +978,9 @@ var riscvGeneratedInstrs = []string{
 	"VWMULVX",
 	"VWREDSUMUVS",
 	"VWREDSUMVS",
+	"VWSLLVI",
+	"VWSLLVV",
+	"VWSLLVX",
 	"VWSUBUVV",
 	"VWSUBUVX",
 	"VWSUBUWV",

asm/aarch64_goobj_test.go

@@ -4,13 +4,14 @@
 package asm
 
 import (
+	"encoding/binary"
 	"os"
 	"os/exec"
 	"path/filepath"
 	"strings"
 	"testing"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
 )
 
 // TestGOObjectAARCH64Structure checks the basic structure of the emitted
@@ -61,6 +62,62 @@ TEXT ·add(SB), NOSPLIT, $0-24
 	}
 }
 
+// TestGOObjectAARCH64PairReloc pins the ADRP-pair relocation shape against
+// the toolchain's own object for the same source: exactly one R_ADDRARM64
+// of Siz 8 at the ADRP word (cmd/internal/obj/arm64/asm7.go adds a single
+// Siz-8 relocation per pair and the linker patches both instructions from
+// it).  gasm's assembler records the ADRP+ADD form as two word relocs; the
+// emitter must coalesce them, not emit two Siz-4 records.
+func TestGOObjectAARCH64PairReloc(t *testing.T) {
+	f, errs := parser.Parse("gv_arm64.s", `
+#include "textflag.h"
+
+TEXT ·getv(SB), NOSPLIT, $0-8
+	MOVD $v<>(SB), R4
+	MOVD R4, ret+0(FP)
+	RET
+
+GLOBL v<>(SB), RODATA, $8
+DATA v<>+0(SB)/8, $7
+`)
+	if len(errs) > 0 {
+		t.Fatalf("parse: %v", errs)
+	}
+	img, err := AssembleFileARM64(f)
+	if err != nil {
+		t.Fatalf("AssembleFileARM64: %v", err)
+	}
+	obj, err := img.GOObjectAARCH64("main", "gv_arm64.s")
+	if err != nil {
+		t.Fatalf("GOObjectAARCH64: %v", err)
+	}
+	v := openGoobj(t, obj)
+	relocs := v.blk(blkReloc)
+	le := binary.LittleEndian
+	// Two DWARF relocs on the lines/DIE symbols, then the code's one pair
+	// relocation.
+	if len(relocs) != 3*23 {
+		t.Fatalf("relocs = %d bytes, want three entries", len(relocs))
+	}
+	cr := relocs[2*23:]
+	if off := int32(le.Uint32(cr[0:])); off != 0 {
+		t.Errorf("pair reloc off = %d, want 0 (the ADRP word)", off)
+	}
+	if siz := cr[4]; siz != 8 {
+		t.Errorf("pair reloc siz = %d, want 8", siz)
+	}
+	if typ := le.Uint16(cr[5:]); typ != relocArm64Addr {
+		t.Errorf("pair reloc type = %d, want %d (R_ADDRARM64)", typ, relocArm64Addr)
+	}
+	if pkg := le.Uint32(cr[15:]); pkg != pkgIdxSelf {
+		t.Errorf("pair reloc PkgIdx = %#x, want pkgIdxSelf", pkg)
+	}
+	// The GLOBL is the first package definition.
+	if sym := le.Uint32(cr[19:]); sym != 0 {
+		t.Errorf("pair reloc SymIdx = %d, want 0 (the GLOBL definition)", sym)
+	}
+}
+
 // TestGOObjectAARCH64Link does an end-to-end link test: it cross-compiles a
 // Go program for arm64, substitutes the gasm-produced object into the package
 // archive, re-links with cmd/link, and verifies the symbol appears in the
@@ -79,6 +136,14 @@ TEXT ·add(SB), NOSPLIT, $0-24
 	ADD R5, R4, R4
 	MOVD R4, ret+16(FP)
 	RET
+
+TEXT ·getv(SB), NOSPLIT, $0-8
+	MOVD $v<>(SB), R4
+	MOVD R4, ret+0(FP)
+	RET
+
+GLOBL v<>(SB), RODATA, $8
+DATA v<>+0(SB)/8, $7
 `
 	if err := os.WriteFile(filepath.Join(dir, "main_arm64.s"), []byte(asmSrc), 0o644); err != nil {
 		t.Fatal(err)
@@ -86,11 +151,15 @@ TEXT ·add(SB), NOSPLIT, $0-24
 	mainSrc := `package main
 
 func add(a, b int64) int64
+func getv() *int64
 
 func main() {
 	if add(20, 22) != 42 {
 		panic("bad add")
 	}
+	if getv() == nil {
+		panic("bad getv")
+	}
 }
 `
 	if err := os.WriteFile(filepath.Join(dir, "main.go"), []byte(mainSrc), 0o644); err != nil {
@@ -109,16 +178,13 @@ func main() {
 	if err != nil {
 		t.Fatalf("baseline build: %v\n%s", err, buildLog)
 	}
-	var pkgArch, work, linkLine, asmObj string
-	for _, line := range strings.Split(string(buildLog), "\n") {
+	var work, linkLine, asmObj string
+	for line := range strings.SplitSeq(string(buildLog), "\n") {
 		switch {
 		case strings.HasPrefix(line, "WORK="):
 			work = strings.TrimPrefix(line, "WORK=")
 		case strings.Contains(line, "/asm ") && strings.Contains(line, "main_arm64.s") && !strings.Contains(line, "-gensymabis"):
 			asmObj = fieldAfter(line, "-o")
-		case strings.Contains(line, "pack r") && strings.Contains(line, "_pkg_.a"):
-			pkgArch = strings.TrimSpace(strings.SplitN(line, "pack r", 2)[1])
-			pkgArch = strings.Fields(strings.SplitN(pkgArch, "#", 2)[0])[0]
 		case strings.Contains(line, "/link ") && strings.Contains(line, "-importcfg"):
 			linkLine = line
 		}
@@ -179,3 +245,104 @@ func main() {
 		t.Error("binary does not contain expected symbol")
 	}
 }
+
+// TestGOObjectAARCH64DataSymbolLink does for symbol-valued DATA fields what
+// the rt0 files do ("DATA _rt0…lib+0(SB)/8, $_rt0…lib(SB)"): the gasm object
+// carries an R_ADDR against the file's own TEXT symbol, the toolchain links
+// it, and the binary is checked for the symbol (no arm64 host to run it).
+func TestGOObjectAARCH64DataSymbolLink(t *testing.T) {
+	goBin, err := exec.LookPath("go")
+	if err != nil {
+		t.Skip("no Go toolchain available")
+	}
+	dir := t.TempDir()
+	asmSrc := `#include "textflag.h"
+GLOBL entry(SB), NOPTR, $8
+DATA entry+0(SB)/8, $·keepme(SB)
+
+TEXT ·keepme(SB), NOSPLIT, $0-0
+	RET
+
+TEXT ·entryptr(SB), NOSPLIT, $0-8
+	MOVD entry+0(SB), R4
+	MOVD R4, ret+0(FP)
+	RET
+`
+	if err := os.WriteFile(filepath.Join(dir, "main_arm64.s"), []byte(asmSrc), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	mainSrc := `package main
+
+func keepme()
+func entryptr() uintptr
+
+func main() {
+	if entryptr() == 0 {
+		panic("the entry word is empty")
+	}
+}
+`
+	if err := os.WriteFile(filepath.Join(dir, "main.go"), []byte(mainSrc), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(dir, "go.mod"), []byte("module a64dlink\n\ngo 1.21\n"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	build := exec.Command(goBin, "build", "-x", "-work", "-o", filepath.Join(dir, "prog"), ".")
+	build.Dir = dir
+	build.Env = append(os.Environ(), "GOARCH=arm64")
+	buildLog, err := build.CombinedOutput()
+	if err != nil {
+		t.Fatalf("baseline build: %v\n%s", err, buildLog)
+	}
+	var work, linkLine, asmObj string
+	for line := range strings.SplitSeq(string(buildLog), "\n") {
+		switch {
+		case strings.HasPrefix(line, "WORK="):
+			work = strings.TrimPrefix(line, "WORK=")
+		case strings.Contains(line, "/asm ") && strings.Contains(line, "main_arm64.s") && !strings.Contains(line, "-gensymabis"):
+			asmObj = fieldAfter(line, "-o")
+		case strings.Contains(line, "/link ") && strings.Contains(line, "-importcfg"):
+			linkLine = line
+		}
+	}
+	if work == "" || asmObj == "" || linkLine == "" {
+		t.Skipf("could not parse build log (work=%q asmObj=%q link=%q)", work, asmObj, linkLine)
+	}
+	defer os.RemoveAll(work)
+	asmObj = strings.ReplaceAll(asmObj, "$WORK", work)
+	linkLine = strings.ReplaceAll(linkLine, "$WORK", work)
+
+	src, err := os.ReadFile(filepath.Join(dir, "main_arm64.s"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	f, errs := parser.Parse("main_arm64.s", string(src))
+	if len(errs) > 0 {
+		t.Fatalf("parse: %v", errs)
+	}
+	img, err := AssembleFileARM64(f)
+	if err != nil {
+		t.Fatalf("AssembleFileARM64: %v", err)
+	}
+	gasmObj, err := img.GOObjectAARCH64("a64dlink", "main_arm64.s")
+	if err != nil {
+		t.Fatalf("GOObjectAARCH64: %v", err)
+	}
+	if err := os.WriteFile(asmObj, gasmObj, 0o644); err != nil {
+		t.Fatalf("write gasm object: %v", err)
+	}
+	linkCmd := exec.Command("bash", "-c", "cd "+dir+" && "+linkLine)
+	linkCmd.Env = append(os.Environ(), "GOARCH=arm64")
+	if out, err := linkCmd.CombinedOutput(); err != nil {
+		t.Fatalf("re-link with gasm object: %v\n%s", err, out)
+	}
+	binData, err := os.ReadFile(filepath.Join(dir, "prog"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(string(binData), "keepme") {
+		t.Error("binary does not contain the keepme symbol")
+	}
+}

asm/arm64_frame.go

@@ -53,7 +53,7 @@ package asm
 import (
 	"strings"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/ast"
+	"sourcedock.dev/petrbalvin/gasm-sdk/ast"
 )
 
 // arm64FrameInfo holds the frame layout derived from a TEXT directive.
@@ -63,6 +63,12 @@ type arm64FrameInfo struct {
 	args     int  // the declared -argsize
 	noSplit  bool // the NOSPLIT flag
 	leaf     bool // no call instructions in the body
+
+	// Stack-split guard state: needSplit mirrors the toolchain, which skips
+	// the check for NOSPLIT functions and auto-marks leaf functions with an
+	// autosize below StackSmall as NOSPLIT.
+	needSplit  bool
+	splitClass int // 0: <=StackSmall, 1: <=StackBig, 2: >StackBig
 }
 
 // arm64ComputeFrame derives the frame layout for a TEXT function.
@@ -80,15 +86,68 @@ func arm64ComputeFrame(t *ast.Text) arm64FrameInfo {
 
 	if fi.frame != 0 || !fi.leaf {
 		fi.autosize = fi.frame + 8 // space for the saved LR
-		if fi.autosize%16 != 0 {
-			// The toolchain aligns to 16: if autosize%16 == 8, add 8;
-			// otherwise add whatever is needed.
+		// The toolchain always adds an extrasize: 8 when the total leaves a
+		// 16-byte alignment gap, another 16 when already aligned.
+		switch fi.autosize % 16 {
+		case 8:
+			fi.autosize += 8
+		case 0:
+			fi.autosize += 16
+		default:
+			// The toolchain rejects unaligned frames; round up so such
+			// sources still assemble.
 			fi.autosize += 16 - (fi.autosize % 16)
 		}
 	}
+	switch {
+	case fi.noSplit:
+	case fi.autosize < stackSmall && fi.leaf:
+		// Auto-NOSPLIT, as the toolchain's leaf mark concludes.
+	default:
+		fi.needSplit = true
+		switch {
+		case fi.autosize <= stackSmall:
+			fi.splitClass = 0
+		case fi.autosize <= stackBig:
+			fi.splitClass = 1
+		default:
+			fi.splitClass = 2
+		}
+	}
 	return fi
 }
 
+// arm64GuardLen returns the byte length of the stack-split guard prefix
+// (zero when the function needs no guard).  The big class materialises
+// framesize-StackSmall into REGTMP, whose MOVZ/MOVK sequence length varies.
+func arm64GuardLen(fi arm64FrameInfo) int {
+	if !fi.needSplit {
+		return 0
+	}
+	switch fi.splitClass {
+	case 0:
+		return 12
+	case 1:
+		return 16
+	default:
+		n, err := arm64LoadImmLen(int64(fi.autosize - stackSmall))
+		if err != nil {
+			return 0
+		}
+		return 4 + n + 4 + 4 + 4 + 4
+	}
+}
+
+// arm64LoadImmLen returns the byte length of the MOVZ/MOVK sequence that
+// loads v into a register.
+func arm64LoadImmLen(v int64) (int, error) {
+	b, err := encodeARM64LoadImm(27, v, "MOVD")
+	if err != nil {
+		return 0, err
+	}
+	return len(b), nil
+}
+
 // arm64IsLeaf reports whether a function contains no call instructions
 // (BL/CALL), matching the toolchain's LEAF mark.
 func arm64IsLeaf(t *ast.Text) bool {
@@ -119,12 +178,99 @@ func arm64Prologue(fi arm64FrameInfo) []byte {
 		)
 	}
 	// Large frame: SUB $autosize, SP, R20; STP (FP,LR), -8(R20); ADD $0, R20, SP; SUB $8, SP, FP
-	return a64WordsLE(
-		a64AddSub(1, 1, 0, 0, uint32(fi.autosize), 31, 20), // SUB $autosize, SP, R20
-		a64LSP(2, 0, 0, -1, 30, 20, 29),                    // STP FP, LR, [R20, #-8] (opc=2 for 64-bit pair)
-		a64AddSub(1, 0, 0, 0, 0, 20, 31),                   // ADD $0, R20, SP (= MOV R20, SP)
-		a64AddSub(1, 1, 0, 0, 8, 31, 29),                   // SUB $8, SP, FP (op=1 for SUB)
+	ws := arm64SubImmWords(uint32(fi.autosize), 20)
+	ws = append(ws,
+		a64LSP(2, 0, 0, -1, 30, 20, 29),  // STP FP, LR, [R20, #-8] (opc=2 for 64-bit pair)
+		a64AddSub(1, 0, 0, 0, 0, 20, 31), // ADD $0, R20, SP (= MOV R20, SP)
+		a64AddSub(1, 1, 0, 0, 8, 31, 29), // SUB $8, SP, FP (op=1 for SUB)
 	)
+	return a64WordsLE(ws...)
+}
+
+// arm64SplitImm12 reports whether the toolchain decomposes ADD/SUB $imm into
+// two imm12 instructions instead of materialising it into REGTMP
+// (asm7.go case 48, the C_ADDCON2 class): the value must fit 24 bits
+// unsigned and be neither encodable as one imm12 (checked by the callers
+// first), nor loadable into a register in a single MOVZ/MOVN word, nor a
+// logical immediate, because conclass tests all three before C_ADDCON2.
+func arm64SplitImm12(imm uint32) bool {
+	if imm > 0xFFFFFF {
+		return false
+	}
+	if _, _, _, ok := arm64Bitmask(uint64(imm), 1); ok {
+		return false
+	}
+	return arm64Movcon(int64(imm)) < 0 && arm64Movcon(^int64(imm)) < 0
+}
+
+// arm64SubImmWords emits SUB $imm, SP, Rd with the toolchain's ladder for an
+// ADD/SUB constant (asm7.go conclass and cases 2, 48, 62 and 13): the
+// immediate form when the value fits imm12 (plain, or shifted left by 12
+// when it is a multiple of 4096); a value with a single 16-bit chunk, a
+// logical immediate, or one wider than 24 bits is materialised into REGTMP
+// (R27) and subtracted in the extended-register form; everything else up to
+// 0xFFFFFF is split into two imm12 instructions:
+//
+//	SUB $(imm&0xfff), SP, Rd
+//	SUB $((imm&0xfff000)>>12)<<12, Rd, Rd
+func arm64SubImmWords(imm uint32, rd uint32) []uint32 {
+	if imm <= 0xFFF {
+		return []uint32{a64AddSub(1, 1, 0, 0, imm, 31, rd)}
+	}
+	if imm <= 4095<<12 && imm&0xFFF == 0 {
+		return []uint32{a64AddSub(1, 1, 0, 1, imm>>12, 31, rd)}
+	}
+	if !arm64SplitImm12(imm) {
+		mov, err := encodeARM64LoadImm(27, int64(imm), "MOVD")
+		if err != nil {
+			mov = nil
+		}
+		return append(wordsOf(mov), arm64DPExtWords(arm64OpSub, 27, 31, rd))
+	}
+	return []uint32{
+		a64AddSub(1, 1, 0, 0, imm&0xFFF, 31, rd),
+		a64AddSub(1, 1, 0, 1, (imm&0xFFF000)>>12, rd, rd),
+	}
+}
+
+// arm64AddImmWords emits ADD $imm, SP, Rd with the same imm12, shifted-imm12,
+// split and REGTMP ladder as arm64SubImmWords.
+func arm64AddImmWords(imm uint32, rd uint32) []uint32 {
+	if imm <= 0xFFF {
+		return []uint32{a64AddSub(1, 0, 0, 0, imm, 31, rd)}
+	}
+	if imm <= 4095<<12 && imm&0xFFF == 0 {
+		return []uint32{a64AddSub(1, 0, 0, 1, imm>>12, 31, rd)}
+	}
+	if !arm64SplitImm12(imm) {
+		mov, err := encodeARM64LoadImm(27, int64(imm), "MOVD")
+		if err != nil {
+			mov = nil
+		}
+		return append(wordsOf(mov), arm64DPExtWords(arm64OpAdd, 27, 31, rd))
+	}
+	return []uint32{
+		a64AddSub(1, 0, 0, 0, imm&0xFFF, 31, rd),
+		a64AddSub(1, 0, 0, 1, (imm&0xFFF000)>>12, rd, rd),
+	}
+}
+
+// arm64RetAddWords emits the frame deallocation of a non-leaf RET with a
+// large frame.  The toolchain adds the frame back with a single instruction:
+// a plain imm12 ADD when autosize fits 12 bits, otherwise the value is
+// materialised into REGTMP and added as a register, so the epilogue never
+// leaves a partially deallocated frame (obj7.go ARET, issue 73259).  The
+// shifted-imm12 and split-imm12 forms are therefore never used here, unlike
+// the leaf epilogue's plain ADD instructions.
+func arm64RetAddWords(autosize uint32) []uint32 {
+	if autosize < 1<<12 {
+		return []uint32{a64AddSub(1, 0, 0, 0, autosize, 31, 31)}
+	}
+	mov, err := encodeARM64LoadImm(27, int64(autosize), "MOVD")
+	if err != nil {
+		mov = nil
+	}
+	return append(wordsOf(mov), arm64DPExtWords(arm64OpAdd, 27, 31, 31))
 }
 
 // arm64Return returns the bytes for a RET: the epilogue (restore FP/LR and
@@ -134,10 +280,8 @@ func arm64Return(fi arm64FrameInfo) []byte {
 	if fi.autosize != 0 {
 		if fi.leaf {
 			// Leaf with frame: ADD $autosize-8, SP, FP; ADD $autosize, SP, SP
-			ws = append(ws,
-				a64AddSub(1, 0, 0, 0, uint32(fi.autosize-8), 31, 29), // ADD $autosize-8, SP, FP
-				a64AddSub(1, 0, 0, 0, uint32(fi.autosize), 31, 31),   // ADD $autosize, SP, SP
-			)
+			ws = append(ws, arm64AddImmWords(uint32(fi.autosize-8), 29)...)
+			ws = append(ws, arm64AddImmWords(uint32(fi.autosize), 31)...)
 		} else if fi.autosize <= 0xf0 {
 			// Non-leaf small frame: LDR FP, [SP, #-8]; LDR.P LR, [SP], #autosize
 			ws = append(ws,
@@ -145,11 +289,11 @@ func arm64Return(fi arm64FrameInfo) []byte {
 				arm64PostLoad(3, 0, int32(fi.autosize), 31, 30), // LDR.P LR, [SP], #autosize
 			)
 		} else {
-			// Large frame: LDP -8(SP), (FP, LR); ADD $autosize, SP, SP
+			// Large frame: LDP -8(SP), (FP, LR), then deallocate.
 			ws = append(ws,
-				a64LSP(2, 0, 1, -1, 30, 31, 29),                    // LDP FP, LR, [SP, #-8] (opc=2 for 64-bit pair)
-				a64AddSub(1, 0, 0, 0, uint32(fi.autosize), 31, 31), // ADD $autosize, SP, SP
+				a64LSP(2, 0, 1, -1, 30, 31, 29), // LDP FP, LR, [SP, #-8] (opc=2 for 64-bit pair)
 			)
+			ws = append(ws, arm64RetAddWords(uint32(fi.autosize))...)
 		}
 	}
 	// RET: BR LR (0xd65f03c0)
@@ -166,22 +310,32 @@ func arm64PrologueSpadjPC(fi arm64FrameInfo) int {
 	if fi.autosize <= 0xf0 {
 		return 4 // MOVD.W instruction decrements SP
 	}
-	return 8 // SUB + STP + MOVD (3 instructions, SP updated at the MOVD)
+	// Large frame: [SUB words][STP][ADD R20, SP]; SP moves at the ADD, whose
+	// position depends on how many words the SUB itself took (immediate,
+	// shifted immediate, the two-word imm12 split, or a materialised REGTMP
+	// sequence).
+	return 4 * (len(arm64SubImmWords(uint32(fi.autosize), 20)) + 1)
 }
 
 // arm64ReturnEpilogueLen returns the byte length of the RET's epilogue up to
-// (but not including) the final RET instruction.
+// (but not including) the final RET instruction.  The lengths are read from
+// the same word-emitting helpers the epilogue uses rather than assumed: the
+// leaf path shares the prologue's immediate ladder, and a materialised
+// autosize costs its MOV words plus the ADD itself.
 func arm64ReturnEpilogueLen(fi arm64FrameInfo) int {
 	if fi.autosize == 0 {
 		return 0
 	}
 	if fi.leaf {
-		return 8 // ADD + ADD
+		return 4 * (len(arm64AddImmWords(uint32(fi.autosize-8), 29)) +
+			len(arm64AddImmWords(uint32(fi.autosize), 31)))
 	}
 	if fi.autosize <= 0xf0 {
 		return 8 // LDR + LDR.P
 	}
-	return 8 // LDP + ADD
+	// LDP + the deallocation emitted by arm64RetAddWords, so the length
+	// tracks whatever the MOVD ladder needs.
+	return 4 + 4*len(arm64RetAddWords(uint32(fi.autosize)))
 }
 
 // arm64ResolvePseudo translates a pseudo-register memory reference into a
@@ -235,3 +389,86 @@ func arm64PostLoad(size, V int, imm9 int32, rn, rt int) uint32 {
 	return uint32(size)<<30 | 7<<27 | uint32(V)<<26 | 1<<22 |
 		1<<10 | (uint32(imm9)&0x1FF)<<12 | uint32(rn&31)<<5 | uint32(rt&31)
 }
+
+// Data-processing (shifted register) base opcodes for the guard blocks.
+const (
+	arm64OpAdd  = 1<<31 | 0<<30 | 0<<29 | 0x0b<<24
+	arm64OpSub  = 1<<31 | 1<<30 | 0<<29 | 0x0b<<24
+	arm64OpSubs = 1<<31 | 1<<30 | 1<<29 | 0x0b<<24
+)
+
+// arm64DPSRWords builds one data-processing (shifted register) word:
+// OP Rm, Rn, Rd in the Go assembler's operand order.
+func arm64DPSRWords(base uint32, rm, rn, rd uint32) uint32 {
+	return base | rm<<16 | rn<<5 | rd
+}
+
+// arm64DPExtWords builds one data-processing (extended register) word, the
+// form the toolchain picks when a large immediate was materialised into
+// REGTMP before the operation: base | 1<<21 | Rm<<16 | UXTX<<13 | Rn<<5 | Rd.
+func arm64DPExtWords(base, rm, rn, rd uint32) uint32 {
+	return base | 1<<21 | rm<<16 | 3<<13 | rn<<5 | rd
+}
+
+// wordsOf converts little-endian instruction bytes back to words.
+func wordsOf(b []byte) []uint32 {
+	ws := make([]uint32, 0, len(b)/4)
+	for i := 0; i+4 <= len(b); i += 4 {
+		ws = append(ws, uint32(b[i])|uint32(b[i+1])<<8|uint32(b[i+2])<<16|uint32(b[i+3])<<24)
+	}
+	return ws
+}
+
+// arm64GuardBytes emits the stack-split guard prefix; blockStart is the
+// function-relative byte address of the morestack block the branches target.
+func arm64GuardBytes(fi arm64FrameInfo, blockStart int) []byte {
+	// MOVD 16(R28), R16 (g.stackguard0)
+	ws := []uint32{a64LSU(3, 0, 1, 2, 28, 16)}
+	br := func(from int, cond uint32) uint32 {
+		return a64BranchCond(int32((blockStart-from)>>2), cond)
+	}
+	switch fi.splitClass {
+	case 0:
+		// CMP R16, RSP in the exact encoding go tool asm emits for it.
+		ws = append(ws, 0xeb3063ff)
+		ws = append(ws, br(8, a64CondLS))
+	case 1:
+		ws = append(ws, a64AddSub(1, 1, 0, 0, uint32(fi.autosize-stackSmall), 31, 17))
+		ws = append(ws, arm64DPSRWords(arm64OpSubs, 16, 17, 31)) // CMP R16, R17
+		ws = append(ws, br(12, a64CondLS))
+	default:
+		mov, err := encodeARM64LoadImm(27, int64(fi.autosize-stackSmall), "MOVD")
+		if err != nil {
+			mov = nil
+		}
+		ws = append(ws, wordsOf(mov)...)
+		ml := len(mov) / 4
+		ws = append(ws, arm64DPExtWords(arm64OpSubs, 27, 31, 17)) // SUBS R17, RSP, R27
+		// The branches sit at fixed byte offsets in the guard prefix: after
+		// the LDR (4), the ml MOV words (4*ml) and the SUBS (4) for B.LO,
+		// then a further B.LO word and the CMP for B.LS.
+		ws = append(ws, br(8+4*ml, a64CondLO))
+		ws = append(ws, arm64DPSRWords(arm64OpSubs, 16, 17, 31)) // CMP R16, R17
+		ws = append(ws, br(16+4*ml, a64CondLS))
+	}
+	return a64WordsLE(ws...)
+}
+
+// arm64MoreStackBlock emits the trailing block: MOVD R30, R3 (save LR),
+// BL runtime.morestack_noctxt, B back to the function start.  The BL carries
+// the R_CALLARM64 relocation.
+func arm64MoreStackBlock(blockStart int) ([]byte, Reloc) {
+	ws := []uint32{
+		1<<31 | 1<<29 | 0x0a<<24 | 30<<16 | 31<<5 | 3, // MOVD R30, R3
+		a64Branch(1, 0), // BL, patched by the linker
+	}
+	bPC := blockStart + 8
+	ws = append(ws, a64Branch(0, int32(-bPC>>2))) // B back to the entry
+	reloc := Reloc{
+		Off:   blockStart + 4,
+		After: blockStart + 8,
+		Name:  "runtime\u00b7morestack_noctxt",
+		Kind:  RelArm64Branch,
+	}
+	return a64WordsLE(ws...), reloc
+}

asm/arm64_reloc_test.go

@@ -0,0 +1,126 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package asm
+
+import (
+	"encoding/binary"
+	"testing"
+
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
+)
+
+// parseArm64File is a helper assembling one arm64 source file.
+func parseArm64File(t *testing.T, src string) *Image {
+	t.Helper()
+	f, errs := parser.Parse("k_arm64.s", src)
+	if len(errs) > 0 {
+		t.Fatalf("parse: %v", errs)
+	}
+	img, err := AssembleFileARM64(f)
+	if err != nil {
+		t.Fatalf("assemble: %v", err)
+	}
+	return img
+}
+
+// TestArm64RelocOffsetsIncludePrologue pins the function-relative relocation
+// offsets of a framed function: the offsets used to exclude the prologue, so
+// every relocation landed on a prologue instruction in the GOOBJ/ELF output.
+// The function calls an external, so it is a non-leaf and carries the
+// stack-split guard (12 bytes, small class) before the prologue.
+func TestArm64RelocOffsetsIncludePrologue(t *testing.T) {
+	img := parseArm64File(t, "TEXT \u00b7f(SB), $16-0\n"+
+		"\tBL ext\u00b7foo(SB)\n"+
+		"\tMOVD $gdata(SB), R5\n"+
+		"\tMOVD $extsym(SB), R6\n"+
+		"\tRET\n"+
+		"GLOBL gdata(SB), $8\n")
+	fn := img.Funcs[0]
+
+	// Layout: 12-byte guard, 12-byte prologue, BL (24), ADRP+ADD (28, 32),
+	// ADRP+ADD (36, 40), 12-byte epilogue with RET, 12-byte morestack block.
+	want := []struct {
+		off      int
+		after    int
+		name     string
+		kind     RelocKind
+		external bool
+	}{
+		{24, 28, "foo", RelArm64Branch, true},
+		{28, 28, "gdata", RelArm64Addr, false},
+		{32, 32, "gdata", RelArm64Addr, false},
+		{36, 36, "extsym", RelArm64Addr, true},
+		{40, 40, "extsym", RelArm64Addr, true},
+		{60, 64, "runtime\u00b7morestack_noctxt", RelArm64Branch, true},
+	}
+	if len(fn.Relocs) != len(want) {
+		t.Fatalf("relocs = %d, want %d", len(fn.Relocs), len(want))
+	}
+	for i, w := range want {
+		r := fn.Relocs[i]
+		if r.Off != w.off || r.After != w.after || r.Name != w.name || r.Kind != w.kind || r.External != w.external {
+			t.Errorf("reloc %d = {off %d after %d name %q kind %d ext %v}, want {off %d after %d name %q kind %d ext %v}",
+				i, r.Off, r.After, r.Name, r.Kind, r.External, w.off, w.after, w.name, w.kind, w.external)
+		}
+	}
+
+	// The BL with a zero offset sits exactly at the first reloc site.
+	code := img.Code[fn.Offset : fn.Offset+fn.Size]
+	if w := binary.LittleEndian.Uint32(code[24:28]); w != 0x94000000 {
+		t.Errorf("BL word = %08x, want 94000000", w)
+	}
+}
+
+// TestArm64SBLoadStoreMatchesToolchain pins the ADRP scratch register
+// (REGTMP, R27) and the LDST64 relocation kind for sym loads and stores,
+// against the bytes go tool asm emits for MOVD sym(SB), R5.
+func TestArm64SBLoadStoreMatchesToolchain(t *testing.T) {
+	img := parseArm64File(t, "TEXT \u00b7ld(SB), NOSPLIT, $0\n"+
+		"\tMOVD sym(SB), R5\n"+
+		"\tMOVD R5, sym(SB)\n"+
+		"\tRET\n"+
+		"GLOBL sym(SB), $8\n")
+	fn := img.Funcs[0]
+	code := img.Code[fn.Offset : fn.Offset+fn.Size]
+
+	// go tool asm: ADRP 0(PC), R27 (9000001b); MOVD (R27), R5 (f9400365);
+	// ADRP 0(PC), R27; MOVD R5, (R27) (f9000365).
+	for off, want := range map[int]uint32{0: 0x9000001b, 4: 0xf9400365, 8: 0x9000001b, 12: 0xf9000365} {
+		if got := binary.LittleEndian.Uint32(code[off : off+4]); got != want {
+			t.Errorf("word at %d = %08x, want %08x", off, got, want)
+		}
+	}
+
+	if len(fn.Relocs) != 2 {
+		t.Fatalf("relocs = %d, want 2", len(fn.Relocs))
+	}
+	for i, w := range []struct{ off, after int }{{0, 8}, {8, 16}} {
+		r := fn.Relocs[i]
+		if r.Kind != RelArm64LDST64 {
+			t.Errorf("reloc %d kind = %d, want RelArm64LDST64 (%d)", i, r.Kind, RelArm64LDST64)
+		}
+		if r.Off != w.off || r.After != w.after {
+			t.Errorf("reloc %d = {off %d after %d}, want {off %d after %d}", i, r.Off, r.After, w.off, w.after)
+		}
+	}
+}
+
+// TestArm64GOObjRelocTypes checks that GOOBJ emission succeeds with the new
+// relocation kinds in play; the detailed layout is covered by the goobj tests.
+func TestArm64GOObjRelocTypes(t *testing.T) {
+	img := parseArm64File(t, "TEXT \u00b7ld(SB), NOSPLIT, $0\n"+
+		"\tMOVD sym(SB), R5\n"+
+		"\tMOVD R5, sym(SB)\n"+
+		"\tRET\n"+
+		"GLOBL sym(SB), $8\n")
+	obj, err := img.GOObjectAARCH64("testpkg", "k_arm64.s")
+	if err != nil {
+		t.Fatalf("GOObjectAARCH64: %v", err)
+	}
+	if len(obj) == 0 {
+		t.Fatal("empty object")
+	}
+	// The detailed layout is covered by the goobj tests; here we only pin
+	// that emission succeeds with the new relocation kinds in play.
+}

asm/assemble_test.go

@@ -4,13 +4,14 @@
 package asm
 
 import (
+	"bytes"
 	"strings"
 	"testing"
 
 	"golang.org/x/arch/x86/x86asm"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/ast"
-	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+	"sourcedock.dev/petrbalvin/gasm-sdk/ast"
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
 )
 
 // firstText parses src and returns its first TEXT function.
@@ -62,7 +63,7 @@ TEXT ·f(SB), NOSPLIT, $0
 	XORQ AX, AX
 loop:
 	ADDQ $1, AX
-	CMPQ $10, AX
+	CMPQ AX, $10
 	JLT loop
 	RET
 `)
@@ -159,6 +160,57 @@ TEXT ·loadarg(SB), NOSPLIT, $0-24
 	}
 }
 
+// TestAssembleFramelessCall verifies the forced base-pointer frame a $0-frame
+// function containing a CALL receives: the PUSHQ BP prologue with no stack
+// adjustment and the x+N(FP) → (N+16)(SP) translation, against the bytes the
+// Go assembler produces.  The push is the frame, so the offset must not count
+// it twice.
+func TestAssembleFramelessCall(t *testing.T) {
+	f, errs := parser.Parse("frameless_call_amd64.s", `
+#include "textflag.h"
+TEXT ·withcall(SB), NOSPLIT, $0-16
+	MOVQ x+0(FP), AX
+	CALL ·other(SB)
+	MOVQ AX, ret+8(FP)
+	RET
+TEXT ·other(SB), NOSPLIT, $0-0
+	RET
+`)
+	if len(errs) > 0 {
+		t.Fatalf("parse: %v", errs)
+	}
+	img, err := AssembleFile(f)
+	if err != nil {
+		t.Fatalf("AssembleFile: %v", err)
+	}
+	code := append([]byte(nil), img.Code[img.Funcs[0].Offset:img.Funcs[0].Offset+img.Funcs[0].Size]...)
+	for _, r := range img.Funcs[0].Relocs {
+		for j := r.Off; j < r.Off+4 && j < len(code); j++ {
+			code[j] = 0
+		}
+	}
+	// From `go tool objdump` of the Go-assembled function:
+	//   PUSHQ BP           55
+	//   MOVQ SP, BP        4889e5
+	//   MOVQ 0x10(SP), AX  488b442410
+	//   CALL other         e800000000
+	//   MOVQ AX, 0x18(SP)  4889442418
+	//   POPQ BP            5d
+	//   RET                c3
+	want := []byte{
+		0x55,
+		0x48, 0x89, 0xe5,
+		0x48, 0x8b, 0x44, 0x24, 0x10,
+		0xe8, 0x00, 0x00, 0x00, 0x00,
+		0x48, 0x89, 0x44, 0x24, 0x18,
+		0x5d,
+		0xc3,
+	}
+	if hexBytes(code) != hexBytes(want) {
+		t.Errorf("frameless CALL FP translation mismatch:\n got:  %s\n want: %s", hexBytes(code), hexBytes(want))
+	}
+}
+
 // TestAssembleFrame verifies a function with a non-zero frame: the Go-style
 // prologue/epilogue and the x+N(FP) → (N+frame+16)(SP) translation, against
 // the bytes the Go assembler produces.
@@ -201,8 +253,8 @@ TEXT ·withframe(SB), NOSPLIT, $16-16
 }
 
 // TestAssembleVexKernel assembles the horizontal-sum reduction the go-flac
-// kernels end with — exercising the VEX moves, shuffle and extract forms
-// through the full parser → encoder path — and checks the output is
+// kernels end with; exercising the VEX moves, shuffle and extract forms
+// through the full parser → encoder path; and checks the output is
 // byte-identical to the Go assembler's.
 func TestAssembleVexKernel(t *testing.T) {
 	fn := firstText(t, `
@@ -317,3 +369,279 @@ end:
 		t.Errorf("jump-folding mismatch:\n got:  %s\n want: %s", hexBytes(code), hexBytes(want))
 	}
 }
+
+// TestAssembleNumericPCJumps pins the numeric ±N(PC) branch operands: N
+// counts instruction statements, skipping labels, in both directions (the
+// runtime's exit loops write JMP -3(PC)), N = 0 parks on the jump itself.
+func TestAssembleNumericPCJumps(t *testing.T) {
+	fn := firstText(t, `
+#include "textflag.h"
+TEXT ·exit(SB), NOSPLIT, $0
+	MOVB $1, AL
+lab:
+	MOVB $2, AL
+	MOVB $3, AL
+	JMP -3(PC)
+	MOVB $4, AL
+park:
+	JMP 0(PC)
+	MOVB $5, AL
+	JMP 2(PC)
+	MOVB $6, AL
+	RET
+`)
+	code, _, err := Assemble(fn)
+	if err != nil {
+		t.Fatalf("Assemble: %v", err)
+	}
+	// From the Go-assembled function:
+	//   MOVB $1, AL       b001
+	//   MOVB $2, AL       b002
+	//   MOVB $3, AL       b003
+	//   JMP -3(PC)        ebf8      (three instructions back, past lab:)
+	//   MOVB $4, AL       b004
+	//   JMP 0(PC)         ebfe      (the park loop)
+	//   MOVB $5, AL       b005
+	//   JMP 2(PC)         eb02      (over MOVB $6 to the RET)
+	//   MOVB $6, AL       b006
+	//   RET               c3
+	want := []byte{
+		0xb0, 0x01,
+		0xb0, 0x02,
+		0xb0, 0x03,
+		0xeb, 0xf8,
+		0xb0, 0x04,
+		0xeb, 0xfe,
+		0xb0, 0x05,
+		0xeb, 0x02,
+		0xb0, 0x06,
+		0xc3,
+	}
+	if hexBytes(code) != hexBytes(want) {
+		t.Errorf("numeric-PC mismatch:\n got:  %s\n want: %s", hexBytes(code), hexBytes(want))
+	}
+}
+
+func TestAssemblePrefetch(t *testing.T) {
+	fn := firstText(t, `
+#include "textflag.h"
+TEXT ·pf(SB), NOSPLIT, $0
+	PREFETCHNTA (AX)
+	PREFETCHT0 (BX)
+	PREFETCHT1 8(CX)
+	PREFETCHT2 -1(AX)(R12*1)
+	RET
+`)
+	code, _, err := Assemble(fn)
+	if err != nil {
+		t.Fatalf("Assemble: %v", err)
+	}
+	got := strings.Join(disasm(t, code), "\n")
+	want := strings.Join([]string{
+		"prefetchnta zmmword ptr [rax]",
+		"prefetcht0 zmmword ptr [rbx]",
+		"prefetcht1 zmmword ptr [rcx+0x8]",
+		"prefetcht2 zmmword ptr [rax+r12-0x1]",
+		"ret",
+	}, "\n")
+	if got != want {
+		t.Errorf("prefetch disassembly mismatch:\n got:\n%s\n want:\n%s", got, want)
+	}
+	// Byte-level expectations: 0F 18 with the variant in the reg field.
+	if hex := hexBytes(code[:3]); hex != "0f 18 00" {
+		t.Errorf("PREFETCHNTA bytes: got %s, want 0f 18 00", hex)
+	}
+	if hex := hexBytes(code[3:6]); hex != "0f 18 0b" {
+		t.Errorf("PREFETCHT0 bytes: got %s, want 0f 18 0b", hex)
+	}
+}
+
+// TestAssembleBareJump checks that a zero-operand jump (which parses, because
+// the parser does not arity-check mnemonics) is rejected with an error rather
+// than panicking in the layout loop, which indexes Operands[0] before the
+// emission pass gets a chance to diagnose the arity.
+func TestAssembleBareJump(t *testing.T) {
+	for _, mnem := range []string{"JE", "JMP", "JLT", "CALL"} {
+		fn := firstText(t, "TEXT ·bare(SB), $16-0\n\t"+mnem+"\n")
+		if _, _, err := Assemble(fn); err == nil {
+			t.Errorf("%s with no operand: expected an error, got none", mnem)
+		}
+	}
+}
+
+// TestSubSPEncodings pins the prologue SUB against the bytes go tool asm
+// emits for SUBQ $size, SP: imm8 for -128..127, the imm32 form for anything
+// larger.  The intermediate 129..255 range used to encode an ADD with a
+// truncated immediate, moving SP the wrong way.
+func TestSubSPEncodings(t *testing.T) {
+	for _, tt := range []struct {
+		size int
+		want []byte
+	}{
+		{8, []byte{0x48, 0x83, 0xEC, 0x08}},
+		{127, []byte{0x48, 0x83, 0xEC, 0x7F}},
+		{128, []byte{0x48, 0x81, 0xEC, 0x80, 0x00, 0x00, 0x00}},
+		{200, []byte{0x48, 0x81, 0xEC, 0xC8, 0x00, 0x00, 0x00}},
+		{255, []byte{0x48, 0x81, 0xEC, 0xFF, 0x00, 0x00, 0x00}},
+		{4096, []byte{0x48, 0x81, 0xEC, 0x00, 0x10, 0x00, 0x00}},
+	} {
+		got := subSP(tt.size)
+		if !bytes.Equal(got, tt.want) {
+			t.Errorf("subSP(%d) = %x, want %x", tt.size, got, tt.want)
+		}
+	}
+}
+
+// TestAssemblePseudoStatements runs LOCK/REP, BYTE/WORD and END through the
+// full statement pipeline, pinned against go tool asm (Go 1.27, amd64).  It
+// asserts the three behaviours the toolchain shows: each prefix statement is
+// a standalone byte with a PC of its own (so a label placed on the LOCK
+// points at the F0), the data pseudo-ops write their literal bytes inline,
+// and END terminates nothing (the statements after it still belong to the
+// function and carry no trace of it).
+func TestAssemblePseudoStatements(t *testing.T) {
+	fn := firstText(t, `
+#include "textflag.h"
+TEXT ·pseudo(SB), NOSPLIT, $0-0
+pfx:
+	LOCK
+	CMPXCHGQ AX, (BX)
+	REP
+	MOVSQ
+	BYTE $0x0f
+	BYTE $0x1f
+	WORD $0x1234
+	END
+	BYTE $0x02
+	RET
+`)
+	code, labels, err := Assemble(fn)
+	if err != nil {
+		t.Fatalf("Assemble: %v", err)
+	}
+	// go tool asm: f0 480fb103 f3 48a5 0f 1f 3412 02 c3
+	want := []byte{
+		0xf0,
+		0x48, 0x0f, 0xb1, 0x03,
+		0xf3, 0x48, 0xa5,
+		0x0f, 0x1f, 0x34, 0x12,
+		0x02, 0xc3,
+	}
+	if hexBytes(code) != hexBytes(want) {
+		t.Errorf("pseudo statements:\n got:  %s\n want: %s", hexBytes(code), hexBytes(want))
+	}
+	// The label sits on the LOCK byte, exactly where the toolchain's PC
+	// listing puts it.
+	if off := labels["pfx"]; off != 0 {
+		t.Errorf("label pfx = %d, want 0 (the LOCK's own byte)", off)
+	}
+	// The trailing BYTE lands where the layout says: after the 8 bytes of
+	// LOCK, CMPXCHGQ, REP and MOVSQ plus the 4 data bytes, END contributing
+	// none.
+	if code[12] != 0x02 {
+		t.Errorf("byte at 12 = %02x, want 02 (the BYTE after END)", code[12])
+	}
+}
+
+// TestAssembleAdjspBalance pins the toolchain's push/pop balance rule over
+// ADJSP: the straight-line sum of the adjustments must be zero at each
+// RET, branches in between counting for nothing (verified against go tool
+// asm: ADJSP $16 before a RET is reported as "unbalanced PUSH/POP", a
+// $16/$-16 pair with a JMP in between assembles).
+func TestAssembleAdjspBalance(t *testing.T) {
+	// Balanced pair with a branch in between, bytes pinned from go tool asm.
+	fn := firstText(t, `
+#include "textflag.h"
+TEXT ·adjsp(SB), NOSPLIT, $0-0
+	ADJSP $16
+	JMP body
+body:
+	ADJSP $-16
+	RET
+`)
+	code, _, err := Assemble(fn)
+	if err != nil {
+		t.Fatalf("Assemble: %v", err)
+	}
+	want := []byte{0x48, 0x83, 0xEC, 0x10, 0xEB, 0x00, 0x48, 0x83, 0xC4, 0x10, 0xC3}
+	if hexBytes(code) != hexBytes(want) {
+		t.Errorf("adjsp pair:\n got:  %s\n want: %s", hexBytes(code), hexBytes(want))
+	}
+
+	// Unbalanced at the RET: the toolchain diagnoses, so must we.
+	_, _, err = Assemble(firstText(t, `
+#include "textflag.h"
+TEXT ·unbalanced(SB), NOSPLIT, $0-0
+	ADJSP $16
+	RET
+`))
+	if err == nil || !strings.Contains(err.Error(), "unbalanced PUSH/POP") {
+		t.Errorf("unbalanced ADJSP: err = %v, want unbalanced PUSH/POP", err)
+	}
+
+	// The check runs per RET: a closed pair before the first RET does not
+	// excuse an open adjustment before the second.
+	_, _, err = Assemble(firstText(t, `
+#include "textflag.h"
+TEXT ·tworet(SB), NOSPLIT, $0-0
+	ADJSP $8
+	ADJSP $-8
+	RET
+mid:
+	ADJSP $8
+	RET
+`))
+	if err == nil || !strings.Contains(err.Error(), "unbalanced PUSH/POP") {
+		t.Errorf("second RET with open ADJSP: err = %v, want unbalanced PUSH/POP", err)
+	}
+
+	// A framed function: the assembler's own prologue and epilogue
+	// contribute matching deltas, so the pair in the body still balances,
+	// and the bytes match go tool asm end to end.
+	fn = firstText(t, `
+#include "textflag.h"
+TEXT ·framed(SB), $16-8
+	ADJSP $8
+	ADJSP $-8
+	RET
+`)
+	code, _, err = Assemble(fn)
+	if err != nil {
+		t.Fatalf("Assemble framed: %v", err)
+	}
+	want = []byte{
+		0x55, 0x48, 0x89, 0xE5, 0x48, 0x83, 0xEC, 0x10, // prologue
+		0x48, 0x83, 0xEC, 0x08, // ADJSP $8
+		0x48, 0x83, 0xC4, 0x08, // ADJSP $-8
+		0x48, 0x83, 0xC4, 0x10, 0x5D, // epilogue
+		0xC3,
+	}
+	if hexBytes(code) != hexBytes(want) {
+		t.Errorf("framed adjsp:\n got:  %s\n want: %s", hexBytes(code), hexBytes(want))
+	}
+}
+
+// TestAssembleRegRange pins the bracketed register range at the statement
+// level: exactly four consecutive same-width vector registers assemble, the
+// toolchain's rejected shapes all report an error.
+func TestAssembleRegRange(t *testing.T) {
+	asm := func(t *testing.T, op string) ([]byte, error) {
+		t.Helper()
+		f, errs := parser.Parse("f_amd64.s", "TEXT \u00b7f(SB), NOSPLIT, $0\n\tV4FMADDPS 17(SP), "+op+", K2, Z0\n\tRET\n")
+		if len(errs) > 0 {
+			t.Fatalf("parse %s: %v", op, errs)
+		}
+		code, _, err := Assemble(f.Decls[0].(*ast.Text))
+		return code, err
+	}
+	for _, op := range []string{"[Z0-Z3]", "[Z4-Z7]", "[Z28-Z31]"} {
+		if _, err := asm(t, op); err != nil {
+			t.Errorf("%s: %v", op, err)
+		}
+	}
+	for _, op := range []string{"[Z0-Z4]", "[Z0-Z2]", "[Z0-Z0]", "[Z4-Z0]", "[Z1-Z0]", "[AX-Z3]", "[Z0-AX]"} {
+		if _, err := asm(t, op); err == nil {
+			t.Errorf("%s: assembled, want an error", op)
+		}
+	}
+}

asm/elf.go

@@ -12,12 +12,11 @@ import (
 // Image: a .text section holding the function bodies, a .data section
 // holding the GLOBL initialisers, a symbol table with one symbol per TEXT
 // and GLOBL (file-local <> symbols are STB_LOCAL, the rest STB_GLOBAL), and
-// a .rela.text relocation table — one R_X86_64_PC32 entry per static-symbol
+// a .rela.text relocation table, one R_X86_64_PC32 entry per static-symbol
 // reference, internal references resolving against the local data symbols
 // and external ones against undefined globals.  The output links with the
 // system toolchain (cc/ld) the way a hand-assembled .o would.
 
-// ELF constants (ELF64, little-endian, System V).
 const (
 	elfClass64 = 2
 	elfDataLSB = 1
@@ -36,18 +35,20 @@ const (
 	shfAlloc     = 2
 	shfExecInstr = 4
 
-	stbLocal  = 0
 	stbGlobal = 1
 
-	sttNotype   = 0
 	sttObject   = 1
 	sttFunc     = 2
 	sttSection  = 3
 	stInfoShift = 4
 
-	shnUndef = 0
-
 	rX8664PC32 = 2
+	// R_X86_64_32 (debug/elf): the absolute 32-bit address of a symbol, the
+	// R_ADDR shape a 4-byte DATA field carries.
+	rX8664Abs32 = 10
+	// R_X86_64_TPOFF32 (debug/elf): the local-exec TLS offset the stack
+	// guard loads from FS.  20 is R_X86_64_TLSLD, a different relocation.
+	rX8664TPOFF32 = 23
 )
 
 // elfSym is one symbol-table entry in construction.
@@ -76,7 +77,7 @@ func (img *Image) ELFObject() ([]byte, error) {
 
 	// Build the symbol table: the null entry and the two section symbols
 	// come first, then the local symbols (static TEXT and GLOBL), then the
-	// globals (exported TEXT and GLOBL, and the undefined externals) — ELF
+	// globals (exported TEXT and GLOBL, and the undefined externals), ELF
 	// requires every local to precede every global, and sh_info records the
 	// boundary.  symIdx maps a symbol name to its index for the relocations.
 	var locals, globals []elfSym
@@ -130,11 +131,19 @@ func (img *Image) ELFObject() ([]byte, error) {
 	type elfRela struct {
 		off    uint64
 		sym    int
+		typ    uint32
 		addend int64
 	}
 	var relas []elfRela
 	for _, fn := range img.Funcs {
 		for _, r := range fn.Relocs {
+			var typ uint32 = rX8664PC32
+			if r.Kind == RelTLSLE {
+				// R_X86_64_TPOFF32 resolves to the local-exec TLS offset and
+				// carries no symbol.
+				relas = append(relas, elfRela{off: uint64(fn.Offset + r.Off), sym: 0, typ: rX8664TPOFF32})
+				continue
+			}
 			idx, ok := symIdx[r.Name]
 			if !ok {
 				return nil, fmt.Errorf("relocation references unknown symbol %q", r.Name)
@@ -142,6 +151,7 @@ func (img *Image) ELFObject() ([]byte, error) {
 			relas = append(relas, elfRela{
 				off: uint64(fn.Offset + r.Off),
 				sym: idx,
+				typ: typ,
 				// R_X86_64_PC32 computes S + A − P with P the patch site; the
 				// assembler measures the symbol from the instruction end,
 				// After − Off bytes past the field, so the addend carries
@@ -151,6 +161,50 @@ func (img *Image) ELFObject() ([]byte, error) {
 		}
 	}
 
+	// The data symbols' symbol-valued DATA fields ("DATA s+0(SB)/8,
+	// $other(SB)") become .rela.data entries: an absolute relocation of the
+	// DATA line's width at the field's data-section offset, S + A with no
+	// PC term.  Widths 4 and 8 have ELF relocation shapes; narrower fields
+	// cannot hold an address, so they are refused rather than truncated.
+	var dataRelas []elfRela
+	for _, d := range img.DataSyms {
+		for _, r := range d.Relocs {
+			idx, ok := symIdx[r.Name]
+			if !ok {
+				return nil, fmt.Errorf("data relocation references unknown symbol %q", r.Name)
+			}
+			var typ uint32
+			switch r.Siz {
+			case 8:
+				typ = rX8664Abs64
+			case 4:
+				typ = rX8664Abs32
+			default:
+				return nil, fmt.Errorf("DATA %q: a symbol value of width %d has no ELF relocation", d.Name, r.Siz)
+			}
+			dataRelas = append(dataRelas, elfRela{
+				off:    uint64(d.Offset + r.Off),
+				sym:    idx,
+				typ:    typ,
+				addend: r.Addend,
+			})
+		}
+	}
+
+	// Section presence: .rela.text only when there are code relocations,
+	// .rela.data only when a DATA line holds a symbol value.
+	hasRela := len(relas) > 0
+	hasDataRela := len(dataRelas) > 0
+	nSections := 6 // NULL, .text, .data, .symtab, .strtab, .shstrtab
+	if hasRela {
+		nSections++
+	}
+	if hasDataRela {
+		nSections++
+	}
+	secSymtab, secStrtab := 3, 4
+	secShstr := nSections - 1
+
 	// Serialise the string tables.
 	stNames := newElfStrtab()
 	for _, s := range syms {
@@ -160,15 +214,12 @@ func (img *Image) ELFObject() ([]byte, error) {
 	for _, n := range []string{".text", ".data", ".symtab", ".strtab", ".rela.text", ".shstrtab"} {
 		stSections.add(n)
 	}
-
-	// Section presence: .rela.text only when there are relocations.
-	hasRela := len(relas) > 0
-	nSections := 6 // NULL, .text, .data, .symtab, .strtab, .shstrtab
-	if hasRela {
-		nSections = 7
+	if hasDataRela {
+		stSections.add(".rela.data")
+	}
+	for _, n := range dwarfSectionNames {
+		stSections.add(n)
 	}
-	secSymtab, secStrtab := 3, 4
-	secShstr := nSections - 1
 
 	// Lay the file out: header, section data, section headers.
 	var out []byte
@@ -204,14 +255,25 @@ func (img *Image) ELFObject() ([]byte, error) {
 	strtabOff := len(out)
 	out = append(out, stNames.bytes()...)
 
-	var relaOff int
+	var relaOff, relaDataOff int
 	if hasRela {
 		align(8)
 		relaOff = len(out)
 		for _, r := range relas {
 			var b [24]byte
 			le.PutUint64(b[0:], r.off)
-			le.PutUint64(b[8:], uint64(r.sym)<<32|rX8664PC32)
+			le.PutUint64(b[8:], uint64(r.sym)<<32|uint64(r.typ))
+			le.PutUint64(b[16:], uint64(r.addend))
+			out = append(out, b[:]...)
+		}
+	}
+	if hasDataRela {
+		align(8)
+		relaDataOff = len(out)
+		for _, r := range dataRelas {
+			var b [24]byte
+			le.PutUint64(b[0:], r.off)
+			le.PutUint64(b[8:], uint64(r.sym)<<32|uint64(r.typ))
 			le.PutUint64(b[16:], uint64(r.addend))
 			out = append(out, b[:]...)
 		}
@@ -220,6 +282,34 @@ func (img *Image) ELFObject() ([]byte, error) {
 	shstrOff := len(out)
 	out = append(out, stSections.bytes()...)
 
+	// DWARF debug sections; the address placeholders they leave are carried
+	// as .rela.debug_info/.rela.debug_line entries the system linker applies.
+	dwAlign := func(n int) {
+		for len(out)%n != 0 {
+			out = append(out, 0)
+		}
+	}
+	dw := appendDWARFSections(&out, img, dwarfSourceName(img), symIdx, dwAlign, cfiAMD64)
+	dwarfStart := 0 // section index of .debug_abbrev, set when DWARF is present
+	if dw != nil {
+		// Five DWARF sections: .debug_abbrev, .debug_info, .debug_line,
+		// .debug_line_str and .debug_frame (the CIE is unconditional, so
+		// the frame section is always present), plus the relocation
+		// sections below when they carry entries.
+		dwarfStart = nSections
+		nSections += 5
+		appendDWARFRelas(&out, dw, rX8664Abs64, dwAlign)
+		if dw.infoRelaCount > 0 {
+			nSections++
+		}
+		if dw.lineRelaCount > 0 {
+			nSections++
+		}
+		if dw.frameRelaCount > 0 {
+			nSections++
+		}
+	}
+
 	align(8)
 	shoff := len(out)
 
@@ -246,8 +336,45 @@ func (img *Image) ELFObject() ([]byte, error) {
 	if hasRela {
 		putSh(".rela.text", shtRela, 0, relaOff, 24*len(relas), secSymtab, secText, 8, 24)
 	}
+	if hasDataRela {
+		putSh(".rela.data", shtRela, 0, relaDataOff, 24*len(dataRelas), secSymtab, secData, 8, 24)
+	}
 	putSh(".shstrtab", shtStrtab, 0, shstrOff, len(stSections.bytes()), 0, 0, 1, 0)
 
+	// DWARF section headers; their indices follow the write order.
+	if dw != nil {
+		// secIdx is a running section index: each putSh below emits the
+		// next header, and the sh_info of a .rela section names the index
+		// of the section it relocates.
+		secIdx := dwarfStart
+		putSh(".debug_abbrev", shtProgbits, 0, dw.abbrevOff, dw.abbrevSize, 0, 0, 1, 0)
+		secIdx++
+		putSh(".debug_info", shtProgbits, 0, dw.infoOff, dw.infoSize, 0, 0, 1, 0)
+		secInfoIdx := secIdx
+		secIdx++
+		if dw.infoRelaCount > 0 {
+			putSh(".rela.debug_info", shtRela, 0, dw.infoRelaOff, 24*dw.infoRelaCount, secSymtab, secInfoIdx, 8, 24)
+			secIdx++
+		}
+		putSh(".debug_line", shtProgbits, 0, dw.lineOff, dw.lineSize, 0, 0, 1, 0)
+		secLineIdx := secIdx
+		secIdx++
+		if dw.lineRelaCount > 0 {
+			putSh(".rela.debug_line", shtRela, 0, dw.lineRelaOff, 24*dw.lineRelaCount, secSymtab, secLineIdx, 8, 24)
+			secIdx++
+		}
+		putSh(".debug_line_str", shtProgbits, 0, dw.lineStrOff, dw.lineStrSize, 0, 0, 1, 0)
+		secIdx++
+		if dw.frameSize > 0 {
+			putSh(".debug_frame", shtProgbits, 0, dw.frameOff, dw.frameSize, 0, 0, 8, 0)
+			secFrameIdx := secIdx
+			secIdx++
+			if dw.frameRelaCount > 0 {
+				putSh(".rela.debug_frame", shtRela, 0, dw.frameRelaOff, 24*dw.frameRelaCount, secSymtab, secFrameIdx, 8, 24)
+			}
+		}
+	}
+
 	// The ELF header.
 	hdr := out[:64]
 	copy(hdr[0:], []byte{0x7f, 'E', 'L', 'F', elfClass64, elfDataLSB, elfVersion, 0})

asm/elf_dwarf.go

@@ -0,0 +1,406 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package asm
+
+import (
+	"encoding/binary"
+)
+
+// DWARF5 section generation for ELF output.  Unlike the GOOBJ path (where
+// the linker assembles the final DWARF), the ELF path must emit complete,
+// self-contained sections because the system linker only performs fixup
+// relocations, not assembly.
+
+// DWARF5 attribute, form and line-table constants (the values the
+// toolchain uses, cmd/internal/dwarf/dwarf_defs.go; the DIE streams below
+// are written against these forms).
+const (
+	dwAtName       = 0x03 // DW_AT_name
+	dwAtStmtList   = 0x10 // DW_AT_stmt_list
+	dwAtLowPC      = 0x11 // DW_AT_low_pc
+	dwAtHighPC     = 0x12 // DW_AT_high_pc
+	dwAtDeclFile   = 0x3a // DW_AT_decl_file
+	dwAtDeclLine   = 0x3b // DW_AT_decl_line
+	dwAtExternal   = 0x3f // DW_AT_external
+	dwAtFrameBase  = 0x40 // DW_AT_frame_base
+	dwTagSubprog   = 0x2e // DW_TAG_subprogram
+	dwTagCompUnit  = 0x11 // DW_TAG_compile_unit
+	dwFormAddr     = 0x01 // DW_FORM_addr
+	dwFormData8    = 0x07 // DW_FORM_data8
+	dwFormString   = 0x08 // DW_FORM_string
+	dwFormData1    = 0x0b // DW_FORM_data1
+	dwFormUdata    = 0x0f // DW_FORM_udata
+	dwFormSecOff   = 0x17 // DW_FORM_sec_offset
+	dwFormExprloc  = 0x18 // DW_FORM_exprloc
+	dwFormLineStrp = 0x1f // DW_FORM_line_strp
+	dwLnctPath     = 0x01 // DW_LNCT_path
+	dwLnctDirIndex = 0x02 // DW_LNCT_directory_index
+)
+
+// dwarfAbbrevTable returns the .debug_abbrev content: a single compilation
+// unit with DW_TAG_compile_unit and DW_TAG_subprogram entries.  The
+// attribute/form pairs must match the DIE streams dwarfBuildInfoSection
+// writes byte for byte, in the same order, or every consumer's parse of
+// .debug_info desynchronises.
+func dwarfAbbrevTable() []byte {
+	var b []byte
+	// Abbrev 1: DW_TAG_compile_unit.
+	b = append(b, 1)                 // abbreviation code
+	b = appendUleb(b, dwTagCompUnit) // DW_TAG_compile_unit
+	b = append(b, 1)                 // DW_CHILDREN_yes
+	b = appendUleb(b, dwAtLowPC)     // DW_AT_low_pc
+	b = appendUleb(b, dwFormAddr)    // DW_FORM_addr
+	b = appendUleb(b, dwAtHighPC)    // DW_AT_high_pc
+	b = appendUleb(b, dwFormData8)   // DW_FORM_data8
+	b = appendUleb(b, dwAtStmtList)  // DW_AT_stmt_list
+	b = appendUleb(b, dwFormSecOff)  // DW_FORM_sec_offset (4 bytes here)
+	b = appendUleb(b, dwAtName)      // DW_AT_name
+	b = appendUleb(b, dwFormString)  // DW_FORM_string
+	b = appendUleb(b, 0)             // end of attributes: attr 0
+	b = appendUleb(b, 0)             // ... paired with form 0
+
+	// Abbrev 2: DW_TAG_subprogram.
+	b = append(b, 2)                 // abbreviation code
+	b = appendUleb(b, dwTagSubprog)  // DW_TAG_subprogram
+	b = append(b, 0)                 // DW_CHILDREN_no
+	b = appendUleb(b, dwAtName)      // DW_AT_name
+	b = appendUleb(b, dwFormString)  // DW_FORM_string
+	b = appendUleb(b, dwAtLowPC)     // DW_AT_low_pc
+	b = appendUleb(b, dwFormAddr)    // DW_FORM_addr
+	b = appendUleb(b, dwAtHighPC)    // DW_AT_high_pc
+	b = appendUleb(b, dwFormData8)   // DW_FORM_data8
+	b = appendUleb(b, dwAtFrameBase) // DW_AT_frame_base
+	b = appendUleb(b, dwFormExprloc) // DW_FORM_exprloc
+	b = appendUleb(b, dwAtDeclFile)  // DW_AT_decl_file
+	b = appendUleb(b, dwFormData1)   // DW_FORM_data1
+	b = appendUleb(b, dwAtDeclLine)  // DW_AT_decl_line
+	b = appendUleb(b, dwFormData1)   // DW_FORM_data1
+	b = appendUleb(b, dwAtExternal)  // DW_AT_external
+	b = appendUleb(b, 0x0c)          // DW_FORM_flag (one byte, 0 or 1)
+	b = appendUleb(b, 0)             // end of attributes: attr 0
+	b = appendUleb(b, 0)             // ... paired with form 0
+
+	// End of table.
+	b = append(b, 0)
+	return b
+}
+
+// dwarfSections holds the generated DWARF section payloads and their
+// relocations (byte offsets within .debug_info and .debug_line that need
+// fixup against .text symbols).
+type dwarfSections struct {
+	debugAbbrev  []byte
+	debugInfo    []byte
+	debugLine    []byte
+	debugLineStr []byte
+	debugFrame   []byte
+	// Relocations for .debug_info: (offset, symbol name, addend).
+	infoRelocs []dwarfReloc
+	// Relocations for .debug_line: (offset, symbol name, addend).
+	lineRelocs []dwarfReloc
+	// Relocations for .debug_frame: (offset, symbol name, addend), one per
+	// FDE initial_location.
+	frameRelocs []dwarfReloc
+}
+
+type dwarfReloc struct {
+	off    uint64
+	name   string
+	addend int64
+}
+
+// emitDWARF generates complete DWARF5 sections for the image.  cfi carries
+// the architecture's .debug_frame register conventions.
+func emitDWARF(img *Image, srcFile string, cfi cfiArch) *dwarfSections {
+	ds := &dwarfSections{}
+	ds.debugAbbrev = dwarfAbbrevTable()
+
+	// Build the string table for .debug_line_str.
+	lineStr := newElfStrtab()
+	lineStr.add(srcFile)
+	ds.debugLineStr = lineStr.bytes()
+
+	// Build .debug_line; the file table references the source name through
+	// its offset in .debug_line_str.
+	ds.debugLine = dwarfBuildLineSection(img, uint32(lineStr.at(srcFile)), ds)
+
+	// Build .debug_info.
+	ds.debugInfo = dwarfBuildInfoSection(img, srcFile, ds)
+
+	// Build .debug_frame.
+	ds.debugFrame = dwarfBuildFrameSection(img, cfi, ds)
+	return ds
+}
+
+// dwarfBuildLineSection builds a complete .debug_line section.  srcStrOff is
+// the source file name's offset in .debug_line_str.
+func dwarfBuildLineSection(img *Image, srcStrOff uint32, ds *dwarfSections) []byte {
+	var b []byte
+	le := binary.LittleEndian
+
+	// We'll build the header first, then the programs, then patch the length.
+	headerStart := len(b)
+	b = append(b, 0, 0, 0, 0) // unit_length (placeholder)
+	b = le.AppendUint16(b, 5) // version (DWARF5)
+	b = append(b, 8)          // address_size
+	b = append(b, 0)          // segment_selector_size
+	b = append(b, 0, 0, 0, 0) // header_length (placeholder)
+
+	// Line program parameters.
+	b = append(b, 1)                     // minimum_instruction_length
+	b = append(b, 1)                     // maximum_ops_per_instruction
+	b = append(b, 1)                     // default_is_stmt
+	b = append(b, byte(dwLineBase&0xFF)) // line_base (-4 as unsigned)
+	b = append(b, uint8(dwLineRange))    // line_range
+	b = append(b, uint8(dwOpcodeBase))   // opcode_base
+	// Standard opcode lengths (opcode 1..opcode_base-1).
+	b = append(b, 0, 1, 1, 1, 1, 0, 0, 0, 1, 0)
+
+	// Directory table (DWARF5 §6.2.4): entry format descriptors followed by
+	// the entries.  One directory, the compilation directory, whose path is
+	// the empty string at .debug_line_str offset 0.
+	b = append(b, 1)                  // directory_entry_format_count
+	b = appendUleb(b, dwLnctPath)     // DW_LNCT_path
+	b = appendUleb(b, dwFormLineStrp) // DW_FORM_line_strp
+	b = appendUleb(b, 1)              // directories_count
+	b = le.AppendUint32(b, 0)         // .debug_line_str offset of ""
+
+	// File table (DWARF5 §6.2.5).  v5 indexes files from 0, so the source
+	// file is entry 0, matching the DW_AT_decl_file value 0 the DIEs carry.
+	b = append(b, 2)                  // file_name_entry_format_count
+	b = appendUleb(b, dwLnctPath)     // DW_LNCT_path
+	b = appendUleb(b, dwFormLineStrp) // DW_FORM_line_strp
+	b = appendUleb(b, dwLnctDirIndex) // DW_LNCT_directory_index
+	b = appendUleb(b, dwFormUdata)    // DW_FORM_udata
+	b = appendUleb(b, 1)              // file_names_count
+	b = le.AppendUint32(b, srcStrOff) // .debug_line_str offset of the source name
+	b = appendUleb(b, 0)              // directory index 0 (the compilation directory)
+
+	headerEnd := len(b)
+
+	// Per-function line programs.
+	for _, fn := range img.Funcs {
+		// LNE_set_address with the function's offset in .text.
+		b = append(b, 0, 9, 2) // extended opcode, length 9, DW_LNE_set_address
+		addrOff := len(b)
+		b = le.AppendUint64(b, 0) // placeholder for address
+		ds.lineRelocs = append(ds.lineRelocs, dwarfReloc{
+			off:    uint64(addrOff),
+			name:   fn.Name,
+			addend: 0,
+		})
+
+		// Build the line entries.
+		pts := make([]LineEntry, 0, len(fn.Lines)+1)
+		if len(fn.Lines) == 0 || fn.Lines[0].Offset > 0 {
+			pts = append(pts, LineEntry{Offset: 0, Line: fn.Line})
+		}
+		pts = append(pts, fn.Lines...)
+
+		line := int64(1)
+		pc := uint64(0)
+		for _, p := range pts {
+			if p.Line == 0 || uint64(p.Offset) < pc {
+				continue
+			}
+			if int64(p.Line) == line {
+				continue
+			}
+			deltaPC := uint64(p.Offset) - pc
+			deltaLC := int64(p.Line) - line
+			b = dwPutPCLCDelta(b, deltaPC, deltaLC)
+			line, pc = int64(p.Line), uint64(p.Offset)
+		}
+
+		// Advance to end of function.
+		if end := uint64(fn.Size) - pc; end > 0 {
+			b = append(b, 2) // DW_LNS_advance_pc
+			b = appendUleb(b, end)
+		}
+		b = append(b, 0, 1, 1) // LNE_end_sequence
+	}
+
+	// Patch unit_length.
+	le.PutUint32(b[headerStart:], uint32(len(b)-headerStart-4))
+	// Patch header_length.  In the v5 header it follows the one-byte
+	// address_size and segment_selector_size (offset 8, not the DWARF2-4
+	// offset 6), and counts from just past itself to the first program
+	// byte.
+	le.PutUint32(b[headerStart+8:], uint32(headerEnd-headerStart-12))
+	return b
+}
+
+// dwarfBuildInfoSection builds a complete .debug_info section.
+func dwarfBuildInfoSection(img *Image, srcFile string, ds *dwarfSections) []byte {
+	var b []byte
+	le := binary.LittleEndian
+
+	cuStart := len(b)
+	b = append(b, 0, 0, 0, 0) // unit_length (placeholder)
+	b = le.AppendUint16(b, 5) // version (DWARF5)
+	b = append(b, 0x01)       // unit_type (DW_UT_compile)
+	b = append(b, 8)          // address_size
+	b = le.AppendUint32(b, 0) // debug_abbrev_offset (0 since single CU)
+
+	// DW_TAG_compile_unit (abbrev 1).
+	b = append(b, 1) // abbreviation code
+	// DW_AT_low_pc: address of .text start.  A data-only image has no
+	// functions to relocate against; its CU covers no code, so the base
+	// stays zero (the DWARF "no base address" value) with no relocation.
+	b = le.AppendUint64(b, 0) // placeholder
+	if len(img.Funcs) > 0 {
+		ds.infoRelocs = append(ds.infoRelocs, dwarfReloc{
+			off:  uint64(len(b) - 8),
+			name: img.Funcs[0].Name,
+		})
+	}
+	// DW_AT_high_pc: size of .text.
+	b = le.AppendUint64(b, uint64(len(img.Code)))
+	// DW_AT_stmt_list: offset into .debug_line (0).
+	b = le.AppendUint32(b, 0)
+	// DW_AT_name: source file name.
+	b = append(b, srcFile...)
+	b = append(b, 0)
+
+	// DW_TAG_subprogram entries (abbrev 2).
+	for _, fn := range img.Funcs {
+		b = append(b, 2) // abbreviation code
+		// DW_AT_name.
+		b = append(b, fn.Name...)
+		b = append(b, 0)
+		// DW_AT_low_pc.
+		addrOff := len(b)
+		b = le.AppendUint64(b, 0) // placeholder
+		ds.infoRelocs = append(ds.infoRelocs, dwarfReloc{
+			off:    uint64(addrOff),
+			name:   fn.Name,
+			addend: 0,
+		})
+		// DW_AT_high_pc: function size.
+		b = le.AppendUint64(b, uint64(fn.Size))
+		// DW_AT_frame_base: DW_OP_call_frame_cfa.
+		b = append(b, 1, 0x9c)
+		// DW_AT_decl_file: the single file-table entry, index 0 (v5 indexes
+		// files from 0).
+		b = append(b, 0)
+		// DW_AT_decl_line.
+		b = append(b, uint8(fn.Line))
+		// DW_AT_external.
+		if fn.Static {
+			b = append(b, 0)
+		} else {
+			b = append(b, 1)
+		}
+	}
+
+	// End of compile unit children.
+	b = append(b, 0)
+
+	// Patch unit_length.
+	le.PutUint32(b[cuStart:], uint32(len(b)-cuStart-4))
+	return b
+}
+
+func appendUleb(b []byte, v uint64) []byte {
+	return binary.AppendUvarint(b, v)
+}
+
+// appendSleb appends v in signed LEB128, the encoding DWARF specifies:
+// two's-complement sign extension, which is NOT Go's zigzag varint
+// (binary.AppendVarint(-8) encodes 15, where DWARF wants 0x78).
+func appendSleb(b []byte, v int64) []byte {
+	for {
+		c := byte(v & 0x7f)
+		v >>= 7
+		if (v == 0 && c&0x40 == 0) || (v == -1 && c&0x40 != 0) {
+			return append(b, c)
+		}
+		b = append(b, c|0x80)
+	}
+}
+
+// cfiArch carries the .debug_frame CIE parameters that differ per
+// architecture: the DWARF register numbers of the stack pointer the initial
+// CFA rule names and of the return address.  The values are the ones the Go
+// linker writes into its own CIE (cmd/link/internal/ld/dwarf.go uses
+// Dwarfregsp and Dwarfreglr; the per-architecture constants live in
+// cmd/link/internal/<arch>/l.go).
+type cfiArch struct {
+	name   string
+	cfaReg byte // the stack-pointer register the initial CFA rule names
+	raReg  byte // the return-address register
+}
+
+var (
+	cfiAMD64   = cfiArch{"amd64", 7, 16}  // RSP, RIP
+	cfiARM64   = cfiArch{"arm64", 31, 30} // SP (X31), LR (X30)
+	cfiRISCV64 = cfiArch{"riscv64", 2, 1} // X2 (sp), X1 (ra)
+	cfiLOONG64 = cfiArch{"loong64", 3, 1} // $r3 (sp), $r1 (ra)
+)
+
+// dwarfBuildFrameSection builds a .debug_frame section with CFI for stack
+// unwinding.  It emits one CIE and one FDE per function, encoding the
+// CFA (Canonical Frame Address) rule changes at each stack-adjustment
+// boundary recorded in FuncLayout.Spadj.
+func dwarfBuildFrameSection(img *Image, cfi cfiArch, ds *dwarfSections) []byte {
+	var b []byte
+	le := binary.LittleEndian
+
+	// CIE (Common Information Entry).
+	cieStart := len(b)
+	b = append(b, 0, 0, 0, 0)            // length (placeholder)
+	b = le.AppendUint32(b, 0xFFFFFFFF)   // CIE marker
+	b = append(b, 3)                     // version (DWARF3, widely supported)
+	b = append(b, 0)                     // augmentation (empty)
+	b = appendUleb(b, 1)                 // code alignment
+	b = appendSleb(b, -8)                // data alignment (-8 for 64-bit)
+	b = appendUleb(b, uint64(cfi.raReg)) // return address register
+	// Initial CFA rule: DW_CFA_def_cfa (SP, 0)
+	b = append(b, 0x0c)                   // DW_CFA_def_cfa
+	b = appendUleb(b, uint64(cfi.cfaReg)) // the architecture's stack pointer
+	b = appendUleb(b, 0)                  // offset: 0
+	b = append(b, 0)                      // DW_CFA_nop (padding)
+	// Patch CIE length.
+	le.PutUint32(b[cieStart:], uint32(len(b)-cieStart-4))
+
+	// FDEs (Frame Description Entries), one per function.
+	for _, fn := range img.Funcs {
+		fdeStart := len(b)
+		b = append(b, 0, 0, 0, 0)                // length (placeholder)
+		b = le.AppendUint32(b, uint32(cieStart)) // CIE pointer (offset from start)
+		// Initial location: function offset in .text, referenced through
+		// the function's symbol so the linker relocates it.
+		ds.frameRelocs = append(ds.frameRelocs, dwarfReloc{
+			off:  uint64(fdeStart + 8),
+			name: fn.Name,
+		})
+		b = le.AppendUint64(b, uint64(fn.Offset))
+		// Address range: function size.
+		b = le.AppendUint64(b, uint64(fn.Size))
+
+		// Emit CFA rule changes at each Spadj boundary.
+		for _, step := range fn.Spadj {
+			if step.Value == 0 {
+				continue
+			}
+			// DW_CFA_def_cfa_offset: set CFA = SP + |delta|.
+			// The delta is negative (stack grows down), so CFA offset = -delta.
+			offset := -step.Value
+			if offset > 0 {
+				b = append(b, 0x0e) // DW_CFA_def_cfa_offset
+				b = appendUleb(b, uint64(offset))
+			}
+		}
+
+		// Pad to alignment.
+		for len(b)%4 != 0 {
+			b = append(b, 0) // DW_CFA_nop
+		}
+
+		// Patch FDE length.
+		le.PutUint32(b[fdeStart:], uint32(len(b)-fdeStart-4))
+	}
+
+	return b
+}

asm/elf_dwarf_sections.go

@@ -0,0 +1,179 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package asm
+
+import "encoding/binary"
+
+// Absolute 64-bit relocation types for the DWARF address fixups, one per
+// supported architecture (the numbers debug/elf carries).
+const (
+	rX8664Abs64   = 1   // R_X86_64_64
+	rAARCH64Abs64 = 257 // R_AARCH64_ABS64
+	rRISCVAbs64   = 2   // R_RISCV_64
+	rLarchAbs64   = 2   // R_LARCH_64
+)
+
+// dwarfELFSections holds the laid-out DWARF sections ready for inclusion
+// in an ELF file.
+type dwarfELFSections struct {
+	abbrevOff, abbrevSize   int
+	infoOff, infoSize       int
+	lineOff, lineSize       int
+	lineStrOff, lineStrSize int
+	frameOff, frameSize     int
+	// .rela.debug_info and .rela.debug_line contents: file offsets and
+	// entry counts (zero count: the section is absent).
+	infoRelaOff, infoRelaCount   int
+	lineRelaOff, lineRelaCount   int
+	frameRelaOff, frameRelaCount int
+	// Relocations for .debug_info address references, offsets relative to
+	// the section start (what an r_offset in .rela.debug_info means).
+	infoRelocs []elfDwarfReloc
+	// Relocations for .debug_line address references, section-relative.
+	lineRelocs []elfDwarfReloc
+	// Relocations for .debug_frame FDE initial locations, section-relative.
+	frameRelocs []elfDwarfReloc
+}
+
+type elfDwarfReloc struct {
+	off    uint64 // offset within the target section
+	sym    int    // symbol index in .symtab
+	addend int64
+}
+
+// appendDWARFSections generates and appends DWARF5 debug sections to the ELF
+// output.  It returns the section offsets/sizes and relocations for the caller
+// to emit section headers and relocation records.
+//
+// symIdx maps function names to their .symtab indices (needed for relocations
+// against .text symbols).  The map uses objectName format (pkg.name); the
+// DWARF code uses bare function names, so we build a reverse lookup.  cfi
+// carries the architecture's .debug_frame register conventions.
+func appendDWARFSections(out *[]byte, img *Image, srcFile string, symIdx map[string]int, align func(int), cfi cfiArch) *dwarfELFSections {
+	// Build a lookup from bare function name to symbol index.
+	nameToIdx := make(map[string]int, len(symIdx))
+	for name, idx := range symIdx {
+		// Strip package prefix: "pkg.name" → "name".
+		if i := len(name) - 1; i >= 0 {
+			for j := len(name) - 1; j >= 0; j-- {
+				if name[j] == '.' {
+					nameToIdx[name[j+1:]] = idx
+					break
+				}
+			}
+		}
+		nameToIdx[name] = idx
+	}
+	ds := emitDWARF(img, srcFile, cfi)
+	if ds == nil || len(ds.debugAbbrev) == 0 {
+		return nil
+	}
+
+	result := &dwarfELFSections{}
+
+	// .debug_abbrev
+	align(1)
+	result.abbrevOff = len(*out)
+	result.abbrevSize = len(ds.debugAbbrev)
+	*out = append(*out, ds.debugAbbrev...)
+
+	// .debug_line_str
+	align(1)
+	result.lineStrOff = len(*out)
+	result.lineStrSize = len(ds.debugLineStr)
+	*out = append(*out, ds.debugLineStr...)
+
+	// .debug_line
+	align(1)
+	result.lineOff = len(*out)
+	result.lineSize = len(ds.debugLine)
+	*out = append(*out, ds.debugLine...)
+	for _, dr := range ds.lineRelocs {
+		if idx, ok := nameToIdx[dr.name]; ok {
+			result.lineRelocs = append(result.lineRelocs, elfDwarfReloc{
+				off:    dr.off,
+				sym:    idx,
+				addend: dr.addend,
+			})
+		}
+	}
+
+	// .debug_info
+	align(1)
+	result.infoOff = len(*out)
+	result.infoSize = len(ds.debugInfo)
+	*out = append(*out, ds.debugInfo...)
+	for _, dr := range ds.infoRelocs {
+		if idx, ok := nameToIdx[dr.name]; ok {
+			result.infoRelocs = append(result.infoRelocs, elfDwarfReloc{
+				off:    dr.off,
+				sym:    idx,
+				addend: dr.addend,
+			})
+		}
+	}
+
+	// .debug_frame: the section header declares alignment 8, so the data is
+	// padded to 8, matching it.
+	if len(ds.debugFrame) > 0 {
+		align(8)
+		result.frameOff = len(*out)
+		result.frameSize = len(ds.debugFrame)
+		*out = append(*out, ds.debugFrame...)
+		for _, dr := range ds.frameRelocs {
+			if idx, ok := nameToIdx[dr.name]; ok {
+				result.frameRelocs = append(result.frameRelocs, elfDwarfReloc{
+					off:    dr.off,
+					sym:    idx,
+					addend: dr.addend,
+				})
+			}
+		}
+	}
+
+	return result
+}
+
+// appendDWARFRelas writes the .rela.debug_info and .rela.debug_line section
+// bodies from the relocations appendDWARFSections recorded, with the
+// architecture's absolute 64-bit relocation type, and records their file
+// offsets and entry counts on dw.  Called after the DWARF sections
+// themselves so the r_offsets (section-relative) need no adjustment.
+func appendDWARFRelas(out *[]byte, dw *dwarfELFSections, abs64 uint32, align func(int)) {
+	le := binary.LittleEndian
+	write := func(relas []elfDwarfReloc) (off, count int) {
+		if len(relas) == 0 {
+			return 0, 0
+		}
+		align(8)
+		off = len(*out)
+		for _, r := range relas {
+			var b [24]byte
+			le.PutUint64(b[0:], r.off)
+			le.PutUint64(b[8:], uint64(r.sym)<<32|uint64(abs64))
+			le.PutUint64(b[16:], uint64(r.addend))
+			*out = append(*out, b[:]...)
+		}
+		return off, len(relas)
+	}
+	dw.infoRelaOff, dw.infoRelaCount = write(dw.infoRelocs)
+	dw.lineRelaOff, dw.lineRelaCount = write(dw.lineRelocs)
+	dw.frameRelaOff, dw.frameRelaCount = write(dw.frameRelocs)
+}
+
+// dwarfSourceName returns the source name the DWARF sections record: the
+// image's source path when the assembler captured one, "gasm.s" otherwise.
+func dwarfSourceName(img *Image) string {
+	if img.SourcePath != "" {
+		return img.SourcePath
+	}
+	return "gasm.s"
+}
+
+// dwarfSectionNames returns the DWARF section names for the string table.
+var dwarfSectionNames = []string{
+	".debug_abbrev", ".debug_info", ".debug_line", ".debug_line_str",
+	".debug_frame", ".rela.debug_info", ".rela.debug_line",
+	".rela.debug_frame",
+}

asm/elf_dwarf_test.go

@@ -0,0 +1,372 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package asm
+
+import (
+	"bytes"
+	"encoding/binary"
+	"testing"
+
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
+)
+
+// ulebIter reads ULEB128 values, the .debug_abbrev and line-header
+// encoding.
+type ulebIter struct {
+	b []byte
+	i int
+}
+
+func (r *ulebIter) uleb(t *testing.T) uint64 {
+	t.Helper()
+	v, n := binary.Uvarint(r.b[r.i:])
+	if n <= 0 {
+		t.Fatalf("bad ULEB at %d", r.i)
+	}
+	r.i += n
+	return v
+}
+
+func (r *ulebIter) byteAt(t *testing.T) byte {
+	t.Helper()
+	if r.i >= len(r.b) {
+		t.Fatalf("read past end at %d", r.i)
+	}
+	c := r.b[r.i]
+	r.i++
+	return c
+}
+
+func (r *ulebIter) uint32At(t *testing.T) uint32 {
+	t.Helper()
+	v := binary.LittleEndian.Uint32(r.b[r.i:])
+	r.i += 4
+	return v
+}
+
+// sleb reads a signed LEB128, the DWARF encoding (sign-extended two's
+// complement, not Go's zigzag varint).
+func (r *ulebIter) sleb(t *testing.T) int64 {
+	t.Helper()
+	var v int64
+	var shift uint
+	for {
+		c := r.byteAt(t)
+		v |= int64(c&0x7f) << shift
+		shift += 7
+		if c&0x80 == 0 {
+			if c&0x40 != 0 {
+				v |= -1 << shift
+			}
+			return v
+		}
+	}
+}
+
+// dwarfAttr is one attribute/form pair of an abbreviation.
+type dwarfAttr struct{ attr, form uint64 }
+
+// dwarfAbbrev is one parsed abbreviation declaration.
+type dwarfAbbrev struct {
+	code     uint64
+	tag      uint64
+	children bool
+	attrs    []dwarfAttr
+}
+
+// parseAbbrevs walks a .debug_abbrev table: abbreviation code, tag,
+// children flag, then attr/form ULEB pairs terminated by a double zero.
+func parseAbbrevs(t *testing.T, b []byte) map[uint64]dwarfAbbrev {
+	t.Helper()
+	out := map[uint64]dwarfAbbrev{}
+	r := &ulebIter{b: b}
+	for {
+		code := r.uleb(t)
+		if code == 0 {
+			return out
+		}
+		ab := dwarfAbbrev{code: code, tag: r.uleb(t)}
+		ab.children = r.byteAt(t) == 1
+		for {
+			attr := r.uleb(t)
+			form := r.uleb(t)
+			if attr == 0 && form == 0 {
+				break
+			}
+			if attr == 0 || form == 0 {
+				t.Fatalf("abbrev %d: half-terminated attr/form pair (%d, %d)", code, attr, form)
+			}
+			ab.attrs = append(ab.attrs, dwarfAttr{attr, form})
+		}
+		out[code] = ab
+	}
+}
+
+func eqAttrs(t *testing.T, ab dwarfAbbrev, want []dwarfAttr) {
+	t.Helper()
+	if len(ab.attrs) != len(want) {
+		t.Fatalf("abbrev %d attrs = %v, want %v", ab.code, ab.attrs, want)
+	}
+	for i, w := range want {
+		if ab.attrs[i] != w {
+			t.Fatalf("abbrev %d attr %d = (%#x, %#x), want (%#x, %#x)", ab.code, i, ab.attrs[i].attr, ab.attrs[i].form, w.attr, w.form)
+		}
+	}
+}
+
+// TestDwarfAbbrevTable walks the abbreviation table as a consumer does and
+// checks the attribute/form sets against the constants the toolchain uses
+// (cmd/internal/dwarf/dwarf_defs.go).  A wrong constant here renames an
+// attribute (0x1b is comp_dir, not low_pc; 0x29 and 0x37 are bounds and
+// count) and a wrong form desynchronises the DIE parse: 0x25 is strx1, one
+// byte, where the writer emits four for a section offset.
+func TestDwarfAbbrevTable(t *testing.T) {
+	abbrev := dwarfAbbrevTable()
+	if len(abbrev) == 0 {
+		t.Fatal("empty abbrev table")
+	}
+	// Must end with a zero byte (end of table).
+	if abbrev[len(abbrev)-1] != 0 {
+		t.Fatalf("abbrev table last byte = %d, want 0", abbrev[len(abbrev)-1])
+	}
+	abs := parseAbbrevs(t, abbrev)
+	if len(abs) != 2 {
+		t.Fatalf("abbreviations = %d, want 2", len(abs))
+	}
+	cu, ok := abs[1]
+	if !ok {
+		t.Fatal("missing abbreviation 1 (compile unit)")
+	}
+	if cu.tag != dwTagCompUnit || !cu.children {
+		t.Errorf("abbrev 1: tag %#x children %v, want compile unit with children", cu.tag, cu.children)
+	}
+	eqAttrs(t, cu, []dwarfAttr{
+		{dwAtLowPC, dwFormAddr},
+		{dwAtHighPC, dwFormData8},
+		{dwAtStmtList, dwFormSecOff},
+		{dwAtName, dwFormString},
+	})
+	sp, ok := abs[2]
+	if !ok {
+		t.Fatal("missing abbreviation 2 (subprogram)")
+	}
+	if sp.tag != dwTagSubprog || sp.children {
+		t.Errorf("abbrev 2: tag %#x children %v, want subprogram without children", sp.tag, sp.children)
+	}
+	eqAttrs(t, sp, []dwarfAttr{
+		{dwAtName, dwFormString},
+		{dwAtLowPC, dwFormAddr},
+		{dwAtHighPC, dwFormData8},
+		{dwAtFrameBase, dwFormExprloc},
+		{dwAtDeclFile, dwFormData1},
+		{dwAtDeclLine, dwFormData1},
+		{dwAtExternal, 0x0c}, // DW_FORM_flag
+	})
+}
+
+// TestDwarfLineHeaderV5 parses the .debug_line header under DWARF5 rules:
+// the directory and file tables are format-descriptor lists, not the
+// DWARF2-4 shape of null-terminated strings, and the file entry references
+// the source name through .debug_line_str.
+func TestDwarfLineHeaderV5(t *testing.T) {
+	src := `#include "textflag.h"
+TEXT ·add(SB), NOSPLIT, $0-24
+	MOVQ a+0(FP), AX
+	MOVQ b+8(FP), BX
+	ADDQ BX, AX
+	MOVQ AX, ret+16(FP)
+	RET
+`
+	f, errs := parser.Parse("test_amd64.s", src)
+	if len(errs) > 0 {
+		t.Fatalf("parse: %v", errs)
+	}
+	img, err := AssembleFile(f)
+	if err != nil {
+		t.Fatalf("assemble: %v", err)
+	}
+
+	ds := emitDWARF(img, "test_amd64.s", cfiAMD64)
+	r := &ulebIter{b: ds.debugLine}
+	r.uint32At(t) // unit_length
+	if v := binary.LittleEndian.Uint16(ds.debugLine[4:]); v != 5 {
+		t.Fatalf("version = %d, want 5", v)
+	}
+	r.i = 6
+	r.byteAt(t)   // address_size
+	r.byteAt(t)   // segment_selector_size
+	r.uint32At(t) // header_length
+	r.byteAt(t)   // minimum_instruction_length
+	r.byteAt(t)   // maximum_ops_per_instruction
+	r.byteAt(t)   // default_is_stmt
+	r.byteAt(t)   // line_base
+	r.byteAt(t)   // line_range
+	opcodeBase := r.byteAt(t)
+	for range int(opcodeBase) - 1 {
+		r.byteAt(t) // standard opcode lengths
+	}
+
+	// Directory table (DWARF5 §6.2.4).
+	if n := r.byteAt(t); n != 1 {
+		t.Fatalf("directory_entry_format_count = %d, want 1", n)
+	}
+	if lnct := r.uleb(t); lnct != dwLnctPath {
+		t.Errorf("directory content type = %#x, want DW_LNCT_path", lnct)
+	}
+	if form := r.uleb(t); form != dwFormLineStrp {
+		t.Errorf("directory form = %#x, want DW_FORM_line_strp", form)
+	}
+	if n := r.uleb(t); n != 1 {
+		t.Fatalf("directories_count = %d, want 1", n)
+	}
+	if off := r.uint32At(t); off != 0 {
+		t.Errorf("compilation directory line_strp = %d, want 0 (the empty string)", off)
+	}
+
+	// File table (DWARF5 §6.2.5).
+	if n := r.byteAt(t); n != 2 {
+		t.Fatalf("file_name_entry_format_count = %d, want 2", n)
+	}
+	if lnct := r.uleb(t); lnct != dwLnctPath {
+		t.Errorf("file content type = %#x, want DW_LNCT_path", lnct)
+	}
+	if form := r.uleb(t); form != dwFormLineStrp {
+		t.Errorf("file path form = %#x, want DW_FORM_line_strp", form)
+	}
+	if lnct := r.uleb(t); lnct != dwLnctDirIndex {
+		t.Errorf("file content type = %#x, want DW_LNCT_directory_index", lnct)
+	}
+	if form := r.uleb(t); form != dwFormUdata {
+		t.Errorf("file dir-index form = %#x, want DW_FORM_udata", form)
+	}
+	if n := r.uleb(t); n != 1 {
+		t.Fatalf("file_names_count = %d, want 1", n)
+	}
+	strOff := r.uint32At(t)
+	if dirIdx := r.uleb(t); dirIdx != 0 {
+		t.Errorf("file directory index = %d, want 0", dirIdx)
+	}
+
+	// The file entry's line_strp must resolve to the source name.
+	end := int(strOff) + len("test_amd64.s")
+	if int(strOff) >= len(ds.debugLineStr) || !bytes.Equal(ds.debugLineStr[strOff:end], []byte("test_amd64.s")) {
+		t.Errorf("file entry line_strp %d does not name the source: %q", strOff, ds.debugLineStr)
+	}
+
+	// The fixed header fields: address_size 8 and a header_length that
+	// points just past the file table (the patch site is offset 8 in the
+	// v5 header, and the field counts from its own end).
+	if ds.debugLine[6] != 8 || ds.debugLine[7] != 0 {
+		t.Errorf("address_size/segment_selector = %d/%d, want 8/0", ds.debugLine[6], ds.debugLine[7])
+	}
+	if hl := binary.LittleEndian.Uint32(ds.debugLine[8:]); hl != uint32(r.i-12) {
+		t.Errorf("header_length = %d, want %d (the byte after the file table is %d)", hl, r.i-12, r.i)
+	}
+}
+
+// TestDwarfFrameCIEArch checks the shared CIE carries each architecture's
+// stack-pointer and return-address registers: the values the Go linker
+// writes (cmd/link/internal/<arch>/l.go dwarfRegSP/dwarfRegLR).
+func TestDwarfFrameCIEArch(t *testing.T) {
+	for _, tc := range []struct {
+		name string
+		cfi  cfiArch
+	}{
+		{"amd64", cfiAMD64},
+		{"arm64", cfiARM64},
+		{"riscv64", cfiRISCV64},
+		{"loong64", cfiLOONG64},
+	} {
+		frame := dwarfBuildFrameSection(&Image{}, tc.cfi, &dwarfSections{})
+		r := &ulebIter{b: frame}
+		r.uint32At(t) // length
+		if cid := r.uint32At(t); cid != 0xFFFFFFFF {
+			t.Errorf("%s: CIE id = %#x, want 0xffffffff", tc.name, cid)
+		}
+		if v := r.byteAt(t); v != 3 {
+			t.Errorf("%s: CIE version = %d, want 3", tc.name, v)
+		}
+		if aug := r.byteAt(t); aug != 0 {
+			t.Errorf("%s: CIE augmentation = %d, want 0", tc.name, aug)
+		}
+		if ca := r.uleb(t); ca != 1 {
+			t.Errorf("%s: code alignment = %d, want 1", tc.name, ca)
+		}
+		if da := r.sleb(t); da != -8 {
+			t.Errorf("%s: data alignment = %d, want -8 (signed LEB128, not zigzag)", tc.name, da)
+		}
+		if ra := r.uleb(t); ra != uint64(tc.cfi.raReg) {
+			t.Errorf("%s: return-address register = %d, want %d", tc.name, ra, tc.cfi.raReg)
+		}
+		if op := r.byteAt(t); op != 0x0c {
+			t.Errorf("%s: expected DW_CFA_def_cfa, got opcode %#x", tc.name, op)
+		}
+		if cfa := r.uleb(t); cfa != uint64(tc.cfi.cfaReg) {
+			t.Errorf("%s: CFA register = %d, want %d", tc.name, cfa, tc.cfi.cfaReg)
+		}
+		if off := r.uleb(t); off != 0 {
+			t.Errorf("%s: CFA offset = %d, want 0", tc.name, off)
+		}
+	}
+}
+
+func TestEmitDWARF(t *testing.T) {
+	src := `#include "textflag.h"
+TEXT ·add(SB), NOSPLIT, $0-24
+	MOVQ a+0(FP), AX
+	MOVQ b+8(FP), BX
+	ADDQ BX, AX
+	MOVQ AX, ret+16(FP)
+	RET
+`
+	f, errs := parser.Parse("test_amd64.s", src)
+	if len(errs) > 0 {
+		t.Fatalf("parse: %v", errs)
+	}
+	img, err := AssembleFile(f)
+	if err != nil {
+		t.Fatalf("assemble: %v", err)
+	}
+
+	ds := emitDWARF(img, "test_amd64.s", cfiAMD64)
+
+	// .debug_abbrev must not be empty and must start with abbrev code 1.
+	if len(ds.debugAbbrev) == 0 {
+		t.Fatal("empty .debug_abbrev")
+	}
+	if ds.debugAbbrev[0] != 1 {
+		t.Fatalf(".debug_abbrev first byte = %d, want 1", ds.debugAbbrev[0])
+	}
+
+	// .debug_info must have a compile unit header (DWARF5 version 5).
+	if len(ds.debugInfo) < 12 {
+		t.Fatalf(".debug_info too short: %d bytes", len(ds.debugInfo))
+	}
+	// Version field at offset 4 (after unit_length).
+	if ds.debugInfo[4] != 5 || ds.debugInfo[5] != 0 {
+		t.Fatalf(".debug_info version = %d, want 5", uint16(ds.debugInfo[4])|uint16(ds.debugInfo[5])<<8)
+	}
+
+	// .debug_line must have a header.
+	if len(ds.debugLine) < 20 {
+		t.Fatalf(".debug_line too short: %d bytes", len(ds.debugLine))
+	}
+	// Version at offset 4.
+	if ds.debugLine[4] != 5 || ds.debugLine[5] != 0 {
+		t.Fatalf(".debug_line version = %d, want 5", uint16(ds.debugLine[4])|uint16(ds.debugLine[5])<<8)
+	}
+
+	// .debug_line_str must contain the source file name.
+	if len(ds.debugLineStr) == 0 {
+		t.Fatal("empty .debug_line_str")
+	}
+
+	// Relocations must reference the function.
+	if len(ds.lineRelocs) == 0 {
+		t.Fatal("no .debug_line relocations")
+	}
+	if len(ds.infoRelocs) == 0 {
+		t.Fatal("no .debug_info relocations")
+	}
+}

asm/elf_test.go

@@ -12,7 +12,8 @@ import (
 	"path/filepath"
 	"testing"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+	"sourcedock.dev/petrbalvin/gasm-sdk/ast"
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
 )
 
 // The object-file tests share one source: two exported functions, one
@@ -52,7 +53,7 @@ func elfTestImage(t *testing.T) *Image {
 }
 
 // TestAssembleFileExternals checks that a reference to a symbol no GLOBL
-// defines is recorded as an external relocation instead of failing — the
+// defines is recorded as an external relocation instead of failing; the
 // raw image leaves the displacement zero, the object emitters carry it.
 func TestAssembleFileExternals(t *testing.T) {
 	img := elfTestImage(t)
@@ -187,7 +188,7 @@ func TestELFObject(t *testing.T) {
 		end := bytes.IndexByte(strtabRaw[stName:], 0)
 		return string(strtabRaw[stName : int(stName)+end])
 	}
-	for i := 0; i < 2; i++ {
+	for i := range 2 {
 		e := raw[i*24 : (i+1)*24]
 		off := binary.LittleEndian.Uint64(e[0:])
 		info := binary.LittleEndian.Uint64(e[8:])
@@ -211,6 +212,75 @@ func TestELFObject(t *testing.T) {
 	}
 }
 
+// TestELFObjectTLSGuardReloc checks that a non-NOSPLIT function's stack
+// guard carries an R_X86_64_TPOFF32 relocation against the null symbol in
+// .rela.text.  The serialisation must honour the record's type field: a
+// hardcoded R_X86_64_PC32 mislinks the TLS load as an ordinary
+// PC-relative reference.
+func TestELFObjectTLSGuardReloc(t *testing.T) {
+	f, errs := parser.Parse("g_amd64.s", `
+#include "textflag.h"
+
+TEXT ·grow(SB), $0
+	CALL ·other(SB)
+	RET
+
+TEXT ·other(SB), NOSPLIT, $0
+	RET
+`)
+	if len(errs) > 0 {
+		t.Fatalf("parse: %v", errs)
+	}
+	img, err := AssembleFile(f)
+	if err != nil {
+		t.Fatalf("AssembleFile: %v", err)
+	}
+	var haveTLS bool
+	for _, fn := range img.Funcs {
+		for _, r := range fn.Relocs {
+			if r.Kind == RelTLSLE {
+				haveTLS = true
+			}
+		}
+	}
+	if !haveTLS {
+		t.Fatal("test source produced no RelTLSLE relocation")
+	}
+	obj, err := img.ELFObject()
+	if err != nil {
+		t.Fatalf("ELFObject: %v", err)
+	}
+	ef, err := elf.NewFile(bytes.NewReader(obj))
+	if err != nil {
+		t.Fatalf("parse emitted object: %v", err)
+	}
+	defer ef.Close()
+	relaSec := ef.Section(".rela.text")
+	if relaSec == nil {
+		t.Fatal("missing .rela.text")
+	}
+	raw, err := relaSec.Data()
+	if err != nil {
+		t.Fatal(err)
+	}
+	found := false
+	for i := 0; i+24 <= len(raw); i += 24 {
+		e := raw[i:]
+		info := binary.LittleEndian.Uint64(e[8:])
+		typ := info & 0xffffffff
+		sym := int(info >> 32)
+		if typ == uint64(elf.R_X86_64_TPOFF32) {
+			found = true
+			if sym != 0 {
+				t.Errorf("TPOFF32 relocation against symbol %d, want 0 (the null symbol)", sym)
+			}
+		}
+	}
+	if !found {
+		t.Errorf("no R_X86_64_TPOFF32 relocation in .rela.text (%d bytes)", len(raw))
+	}
+}
+
 // TestELFObjectNoRelocations checks a file with no static-symbol references
 // emits a valid object without a .rela.text section.
 func TestELFObjectNoRelocations(t *testing.T) {
@@ -253,6 +323,238 @@ TEXT ·nop(SB), NOSPLIT, $0
 	}
 }
 
+// elfSectionHeaderCount returns the e_shnum the ELF header declares.
+func elfSectionHeaderCount(t *testing.T, obj []byte) int {
+	t.Helper()
+	return int(binary.LittleEndian.Uint16(obj[60:]))
+}
+
+// checkELFSectionAccounting verifies the number of section headers the
+// writer physically laid out equals e_shnum: every DWARF section written
+// after .shstrtab must be counted, or the last ones (always .debug_frame)
+// are invisible to every consumer, debug/elf included.
+func checkELFSectionAccounting(t *testing.T, obj []byte) {
+	t.Helper()
+	shoff := int(binary.LittleEndian.Uint64(obj[40:]))
+	shentsize := int(binary.LittleEndian.Uint16(obj[58:]))
+	shnum := elfSectionHeaderCount(t, obj)
+	if shentsize != 64 {
+		t.Fatalf("e_shentsize = %d, want 64", shentsize)
+	}
+	if (len(obj)-shoff)%shentsize != 0 {
+		t.Fatalf("section header table is not a whole number of entries: shoff=%d len=%d", shoff, len(obj))
+	}
+	if present := (len(obj) - shoff) / shentsize; present != shnum {
+		t.Errorf("e_shnum = %d but %d section headers are laid out", shnum, present)
+	}
+}
+
+// TestELFDWARFSectionAccounting runs the header accounting check over all
+// four architecture emitters, and additionally checks the .debug_frame
+// section is visible (its data aligned as its header declares).
+func TestELFDWARFSectionAccounting(t *testing.T) {
+	parse := func(name, src string) *ast.File {
+		f, errs := parser.Parse(name, src)
+		if len(errs) > 0 {
+			t.Fatalf("parse %s: %v", name, errs)
+		}
+		return f
+	}
+	cases := []struct {
+		name string
+		img  *Image
+		emit func(*Image) ([]byte, error)
+	}{
+		{"amd64", elfTestImage(t), (*Image).ELFObject},
+		{"arm64", mustImage(t, func() (*Image, error) {
+			return AssembleFileARM64(parse("k_arm64.s", `
+#include "textflag.h"
+
+TEXT ·add(SB), NOSPLIT, $0-24
+	MOVD a+0(FP), R4
+	MOVD b+8(FP), R5
+	ADD R5, R4, R4
+	MOVD R4, ret+16(FP)
+	RET
+`))
+		}), (*Image).ELFAARCH64Object},
+		{"riscv64", mustImage(t, func() (*Image, error) {
+			return AssembleFileRISCV(parse("k_riscv64.s", `
+#include "textflag.h"
+
+TEXT ·sb(SB), NOSPLIT, $0-0
+	MOV $answer<>(SB), X10
+	RET
+
+GLOBL answer<>(SB), RODATA, $8
+DATA answer<>+0(SB)/8, $42
+`))
+		}), (*Image).ELFRISCVObject},
+		{"loong64", mustImage(t, func() (*Image, error) {
+			return AssembleFileLOONG64(parse("k_loong64.s", `
+#include "textflag.h"
+
+TEXT ·add(SB), NOSPLIT, $0-24
+	MOVV a+0(FP), R4
+	MOVV b+8(FP), R5
+	ADDV R5, R4, R4
+	MOVV R4, ret+16(FP)
+	RET
+`))
+		}), (*Image).ELFLOONG64Object},
+	}
+	for _, tc := range cases {
+		obj, err := tc.emit(tc.img)
+		if err != nil {
+			t.Fatalf("%s: emit: %v", tc.name, err)
+		}
+		checkELFSectionAccounting(t, obj)
+		ef, err := elf.NewFile(bytes.NewReader(obj))
+		if err != nil {
+			t.Fatalf("%s: parse emitted object: %v", tc.name, err)
+		}
+		frame := ef.Section(".debug_frame")
+		if frame == nil {
+			t.Errorf("%s: .debug_frame invisible to debug/elf (e_shnum too small?)", tc.name)
+			ef.Close()
+			continue
+		}
+		if frame.Offset%8 != 0 || frame.Addralign != 8 {
+			t.Errorf("%s: .debug_frame offset %d align %d, want offset%%8==0 align 8", tc.name, frame.Offset, frame.Addralign)
+		}
+		ef.Close()
+	}
+}
+
+func mustImage(t *testing.T, f func() (*Image, error)) *Image {
+	t.Helper()
+	img, err := f()
+	if err != nil {
+		t.Fatal(err)
+	}
+	return img
+}
+
+// TestELFDWARFRelocations checks the .rela.debug_info and .rela.debug_line
+// sections exist and carry absolute 64-bit relocations against the
+// function symbols, with r_offsets inside their target sections.
+func TestELFDWARFRelocations(t *testing.T) {
+	img := elfTestImage(t)
+	obj, err := img.ELFObject()
+	if err != nil {
+		t.Fatalf("ELFObject: %v", err)
+	}
+	ef, err := elf.NewFile(bytes.NewReader(obj))
+	if err != nil {
+		t.Fatalf("parse emitted object: %v", err)
+	}
+	defer ef.Close()
+	// The DWARF must record the assembled file's path (threaded through
+	// Image.SourcePath), not a placeholder name.
+	info, err := ef.Section(".debug_info").Data()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if img.SourcePath != "t_amd64.s" || !bytes.Contains(info, []byte(img.SourcePath)) {
+		t.Errorf("DWARF compilation unit does not name the source %q", img.SourcePath)
+	}
+	for _, tc := range []struct {
+		rela   string
+		target string
+		want   uint32
+	}{
+		{".rela.debug_info", ".debug_info", rX8664Abs64},
+		{".rela.debug_line", ".debug_line", rX8664Abs64},
+		{".rela.debug_frame", ".debug_frame", rX8664Abs64},
+	} {
+		rs := ef.Section(tc.rela)
+		if rs == nil {
+			t.Fatalf("missing %s", tc.rela)
+		}
+		if rs.Type != elf.SHT_RELA {
+			t.Errorf("%s: type %v, want SHT_RELA", tc.rela, rs.Type)
+		}
+		target := ef.Section(tc.target)
+		if target == nil {
+			t.Fatalf("missing %s", tc.target)
+		}
+		if rs.Link == 0 || ef.Sections[rs.Info] != target {
+			t.Errorf("%s: link %d info %d, want the symtab and %s", tc.rela, rs.Link, rs.Info, tc.target)
+		}
+		b, err := rs.Data()
+		if err != nil {
+			t.Fatal(err)
+		}
+		// .debug_line has one address per function; .debug_info adds the
+		// compile unit's own low_pc.
+		want := len(img.Funcs)
+		if tc.target == ".debug_info" {
+			want++
+		}
+		if len(b)/24 != want {
+			t.Errorf("%s: %d entries, want %d", tc.rela, len(b)/24, want)
+		}
+		for i := 0; i+24 <= len(b); i += 24 {
+			r_offset := binary.LittleEndian.Uint64(b[i:])
+			info := binary.LittleEndian.Uint64(b[i+8:])
+			typ := uint32(info)
+			sym := int(info >> 32)
+			if typ != tc.want {
+				t.Errorf("%s entry %d: type %d, want R_X86_64_64 (%d)", tc.rela, i/24, typ, tc.want)
+			}
+			if r_offset >= uint64(target.Size) {
+				t.Errorf("%s entry %d: r_offset %d outside %s (%d bytes)", tc.rela, i/24, r_offset, tc.target, target.Size)
+			}
+			if sym == 0 {
+				t.Errorf("%s entry %d: against the null symbol", tc.rela, i/24)
+			}
+		}
+	}
+}
+
+// TestELFDataOnly checks a source with GLOBL data and no TEXT emits a valid
+// ELF object: the DWARF compilation unit of a code-less image has no
+// function to relocate against and must not reach for one.
+func TestELFDataOnly(t *testing.T) {
+	f, errs := parser.Parse("d0_amd64.s", `
+GLOBL table<>(SB), RODATA, $8
+DATA table<>+0(SB)/8, $12345
+`)
+	if len(errs) > 0 {
+		t.Fatalf("parse: %v", errs)
+	}
+	img, err := AssembleFile(f)
+	if err != nil {
+		t.Fatalf("AssembleFile: %v", err)
+	}
+	obj, err := img.ELFObject()
+	if err != nil {
+		t.Fatalf("ELFObject: %v", err)
+	}
+	checkELFSectionAccounting(t, obj)
+	ef, err := elf.NewFile(bytes.NewReader(obj))
+	if err != nil {
+		t.Fatalf("parse emitted object: %v", err)
+	}
+	defer ef.Close()
+	syms, err := ef.Symbols()
+	if err != nil {
+		t.Fatal(err)
+	}
+	found := false
+	for _, s := range syms {
+		if s.Name == "table" && s.Size == 8 {
+			found = true
+		}
+	}
+	if !found {
+		t.Errorf("data symbol table missing: %v", syms)
+	}
+	if ef.Section(".rela.debug_info") != nil || ef.Section(".rela.debug_line") != nil {
+		t.Error("data-only image must not emit DWARF address relocations")
+	}
+}
+
 // TestELFLinkAndRun is the end-to-end check: assemble the test functions,
 // link the emitted object with a C driver that defines the external symbol,
 // and run the result.  Skipped when no C compiler is available.
@@ -307,4 +609,252 @@ int main(void) {
 	if got := string(run); got != "42 42 7\n" {
 		t.Errorf("output %q, want \"42 42 7\\n\"", got)
 	}
+
+	// The DWARF addresses must have resolved at link time: the .debug_info
+	// placeholders were carried by .rela.debug_info, so every subprogram's
+	// low_pc must now equal its linked symbol address.
+	bin, err := os.ReadFile(appPath)
+	if err != nil {
+		t.Fatal(err)
+	}
+	lef, err := elf.NewFile(bytes.NewReader(bin))
+	if err != nil {
+		t.Fatalf("parse linked binary: %v", err)
+	}
+	defer lef.Close()
+	syms, err := lef.Symbols()
+	if err != nil {
+		t.Fatal(err)
+	}
+	addrByName := map[string]uint64{}
+	for _, s := range syms {
+		if elf.ST_TYPE(s.Info) == elf.STT_FUNC && s.Value != 0 {
+			addrByName[s.Name] = s.Value
+		}
+	}
+	lowPCs := dwarfSubprogramLowPCs(t, lef)
+	if len(lowPCs) == 0 {
+		t.Fatal("no subprogram DW_AT_low_pc parsed from the linked binary")
+	}
+	for name, pc := range lowPCs {
+		addr, ok := addrByName[name]
+		if !ok {
+			t.Errorf("subprogram %q not in the linked symbol table", name)
+			continue
+		}
+		if pc != addr {
+			t.Errorf("subprogram %q: DW_AT_low_pc = %#x, linked address %#x (DWARF relocation unresolved)", name, pc, addr)
+		}
+	}
+}
+
+// dwarfSubprogramLowPCs walks the linked binary's .debug_info with its own
+// .debug_abbrev and returns each DW_TAG_subprogram's DW_AT_low_pc by name.
+func dwarfSubprogramLowPCs(t *testing.T, ef *elf.File) map[string]uint64 {
+	t.Helper()
+	abbrevSec := ef.Section(".debug_abbrev")
+	infoSec := ef.Section(".debug_info")
+	if abbrevSec == nil || infoSec == nil {
+		t.Fatal("linked binary lacks .debug_abbrev or .debug_info")
+	}
+	abbrev, err := abbrevSec.Data()
+	if err != nil {
+		t.Fatal(err)
+	}
+	info, err := infoSec.Data()
+	if err != nil {
+		t.Fatal(err)
+	}
+	abs := parseAbbrevs(t, abbrev)
+	le := binary.LittleEndian
+	out := map[string]uint64{}
+	r := &ulebIter{b: info}
+	r.uint32At(t) // unit_length
+	if v := le.Uint16(info[4:]); v != 5 {
+		t.Fatalf(".debug_info version %d, want 5", v)
+	}
+	r.i = 6
+	r.byteAt(t)   // unit_type
+	r.byteAt(t)   // address_size
+	r.uint32At(t) // debug_abbrev_offset
+	var name string
+	var lowPC uint64
+	for r.i < len(r.b) {
+		code := r.uleb(t)
+		if code == 0 {
+			continue // end of the CU's children
+		}
+		ab, ok := abs[code]
+		if !ok {
+			t.Fatalf("unknown abbreviation code %d", code)
+		}
+		name, lowPC = "", 0
+		for _, a := range ab.attrs {
+			switch a.attr {
+			case dwAtName:
+				readFormKeep(t, r, a.form, &name, nil)
+			case dwAtLowPC:
+				readFormKeep(t, r, a.form, nil, &lowPC)
+			default:
+				readFormSkip(t, r, a.form)
+			}
+		}
+		if ab.tag == dwTagSubprog && name != "" {
+			out[name] = lowPC
+		}
+	}
+	return out
+}
+
+// readFormKeep reads one DIE attribute value, keeping a string or an
+// address into the pointer it was given (nil keeps nothing).
+func readFormKeep(t *testing.T, r *ulebIter, form uint64, name *string, addr *uint64) {
+	t.Helper()
+	switch form {
+	case dwFormString:
+		end := r.i
+		for end < len(r.b) && r.b[end] != 0 {
+			end++
+		}
+		if name != nil {
+			*name = string(r.b[r.i:end])
+		}
+		r.i = end + 1
+	case dwFormAddr:
+		if addr != nil {
+			*addr = binary.LittleEndian.Uint64(r.b[r.i:])
+		}
+		r.i += 8
+	default:
+		readFormSkip(t, r, form)
+	}
+}
+
+func readFormSkip(t *testing.T, r *ulebIter, form uint64) {
+	t.Helper()
+	switch form {
+	case dwFormString:
+		for r.i < len(r.b) && r.b[r.i] != 0 {
+			r.i++
+		}
+		r.i++
+	case dwFormAddr, dwFormData8:
+		r.i += 8
+	case dwFormSecOff:
+		r.i += 4
+	case dwFormExprloc:
+		r.i += int(r.uleb(t))
+	case dwFormData1, 0x0c:
+		r.i++
+	default:
+		t.Fatalf("unsupported form %#x", form)
+	}
+}
+
+// TestELFObjectDataRelocation checks that a symbol-valued DATA field ("DATA
+// s+0(SB)/8, $other(SB)") reaches the ELF object as a .rela.data entry: an
+// absolute 64-bit relocation at the field's offset within .data, against
+// the named symbol, external targets included.
+func TestELFObjectDataRelocation(t *testing.T) {
+	f, errs := parser.Parse("t_amd64.s", `#include "textflag.h"
+TEXT ·Keep(SB), NOSPLIT, $0-8
+	RET
+GLOBL holder(SB), NOPTR, $24
+DATA holder+0(SB)/8, $·Keep+5(SB)
+DATA holder+8(SB)/8, $holder(SB)
+DATA holder+16(SB)/8, $extvar(SB)
+`)
+	if len(errs) > 0 {
+		t.Fatalf("parse: %v", errs)
+	}
+	img, err := AssembleFile(f)
+	if err != nil {
+		t.Fatalf("assemble: %v", err)
+	}
+	obj, err := img.ELFObject()
+	if err != nil {
+		t.Fatalf("ELFObject: %v", err)
+	}
+	ef, err := elf.NewFile(bytes.NewReader(obj))
+	if err != nil {
+		t.Fatalf("parse emitted object: %v", err)
+	}
+	defer ef.Close()
+	relaData := ef.Section(".rela.data")
+	if relaData == nil {
+		t.Fatal("missing .rela.data section")
+	}
+	if relaData.Link == 0 || ef.Sections[relaData.Link].Name != ".symtab" {
+		t.Errorf(".rela.data sh_link = %d, want the .symtab index", relaData.Link)
+	}
+	if ef.Sections[relaData.Info].Name != ".data" {
+		t.Errorf(".rela.data sh_info = %d, want the .data index", relaData.Info)
+	}
+	relas, err := relaData.Data()
+	if err != nil {
+		t.Fatal(err)
+	}
+	var got []struct {
+		off    uint64
+		sym    uint32
+		typ    uint32
+		addend int64
+	}
+	for i := 0; i+24 <= len(relas); i += 24 {
+		got = append(got, struct {
+			off    uint64
+			sym    uint32
+			typ    uint32
+			addend int64
+		}{
+			off: binary.LittleEndian.Uint64(relas[i:]),
+			// r_info packs the type in the low dword and the symbol index
+			// in the high dword.
+			typ:    binary.LittleEndian.Uint32(relas[i+8:]),
+			sym:    binary.LittleEndian.Uint32(relas[i+12:]),
+			addend: int64(binary.LittleEndian.Uint64(relas[i+16:])),
+		})
+	}
+	// debug/elf hides the table's null entry, so raw index s names syms[s-1].
+	syms, err := ef.Symbols()
+	if err != nil {
+		t.Fatal(err)
+	}
+	name := func(idx uint32) string {
+		if idx >= 1 && int(idx) <= len(syms) {
+			return syms[idx-1].Name
+		}
+		return ""
+	}
+	// The offsets are data-section-relative: the field's DATA offset plus
+	// the symbol's position in .data (the layout aligns each symbol to 16).
+	base := uint64(0)
+	for _, d := range img.DataSyms {
+		if d.Name == "holder" {
+			base = uint64(d.Offset)
+		}
+	}
+	want := []struct {
+		off    uint64
+		typ    uint32
+		addend int64
+		target string
+	}{
+		{off: base + 0, typ: uint32(elf.R_X86_64_64), addend: 5, target: "Keep"},
+		{off: base + 8, typ: uint32(elf.R_X86_64_64), addend: 0, target: "holder"},
+		{off: base + 16, typ: uint32(elf.R_X86_64_64), addend: 0, target: "extvar"},
+	}
+	if len(got) != len(want) {
+		t.Fatalf(".rela.data entries = %d, want %d", len(got), len(want))
+	}
+	for i, w := range want {
+		g := got[i]
+		if g.off != w.off || g.typ != w.typ || g.addend != w.addend {
+			t.Errorf("entry %d = {off %d typ %d addend %d}, want {off %d typ %d addend %d}",
+				i, g.off, g.typ, g.addend, w.off, w.typ, w.addend)
+		}
+		if n := name(g.sym); n != w.target {
+			t.Errorf("entry %d names %q, want %q", i, n, w.target)
+		}
+	}
 }

asm/elfarm64.go

@@ -15,13 +15,19 @@ const (
 
 	// AArch64 relocation types (the ELF psABI).
 	rArm64PrelPgHi21   = 275 // R_AARCH64_ADR_PREL_PG_HI21 (ADRP page)
-	rArm64AddAbsLo12NC = 277 // R_AARCH64_ADD_ABS_LO12_NC (ADD/STR/LDR page offset)
+	rArm64AddAbsLo12NC = 277 // R_AARCH64_ADD_ABS_LO12_NC (ADD page offset)
+	rArm64Call26       = 283 // R_AARCH64_CALL26 (BL instruction)
+	rArm64Ldst64Lo12NC = 286 // R_AARCH64_LDST64_ABS_LO12_NC (64-bit LDR/STR page offset)
+	// R_AARCH64_ABS32 (debug/elf 258): the absolute 32-bit address of a
+	// symbol, the R_ADDR shape a 4-byte DATA field carries.  ABS64 (257)
+	// lives with the DWARF fixup constants as rAARCH64Abs64.
+	rArm64Abs32 = 258
 )
 
 // ELFAARCH64Object returns the image as an ELF64 relocatable object file for
 // AArch64 (EM_AARCH64, 64-bit, little-endian).  The structure mirrors the
-// amd64 and RISC-V ELF emitters: .text, .data, .symtab, .strtab and an
-// optional .rela.text.
+// amd64 and RISC-V ELF emitters: .text, .data, .symtab, .strtab, an
+// optional .rela.text and an optional .rela.data.
 func (img *Image) ELFAARCH64Object() ([]byte, error) {
 	le := binary.LittleEndian
 
@@ -79,8 +85,19 @@ func (img *Image) ELFAARCH64Object() ([]byte, error) {
 	}
 
 	// Build relocations.  Each SB reference is an ADRP pair:
-	//   ADRP Rd, 0  → R_AARCH64_ADR_PREL_PG_HI21
-	//   ADD/LDR/STR  → R_AARCH64_ADD_ABS_LO12_NC
+	//   ADRP Rd, 0  → R_AARCH64_ADR_PREL_PG_HI21 at the ADRP
+	//   ADD         → R_AARCH64_ADD_ABS_LO12_NC at the ADD word
+	//   LDR/STR X   → R_AARCH64_LDST64_ABS_LO12_NC at the LDR/STR word
+	//   BL          → R_AARCH64_CALL26
+	// cmd/link's own conversion emits the HI21 at sectoff and the LO12 at
+	// sectoff+4 (cmd/link/internal/arm64/asm.go), so the ADD or load word
+	// carries the page-offset relocation, never a second HI21.  The
+	// assembler records two RelArm64Addr relocs per ADRP+ADD pair (one per
+	// word), so the second of the pair is consumed here.
+	// Addends stay raw: ADR_PREL_PG_HI21 and the ABS_LO12_NC forms resolve
+	// against S+A, and CALL26 branches take the branch instruction's own
+	// place as the PC-relative base, so subtracting the field width (the
+	// amd64 R_PCREL convention) would misplace every branch by 4 bytes.
 	type elfRela struct {
 		off    uint64
 		typ    uint32
@@ -89,25 +106,81 @@ func (img *Image) ELFAARCH64Object() ([]byte, error) {
 	}
 	var relas []elfRela
 	for _, fn := range img.Funcs {
-		for _, r := range fn.Relocs {
+		for i := 0; i < len(fn.Relocs); i++ {
+			r := fn.Relocs[i]
 			idx, ok := symIdx[r.Name]
 			if !ok {
 				return nil, fmt.Errorf("relocation references unknown symbol %q", r.Name)
 			}
-			typ := uint32(rArm64PrelPgHi21)
-			if r.Kind == RelArm64Addr && r.Off%4 == 4 {
-				// The second instruction in an ADRP pair uses ADD_ABS_LO12_NC.
-				typ = rArm64AddAbsLo12NC
+			switch r.Kind {
+			case RelArm64Branch:
+				relas = append(relas, elfRela{
+					off: uint64(fn.Offset + r.Off), typ: rArm64Call26, sym: idx, addend: r.Addend,
+				})
+			case RelArm64Addr:
+				// ADRP+ADD: the pair's second reloc (at Off+4) is the
+				// assembler's twin of the same pair; skip it.
+				relas = append(relas,
+					elfRela{off: uint64(fn.Offset + r.Off), typ: rArm64PrelPgHi21, sym: idx, addend: r.Addend},
+					elfRela{off: uint64(fn.Offset + r.Off + 4), typ: rArm64AddAbsLo12NC, sym: idx, addend: r.Addend},
+				)
+				i++
+			case RelArm64LDST64:
+				// ADRP+LDR/STR: one assembler reloc covers the pair.
+				relas = append(relas,
+					elfRela{off: uint64(fn.Offset + r.Off), typ: rArm64PrelPgHi21, sym: idx, addend: r.Addend},
+					elfRela{off: uint64(fn.Offset + r.Off + 4), typ: rArm64Ldst64Lo12NC, sym: idx, addend: r.Addend},
+				)
+			default:
+				return nil, fmt.Errorf("relocation kind %v unsupported in ELF emission", r.Kind)
 			}
-			relas = append(relas, elfRela{
-				off:    uint64(fn.Offset + r.Off),
-				typ:    typ,
+		}
+	}
+
+	// The data symbols' symbol-valued DATA fields ("DATA s+0(SB)/8,
+	// $other(SB)") become .rela.data entries: an absolute relocation of the
+	// DATA line's width at the field's data-section offset, S + A with no
+	// PC term.  Widths 4 and 8 have ELF relocation shapes; narrower fields
+	// cannot hold an address, so they are refused rather than truncated.
+	var dataRelas []elfRela
+	for _, d := range img.DataSyms {
+		for _, r := range d.Relocs {
+			idx, ok := symIdx[r.Name]
+			if !ok {
+				return nil, fmt.Errorf("data relocation references unknown symbol %q", r.Name)
+			}
+			var typ uint32
+			switch r.Siz {
+			case 8:
+				typ = rAARCH64Abs64
+			case 4:
+				typ = rArm64Abs32
+			default:
+				return nil, fmt.Errorf("DATA %q: a symbol value of width %d has no ELF relocation", d.Name, r.Siz)
+			}
+			dataRelas = append(dataRelas, elfRela{
+				off:    uint64(d.Offset + r.Off),
 				sym:    idx,
-				addend: r.Addend - int64(r.After-r.Off),
+				typ:    typ,
+				addend: r.Addend,
 			})
 		}
 	}
 
+	// Section presence: .rela.text only when there are code relocations,
+	// .rela.data only when a DATA line holds a symbol value.
+	hasRela := len(relas) > 0
+	hasDataRela := len(dataRelas) > 0
+	nSections := 6
+	if hasRela {
+		nSections++
+	}
+	if hasDataRela {
+		nSections++
+	}
+	secSymtab, secStrtab := 3, 4
+	secShstr := nSections - 1
+
 	// String tables.
 	stNames := newElfStrtab()
 	for _, s := range syms {
@@ -117,14 +190,12 @@ func (img *Image) ELFAARCH64Object() ([]byte, error) {
 	for _, n := range []string{".text", ".data", ".symtab", ".strtab", ".rela.text", ".shstrtab"} {
 		stSections.add(n)
 	}
-
-	hasRela := len(relas) > 0
-	nSections := 6
-	if hasRela {
-		nSections = 7
+	if hasDataRela {
+		stSections.add(".rela.data")
+	}
+	for _, n := range dwarfSectionNames {
+		stSections.add(n)
 	}
-	secSymtab, secStrtab := 3, 4
-	secShstr := nSections - 1
 
 	// Layout.
 	var out []byte
@@ -160,7 +231,7 @@ func (img *Image) ELFAARCH64Object() ([]byte, error) {
 	strtabOff := len(out)
 	out = append(out, stNames.bytes()...)
 
-	var relaOff int
+	var relaOff, relaDataOff int
 	if hasRela {
 		align(8)
 		relaOff = len(out)
@@ -172,10 +243,49 @@ func (img *Image) ELFAARCH64Object() ([]byte, error) {
 			out = append(out, b[:]...)
 		}
 	}
+	if hasDataRela {
+		align(8)
+		relaDataOff = len(out)
+		for _, r := range dataRelas {
+			var b [24]byte
+			le.PutUint64(b[0:], r.off)
+			le.PutUint64(b[8:], uint64(r.sym)<<32|uint64(r.typ))
+			le.PutUint64(b[16:], uint64(r.addend))
+			out = append(out, b[:]...)
+		}
+	}
 
 	shstrOff := len(out)
 	out = append(out, stSections.bytes()...)
 
+	// DWARF debug sections; the address placeholders they leave are carried
+	// as .rela.debug_info/.rela.debug_line entries the system linker applies.
+	dwAlign := func(n int) {
+		for len(out)%n != 0 {
+			out = append(out, 0)
+		}
+	}
+	dw := appendDWARFSections(&out, img, dwarfSourceName(img), symIdx, dwAlign, cfiARM64)
+	dwarfStart := 0 // section index of .debug_abbrev, set when DWARF is present
+	if dw != nil {
+		// Five DWARF sections: .debug_abbrev, .debug_info, .debug_line,
+		// .debug_line_str and .debug_frame (the CIE is unconditional, so
+		// the frame section is always present), plus the relocation
+		// sections below when they carry entries.
+		dwarfStart = nSections
+		nSections += 5
+		appendDWARFRelas(&out, dw, rAARCH64Abs64, dwAlign)
+		if dw.infoRelaCount > 0 {
+			nSections++
+		}
+		if dw.lineRelaCount > 0 {
+			nSections++
+		}
+		if dw.frameRelaCount > 0 {
+			nSections++
+		}
+	}
+
 	align(8)
 	shoff := len(out)
 
@@ -201,7 +311,43 @@ func (img *Image) ELFAARCH64Object() ([]byte, error) {
 	if hasRela {
 		putSh(".rela.text", shtRela, 0, relaOff, 24*len(relas), secSymtab, secText, 8, 24)
 	}
+	if hasDataRela {
+		putSh(".rela.data", shtRela, 0, relaDataOff, 24*len(dataRelas), secSymtab, secData, 8, 24)
+	}
 	putSh(".shstrtab", shtStrtab, 0, shstrOff, len(stSections.bytes()), 0, 0, 1, 0)
+	// DWARF section headers; their indices follow the write order.
+	if dw != nil {
+		// secIdx is a running section index: each putSh below emits the
+		// next header, and the sh_info of a .rela section names the index
+		// of the section it relocates.
+		secIdx := dwarfStart
+		putSh(".debug_abbrev", shtProgbits, 0, dw.abbrevOff, dw.abbrevSize, 0, 0, 1, 0)
+		secIdx++
+		putSh(".debug_info", shtProgbits, 0, dw.infoOff, dw.infoSize, 0, 0, 1, 0)
+		secInfoIdx := secIdx
+		secIdx++
+		if dw.infoRelaCount > 0 {
+			putSh(".rela.debug_info", shtRela, 0, dw.infoRelaOff, 24*dw.infoRelaCount, secSymtab, secInfoIdx, 8, 24)
+			secIdx++
+		}
+		putSh(".debug_line", shtProgbits, 0, dw.lineOff, dw.lineSize, 0, 0, 1, 0)
+		secLineIdx := secIdx
+		secIdx++
+		if dw.lineRelaCount > 0 {
+			putSh(".rela.debug_line", shtRela, 0, dw.lineRelaOff, 24*dw.lineRelaCount, secSymtab, secLineIdx, 8, 24)
+			secIdx++
+		}
+		putSh(".debug_line_str", shtProgbits, 0, dw.lineStrOff, dw.lineStrSize, 0, 0, 1, 0)
+		secIdx++
+		if dw.frameSize > 0 {
+			putSh(".debug_frame", shtProgbits, 0, dw.frameOff, dw.frameSize, 0, 0, 8, 0)
+			secFrameIdx := secIdx
+			secIdx++
+			if dw.frameRelaCount > 0 {
+				putSh(".rela.debug_frame", shtRela, 0, dw.frameRelaOff, 24*dw.frameRelaCount, secSymtab, secFrameIdx, 8, 24)
+			}
+		}
+	}
 
 	// ELF header.
 	hdr := out[:64]

asm/elfarm64_test.go

@@ -6,9 +6,10 @@ package asm
 import (
 	"bytes"
 	"debug/elf"
+	"encoding/binary"
 	"testing"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
 )
 
 // TestELFAARCH64Object checks the structure of the emitted AArch64 ELF64
@@ -28,6 +29,7 @@ TEXT ·add(SB), NOSPLIT, $0-24
 
 TEXT ·getanswer(SB), NOSPLIT, $0-8
 	MOVD answer<>(SB), R4
+	MOVD $answer<>(SB), R5
 	MOVD R4, ret+0(FP)
 	RET
 
@@ -102,8 +104,63 @@ DATA answer<>+0(SB)/8, $42
 	// Check that .rela.text exists (getanswer has SB reference).
 	relaText := ef.Section(".rela.text")
 	if relaText == nil {
-		t.Error("missing .rela.text section")
+		t.Fatal("missing .rela.text section")
 	}
+
+	// The SB references of getanswer form two ADRP pairs: the load
+	// (MOVD answer<>(SB), R4) is ADRP+LDR carrying HI21 at the ADRP and
+	// LDST64_ABS_LO12_NC at the LDR word, and the address-of
+	// (MOVD $answer<>(SB), R5) is ADRP+ADD carrying HI21 and
+	// ADD_ABS_LO12_NC.  cmd/link's own conversion emits exactly this
+	// sectoff / sectoff+4 pairing; a second HI21 at the ADD or LDR word
+	// corrupts the pair.
+	raw, err := relaText.Data()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(raw)%24 != 0 || len(raw)/24 != 4 {
+		t.Fatalf(".rela.text has %d bytes, want four 24-byte entries", len(raw))
+	}
+	wantRela := []struct {
+		typ elf.R_AARCH64
+		off uint64 // relative to the getanswer function start
+	}{
+		{elf.R_AARCH64_ADR_PREL_PG_HI21, 0},
+		{elf.R_AARCH64_LDST64_ABS_LO12_NC, 4},
+		{elf.R_AARCH64_ADR_PREL_PG_HI21, 8},
+		{elf.R_AARCH64_ADD_ABS_LO12_NC, 12},
+	}
+	getanswer := byNameElf(t, ef, "getanswer")
+	for i, w := range wantRela {
+		e := raw[i*24 : (i+1)*24]
+		off := binary.LittleEndian.Uint64(e[0:])
+		info := binary.LittleEndian.Uint64(e[8:])
+		typ := elf.R_AARCH64(info & 0xffffffff)
+		sym := int(info >> 32)
+		if typ != w.typ || off != getanswer.Value+w.off {
+			t.Errorf("reloc %d: type %v off %d, want %v at %d", i, typ, off, w.typ, getanswer.Value+w.off)
+		}
+		if sym != 3 { // NULL, .text, .data, then the first local: answer
+			t.Errorf("reloc %d: symbol index %d, want 3 (answer)", i, sym)
+		}
+	}
+}
+
+// byNameElf returns the symbol table entry for name from the raw .symtab,
+// which carries every entry including the null and section symbols in order.
+func byNameElf(t *testing.T, ef *elf.File, name string) elf.Symbol {
+	t.Helper()
+	syms, err := ef.Symbols()
+	if err != nil {
+		t.Fatalf("symbols: %v", err)
+	}
+	for _, s := range syms {
+		if s.Name == name {
+			return s
+		}
+	}
+	t.Fatalf("symbol %q not found", name)
+	return elf.Symbol{}
 }
 
 // TestELFAARCH64ObjectNoRelocations checks the ELF output when there are no
@@ -140,3 +197,118 @@ TEXT ·add(SB), NOSPLIT, $0-24
 		t.Error("unexpected .rela.text section when there are no relocations")
 	}
 }
+
+// TestELFAARCH64ObjectDataRelocation checks that a symbol-valued DATA field
+// ("DATA s+0(SB)/8, $other(SB)") reaches the AArch64 ELF object as a
+// .rela.data entry: an R_AARCH64_ABS64 (ABS32 for a width-4 field) at the
+// field's offset within .data, against the named symbol, external targets
+// included.
+func TestELFAARCH64ObjectDataRelocation(t *testing.T) {
+	f, errs := parser.Parse("t_arm64.s", `#include "textflag.h"
+TEXT ·Keep(SB), NOSPLIT, $0-0
+	RET
+GLOBL holder(SB), NOPTR, $32
+DATA holder+0(SB)/8, $·Keep+5(SB)
+DATA holder+8(SB)/8, $holder(SB)
+DATA holder+16(SB)/8, $extvar(SB)
+DATA holder+24(SB)/4, $Keep(SB)
+`)
+	if len(errs) > 0 {
+		t.Fatalf("parse: %v", errs)
+	}
+	img, err := AssembleFileARM64(f)
+	if err != nil {
+		t.Fatalf("AssembleFileARM64: %v", err)
+	}
+	obj, err := img.ELFAARCH64Object()
+	if err != nil {
+		t.Fatalf("ELFAARCH64Object: %v", err)
+	}
+	checkELFSectionAccounting(t, obj)
+	ef, err := elf.NewFile(bytes.NewReader(obj))
+	if err != nil {
+		t.Fatalf("parse emitted object: %v", err)
+	}
+	defer ef.Close()
+	relaData := ef.Section(".rela.data")
+	if relaData == nil {
+		t.Fatal("missing .rela.data section")
+	}
+	if relaData.Type != elf.SHT_RELA {
+		t.Errorf(".rela.data type = %v, want SHT_RELA", relaData.Type)
+	}
+	if relaData.Link == 0 || ef.Sections[relaData.Link].Name != ".symtab" {
+		t.Errorf(".rela.data sh_link = %d, want the .symtab index", relaData.Link)
+	}
+	if ef.Sections[relaData.Info].Name != ".data" {
+		t.Errorf(".rela.data sh_info = %d, want the .data index", relaData.Info)
+	}
+	relas, err := relaData.Data()
+	if err != nil {
+		t.Fatal(err)
+	}
+	var got []struct {
+		off    uint64
+		sym    uint32
+		typ    uint32
+		addend int64
+	}
+	for i := 0; i+24 <= len(relas); i += 24 {
+		got = append(got, struct {
+			off    uint64
+			sym    uint32
+			typ    uint32
+			addend int64
+		}{
+			off: binary.LittleEndian.Uint64(relas[i:]),
+			// r_info packs the type in the low dword and the symbol index
+			// in the high dword.
+			typ:    binary.LittleEndian.Uint32(relas[i+8:]),
+			sym:    binary.LittleEndian.Uint32(relas[i+12:]),
+			addend: int64(binary.LittleEndian.Uint64(relas[i+16:])),
+		})
+	}
+	// debug/elf hides the table's null entry, so raw index s names syms[s-1].
+	syms, err := ef.Symbols()
+	if err != nil {
+		t.Fatal(err)
+	}
+	name := func(idx uint32) string {
+		if idx >= 1 && int(idx) <= len(syms) {
+			return syms[idx-1].Name
+		}
+		return ""
+	}
+	// The offsets are data-section-relative: the field's DATA offset plus
+	// the symbol's position in .data (the layout aligns each symbol to 16).
+	base := uint64(0)
+	for _, d := range img.DataSyms {
+		if d.Name == "holder" {
+			base = uint64(d.Offset)
+		}
+	}
+	want := []struct {
+		off    uint64
+		typ    uint32
+		addend int64
+		target string
+	}{
+		{off: base + 0, typ: uint32(elf.R_AARCH64_ABS64), addend: 5, target: "Keep"},
+		{off: base + 8, typ: uint32(elf.R_AARCH64_ABS64), addend: 0, target: "holder"},
+		{off: base + 16, typ: uint32(elf.R_AARCH64_ABS64), addend: 0, target: "extvar"},
+		{off: base + 24, typ: uint32(elf.R_AARCH64_ABS32), addend: 0, target: "Keep"},
+	}
+	if len(got) != len(want) {
+		t.Fatalf(".rela.data entries = %d, want %d", len(got), len(want))
+	}
+	for i, w := range want {
+		g := got[i]
+		if g.off != w.off || g.typ != w.typ || g.addend != w.addend {
+			t.Errorf("entry %d = {off %d typ %d addend %d}, want {off %d typ %d addend %d}",
+				i, g.off, g.typ, g.addend, w.off, w.typ, w.addend)
+		}
+		if n := name(g.sym); n != w.target {
+			t.Errorf("entry %d names %q, want %q", i, n, w.target)
+		}
+	}
+}

asm/elfloong64.go

@@ -13,15 +13,26 @@ import (
 const (
 	emLOONGARCH = 258 // EM_LOONGARCH
 
+	// EF_LOONGARCH_ABI_DOUBLE_FLOAT | EF_LOONGARCH_OBJABI_V1: the flags the
+	// Go toolchain writes (cmd/link/internal/ld/elf.go: Flags = 0x43 for
+	// Loong64).  System linkers refuse to merge ET_REL objects whose float
+	// ABI differs, so 0 (soft-float) would make the object unlinkable.
+	efLarchAbiDoubleObjV1 = 0x43
+
 	// LoongArch relocation types (the ELF psABI).
 	rLarchPCALAHI20 = 71 // R_LARCH_PCALA_HI20 (pcalau12i)
 	rLarchPCALALO12 = 72 // R_LARCH_PCALA_LO12 (addi.d/ld/st)
+	rLarchB26       = 66 // R_LARCH_B26 (b/bl, matches the Go linker's mapping)
+	// R_LARCH_32 (debug/elf 1): the absolute 32-bit address of a symbol,
+	// the R_ADDR shape a 4-byte DATA field carries.  R_LARCH_64 (2) lives
+	// with the DWARF fixup constants as rLarchAbs64.
+	rLarchAbs32 = 1
 )
 
 // ELFLOONG64Object returns the image as an ELF64 relocatable object file for
 // LoongArch (EM_LOONGARCH, 64-bit, little-endian).  The structure mirrors the
-// amd64 and RISC-V ELF emitters: .text, .data, .symtab, .strtab and an
-// optional .rela.text.
+// amd64 and RISC-V ELF emitters: .text, .data, .symtab, .strtab, an
+// optional .rela.text and an optional .rela.data.
 func (img *Image) ELFLOONG64Object() ([]byte, error) {
 	le := binary.LittleEndian
 
@@ -95,18 +106,65 @@ func (img *Image) ELFLOONG64Object() ([]byte, error) {
 				return nil, fmt.Errorf("relocation references unknown symbol %q", r.Name)
 			}
 			typ := uint32(rLarchPCALAHI20)
-			if r.Kind == RelLoong64AddrLo {
+			switch r.Kind {
+			case RelLoong64AddrLo:
 				typ = rLarchPCALALO12
+			case RelLoong64Branch:
+				typ = rLarchB26
 			}
 			relas = append(relas, elfRela{
 				off:    uint64(fn.Offset + r.Off),
 				typ:    typ,
 				sym:    idx,
-				addend: r.Addend - int64(r.After-r.Off),
+				addend: r.Addend,
 			})
 		}
 	}
 
+	// The data symbols' symbol-valued DATA fields ("DATA s+0(SB)/8,
+	// $other(SB)") become .rela.data entries: an absolute relocation of the
+	// DATA line's width at the field's data-section offset, S + A with no
+	// PC term.  Widths 4 and 8 have ELF relocation shapes; narrower fields
+	// cannot hold an address, so they are refused rather than truncated.
+	var dataRelas []elfRela
+	for _, d := range img.DataSyms {
+		for _, r := range d.Relocs {
+			idx, ok := symIdx[r.Name]
+			if !ok {
+				return nil, fmt.Errorf("data relocation references unknown symbol %q", r.Name)
+			}
+			var typ uint32
+			switch r.Siz {
+			case 8:
+				typ = rLarchAbs64
+			case 4:
+				typ = rLarchAbs32
+			default:
+				return nil, fmt.Errorf("DATA %q: a symbol value of width %d has no ELF relocation", d.Name, r.Siz)
+			}
+			dataRelas = append(dataRelas, elfRela{
+				off:    uint64(d.Offset + r.Off),
+				sym:    idx,
+				typ:    typ,
+				addend: r.Addend,
+			})
+		}
+	}
+
+	// Section presence: .rela.text only when there are code relocations,
+	// .rela.data only when a DATA line holds a symbol value.
+	hasRela := len(relas) > 0
+	hasDataRela := len(dataRelas) > 0
+	nSections := 6
+	if hasRela {
+		nSections++
+	}
+	if hasDataRela {
+		nSections++
+	}
+	secSymtab, secStrtab := 3, 4
+	secShstr := nSections - 1
+
 	// String tables.
 	stNames := newElfStrtab()
 	for _, s := range syms {
@@ -116,14 +174,12 @@ func (img *Image) ELFLOONG64Object() ([]byte, error) {
 	for _, n := range []string{".text", ".data", ".symtab", ".strtab", ".rela.text", ".shstrtab"} {
 		stSections.add(n)
 	}
-
-	hasRela := len(relas) > 0
-	nSections := 6
-	if hasRela {
-		nSections = 7
+	if hasDataRela {
+		stSections.add(".rela.data")
+	}
+	for _, n := range dwarfSectionNames {
+		stSections.add(n)
 	}
-	secSymtab, secStrtab := 3, 4
-	secShstr := nSections - 1
 
 	// Layout.
 	var out []byte
@@ -159,7 +215,7 @@ func (img *Image) ELFLOONG64Object() ([]byte, error) {
 	strtabOff := len(out)
 	out = append(out, stNames.bytes()...)
 
-	var relaOff int
+	var relaOff, relaDataOff int
 	if hasRela {
 		align(8)
 		relaOff = len(out)
@@ -171,10 +227,47 @@ func (img *Image) ELFLOONG64Object() ([]byte, error) {
 			out = append(out, b[:]...)
 		}
 	}
+	if hasDataRela {
+		align(8)
+		relaDataOff = len(out)
+		for _, r := range dataRelas {
+			var b [24]byte
+			le.PutUint64(b[0:], r.off)
+			le.PutUint64(b[8:], uint64(r.sym)<<32|uint64(r.typ))
+			le.PutUint64(b[16:], uint64(r.addend))
+			out = append(out, b[:]...)
+		}
+	}
 
 	shstrOff := len(out)
 	out = append(out, stSections.bytes()...)
 
+	dwAlign := func(n int) {
+		for len(out)%n != 0 {
+			out = append(out, 0)
+		}
+	}
+	dw := appendDWARFSections(&out, img, dwarfSourceName(img), symIdx, dwAlign, cfiLOONG64)
+	dwarfStart := 0 // section index of .debug_abbrev, set when DWARF is present
+	if dw != nil {
+		// Five DWARF sections: .debug_abbrev, .debug_info, .debug_line,
+		// .debug_line_str and .debug_frame (the CIE is unconditional, so
+		// the frame section is always present), plus the relocation
+		// sections below when they carry entries.
+		dwarfStart = nSections
+		nSections += 5
+		appendDWARFRelas(&out, dw, rLarchAbs64, dwAlign)
+		if dw.infoRelaCount > 0 {
+			nSections++
+		}
+		if dw.lineRelaCount > 0 {
+			nSections++
+		}
+		if dw.frameRelaCount > 0 {
+			nSections++
+		}
+	}
+
 	align(8)
 	shoff := len(out)
 
@@ -200,7 +293,43 @@ func (img *Image) ELFLOONG64Object() ([]byte, error) {
 	if hasRela {
 		putSh(".rela.text", shtRela, 0, relaOff, 24*len(relas), secSymtab, secText, 8, 24)
 	}
+	if hasDataRela {
+		putSh(".rela.data", shtRela, 0, relaDataOff, 24*len(dataRelas), secSymtab, secData, 8, 24)
+	}
 	putSh(".shstrtab", shtStrtab, 0, shstrOff, len(stSections.bytes()), 0, 0, 1, 0)
+	// DWARF section headers; their indices follow the write order.
+	if dw != nil {
+		// secIdx is a running section index: each putSh below emits the
+		// next header, and the sh_info of a .rela section names the index
+		// of the section it relocates.
+		secIdx := dwarfStart
+		putSh(".debug_abbrev", shtProgbits, 0, dw.abbrevOff, dw.abbrevSize, 0, 0, 1, 0)
+		secIdx++
+		putSh(".debug_info", shtProgbits, 0, dw.infoOff, dw.infoSize, 0, 0, 1, 0)
+		secInfoIdx := secIdx
+		secIdx++
+		if dw.infoRelaCount > 0 {
+			putSh(".rela.debug_info", shtRela, 0, dw.infoRelaOff, 24*dw.infoRelaCount, secSymtab, secInfoIdx, 8, 24)
+			secIdx++
+		}
+		putSh(".debug_line", shtProgbits, 0, dw.lineOff, dw.lineSize, 0, 0, 1, 0)
+		secLineIdx := secIdx
+		secIdx++
+		if dw.lineRelaCount > 0 {
+			putSh(".rela.debug_line", shtRela, 0, dw.lineRelaOff, 24*dw.lineRelaCount, secSymtab, secLineIdx, 8, 24)
+			secIdx++
+		}
+		putSh(".debug_line_str", shtProgbits, 0, dw.lineStrOff, dw.lineStrSize, 0, 0, 1, 0)
+		secIdx++
+		if dw.frameSize > 0 {
+			putSh(".debug_frame", shtProgbits, 0, dw.frameOff, dw.frameSize, 0, 0, 8, 0)
+			secFrameIdx := secIdx
+			secIdx++
+			if dw.frameRelaCount > 0 {
+				putSh(".rela.debug_frame", shtRela, 0, dw.frameRelaOff, 24*dw.frameRelaCount, secSymtab, secFrameIdx, 8, 24)
+			}
+		}
+	}
 
 	// ELF header.
 	hdr := out[:64]
@@ -211,7 +340,7 @@ func (img *Image) ELFLOONG64Object() ([]byte, error) {
 	le.PutUint64(hdr[24:], 0)
 	le.PutUint64(hdr[32:], 0)
 	le.PutUint64(hdr[40:], uint64(shoff))
-	le.PutUint32(hdr[48:], 0)
+	le.PutUint32(hdr[48:], efLarchAbiDoubleObjV1)
 	le.PutUint16(hdr[52:], 64)
 	le.PutUint16(hdr[54:], 0)
 	le.PutUint16(hdr[56:], 0)

asm/elfloong64_test.go

@@ -9,7 +9,7 @@ import (
 	"encoding/binary"
 	"testing"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
 )
 
 // TestELFLOONG64Object checks the structure of the emitted LoongArch ELF64
@@ -55,6 +55,11 @@ DATA answer<>+0(SB)/8, $42
 	if ef.Type != elf.ET_REL || ef.Machine != elf.EM_LOONGARCH {
 		t.Errorf("type/machine = %v/%v, want ET_REL/EM_LOONGARCH", ef.Type, ef.Machine)
 	}
+	// The double-float ABI plus OBJABI_V1 flags the Go toolchain writes;
+	// system linkers refuse ABI-mismatched merges.
+	if flags := binary.LittleEndian.Uint32(obj[48:]); flags != efLarchAbiDoubleObjV1 {
+		t.Errorf("e_flags = %#x, want %#x (double-float, OBJABI_V1)", flags, efLarchAbiDoubleObjV1)
+	}
 
 	text := ef.Section(".text")
 	data := ef.Section(".data")
@@ -139,7 +144,7 @@ DATA answer<>+0(SB)/8, $42
 		t.Fatalf(".rela.text has %d bytes, want two 24-byte entries", len(raw))
 	}
 	le := binary.LittleEndian
-	for i := 0; i < 2; i++ {
+	for i := range 2 {
 		e := raw[i*24 : (i+1)*24]
 		off := le.Uint64(e[0:])
 		info := le.Uint64(e[8:])
@@ -198,3 +203,160 @@ TEXT ·nop(SB), NOSPLIT, $0
 		t.Error("function symbol nop not found")
 	}
 }
+
+// TestELFLOONG64BranchRelocation checks that the morestack call and an
+// internal CALL both carry R_LARCH_B26 in the emitted object, matching the
+// Go linker's mapping of its call relocation.
+func TestELFLOONG64BranchRelocation(t *testing.T) {
+	f, errs := parser.Parse("k_loong64.s", "TEXT \u00b7callbig(SB), $8192-0\n\tCALL \u00b7other(SB)\n\tRET\nTEXT \u00b7other(SB), NOSPLIT, $0\n\tRET\n")
+	if len(errs) > 0 {
+		t.Fatalf("parse: %v", errs)
+	}
+	img, err := AssembleFileLOONG64(f)
+	if err != nil {
+		t.Fatalf("AssembleFileLOONG64: %v", err)
+	}
+	obj, err := img.ELFLOONG64Object()
+	if err != nil {
+		t.Fatalf("ELFLOONG64Object: %v", err)
+	}
+	ef, err := elf.NewFile(bytes.NewReader(obj))
+	if err != nil {
+		t.Fatalf("parse emitted object: %v", err)
+	}
+	defer ef.Close()
+	relaSec := ef.Section(".rela.text")
+	if relaSec == nil {
+		t.Fatal("missing .rela.text")
+	}
+	raw, err := relaSec.Data()
+	if err != nil {
+		t.Fatal(err)
+	}
+	// The guard's morestack call plus the body's CALL to other.
+	if len(raw)%24 != 0 || len(raw)/24 != 2 {
+		t.Fatalf(".rela.text has %d bytes, want two 24-byte entries", len(raw))
+	}
+	le := binary.LittleEndian
+	for i := range 2 {
+		info := le.Uint64(raw[i*24+8:])
+		if elf.R_LARCH(info&0xffffffff) != elf.R_LARCH_B26 {
+			t.Errorf("relocation %d type = %v, want R_LARCH_B26", i, elf.R_LARCH(info&0xffffffff))
+		}
+	}
+}
+
+// TestELFLOONG64ObjectDataRelocation checks that a symbol-valued DATA field
+// ("DATA s+0(SB)/8, $other(SB)") reaches the LoongArch ELF object as a
+// .rela.data entry: an R_LARCH_64 (R_LARCH_32 for a width-4 field) at the
+// field's offset within .data, against the named symbol, external targets
+// included.
+func TestELFLOONG64ObjectDataRelocation(t *testing.T) {
+	f, errs := parser.Parse("t_loong64.s", `#include "textflag.h"
+TEXT ·Keep(SB), NOSPLIT, $0-0
+	RET
+GLOBL holder(SB), NOPTR, $32
+DATA holder+0(SB)/8, $·Keep+5(SB)
+DATA holder+8(SB)/8, $holder(SB)
+DATA holder+16(SB)/8, $extvar(SB)
+DATA holder+24(SB)/4, $Keep(SB)
+`)
+	if len(errs) > 0 {
+		t.Fatalf("parse: %v", errs)
+	}
+	img, err := AssembleFileLOONG64(f)
+	if err != nil {
+		t.Fatalf("AssembleFileLOONG64: %v", err)
+	}
+	obj, err := img.ELFLOONG64Object()
+	if err != nil {
+		t.Fatalf("ELFLOONG64Object: %v", err)
+	}
+	checkELFSectionAccounting(t, obj)
+	ef, err := elf.NewFile(bytes.NewReader(obj))
+	if err != nil {
+		t.Fatalf("parse emitted object: %v", err)
+	}
+	defer ef.Close()
+	relaData := ef.Section(".rela.data")
+	if relaData == nil {
+		t.Fatal("missing .rela.data section")
+	}
+	if relaData.Type != elf.SHT_RELA {
+		t.Errorf(".rela.data type = %v, want SHT_RELA", relaData.Type)
+	}
+	if relaData.Link == 0 || ef.Sections[relaData.Link].Name != ".symtab" {
+		t.Errorf(".rela.data sh_link = %d, want the .symtab index", relaData.Link)
+	}
+	if ef.Sections[relaData.Info].Name != ".data" {
+		t.Errorf(".rela.data sh_info = %d, want the .data index", relaData.Info)
+	}
+	relas, err := relaData.Data()
+	if err != nil {
+		t.Fatal(err)
+	}
+	var got []struct {
+		off    uint64
+		sym    uint32
+		typ    uint32
+		addend int64
+	}
+	for i := 0; i+24 <= len(relas); i += 24 {
+		got = append(got, struct {
+			off    uint64
+			sym    uint32
+			typ    uint32
+			addend int64
+		}{
+			off: binary.LittleEndian.Uint64(relas[i:]),
+			// r_info packs the type in the low dword and the symbol index
+			// in the high dword.
+			typ:    binary.LittleEndian.Uint32(relas[i+8:]),
+			sym:    binary.LittleEndian.Uint32(relas[i+12:]),
+			addend: int64(binary.LittleEndian.Uint64(relas[i+16:])),
+		})
+	}
+	// debug/elf hides the table's null entry, so raw index s names syms[s-1].
+	syms, err := ef.Symbols()
+	if err != nil {
+		t.Fatal(err)
+	}
+	name := func(idx uint32) string {
+		if idx >= 1 && int(idx) <= len(syms) {
+			return syms[idx-1].Name
+		}
+		return ""
+	}
+	// The offsets are data-section-relative: the field's DATA offset plus
+	// the symbol's position in .data (the layout aligns each symbol to 16).
+	base := uint64(0)
+	for _, d := range img.DataSyms {
+		if d.Name == "holder" {
+			base = uint64(d.Offset)
+		}
+	}
+	want := []struct {
+		off    uint64
+		typ    uint32
+		addend int64
+		target string
+	}{
+		{off: base + 0, typ: uint32(elf.R_LARCH_64), addend: 5, target: "Keep"},
+		{off: base + 8, typ: uint32(elf.R_LARCH_64), addend: 0, target: "holder"},
+		{off: base + 16, typ: uint32(elf.R_LARCH_64), addend: 0, target: "extvar"},
+		{off: base + 24, typ: uint32(elf.R_LARCH_32), addend: 0, target: "Keep"},
+	}
+	if len(got) != len(want) {
+		t.Fatalf(".rela.data entries = %d, want %d", len(got), len(want))
+	}
+	for i, w := range want {
+		g := got[i]
+		if g.off != w.off || g.typ != w.typ || g.addend != w.addend {
+			t.Errorf("entry %d = {off %d typ %d addend %d}, want {off %d typ %d addend %d}",
+				i, g.off, g.typ, g.addend, w.off, w.typ, w.addend)
+		}
+		if n := name(g.sym); n != w.target {
+			t.Errorf("entry %d names %q, want %q", i, n, w.target)
+		}
+	}
+}

asm/elfriscv.go

@@ -13,17 +13,27 @@ import (
 const (
 	emRISCV = 243 // EM_RISCV
 
+	// EF_RISCV_FLOAT_ABI_DOUBLE: the double-precision float ABI the Go
+	// toolchain targets (cmd/link/internal/ld/elf.go writes Flags = 0x4 for
+	// RISCV64).  System linkers refuse to merge ET_REL objects whose float
+	// ABI differs, so 0 (soft-float) would make the object unlinkable.
+	efRISCVFloatAbiDouble = 0x4
+
 	// RISC-V relocation types.
-	rRISCV32         = 1
 	rRISCVJAL        = 17 // R_RISCV_JAL
 	rRISCVPCRELHI20  = 23 // R_RISCV_PCREL_HI20
 	rRISCVPCRELLO12I = 24 // R_RISCV_PCREL_LO12_I
 	rRISCVPCRELLO12S = 25 // R_RISCV_PCREL_LO12_S
+	// R_RISCV_32 (debug/elf 1): the absolute 32-bit address of a symbol,
+	// the R_ADDR shape a 4-byte DATA field carries.  R_RISCV_64 (2) lives
+	// with the DWARF fixup constants as rRISCVAbs64.
+	rRISVCAbs32 = 1
 )
 
 // ELFRISCVObject returns the image as an ELF64 relocatable object file for
 // RISC-V (EM_RISCV, 64-bit, little-endian).  The structure mirrors the amd64
-// ELF emission: .text, .data, .symtab, .strtab and optional .rela.text.
+// ELF emission: .text, .data, .symtab, .strtab, an optional .rela.text and
+// an optional .rela.data.
 func (img *Image) ELFRISCVObject() ([]byte, error) {
 	le := binary.LittleEndian
 
@@ -83,9 +93,14 @@ func (img *Image) ELFRISCVObject() ([]byte, error) {
 	// Build relocations.  Each SB reference is an AUIPC + second-instruction
 	// pair carrying a single relocation kind; the ELF writer expands it into
 	// the R_RISCV_PCREL_HI20 + R_RISCV_PCREL_LO12_I/S pair the psABI expects.
-	// The HI20 carries the symbol addend; the LO12 addend is zero, matching
-	// cmd/link's own ELF conversion (the LO12 resolves against the HI20's
-	// AUIPC location).
+	// The HI20 carries the symbol and its addend.  The LO12's symbol must
+	// denote the AUIPC site the HI20 relocates (psABI §8.4.9: the pair is
+	// resolved against the label of the AUIPC, not the target symbol;
+	// cmd/link generates one local text symbol per AUIPC for exactly this,
+	// cmd/link/internal/riscv64/asm.go).  The .text section symbol with the
+	// AUIPC's section-relative offset as addend gives S + A = the AUIPC
+	// address, which is that label.
+	const secSymText = 1 // syms[1], the .text section symbol
 	type elfRela struct {
 		off    uint64
 		typ    uint32
@@ -99,27 +114,70 @@ func (img *Image) ELFRISCVObject() ([]byte, error) {
 			if !ok {
 				return nil, fmt.Errorf("relocation references unknown symbol %q", r.Name)
 			}
+			auipc := int64(fn.Offset + r.Off)
 			switch r.Kind {
 			case RelRISCVPCRELIType:
 				relas = append(relas,
 					elfRela{off: uint64(fn.Offset + r.Off), typ: rRISCVPCRELHI20, sym: idx, addend: r.Addend},
-					elfRela{off: uint64(fn.Offset + r.Off + 4), typ: rRISCVPCRELLO12I, sym: idx, addend: 0},
+					elfRela{off: uint64(fn.Offset + r.Off + 4), typ: rRISCVPCRELLO12I, sym: secSymText, addend: auipc},
 				)
 			case RelRISCVPCRELSType:
 				relas = append(relas,
 					elfRela{off: uint64(fn.Offset + r.Off), typ: rRISCVPCRELHI20, sym: idx, addend: r.Addend},
-					elfRela{off: uint64(fn.Offset + r.Off + 4), typ: rRISCVPCRELLO12S, sym: idx, addend: 0},
+					elfRela{off: uint64(fn.Offset + r.Off + 4), typ: rRISCVPCRELLO12S, sym: secSymText, addend: auipc},
 				)
 			case RelRISCVJal:
 				relas = append(relas, elfRela{off: uint64(fn.Offset + r.Off), typ: rRISCVJAL, sym: idx, addend: r.Addend})
-			case RelPCRelAbs:
-				relas = append(relas, elfRela{off: uint64(fn.Offset + r.Off), typ: rRISCV32, sym: idx, addend: r.Addend})
 			default:
 				return nil, fmt.Errorf("relocation kind %v unsupported in ELF emission", r.Kind)
 			}
 		}
 	}
 
+	// The data symbols' symbol-valued DATA fields ("DATA s+0(SB)/8,
+	// $other(SB)") become .rela.data entries: an absolute relocation of the
+	// DATA line's width at the field's data-section offset, S + A with no
+	// PC term.  Widths 4 and 8 have ELF relocation shapes; narrower fields
+	// cannot hold an address, so they are refused rather than truncated.
+	var dataRelas []elfRela
+	for _, d := range img.DataSyms {
+		for _, r := range d.Relocs {
+			idx, ok := symIdx[r.Name]
+			if !ok {
+				return nil, fmt.Errorf("data relocation references unknown symbol %q", r.Name)
+			}
+			var typ uint32
+			switch r.Siz {
+			case 8:
+				typ = rRISCVAbs64
+			case 4:
+				typ = rRISVCAbs32
+			default:
+				return nil, fmt.Errorf("DATA %q: a symbol value of width %d has no ELF relocation", d.Name, r.Siz)
+			}
+			dataRelas = append(dataRelas, elfRela{
+				off:    uint64(d.Offset + r.Off),
+				sym:    idx,
+				typ:    typ,
+				addend: r.Addend,
+			})
+		}
+	}
+
+	// Section presence: .rela.text only when there are code relocations,
+	// .rela.data only when a DATA line holds a symbol value.
+	hasRela := len(relas) > 0
+	hasDataRela := len(dataRelas) > 0
+	nSections := 6
+	if hasRela {
+		nSections++
+	}
+	if hasDataRela {
+		nSections++
+	}
+	secSymtab, secStrtab := 3, 4
+	secShstr := nSections - 1
+
 	// String tables.
 	stNames := newElfStrtab()
 	for _, s := range syms {
@@ -129,14 +187,12 @@ func (img *Image) ELFRISCVObject() ([]byte, error) {
 	for _, n := range []string{".text", ".data", ".symtab", ".strtab", ".rela.text", ".shstrtab"} {
 		stSections.add(n)
 	}
-
-	hasRela := len(relas) > 0
-	nSections := 6
-	if hasRela {
-		nSections = 7
+	if hasDataRela {
+		stSections.add(".rela.data")
+	}
+	for _, n := range dwarfSectionNames {
+		stSections.add(n)
 	}
-	secSymtab, secStrtab := 3, 4
-	secShstr := nSections - 1
 
 	// Layout.
 	var out []byte
@@ -172,7 +228,7 @@ func (img *Image) ELFRISCVObject() ([]byte, error) {
 	strtabOff := len(out)
 	out = append(out, stNames.bytes()...)
 
-	var relaOff int
+	var relaOff, relaDataOff int
 	if hasRela {
 		align(8)
 		relaOff = len(out)
@@ -184,10 +240,47 @@ func (img *Image) ELFRISCVObject() ([]byte, error) {
 			out = append(out, b[:]...)
 		}
 	}
+	if hasDataRela {
+		align(8)
+		relaDataOff = len(out)
+		for _, r := range dataRelas {
+			var b [24]byte
+			le.PutUint64(b[0:], r.off)
+			le.PutUint64(b[8:], uint64(r.sym)<<32|uint64(r.typ))
+			le.PutUint64(b[16:], uint64(r.addend))
+			out = append(out, b[:]...)
+		}
+	}
 
 	shstrOff := len(out)
 	out = append(out, stSections.bytes()...)
 
+	dwAlign := func(n int) {
+		for len(out)%n != 0 {
+			out = append(out, 0)
+		}
+	}
+	dw := appendDWARFSections(&out, img, dwarfSourceName(img), symIdx, dwAlign, cfiRISCV64)
+	dwarfStart := 0 // section index of .debug_abbrev, set when DWARF is present
+	if dw != nil {
+		// Five DWARF sections: .debug_abbrev, .debug_info, .debug_line,
+		// .debug_line_str and .debug_frame (the CIE is unconditional, so
+		// the frame section is always present), plus the relocation
+		// sections below when they carry entries.
+		dwarfStart = nSections
+		nSections += 5
+		appendDWARFRelas(&out, dw, rRISCVAbs64, dwAlign)
+		if dw.infoRelaCount > 0 {
+			nSections++
+		}
+		if dw.lineRelaCount > 0 {
+			nSections++
+		}
+		if dw.frameRelaCount > 0 {
+			nSections++
+		}
+	}
+
 	align(8)
 	shoff := len(out)
 
@@ -213,7 +306,43 @@ func (img *Image) ELFRISCVObject() ([]byte, error) {
 	if hasRela {
 		putSh(".rela.text", shtRela, 0, relaOff, 24*len(relas), secSymtab, secText, 8, 24)
 	}
+	if hasDataRela {
+		putSh(".rela.data", shtRela, 0, relaDataOff, 24*len(dataRelas), secSymtab, secData, 8, 24)
+	}
 	putSh(".shstrtab", shtStrtab, 0, shstrOff, len(stSections.bytes()), 0, 0, 1, 0)
+	// DWARF section headers; their indices follow the write order.
+	if dw != nil {
+		// secIdx is a running section index: each putSh below emits the
+		// next header, and the sh_info of a .rela section names the index
+		// of the section it relocates.
+		secIdx := dwarfStart
+		putSh(".debug_abbrev", shtProgbits, 0, dw.abbrevOff, dw.abbrevSize, 0, 0, 1, 0)
+		secIdx++
+		putSh(".debug_info", shtProgbits, 0, dw.infoOff, dw.infoSize, 0, 0, 1, 0)
+		secInfoIdx := secIdx
+		secIdx++
+		if dw.infoRelaCount > 0 {
+			putSh(".rela.debug_info", shtRela, 0, dw.infoRelaOff, 24*dw.infoRelaCount, secSymtab, secInfoIdx, 8, 24)
+			secIdx++
+		}
+		putSh(".debug_line", shtProgbits, 0, dw.lineOff, dw.lineSize, 0, 0, 1, 0)
+		secLineIdx := secIdx
+		secIdx++
+		if dw.lineRelaCount > 0 {
+			putSh(".rela.debug_line", shtRela, 0, dw.lineRelaOff, 24*dw.lineRelaCount, secSymtab, secLineIdx, 8, 24)
+			secIdx++
+		}
+		putSh(".debug_line_str", shtProgbits, 0, dw.lineStrOff, dw.lineStrSize, 0, 0, 1, 0)
+		secIdx++
+		if dw.frameSize > 0 {
+			putSh(".debug_frame", shtProgbits, 0, dw.frameOff, dw.frameSize, 0, 0, 8, 0)
+			secFrameIdx := secIdx
+			secIdx++
+			if dw.frameRelaCount > 0 {
+				putSh(".rela.debug_frame", shtRela, 0, dw.frameRelaOff, 24*dw.frameRelaCount, secSymtab, secFrameIdx, 8, 24)
+			}
+		}
+	}
 
 	// ELF header.
 	hdr := out[:64]
@@ -224,7 +353,7 @@ func (img *Image) ELFRISCVObject() ([]byte, error) {
 	le.PutUint64(hdr[24:], 0)
 	le.PutUint64(hdr[32:], 0)
 	le.PutUint64(hdr[40:], uint64(shoff))
-	le.PutUint32(hdr[48:], 0)
+	le.PutUint32(hdr[48:], efRISCVFloatAbiDouble)
 	le.PutUint16(hdr[52:], 64)
 	le.PutUint16(hdr[54:], 0)
 	le.PutUint16(hdr[56:], 0)

asm/elfriscv_test.go

@@ -0,0 +1,128 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package asm
+
+import (
+	"bytes"
+	"debug/elf"
+	"encoding/binary"
+	"testing"
+
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
+)
+
+// TestELFRISCVObjectDataRelocation checks that a symbol-valued DATA field
+// ("DATA s+0(SB)/8, $other(SB)") reaches the RISC-V ELF object as a
+// .rela.data entry: an R_RISCV_64 (R_RISCV_32 for a width-4 field) at the
+// field's offset within .data, against the named symbol, external targets
+// included.
+func TestELFRISCVObjectDataRelocation(t *testing.T) {
+	f, errs := parser.Parse("t_riscv64.s", `#include "textflag.h"
+TEXT ·Keep(SB), NOSPLIT, $0-0
+	RET
+GLOBL holder(SB), NOPTR, $32
+DATA holder+0(SB)/8, $·Keep+5(SB)
+DATA holder+8(SB)/8, $holder(SB)
+DATA holder+16(SB)/8, $extvar(SB)
+DATA holder+24(SB)/4, $Keep(SB)
+`)
+	if len(errs) > 0 {
+		t.Fatalf("parse: %v", errs)
+	}
+	img, err := AssembleFileRISCV(f)
+	if err != nil {
+		t.Fatalf("AssembleFileRISCV: %v", err)
+	}
+	obj, err := img.ELFRISCVObject()
+	if err != nil {
+		t.Fatalf("ELFRISCVObject: %v", err)
+	}
+	checkELFSectionAccounting(t, obj)
+	ef, err := elf.NewFile(bytes.NewReader(obj))
+	if err != nil {
+		t.Fatalf("parse emitted object: %v", err)
+	}
+	defer ef.Close()
+	relaData := ef.Section(".rela.data")
+	if relaData == nil {
+		t.Fatal("missing .rela.data section")
+	}
+	if relaData.Type != elf.SHT_RELA {
+		t.Errorf(".rela.data type = %v, want SHT_RELA", relaData.Type)
+	}
+	if relaData.Link == 0 || ef.Sections[relaData.Link].Name != ".symtab" {
+		t.Errorf(".rela.data sh_link = %d, want the .symtab index", relaData.Link)
+	}
+	if ef.Sections[relaData.Info].Name != ".data" {
+		t.Errorf(".rela.data sh_info = %d, want the .data index", relaData.Info)
+	}
+	relas, err := relaData.Data()
+	if err != nil {
+		t.Fatal(err)
+	}
+	var got []struct {
+		off    uint64
+		sym    uint32
+		typ    uint32
+		addend int64
+	}
+	for i := 0; i+24 <= len(relas); i += 24 {
+		got = append(got, struct {
+			off    uint64
+			sym    uint32
+			typ    uint32
+			addend int64
+		}{
+			off: binary.LittleEndian.Uint64(relas[i:]),
+			// r_info packs the type in the low dword and the symbol index
+			// in the high dword.
+			typ:    binary.LittleEndian.Uint32(relas[i+8:]),
+			sym:    binary.LittleEndian.Uint32(relas[i+12:]),
+			addend: int64(binary.LittleEndian.Uint64(relas[i+16:])),
+		})
+	}
+	// debug/elf hides the table's null entry, so raw index s names syms[s-1].
+	syms, err := ef.Symbols()
+	if err != nil {
+		t.Fatal(err)
+	}
+	name := func(idx uint32) string {
+		if idx >= 1 && int(idx) <= len(syms) {
+			return syms[idx-1].Name
+		}
+		return ""
+	}
+	// The offsets are data-section-relative: the field's DATA offset plus
+	// the symbol's position in .data (the layout aligns each symbol to 16).
+	base := uint64(0)
+	for _, d := range img.DataSyms {
+		if d.Name == "holder" {
+			base = uint64(d.Offset)
+		}
+	}
+	want := []struct {
+		off    uint64
+		typ    uint32
+		addend int64
+		target string
+	}{
+		{off: base + 0, typ: uint32(elf.R_RISCV_64), addend: 5, target: "Keep"},
+		{off: base + 8, typ: uint32(elf.R_RISCV_64), addend: 0, target: "holder"},
+		{off: base + 16, typ: uint32(elf.R_RISCV_64), addend: 0, target: "extvar"},
+		{off: base + 24, typ: uint32(elf.R_RISCV_32), addend: 0, target: "Keep"},
+	}
+	if len(got) != len(want) {
+		t.Fatalf(".rela.data entries = %d, want %d", len(got), len(want))
+	}
+	for i, w := range want {
+		g := got[i]
+		if g.off != w.off || g.typ != w.typ || g.addend != w.addend {
+			t.Errorf("entry %d = {off %d typ %d addend %d}, want {off %d typ %d addend %d}",
+				i, g.off, g.typ, g.addend, w.off, w.typ, w.addend)
+		}
+		if n := name(g.sym); n != w.target {
+			t.Errorf("entry %d names %q, want %q", i, n, w.target)
+		}
+	}
+}

asm/encodable.go

@@ -0,0 +1,128 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package asm
+
+import "strings"
+
+// Encodable reports whether the amd64 encoder knows how to encode the
+// mnemonic.  It mirrors the dispatch in (*enc).encode: the fixed-name
+// instructions, conditional jumps, the CMOV/SET condition families, the
+// VEX/EVEX/opmask/gather/scatter vector paths, the legacy SSE tables and the
+// explicit scalar cases.  A mnemonic that parses (is in the architecture
+// table) but is not encodable would otherwise surface only at assembly time,
+// deep inside a build; the linter uses this predicate to flag it at edit
+// time.
+func Encodable(mnemonic string) bool {
+	upper := strings.ToUpper(mnemonic)
+
+	// Fixed-name instructions (no size suffix).
+	switch upper {
+	case "RET", "NOP", "CALL", "JMP",
+		"POPFQ", "PUSHFQ", "INT", "LDMXCSR", "STMXCSR", "CMPSD", "SHA256RNDS2",
+		// The literal-data pseudo-ops, the accepted-and-ignored END and
+		// bookkeeping statements, and the SP adjust.
+		"BYTE", "WORD", "LONG", "QUAD", "END", "ADJSP", "FUNCDATA", "PCDATA":
+		return true
+	}
+	if _, ok := noOperandTable[upper]; ok {
+		return true
+	}
+	if _, ok := condCode(upper); ok {
+		return true
+	}
+
+	// VEX/EVEX and friends: the trailing B/W/L/Q/D is part of the mnemonic.
+	base, _, err := parseEvexSuffix(upper)
+	if err != nil {
+		return false
+	}
+	if isVex(base) || isEvex(base) || isKOp(base) || isGather(base) || isScatter(base) ||
+		base == "KMOVW" || base == "KMOVQ" || base == "KMOVB" || base == "KMOVD" {
+		return true
+	}
+
+	// CMOV carries size then condition (CMOVLGT); SET carries the condition
+	// alone (SETNE).  The size letter is checked exactly as encodeCmov does,
+	// so a spelling like CMOVBGT is not reported encodable when Encode
+	// would reject it.
+	if rest, ok := strings.CutPrefix(upper, "CMOV"); ok && len(rest) >= 2 {
+		switch rest[0] {
+		case 'W', 'L', 'Q':
+			if _, ok := jccMap[rest[1:]]; ok {
+				return true
+			}
+		}
+	}
+	if rest, ok := strings.CutPrefix(upper, "SET"); ok {
+		if _, ok := jccMap[rest]; ok {
+			return true
+		}
+	}
+
+	// Legacy SSE shuffles and packed binaries dispatch on the full name; so
+	// do the imm8-controlled instructions, the lane extracts and inserts and
+	// the packed integer shifts (their trailing width letters belong to the
+	// mnemonic).
+	if _, ok := sseShufTable[upper]; ok {
+		return true
+	}
+	if _, ok := sseBinTable[upper]; ok {
+		return true
+	}
+	if _, ok := sseImm3Table[upper]; ok {
+		return true
+	}
+	if _, ok := sseExtractTable[upper]; ok {
+		return true
+	}
+	if _, ok := sseInsertTable[upper]; ok {
+		return true
+	}
+	if _, ok := sseShiftImm[upper]; ok {
+		return true
+	}
+
+	// The size-suffix split: retry the tables and the scalar switch on the
+	// base.
+	base2, size := splitSize(upper)
+	if size == 0 {
+		size = 8
+	}
+	_ = size
+	if base2 != upper {
+		if _, ok := sseBinTable[base2]; ok {
+			return true
+		}
+	}
+	switch base2 {
+	case "MOV", "MOVD",
+		"ADD", "SUB", "AND", "OR", "XOR", "CMP", "ADC", "SBB",
+		"TEST",
+		"LEA",
+		"INC", "DEC", "NEG", "NOT", "MUL", "DIV", "IDIV",
+		"SHL", "SHR", "SAR", "SAL", "ROL", "ROR", "RCL", "RCR",
+		"BT", "BTS", "BTR", "BTC",
+		"XCHG", "CMPXCHG", "XADD", "CRC32", "ADCX", "ADOX",
+		"MOVS", "STOS",
+		"IMUL", "IMUL3",
+		"PUSH", "POP",
+		"BSF", "BSR", "LZCNT", "TZCNT", "POPCNT",
+		"BSWAP",
+		"PREFETCHNTA", "PREFETCHT0", "PREFETCHT1", "PREFETCHT2",
+		"MOVBLZX", "MOVBQZX", "MOVWLZX", "MOVWQZX", "MOVWLSX", "MOVLQSX",
+		"MOVBWZX", "MOVBWSX", "MOVBLSX", "MOVBQSX", "MOVWQSX", "MOVLQZX",
+		"CVTSL2SD", "CVTSQ2SD",
+		"CVTSD2S", "CVTTSD2S", "CVTSS2S", "CVTTSS2S",
+		"FMOVD",
+		"MOVOU", "MOVO", "MOVOA", "MOVUPS", "MOVAPS", "MOVUPD", "MOVAPD", "MOVSD", "MOVSS":
+		return true
+	}
+	// Full-name dispatches the size split would eat (a trailing width
+	// letter that is part of the mnemonic).
+	switch upper {
+	case "PMOVMSKB":
+		return true
+	}
+	return false
+}

asm/encode.go

@@ -5,6 +5,8 @@ package asm
 
 import (
 	"fmt"
+	"math"
+	"strconv"
 	"strings"
 )
 
@@ -21,6 +23,39 @@ func Encode(mnemonic string, ops ...Operand) ([]byte, error) {
 type enc struct {
 	out     []byte
 	patches []encPatch // disp32 fields awaiting static-symbol resolution
+
+	// FloatPool collects the pooled constants the floating-point
+	// immediates reference, in first-use order.
+	floatPool     []floatPoolEntry
+	floatPoolSeen map[string]bool
+}
+
+// floatPoolEntry is one pooled floating-point constant: the symbol name
+// the emitted RIP-relative load refers to and its IEEE-754 bytes.
+type floatPoolEntry struct {
+	name string
+	data []byte
+}
+
+// addFloatPool records a pooled constant, deduplicated by symbol name.
+func (e *enc) addFloatPool(name string, bits uint64, width int) {
+	if e.floatPoolSeen == nil {
+		e.floatPoolSeen = map[string]bool{}
+	}
+	if e.floatPoolSeen[name] {
+		return
+	}
+	e.floatPoolSeen[name] = true
+	data := make([]byte, width)
+	for i := range width {
+		data[i] = byte(bits >> (8 * i))
+	}
+	e.floatPool = append(e.floatPool, floatPoolEntry{name: name, data: data})
+}
+
+// floatPoolList returns the pooled constants in first-use order.
+func (e *enc) floatPoolList() []floatPoolEntry {
+	return e.floatPool
 }
 
 // encPatch marks a 4-byte displacement field in enc.out that must receive the
@@ -29,6 +64,7 @@ type encPatch struct {
 	off    int
 	name   string
 	addend int64
+	tls    bool // a TLS slot offset: the patch is R_TLSLE with no symbol
 }
 
 func (e *enc) encode(mnem string, ops []Operand) error {
@@ -40,14 +76,74 @@ func (e *enc) encode(mnem string, ops []Operand) error {
 		return e.encodeRet()
 	case upper == "NOP":
 		return e.emit(&instr{opcode: []byte{0x90}, modrm: -1, sib: -1})
-	case upper == "CALL":
-		return e.encodeJmpRel(ops, []byte{0xE8})
-	case upper == "JMP":
-		return e.encodeJmpRel(ops, []byte{0xE9})
+	case upper == "CALL" || upper == "JMP":
+		// Through a register or memory: FF /2 (CALL) or FF /4 (JMP).
+		// Anything else is a rel32 against a label resolved by the assembler.
+		if len(ops) == 1 {
+			switch ops[0].(type) {
+			case Reg, Mem:
+				return e.encodeIndirectBranch(upper, ops)
+			}
+		}
+		opcode := []byte{0xE8}
+		if upper == "JMP" {
+			opcode = []byte{0xE9}
+		}
+		return e.encodeJmpRel(ops, opcode)
 	}
 	if cc, ok := condCode(upper); ok {
 		return e.encodeJcc(cc, ops)
 	}
+	// No-operand system and string-control instructions (CPUID, RDTSC,
+	// SYSCALL, the fences, UNDEF, …).
+	if op, ok := noOperandTable[upper]; ok {
+		if len(ops) != 0 {
+			return fmt.Errorf("%s takes no operands, got %d", upper, len(ops))
+		}
+		return e.emit(&instr{opcode: op, modrm: -1, sib: -1})
+	}
+	// POPFQ/PUSHFQ are exact names: the bare POPF/PUSHF and the L spellings
+	// are rejected by go tool asm in 64-bit mode, so they stay unsupported.
+	switch upper {
+	case "POPFQ":
+		if len(ops) != 0 {
+			return fmt.Errorf("POPFQ takes no operands, got %d", len(ops))
+		}
+		return e.emit(&instr{opcode: []byte{0x9D}, modrm: -1, sib: -1})
+	case "PUSHFQ":
+		if len(ops) != 0 {
+			return fmt.Errorf("PUSHFQ takes no operands, got %d", len(ops))
+		}
+		return e.emit(&instr{opcode: []byte{0x9C}, modrm: -1, sib: -1})
+	case "INT":
+		return e.encodeInt(ops)
+	case "LDMXCSR":
+		return e.encodeMxcsr(2, ops)
+	case "STMXCSR":
+		return e.encodeMxcsr(3, ops)
+	// CMPSD is the scalar double compare, whose predicate immediate comes
+	// LAST in Plan 9 order (src, dst, $imm).
+	case "CMPSD":
+		return e.encodeCmpsd(ops)
+	// SHA256RNDS2 carries the round constant in a literal X0 first operand.
+	case "SHA256RNDS2":
+		return e.encodeSha256rnds2(ops)
+	// BYTE, WORD, LONG and QUAD write the immediate into the text stream
+	// itself: 1, 2, 4 or 8 literal bytes, little-endian.  END is accepted
+	// and ignored.  ADJSP adjusts SP by the immediate, sign-chosen between
+	// the SUBQ and ADDQ forms.
+	case "BYTE", "WORD", "LONG", "QUAD":
+		return e.encodeData(upper, ops)
+	case "END":
+		return e.encodeEnd(ops)
+	case "ADJSP":
+		return e.encodeAdjsp(ops)
+	// The runtime's bookkeeping statements carry no text bytes: go tool asm
+	// records FUNCDATA and PCDATA in the program list only, so the encoded
+	// body shows nothing, on every architecture.
+	case "FUNCDATA", "PCDATA":
+		return e.encodeFuncdata(upper, ops)
+	}
 
 	// VEX (AVX/AVX2) and EVEX (AVX-512) instructions: the trailing
 	// B/W/L/Q/D is part of the mnemonic, not a size suffix, so dispatch
@@ -57,7 +153,9 @@ func (e *enc) encode(mnem string, ops []Operand) error {
 	if err != nil {
 		return err
 	}
-	if isVex(base) || isEvex(base) || isKOp(base) || isGather(base) || isScatter(base) || base == "KMOVW" || base == "KMOVQ" {
+	if isVex(base) || isEvex(base) || isKOp(base) || isGather(base) || isScatter(base) ||
+		isEvexPrefGather(base) ||
+		base == "KMOVW" || base == "KMOVQ" || base == "KMOVB" || base == "KMOVD" {
 		return e.encodeVec(base, ops, sfx)
 	}
 	if sfx.any() {
@@ -76,37 +174,345 @@ func (e *enc) encode(mnem string, ops []Operand) error {
 	if size == 0 {
 		size = 8 // default operand size in 64-bit mode (e.g. PUSHQ)
 	}
+	// Legacy SSE imm8 shuffles whose names end in W/H (PSHUFLW,
+	// PSHUFHW) must dispatch BEFORE the size-suffix split, and the
+	// others ride along.
+	if m, ok := sseShufTable[upper]; ok {
+		return e.encodeSSEShuf(m, ops)
+	}
+	// Legacy SSE packed binaries dispatch on the full name: the packed
+	// integer mnemonics carry real width suffixes (PADDB/PCMPGTW/...),
+	// which the size split must not eat.  A floating-point immediate
+	// rewrites into a pooled-constant read on the scalar members.
+	if m, ok := sseBinTable[upper]; ok {
+		if f, isFloat := floatImmOperand(ops); isFloat {
+			return e.encodeSSEFloatBin(upper, m, f, ops)
+		}
+		return e.encodeSSEBin(m, ops)
+	}
+	if m, ok := sseBinTable[base]; ok {
+		if f, isFloat := floatImmOperand(ops); isFloat {
+			return e.encodeSSEFloatBin(upper, m, f, ops)
+		}
+		return e.encodeSSEBin(m, ops)
+	}
+	// The imm8-controlled legacy instructions, the lane extracts and inserts
+	// and the packed integer shifts all dispatch on the full name: a trailing
+	// width letter here belongs to the mnemonic, not to the size split.
+	if m, ok := sseImm3Table[upper]; ok {
+		return e.encodeSSEImm3(m, ops)
+	}
+	if m, ok := sseExtractTable[upper]; ok {
+		return e.encodeSSEExtract(m, ops)
+	}
+	if m, ok := sseInsertTable[upper]; ok {
+		return e.encodeSSEInsert(m, ops)
+	}
+	if _, ok := sseShiftImm[upper]; ok {
+		return e.encodeSSEShift(upper, ops)
+	}
+	// PMOVMSKB ends in a width letter the size split would eat, so it
+	// dispatches on the full name like the packed binaries above.
+	if upper == "PMOVMSKB" {
+		return e.encodePmovmskb(upper, ops)
+	}
 	switch base {
 	case "MOV":
 		return e.encodeMov(ops, size)
-	case "ADD", "SUB", "AND", "OR", "XOR", "CMP":
+	// MOVD is the Go assembler's alias of MOVQ: the same byte forms, 64-bit
+	// REX.W and all.
+	case "MOVD":
+		return e.encodeMov(ops, 8)
+	case "ADD", "SUB", "AND", "OR", "XOR", "CMP", "ADC", "SBB":
 		return e.encodeALU(aluOp[base], ops, size)
 	case "TEST":
 		return e.encodeTest(ops, size)
 	case "LEA":
 		return e.encodeLea(ops, size)
-	case "INC", "DEC", "NEG", "NOT":
+	case "INC", "DEC", "NEG", "NOT", "MUL", "DIV", "IDIV":
 		return e.encodeUnary(unaryOp[base], ops, size)
-	case "SHL", "SHR", "SAR":
-		return e.encodeShift(shiftOp[base], ops, size)
+	case "SHL", "SHR", "SAR", "SAL", "ROL", "ROR", "RCL", "RCR":
+		return e.encodeShift(base, ops, size)
+	case "BT", "BTS", "BTR", "BTC":
+		return e.encodeBitTest(base, ops, size)
+	case "XCHG":
+		return e.encodeExchange(ops, size)
+	case "CMPXCHG":
+		return e.encodeRegRegOp(0xB0, 0xB1, base, ops, size)
+	case "XADD":
+		return e.encodeRegRegOp(0xC0, 0xC1, base, ops, size)
+	case "CRC32":
+		return e.encodeCrc32(ops, size)
+	case "ADCX":
+		return e.encodeCarryExt(0x66, ops, size)
+	case "ADOX":
+		return e.encodeCarryExt(0xF3, ops, size)
+	case "MOVS", "STOS":
+		return e.encodeStringOp(base, ops, size)
 	case "IMUL", "IMUL3":
 		return e.encodeImul(ops, size)
 	case "PUSH":
-		return e.encodePushPop(ops, true)
+		return e.encodePushPop(ops, size, true)
 	case "POP":
-		return e.encodePushPop(ops, false)
-	case "LZCNT", "TZCNT":
+		return e.encodePushPop(ops, size, false)
+	case "BSF", "BSR", "LZCNT", "TZCNT", "POPCNT":
 		return e.encodeCount(base, ops, size)
-	case "MOVBLZX", "MOVBQZX", "MOVWLZX", "MOVWQZX", "MOVWLSX", "MOVLQSX":
+	case "BSWAP":
+		return e.encodeBswap(ops, size)
+	case "PREFETCHNTA", "PREFETCHT0", "PREFETCHT1", "PREFETCHT2":
+		return e.encodePrefetch(base, ops)
+	case "MOVBLZX", "MOVBQZX", "MOVWLZX", "MOVWQZX", "MOVWLSX", "MOVLQSX",
+		"MOVBWZX", "MOVBWSX", "MOVBLSX", "MOVBQSX", "MOVWQSX", "MOVLQZX":
 		return e.encodeMovExtend(base, ops)
 	case "CVTSL2SD", "CVTSQ2SD":
 		return e.encodeCvtsi2sd(base == "CVTSQ2SD", ops)
-	case "MOVOU", "MOVO", "MOVUPS", "MOVAPS", "MOVUPD", "MOVAPD", "MOVSD", "MOVSS":
+	case "CVTSD2S", "CVTTSD2S", "CVTSS2S", "CVTTSS2S":
+		return e.encodeCvtInt(base, ops, size)
+	case "FMOVD":
+		return e.encodeFmov(ops)
+	case "MOVSD", "MOVSS":
+		if f, isFloat := floatImmOperand(ops); isFloat {
+			return e.encodeSSEFloatMove(upper, f, ops)
+		}
+		return e.encodeSSEMove(sseMoveTable[base], ops)
+	case "MOVOU", "MOVO", "MOVOA", "MOVUPS", "MOVAPS", "MOVUPD", "MOVAPD":
 		return e.encodeSSEMove(sseMoveTable[base], ops)
 	}
 	return fmt.Errorf("unsupported instruction %q", mnem)
 }
 
+// encodePrefetch emits the 0F 18 /r prefetch hints: the reg field selects
+// the locality (NTA=0, T0=1, T1=2, T2=3) and the single operand is memory.
+func (e *enc) encodePrefetch(base string, ops []Operand) error {
+	if len(ops) != 1 {
+		return fmt.Errorf("%s expects one memory operand", base)
+	}
+	m, ok := ops[0].(Mem)
+	if !ok {
+		return fmt.Errorf("%s requires a memory operand", base)
+	}
+	i := newInstr(0, []byte{0x0F, 0x18})
+	if err := setMem(i, prefetchVariant[base], m); err != nil {
+		return err
+	}
+	return e.emit(i)
+}
+
+var prefetchVariant = map[string]int{
+	"PREFETCHNTA": 0,
+	"PREFETCHT0":  1,
+	"PREFETCHT1":  2,
+	"PREFETCHT2":  3,
+}
+
+// dataWidth is the literal byte count of each data-emission pseudo-op.
+var dataWidth = map[string]int{
+	"BYTE": 1,
+	"WORD": 2,
+	"LONG": 4,
+	"QUAD": 8,
+}
+
+// encodeData emits the literal-data pseudo-ops: BYTE, WORD, LONG and QUAD
+// write the immediate into the text stream as 1, 2, 4 or 8 bytes,
+// little-endian, with no opcode lookup.  The value is truncated to the
+// width rather than range-checked, exactly as go tool asm behaves (BYTE
+// $0x1FF emits FF, WORD $0x12345 emits 45 23, both without an error), and
+// exactly one immediate is accepted: the toolchain rejects a list such as
+// BYTE $1, $2, $3.
+func (e *enc) encodeData(mnem string, ops []Operand) error {
+	if len(ops) != 1 {
+		return fmt.Errorf("%s expects 1 immediate operand, got %d", mnem, len(ops))
+	}
+	imm, ok := ops[0].(Imm)
+	if !ok {
+		return fmt.Errorf("%s requires an integer immediate", mnem)
+	}
+	width := dataWidth[mnem]
+	out := make([]byte, width)
+	u := uint64(imm)
+	for i := range width {
+		out[i] = byte(u >> (8 * i))
+	}
+	e.out = append(e.out, out...)
+	return nil
+}
+
+// encodeFuncdata accepts-and-ignores the runtime bookkeeping statements:
+// FUNCDATA $n, sym(SB) and PCDATA $n, $m.  go tool asm emits no text bytes
+// for either (the entries live in the object's ancillary tables, not the
+// function body), and the operand shapes it takes are exactly these: an
+// integer count first, then a symbol reference for FUNCDATA and an integer
+// value for PCDATA.  The other architectures accept-and-ignore the same
+// statements; amd64 now matches.
+func (e *enc) encodeFuncdata(upper string, ops []Operand) error {
+	if len(ops) != 2 {
+		return fmt.Errorf("%s expects 2 operands, got %d", upper, len(ops))
+	}
+	if _, ok := ops[0].(Imm); !ok {
+		return fmt.Errorf("%s: first operand must be an integer immediate", upper)
+	}
+	switch upper {
+	case "FUNCDATA":
+		if _, ok := ops[1].(sbMem); !ok {
+			return fmt.Errorf("FUNCDATA: second operand must be a symbol reference")
+		}
+	case "PCDATA":
+		if _, ok := ops[1].(Imm); !ok {
+			return fmt.Errorf("PCDATA: second operand must be an integer immediate")
+		}
+	}
+	return nil
+}
+
+// encodeEnd accepts-and-ignores END.  go tool asm drops the statement
+// entirely: the AEND Prog is skipped when the program list is flushed, so
+// the statements after an END still belong to the same function and the
+// encoded body carries no trace of it, whatever operands follow the name
+// (the toolchain takes END $0 and END AX alike).  Zero bytes, no effect.
+func (e *enc) encodeEnd(ops []Operand) error {
+	return nil
+}
+
+// encodeAdjsp emits ADJSP $imm: a positive value is SUBQ $imm, SP, a
+// negative one ADDQ $-imm, SP, in the imm8 or imm32 form the magnitude
+// picks (the same selection subSP and addSP make for the frame).  go tool
+// asm refuses ADJSP $0 outright, so a zero value is an error here too; the
+// statement's effect on the SP balance is checked by the function-level
+// assembly (checkAdjspBalance), as the toolchain's push/pop walk does.
+func (e *enc) encodeAdjsp(ops []Operand) error {
+	if len(ops) != 1 {
+		return fmt.Errorf("ADJSP expects 1 immediate operand, got %d", len(ops))
+	}
+	imm, ok := ops[0].(Imm)
+	if !ok {
+		return fmt.Errorf("ADJSP requires an integer immediate")
+	}
+	switch v := int(imm); {
+	case v > 0:
+		e.out = append(e.out, subSP(v)...)
+	case v < 0:
+		e.out = append(e.out, addSP(-v)...)
+	default:
+		return fmt.Errorf("ADJSP $0 has no encoding")
+	}
+	return nil
+}
+
+// --- floating-point immediates ----------------------------------------------
+
+// sseFloatImm lists the mnemonics whose first operand may be a floating-point
+// immediate, the set go tool asm rewrites into a pooled-constant read: the
+// scalar moves, the four scalar arithmetic pairs and the scalar compares.
+// The packed members and the uniform forms (MAXSD, MINSD, SQRTSD, CMPSD)
+// reject the immediate in the toolchain and are absent here on purpose.
+var sseFloatImm = map[string]bool{
+	"MOVSD": true, "MOVSS": true,
+	"ADDSD": true, "ADDSS": true,
+	"SUBSD": true, "SUBSS": true,
+	"MULSD": true, "MULSS": true,
+	"DIVSD": true, "DIVSS": true,
+	"COMISD": true, "COMISS": true,
+	"UCOMISD": true, "UCOMISS": true,
+}
+
+// floatImmOperand reports whether the operand list opens with a
+// floating-point immediate in the two-operand spelling (imm, dst).
+func floatImmOperand(ops []Operand) (FloatImm, bool) {
+	if len(ops) != 2 {
+		return FloatImm{}, false
+	}
+	f, ok := ops[0].(FloatImm)
+	return f, ok
+}
+
+// floatPoolValue evaluates a floating-point immediate at the width its
+// mnemonic encodes and names the pool constant the toolchain synthesises:
+// $f64.<16 hex> for the doubles, $f32.<8 hex> for the singles (the float32
+// rounding of the parsed value).  The name carries the IEEE-754 bits; the
+// section holds them little-endian.
+func floatPoolValue(mnem string, f FloatImm) (bits uint64, name string, err error) {
+	v, err := strconv.ParseFloat(f.Text, 64)
+	if err != nil {
+		return 0, "", fmt.Errorf("invalid floating-point immediate %q", f.Text)
+	}
+	if f.Neg {
+		v = -v
+	}
+	if strings.HasSuffix(mnem, "D") {
+		bits = math.Float64bits(v)
+		return bits, fmt.Sprintf("$f64.%016x", bits), nil
+	}
+	bits = uint64(math.Float32bits(float32(v)))
+	return bits, fmt.Sprintf("$f32.%08x", bits), nil
+}
+
+// encodeSSEFloatMove encodes MOVSD/MOVSS with a floating-point immediate
+// source.  A positive zero needs no memory read: the toolchain emits
+// XORPS dst, dst.  Anything else loads the pooled constant RIP-relative
+// ($f64.<hex>(SB) / $f32.<hex>(SB)), the displacement a patch site the
+// file-level layout or the linker resolves.
+func (e *enc) encodeSSEFloatMove(mnem string, f FloatImm, ops []Operand) error {
+	if !sseFloatImm[mnem] {
+		return fmt.Errorf("%s does not take a floating-point immediate", mnem)
+	}
+	dst, ok := ops[1].(Reg)
+	if !ok || !dst.isVec() {
+		return fmt.Errorf("%s: destination must be a vector register", mnem)
+	}
+	bits, name, err := floatPoolValue(mnem, f)
+	if err != nil {
+		return err
+	}
+	e.addFloatPool(name, bits, mwidth(mnem))
+	if bits == 0 {
+		i := &instr{opcode: []byte{0x0F, 0x57}, modrm: -1, sib: -1} // XORPS
+		if err := setRM(i, dst, dst, 8); err != nil {
+			return err
+		}
+		return e.emit(i)
+	}
+	m := sseMoveTable[mnem]
+	i := &instr{prefix: m.prefix, opcode: []byte{0x0F, m.load}, modrm: -1, sib: -1}
+	if err := setRM(i, dst, sbMem{size: mwidth(mnem), name: name}, 8); err != nil {
+		return err
+	}
+	return e.emit(i)
+}
+
+// encodeSSEFloatBin encodes the scalar arithmetic and compare mnemonics with
+// a floating-point immediate source: the constant is read from the pool into
+// the instruction's r/m side (reg = destination), the rewrite go tool asm
+// performs at the source level.
+func (e *enc) encodeSSEFloatBin(mnem string, m sseBin, f FloatImm, ops []Operand) error {
+	if !sseFloatImm[mnem] {
+		return fmt.Errorf("%s does not take a floating-point immediate", mnem)
+	}
+	dst, ok := ops[1].(Reg)
+	if !ok || !dst.isVec() {
+		return fmt.Errorf("%s: destination must be a vector register", mnem)
+	}
+	bits, name, err := floatPoolValue(mnem, f)
+	if err != nil {
+		return err
+	}
+	e.addFloatPool(name, bits, mwidth(mnem))
+	i := &instr{prefix: m.prefix, opcode: []byte{0x0F, m.op}, modrm: -1, sib: -1}
+	if err := setRM(i, dst, sbMem{size: mwidth(mnem), name: name}, 8); err != nil {
+		return err
+	}
+	return e.emit(i)
+}
+
+// mwidth returns the operand width a scalar SSE mnemonic encodes: the double
+// spellings end in D, the single spellings in S.
+func mwidth(mnem string) int {
+	if strings.HasSuffix(mnem, "D") {
+		return 8
+	}
+	return 4
+}
+
 // splitSize separates a trailing B/W/L/Q size suffix from the mnemonic.
 func splitSize(upper string) (base string, size int) {
 	if upper == "" {
@@ -136,7 +542,7 @@ func (e *enc) encodeVec(upper string, ops []Operand, sfx evexSuffix) error {
 	if ss, ok := scatterTable[upper]; ok {
 		return e.encodeScatter(upper, ss, ops, sfx)
 	}
-	if upper == "KMOVW" || upper == "KMOVQ" {
+	if upper == "KMOVW" || upper == "KMOVQ" || upper == "KMOVB" || upper == "KMOVD" {
 		if sfx.any() {
 			return fmt.Errorf("%s takes no EVEX suffixes", upper)
 		}
@@ -173,6 +579,7 @@ type instr struct {
 	disp      []byte
 	imm       []byte
 	sb        *sbRef // static-symbol displacement in disp, awaiting resolution
+	tls       bool   // the displacement is a TLS slot offset, patched R_TLSLE
 }
 
 // sbRef records that an instruction's displacement refers to a static symbol
@@ -215,6 +622,9 @@ func (e *enc) emit(i *instr) error {
 	if i.sb != nil {
 		e.patches = append(e.patches, encPatch{off: len(e.out), name: i.sb.name, addend: i.sb.addend})
 	}
+	if i.tls {
+		e.patches = append(e.patches, encPatch{off: len(e.out), tls: true})
+	}
 	e.out = append(e.out, i.disp...)
 	e.out = append(e.out, i.imm...)
 	return nil
@@ -241,7 +651,7 @@ func setRM(i *instr, reg Reg, rm Operand, opSize int) error {
 }
 
 // setRMDigit fills in the ModR/M for an instruction whose reg field is an
-// opcode /digit extension (0–7), which carries none of the register REX rules.
+// opcode /digit extension (0-7), which carries none of the register REX rules.
 func setRMDigit(i *instr, digit int, rm Operand, opSize int) error {
 	return setRMReg(i, digit, false, false, rm, opSize)
 }
@@ -269,12 +679,30 @@ func setRMReg(i *instr, regField int, rexR, regForced bool, rm Operand, opSize i
 		i.disp = le32(0)
 		i.sb = &sbRef{name: r.name, addend: r.addend}
 		return nil
+	case TLSMem:
+		// off(TLS): the segment-prefixed absolute access, mod=00 with the
+		// SIB escape's disp32 absolute form.  The displacement is the TLS
+		// slot offset, patched by the linker's TLS relocation.
+		i.prefix = r.Seg
+		i.modrm = 0x04 | regField<<3
+		i.sib = 0x25
+		i.disp = le32(r.Disp)
+		i.tls = true
+		return nil
+	case SegAbs:
+		// 0x30(GS): the segment override with the SIB escape's disp32
+		// absolute form, no relocation.
+		setSegAbs(i, regField, r)
+		return nil
 	default:
 		return fmt.Errorf("invalid r/m operand %T", rm)
 	}
 }
 
 func setMem(i *instr, regField int, m Mem) error {
+	if m.Seg != 0 {
+		i.prefix = m.Seg
+	}
 	modrm, sib, disp, xBit, bBit, err := memComponents(regField, m)
 	if err != nil {
 		return err
@@ -287,16 +715,39 @@ func setMem(i *instr, regField int, m Mem) error {
 	return nil
 }
 
+// setSegAbs assembles a segment-absolute operand, 0x30(GS): the segment
+// override with the mod=00 SIB escape's disp32 absolute form and no
+// relocation.
+func setSegAbs(i *instr, regField int, m SegAbs) {
+	i.prefix = m.Seg
+	i.modrm = 0x04 | regField<<3
+	i.sib = 0x25
+	i.disp = le32(m.Disp)
+}
+
 // memComponents computes the ModR/M byte (with the given reg field), the SIB
 // byte (-1 if none), the displacement bytes, and the high index/base bits, for
 // a memory operand.  It is shared by the REX (scalar) and VEX (vector) paths.
 func memComponents(regField int, m Mem) (modrm, sib int, disp []byte, xBit, bBit int, err error) {
 	sib = -1
+	// A displacement wider than int32 fits no encoding form; truncating it
+	// would address a different location, and go tool asm reports "offset
+	// too large" for the same operand.
+	if m.Disp < -(1<<31) || m.Disp > (1<<31)-1 {
+		return 0, -1, nil, 0, 0, fmt.Errorf("displacement %d does not fit in 32 bits", m.Disp)
+	}
 	// RIP-relative: neither base nor index.
 	if !m.HasBase && !m.HasIndex {
 		return regField<<3 | 0x05, -1, le32(m.Disp), 0, 0, nil // mod=00, rm=101
 	}
 
+	// The SIB scale field only encodes 1/2/4/8; the Go assembler rejects
+	// anything else ("bad scale: 16"), so a silent fallback to scale 1 here
+	// would mis-assemble the operand instead of reporting it.
+	if m.HasIndex && m.Scale != 1 && m.Scale != 2 && m.Scale != 4 && m.Scale != 8 {
+		return 0, -1, nil, 0, 0, fmt.Errorf("bad scale: %d", m.Scale)
+	}
+
 	needSIB := m.HasIndex || (m.HasBase && m.Base.idx&7 == 4)
 
 	var mod int
@@ -369,7 +820,7 @@ func le16(v int64) []byte {
 func le64(v int64) []byte {
 	u := uint64(v)
 	b := make([]byte, 8)
-	for i := 0; i < 8; i++ {
+	for i := range 8 {
 		b[i] = byte(u >> (8 * i))
 	}
 	return b

asm/evex_test.go

@@ -16,7 +16,7 @@ import (
 // kernels use: NDS arithmetic, immediate and variable shifts, shuffles with
 // an immediate, lane extracts, narrowing stores, broadcasts from a GPR or
 // memory, mask destinations, mask moves, disp8×N compression and the 5-bit
-// register fields (X/Y 16–31, Z 0–31).
+// register fields (X/Y 16-31, Z 0-31).
 func TestEvexGroundTruth(t *testing.T) {
 	cases := []struct {
 		name string
@@ -38,6 +38,15 @@ func TestEvexGroundTruth(t *testing.T) {
 		{"VADDPD Z11,Z10,Z10", "VADDPD", []Operand{vreg(t, "Z11"), vreg(t, "Z10"), vreg(t, "Z10")}, "6251ad4858d3"},
 		{"VMULPD Z13,Z12,Z12", "VMULPD", []Operand{vreg(t, "Z13"), vreg(t, "Z12"), vreg(t, "Z12")}, "62519d4859e5"},
 		{"VFMADD231PD Z14,Z12,Z10", "VFMADD231PD", []Operand{vreg(t, "Z14"), vreg(t, "Z12"), vreg(t, "Z10")}, "62529d48b8d6"},
+		// The qword OR spelling always encodes through EVEX.
+		{"VPORQ Y0,Y1,Y2", "VPORQ", []Operand{vreg(t, "Y0"), vreg(t, "Y1"), vreg(t, "Y2")}, "62f1f528ebd0"},
+		{"VPORQ X0,X1,X2", "VPORQ", []Operand{vreg(t, "X0"), vreg(t, "X1"), vreg(t, "X2")}, "62f1f508ebd0"},
+		// Byte permute and population count.
+		{"VPERMI2B X0,X1,X2", "VPERMI2B", []Operand{vreg(t, "X0"), vreg(t, "X1"), vreg(t, "X2")}, "62f2750875d0"},
+		{"VPOPCNTB X0,X1", "VPOPCNTB", []Operand{vreg(t, "X0"), vreg(t, "X1")}, "62f27d0854c8"},
+		{"VPOPCNTD X0,X1", "VPOPCNTD", []Operand{vreg(t, "X0"), vreg(t, "X1")}, "62f27d0855c8"},
+		{"VPOPCNTD Y0,Y1", "VPOPCNTD", []Operand{vreg(t, "Y0"), vreg(t, "Y1")}, "62f27d2855c8"},
+		{"VPOPCNTQ X0,X1", "VPOPCNTQ", []Operand{vreg(t, "X0"), vreg(t, "X1")}, "62f2fd0855c8"},
 		// Align (NDS + imm8).
 		{"VALIGND $12,Z12,Z0,Z1", "VALIGND", []Operand{Imm(12), vreg(t, "Z12"), vreg(t, "Z0"), vreg(t, "Z1")}, "62d37d4803cc0c"},
 		{"VALIGND $15,Z9,Z0,Z1", "VALIGND", []Operand{Imm(15), vreg(t, "Z9"), vreg(t, "Z0"), vreg(t, "Z1")}, "62d37d4803c90f"},
@@ -52,13 +61,23 @@ func TestEvexGroundTruth(t *testing.T) {
 		{"KMOVW K1,CX", "KMOVW", []Operand{vreg(t, "K1"), CX}, "c5f893c9"},
 		{"KMOVW K1,R12", "KMOVW", []Operand{vreg(t, "K1"), vreg(t, "R12")}, "c57893e1"},
 		{"KTESTW K1,K1", "KTESTW", []Operand{vreg(t, "K1"), vreg(t, "K1")}, "c5f899c9"},
+		{"KMOVB K1,K2", "KMOVB", []Operand{vreg(t, "K1"), vreg(t, "K2")}, "c5f990d1"},
+		{"KMOVB AX,K1", "KMOVB", []Operand{AX, vreg(t, "K1")}, "c5f992c8"},
+		{"KMOVB K1,AX", "KMOVB", []Operand{vreg(t, "K1"), AX}, "c5f993c1"},
+		{"KMOVB K1,(AX)", "KMOVB", []Operand{vreg(t, "K1"), Ptr(AX, 0, 1)}, "c5f99108"},
+		{"KMOVD K1,K2", "KMOVD", []Operand{vreg(t, "K1"), vreg(t, "K2")}, "c4e1f990d1"},
+		{"KMOVD AX,K1", "KMOVD", []Operand{AX, vreg(t, "K1")}, "c5fb92c8"},
+		{"KMOVD K1,AX", "KMOVD", []Operand{vreg(t, "K1"), AX}, "c5fb93c1"},
+		{"KMOVD K1,(AX)", "KMOVD", []Operand{vreg(t, "K1"), Ptr(AX, 0, 4)}, "c4e1f99108"},
+		{"KMOVB (AX),K1", "KMOVB", []Operand{Ptr(AX, 0, 1), vreg(t, "K1")}, "c5f99008"},
+		{"KMOVQ (AX),K1", "KMOVQ", []Operand{Ptr(AX, 0, 8), vreg(t, "K1")}, "c4e1f89008"},
 		// Moves, incl. disp8×N (64 for a 512-bit operand).
 		{"VMOVDQU32 (SI)(R15*4),Z3", "VMOVDQU32", []Operand{Idx(SI, vreg(t, "R15"), 4, 0, 64), vreg(t, "Z3")}, "62b17e486f1cbe"},
 		{"VMOVDQU32 4(SI)(AX*1),Z4", "VMOVDQU32", []Operand{Idx(SI, AX, 1, 4, 64), vreg(t, "Z4")}, "62f17e486fa40604000000"},
 		{"VMOVDQU32 16(SI)(R15*4),Z4", "VMOVDQU32", []Operand{Idx(SI, vreg(t, "R15"), 4, 16, 64), vreg(t, "Z4")}, "62b17e486fa4be10000000"},
 		{"VMOVDQU32 Z0,4(SI)(AX*1)", "VMOVDQU32", []Operand{vreg(t, "Z0"), Idx(SI, AX, 1, 4, 64)}, "62f17e487f840604000000"},
 		{"VMOVDQU32 Z3,(DI)(R15*4)", "VMOVDQU32", []Operand{vreg(t, "Z3"), Idx(DI, vreg(t, "R15"), 4, 0, 64)}, "62b17e487f1cbf"},
-		// VMOVDQU64 — the W1 qword variant.
+		// VMOVDQU64; the W1 qword variant.
 		{"VMOVDQU64 (SI)(R15*4),Z3", "VMOVDQU64", []Operand{Idx(SI, vreg(t, "R15"), 4, 0, 64), vreg(t, "Z3")}, "62b1fe486f1cbe"},
 		{"VMOVDQU64 Z0,4(SI)(AX*1)", "VMOVDQU64", []Operand{vreg(t, "Z0"), Idx(SI, AX, 1, 4, 64)}, "62f1fe487f840604000000"},
 		{"VMOVDQU64 Z1,Z2", "VMOVDQU64", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f1fe487fca"},
@@ -77,7 +96,7 @@ func TestEvexGroundTruth(t *testing.T) {
 		{"VPSHUFB Z1,Z2,Z3", "VPSHUFB", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f26d4800d9"},
 		{"VMOVDQU8 Z1,Z2", "VMOVDQU8", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f17f487fca"},
 		{"VMOVDQU16 Z1,Z2", "VMOVDQU16", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f1ff487fca"},
-		// Indices 16–31: rm[4] rides in X̄ for register operands.
+		// Indices 16-31: rm[4] rides in X̄ for register operands.
 		{"VPSHUFD $1,X16,X17", "VPSHUFD", []Operand{Imm(1), vreg(t, "X16"), vreg(t, "X17")}, "62a17d0870c801"},
 		{"VMOVUPD (DI),Z14", "VMOVUPD", []Operand{Ptr(DI, 0, 64), vreg(t, "Z14")}, "6271fd481037"},
 		{"VMOVUPD 64(DI),Z14", "VMOVUPD", []Operand{Ptr(DI, 64, 64), vreg(t, "Z14")}, "6271fd48107701"},
@@ -96,7 +115,7 @@ func TestEvexGroundTruth(t *testing.T) {
 		{"VPBROADCASTD 4(SI),Z10", "VPBROADCASTD", []Operand{Ptr(SI, 4, 4), vreg(t, "Z10")}, "62727d48585601"},
 		{"VPBROADCASTQ R8,X31", "VPBROADCASTQ", []Operand{vreg(t, "R8"), vreg(t, "X31")}, "6242fd087cf8"},
 		{"VPBROADCASTQ AX,Z9", "VPBROADCASTQ", []Operand{AX, vreg(t, "Z9")}, "6272fd487cc8"},
-		// Register indices 16–31 exist only in EVEX encodings.
+		// Register indices 16-31 exist only in EVEX encodings.
 		{"VPBROADCASTD AX,Y30", "VPBROADCASTD", []Operand{AX, vreg(t, "Y30")}, "62627d287cf0"},
 		// Packed double arithmetic / unpack (EVEX forms carry W=1).
 		{"VSUBPD Z1,Z2,Z3", "VSUBPD", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f1ed485cd9"},
@@ -107,7 +126,7 @@ func TestEvexGroundTruth(t *testing.T) {
 		{"VUNPCKHPD Z1,Z2,Z3", "VUNPCKHPD", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f1ed4815d9"},
 		{"VSUBPD 64(AX),Z1,Z2", "VSUBPD", []Operand{Ptr(AX, 64, 64), vreg(t, "Z1"), vreg(t, "Z2")}, "62f1f5485c5001"},
 		{"VSUBPD Z17,Z18,Z19", "VSUBPD", []Operand{vreg(t, "Z17"), vreg(t, "Z18"), vreg(t, "Z19")}, "62a1ed405cd9"},
-		// VMOVDDUP — duplicate the low double; disp8×N = 64 at 512 bits, and
+		// VMOVDDUP; duplicate the low double; disp8×N = 64 at 512 bits, and
 		// X16/X17 force EVEX (the mod=11 rm[4] extension rides in X̄).
 		{"VMOVDDUP Z1,Z2", "VMOVDDUP", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f1ff4812d1"},
 		{"VMOVDDUP 64(AX),Z1", "VMOVDDUP", []Operand{Ptr(AX, 64, 64), vreg(t, "Z1")}, "62f1ff48124801"},
@@ -150,7 +169,7 @@ func TestEvexGroundTruth(t *testing.T) {
 	}
 }
 
-// TestEvexMasking checks the AVX-512 mask operand (K1–K7, placed freely among
+// TestEvexMasking checks the AVX-512 mask operand (K1-K7, placed freely among
 // the operands) and the .Z zeroing suffix, byte for byte against the Go
 // assembler.
 func TestEvexMasking(t *testing.T) {
@@ -241,11 +260,11 @@ func TestEvexMasking(t *testing.T) {
 	}
 }
 
-// TestEvexExtendedGroundTruth covers the wider EVEX/AVX-512 set — ternary
+// TestEvexExtendedGroundTruth covers the wider EVEX/AVX-512 set; ternary
 // logic, lane shuffles/inserts/extracts, compares with a K destination,
 // permutes, the wider integer families, expand/compress, broadcasts,
 // rotates and word shifts, the opmask instructions, the EVEX suffixes
-// (rounding/SAE/broadcast) and the aligned/scalar moves — byte for byte
+// (rounding/SAE/broadcast) and the aligned/scalar moves; byte for byte
 // against the Go assembler.
 func TestEvexExtendedGroundTruth(t *testing.T) {
 	mem64 := func(base Reg) Operand { return Ptr(base, 0, 64) }
@@ -275,7 +294,7 @@ func TestEvexExtendedGroundTruth(t *testing.T) {
 		{"VMULPD.RZ_SAE.Z", "VMULPD.RZ_SAE.Z", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "K1"), vreg(t, "Z3")}, "62f1edf959d9"},
 		{"VMAXPD.SAE", "VMAXPD.SAE", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f1ed585fd9"},
 		{"VADDPD.BCST", "VADDPD.BCST", []Operand{mem64(AX), vreg(t, "Z1"), vreg(t, "Z2")}, "62f1f5585810"},
-		// Packed single arithmetic (same opcodes, no mandatory prefix) —
+		// Packed single arithmetic (same opcodes, no mandatory prefix);
 		// ZMM, YMM and XMM widths, rounding and broadcast.
 		{"VADDPS", "VADDPS", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f16c4858d9"},
 		{"VMULPS", "VMULPS", []Operand{vreg(t, "Y1"), vreg(t, "Y2"), vreg(t, "Y3")}, "c5ec59d9"},
@@ -319,6 +338,43 @@ func TestEvexExtendedGroundTruth(t *testing.T) {
 		{"KORTESTD", "KORTESTD", []Operand{vreg(t, "K1"), vreg(t, "K2")}, "c4e1f998d1"},
 		{"KMOVQ k,k", "KMOVQ", []Operand{vreg(t, "K1"), vreg(t, "K2")}, "c4e1f890d1"},
 		{"KMOVQ gpr,k", "KMOVQ", []Operand{BX, vreg(t, "K1")}, "c4e1fb92cb"},
+		// Completed opmask families (ANDN, NOT, OR/XOR word+qword, TEST,
+		// word-width shifts; byte-exact against go tool asm).
+		{"KANDNW", "KANDNW", []Operand{vreg(t, "K1"), vreg(t, "K2"), vreg(t, "K3")}, "c5ec42d9"},
+		{"KANDNB", "KANDNB", []Operand{vreg(t, "K4"), vreg(t, "K5"), vreg(t, "K6")}, "c5d542f4"},
+		{"KANDND", "KANDND", []Operand{vreg(t, "K1"), vreg(t, "K2"), vreg(t, "K3")}, "c4e1ed42d9"},
+		{"KANDNQ", "KANDNQ", []Operand{vreg(t, "K4"), vreg(t, "K5"), vreg(t, "K6")}, "c4e1d442f4"},
+		{"KANDD", "KANDD", []Operand{vreg(t, "K1"), vreg(t, "K2"), vreg(t, "K3")}, "c4e1ed41d9"},
+		{"KADDD", "KADDD", []Operand{vreg(t, "K4"), vreg(t, "K5"), vreg(t, "K6")}, "c4e1d54af4"},
+		{"KNOTW", "KNOTW", []Operand{vreg(t, "K1"), vreg(t, "K2")}, "c5f844d1"},
+		{"KNOTD", "KNOTD", []Operand{vreg(t, "K3"), vreg(t, "K4")}, "c4e1f944e3"},
+		{"KNOTQ", "KNOTQ", []Operand{vreg(t, "K5"), vreg(t, "K6")}, "c4e1f844f5"},
+		{"KORW", "KORW", []Operand{vreg(t, "K1"), vreg(t, "K2"), vreg(t, "K3")}, "c5ec45d9"},
+		{"KORQ", "KORQ", []Operand{vreg(t, "K4"), vreg(t, "K5"), vreg(t, "K6")}, "c4e1d445f4"},
+		{"KXNORB", "KXNORB", []Operand{vreg(t, "K1"), vreg(t, "K2"), vreg(t, "K3")}, "c5ed46d9"},
+		{"KXORW", "KXORW", []Operand{vreg(t, "K1"), vreg(t, "K2"), vreg(t, "K3")}, "c5ec47d9"},
+		{"KXORQ", "KXORQ", []Operand{vreg(t, "K4"), vreg(t, "K5"), vreg(t, "K6")}, "c4e1d447f4"},
+		{"KORTESTW", "KORTESTW", []Operand{vreg(t, "K1"), vreg(t, "K2")}, "c5f898d1"},
+		{"KORTESTB", "KORTESTB", []Operand{vreg(t, "K3"), vreg(t, "K4")}, "c5f998e3"},
+		{"KORTESTQ", "KORTESTQ", []Operand{vreg(t, "K5"), vreg(t, "K6")}, "c4e1f898f5"},
+		{"KTESTW", "KTESTW", []Operand{vreg(t, "K1"), vreg(t, "K2")}, "c5f899d1"},
+		{"KTESTD", "KTESTD", []Operand{vreg(t, "K3"), vreg(t, "K4")}, "c4e1f999e3"},
+		{"KSHIFTLB", "KSHIFTLB", []Operand{Imm(1), vreg(t, "K1"), vreg(t, "K2")}, "c4e37932d101"},
+		{"KSHIFTLD", "KSHIFTLD", []Operand{Imm(2), vreg(t, "K3"), vreg(t, "K4")}, "c4e37933e302"},
+		{"KSHIFTLQ", "KSHIFTLQ", []Operand{Imm(3), vreg(t, "K5"), vreg(t, "K6")}, "c4e3f933f503"},
+		{"KSHIFTRB", "KSHIFTRB", []Operand{Imm(4), vreg(t, "K1"), vreg(t, "K2")}, "c4e37930d104"},
+		{"KSHIFTRW", "KSHIFTRW", []Operand{Imm(5), vreg(t, "K3"), vreg(t, "K4")}, "c4e3f930e305"},
+		{"KSHIFTRQ", "KSHIFTRQ", []Operand{Imm(6), vreg(t, "K5"), vreg(t, "K6")}, "c4e3f931f506"},
+		// Integer compares with an opmask destination (0F3A map, the
+		// go-bzip2 partition kernel's classify instructions).
+		{"VPCMPUB", "VPCMPUB", []Operand{Imm(1), vreg(t, "X1"), vreg(t, "X0"), vreg(t, "K1")}, "62f37d083ec901"},
+		{"VPCMPB", "VPCMPB", []Operand{Imm(2), vreg(t, "Y2"), vreg(t, "Y3"), vreg(t, "K2")}, "62f365283fd202"},
+		{"VPCMPUW", "VPCMPUW", []Operand{Imm(5), vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "K3")}, "62f3ed483ed905"},
+		{"VPCMPW", "VPCMPW", []Operand{Imm(6), vreg(t, "X3"), vreg(t, "X4"), vreg(t, "K4")}, "62f3dd083fe306"},
+		{"VPCMPD", "VPCMPD", []Operand{Imm(0), vreg(t, "Y1"), vreg(t, "Y2"), vreg(t, "K1")}, "62f36d281fc900"},
+		{"VPCMPUD", "VPCMPUD", []Operand{Imm(1), vreg(t, "Z2"), vreg(t, "Z3"), vreg(t, "K2")}, "62f365481ed201"},
+		{"VPCMPQ", "VPCMPQ", []Operand{Imm(2), vreg(t, "X1"), vreg(t, "X2"), vreg(t, "K3")}, "62f3ed081fd902"},
+		{"VPCMPUQ", "VPCMPUQ", []Operand{Imm(3), vreg(t, "Y3"), vreg(t, "Y4"), vreg(t, "K4")}, "62f3dd281ee303"},
 		// Lane extract / insert.
 		{"VEXTRACTF32X4", "VEXTRACTF32X4", []Operand{Imm(1), vreg(t, "Y1"), vreg(t, "X2")}, "62f37d2819ca01"},
 		{"VEXTRACTI64X2", "VEXTRACTI64X2", []Operand{Imm(1), vreg(t, "Y1"), vreg(t, "X2")}, "62f3fd2839ca01"},
@@ -362,8 +418,8 @@ func TestEvexExtendedGroundTruth(t *testing.T) {
 }
 
 // TestEvexHelperGroundTruth covers the floating-point helper and conversion
-// tail of the EVEX set — reciprocals, rsqrt, getexp/getmant, scalef,
-// rndscale, reduce, fixupimm, range, fpclass, the remaining conversions —
+// tail of the EVEX set; reciprocals, rsqrt, getexp/getmant, scalef,
+// rndscale, reduce, fixupimm, range, fpclass, the remaining conversions;
 // plus gather/scatter with VSIB addressing, byte for byte against the Go
 // assembler.
 func TestEvexHelperGroundTruth(t *testing.T) {
@@ -465,9 +521,9 @@ func TestEvexHelperGroundTruth(t *testing.T) {
 }
 
 // TestEvexGprGroundTruth covers the scalar conversions between vector and
-// general-purpose registers — the signed and truncated VCVT{,T}S{D,S}2SI
+// general-purpose registers; the signed and truncated VCVT{,T}S{D,S}2SI
 // forms (VEX and EVEX), the unsigned EVEX-only forms, and the GPR-to-vector
-// VCVTSI2*/VCVTUSI2* forms with the preserved vector source in vvvv — byte
+// VCVTSI2*/VCVTUSI2* forms with the preserved vector source in vvvv; byte
 // for byte against the Go assembler, including memory sources and extended
 // GPRs.
 func TestEvexGprGroundTruth(t *testing.T) {
@@ -638,6 +694,15 @@ func TestEvexErrors(t *testing.T) {
 		{"align arity", "VALIGND", []Operand{Imm(1), vreg(t, "Z0"), vreg(t, "Z1")}},
 		// VEX-only mnemonics reject registers only EVEX can encode.
 		{"VMOVMSKPS X16", "VMOVMSKPS", []Operand{vreg(t, "X16"), AX}},
+		// The scalar EVEX move matches its VEX twin and the Go assembler:
+		// XMM↔memory only, never reg-reg and never a wider register (the
+		// toolchain rejects every one of these shapes).
+		{"VMOVSS X1,X2", "VMOVSS", []Operand{vreg(t, "X1"), vreg(t, "X2")}},
+		{"VMOVSS X16,X2", "VMOVSS", []Operand{vreg(t, "X16"), vreg(t, "X2")}},
+		{"VMOVSS Y1,(AX)", "VMOVSS", []Operand{vreg(t, "Y1"), Ptr(AX, 0, 4)}},
+		{"VMOVSS Z1,Z2", "VMOVSS", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}},
+		{"VMOVSS Z1,(AX)", "VMOVSS", []Operand{vreg(t, "Z1"), Ptr(AX, 0, 4)}},
+		{"VMOVSS (AX),Z2", "VMOVSS", []Operand{Ptr(AX, 0, 4), vreg(t, "Z2")}},
 	}
 	for _, c := range cases {
 		if _, err := Encode(c.mnem, c.ops...); err == nil {
@@ -656,3 +721,216 @@ func hexCompact(b []byte) string {
 	}
 	return string(out)
 }
+
+// TestAvx512CorpusFamilies pins representative encodings of the AVX-512
+// families the toolchain's avx512enc corpus exercises: the bytes are the
+// go tool asm output for exactly these operands, and the same families are
+// covered end to end by the avx512_amd64.s differential kernel.
+func TestAvx512CorpusFamilies(t *testing.T) {
+	vsib := func(base, idx string, scale int) Operand {
+		return Idx(vreg(t, base), vreg(t, idx), scale, 0, 0)
+	}
+	cases := []struct {
+		name string
+		mnem string
+		ops  []Operand
+		want string
+	}{
+		// AES rounds (EVEX NDS, VEX twin routed by operand width).
+		{"VAESDEC Z", "VAESDEC", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f26d48ded9"},
+		// Integer VNNI and the bit algorithm group.
+		{"VPDPBUSD", "VPDPBUSD", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "K2"), vreg(t, "Z3")}, "62f26d4a50d9"},
+		{"VPOPCNTW", "VPOPCNTW", []Operand{vreg(t, "Z1"), vreg(t, "K3"), vreg(t, "Z2")}, "62f2fd4b54d1"},
+		{"VPCONFLICTD", "VPCONFLICTD", []Operand{vreg(t, "Z1"), vreg(t, "K1"), vreg(t, "Z2")}, "62f27d49c4d1"},
+		{"VPLZCNTQ masked", "VPLZCNTQ", []Operand{vreg(t, "Z7"), vreg(t, "K1"), vreg(t, "Z8")}, "6272fd4944c7"},
+		{"VPERMT2B", "VPERMT2B", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "K1"), vreg(t, "Z3")}, "62f26d497dd9"},
+		{"VPMULTISHIFTQB", "VPMULTISHIFTQB", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "K3"), vreg(t, "Z4")}, "62f2ed4b83e1"},
+		{"VDBPSADBW", "VDBPSADBW", []Operand{Imm(3), vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "K3"), vreg(t, "Z3")}, "62f36d4b42d903"},
+		{"VPSHUFBITQMB", "VPSHUFBITQMB", []Operand{vreg(t, "Z9"), vreg(t, "Z10"), vreg(t, "K3")}, "62d22d488fd9"},
+		{"VPTESTNMQ", "VPTESTNMQ", []Operand{vreg(t, "Z13"), vreg(t, "Z14"), vreg(t, "K5")}, "62d28e4827ed"},
+		// Permutations: immediate and register counts.
+		{"VALIGNQ", "VALIGNQ", []Operand{Imm(3), vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "K1"), vreg(t, "Z3")}, "62f3ed4903d903"},
+		{"VPERMQ imm", "VPERMQ", []Operand{Imm(1), vreg(t, "Z1"), vreg(t, "K2"), vreg(t, "Z2")}, "62f3fd4a00d101"},
+		{"VPERMQ reg", "VPERMQ", []Operand{vreg(t, "Z3"), vreg(t, "Z4"), vreg(t, "K2"), vreg(t, "Z5")}, "62f2dd4a36eb"},
+		{"VPERMPD reg", "VPERMPD", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f2ed4816d9"},
+		{"VPERMILPS imm", "VPERMILPS", []Operand{Imm(5), vreg(t, "Z9"), vreg(t, "K2"), vreg(t, "Z10")}, "62537d4a04d105"},
+		{"VPERMILPS reg", "VPERMILPS", []Operand{vreg(t, "Z11"), vreg(t, "Z12"), vreg(t, "K2"), vreg(t, "Z13")}, "62521d4a0ceb"},
+		// Shifts: immediate, register-count and memory-count forms; the
+		// count source carries its own XMM tuple width.
+		{"VPSLLW imm mask", "VPSLLW", []Operand{Imm(3), vreg(t, "Z1"), vreg(t, "K2"), vreg(t, "Z2")}, "62f16d4a71f103"},
+		{"VPSLLD reg count", "VPSLLD", []Operand{vreg(t, "X1"), vreg(t, "Z2"), vreg(t, "K1"), vreg(t, "Z3")}, "62f16d49f2d9"},
+		{"VPSLLDQ", "VPSLLDQ", []Operand{Imm(9), vreg(t, "Z7"), vreg(t, "Z8")}, "62f13d4873ff09"},
+		{"VPSRLDQ mem", "VPSRLDQ", []Operand{Imm(11), Ptr(SI, 16, 16), vreg(t, "Z4")}, "62f15d48739e100000000b"},
+		{"VPSRLVW", "VPSRLVW", []Operand{vreg(t, "Z3"), vreg(t, "Z4"), vreg(t, "K1"), vreg(t, "Z5")}, "62f2dd4910eb"},
+		// Conversions and shuffles with the F2 prefix and no prefix.
+		{"VCVTUDQ2PS", "VCVTUDQ2PS", []Operand{vreg(t, "Z1"), vreg(t, "K1"), vreg(t, "Z2")}, "62f17f497ad1"},
+		{"VSHUFPS", "VSHUFPS", []Operand{Imm(2), vreg(t, "Z4"), vreg(t, "Z5"), vreg(t, "K1"), vreg(t, "Z6")}, "62f15449c6f402"},
+		// Gather and scatter prefetch hints (memory-only, /digit in reg).
+		{"VGATHERPF0DPD", "VGATHERPF0DPD", []Operand{vreg(t, "K5"), vsib("R10", "Y29", 8)}, "6292fd45c60cea"},
+		{"VSCATTERPF1DPS", "VSCATTERPF1DPS", []Operand{vreg(t, "K2"), vsib("R10", "Z28", 4)}, "62927d42c634a2"},
+		// Opmask broadcasts and the K logic.
+		{"VPBROADCASTMB2Q", "VPBROADCASTMB2Q", []Operand{vreg(t, "K1"), vreg(t, "Z2")}, "62f2fe482ad1"},
+		{"VPBROADCASTMW2D", "VPBROADCASTMW2D", []Operand{vreg(t, "K3"), vreg(t, "Z4")}, "62f27e483ae3"},
+		{"KUNPCKWD", "KUNPCKWD", []Operand{vreg(t, "K6"), vreg(t, "K4"), vreg(t, "K1")}, "c5dc4bce"},
+		{"KADDB", "KADDB", []Operand{vreg(t, "K2"), vreg(t, "K3"), vreg(t, "K5")}, "c5e54aea"},
+		// Lane extracts to general registers (EVEX and VEX routes).
+		{"VPEXTRB", "VPEXTRB", []Operand{Imm(3), vreg(t, "X26"), AX}, "62637d0814d003"},
+		{"VPEXTRD", "VPEXTRD", []Operand{Imm(1), vreg(t, "X26"), vreg(t, "R9")}, "62437d0816d101"},
+		{"VPEXTRD vex", "VPEXTRD", []Operand{Imm(1), vreg(t, "X2"), DI}, "c4e37916d701"},
+		{"VPINSRQ", "VPINSRQ", []Operand{Imm(1), DI, vreg(t, "X3"), vreg(t, "X4")}, "c4e3e122e701"},
+		// Moves: masked unaligned, masked scalar register form, half moves
+		// and non-temporal stores.
+		{"VMOVUPS mask", "VMOVUPS", []Operand{vreg(t, "Z1"), vreg(t, "K2"), vreg(t, "Z3")}, "62f17c4a11cb"},
+		{"VMOVSD 3op", "VMOVSD", []Operand{vreg(t, "X14"), vreg(t, "X5"), vreg(t, "K3"), vreg(t, "X22")}, "6231d70b11f6"},
+		{"VMOVSS 3op", "VMOVSS", []Operand{vreg(t, "X18"), vreg(t, "X3"), vreg(t, "K2"), vreg(t, "X25")}, "6281660a11d1"},
+		{"VMOVHPS insert", "VMOVHPS", []Operand{Ptr(SI, 0, 8), vreg(t, "X18"), vreg(t, "X19")}, "62e16c00161e"},
+		{"VMOVHPS store", "VMOVHPS", []Operand{vreg(t, "X20"), Ptr(SI, 8, 8)}, "62e17c08176601"},
+		{"VMOVLHPS", "VMOVLHPS", []Operand{vreg(t, "X16"), vreg(t, "X5"), vreg(t, "X17")}, "62a1540816c8"},
+		{"VMOVNTDQ", "VMOVNTDQ", []Operand{vreg(t, "Z7"), Ptr(SI, 0, 64)}, "62f17d48e73e"},
+		{"VMOVNTDQA", "VMOVNTDQA", []Operand{Ptr(SI, 64, 64), vreg(t, "Z8")}, "62727d482a4601"},
+		{"VMOVNTPS", "VMOVNTPS", []Operand{vreg(t, "Z9"), Ptr(SI, 0, 64)}, "62717c482b0e"},
+		// Scalar compares with and without the 66 prefix.
+		{"VCOMISD", "VCOMISD", []Operand{vreg(t, "X5"), vreg(t, "X6")}, "c5f92ff5"},
+		{"VUCOMISS", "VUCOMISS", []Operand{vreg(t, "X7"), vreg(t, "X8")}, "c5782ec7"},
+		// Floating point helpers.
+		{"VSQRTSD", "VSQRTSD", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "K1"), vreg(t, "X3")}, "62f1ef0951d9"},
+		{"VEXP2PD", "VEXP2PD", []Operand{vreg(t, "Z5"), vreg(t, "K1"), vreg(t, "Z6")}, "62f2fd49c8f5"},
+		{"VRCP28SD", "VRCP28SD", []Operand{vreg(t, "X9"), vreg(t, "X8"), vreg(t, "K1"), vreg(t, "X10")}, "6252bd09cbd1"},
+		{"VBROADCASTF32X2", "VBROADCASTF32X2", []Operand{vreg(t, "X1"), vreg(t, "K1"), vreg(t, "Z2")}, "62f27d4919d1"},
+		{"VPCOMPRESSB", "VPCOMPRESSB", []Operand{vreg(t, "Z1"), vreg(t, "K1"), Ptr(SI, 0, 64)}, "62f27d49630e"},
+	}
+	for _, c := range cases {
+		code, err := Encode(c.mnem, c.ops...)
+		if err != nil {
+			t.Errorf("%s: Encode: %v", c.name, err)
+			continue
+		}
+		if got := hexCompact(code); got != c.want {
+			t.Errorf("%s: got %s, want %s", c.name, got, c.want)
+		}
+	}
+}
+
+// TestEvexQuadRegisterGroundTruth pins the quad-register instructions (the
+// 4FMAPS and 4VNNIW families) byte for byte against go tool asm: the memory
+// source keeps r/m, the bracketed list's LOW register travels the inverted
+// 5-bit V'VVVV field, the destination sits in reg, the opmask rides aaa and
+// the vector length follows the destination (L'L=512 for the ZMM forms,
+// 128 for the scalar ones) while the disp8×N multiplier stays 16 for every
+// member.  The x86 decoder has no view of these forms, so no decode check
+// runs.
+func TestEvexQuadRegisterGroundTruth(t *testing.T) {
+	sp := vreg(t, "RSP")
+	cases := []struct {
+		name string
+		mnem string
+		ops  []Operand
+		want string
+	}{
+		{"V4FMADDPS 17(SP) [Z0-Z3] K2 Z0", "V4FMADDPS",
+			[]Operand{Ptr(sp, 17, 8), RegList{vreg(t, "Z0"), vreg(t, "Z3")}, vreg(t, "K2"), vreg(t, "Z0")},
+			"62f27f4a9a842411000000"},
+		{"V4FMADDPS [Z10-Z13]", "V4FMADDPS",
+			[]Operand{Ptr(sp, 17, 8), RegList{vreg(t, "Z10"), vreg(t, "Z13")}, vreg(t, "K2"), vreg(t, "Z0")},
+			"62f22f4a9a842411000000"},
+		{"V4FMADDPS [Z20-Z23]", "V4FMADDPS",
+			[]Operand{Ptr(sp, 17, 8), RegList{vreg(t, "Z20"), vreg(t, "Z23")}, vreg(t, "K2"), vreg(t, "Z0")},
+			"62f25f429a842411000000"},
+		{"V4FMADDPS Z8 dst", "V4FMADDPS",
+			[]Operand{Ptr(sp, 17, 8), RegList{vreg(t, "Z0"), vreg(t, "Z3")}, vreg(t, "K2"), vreg(t, "Z8")},
+			"62727f4a9a842411000000"},
+		{"V4FMADDPS disp8x16", "V4FMADDPS",
+			[]Operand{Ptr(sp, 64, 8), RegList{vreg(t, "Z0"), vreg(t, "Z3")}, vreg(t, "K2"), vreg(t, "Z0")},
+			"62f27f4a9a442404"},
+		{"V4FMADDPS unmasked", "V4FMADDPS",
+			[]Operand{Ptr(sp, 17, 8), RegList{vreg(t, "Z0"), vreg(t, "Z3")}, vreg(t, "Z0")},
+			"62f27f489a842411000000"},
+		{"V4FMADDSS 7(AX) [X0-X3] K5 X22", "V4FMADDSS",
+			[]Operand{Ptr(AX, 7, 8), RegList{vreg(t, "X0"), vreg(t, "X3")}, vreg(t, "K5"), vreg(t, "X22")},
+			"62e27f0d9bb007000000"},
+		{"V4FMADDSS (DI)", "V4FMADDSS",
+			[]Operand{Ptr(DI, 0, 8), RegList{vreg(t, "X0"), vreg(t, "X3")}, vreg(t, "K5"), vreg(t, "X22")},
+			"62e27f0d9b37"},
+		{"V4FMADDSS [X10-X13]", "V4FMADDSS",
+			[]Operand{Ptr(AX, 7, 8), RegList{vreg(t, "X10"), vreg(t, "X13")}, vreg(t, "K5"), vreg(t, "X22")},
+			"62e22f0d9bb007000000"},
+		{"V4FMADDSS [X20-X23]", "V4FMADDSS",
+			[]Operand{Ptr(AX, 7, 8), RegList{vreg(t, "X20"), vreg(t, "X23")}, vreg(t, "K5"), vreg(t, "X22")},
+			"62e25f059bb007000000"},
+		{"V4FMADDSS X30 dst", "V4FMADDSS",
+			[]Operand{Ptr(AX, 7, 8), RegList{vreg(t, "X0"), vreg(t, "X3")}, vreg(t, "K5"), vreg(t, "X30")},
+			"62627f0d9bb007000000"},
+		{"V4FMADDSS X3 dst", "V4FMADDSS",
+			[]Operand{Ptr(AX, 7, 8), RegList{vreg(t, "X0"), vreg(t, "X3")}, vreg(t, "K5"), vreg(t, "X3")},
+			"62f27f0d9b9807000000"},
+		{"V4FMADDSS disp8x16", "V4FMADDSS",
+			[]Operand{Ptr(AX, 16, 8), RegList{vreg(t, "X20"), vreg(t, "X23")}, vreg(t, "K5"), vreg(t, "X30")},
+			"62625f059b7001"},
+		{"V4FNMADDPS", "V4FNMADDPS",
+			[]Operand{Ptr(sp, 17, 8), RegList{vreg(t, "Z0"), vreg(t, "Z3")}, vreg(t, "K2"), vreg(t, "Z0")},
+			"62f27f4aaa842411000000"},
+		{"V4FNMADDSS", "V4FNMADDSS",
+			[]Operand{Ptr(AX, 7, 8), RegList{vreg(t, "X0"), vreg(t, "X3")}, vreg(t, "K5"), vreg(t, "X22")},
+			"62e27f0dabb007000000"},
+		{"VP4DPWSSD", "VP4DPWSSD",
+			[]Operand{Ptr(sp, 17, 8), RegList{vreg(t, "Z0"), vreg(t, "Z3")}, vreg(t, "K2"), vreg(t, "Z0")},
+			"62f27f4a52842411000000"},
+		{"VP4DPWSSDS unmasked", "VP4DPWSSDS",
+			[]Operand{Ptr(sp, 17, 8), RegList{vreg(t, "Z0"), vreg(t, "Z3")}, vreg(t, "Z0")},
+			"62f27f4853842411000000"},
+	}
+	for _, c := range cases {
+		code, err := Encode(c.mnem, c.ops...)
+		if err != nil {
+			t.Errorf("%s: Encode: %v", c.name, err)
+			continue
+		}
+		if got := hexCompact(code); got != c.want {
+			t.Errorf("%s: got %s, want %s", c.name, got, c.want)
+		}
+	}
+}
+
+// TestEvexQuadRegisterErrors pins the operand shapes the toolchain rejects:
+// the register class the list and the destination take is fixed per
+// instruction, the source is memory only, the opmask slot is positional and
+// the list's low register owns V'VVVV.
+func TestEvexQuadRegisterErrors(t *testing.T) {
+	sp := vreg(t, "RSP")
+	list := func(lo, hi string) RegList {
+		return RegList{vreg(t, lo), vreg(t, hi)}
+	}
+	cases := []struct {
+		name string
+		mnem string
+		ops  []Operand
+	}{
+		{"X list on the PS form", "V4FMADDPS",
+			[]Operand{Ptr(sp, 0, 8), list("X0", "X3"), vreg(t, "K2"), vreg(t, "Z0")}},
+		{"Z list on the SS form", "V4FMADDSS",
+			[]Operand{Ptr(AX, 0, 8), list("Z0", "Z3"), vreg(t, "K5"), vreg(t, "X22")}},
+		{"Y destination", "V4FMADDPS",
+			[]Operand{Ptr(sp, 0, 8), list("Z0", "Z3"), vreg(t, "K2"), vreg(t, "Y0")}},
+		{"register source", "V4FMADDPS",
+			[]Operand{vreg(t, "Z1"), list("Z0", "Z3"), vreg(t, "K2"), vreg(t, "Z0")}},
+		{"non-mask third operand", "V4FMADDPS",
+			[]Operand{Ptr(sp, 0, 8), list("Z0", "Z3"), vreg(t, "Z4"), vreg(t, "Z0")}},
+		{"k0 mask", "V4FMADDPS",
+			[]Operand{Ptr(sp, 0, 8), list("Z0", "Z3"), vreg(t, "K0"), vreg(t, "Z0")}},
+		{"K after the destination", "V4FMADDPS",
+			[]Operand{Ptr(sp, 0, 8), list("Z0", "Z3"), vreg(t, "Z0"), vreg(t, "K2")}},
+		{"zeroing without a mask", "V4FMADDPS.Z",
+			[]Operand{Ptr(sp, 0, 8), list("Z0", "Z3"), vreg(t, "Z0")}},
+		{"SAE suffix", "V4FMADDPS.SAE",
+			[]Operand{Ptr(sp, 0, 8), list("Z0", "Z3"), vreg(t, "K2"), vreg(t, "Z0")}},
+		{"high index source", "VP4DPWSSD",
+			[]Operand{Idx(DI, vreg(t, "X16"), 1, 0, 8), list("Z0", "Z3"), vreg(t, "K2"), vreg(t, "Z0")}},
+		{"short operand list", "V4FMADDPS",
+			[]Operand{Ptr(sp, 0, 8), list("Z0", "Z3")}},
+	}
+	for _, c := range cases {
+		if _, err := Encode(c.mnem, c.ops...); err == nil {
+			t.Errorf("%s: expected an error, got none", c.name)
+		}
+	}
+}

asm/goobj.go

@@ -14,8 +14,8 @@ import (
 	"sync"
 )
 
-// This file emits GOOBJ — the Go toolchain's object format, which cmd/link
-// consumes directly — so gasm-assembled functions drop into a go build
+// This file emits GOOBJ, the Go toolchain's object format, which cmd/link
+// consumes directly, so gasm-assembled functions drop into a go build
 // without the Go assembler.  The layout follows cmd/internal/goobj: a
 // toolchain preamble ("go object ...\n!\n"), the go120ld header with its
 // block offsets, a string table, symbol definitions, the relocation /
@@ -33,7 +33,10 @@ import (
 // methods supply the toolchain preamble, the MinLC (pc-value delta unit)
 // and the relocation-type mapping for code relocations.
 
-// GOOBJ block indices (cmd/internal/goobj).
+// GOOBJ block indices (cmd/internal/goobj).  These MUST match the real
+// archive layout: the emitter writes the header offsets per index and the
+// reader (groundtruth, goobj_resolve) parses real Go archives with them.
+// blkAutolib is unused by the emitter but still defines index 0.
 const (
 	blkAutolib = iota
 	blkPkgIdx
@@ -65,11 +68,12 @@ const (
 	kindSDWARFLINES = 20
 )
 
-// Symbol flags (cmd/internal/goobj).
+// Symbol flags (cmd/internal/goobj).  The linkname flag is set only for
+// //go:linkname symbols (and main.main); ordinary assembly symbols carry
+// none, matching cmd/asm's output.
 const (
 	symFlagDupok   = 0x01
 	symFlagNoSplit = 0x10
-	symFlag2Link   = 0x10 // asm objects flag every named symbol as linkname
 	symABIStatic   = 0xffff
 )
 
@@ -91,10 +95,12 @@ const (
 )
 
 // Relocation types (cmd/internal/objabi).
-// R_PCREL and R_ADDR are stable across Go versions.
+// R_ADDR, R_CALL, R_PCREL and R_TLS_LE are stable across Go versions.
 const (
-	relocPCRel = 14 // R_PCREL
 	relocAddr  = 1  // R_ADDR
+	relocCall  = 7  // R_CALL
+	relocPCRel = 14 // R_PCREL
+	relocTLSLE = 15 // R_TLS_LE
 )
 
 // relocDWTXTADDRU4 returns the R_DWTXTADDR_U4 relocation type for the
@@ -137,10 +143,30 @@ func isGo127OrLater() bool {
 
 // Special package indices for symbol references.
 const (
-	pkgIdxNone = 0x7fffffff
-	pkgIdxSelf = 0x7ffffffb
+	pkgIdxNone    = 0x7fffffff
+	pkgIdxSelf    = 0x7ffffffb
+	pkgIdxBuiltin = 0x7ffffffc
 )
 
+// goobjBuiltinMorestackNoctxt is the index of runtime.morestack_noctxt in
+// cmd/internal/goobj/builtinlist.go of the toolchain the object targets
+// (246 since Go 1.25; the list is append-only).
+const goobjBuiltinMorestackNoctxt = 246
+
+// goobjBuiltinMorestack is the builtin reference the toolchain emits for the
+// stack-guard call.
+var goobjBuiltinMorestack = "runtime\u00b7morestack_noctxt"
+
+// isCallReloc reports whether k is one of the per-arch call relocations a
+// direct branch to a TEXT symbol carries.
+func isCallReloc(k RelocKind) bool {
+	switch k {
+	case RelCall, RelRISCVJal, RelArm64Branch, RelLoong64Branch:
+		return true
+	}
+	return false
+}
+
 const goobjMagic = "\x00go120ld"
 
 // goSym is one symbol definition under construction.
@@ -175,14 +201,24 @@ type dwarfRelocSet struct {
 // does with its -p flag).  srcPath names the source file recorded in the
 // object's file table and line tables.  The toolchain's object preamble is
 // captured from the installed go tool asm, so the output links with the
-// toolchain it was produced on — exactly like a real assembly object.
+// toolchain it was produced on, exactly like a real assembly object.
 func (img *Image) GOObject(pkgPath, srcPath string) ([]byte, error) {
 	pre, err := toolchainObjectPreamble()
 	if err != nil {
 		return nil, err
 	}
-	// amd64: MinLC 1, R_PCREL for the code relocations.
-	return img.emitGOObject(pkgPath, srcPath, pre, 1, func(Reloc) (uint16, uint8) { return relocPCRel, 4 })
+	// amd64: MinLC 1, R_PCREL for displacements, R_CALL for calls and
+	// R_TLS_LE for the stack-guard TLS load.
+	return img.emitGOObject(pkgPath, srcPath, pre, 1, func(r Reloc) (uint16, uint8) {
+		switch r.Kind {
+		case RelCall:
+			return relocCall, 4
+		case RelTLSLE:
+			return relocTLSLE, 4
+		default:
+			return relocPCRel, 4
+		}
+	})
 }
 
 // emitGOObject assembles the GOOBJ payload for any architecture.  pre is
@@ -195,7 +231,7 @@ func (img *Image) emitGOObject(pkgPath, srcPath string, pre []byte, minLC int, r
 		return nil, fmt.Errorf("GOOBJ emission requires a package path (-p)")
 	}
 
-	// The non-package definitions first — the DWARF symbols reference the
+	// The non-package definitions first, the DWARF symbols reference the
 	// functions by these indices: per function the four pc-value tables
 	// and the function itself, as cmd/asm lays them out.
 	type npSym struct {
@@ -244,7 +280,7 @@ func (img *Image) emitGOObject(pkgPath, srcPath string, pre []byte, minLC int, r
 			// their relocations cover whole AUIPC/pcalau12i pairs, so
 			// zeroing r.Off would erase the opcode/register bits the linker
 			// preserves when it patches only the immediate.
-			if r.Kind != RelPCRel32 {
+			if r.Kind != RelPCRel32 && r.Kind != RelCall {
 				continue
 			}
 			if r.Off >= 0 && r.Off+4 <= len(code) {
@@ -252,7 +288,7 @@ func (img *Image) emitGOObject(pkgPath, srcPath string, pre []byte, minLC int, r
 			}
 		}
 		nps = append(nps, npSym{
-			sym:  goSym{name: name, abi: abi, typ: kindSTEXT, flag: flag, flag2: symFlag2Link, size: uint32(fn.Size)},
+			sym:  goSym{name: name, abi: abi, typ: kindSTEXT, flag: flag, size: uint32(fn.Size)},
 			data: code,
 		})
 	}
@@ -282,7 +318,7 @@ func (img *Image) emitGOObject(pkgPath, srcPath string, pre []byte, minLC int, r
 			abi = symABIStatic
 		}
 		defIdx[d.Name] = len(defs)
-		defs = append(defs, goSym{name: name, abi: abi, typ: typ, flag: flag, flag2: symFlag2Link, size: uint32(d.Size)})
+		defs = append(defs, goSym{name: name, abi: abi, typ: typ, flag: flag, size: uint32(d.Size)})
 		defData = append(defData, img.Data[d.Offset:d.Offset+d.Size])
 	}
 	fnFiIdx := make([]int, len(img.Funcs))
@@ -317,6 +353,13 @@ func (img *Image) emitGOObject(pkgPath, srcPath string, pre []byte, minLC int, r
 		)
 	}
 
+	// Index the non-package TEXT definitions by short name for the internal
+	// call references.
+	textNpIdx := map[string]int{}
+	for i, fn := range img.Funcs {
+		textNpIdx[fn.Name] = fnNpIdx[i]
+	}
+
 	// Resolve external symbol references (cross-package).  Build the
 	// package index table and determine each external symbol's SymIdx
 	// by reading the target package's export data.
@@ -324,10 +367,20 @@ func (img *Image) emitGOObject(pkgPath, srcPath string, pre []byte, minLC int, r
 	var extPkgIdx map[string]int
 	var extSymIdx map[string]int
 	if len(img.Externals) > 0 {
-		var err error
-		extPkgTable, extPkgIdx, extSymIdx, err = resolveExternalSymbols(img.Externals)
-		if err != nil {
-			return nil, fmt.Errorf("GOOBJ emission: resolving external symbols: %w", err)
+		// The morestack call is a builtin reference, not a resolved external.
+		var need []string
+		for _, n := range img.Externals {
+			if n == goobjBuiltinMorestack {
+				continue
+			}
+			need = append(need, n)
+		}
+		if len(need) > 0 {
+			var err error
+			extPkgTable, extPkgIdx, extSymIdx, err = resolveExternalSymbols(need)
+			if err != nil {
+				return nil, fmt.Errorf("GOOBJ emission: resolving external symbols: %w", err)
+			}
 		}
 	}
 
@@ -339,6 +392,31 @@ func (img *Image) emitGOObject(pkgPath, srcPath string, pre []byte, minLC int, r
 		si := len(defs) + fnNpIdx[i]
 		for _, r := range fn.Relocs {
 			typ, size := relocField(r)
+			if r.Kind == RelTLSLE {
+				// The TLS load has no symbol: {0, 0} is the nil ref.
+				var rec [23]byte
+				binary.LittleEndian.PutUint32(rec[0:], uint32(int32(r.Off)))
+				rec[4] = size
+				binary.LittleEndian.PutUint16(rec[5:], typ)
+				binary.LittleEndian.PutUint64(rec[7:], uint64(r.Addend))
+				binary.LittleEndian.PutUint32(rec[15:], 0)
+				binary.LittleEndian.PutUint32(rec[19:], 0)
+				symRelocs[si] = append(symRelocs[si], rec[:]...)
+				continue
+			}
+			if r.External && r.Name == goobjBuiltinMorestack {
+				// The stack-guard morestack call uses the toolchain's
+				// builtin reference.
+				var rec [23]byte
+				binary.LittleEndian.PutUint32(rec[0:], uint32(int32(r.Off)))
+				rec[4] = size
+				binary.LittleEndian.PutUint16(rec[5:], typ)
+				binary.LittleEndian.PutUint64(rec[7:], uint64(r.Addend))
+				binary.LittleEndian.PutUint32(rec[15:], pkgIdxBuiltin)
+				binary.LittleEndian.PutUint32(rec[19:], goobjBuiltinMorestackNoctxt)
+				symRelocs[si] = append(symRelocs[si], rec[:]...)
+				continue
+			}
 			if r.External {
 				// Split package-qualified name: "runtime·morestack" → runtime, morestack.
 				pkg, name := splitQualified(r.Name)
@@ -363,20 +441,83 @@ func (img *Image) emitGOObject(pkgPath, srcPath string, pre []byte, minLC int, r
 				symRelocs[si] = append(symRelocs[si], rec[:]...)
 				continue
 			}
+			pkg := uint32(pkgIdxSelf)
 			di, ok := defIdx[r.Name]
 			if !ok {
-				return nil, fmt.Errorf("GOOBJ emission: reference to unknown symbol %q", r.Name)
+				// A call to a TEXT function of the same file references the
+				// non-package definition table.
+				ni, isText := textNpIdx[r.Name]
+				if !isText || !isCallReloc(r.Kind) {
+					return nil, fmt.Errorf("GOOBJ emission: reference to unknown symbol %q", r.Name)
+				}
+				pkg = pkgIdxNone
+				di = ni
 			}
 			var rec [23]byte
 			binary.LittleEndian.PutUint32(rec[0:], uint32(int32(r.Off)))
 			rec[4] = size // field width
 			binary.LittleEndian.PutUint16(rec[5:], typ)
 			binary.LittleEndian.PutUint64(rec[7:], uint64(r.Addend))
-			binary.LittleEndian.PutUint32(rec[15:], pkgIdxSelf)
+			binary.LittleEndian.PutUint32(rec[15:], pkg)
 			binary.LittleEndian.PutUint32(rec[19:], uint32(di))
 			symRelocs[si] = append(symRelocs[si], rec[:]...)
 		}
 	}
+	// The data symbols' own relocations: the symbol-valued DATA fields
+	// ("DATA s+0(SB)/8, $other(SB)").  The toolchain patches each field
+	// with the target's absolute address through an R_ADDR of the DATA
+	// line's width, on every architecture (the code relocations are
+	// per-architecture PC-relative shapes; a data pointer word is not), so
+	// this mapping bypasses relocField.  The definitions were appended in
+	// DataSyms order, so data symbol i is definition index i.
+	for i, d := range img.DataSyms {
+		for _, r := range d.Relocs {
+			if r.Kind != RelAddr {
+				return nil, fmt.Errorf("GOOBJ emission: data symbol %q carries a non-data relocation", d.Name)
+			}
+			var rec [23]byte
+			binary.LittleEndian.PutUint32(rec[0:], uint32(int32(r.Off)))
+			rec[4] = r.Siz
+			binary.LittleEndian.PutUint16(rec[5:], relocAddr)
+			binary.LittleEndian.PutUint64(rec[7:], uint64(r.Addend))
+			switch {
+			case r.External && r.Name == goobjBuiltinMorestack:
+				binary.LittleEndian.PutUint32(rec[15:], pkgIdxBuiltin)
+				binary.LittleEndian.PutUint32(rec[19:], goobjBuiltinMorestackNoctxt)
+			case r.External:
+				pkg, name := splitQualified(r.Name)
+				if pkg == "" {
+					return nil, fmt.Errorf("GOOBJ emission: external symbol %q has no package prefix", r.Name)
+				}
+				pIdx, ok := extPkgIdx[pkg]
+				if !ok {
+					return nil, fmt.Errorf("GOOBJ emission: package %q not resolved", pkg)
+				}
+				sIdx, ok := extSymIdx[pkg+"·"+name]
+				if !ok {
+					return nil, fmt.Errorf("GOOBJ emission: symbol %s·%s not resolved", pkg, name)
+				}
+				binary.LittleEndian.PutUint32(rec[15:], uint32(pIdx))
+				binary.LittleEndian.PutUint32(rec[19:], uint32(sIdx))
+			default:
+				if di, ok := defIdx[r.Name]; ok {
+					binary.LittleEndian.PutUint32(rec[15:], pkgIdxSelf)
+					binary.LittleEndian.PutUint32(rec[19:], uint32(di))
+					break
+				}
+				// A DATA field may hold the address of a TEXT function of
+				// the same file (the rt0 lib entry spelling), which is a
+				// non-package definition.
+				ni, isText := textNpIdx[r.Name]
+				if !isText {
+					return nil, fmt.Errorf("GOOBJ emission: reference to unknown symbol %q", r.Name)
+				}
+				binary.LittleEndian.PutUint32(rec[15:], pkgIdxNone)
+				binary.LittleEndian.PutUint32(rec[19:], uint32(ni))
+			}
+			symRelocs[i] = append(symRelocs[i], rec[:]...)
+		}
+	}
 	// The DWARF symbols' own relocations (the function address references).
 	for _, ds := range dwarfRelocs {
 		for _, r := range ds.relocs {
@@ -417,7 +558,7 @@ func (img *Image) emitGOObject(pkgPath, srcPath string, pre []byte, minLC int, r
 	// The string table.  Absolute offsets: it starts right after the
 	// 96-byte header (magic, fingerprint, flags, the 19 block offsets).
 	const headerSize = 8 + 8 + 4 + 4*(blkEnd+1)
-	strTab := []byte{}
+	var strTab []byte
 	strOff := map[string]uint32{}
 	addStr := func(s string) {
 		if _, ok := strOff[s]; ok {
@@ -464,7 +605,7 @@ func (img *Image) emitGOObject(pkgPath, srcPath string, pre []byte, minLC int, r
 	auxIdxBlk := make([]byte, 0, 4*(nsyms+1))
 	dataIdxBlk := make([]byte, 0, 4*(nsyms+1))
 	var nr, na, nd uint32
-	for si := 0; si < nsyms; si++ {
+	for si := range nsyms {
 		relocIdxBlk = binary.LittleEndian.AppendUint32(relocIdxBlk, nr)
 		auxIdxBlk = binary.LittleEndian.AppendUint32(auxIdxBlk, na)
 		dataIdxBlk = binary.LittleEndian.AppendUint32(dataIdxBlk, nd)
@@ -505,7 +646,7 @@ func (img *Image) emitGOObject(pkgPath, srcPath string, pre []byte, minLC int, r
 	// The fingerprint stays zero, as cmd/asm leaves it.
 	binary.LittleEndian.PutUint32(payload[16:], 4) // ObjFlagFromAssembly
 	off := uint32(headerSize + len(strTab))
-	for i := 0; i < blkEnd; i++ {
+	for i := range blkEnd {
 		binary.LittleEndian.PutUint32(payload[20+4*i:], off)
 		off += uint32(len(blocks[i]))
 	}

asm/goobj_dwarf.go

@@ -139,10 +139,7 @@ func dwSelectOpcode(deltaPC uint64, deltaLC int64) int64 {
 		return int64(dwOpcodeBase) + (deltaLC - dwLineBase) + dwLineRange*int64(deltaPC)
 	default:
 		if deltaPC <= uint64(dwPCRange) {
-			op := int64(dwOpcodeBase) + (dwLineRange - 1) + dwLineRange*int64(deltaPC)
-			if op > 255 {
-				op = 255
-			}
+			op := min(int64(dwOpcodeBase)+(dwLineRange-1)+dwLineRange*int64(deltaPC), 255)
 			return op
 		}
 		switch deltaPC - uint64(dwPCRange) {

asm/goobj_resolve.go

@@ -12,24 +12,6 @@ import (
 	"strings"
 )
 
-// readGOOBJSymbols reads the GOOBJ symbol definitions from a compiled Go
-// package's export file.  The file is an ar archive containing a __.PKGDEF
-// member whose payload is the "go object ...\n!\n" preamble followed by the
-// GOOBJ data.  The function returns the symbol names in definition order
-// (the order they appear in blkSymdef), which matches the SymIdx the linker
-// expects for cross-package references.
-func readGOOBJSymbols(exportPath string) ([]string, error) {
-	data, err := os.ReadFile(exportPath)
-	if err != nil {
-		return nil, err
-	}
-	goobj, err := extractGOOBJ(data)
-	if err != nil {
-		return nil, fmt.Errorf("%s: %w", exportPath, err)
-	}
-	return goobj.symbols(), nil
-}
-
 // exportPath returns the export file path for a given import path by running
 // "go list -export".  The result is cached so repeated calls for the same
 // package are fast.
@@ -58,8 +40,11 @@ func exportPath(importPath string) (string, error) {
 //
 // refs maps package import paths to the symbol names referenced from that
 // package.  The returned pkgIdx maps each import path to its position in
-// the blkPkgIdx table (0-based), and symIdx gives each symbol's index within
-// its package.
+// the blkPkgIdx table, which reserves index 0 for the dummy invalid
+// package (cmd/internal/obj/sym.go: "0 is invalid index"; the loader's
+// reader loop starts at 1), so package i sits at block index i+1 and its
+// relocations carry i+1.  symIdx gives each symbol's index within its
+// package.
 func resolveExternalGOOBJ(refs map[string][]string) (pkgIdx map[string]int, symIdx map[string]int, err error) {
 	pkgIdx = make(map[string]int, len(refs))
 	symIdx = make(map[string]int)
@@ -68,7 +53,9 @@ func resolveExternalGOOBJ(refs map[string][]string) (pkgIdx map[string]int, symI
 	packages := sortedPkgRefs(refs)
 
 	for i, pkg := range packages {
-		pkgIdx[pkg.path] = i
+		// Block index 0 is the dummy invalid package; the first real
+		// package starts at 1.
+		pkgIdx[pkg.path] = i + 1
 		exp, err := exportPath(pkg.path)
 		if err != nil {
 			return nil, nil, err
@@ -102,7 +89,7 @@ func sortedPkgRefs(refs map[string][]string) []pkgRef {
 	for pkg, syms := range refs {
 		pkgs = append(pkgs, pkgRef{pkg, syms})
 	}
-	// Simple insertion sort — the list is tiny (usually 1–3 packages).
+	// Simple insertion sort, the list is tiny (usually 1-3 packages).
 	for i := 1; i < len(pkgs); i++ {
 		for j := i; j > 0 && pkgs[j-1].path > pkgs[j].path; j-- {
 			pkgs[j-1], pkgs[j] = pkgs[j], pkgs[j-1]
@@ -163,40 +150,65 @@ func parseArDecimal(b []byte) int {
 }
 
 // goobjFile is a parsed GOOBJ file: the string table and the symbol-definition
-// block.
+// blocks.  The hashed blocks are kept raw: their symbols carry no names, only
+// the loader needs their counts.
 type goobjFile struct {
-	strTab []byte // string table, at headerSize + n
-	symdef []byte // blkSymdef raw block
-	npdef  []byte // blkNonpkgdef raw block
+	strTab   []byte // string table, at headerSize + n
+	symdef   []byte // blkSymdef raw block
+	hashed64 []byte // blkHashed64def raw block
+	hashed   []byte // blkHasheddef raw block
+	npdef    []byte // blkNonpkgdef raw block
 }
 
-// symbols returns all symbol names in definition order by scanning the
-// symdef and nonpkgdef blocks and resolving each name through the string
-// table.  Package definitions (blkSymdef) use fully-qualified names like
-// "runtime.morestack"; non-package definitions (blkNonpkgdef) use bare
-// names like "morestack".  This combined list matches the index the
-// linker expects for cross-package references.
+// loaderIndexBase returns the index the first nonpkgdef symbol occupies in the
+// loader's per-object symbol array.  cmd/link lays the definition blocks out as
+// symdef, hashed64def, hasheddef, nonpkgdef, nonpkgref (loader.go: preloadSyms
+// fills r.syms in exactly that order, and resolve() indexes PkgIdxNone and
+// cross-package SymIdx into it), so a symbol found in blkNonpkgdef carries the
+// three leading blocks' symbol counts as its base.
+func (f *goobjFile) loaderIndexBase() int {
+	return len(f.symdef)/recSymSize + len(f.hashed64)/recSymSize + len(f.hashed)/recSymSize
+}
+
+// symbols returns the names of the symdef and nonpkgdef blocks in
+// definition order.  Package definitions (blkSymdef) use fully-qualified
+// names like "runtime.morestack"; non-package definitions (blkNonpkgdef)
+// use bare names like "morestack".  For lookups by index prefer
+// findSymbol: it adds the hashed blocks' count the loader's array
+// interleaves between the two.
 func (f *goobjFile) symbols() []string {
 	return append(f.defNames(), f.npdefNames()...)
 }
 
-// findSymbol returns the index of a symbol within the combined symbol list,
-// or -1 if not found.  It first tries the fully-qualified name (pkg.name),
-// then the bare name.
+// findSymbol returns the index of a symbol within the loader's per-object
+// symbol array, or -1 if not found.  It first tries the fully-qualified
+// name (pkg.name), then the bare name (assembly objects store dotless
+// names, e.g. runtime's "gogo", for symbols other packages reach through
+// a linkname).
 func (f *goobjFile) findSymbol(pkg, name string) int {
+	base := f.loaderIndexBase()
 	qualified := pkg + "." + name
-	syms := f.symbols()
-	for i, s := range syms {
+	for i, s := range f.defNames() {
 		if s == qualified {
 			return i
 		}
 	}
-	// Try bare name (for non-package definitions).
-	for i, s := range syms {
+	for i, s := range f.npdefNames() {
+		if s == qualified {
+			return base + i
+		}
+	}
+	// Try bare name (for dotless assembly definitions).
+	for i, s := range f.defNames() {
 		if s == name {
 			return i
 		}
 	}
+	for i, s := range f.npdefNames() {
+		if s == name {
+			return base + i
+		}
+	}
 	return -1
 }
 
@@ -210,18 +222,22 @@ func (f *goobjFile) npdefNames() []string {
 	return f.readSymNames(f.npdef)
 }
 
+// recSymSize is the size of one Sym record in the definition blocks
+// (goobj.SymSize: stringRefSize + 2 + 1 + 1 + 1 + 4 + 4).
+const recSymSize = 21
+
 // readSymNames reads symbol names from a symdef/nonpkgdef block.  Each record
 // is 21 bytes: nameLen (u32), nameOff (u32), abi (u16), typ, flag, flag2,
 // size (u32), align (u32).  nameOff is an absolute offset into the string
 // table.
 func (f *goobjFile) readSymNames(block []byte) []string {
-	const recSize = 21
+	const recSize = recSymSize
 	if len(block) < recSize {
 		return nil
 	}
 	n := len(block) / recSize
 	names := make([]string, 0, n)
-	for i := 0; i < n; i++ {
+	for i := range n {
 		rec := block[i*recSize : (i+1)*recSize]
 		nameLen := binary.LittleEndian.Uint32(rec[0:4])
 		nameOff := binary.LittleEndian.Uint32(rec[4:8])
@@ -265,16 +281,18 @@ func parseGOOBJ(data []byte) (*goobjFile, error) {
 	//   [16:20] flags
 	//   [20:96] 19 × uint32 offsets
 	var offs [blkEnd + 1]uint32
-	for i := 0; i <= blkEnd; i++ {
+	for i := range blkEnd + 1 {
 		offs[i] = binary.LittleEndian.Uint32(payload[20+4*i:])
 	}
 	// The string table lives at headerSize.
 	strTabStart := uint32(goobjHeaderSize)
 
 	f := &goobjFile{
-		strTab: payload[strTabStart:offs[0]],
-		symdef: blockSlice(payload, offs, blkSymdef, blkSymdef+1),
-		npdef:  blockSlice(payload, offs, blkNonpkgdef, blkNonpkgdef+1),
+		strTab:   payload[strTabStart:offs[0]],
+		symdef:   blockSlice(payload, offs, blkSymdef, blkSymdef+1),
+		hashed64: blockSlice(payload, offs, blkHashed64def, blkHashed64def+1),
+		hashed:   blockSlice(payload, offs, blkHasheddef, blkHasheddef+1),
+		npdef:    blockSlice(payload, offs, blkNonpkgdef, blkNonpkgdef+1),
 	}
 	return f, nil
 }
@@ -324,25 +342,29 @@ func resolveExternalSymbols(externals []string) (pkgTable []string, pkgIdxMap ma
 		return nil, nil, nil, err
 	}
 
-	// Build the package table in pkgIdx order.
+	// Build the package table in pkgIdx order.  The indices are 1-based
+	// (0 is the dummy invalid package, written by the emitter itself), so
+	// the table without the dummy is indexed one below.
 	pkgTable = make([]string, len(pkgIdx1))
 	for pkg, idx := range pkgIdx1 {
-		pkgTable[idx] = pkg
+		pkgTable[idx-1] = pkg
 	}
 
 	return pkgTable, pkgIdx1, symIdx1, nil
 }
 
 // splitQualified splits a qualified Go symbol name (pkgpath·name) into its
-// package path and local name.  The separator is the middle dot (U+00B7).
-// If no separator is found, the symbol is assumed to be in the current
-// package (empty pkg).
+// package path and local name.  The separator is the middle dot (U+00B7),
+// whose UTF-8 encoding is two bytes, so the search must be string-based:
+// IndexByte would match only the second byte and leave the lead byte on
+// the package path.  If no separator is found, the symbol is assumed to be
+// in the current package (empty pkg).
 func splitQualified(full string) (pkg, name string) {
-	if idx := strings.IndexByte(full, '\u00b7'); idx >= 0 {
-		return full[:idx], full[idx+len("\u00b7"):]
+	if before, after, ok := strings.Cut(full, "\u00b7"); ok {
+		return before, after
 	}
-	if idx := strings.IndexByte(full, '.'); idx >= 0 {
-		return full[:idx], full[idx+1:]
+	if before, after, ok := strings.Cut(full, "."); ok {
+		return before, after
 	}
 	return "", full
 }

asm/goobj_resolve_test.go

@@ -56,8 +56,12 @@ func TestResolveExternalSymbols(t *testing.T) {
 	if err != nil {
 		t.Fatalf("resolveExternalGOOBJ: %v", err)
 	}
-	if len(pkgIdx) != 1 || pkgIdx["runtime"] != 0 {
-		t.Errorf("pkgIdx = %v, want runtime→0", pkgIdx)
+	if len(pkgIdx) != 1 || pkgIdx["runtime"] != 1 {
+		// Index 0 is the dummy invalid package in the blkPkgIdx table;
+		// the loader's reader loop starts at 1 (cmd/link/internal/
+		// loader/loader.go: "PkgIdx 0 is a dummy invalid package"), so
+		// the first real package must carry index 1.
+		t.Errorf("pkgIdx = %v, want runtime→1", pkgIdx)
 	}
 	if _, ok := symIdx["runtime·g0"]; !ok {
 		t.Errorf("symIdx missing runtime·g0, got %v", symIdx)

asm/goobj_test.go

@@ -12,7 +12,7 @@ import (
 	"strings"
 	"testing"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
 )
 
 // goobjView is a minimal parsed view of a GOOBJ payload, enough to check
@@ -117,7 +117,9 @@ DATA mask<>+8(SB)/8, $0x800f0e0d0c0b0a09
 	if len(defs) != 7 {
 		t.Fatalf("symdefs = %d, want 7", len(defs))
 	}
-	if defs[0].name != "mask" || defs[0].abi != 0xffff || defs[0].typ != kindSRODATA || defs[0].size != 16 || defs[0].flag2 != symFlag2Link {
+	// The linkname flag stays clear: the toolchain sets it only for
+	// //go:linkname symbols, and an ordinary static GLOBL is not one.
+	if defs[0].name != "mask" || defs[0].abi != 0xffff || defs[0].typ != kindSRODATA || defs[0].size != 16 || defs[0].flag2 != 0 {
 		t.Errorf("mask symbol = %+v", defs[0])
 	}
 	if defs[1].name != "" || defs[1].typ != kindSDATA || defs[1].size != 28 {
@@ -158,8 +160,8 @@ DATA mask<>+8(SB)/8, $0x800f0e0d0c0b0a09
 		t.Errorf("funcinfo bytes %x", fi)
 	}
 
-	// The pc-value tables of addq (non-package indices 0–3, so global
-	// indices 7–10): pcsp a flat zero over the whole function, pcinline a
+	// The pc-value tables of addq (non-package indices 0-3, so global
+	// indices 7-10): pcsp a flat zero over the whole function, pcinline a
 	// flat -1, both with the pc delta in MinLC (1) units.
 	pcsp := data[le.Uint32(didx[4*7:]):]
 	if got := pcsp[:3]; !bytes.Equal(got, []byte{0x02, 19, 0x00}) {
@@ -294,7 +296,7 @@ TEXT ·framed(SB), NOSPLIT, $8-0
 	}
 	for i := range wantPCs {
 		if pcs[i] != wantPCs[i] || vals[i] != wantVals[i] {
-			t.Errorf("pcsp[%d] = (%d,%d), want (%d,%d) — all: %v %v", i, pcs[i], vals[i], wantPCs[i], wantVals[i], pcs, vals)
+			t.Errorf("pcsp[%d] = (%d,%d), want (%d,%d); all: %v %v", i, pcs[i], vals[i], wantPCs[i], wantVals[i], pcs, vals)
 		}
 	}
 	// The last two steps unwind the epilogue to zero.
@@ -333,7 +335,7 @@ TEXT ·useext(SB), NOSPLIT, $0-8
 
 // TestGOObjectLinkAndRun is the end-to-end check: assemble the test
 // functions to a GOOBJ, swap it into a go build in place of the toolchain's
-// assembly object, link, and run — the output must match the baseline
+// assembly object, link, and run; the output must match the baseline
 // binary the Go assembler produced.  Skipped when no Go toolchain is
 // available.
 func TestGOObjectLinkAndRun(t *testing.T) {
@@ -393,7 +395,7 @@ func main() {
 	}
 	var work string
 	var asmObj, pkgArch, linkLine string
-	for _, line := range strings.Split(string(buildLog), "\n") {
+	for line := range strings.SplitSeq(string(buildLog), "\n") {
 		switch {
 		case strings.HasPrefix(line, "WORK="):
 			work = strings.TrimPrefix(line, "WORK=")
@@ -461,7 +463,7 @@ func main() {
 	newArch := filepath.Join(dir, "pkg.a")
 	args := []string{"tool", "pack", "c", newArch}
 	seen := map[string]bool{}
-	for _, m := range strings.Fields(string(listOut)) {
+	for m := range strings.FieldsSeq(string(listOut)) {
 		if seen[m] {
 			continue
 		}
@@ -511,3 +513,188 @@ func fieldAfter(line, flag string) string {
 	}
 	return ""
 }
+
+// TestGOObjectExternalPackageLink is the cross-package end-to-end check: a
+// GOOBJ whose code references a real external package symbol (runtime's
+// morestack, a plain reference rather than the builtin noctxt form) must
+// carry a package index that points past the blkPkgIdx table's dummy entry
+// 0, and the object must link against the real runtime.  Pre-fix, the
+// relocations carried block index 0, which the loader never fills, so the
+// reference resolved against whatever object was loaded first and the link
+// failed.  The binary is not run: morestack returns to the call site's
+// stack check, which a hand-written caller has none of.
+func TestGOObjectExternalPackageLink(t *testing.T) {
+	goBin, err := exec.LookPath("go")
+	if err != nil {
+		t.Skip("no Go toolchain available")
+	}
+	dir := t.TempDir()
+
+	const asmSrc = `
+#include "textflag.h"
+
+TEXT ·fn(SB), NOSPLIT, $0-0
+	CALL ·helper(SB)
+	RET
+
+TEXT ·helper(SB), NOSPLIT, $0-0
+	RET
+`
+	const mainSrc = `package main
+
+func fn()
+func helper()
+
+func main() {
+	fn()
+	helper()
+}
+`
+	if err := os.WriteFile(filepath.Join(dir, "main_amd64.s"), []byte(asmSrc), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(dir, "main.go"), []byte(mainSrc), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(dir, "go.mod"), []byte("module extlink\n\ngo 1.27\n"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	// Capture the build the toolchain performs and re-run only its link
+	// step with our object swapped into the package archive, mirroring
+	// TestGOObjectLinkAndRun.
+	build := exec.Command(goBin, "build", "-x", "-work", "-o", filepath.Join(dir, "prog"), ".")
+	build.Dir = dir
+	buildLog, err := build.CombinedOutput()
+	if err != nil {
+		t.Fatalf("baseline build: %v\n%s", err, buildLog)
+	}
+	var work, linkLine, asmObj, pkgArch string
+	for line := range strings.SplitSeq(string(buildLog), "\n") {
+		switch {
+		case strings.HasPrefix(line, "WORK="):
+			work = strings.TrimPrefix(line, "WORK=")
+		case strings.Contains(line, "/asm ") && strings.Contains(line, "main_amd64.s") && !strings.Contains(line, "-gensymabis"):
+			asmObj = fieldAfter(line, "-o")
+		case strings.Contains(line, "pack r") && strings.Contains(line, "_pkg_.a"):
+			pkgArch = strings.TrimSpace(strings.SplitN(line, "pack r", 2)[1])
+			pkgArch = strings.Fields(strings.SplitN(pkgArch, "#", 2)[0])[0]
+		case strings.Contains(line, "/link ") && strings.Contains(line, "-importcfg"):
+			linkLine = line
+		}
+	}
+	if work == "" || asmObj == "" || pkgArch == "" || linkLine == "" {
+		t.Skipf("could not parse build log (work=%q asmObj=%q)", work, asmObj)
+	}
+	defer os.RemoveAll(work)
+	asmObj = strings.ReplaceAll(asmObj, "$WORK", work)
+	pkgArch = strings.ReplaceAll(pkgArch, "$WORK", work)
+
+	// Assemble the source with gasm, then retarget fn's internal call at
+	// a real external package symbol: the reloc's qualified name drives
+	// the export-data resolution the way a source-level runtime·sym(SB)
+	// reference would.
+	f, errs := parser.Parse("main_amd64.s", asmSrc)
+	if len(errs) > 0 {
+		t.Fatalf("parse: %v", errs)
+	}
+	img, err := AssembleFile(f)
+	if err != nil {
+		t.Fatalf("AssembleFile: %v", err)
+	}
+	fn := &img.Funcs[0]
+	for i := range fn.Relocs {
+		fn.Relocs[i].Name = "runtime\u00b7morestack"
+		fn.Relocs[i].External = true
+	}
+	img.Externals = []string{"runtime\u00b7morestack"}
+	obj, err := img.GOObject("main", "main_amd64.s")
+	if err != nil {
+		t.Fatalf("GOObject: %v", err)
+	}
+
+	// Structural check: the blkPkgIdx block reserves entry 0 for the
+	// dummy invalid package and places runtime at entry 1, and fn's call
+	// relocation carries PkgIdx 1.
+	v := openGoobj(t, obj)
+	pkgBlk := v.blk(blkPkgIdx)
+	if len(pkgBlk) != 2*8 {
+		t.Fatalf("blkPkgIdx = %d bytes, want two entries", len(pkgBlk))
+	}
+	le := binary.LittleEndian
+	strEntry := func(i int) string {
+		e := pkgBlk[i*8 : (i+1)*8]
+		return v.str(le.Uint32(e[4:]), le.Uint32(e[0:]))
+	}
+	if s := strEntry(0); s != "" {
+		t.Errorf("blkPkgIdx[0] = %q, want the dummy empty package", s)
+	}
+	if s := strEntry(1); s != "runtime" {
+		t.Errorf("blkPkgIdx[1] = %q, want runtime", s)
+	}
+	relocs := v.blk(blkReloc)
+	// fn is the last non-package symbol (two functions, four pc tables
+	// each); its one reloc is the final record.
+	fnRec := relocs[len(relocs)-23:]
+	if pIdx := le.Uint32(fnRec[15:]); pIdx != 1 {
+		t.Errorf("external reloc PkgIdx = %d, want 1 (runtime)", pIdx)
+	}
+
+	// Swap the object into the package archive and link with cmd/link;
+	// the link line consumes the archive, not the loose object file.
+	membersDir := filepath.Join(dir, "members")
+	if err := os.MkdirAll(membersDir, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	extract := exec.Command(goBin, "tool", "pack", "x", pkgArch)
+	extract.Dir = membersDir
+	if out, err := extract.CombinedOutput(); err != nil {
+		t.Fatalf("pack x: %v\n%s", err, out)
+	}
+	member := filepath.Join(membersDir, filepath.Base(asmObj))
+	if err := os.Chmod(member, 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(member, obj, 0o644); err != nil {
+		t.Fatal(err)
+	}
+	listCmd := exec.Command(goBin, "tool", "pack", "t", pkgArch)
+	listOut, err := listCmd.CombinedOutput()
+	if err != nil {
+		t.Fatalf("pack t: %v\n%s", err, listOut)
+	}
+	newArch := filepath.Join(dir, "pkg.a")
+	args := []string{"tool", "pack", "c", newArch}
+	seen := map[string]bool{}
+	for m := range strings.FieldsSeq(string(listOut)) {
+		if seen[m] {
+			continue
+		}
+		seen[m] = true
+		if err := os.Chmod(filepath.Join(membersDir, m), 0o644); err != nil {
+			t.Fatal(err)
+		}
+		args = append(args, filepath.Join(membersDir, m))
+	}
+	pack := exec.Command(goBin, args...)
+	pack.Dir = membersDir
+	if out, err := pack.CombinedOutput(); err != nil {
+		t.Fatalf("pack c: %v\n%s", err, out)
+	}
+	linkLine = strings.ReplaceAll(linkLine, "$WORK", work)
+	linkLine = strings.ReplaceAll(linkLine, pkgArch, newArch)
+	linkLine = strings.ReplaceAll(linkLine, filepath.Join(work, "b001", "exe", "a.out"), filepath.Join(dir, "prog2"))
+	linkCmd := exec.Command("sh", "-c", "cd "+dir+" && "+linkLine)
+	if out, err := linkCmd.CombinedOutput(); err != nil {
+		t.Fatalf("re-link with gasm object: %v\n%s", err, out)
+	}
+
+	// The call must have resolved to the real runtime symbol.
+	dump, err := exec.Command(goBin, "tool", "objdump", "-s", "main.fn", filepath.Join(dir, "prog2")).CombinedOutput()
+	if err != nil {
+		t.Fatalf("objdump main.fn: %v\n%s", err, dump)
+	}
+	if !bytes.Contains(dump, []byte("runtime.morestack")) {
+		t.Errorf("main.fn does not call runtime.morestack:\n%s", dump)
+	}
+}

asm/goobjarm64.go

@@ -13,25 +13,54 @@ import (
 )
 
 // GOObjectAARCH64 emits a GOOBJ object file for AArch64.  The layout is
-// the shared one in goobj.go — the toolchain preamble, the go120ld header
+// the shared one in goobj.go, the toolchain preamble, the go120ld header
 // with its block offsets, the string table, the symbol definitions and the
-// reloc/aux/data index arrays — with the arm64 preamble, the MinLC of 4
-// for the pc-value deltas, and R_ADDRARM64 relocation types for the
-// ADRP+ADD/LDR/STR address pairs.
+// reloc/aux/data index arrays, with the arm64 preamble, the MinLC of 4
+// for the pc-value deltas, and the arm64 relocation types for the ADRP
+// pairs and BL calls.
+//
+// The toolchain records one relocation per ADRP pair: a single R_ADDRARM64
+// or R_ARM64_PCREL_LDST64 of Siz 8 at the ADRP word, from which the linker
+// patches both instructions of the pair (cmd/internal/obj/arm64/asm7.go,
+// the ADRP cases: one AddRel with Off at the pair's pc and Siz 8).  gasm's
+// assembler records the ADRP+ADD form as two word relocs, so the second
+// word's twin is dropped here before emission.
 func (img *Image) GOObjectAARCH64(pkgPath, srcPath string) ([]byte, error) {
 	pre, err := toolchainObjectPreambleAARCH64()
 	if err != nil {
 		return nil, err
 	}
-	return img.emitGOObject(pkgPath, srcPath, pre, 4, func(r Reloc) (uint16, uint8) {
-		return relocArm64Addr, 4
+	coalesced := *img
+	coalesced.Funcs = append([]FuncLayout(nil), img.Funcs...)
+	for i := range coalesced.Funcs {
+		rs := coalesced.Funcs[i].Relocs
+		var keep []Reloc
+		for j := 0; j < len(rs); j++ {
+			keep = append(keep, rs[j])
+			if rs[j].Kind == RelArm64Addr && j+1 < len(rs) &&
+				rs[j+1].Kind == RelArm64Addr && rs[j+1].Off == rs[j].Off+4 {
+				j++ // the ADD word's twin: the Siz-8 pair reloc covers it
+			}
+		}
+		coalesced.Funcs[i].Relocs = keep
+	}
+	return coalesced.emitGOObject(pkgPath, srcPath, pre, 4, func(r Reloc) (uint16, uint8) {
+		switch r.Kind {
+		case RelArm64Branch:
+			return relocArm64Branch, 4
+		case RelArm64LDST64:
+			return relocArm64LDST64, 8
+		default:
+			return relocArm64Addr, 8
+		}
 	})
 }
 
-// arm64 relocation types (cmd/internal/objabi).  R_ADDRARM64 resolves an
-// ADRP+ADD/LDR/STR pair to a symbol's address.
+// arm64 relocation types (cmd/internal/objabi).
 const (
-	relocArm64Addr = 9 // R_ADDRARM64
+	relocArm64Addr   = 3  // R_ADDRARM64, ADRP+ADD pair
+	relocArm64Branch = 9  // R_CALLARM64, BL instruction
+	relocArm64LDST64 = 40 // R_ARM64_PCREL_LDST64, ADRP+LDR/STR pair
 )
 
 // toolchainObjectPreambleAARCH64 returns the "go object ...\n!\n" header

asm/goobjloong64.go

@@ -13,9 +13,9 @@ import (
 )
 
 // GOObjectLOONG64 emits a GOOBJ object file for LoongArch.  The layout is
-// the shared one in goobj.go — the toolchain preamble, the go120ld header
+// the shared one in goobj.go, the toolchain preamble, the go120ld header
 // with its block offsets, the string table, the symbol definitions and the
-// reloc/aux/data index arrays — with the loong64 preamble, the MinLC of 4
+// reloc/aux/data index arrays, with the loong64 preamble, the MinLC of 4
 // for the pc-value deltas, and R_LOONG64_ADDR_HI/LO relocation types for
 // the pcalau12i+addi.d address pairs.
 func (img *Image) GOObjectLOONG64(pkgPath, srcPath string) ([]byte, error) {
@@ -25,11 +25,16 @@ func (img *Image) GOObjectLOONG64(pkgPath, srcPath string) ([]byte, error) {
 	}
 	return img.emitGOObject(pkgPath, srcPath, pre, 4, func(r Reloc) (uint16, uint8) {
 		// A pcalau12i+addi.d pair: the high part carries
-		// R_LOONG64_ADDR_HI, the low part R_LOONG64_ADDR_LO.
-		if r.Kind == RelLoong64AddrLo {
+		// R_LOONG64_ADDR_HI, the low part R_LOONG64_ADDR_LO; the guard's
+		// morestack call carries R_CALLLOONG64.
+		switch {
+		case r.Kind == RelLoong64AddrLo:
 			return relocLoong64AddrLo, 4
+		case r.Kind == RelLoong64Branch:
+			return relocCallLoong64, 4
+		default:
+			return relocLoong64AddrHi, 4
 		}
-		return relocLoong64AddrHi, 4
 	})
 }
 
@@ -39,6 +44,7 @@ func (img *Image) GOObjectLOONG64(pkgPath, srcPath string) ([]byte, error) {
 const (
 	relocLoong64AddrHi = 77 // R_LOONG64_ADDR_HI
 	relocLoong64AddrLo = 78 // R_LOONG64_ADDR_LO
+	relocCallLoong64   = 84 // R_CALLLOONG64
 )
 
 // toolchainObjectPreambleLOONG64 returns the "go object ...\n!\n" header

asm/goobjriscv.go

@@ -13,9 +13,9 @@ import (
 )
 
 // GOObjectRISCV emits a GOOBJ object file for RISC-V.  The layout is the
-// shared one in goobj.go — the toolchain preamble, the go120ld header with
+// shared one in goobj.go, the toolchain preamble, the go120ld header with
 // its block offsets, the string table, the symbol definitions and the
-// reloc/aux/data index arrays — with the RISC-V preamble, the MinLC of 2 for
+// reloc/aux/data index arrays, with the RISC-V preamble, the MinLC of 2 for
 // the pc-value deltas, and the single R_RISCV_PCREL_ITYPE/STYPE relocation
 // per AUIPC pair, matching `go tool asm`'s model (each pair is one 8-byte
 // relocation, not the ELF HI20/LO12 pair).

asm/guard_test.go

@@ -0,0 +1,329 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package asm
+
+import (
+	"bytes"
+	"encoding/hex"
+	"strings"
+	"testing"
+
+	"sourcedock.dev/petrbalvin/gasm-sdk/ast"
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
+)
+
+// The expected bytes are pinned from `go tool asm` output (Go 1.27, amd64,
+// verified with go tool objdump): the stack-split guard classes, the morestack
+// block and the auto-NOSPLIT leaf behaviour.
+func TestStackGuardBytes(t *testing.T) {
+	for _, tt := range []struct {
+		name string
+		src  string
+		want string
+	}{
+		{"leafsmall", "TEXT \u00b7leafsmall(SB), $16-0\n\tRET\n",
+			"554889e54883ec104883c4105dc3"},
+		{"leafmed", "TEXT \u00b7leafmed(SB), $256-0\n\tRET\n",
+			"644c8b3425000000004c8da42478ffffff4d3b66107614554889e54881ec000100004881c4000100005dc3e800000000ebce"},
+		{"leafbig", "TEXT \u00b7leafbig(SB), $8192-0\n\tRET\n",
+			"644c8b3425000000004989e44981ec881f0000721a4d3b66107614554889e54881ec002000004881c4002000005dc3e800000000ebca"},
+		// Class 2 with a body long enough that the underflow JB relaxes to
+		// rel32: its displacement must span the real 6-byte JB, else the
+		// branch lands 4 bytes past the morestack block, inside the CALL
+		// displacement field.
+		{"leafbiglong", "TEXT \u00b7leafbiglong(SB), $8192-0\n" + strings.Repeat("\tMOVQ AX, BX\n", 40) + "\tRET\n",
+			"644c8b3425000000004989e44981ec881f00000f82960000004d3b66100f868c000000554889e54881ec00200000" + strings.Repeat("4889c3", 40) + "4881c4002000005dc3e800000000e947ffffff"},
+		{"callsmall", "TEXT \u00b7callsmall(SB), $16-0\n\tCALL \u00b7other(SB)\n\tRET\nTEXT \u00b7other(SB), NOSPLIT, $0\n\tRET\n",
+			"644c8b342500000000493b66107613554889e54883ec10e8000000004883c4105dc3e800000000ebd7"},
+		{"nosplit", "TEXT \u00b7nosplit(SB), NOSPLIT, $16-0\n\tRET\n",
+			"554889e54883ec104883c4105dc3"},
+	} {
+		f, errs := parser.Parse("g_amd64.s", tt.src)
+		if len(errs) > 0 {
+			t.Fatalf("%s: parse: %v", tt.name, errs)
+		}
+		img, err := AssembleFile(f)
+		if err != nil {
+			t.Fatalf("%s: assemble: %v", tt.name, err)
+		}
+		fn := img.Funcs[0]
+		// The toolchain's object leaves every relocation field zero for the
+		// linker, while the gasm image resolves file-internal references, so
+		// the comparison masks the patch sites the way verify's ground truth
+		// does.
+		code := append([]byte(nil), img.Code[fn.Offset:fn.Offset+fn.Size]...)
+		for _, r := range fn.Relocs {
+			for j := r.Off; j < r.Off+4 && j < len(code); j++ {
+				code[j] = 0
+			}
+		}
+		got := hex.EncodeToString(code)
+		if got != tt.want {
+			t.Errorf("%s:\n  got  %s\n  want %s", tt.name, got, tt.want)
+		}
+	}
+}
+
+// TestStackGuardRelocs checks the guard's patch sites: the TLS slot and the
+// morestack call.
+func TestStackGuardRelocs(t *testing.T) {
+	f, errs := parser.Parse("g_amd64.s", "TEXT \u00b7f(SB), $256-0\n\tRET\n")
+	if len(errs) > 0 {
+		t.Fatalf("parse: %v", errs)
+	}
+	img, err := AssembleFile(f)
+	if err != nil {
+		t.Fatalf("assemble: %v", err)
+	}
+	relocs := img.Funcs[0].Relocs
+	if len(relocs) != 2 {
+		t.Fatalf("relocs = %d, want 2", len(relocs))
+	}
+	tls, call := relocs[0], relocs[1]
+	if tls.Kind != RelTLSLE || tls.Off != 5 || tls.Name != "" || tls.External {
+		t.Errorf("tls reloc = %+v, want RelTLSLE at 5 with no symbol", tls)
+	}
+	if call.Kind != RelCall || call.Name != "runtime\u00b7morestack_noctxt" || !call.External {
+		t.Errorf("call reloc = %+v, want RelCall to runtime.morestack_noctxt", call)
+	}
+}
+
+// TestStackGuardGOObj emissions succeed with the guard's TLS and builtin
+// references in play.
+func TestStackGuardGOObj(t *testing.T) {
+	f, errs := parser.Parse("g_amd64.s", "TEXT \u00b7f(SB), $256-0\n\tCALL \u00b7helper(SB)\n\tRET\nTEXT \u00b7helper(SB), NOSPLIT, $0\n\tRET\n")
+	if len(errs) > 0 {
+		t.Fatalf("parse: %v", errs)
+	}
+	img, err := AssembleFile(f)
+	if err != nil {
+		t.Fatalf("assemble: %v", err)
+	}
+	obj, err := img.GOObject("testpkg", "g_amd64.s")
+	if err != nil {
+		t.Fatalf("GOObject: %v", err)
+	}
+	if !bytes.Contains(obj, []byte("go120ld")) {
+		t.Fatal("object lacks the GOOBJ magic")
+	}
+}
+
+// The arm64 stack-split guard, pinned from `go tool asm` (Go 1.27, arm64):
+// the guard classes, the auto-NOSPLIT leaf behaviour and the morestack
+// block.  Relocation fields are masked: the toolchain's object leaves them
+// zero for the linker, the gasm image resolves file-internal references.
+func TestStackGuardBytesARM64(t *testing.T) {
+	for _, tt := range []struct {
+		name string
+		src  string
+		want string
+	}{
+		{"leafsmall", "TEXT \u00b7leafsmall(SB), $16-0\n\tRET\n",
+			"fe0f1ef8fd831ff8fd2300d1fd630091ff830091c0035fd6"},
+		{"leafmed", "TEXT \u00b7leafmed(SB), $256-0\n\tRET\n",
+			"900b40f9f14302d13f0210eb09010054f44304d19dfa3fa99f020091fd2300d1fd230491ff430491c0035fd6e3031eaa00000000f3ffff17"},
+		{"leafbig", "TEXT \u00b7leafbig(SB), $8192-0\n\tRET\n",
+			"900b40f91bf283d2f1633beba30100543f0210eb690100541b0284d2f4633bcb9dfa3fa99f020091fd2300d11b0184d2fd633b8b1b0284d2ff633b8bc0035fd6e3031eaa00000000eeffff17"},
+		{"callsmall", "TEXT \u00b7callsmall(SB), $16-0\n\tCALL \u00b7other(SB)\n\tRET\nTEXT \u00b7other(SB), NOSPLIT, $0\n\tRET\n",
+			"900b40f9ff6330eb09010054fe0f1ef8fd831ff8fd2300d100000000fd835ff8fe0742f8c0035fd6e3031eaa00000000f4ffff17"},
+		{"nosplit", "TEXT \u00b7nosplit(SB), NOSPLIT, $16-0\n\tRET\n",
+			"fe0f1ef8fd831ff8fd2300d1fd630091ff830091c0035fd6"},
+	} {
+		f, errs := parser.Parse("g_arm64.s", tt.src)
+		if len(errs) > 0 {
+			t.Fatalf("%s: parse: %v", tt.name, errs)
+		}
+		img, err := AssembleFileARM64(f)
+		if err != nil {
+			t.Fatalf("%s: assemble: %v", tt.name, err)
+		}
+		fn := img.Funcs[0]
+		code := append([]byte(nil), img.Code[fn.Offset:fn.Offset+fn.Size]...)
+		for _, r := range fn.Relocs {
+			for j := r.Off; j < r.Off+4 && j < len(code); j++ {
+				code[j] = 0
+			}
+		}
+		got := hex.EncodeToString(code)
+		if got != tt.want {
+			t.Errorf("%s:\n  got  %s\n  want %s", tt.name, got, tt.want)
+		}
+	}
+}
+
+// TestStackGuardBranchTargetsARM64 checks the class-2 guard's branch
+// positions for a frame whose guard constant needs two MOV words: the
+// displacements must be computed from byte offsets (8+4*ml and 16+4*ml), so
+// both branches land on the morestack block rather than inside the body.
+// The frame size makes the toolchain switch its own prologue decomposition,
+// so the assertion is on the branch targets, not pinned bytes.
+func TestStackGuardBranchTargetsARM64(t *testing.T) {
+	f, errs := parser.Parse("g_arm64.s", "TEXT \u00b7f(SB), $65664-0\n\tCALL \u00b7other(SB)\n\tRET\nTEXT \u00b7other(SB), NOSPLIT, $0\n\tRET\n")
+	if len(errs) > 0 {
+		t.Fatalf("parse: %v", errs)
+	}
+	img, err := AssembleFileARM64(f)
+	if err != nil {
+		t.Fatalf("assemble: %v", err)
+	}
+	fn := img.Funcs[0]
+	code := img.Code[fn.Offset : fn.Offset+fn.Size]
+	if len(code)%4 != 0 {
+		t.Fatalf("function size %d is not a word multiple", len(code))
+	}
+	// autosize = 65680, so the guard materialises 65552 = MOVZ+MOVK: ml = 2
+	// and the branches sit at bytes 16 and 24 of the guard prefix.
+	const morestackBlock = 12 // MOVD R30, R3; BL; B back
+	blockStart := len(code) - morestackBlock
+	check := func(name string, off int) {
+		t.Helper()
+		w := leWord(code[off:])
+		imm19 := int32(w>>5) & 0x7FFFF
+		if imm19&(1<<18) != 0 {
+			imm19 -= 1 << 19
+		}
+		if target := off + int(imm19)*4; target != blockStart {
+			t.Errorf("%s at byte %d targets byte %d, want the morestack block at %d", name, off, target, blockStart)
+		}
+	}
+	check("B.LO", 16)
+	check("B.LS", 24)
+}
+
+// The riscv64 stack-split guard, pinned from `go tool asm` (Go 1.27,
+// riscv64): the morestack call sits between the guard and the body, and the
+// guard branches forward over it.  Relocation fields are masked.
+func TestStackGuardBytesRISCV64(t *testing.T) {
+	for _, tt := range []struct {
+		name string
+		src  string
+		want string
+	}{
+		{"leafsmall", "TEXT \u00b7leafsmall(SB), $16-0\n\tRET\n",
+			"03b30d0163662300000000006ff05fff233411fe211106e08260610167800000"},
+		{"leafmed", "TEXT \u00b7leafmed(SB), $256-0\n\tRET\n",
+			"03b30d01930381f763667300000000006ff01fff233c11ee130181ef06e082601301811067800000"},
+		{"leafbig", "TEXT \u00b7leafbig(SB), $8192-0\n\tRET\n",
+			"03b30d0189639b8383f863697100f97f9b8f8f07b303f10163667300000000006ff01ffef97f8a9f23bc1ffef97fe13f7e9106e08260896fa12f7e9167800000"},
+		{"frameless", "TEXT \u00b7frameless(SB), $0-0\n\tCALL \u00b7other(SB)\n\tRET\nTEXT \u00b7other(SB), NOSPLIT, $0\n\tRET\n",
+			"03b30d0163662300000000006ff05fff233c11fe611106e0000000008260210167800000"},
+		{"nosplit", "TEXT \u00b7nosplit(SB), NOSPLIT, $16-0\n\tRET\n",
+			"233411fe211106e08260610167800000"},
+	} {
+		f, errs := parser.Parse("g_riscv64.s", tt.src)
+		if len(errs) > 0 {
+			t.Fatalf("%s: parse: %v", tt.name, errs)
+		}
+		img, err := AssembleFileRISCV(f)
+		if err != nil {
+			t.Fatalf("%s: assemble: %v", tt.name, err)
+		}
+		fn := img.Funcs[0]
+		code := append([]byte(nil), img.Code[fn.Offset:fn.Offset+fn.Size]...)
+		for _, r := range fn.Relocs {
+			for j := r.Off; j < r.Off+4 && j < len(code); j++ {
+				code[j] = 0
+			}
+		}
+		got := hex.EncodeToString(code)
+		if got != tt.want {
+			t.Errorf("%s:\n  got  %s\n  want %s", tt.name, got, tt.want)
+		}
+	}
+}
+
+// The loong64 stack-split guard, pinned from `go tool asm` (Go 1.27,
+// loong64): every guard class (including the medium class with the
+// materialised constant and the big class with the ORI-less constants), the
+// auto-NOSPLIT leaf behaviour, the large-frame R30 prologue/epilogue forms
+// and the morestack block.  Relocation fields are masked.
+func TestStackGuardBytesLOONG64(t *testing.T) {
+	for _, tt := range []struct {
+		name string
+		src  string
+		want string
+	}{
+		{"leafsmall", "TEXT \u00b7leafsmall(SB), $16-0\n\tRET\n",
+			"61a0ff2963a0ff026100c0296360c0022000004c"},
+		{"leafmed", "TEXT \u00b7leafmed(SB), $256-0\n\tRET\n",
+			"d442c02878e0fd0294e21200801a004061e0fb2963e0fb026100c0296320c4022000004c3f00150000000000ffd7ff53"},
+		{"nosplit", "TEXT \u00b7nosplit(SB), NOSPLIT, $16-0\n\tRET\n",
+			"61a0ff2963a0ff026100c0296360c0022000004c"},
+		// The LR store leaves the 12-bit store-offset range while the SP
+		// adjust immediate still fits, and the epilogue adjusts through a
+		// single ORI.
+		{"fit2048", "TEXT \u00b7fit2048(SB), $2040-0\n\tRET\n",
+			"d442c0287800e20294e21200802600401e000014de8f1000c103e0296300e0026100c0291e00a00363f810002000004c3f00150000000000ffcbff53"},
+		// Medium class at the materialisation boundary (off = 2048 still
+		// immediate, 2049+ goes through R30).
+		{"med2048off", "TEXT \u00b7med2048off(SB), $2168-0\n\tRET\n",
+			"d442c0287800e00294e21200802e0040feffff15de8f1000c103de29feffff15de039e0363f810006100c0291e00a20363f810002000004c3f00150000000000ffc3ff53"},
+		{"medmat", "TEXT \u00b7medmat(SB), $2176-0\n\tRET\n",
+			"d442c028feffff15dee39f0378f8100094e21200802e0040feffff15de8f1000c1e3dd29feffff15dee39d0363f810006100c0291e20a20363f810002000004c3f00150000000000ffbbff53"},
+		// Big class with the rounding-split store and the floor-split adjust.
+		{"leafbig", "TEXT \u00b7leafbig(SB), $8192-0\n\tRET\n",
+			"d442c0283e000014de23be0378f8120000470044deffff15dee3810378f8100094e2120080320040deffff15de8f1000c1e3ff29beffff15dee3bf0363f810006100c0295e000014de23800363f810002000004c3f00150000000000ffa7ff53"},
+		// Zero low 12 bits drop the ORI from the store, the adjust and the
+		// epilogue materialisation.
+		{"bigzero", "TEXT \u00b7bigzero(SB), $4088-0\n\tRET\n",
+			"d442c028feffff15de03820378f8100094e21200802a0040feffff15de8f1000c103c029feffff1563f810006100c0293e00001463f810002000004c3f00150000000000ffbfff53"},
+		// Big class whose first constant has a zero high part: a single ORI.
+		{"big3976", "TEXT \u00b7big3976(SB), $4096-0\n\tRET\n",
+			"d442c0281e20be0378f8120000470044feffff15dee3810378f8100094e2120080320040feffff15de8f1000c1e3ff29deffff15dee3bf0363f810006100c0293e000014de23800363f810002000004c3f00150000000000ffabff53"},
+		// Big class at a multiple of 4096: both guard constants lose their
+		// ORI word.
+		{"giantlo0", "TEXT \u00b7giantlo0(SB), $4216-0\n\tRET\n",
+			"d442c0283e00001478f8120000430044feffff1578f8100094e2120080320040feffff15de8f1000c103fe29deffff15de03be0363f810006100c0293e000014de03820363f810002000004c3f00150000000000ffafff53"},
+		// Non-leaf big frame: the body call plus the LR restore epilogue.
+		{"callbig", "TEXT \u00b7callbig(SB), $8192-0\n\tCALL \u00b7other(SB)\n\tRET\nTEXT \u00b7other(SB), NOSPLIT, $0\n\tRET\n",
+			"d442c0283e000014de23be0378f81200004f0044deffff15dee3810378f8100094e21200803a0040deffff15de8f1000c1e3ff29beffff15dee3bf0363f810006100c029000000006100c0285e000014de23800363f810002000004c3f00150000000000ff9fff53"},
+	} {
+		f, errs := parser.Parse("g_loong64.s", tt.src)
+		if len(errs) > 0 {
+			t.Fatalf("%s: parse: %v", tt.name, errs)
+		}
+		img, err := AssembleFileLOONG64(f)
+		if err != nil {
+			t.Fatalf("%s: assemble: %v", tt.name, err)
+		}
+		fn := img.Funcs[0]
+		code := append([]byte(nil), img.Code[fn.Offset:fn.Offset+fn.Size]...)
+		for _, r := range fn.Relocs {
+			for j := r.Off; j < r.Off+4 && j < len(code); j++ {
+				code[j] = 0
+			}
+		}
+		got := hex.EncodeToString(code)
+		if got != tt.want {
+			t.Errorf("%s:\n  got  %s\n  want %s", tt.name, got, tt.want)
+		}
+	}
+}
+
+// TestStackGuardGOObjInternalCall checks that GOOBJ emission succeeds when a
+// guarded function calls a TEXT symbol of the same file, for every arch's
+// call relocation kind.
+func TestStackGuardGOObjInternalCall(t *testing.T) {
+	for _, tt := range []struct {
+		src      string
+		assemble func(*ast.File, ...AssembleOption) (*Image, error)
+	}{
+		{"g_amd64.s", AssembleFile},
+		{"g_arm64.s", func(f *ast.File, _ ...AssembleOption) (*Image, error) { return AssembleFileARM64(f) }},
+		{"g_riscv64.s", func(f *ast.File, _ ...AssembleOption) (*Image, error) { return AssembleFileRISCV(f) }},
+		{"g_loong64.s", func(f *ast.File, _ ...AssembleOption) (*Image, error) { return AssembleFileLOONG64(f) }},
+	} {
+		f, errs := parser.Parse(tt.src, "TEXT \u00b7callsmall(SB), $16-0\n\tCALL \u00b7other(SB)\n\tRET\nTEXT \u00b7other(SB), NOSPLIT, $0\n\tRET\n")
+		if len(errs) > 0 {
+			t.Fatalf("%s: parse: %v", tt.src, errs)
+		}
+		img, err := tt.assemble(f)
+		if err != nil {
+			t.Fatalf("%s: assemble: %v", tt.src, err)
+		}
+		if _, err := img.GOObject("testpkg", tt.src); err != nil {
+			t.Errorf("%s: GOObject: %v", tt.src, err)
+		}
+	}
+}

asm/integration_test.go

@@ -16,16 +16,16 @@ import (
 
 	"golang.org/x/arch/x86/x86asm"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
 )
 
-// TestAssembleGoFlacAVX2Kernel assembles the whole production AVX2 kernel —
-// all functions plus the file-local mask24 constant — and checks that every
+// TestAssembleGoFlacAVX2Kernel assembles the whole production AVX2 kernel;
+// all functions plus the file-local mask24 constant; and checks that every
 // static-symbol load resolves to the right bytes in the image.
 func TestAssembleGoFlacAVX2Kernel(t *testing.T) {
 	path := "../../go-libraries/go-flac/avx2_amd64.s"
 	if _, err := os.Stat(path); err != nil {
-		t.Skip("go-libraries repository not present next to gasm-devkit")
+		t.Skip("go-libraries repository not present next to gasm-sdk")
 	}
 	src, err := os.ReadFile(path)
 	if err != nil {
@@ -81,12 +81,12 @@ func TestAssembleGoFlacAVX2Kernel(t *testing.T) {
 }
 
 // TestAssembleGoFlacAVX512Kernel assembles the whole production AVX-512
-// kernel — all functions plus the file-global idx16 constant — and checks
+// kernel, all functions plus the file-global idx16 constant, and checks
 // that the static-symbol load resolves to the right bytes in the image.
 func TestAssembleGoFlacAVX512Kernel(t *testing.T) {
 	path := "../../go-libraries/go-flac/avx512_amd64.s"
 	if _, err := os.Stat(path); err != nil {
-		t.Skip("go-libraries repository not present next to gasm-devkit")
+		t.Skip("go-libraries repository not present next to gasm-sdk")
 	}
 	src, err := os.ReadFile(path)
 	if err != nil {

asm/kernels_differential_test.go

@@ -0,0 +1,219 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package asm
+
+import (
+	"bytes"
+	"encoding/binary"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"runtime"
+	"strings"
+	"testing"
+
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
+)
+
+// The differential kernels for the DATA-path and front-end gaps are kept in
+// testdata/verify beside the campaign's other kernels; the verify package's
+// suites are not open to the asm package, so this test is their runner: each
+// kernel assembles through gasm and through go tool asm, and the functions'
+// bytes must agree with the relocation sites masked on both sides.
+
+// toolAsmObject assembles path with the installed toolchain's assembler for
+// goarch ("" = the host) and returns the object bytes.
+func toolAsmObject(t *testing.T, path, goarch string) []byte {
+	t.Helper()
+	goBin, err := exec.LookPath("go")
+	if err != nil {
+		t.Skip("no Go toolchain available")
+	}
+	out, err := exec.Command(goBin, "env", "GOROOT").Output()
+	if err != nil {
+		t.Fatalf("go env GOROOT: %v", err)
+	}
+	includeDir := filepath.Join(strings.TrimSpace(string(out)), "pkg", "include")
+
+	pkg := strings.TrimSuffix(filepath.Base(path), ".s")
+	pkg = strings.TrimSuffix(pkg, "_amd64")
+	pkg = strings.TrimSuffix(pkg, "_arm64")
+
+	objPath := filepath.Join(t.TempDir(), "oracle.o")
+	cmd := exec.Command(goBin, "tool", "asm", "-I", includeDir, "-p", pkg, "-o", objPath, path)
+	if goarch != "" {
+		environ := os.Environ()
+		env := make([]string, 0, len(environ)+1)
+		for _, e := range environ {
+			if !strings.HasPrefix(e, "GOARCH=") {
+				env = append(env, e)
+			}
+		}
+		cmd.Env = append(env, "GOARCH="+goarch)
+	}
+	if out, err := cmd.CombinedOutput(); err != nil {
+		t.Fatalf("go tool asm %s: %v\n%s", filepath.Base(path), err, out)
+	}
+	obj, err := os.ReadFile(objPath)
+	if err != nil {
+		t.Fatal(err)
+	}
+	return obj
+}
+
+// oracleFuncCode extracts the non-package TEXT functions' code bytes from a
+// toolchain object, keyed by the name the object records (pkg.name).  Each
+// function's span is its own symbol size: a toolchain object that follows
+// the text with data symbols (the synthesised float-constant pool) would
+// otherwise fold them into the last function's bytes.
+func oracleFuncCode(t *testing.T, obj []byte) map[string][]byte {
+	t.Helper()
+	v := openGoobj(t, obj)
+	le := binary.LittleEndian
+	const symSize = 21
+	nps := v.syms(blkNonpkgdef)
+	data := v.blk(blkData)
+	didx := v.blk(blkDataIdx)
+	preceding := 0
+	for _, bi := range []int{blkSymdef, blkHashed64def, blkHasheddef} {
+		preceding += len(v.blk(bi)) / symSize
+	}
+	out := make(map[string][]byte, len(nps))
+	for i, s := range nps {
+		if s.typ != kindSTEXT {
+			continue
+		}
+		start := le.Uint32(didx[4*(preceding+i):])
+		out[s.name] = data[start : start+s.size]
+	}
+	return out
+}
+
+// maskCode zeroes every relocation field, the way the toolchain's object
+// leaves them for the linker.
+func maskCode(code []byte, relocs []Reloc) []byte {
+	for _, r := range relocs {
+		for j := r.Off; j < r.Off+4 && j < len(code); j++ {
+			code[j] = 0
+		}
+	}
+	return code
+}
+
+// code assembles src for amd64 and returns the image's code bytes.
+func code(path, src string) []byte {
+	f, errs := parser.Parse(path, src)
+	if len(errs) > 0 {
+		return nil
+	}
+	img, err := AssembleFile(f)
+	if err != nil {
+		return nil
+	}
+	return img.Code
+}
+
+// TestDifferentialKernels pins the new kernels against the oracle.
+func TestDifferentialKernels(t *testing.T) {
+	if runtime.GOARCH != "amd64" {
+		t.Skip("the amd64 kernels assume an amd64 host assembler default")
+	}
+	for _, k := range []struct {
+		path   string
+		goarch string
+		arm64  bool
+	}{
+		{filepath.Join("..", "testdata", "verify", "datarel_amd64.s"), "", false},
+		{filepath.Join("..", "testdata", "verify", "divslash_amd64.s"), "", false},
+		{filepath.Join("..", "testdata", "verify", "semicolons_amd64.s"), "", false},
+		{filepath.Join("..", "testdata", "verify", "quadreg_amd64.s"), "", false},
+		{filepath.Join("..", "testdata", "verify", "floatimm_amd64.s"), "", false},
+		{filepath.Join("..", "testdata", "verify", "bookkeep_amd64.s"), "", false},
+		{filepath.Join("..", "testdata", "verify", "forms_amd64.s"), "", false},
+		{filepath.Join("..", "testdata", "verify", "datarel_arm64.s"), "arm64", true},
+		{filepath.Join("..", "testdata", "verify", "divslash_arm64.s"), "arm64", true},
+	} {
+		t.Run(filepath.Base(k.path), func(t *testing.T) {
+			src, err := os.ReadFile(k.path)
+			if err != nil {
+				t.Fatalf("read: %v", err)
+			}
+			f, errs := parser.Parse(k.path, string(src))
+			if len(errs) > 0 {
+				t.Fatalf("parse: %v", errs)
+			}
+			var img *Image
+			if k.arm64 {
+				img, err = AssembleFileARM64(f)
+			} else {
+				img, err = AssembleFile(f)
+			}
+			if err != nil {
+				t.Fatalf("assemble: %v", err)
+			}
+			gt := oracleFuncCode(t, toolAsmObject(t, k.path, k.goarch))
+			// The oracle keys its functions by the qualified object name
+			// (pkg.name); match on the local part.
+			byLocal := make(map[string][]byte, len(gt))
+			for name, code := range gt {
+				if _, after, ok := strings.Cut(name, "."); ok {
+					name = after
+				}
+				byLocal[name] = code
+			}
+
+			matched := 0
+			for _, fn := range img.Funcs {
+				gasmCode := maskCode(append([]byte(nil), img.Code[fn.Offset:fn.Offset+fn.Size]...), fn.Relocs)
+				goCode, ok := byLocal[fn.Name]
+				if !ok {
+					t.Errorf("%s: not in ground truth (%d functions: %v)", fn.Name, len(gt), keysOf(byLocal))
+					continue
+				}
+				goCode = maskCode(append([]byte(nil), goCode...), fn.Relocs)
+				cmpLen := min(len(goCode), len(gasmCode))
+				if !bytes.Equal(gasmCode[:cmpLen], goCode[:cmpLen]) {
+					t.Errorf("%s: MISMATCH gasm=%d go=%d bytes\ngasm %x\ngo   %x", fn.Name, len(gasmCode), len(goCode), gasmCode, goCode)
+					continue
+				}
+				for _, b := range goCode[len(gasmCode):] {
+					if b != 0 {
+						t.Errorf("%s: non-zero trailing bytes in go tool asm output", fn.Name)
+						break
+					}
+				}
+				matched++
+				t.Logf("%s: MATCH (%d bytes)", fn.Name, len(gasmCode))
+			}
+			if matched == 0 {
+				t.Fatal("no functions matched")
+			}
+		})
+	}
+}
+
+func keysOf(m map[string][]byte) []string {
+	out := make([]string, 0, len(m))
+	for k := range m {
+		out = append(out, k)
+	}
+	return out
+}
+
+// TestSemicolonSpellingParity pins that the ';' statement separator changes
+// nothing about the encoding: the one-line spelling assembles to exactly the
+// bytes of the same statements written one per line.
+func TestSemicolonSpellingParity(t *testing.T) {
+	for _, tt := range []struct{ one, two string }{
+		{"\tROLQ $3, DI; ROLQ $13, DI\n", "\tROLQ $3, DI\n\tROLQ $13, DI\n"},
+		{"\tREP; MOVSQ\n", "\tREP\n\tMOVSQ\n"},
+		{"\tXORQ AX, AX; XORQ CX, CX\n", "\tXORQ AX, AX\n\tXORQ CX, CX\n"},
+	} {
+		one := code("t.s", "TEXT \u00b7f(SB), NOSPLIT, $0\n"+tt.one+"\tRET\n")
+		two := code("t.s", "TEXT \u00b7f(SB), NOSPLIT, $0\n"+tt.two+"\tRET\n")
+		if !bytes.Equal(one, two) {
+			t.Errorf("semicolon spelling %q: %x, want the two-line bytes %x", tt.one, one, two)
+		}
+	}
+}

asm/l64_goobj_test.go

@@ -12,7 +12,7 @@ import (
 	"strings"
 	"testing"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
 )
 
 // TestGOObjectLOONG64Structure checks the emitted loong64 object's blocks:
@@ -94,9 +94,9 @@ DATA ·table<>+0(SB)/8, $0x1122334455667788
 	}
 
 	// The debug_line program: LNE_set_address (the R_ADDR relocation
-	// carries the function address), then one row per line change — the
+	// carries the function address), then one row per line change; the
 	// TEXT is on line 4 (a leading blank line precedes the include), the
-	// instructions on lines 5–9 — an advance to the 20-byte end and an
+	// instructions on lines 5-9; an advance to the 20-byte end and an
 	// end-of-sequence.
 	linesOff := le.Uint32(dataIdx[4*2:])
 	lines := dataBlk[linesOff : linesOff+21]
@@ -214,7 +214,7 @@ func main() {
 		t.Fatalf("baseline build: %v\n%s", err, buildLog)
 	}
 	var pkgArch, work, linkLine, asmObj string
-	for _, line := range strings.Split(string(buildLog), "\n") {
+	for line := range strings.SplitSeq(string(buildLog), "\n") {
 		switch {
 		case strings.HasPrefix(line, "WORK="):
 			work = strings.TrimPrefix(line, "WORK=")
@@ -278,7 +278,7 @@ func main() {
 	newArch := filepath.Join(dir, "pkg.a")
 	args := []string{"tool", "pack", "c", newArch}
 	seen := map[string]bool{}
-	for _, m := range strings.Fields(string(listOut)) {
+	for m := range strings.FieldsSeq(string(listOut)) {
 		if seen[m] {
 			continue
 		}

asm/link.go

@@ -5,9 +5,11 @@ package asm
 
 import (
 	"fmt"
+	"math"
 	"sort"
+	"strconv"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/ast"
+	"sourcedock.dev/petrbalvin/gasm-sdk/ast"
 )
 
 // Image is an assembled file: the function bodies laid out in source order,
@@ -15,7 +17,7 @@ import (
 // file-local static symbols are encoded RIP-relative and resolved within the
 // image, so the raw bytes are self-consistent and executable at any base
 // address; references to external symbols are recorded as relocations
-// (Funcs[i].Relocs, Externals) and left unresolved — the object-file
+// (Funcs[i].Relocs, Externals) and left unresolved, the object-file
 // emitters turn them into linker relocations.
 type Image struct {
 	Code      []byte         // concatenated function bodies
@@ -24,6 +26,10 @@ type Image struct {
 	Symbols   map[string]int // static symbol → byte offset within the image
 	DataSyms  []DataSymbol   // GLOBL symbols, in layout order
 	Externals []string       // referenced but undefined symbols, sorted
+	// SourcePath is the assembled file's path, recorded in the DWARF
+	// sections in place of a placeholder name.  Empty when the image was
+	// not built from a named file.
+	SourcePath string
 }
 
 // FuncLayout describes one assembled function within an Image.
@@ -80,32 +86,45 @@ func (fl *FuncLayout) LineAt(offset int) int {
 	return 0
 }
 
-// Reloc is one static-symbol reference within a function body: the disp32
-// field at Off (function-relative) must reach the symbol plus Addend,
-// measured from After, the address just past the instruction.  An External
-// relocation names a symbol no GLOBL in the file defines; the object-file
-// emitters carry it into the output's relocation table.
-// RelocKind discriminates the type of relocation needed.
+// RelocKind discriminates the relocation a static-symbol reference needs;
+// the encoders record one per SB reference, and the object-file emitters map
+// it to their format's relocation type.
 type RelocKind int
 
 const (
 	RelPCRel32         RelocKind = iota // 32-bit PC-relative (amd64)
+	RelCall                             // R_CALL: CALL to a function symbol (amd64)
+	RelTLSLE                            // R_TLS_LE: local-exec TLS load, no symbol (amd64 guard)
 	RelRISCVPCRELIType                  // R_RISCV_PCREL_ITYPE (AUIPC + I-type pair)
 	RelRISCVPCRELSType                  // R_RISCV_PCREL_STYPE (AUIPC + S-type pair)
 	RelRISCVJal                         // R_RISCV_JAL (J-type call)
-	RelPCRelAbs                         // 32-bit absolute (R_RISCV_32)
 	RelLoong64AddrHi                    // R_LOONG64_ADDR_HI (pcalau12i)
 	RelLoong64AddrLo                    // R_LOONG64_ADDR_LO (addi.d/ld/st)
-	RelArm64Addr                        // R_ADDRARM64 (ADRP + ADD/LDR/STR pair)
+	RelArm64Addr                        // R_ADDRARM64 (ADRP + ADD pair)
+	RelArm64Branch                      // R_CALLARM64 (BL instruction)
+	RelArm64LDST64                      // R_ARM64_PCREL_LDST64 (ADRP + 64-bit LDR/STR pair)
+	RelLoong64Branch                    // R_CALLLOONG64 (BL instruction)
+	RelAddr                             // R_ADDR: the absolute address of a symbol held in a DATA field
 )
 
 type Reloc struct {
+	// Off is the function-relative offset of the field the linker patches
+	// and After the address just past the instruction, the base the
+	// assembler measures PC-relative displacements from.  Name plus
+	// Addend select the target: the symbol plus the byte offset.  An
+	// External relocation names a symbol no GLOBL in the file defines;
+	// the object-file emitters carry it into the output's relocation
+	// table.  Siz is the width of the patched field and is set only for
+	// data-field relocations (RelAddr, Off relative to the data symbol),
+	// whose width is the DATA line's; code relocations take their width
+	// from the architecture's instruction encoding.
 	Off      int
 	After    int
 	Name     string
 	Addend   int64
 	External bool
 	Kind     RelocKind
+	Siz      uint8
 }
 
 // DataSymbol describes one GLOBL symbol laid out in the data section.
@@ -117,6 +136,11 @@ type DataSymbol struct {
 	Static bool // the <> marker: file-local, not exported
 	Rodata bool // the RODATA flag: read-only data
 	Dupok  bool // the DUPOK flag: duplicate-OK
+	// Relocs carries the symbol-valued DATA initialisers ("DATA s+0(SB)/8,
+	// $other(SB)"): fields of this symbol's data that hold another symbol's
+	// address, resolved by the linker.  Off is relative to the symbol's
+	// data start.
+	Relocs []Reloc
 }
 
 // Bytes returns the whole image: code, then data.
@@ -126,14 +150,26 @@ func (img *Image) Bytes() []byte {
 	return append(out, img.Data...)
 }
 
+// AssembleOption adjusts the file-level assembly context.
+type AssembleOption func(*linkInfo)
+
+// WithGOOS selects the target operating system for the forms that depend on
+// it, the TLS access shape above all: linux and freebsd take the
+// one-instruction form, windows and plan9 keep the two-instruction load.
+func WithGOOS(goos string) AssembleOption {
+	return func(l *linkInfo) {
+		l.goos = goos
+	}
+}
+
 // AssembleFile assembles every TEXT function of a parsed file and lays out
 // its static symbols (GLOBL/DATA) in a data section behind the code.  Each
 // reference to a file-local static symbol becomes a RIP-relative load whose
 // displacement is resolved against that layout; a reference to a symbol no
 // GLOBL defines is recorded as an external relocation (Externals) with its
-// displacement left zero — the object-file emitters resolve it at link
+// displacement left zero, the object-file emitters resolve it at link
 // time, while the raw image (Bytes) cannot represent it.
-func AssembleFile(f *ast.File) (*Image, error) {
+func AssembleFile(f *ast.File, opts ...AssembleOption) (*Image, error) {
 	dataSyms, err := collectData(f)
 	if err != nil {
 		return nil, err
@@ -142,9 +178,21 @@ func AssembleFile(f *ast.File) (*Image, error) {
 	for _, d := range dataSyms {
 		known[d.name] = true
 	}
+	// TEXT symbols are file-level definitions too: a symbol immediate
+	// ($fn(SB)) may name one, exactly as a data reference names a GLOBL.
+	for _, d := range f.Decls {
+		if t, ok := d.(*ast.Text); ok {
+			known[t.Name.Name] = true
+		}
+	}
 	link := &linkInfo{symbols: known, allowExternal: true}
+	for _, o := range opts {
+		o(link)
+	}
+	poolSeen := map[string]bool{}
 
-	img := &Image{Symbols: map[string]int{}}
+	img := &Image{Symbols: map[string]int{}, SourcePath: f.Path}
+	textOff := map[string]int{}
 	type asmFunc struct {
 		name    string
 		patches []sbPatch
@@ -155,7 +203,26 @@ func AssembleFile(f *ast.File) (*Image, error) {
 		if !ok {
 			continue
 		}
-		code, patches, labels, steps, lines, err := assemble(t, link)
+		code, patches, labels, steps, lines, pool, err := assemble(t, link)
+		if err != nil {
+			return nil, fmt.Errorf("%s: %w", t.Name.Name, err)
+		}
+		// The pooled floating-point constants join the declared data as
+		// read-only symbols, deduplicated across the file (the toolchain
+		// synthesises the same symbols into its rodata).
+		for _, entry := range pool {
+			if poolSeen[entry.name] {
+				continue
+			}
+			poolSeen[entry.name] = true
+			dataSyms = append(dataSyms, dataSym{
+				name:   entry.name,
+				buf:    entry.data,
+				size:   len(entry.data),
+				rodata: true,
+				dupok:  true,
+			})
+		}
 		if err != nil {
 			return nil, fmt.Errorf("%s: %w", t.Name.Name, err)
 		}
@@ -182,6 +249,7 @@ func AssembleFile(f *ast.File) (*Image, error) {
 		for _, s := range steps {
 			fl.Spadj = append(fl.Spadj, SpadjStep{PC: s.pc, Value: s.value})
 		}
+		textOff[t.Name.Name] = len(img.Code)
 		img.Funcs = append(img.Funcs, fl)
 		img.Code = append(img.Code, code...)
 		funcs = append(funcs, asmFunc{name: t.Name.Name, patches: patches})
@@ -214,13 +282,27 @@ func AssembleFile(f *ast.File) (*Image, error) {
 		base := img.Funcs[i].Offset
 		code := img.Code[base : base+img.Funcs[i].Size]
 		for _, p := range fn.patches {
-			reloc := Reloc{Off: p.off, After: p.after, Name: p.name, Addend: p.addend}
+			reloc := Reloc{Off: p.off, After: p.after, Name: p.name, Addend: p.addend, Kind: p.kind}
+			if p.kind == RelTLSLE {
+				// The TLS slot has no symbol: the linker fills the offset
+				// from the runtime's TLS layout.
+				img.Funcs[i].Relocs = append(img.Funcs[i].Relocs, reloc)
+				continue
+			}
 			if imgOff, ok := img.Symbols[p.name]; ok {
 				rel := int64(imgOff) + p.addend - int64(base+p.after)
 				if rel < -1<<31 || rel >= 1<<31 {
 					return nil, fmt.Errorf("%s: displacement to %q out of rel32 range", fn.name, p.name)
 				}
 				copy(code[p.off:p.off+4], le32(rel))
+			} else if imgOff, ok := textOff[p.name]; ok {
+				// A CALL to a TEXT function of the same file: resolve the
+				// displacement against the function's layout position.
+				rel := int64(imgOff) + p.addend - int64(base+p.after)
+				if rel < -1<<31 || rel >= 1<<31 {
+					return nil, fmt.Errorf("%s: displacement to %q out of rel32 range", fn.name, p.name)
+				}
+				copy(code[p.off:p.off+4], le32(rel))
 			} else {
 				reloc.External = true
 				externals[p.name] = true
@@ -228,6 +310,22 @@ func AssembleFile(f *ast.File) (*Image, error) {
 			img.Funcs[i].Relocs = append(img.Funcs[i].Relocs, reloc)
 		}
 	}
+	// The data symbols' symbol-valued DATA fields resolve the same way the
+	// code references do: a name the file defines (GLOBL or TEXT) stays an
+	// internal reference the emitters resolve, anything else is external.
+	// img.DataSyms was laid out in dataSyms order, so the indexes line up.
+	for i := range img.DataSyms {
+		for _, r := range dataSyms[i].relocs {
+			reloc := r
+			if _, ok := img.Symbols[reloc.Name]; !ok {
+				if _, ok := textOff[reloc.Name]; !ok {
+					reloc.External = true
+					externals[reloc.Name] = true
+				}
+			}
+			img.DataSyms[i].Relocs = append(img.DataSyms[i].Relocs, reloc)
+		}
+	}
 	for name := range externals {
 		img.Externals = append(img.Externals, name)
 	}
@@ -244,17 +342,34 @@ func AssembleFileRISCV(f *ast.File) (*Image, error) {
 	if err != nil {
 		return nil, err
 	}
+	// The pooled $i64 constants the wide MOV immediate loads refer to join
+	// the declared data as read-only symbols, deduplicated across the file
+	// (the toolchain synthesises the same symbols into its rodata).
+	litSeen := map[string]bool{}
 
-	img := &Image{Symbols: map[string]int{}}
+	img := &Image{Symbols: map[string]int{}, SourcePath: f.Path}
 	for _, d := range f.Decls {
 		t, ok := d.(*ast.Text)
 		if !ok {
 			continue
 		}
-		code, labels, relocs, lines, spadj, err := assembleRISCV(t)
+		code, labels, relocs, lines, spadj, lits, err := assembleRISCV(t)
 		if err != nil {
 			return nil, fmt.Errorf("%s: %w", t.Name.Name, err)
 		}
+		for _, lit := range lits {
+			if litSeen[lit.Name] {
+				continue
+			}
+			litSeen[lit.Name] = true
+			dataSyms = append(dataSyms, dataSym{
+				name:   lit.Name,
+				buf:    lit.Data,
+				size:   len(lit.Data),
+				rodata: true,
+				dupok:  true,
+			})
+		}
 		fl := FuncLayout{
 			Name:   t.Name.Name,
 			Pkg:    t.Name.Pkg,
@@ -302,6 +417,7 @@ func AssembleFileRISCV(f *ast.File) (*Image, error) {
 		})
 	}
 
+	markExternals(img, dataSyms)
 	return img, nil
 }
 
@@ -316,7 +432,7 @@ func AssembleFileLOONG64(f *ast.File) (*Image, error) {
 		return nil, err
 	}
 
-	img := &Image{Symbols: map[string]int{}}
+	img := &Image{Symbols: map[string]int{}, SourcePath: f.Path}
 	for _, d := range f.Decls {
 		t, ok := d.(*ast.Text)
 		if !ok {
@@ -373,9 +489,55 @@ func AssembleFileLOONG64(f *ast.File) (*Image, error) {
 		})
 	}
 
+	markExternals(img, dataSyms)
 	return img, nil
 }
 
+// markExternals identifies relocations that reference symbols not defined in
+// the file (neither a GLOBL/DATA symbol nor a TEXT function) and records them
+// as external.  The non-amd64 architectures emit relocations for every SB
+// reference; this post-processing step distinguishes file-local from external.
+func markExternals(img *Image, dataSyms []dataSym) {
+	known := make(map[string]bool, len(dataSyms)+len(img.Funcs))
+	for _, d := range dataSyms {
+		known[d.name] = true
+	}
+	for _, fn := range img.Funcs {
+		known[fn.Name] = true
+	}
+	externals := map[string]bool{}
+	for i := range img.Funcs {
+		for j := range img.Funcs[i].Relocs {
+			r := &img.Funcs[i].Relocs[j]
+			if !known[r.Name] {
+				r.External = true
+				externals[r.Name] = true
+			}
+		}
+	}
+	// The declared data symbols carry the file's own relocations (the
+	// symbol-valued DATA fields); the layouts appended img.DataSyms in
+	// dataSyms order, so the indexes line up.  The trailing entries (the
+	// pooled arm64 literals) have no source relocations.
+	for i := range img.DataSyms {
+		if i >= len(dataSyms) {
+			break
+		}
+		for _, r := range dataSyms[i].relocs {
+			reloc := r
+			if !known[reloc.Name] {
+				reloc.External = true
+				externals[reloc.Name] = true
+			}
+			img.DataSyms[i].Relocs = append(img.DataSyms[i].Relocs, reloc)
+		}
+	}
+	for name := range externals {
+		img.Externals = append(img.Externals, name)
+	}
+	sort.Strings(img.Externals)
+}
+
 // dataSym is one GLOBL symbol and its DATA initialiser.
 type dataSym struct {
 	name   string
@@ -385,81 +547,155 @@ type dataSym struct {
 	static bool
 	rodata bool
 	dupok  bool
+	// relocs are the symbol-valued DATA fields, in declaration order; Off
+	// is relative to the symbol's data start.
+	relocs []Reloc
 }
 
 // collectData gathers the file's static symbols (GLOBL) and their initial
-// contents (DATA) into byte buffers, in declaration order.
+// contents (DATA) into byte buffers.  Two passes: the Plan 9 convention puts
+// every DATA line before its symbol's GLOBL, so the symbols are registered
+// before the initialisers are applied.
 func collectData(f *ast.File) ([]dataSym, error) {
 	index := map[string]int{}
 	var syms []dataSym
 	for _, d := range f.Decls {
-		switch dd := d.(type) {
-		case *ast.Globl:
-			if dd.Name == nil || dd.Name.Pseudo != "SB" {
-				continue
-			}
-			name := dd.Name.Name
-			if _, dup := index[name]; dup {
-				return nil, fmt.Errorf("duplicate GLOBL %q", name)
-			}
-			size := 0
-			if dd.Size != nil && dd.Size.Imm.HasVal {
-				size = int(dd.Size.Imm.Val)
-			}
-			index[name] = len(syms)
-			ds := dataSym{
-				name:   name,
-				pkg:    dd.Name.Pkg,
-				buf:    make([]byte, size),
-				size:   size,
-				static: dd.Name.Static,
-			}
-			for _, f := range dd.Flags {
-				switch f {
-				case "RODATA":
-					ds.rodata = true
-				case "DUPOK":
-					ds.dupok = true
-				case "1":
-					ds.dupok = true
-				case "8":
-					ds.rodata = true
-				case "9":
-					ds.dupok = true
-					ds.rodata = true
+		gd, ok := d.(*ast.Globl)
+		if !ok {
+			continue
+		}
+		if gd.Name == nil || gd.Name.Pseudo != "SB" {
+			continue
+		}
+		name := gd.Name.Name
+		if _, dup := index[name]; dup {
+			return nil, fmt.Errorf("duplicate GLOBL %q", name)
+		}
+		size := 0
+		if gd.Size != nil && gd.Size.Imm.HasVal {
+			size = int(gd.Size.Imm.Val)
+		}
+		index[name] = len(syms)
+		ds := dataSym{
+			name:   name,
+			pkg:    gd.Name.Pkg,
+			buf:    make([]byte, size),
+			size:   size,
+			static: gd.Name.Static,
+		}
+		for _, f := range gd.Flags {
+			switch f {
+			case "RODATA":
+				ds.rodata = true
+			case "DUPOK":
+				ds.dupok = true
+			default:
+				// Legacy numeric flag constants (runtime/textflag.h):
+				// DUPOK is 2, RODATA is 8; combinations arrive as one
+				// number (e.g. 10 = RODATA|DUPOK).
+				if n, err := strconv.Atoi(f); err == nil {
+					if n&2 != 0 {
+						ds.dupok = true
+					}
+					if n&8 != 0 {
+						ds.rodata = true
+					}
 				}
 			}
-			syms = append(syms, ds)
-
-		case *ast.Data:
-			if dd.Name == nil || dd.Name.Pseudo != "SB" {
-				continue
+		}
+		syms = append(syms, ds)
+	}
+	for _, d := range f.Decls {
+		dd, ok := d.(*ast.Data)
+		if !ok {
+			continue
+		}
+		if dd.Name == nil || dd.Name.Pseudo != "SB" {
+			continue
+		}
+		i, ok := index[dd.Name.Name]
+		if !ok {
+			return nil, fmt.Errorf("DATA %q: no matching GLOBL", dd.Name.Name)
+		}
+		if dd.Value == nil {
+			return nil, fmt.Errorf("DATA %q: missing value", dd.Name.Name)
+		}
+		w := dd.Width
+		off := dd.Name.Offset
+		buf := syms[i].buf
+		if off < 0 || off+int64(w) > int64(len(buf)) {
+			return nil, fmt.Errorf("DATA %q+%d/%d exceeds GLOBL size %d", dd.Name.Name, off, w, len(buf))
+		}
+		// A symbol value ("DATA s+0(SB)/8, $other(SB)", the rt0 spelling)
+		// leaves the field zero and records a relocation against the named
+		// symbol: the linker patches the absolute address at this data
+		// offset.  The toolchain emits the same shape, an R_ADDR of the
+		// DATA width with the value's offset as the addend, on every
+		// architecture.
+		if sym := dd.Value.Imm.Sym; !dd.Value.Imm.HasVal && sym != nil {
+			syms[i].relocs = append(syms[i].relocs, Reloc{
+				Off:    int(off),
+				Name:   sym.Name,
+				Addend: sym.Offset,
+				Kind:   RelAddr,
+				Siz:    uint8(w),
+			})
+			continue
+		}
+		// A string or rune value ("DATA s+0(SB)/20, $"text"") writes its
+		// bytes into the field and leaves the rest zero, the toolchain's
+		// WriteString: the declared width must hold every byte, and any
+		// width is legal.
+		if s := dd.Value.Imm.Str; s != "" && !dd.Value.Imm.HasVal {
+			text, err := strconv.Unquote(s)
+			if err != nil {
+				return nil, fmt.Errorf("DATA %q: invalid string value %s", dd.Name.Name, s)
 			}
-			i, ok := index[dd.Name.Name]
-			if !ok {
-				return nil, fmt.Errorf("DATA %q: no matching GLOBL", dd.Name.Name)
+			if len(text) > w {
+				return nil, fmt.Errorf("DATA %q: string of %d bytes does not fit width %d", dd.Name.Name, len(text), w)
 			}
-			if dd.Value == nil || !dd.Value.Imm.HasVal {
-				return nil, fmt.Errorf("DATA %q: value must be an integer immediate", dd.Name.Name)
+			copy(buf[off:], text)
+			continue
+		}
+		// A floating-point value stores its IEEE-754 bits: /4 the float32
+		// rounding of the parsed double, /8 the full 64 bits, the
+		// toolchain's WriteFloat32 and WriteFloat64.
+		if f := dd.Value.Imm.Float; f != "" && !dd.Value.Imm.HasVal {
+			num, err := strconv.ParseFloat(f, 64)
+			if err != nil {
+				return nil, fmt.Errorf("DATA %q: invalid floating-point value %q", dd.Name.Name, f)
 			}
-			w := dd.Width
-			switch w {
-			case 1, 2, 4, 8:
-			default:
-				return nil, fmt.Errorf("DATA %q: invalid width %d (want 1, 2, 4 or 8)", dd.Name.Name, w)
-			}
-			off := dd.Name.Offset
-			buf := syms[i].buf
-			if off < 0 || off+int64(w) > int64(len(buf)) {
-				return nil, fmt.Errorf("DATA %q+%d/%d exceeds GLOBL size %d", dd.Name.Name, off, w, len(buf))
-			}
-			v := dd.Value.Imm.Val
 			if dd.Value.Imm.Neg {
-				v = -v
+				num = -num
 			}
-			for j := 0; j < w; j++ {
+			var v uint64
+			switch w {
+			case 4:
+				v = uint64(math.Float32bits(float32(num)))
+			case 8:
+				v = math.Float64bits(num)
+			default:
+				return nil, fmt.Errorf("DATA %q: invalid width %d for a float (want 4 or 8)", dd.Name.Name, w)
+			}
+			for j := range w {
 				buf[off+int64(j)] = byte(v >> (8 * j))
 			}
+			continue
+		}
+		if !dd.Value.Imm.HasVal {
+			return nil, fmt.Errorf("DATA %q: value must be an integer immediate or a symbol address", dd.Name.Name)
+		}
+		switch w {
+		case 1, 2, 4, 8:
+		default:
+			return nil, fmt.Errorf("DATA %q: invalid width %d (want 1, 2, 4 or 8)", dd.Name.Name, w)
+		}
+		v := dd.Value.Imm.Val
+		if dd.Value.Imm.Neg {
+			v = -v
+		}
+		for j := range w {
+			buf[off+int64(j)] = byte(v >> (8 * j))
 		}
 	}
 	return syms, nil

asm/link_test.go

@@ -4,14 +4,19 @@
 package asm
 
 import (
+	"encoding/binary"
+	"fmt"
+	"os"
+	"os/exec"
+	"path/filepath"
 	"strings"
 	"testing"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
 )
 
-// TestAssembleFileStaticData checks the whole-image layout — code, padding
-// and the data section — and that the RIP-relative displacements of static
+// TestAssembleFileStaticData checks the whole-image layout; code, padding
+// and the data section; and that the RIP-relative displacements of static
 // symbol loads resolve to the right bytes.
 func TestAssembleFileStaticData(t *testing.T) {
 	f, errs := parser.Parse("d_amd64.s", `
@@ -128,3 +133,361 @@ DATA x<>+0(SB)/4, $1
 		t.Errorf("single-function SB: error %v, want a file-level-assembly error", err)
 	}
 }
+
+// TestCollectDataNumericFlags pins the numeric GLOBL flag constants from
+// runtime/textflag.h: DUPOK is 2, RODATA is 8, and combinations arrive as
+// one number (9 = NOPROF|RODATA, 10 = RODATA|DUPOK).
+func TestCollectDataNumericFlags(t *testing.T) {
+	tests := []struct {
+		flags  string
+		rodata bool
+		dupok  bool
+	}{
+		{"2", false, true},
+		{"8", true, false},
+		{"9", true, false}, // NOPROF|RODATA, not DUPOK
+		{"10", true, true}, // RODATA|DUPOK
+		{"RODATA", true, false},
+		{"DUPOK", false, true},
+		{"RODATA|DUPOK", true, true},
+	}
+	for _, tt := range tests {
+		src := "TEXT \u00b7f(SB), NOSPLIT, $0\n\tRET\nGLOBL sym(SB), " + tt.flags + ", $8\n"
+		f, errs := parser.Parse("f_amd64.s", src)
+		if len(errs) > 0 {
+			t.Fatalf("parse %q: %v", tt.flags, errs)
+		}
+		img, err := AssembleFile(f)
+		if err != nil {
+			t.Fatalf("assemble %q: %v", tt.flags, err)
+		}
+		if len(img.DataSyms) != 1 {
+			t.Fatalf("%q: data syms = %d, want 1", tt.flags, len(img.DataSyms))
+		}
+		d := img.DataSyms[0]
+		if d.Rodata != tt.rodata || d.Dupok != tt.dupok {
+			t.Errorf("flags %q: rodata=%v dupok=%v, want rodata=%v dupok=%v",
+				tt.flags, d.Rodata, d.Dupok, tt.rodata, tt.dupok)
+		}
+	}
+}
+
+// TestCollectDataSymbolValue covers the symbol-valued DATA field ("DATA
+// s+0(SB)/8, $other(SB)", the rt0 spelling): the field stays zero in the
+// image and the relocation is recorded against the named symbol, whatever
+// the file defines (a TEXT function, a GLOBL) or leaves external.
+func TestCollectDataSymbolValue(t *testing.T) {
+	src := `#include "textflag.h"
+TEXT ·Keep(SB), NOSPLIT, $0-8
+	MOVQ target+0(FP), AX
+	RET
+GLOBL holder(SB), NOPTR, $32
+DATA holder+0(SB)/8, $·Keep(SB)
+DATA holder+8(SB)/8, $·Keep+5(SB)
+DATA holder+16(SB)/8, $holder(SB)
+GLOBL spare(SB), NOPTR, $8
+DATA spare+0(SB)/8, $extvar(SB)
+`
+	f, errs := parser.Parse("f_amd64.s", src)
+	if len(errs) > 0 {
+		t.Fatalf("parse: %v", errs)
+	}
+	img, err := AssembleFile(f)
+	if err != nil {
+		t.Fatalf("assemble: %v", err)
+	}
+	byName := map[string]DataSymbol{}
+	for _, d := range img.DataSyms {
+		byName[d.Name] = d
+	}
+	want := []struct {
+		sym    string
+		off    int
+		name   string
+		addend int64
+		ext    bool
+	}{
+		{"holder", 0, "Keep", 0, false},
+		{"holder", 8, "Keep", 5, false},
+		{"holder", 16, "holder", 0, false},
+		{"spare", 0, "extvar", 0, true},
+	}
+	var flat []struct {
+		sym string
+		r   Reloc
+	}
+	for _, d := range img.DataSyms {
+		for _, r := range d.Relocs {
+			flat = append(flat, struct {
+				sym string
+				r   Reloc
+			}{d.Name, r})
+		}
+	}
+	if len(flat) != len(want) {
+		t.Fatalf("data relocations = %d, want %d", len(flat), len(want))
+	}
+	for i, w := range want {
+		g := flat[i]
+		r := g.r
+		if g.sym != w.sym {
+			t.Errorf("relocation %d sits on %q, want %q", i, g.sym, w.sym)
+			continue
+		}
+		if r.Off != w.off || r.Name != w.name || r.Addend != w.addend || r.External != w.ext {
+			t.Errorf("relocation %d = {+%d %q addend %d ext %v}, want {+%d %q addend %d ext %v}",
+				i, r.Off, r.Name, r.Addend, r.External, w.off, w.name, w.addend, w.ext)
+		}
+		if r.Kind != RelAddr {
+			t.Errorf("relocation %d kind = %v, want RelAddr", i, r.Kind)
+		}
+		if r.Siz != 8 {
+			t.Errorf("relocation %d siz = %d, want 8", i, r.Siz)
+		}
+	}
+	// The fields themselves stay zero: only the linker fills them.
+	for _, b := range img.Data {
+		if b != 0 {
+			t.Fatal("data section is not all zero before relocation")
+		}
+	}
+	if len(img.Externals) != 1 || img.Externals[0] != "extvar" {
+		t.Errorf("Externals = %v, want [extvar]", img.Externals)
+	}
+}
+
+// TestGOObjectDataSymbolReloc pins the GOOBJ record a symbol-valued DATA
+// field produces, against the shape the toolchain emits for the same
+// source: an R_ADDR of the DATA width at the field offset, pkgIdxNone plus
+// the non-package definition index when the target is the file's own TEXT
+// function (the rt0 lib entry spelling).
+func TestGOObjectDataSymbolReloc(t *testing.T) {
+	f, errs := parser.Parse("f_amd64.s", `#include "textflag.h"
+TEXT ·Keep(SB), NOSPLIT, $0-8
+	RET
+GLOBL holder(SB), NOPTR, $16
+DATA holder+0(SB)/8, $·Keep+5(SB)
+`)
+	if len(errs) > 0 {
+		t.Fatalf("parse: %v", errs)
+	}
+	img, err := AssembleFile(f)
+	if err != nil {
+		t.Fatalf("assemble: %v", err)
+	}
+	obj, err := img.GOObject("main", "f_amd64.s")
+	if err != nil {
+		t.Fatalf("GOObject: %v", err)
+	}
+	v := openGoobj(t, obj)
+	// Walk every relocation record; the data record is the one of Siz 8
+	// and type R_ADDR.
+	var off, add int64
+	var pkg, sym uint32
+	found := false
+	for data := v.blk(blkReloc); len(data) >= 23; data = data[23:] {
+		if data[4] != 8 || binary.LittleEndian.Uint16(data[5:]) != relocAddr {
+			continue
+		}
+		found = true
+		off = int64(int32(binary.LittleEndian.Uint32(data[0:])))
+		add = int64(binary.LittleEndian.Uint64(data[7:]))
+		pkg = binary.LittleEndian.Uint32(data[15:])
+		sym = binary.LittleEndian.Uint32(data[19:])
+		break
+	}
+	if !found {
+		t.Fatal("no data relocation record in the object")
+	}
+	if off != 0 || add != 5 {
+		t.Errorf("data reloc = {off %d addend %d}, want {off 0 addend 5}", off, add)
+	}
+	if pkg != pkgIdxNone {
+		t.Errorf("data reloc pkg = %#x, want pkgIdxNone (the TEXT function)", pkg)
+	}
+	// The function's non-package definition index: the four pc tables
+	// precede it, so index 4.
+	if sym != 4 {
+		t.Errorf("data reloc sym = %d, want 4", sym)
+	}
+}
+
+// TestGOObjectDataSymbolLink is the end-to-end proof for symbol-valued DATA
+// fields: the gasm object is substituted for the toolchain's and re-linked,
+// then executed, and the linked data word must hold the real address of the
+// function the DATA line named (runtime.FuncForPC identifies it).
+func TestGOObjectDataSymbolLink(t *testing.T) {
+	goBin, err := exec.LookPath("go")
+	if err != nil {
+		t.Skip("no Go toolchain available")
+	}
+	dir := t.TempDir()
+	asmSrc := `#include "textflag.h"
+GLOBL entry(SB), NOPTR, $8
+DATA entry+0(SB)/8, $·keepme(SB)
+
+TEXT ·keepme(SB), NOSPLIT, $0-0
+	RET
+
+TEXT ·entryptr(SB), NOSPLIT, $0-8
+	MOVQ entry+0(SB), AX
+	MOVQ AX, ret+0(FP)
+	RET
+`
+	if err := os.WriteFile(filepath.Join(dir, "main_amd64.s"), []byte(asmSrc), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	mainSrc := `package main
+
+import "runtime"
+
+func keepme()
+func entryptr() uintptr
+
+func main() {
+	pc := entryptr()
+	fn := runtime.FuncForPC(pc)
+	if fn == nil {
+		panic("the entry word does not point at a function")
+	}
+	if fn.Name() != "main.keepme" {
+		panic("the entry word points at " + fn.Name())
+	}
+}
+`
+	if err := os.WriteFile(filepath.Join(dir, "main.go"), []byte(mainSrc), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(dir, "go.mod"), []byte("module dlink\n\ngo 1.21\n"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	// Capture the build: the package archive's asm object and the link line.
+	build := exec.Command(goBin, "build", "-x", "-work", "-o", filepath.Join(dir, "prog"), ".")
+	build.Dir = dir
+	buildLog, err := build.CombinedOutput()
+	if err != nil {
+		t.Fatalf("baseline build: %v\n%s", err, buildLog)
+	}
+	var work, linkLine, asmObj string
+	for line := range strings.SplitSeq(string(buildLog), "\n") {
+		switch {
+		case strings.HasPrefix(line, "WORK="):
+			work = strings.TrimPrefix(line, "WORK=")
+		case strings.Contains(line, "/asm ") && strings.Contains(line, "main_amd64.s") && !strings.Contains(line, "-gensymabis"):
+			asmObj = fieldAfter(line, "-o")
+		case strings.Contains(line, "/link ") && strings.Contains(line, "-importcfg"):
+			linkLine = line
+		}
+	}
+	if work == "" || asmObj == "" || linkLine == "" {
+		t.Skipf("could not parse build log (work=%q asmObj=%q link=%q)", work, asmObj, linkLine)
+	}
+	defer os.RemoveAll(work)
+	asmObj = strings.ReplaceAll(asmObj, "$WORK", work)
+	linkLine = strings.ReplaceAll(linkLine, "$WORK", work)
+
+	// Assemble the same source with gasm and substitute the object.
+	src, err := os.ReadFile(filepath.Join(dir, "main_amd64.s"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	f, errs := parser.Parse("main_amd64.s", string(src))
+	if len(errs) > 0 {
+		t.Fatalf("parse: %v", errs)
+	}
+	img, err := AssembleFile(f)
+	if err != nil {
+		t.Fatalf("AssembleFile: %v", err)
+	}
+	gasmObj, err := img.GOObject("dlink", "main_amd64.s")
+	if err != nil {
+		t.Fatalf("GOObject: %v", err)
+	}
+	if err := os.WriteFile(asmObj, gasmObj, 0o644); err != nil {
+		t.Fatalf("write gasm object: %v", err)
+	}
+	linkCmd := exec.Command("bash", "-c", "cd "+dir+" && "+linkLine)
+	if out, err := linkCmd.CombinedOutput(); err != nil {
+		t.Fatalf("re-link with gasm object: %v\n%s", err, out)
+	}
+
+	// The linked program must run and find the right function behind the
+	// data word.
+	out, err := exec.Command(filepath.Join(dir, "prog")).CombinedOutput()
+	if err != nil {
+		t.Fatalf("linked program failed: %v\n%s", err, out)
+	}
+}
+
+// TestCollectDataFloatAndStringValues covers the non-integer DATA values the
+// runtime's math and asm files use: floating-point initialisers store their
+// IEEE-754 bits (/4 the float32 rounding, /8 the full double) and string
+// initialisers write their bytes zero-padded within the declared width.
+func TestCollectDataFloatAndStringValues(t *testing.T) {
+	src := `#include "textflag.h"
+TEXT ·Keep(SB), NOSPLIT, $0-8
+	RET
+GLOBL vals<>(SB), RODATA, $44
+DATA vals<>+0(SB)/8, $0.5
+DATA vals<>+8(SB)/8, $-1.0
+DATA vals<>+16(SB)/4, $1.5
+DATA vals<>+20(SB)/16, $"call frame too "
+DATA vals<>+36(SB)/4, $"hi"
+`
+	f, errs := parser.Parse("fvals_amd64.s", src)
+	if len(errs) > 0 {
+		t.Fatalf("parse: %v", errs)
+	}
+	img, err := AssembleFile(f)
+	if err != nil {
+		t.Fatalf("AssembleFile: %v", err)
+	}
+	byName := map[string]DataSymbol{}
+	for _, d := range img.DataSyms {
+		byName[d.Name] = d
+	}
+	d := byName["vals"]
+	if d.Size != 44 {
+		t.Fatalf("vals size = %d, want 44", d.Size)
+	}
+	buf := img.Data[d.Offset : d.Offset+44]
+	// 0.5 = 0x3FE0000000000000, -1.0 = 0xBFF0000000000000 (float64);
+	// 1.5 = 0x3FC00000 (float32).
+	for _, c := range []struct {
+		off  int
+		want []byte
+	}{
+		{0, []byte{0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xE0, 0x3F}},
+		{8, []byte{0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xF0, 0xBF}},
+		{16, []byte{0x00, 0x00, 0xC0, 0x3F}},
+		{20, []byte("call frame too ")},
+		{36, []byte{'h', 'i', 0x00, 0x00}},
+	} {
+		if string(buf[c.off:c.off+len(c.want)]) != string(c.want) {
+			t.Errorf("vals+%d: got % x, want % x", c.off, buf[c.off:c.off+len(c.want)], c.want)
+		}
+	}
+}
+
+// TestCollectDataValueErrors pins the value-kind width rules: a float needs
+// width 4 or 8, a string must fit its declared width, and a bad float
+// literal is diagnosed rather than stored.
+func TestCollectDataValueErrors(t *testing.T) {
+	cases := []string{
+		`GLOBL v<>(SB), RODATA, $4
+DATA v<>+0(SB)/1, $0.5`,
+		`GLOBL v<>(SB), RODATA, $2
+DATA v<>+0(SB)/2, $"toolarge"`,
+	}
+	for i, src := range cases {
+		full := "#include \"textflag.h\"\nTEXT ·Keep(SB), NOSPLIT, $0-8\n\tRET\n" + src
+		f, errs := parser.Parse(fmt.Sprintf("verr%d_amd64.s", i), full)
+		if len(errs) > 0 {
+			t.Fatalf("case %d parse: %v", i, errs)
+		}
+		if _, err := AssembleFile(f); err == nil {
+			t.Errorf("case %d: expected an error, got none", i)
+		}
+	}
+}

asm/loong64_encode.go

@@ -9,7 +9,7 @@ package asm
 // an opcode constant, and the format selects the bit layout.  The opcode
 // constants and formats are transcribed from the Go toolchain's own loong64
 // backend (cmd/internal/obj/loong64), so the emitted bytes match `go tool asm`
-// exactly — the ground-truth oracle for the verify suite.
+// exactly, the ground-truth oracle for the verify suite.
 //
 // All LoongArch instructions are 32 bits, little-endian.  The formats used
 // here (per the LoongArch Volume I specification):
@@ -30,9 +30,14 @@ package asm
 // of the immediate and register fields), mirroring the toolchain's OP_*
 // helpers, so each l64* function only ORs its fields in.
 
+import (
+	"maps"
+	"strings"
+)
+
 // loong64RegNum returns the 5-bit register number for a LoongArch register
-// name: R0–R31 (integer), F0–F31 (floating point), FCC0–FCC7 (condition
-// flags), FCSR0–FCSR31 (control/status) and the ABI aliases the runtime's
+// name: R0-R31 (integer), F0-F31 (floating point), FCC0-FCC7 (condition
+// flags), FCSR0-FCSR31 (control/status) and the ABI aliases the runtime's
 // assembly uses.  Returns -1 for an unrecognised name.
 func loong64RegNum(name string) int {
 	switch name {
@@ -101,12 +106,17 @@ func loong64RegNum(name string) int {
 	case "R31", "S8":
 		return 31
 	}
-	// F0–F31, FCC0–FCC7, FCSR0–FCSR31.
+	// F0-F31, FCC0-FCC7, FCSR0-FCSR31.  The LSX/LASX vector banks (V0-V31,
+	// X0-X31) are deliberately NOT accepted here: they are a separate
+	// register class, and the toolchain rejects V/X names wherever an
+	// integer or FP register is expected (GOARCH=loong64 go tool asm reports
+	// "unrecognized instruction" for `BEQZ X0`).  Vector operands are
+	// resolved only through loong64VecRegNum.
 	if len(name) >= 4 && name[:4] == "FCSR" {
-		return loong64RegSpecial(name[4:], "FCSR", 31)
+		return loong64RegSpecial(name[4:], 31)
 	}
 	if len(name) >= 3 && name[:3] == "FCC" {
-		return loong64RegSpecial(name[3:], "FCC", 7)
+		return loong64RegSpecial(name[3:], 7)
 	}
 	if len(name) < 2 {
 		return -1
@@ -129,7 +139,7 @@ func loong64RegNum(name string) int {
 }
 
 // loong64RegSpecial parses a numbered FCC/FCSR register.
-func loong64RegSpecial(digits, prefix string, max int) int {
+func loong64RegSpecial(digits string, max int) int {
 	if digits == "" {
 		return -1
 	}
@@ -146,6 +156,19 @@ func loong64RegSpecial(digits, prefix string, max int) int {
 	return -1
 }
 
+// loong64VecRegNum resolves an LSX/LASX vector register name (V0-V31 or
+// X0-X31) to its 5-bit number, or -1.  The vector banks are a register class
+// of their own: the toolchain accepts them only in the vector operands of the
+// LSX/LASX instructions (GOARCH=loong64 go tool asm assembles `VADDV V0, V1,
+// V2` and `XVADDV X0, X1, X2`, and rejects `VADDV R4, R5, R6`), so the V/X
+// spellings never reach the integer/FP resolver.
+func loong64VecRegNum(name string) int {
+	if len(name) < 2 || (name[0] != 'V' && name[0] != 'X') {
+		return -1
+	}
+	return loong64RegSpecial(name[1:], 31)
+}
+
 // ---- format helpers ----
 
 // l64rrr encodes a 3R instruction: op | rk<<10 | rj<<5 | rd.
@@ -197,7 +220,9 @@ func l64rrrr(op uint32, r1, r2, r3, r4 int) uint32 {
 }
 
 // l64irir encodes a BSTRINS/BSTRPICK instruction: op | msb<<16 | rj<<5 | lsb<<10 | rd.
-// The msb/lsb fields are 6 bits wide (0–63) and are validated by the caller.
+// The msb/lsb fields are 6 bits wide and are inserted unmasked: the caller
+// must have validated them (0..31 for the .w forms, 0..63 for the .d forms,
+// lsb <= msb), the same rule the toolchain enforces as "illegal bit number".
 func l64irir(op uint32, msb, rj, lsb, rd int) uint32 {
 	return op | uint32(msb)<<16 | uint32(rj&0x1f)<<5 | uint32(lsb)<<10 | uint32(rd&0x1f)
 }
@@ -243,7 +268,7 @@ const (
 	l64Firr14                   // 2RI14 (ldptr/stptr)
 	l64Firr16                   // 2RI16 (addu16i.d)
 	l64Fir20                    // 2RI20 (lu12i.w, lu32i.d, pcalau12i, pcaddu12i)
-	l64Frrrr                    // 4R (fmadd/fmsub/fnmadd/fnmsub)
+	l64Frrrr                    // 4R (fmadd/fmsub/fnmadd/fnmsub, fsel)
 	l64Firir                    // bstrins/bstrpick
 	l64Firrr                    // alsl
 	l64Fi15                     // syscall/break/dbar
@@ -251,6 +276,9 @@ const (
 	l64Frdtime                  // rdtime (rd at bits [9:5], rj at bits [4:0])
 	l64Fshift                   // 2RI12 with a 5/6-bit shift immediate
 	l64Fpreld                   // preld (2RI12 + 5-bit hint)
+	l64Fvvv                     // 3R vector (LSX/LASX): op | vk<<10 | vj<<5 | vd
+	l64Fvcf                     // vector-to-condition: op | subop<<10 | vj<<5 | fcc
+	l64Fvvvv                    // 4R vector shuffle: op | va<<15 | vk<<10 | vj<<5 | vd
 )
 
 // l64Enc is one instruction's encoding: its bit layout (format) and the
@@ -273,12 +301,74 @@ type l64DualEnc struct {
 var l64DualTable = map[string]l64DualEnc{}
 
 // l64InstrTable maps LoongArch mnemonics (as the Go assembler spells them)
-// to their encoding.  SIMD (LSX/LASX: V*/XV*) instructions are not covered
-// yet; the base integer, memory and floating-point ISA is complete.
+// to their encoding.
 var l64InstrTable = map[string]l64Enc{}
 
+// l64Vec3Enc pairs a vector opcode with its register bank: false = LSX
+// (V0-V31), true = LASX (X0-X31).  The toolchain accepts one bank per
+// spelling: GOARCH=loong64 go tool asm assembles `VADDV V1, V2, V3` and
+// `XVADDV X1, X2, X3`, and rejects the crossed spellings.
+type l64Vec3Enc struct {
+	op   uint32
+	lasx bool
+}
+
+// l64VecImmEnc carries the immediate-form encoding of a vector mnemonic:
+// the opcode, the bank, the accepted immediate range, the bias the toolchain
+// adds (vsrai.b encodes imm+8) and the mask of the encoded field (vseqi.b
+// keeps a 5-bit two's-complement value, vseqi.d a 7-bit one).
+type l64VecImmEnc struct {
+	op       uint32
+	lasx     bool
+	min, max int
+	bias     int
+	mask     int
+}
+
+// l64VecBank marks the LSX/LASX mnemonics and records which register bank
+// each accepts; presence in the map routes the mnemonic through the vector
+// dispatcher rather than the integer/FP formats.
+var l64VecBank = map[string]bool{}
+
+// l64VecImmInfo mirrors l64VecImmTable for the dispatcher.
+var l64VecImmInfo = map[string]l64VecImmEnc{}
+
+// l64Vec2R marks the two-operand vector mnemonics (INSTR vj, vd, such as
+// vpcnt.v).
+var l64Vec2R = map[string]bool{}
+
+// l64Vec4R marks the four-operand vector mnemonics (INSTR va, vk, vj, vd,
+// such as vshuf.b).
+var l64Vec4R = map[string]bool{}
+
+// l64VmovqOps holds the VMOVQ/XVMOVQ opcode constants (pre-shifted to bit
+// 15), read off `go tool objdump` of GOARCH=loong64 `go tool asm` kernels.
+type l64VmovqEnc struct {
+	ld, st, ldx, stx           uint32 // plain and indexed load/store
+	replB, replH, replW, replD uint32 // vldrepl: load and replicate element
+	pickS, pickU               uint32 // vpickve2gr.{,u} element extract
+	ins                        uint32 // vinsgr2vr element insert
+	dup                        uint32 // vreplgr2vr duplicate (width in [11:10])
+	move                       uint32 // vori.b/xvori.b $0 register move
+}
+
+var l64VmovqTable = map[bool]l64VmovqEnc{
+	false: { // VMOVQ, the LSX (V) bank
+		ld: 0x5800 << 15, st: 0x5880 << 15, ldx: 0x7080 << 15, stx: 0x7088 << 15,
+		replB: 0x6100 << 15, replH: 0x6080 << 15, replW: 0x6040 << 15, replD: 0x6020 << 15,
+		pickS: 0xE5DF << 15, pickU: 0xE5E7 << 15,
+		ins: 0xE5D7 << 15, dup: 0xE53E << 15, move: 0xE65A << 15,
+	},
+	true: { // XVMOVQ, the LASX (X) bank
+		ld: 0x5900 << 15, st: 0x5980 << 15, ldx: 0x7090 << 15, stx: 0x7098 << 15,
+		replB: 0x6500 << 15, replH: 0x6480 << 15, replW: 0x6440 << 15, replD: 0x6420 << 15,
+		pickS: 0xEDDF << 15, pickU: 0xEDE7 << 15,
+		ins: 0xEDD7 << 15, dup: 0xED3E << 15, move: 0xEE5A << 15,
+	},
+}
+
 func init() {
-	// 3R — integer.
+	// 3R, integer.
 	rrr := map[string]uint32{
 		"ADD": 0x20 << 15, "ADDW": 0x20 << 15, "ADDV": 0x21 << 15, "ADDVU": 0x21 << 15,
 		"SUB": 0x22 << 15, "SUBW": 0x22 << 15, "SUBV": 0x23 << 15, "SUBVU": 0x23 << 15,
@@ -298,7 +388,7 @@ func init() {
 		"CRCWBW": 0x48 << 15, "CRCWHW": 0x49 << 15, "CRCWWW": 0x4a << 15, "CRCWVW": 0x4b << 15,
 		"CRCCWBW": 0x4c << 15, "CRCCWHW": 0x4d << 15, "CRCCWWW": 0x4e << 15, "CRCCWVW": 0x4f << 15,
 	}
-	// 3R — floating point.
+	// 3R, floating point.
 	rrr["MULF"] = 0x209 << 15
 	rrr["MULD"] = 0x20a << 15
 	rrr["DIVF"] = 0x20d << 15
@@ -356,6 +446,18 @@ func init() {
 		"FTINTRZVF": 0x46a9 << 10, "FTINTRZVD": 0x46aa << 10,
 		"FTINTRNEWF": 0x46b1 << 10, "FTINTRNEWD": 0x46b2 << 10,
 		"FTINTRNEVF": 0x46b9 << 10, "FTINTRNEVD": 0x46ba << 10,
+		// LSX: convert a 64-bit integer lane to a double float.  The operand
+		// bank is the FP registers (the toolchain spells it `FFINTDV F0, F1`),
+		// so the entry stays on the 2R integer/FP format.
+		"FFINTDV": 0x474a << 10,
+		// The rest of the scalar conversions (all F-bank, 2R).
+		"FFINTFW": 0x4744 << 10, // ffint.s.w
+		"FFINTFV": 0x4746 << 10, // ffint.s.l
+		"FFINTDW": 0x4748 << 10, // ffint.d.w
+		"FTINTWF": 0x46c1 << 10, // ftint.w.s
+		"FTINTWD": 0x46c2 << 10, // ftint.w.d
+		"FTINTVF": 0x46c9 << 10, // ftint.l.s
+		"FTINTVD": 0x46ca << 10, // ftint.l.d
 	}
 	for m, op := range rr {
 		l64InstrTable[m] = l64Enc{format: l64Frr, op: op}
@@ -368,7 +470,7 @@ func init() {
 	// The dual-form arithmetic mnemonics (register 3R + immediate 2RI12),
 	// selected by the operand kind; the shift mnemonics pair the 3R form
 	// with a 5/6-bit shift immediate.
-	for m, e := range map[string]l64DualEnc{
+	maps.Copy(l64DualTable, map[string]l64DualEnc{
 		"ADD":   {rrr: 0x20 << 15, imm: 0x00a << 22},
 		"ADDW":  {rrr: 0x20 << 15, imm: 0x00a << 22},
 		"ADDV":  {rrr: 0x21 << 15, imm: 0x00b << 22},
@@ -386,16 +488,14 @@ func init() {
 		"SRLV":  {rrr: 0x32 << 15, imm: 0x0045 << 16, shift: true},
 		"SRAV":  {rrr: 0x33 << 15, imm: 0x0049 << 16, shift: true},
 		"ROTRV": {rrr: 0x37 << 15, imm: 0x004d << 16, shift: true},
-	} {
-		l64DualTable[m] = e
-	}
+	})
 
-	// 2RI12 — pure immediate arithmetic (LU52ID has no register form).
+	// 2RI12, pure immediate arithmetic (LU52ID has no register form).
 	l64InstrTable["LU52ID"] = l64Enc{format: l64Firr, op: 0x00c << 22}
 	// ADDV16 (addu16i.d): 2RI16 with the immediate shifted right by 16.
 	l64InstrTable["ADDV16"] = l64Enc{format: l64Firr16, op: 0x4 << 26}
 
-	// 2RI14 — LL/SC are aliased by the Go assembler to the pointer loads and
+	// 2RI14, LL/SC are aliased by the Go assembler to the pointer loads and
 	// stores (ldptr/stptr), with the offset scaled by 4.
 	l64InstrTable["MOVWP"] = l64Enc{format: l64Firr14, op: 0x25 << 24} // stptr.w
 	l64InstrTable["MOVVP"] = l64Enc{format: l64Firr14, op: 0x27 << 24} // stptr.d
@@ -414,18 +514,20 @@ func init() {
 	// LUI is the Plan 9 spelling of lu12i.w.
 	l64InstrTable["LUI"] = l64Enc{format: l64Fir20, op: 0x0a << 25}
 
-	// 4R — fused multiply-add.
+	// 4R, fused multiply-add, and FSEL (fsel.d: the first operand is a FCC
+	// condition flag, the layout matches the 4R shape).
 	rrrr := map[string]uint32{
 		"FMADDF": 0x81 << 20, "FMADDD": 0x82 << 20,
 		"FMSUBF": 0x85 << 20, "FMSUBD": 0x86 << 20,
 		"FNMADDF": 0x89 << 20, "FNMADDD": 0x8a << 20,
 		"FNMSUBF": 0x8d << 20, "FNMSUBD": 0x8e << 20,
+		"FSEL": 0x340 << 18,
 	}
 	for m, op := range rrrr {
 		l64InstrTable[m] = l64Enc{format: l64Frrrr, op: op}
 	}
 
-	// IRIR — bit-field insert/extract.
+	// IRIR, bit-field insert/extract.
 	irir := map[string]uint32{
 		"BSTRINSW":  0x3<<21 | 0x0<<15,
 		"BSTRINSV":  0x2 << 22,
@@ -436,7 +538,7 @@ func init() {
 		l64InstrTable[m] = l64Enc{format: l64Firir, op: op}
 	}
 
-	// 3RI2 — ALSL.
+	// 3RI2, ALSL.
 	irrr := map[string]uint32{
 		"ALSLW": 0x2 << 17, "ALSLWU": 0x3 << 17, "ALSLV": 0x16 << 17,
 	}
@@ -452,7 +554,11 @@ func init() {
 	// PRELD.
 	l64InstrTable["PRELD"] = l64Enc{format: l64Fpreld, op: 0x0ab << 22}
 
-	// Atomics — 3R with the AM field order (rk=value, rj=address, rd=result).
+	// Atomics, 3R with the AM field order (rk=value, rj=address, rd=result).
+	// The toolchain's form is three operands, `AMADDW rk, (rj), rd`
+	// (cmd/asm/internal/asm/testdata/loong64enc1.s and
+	// internal/runtime/atomic/atomic_loong64.s); the two-register spelling
+	// is rejected by the oracle.
 	am := map[string]uint32{
 		"AMSWAPB": 0x070B8 << 15, "AMSWAPH": 0x070B9 << 15,
 		"AMSWAPW": 0x070C0 << 15, "AMSWAPV": 0x070C1 << 15,
@@ -470,14 +576,512 @@ func init() {
 		"AMSWAPDBW": 0x070D2 << 15, "AMSWAPDBV": 0x070D3 << 15,
 		"AMCASDBB": 0x070B4 << 15, "AMCASDBH": 0x070B5 << 15,
 		"AMCASDBW": 0x070B6 << 15, "AMCASDBV": 0x070B7 << 15,
+		// The _dbar (acquire/release) add, and, or variants: opcodes read off
+		// `go tool objdump` of `AMADDDBW R14, (R13), R12` and friends.
+		"AMADDDBW": 0x070D4 << 15, "AMADDDBV": 0x070D5 << 15,
+		"AMANDDBW": 0x070D6 << 15, "AMANDDBV": 0x070D7 << 15,
+		"AMORDBW": 0x070D8 << 15, "AMORDBV": 0x070D9 << 15,
+		// The remaining _dbar exchange variants (loong64enc1.s).
+		"AMXORDBW": 0x070DA << 15, "AMXORDBV": 0x070DB << 15,
+		"AMMAXDBW": 0x070DC << 15, "AMMAXDBV": 0x070DD << 15,
+		"AMMINDBW": 0x070DE << 15, "AMMINDBV": 0x070DF << 15,
+		"AMMAXDBWU": 0x070E0 << 15, "AMMAXDBVU": 0x070E1 << 15,
+		"AMMINDBWU": 0x070E2 << 15, "AMMINDBVU": 0x070E3 << 15,
 	}
 	for m, op := range am {
 		l64InstrTable[m] = l64Enc{format: l64Fam, op: op}
 	}
+
+	// ---- LSX/LASX (V*/XV*) ----
+	// Every opcode below was read off `go tool objdump` of a GOARCH=loong64
+	// `go tool asm` kernel (the toolchain's own loong64enc1.s cross-checks
+	// most of them), not assumed from the LoongArch manual.
+
+	// Three vector registers: INSTR vk, vj, vd (or INSTR vk, vd with
+	// vj = vd).  l64Vec3Enc.lasx selects the register bank the toolchain
+	// accepts: LSX spellings take V0-V31, LASX spellings X0-X31.
+	vec3 := map[string]l64Vec3Enc{
+		"VADDW": {0xE016 << 15, false}, "VADDV": {0xE017 << 15, false},
+		"VANDV": {0xE24C << 15, false}, "VXORV": {0xE24E << 15, false},
+		"VSEQB": {0xE000 << 15, false}, "VSEQV": {0xE003 << 15, false},
+		"VSRAB": {0xE1D8 << 15, false}, "VROTRW": {0xE1DE << 15, false},
+		"XVADDV": {0xE817 << 15, true},
+		"XVANDV": {0xEA4C << 15, true}, "XVXORV": {0xEA4E << 15, true},
+		"XVSEQB": {0xE800 << 15, true}, "XVSEQV": {0xE803 << 15, true},
+	}
+
+	// The integer and FP add/subtract families: [X]VADD and [X]VSUB by lane
+	// width, plus the [X]VSADD/[X]VSSUB saturating pairs.
+	// Opcodes transcribed from the toolchain's loong64enc1.s.
+	addsub := map[string]l64Vec3Enc{
+		"VADDB": {0xE014 << 15, false}, "VADDH": {0xE015 << 15, false},
+		"VADDD": {0xE262 << 15, false}, "VADDF": {0xE261 << 15, false},
+		"VADDQ": {0xE25A << 15, false},
+		"VSUBB": {0xE018 << 15, false}, "VSUBH": {0xE019 << 15, false},
+		"VSUBW": {0xE01A << 15, false}, "VSUBV": {0xE01B << 15, false},
+		"VSUBQ": {0xE25B << 15, false},
+		"VSUBF": {0xE265 << 15, false}, "VSUBD": {0xE266 << 15, false},
+		"VSADDB": {0xE08C << 15, false}, "VSADDH": {0xE08D << 15, false},
+		"VSADDW": {0xE08E << 15, false}, "VSADDV": {0xE08F << 15, false},
+		"VSADDBU": {0xE094 << 15, false}, "VSADDHU": {0xE095 << 15, false},
+		"VSADDWU": {0xE096 << 15, false}, "VSADDVU": {0xE097 << 15, false},
+		"VSSUBB": {0xE090 << 15, false}, "VSSUBH": {0xE091 << 15, false},
+		"VSSUBW": {0xE092 << 15, false}, "VSSUBV": {0xE093 << 15, false},
+		"VSSUBBU": {0xE098 << 15, false}, "VSSUBHU": {0xE099 << 15, false},
+		"VSSUBWU": {0xE09A << 15, false}, "VSSUBVU": {0xE09B << 15, false},
+		"XVADDB": {0xE814 << 15, true}, "XVADDH": {0xE815 << 15, true},
+		"XVADDW": {0xE816 << 15, true},
+		"XVADDD": {0xEA62 << 15, true}, "XVADDF": {0xEA61 << 15, true},
+		"XVADDQ": {0xEA5A << 15, true},
+		"XVSUBB": {0xE818 << 15, true}, "XVSUBH": {0xE819 << 15, true},
+		"XVSUBW": {0xE81A << 15, true}, "XVSUBV": {0xE81B << 15, true},
+		"XVSUBQ": {0xEA5B << 15, true},
+		"XVSUBF": {0xEA65 << 15, true}, "XVSUBD": {0xEA66 << 15, true},
+		"XVSADDB": {0xE88C << 15, true}, "XVSADDH": {0xE88D << 15, true},
+		"XVSADDW": {0xE88E << 15, true}, "XVSADDV": {0xE88F << 15, true},
+		"XVSADDBU": {0xE894 << 15, true}, "XVSADDHU": {0xE895 << 15, true},
+		"XVSADDWU": {0xE896 << 15, true}, "XVSADDVU": {0xE897 << 15, true},
+		"XVSSUBB": {0xE890 << 15, true}, "XVSSUBH": {0xE891 << 15, true},
+		"XVSSUBW": {0xE892 << 15, true}, "XVSSUBV": {0xE893 << 15, true},
+		"XVSSUBBU": {0xE898 << 15, true}, "XVSSUBHU": {0xE899 << 15, true},
+		"XVSSUBWU": {0xE89A << 15, true}, "XVSSUBVU": {0xE89B << 15, true},
+	}
+
+	// The multiply families: plain and high-half [X]VMUL/[X]VMUH, the
+	// widening [X]VMULW{EV,OD} ladder and its accumulating [X]VMADDW twins,
+	// plus the [X]VMADD/[X]VMSUB fused multiply-add and the [X]VDIV/[X]VMOD
+	// divide and modulo pairs.
+	muldiv := map[string]l64Vec3Enc{
+		"VMULB": {0xE108 << 15, false}, "VMULH": {0xE109 << 15, false},
+		"VMULW": {0xE10A << 15, false}, "VMULV": {0xE10B << 15, false},
+		"VMUHB": {0xE10C << 15, false}, "VMUHH": {0xE10D << 15, false},
+		"VMUHW": {0xE10E << 15, false}, "VMUHV": {0xE10F << 15, false},
+		"VMUHBU": {0xE110 << 15, false}, "VMUHHU": {0xE111 << 15, false},
+		"VMUHWU": {0xE112 << 15, false}, "VMUHVU": {0xE113 << 15, false},
+		"VMULWEVHB": {0xE120 << 15, false}, "VMULWEVWH": {0xE121 << 15, false},
+		"VMULWEVVW": {0xE122 << 15, false}, "VMULWEVQV": {0xE123 << 15, false},
+		"VMULWODHB": {0xE124 << 15, false}, "VMULWODWH": {0xE125 << 15, false},
+		"VMULWODVW": {0xE126 << 15, false}, "VMULWODQV": {0xE127 << 15, false},
+		"VMULWEVHBU": {0xE130 << 15, false}, "VMULWEVWHU": {0xE131 << 15, false},
+		"VMULWEVVWU": {0xE132 << 15, false}, "VMULWEVQVU": {0xE133 << 15, false},
+		"VMULWODHBU": {0xE134 << 15, false}, "VMULWODWHU": {0xE135 << 15, false},
+		"VMULWODVWU": {0xE136 << 15, false}, "VMULWODQVU": {0xE137 << 15, false},
+		"VMULWEVHBUB": {0xE140 << 15, false}, "VMULWEVWHUH": {0xE141 << 15, false},
+		"VMULWEVVWUW": {0xE142 << 15, false}, "VMULWEVQVUV": {0xE143 << 15, false},
+		"VMULWODHBUB": {0xE144 << 15, false}, "VMULWODWHUH": {0xE145 << 15, false},
+		"VMULWODVWUW": {0xE146 << 15, false}, "VMULWODQVUV": {0xE147 << 15, false},
+		"VMADDB": {0xE150 << 15, false}, "VMADDH": {0xE151 << 15, false},
+		"VMADDW": {0xE152 << 15, false}, "VMADDV": {0xE153 << 15, false},
+		"VMSUBB": {0xE154 << 15, false}, "VMSUBH": {0xE155 << 15, false},
+		"VMSUBW": {0xE156 << 15, false}, "VMSUBV": {0xE157 << 15, false},
+		"VMADDWEVHB": {0xE158 << 15, false}, "VMADDWEVWH": {0xE159 << 15, false},
+		"VMADDWEVVW": {0xE15A << 15, false}, "VMADDWEVQV": {0xE15B << 15, false},
+		"VMADDWODHB": {0xE15C << 15, false}, "VMADDWODWH": {0xE15D << 15, false},
+		"VMADDWODVW": {0xE15E << 15, false}, "VMADDWODQV": {0xE15F << 15, false},
+		"VMADDWEVHBU": {0xE168 << 15, false}, "VMADDWEVWHU": {0xE169 << 15, false},
+		"VMADDWEVVWU": {0xE16A << 15, false}, "VMADDWEVQVU": {0xE16B << 15, false},
+		"VMADDWODHBU": {0xE16C << 15, false}, "VMADDWODWHU": {0xE16D << 15, false},
+		"VMADDWODVWU": {0xE16E << 15, false}, "VMADDWODQVU": {0xE16F << 15, false},
+		"VMADDWEVHBUB": {0xE178 << 15, false}, "VMADDWEVWHUH": {0xE179 << 15, false},
+		"VMADDWEVVWUW": {0xE17A << 15, false}, "VMADDWEVQVUV": {0xE17B << 15, false},
+		"VMADDWODHBUB": {0xE17C << 15, false}, "VMADDWODWHUH": {0xE17D << 15, false},
+		"VMADDWODVWUW": {0xE17E << 15, false}, "VMADDWODQVUV": {0xE17F << 15, false},
+		"VDIVB": {0xE1C0 << 15, false}, "VDIVH": {0xE1C1 << 15, false},
+		"VDIVW": {0xE1C2 << 15, false}, "VDIVV": {0xE1C3 << 15, false},
+		"VMODB": {0xE1C4 << 15, false}, "VMODH": {0xE1C5 << 15, false},
+		"VMODW": {0xE1C6 << 15, false}, "VMODV": {0xE1C7 << 15, false},
+		"VDIVBU": {0xE1C8 << 15, false}, "VDIVHU": {0xE1C9 << 15, false},
+		"VDIVWU": {0xE1CA << 15, false}, "VDIVVU": {0xE1CB << 15, false},
+		"VMODBU": {0xE1CC << 15, false}, "VMODHU": {0xE1CD << 15, false},
+		"VMODWU": {0xE1CE << 15, false}, "VMODVU": {0xE1CF << 15, false},
+		"VMULF": {0xE271 << 15, false}, "VMULD": {0xE272 << 15, false},
+		"VDIVF": {0xE275 << 15, false}, "VDIVD": {0xE276 << 15, false},
+		"XVMULB": {0xE908 << 15, true}, "XVMULH": {0xE909 << 15, true},
+		"XVMULW": {0xE90A << 15, true}, "XVMULV": {0xE90B << 15, true},
+		"XVMUHB": {0xE90C << 15, true}, "XVMUHH": {0xE90D << 15, true},
+		"XVMUHW": {0xE90E << 15, true}, "XVMUHV": {0xE90F << 15, true},
+		"XVMUHBU": {0xE910 << 15, true}, "XVMUHHU": {0xE911 << 15, true},
+		"XVMUHWU": {0xE912 << 15, true}, "XVMUHVU": {0xE913 << 15, true},
+		"XVMULWEVHB": {0xE920 << 15, true}, "XVMULWEVWH": {0xE921 << 15, true},
+		"XVMULWEVVW": {0xE922 << 15, true}, "XVMULWEVQV": {0xE923 << 15, true},
+		"XVMULWODHB": {0xE924 << 15, true}, "XVMULWODWH": {0xE925 << 15, true},
+		"XVMULWODVW": {0xE926 << 15, true}, "XVMULWODQV": {0xE927 << 15, true},
+		"XVMULWEVHBU": {0xE930 << 15, true}, "XVMULWEVWHU": {0xE931 << 15, true},
+		"XVMULWEVVWU": {0xE932 << 15, true}, "XVMULWEVQVU": {0xE933 << 15, true},
+		"XVMULWODHBU": {0xE934 << 15, true}, "XVMULWODWHU": {0xE935 << 15, true},
+		"XVMULWODVWU": {0xE936 << 15, true}, "XVMULWODQVU": {0xE937 << 15, true},
+		"XVMULWEVHBUB": {0xE940 << 15, true}, "XVMULWEVWHUH": {0xE941 << 15, true},
+		"XVMULWEVVWUW": {0xE942 << 15, true}, "XVMULWEVQVUV": {0xE943 << 15, true},
+		"XVMULWODHBUB": {0xE944 << 15, true}, "XVMULWODWHUH": {0xE945 << 15, true},
+		"XVMULWODVWUW": {0xE946 << 15, true}, "XVMULWODQVUV": {0xE947 << 15, true},
+		"XVMADDB": {0xE950 << 15, true}, "XVMADDH": {0xE951 << 15, true},
+		"XVMADDW": {0xE952 << 15, true}, "XVMADDV": {0xE953 << 15, true},
+		"XVMSUBB": {0xE954 << 15, true}, "XVMSUBH": {0xE955 << 15, true},
+		"XVMSUBW": {0xE956 << 15, true}, "XVMSUBV": {0xE957 << 15, true},
+		"XVMADDWEVHB": {0xE958 << 15, true}, "XVMADDWEVWH": {0xE959 << 15, true},
+		"XVMADDWEVVW": {0xE95A << 15, true}, "XVMADDWEVQV": {0xE95B << 15, true},
+		"XVMADDWODHB": {0xE95C << 15, true}, "XVMADDWODWH": {0xE95D << 15, true},
+		"XVMADDWODVW": {0xE95E << 15, true}, "XVMADDWODQV": {0xE95F << 15, true},
+		"XVMADDWEVHBU": {0xE968 << 15, true}, "XVMADDWEVWHU": {0xE969 << 15, true},
+		"XVMADDWEVVWU": {0xE96A << 15, true}, "XVMADDWEVQVU": {0xE96B << 15, true},
+		"XVMADDWODHBU": {0xE96C << 15, true}, "XVMADDWODWHU": {0xE96D << 15, true},
+		"XVMADDWODVWU": {0xE96E << 15, true}, "XVMADDWODQVU": {0xE96F << 15, true},
+		"XVMADDWEVHBUB": {0xE978 << 15, true}, "XVMADDWEVWHUH": {0xE979 << 15, true},
+		"XVMADDWEVVWUW": {0xE97A << 15, true}, "XVMADDWEVQVUV": {0xE97B << 15, true},
+		"XVMADDWODHBUB": {0xE97C << 15, true}, "XVMADDWODWHUH": {0xE97D << 15, true},
+		"XVMADDWODVWUW": {0xE97E << 15, true}, "XVMADDWODQVUV": {0xE97F << 15, true},
+		"XVDIVB": {0xE9C0 << 15, true}, "XVDIVH": {0xE9C1 << 15, true},
+		"XVDIVW": {0xE9C2 << 15, true}, "XVDIVV": {0xE9C3 << 15, true},
+		"XVMODB": {0xE9C4 << 15, true}, "XVMODH": {0xE9C5 << 15, true},
+		"XVMODW": {0xE9C6 << 15, true}, "XVMODV": {0xE9C7 << 15, true},
+		"XVDIVBU": {0xE9C8 << 15, true}, "XVDIVHU": {0xE9C9 << 15, true},
+		"XVDIVWU": {0xE9CA << 15, true}, "XVDIVVU": {0xE9CB << 15, true},
+		"XVMODBU": {0xE9CC << 15, true}, "XVMODHU": {0xE9CD << 15, true},
+		"XVMODWU": {0xE9CE << 15, true}, "XVMODVU": {0xE9CF << 15, true},
+		"XVMULF": {0xEA71 << 15, true}, "XVMULD": {0xEA72 << 15, true},
+		"XVDIVF": {0xEA75 << 15, true}, "XVDIVD": {0xEA76 << 15, true},
+	}
+
+	// The lane-wise shifts and rotates (three-register forms; the immediate
+	// forms live in l64VecImmInfo), the interleave families, the bit
+	// clear/set/rev register forms, the remaining logic and compare
+	// spellings, the widening add/subtract ladder and the vector FP
+	// arithmetic.
+	vecmisc := map[string]l64Vec3Enc{
+		"VSLLB": {0xE1D0 << 15, false}, "VSLLH": {0xE1D1 << 15, false},
+		"VSLLW": {0xE1D2 << 15, false}, "VSLLV": {0xE1D3 << 15, false},
+		"VSRLB": {0xE1D4 << 15, false}, "VSRLH": {0xE1D5 << 15, false},
+		"VSRLW": {0xE1D6 << 15, false}, "VSRLV": {0xE1D7 << 15, false},
+		"VSRAH": {0xE1D9 << 15, false}, "VSRAW": {0xE1DA << 15, false},
+		"VSRAV":  {0xE1DB << 15, false},
+		"VROTRB": {0xE1DC << 15, false}, "VROTRH": {0xE1DD << 15, false},
+		"VROTRV": {0xE1DF << 15, false},
+		"VILVLB": {0xE234 << 15, false}, "VILVLH": {0xE235 << 15, false},
+		"VILVLW": {0xE236 << 15, false}, "VILVLV": {0xE237 << 15, false},
+		"VILVHB": {0xE238 << 15, false}, "VILVHH": {0xE239 << 15, false},
+		"VILVHW": {0xE23A << 15, false}, "VILVHV": {0xE23B << 15, false},
+		"VBITCLRB": {0xE218 << 15, false}, "VBITCLRH": {0xE219 << 15, false},
+		"VBITCLRW": {0xE21A << 15, false}, "VBITCLRV": {0xE21B << 15, false},
+		"VBITSETB": {0xE21C << 15, false}, "VBITSETH": {0xE21D << 15, false},
+		"VBITSETW": {0xE21E << 15, false}, "VBITSETV": {0xE21F << 15, false},
+		"VBITREVB": {0xE220 << 15, false}, "VBITREVH": {0xE221 << 15, false},
+		"VBITREVW": {0xE222 << 15, false}, "VBITREVV": {0xE223 << 15, false},
+		"VORV": {0xE24D << 15, false}, "VNORV": {0xE24F << 15, false},
+		"VANDNV": {0xE250 << 15, false}, "VORNV": {0xE251 << 15, false},
+		"VSEQH": {0xE001 << 15, false}, "VSEQW": {0xE002 << 15, false},
+		"VSLTB": {0xE00C << 15, false}, "VSLTH": {0xE00D << 15, false},
+		"VSLTW": {0xE00E << 15, false}, "VSLTV": {0xE00F << 15, false},
+		"VSLTBU": {0xE010 << 15, false}, "VSLTHU": {0xE011 << 15, false},
+		"VSLTWU": {0xE012 << 15, false}, "VSLTVU": {0xE013 << 15, false},
+		"VADDWEVHB": {0xE03C << 15, false}, "VADDWEVWH": {0xE03D << 15, false},
+		"VADDWEVVW": {0xE03E << 15, false}, "VADDWEVQV": {0xE03F << 15, false},
+		"VSUBWEVHB": {0xE040 << 15, false}, "VSUBWEVWH": {0xE041 << 15, false},
+		"VSUBWEVVW": {0xE042 << 15, false}, "VSUBWEVQV": {0xE043 << 15, false},
+		"VADDWODHB": {0xE044 << 15, false}, "VADDWODWH": {0xE045 << 15, false},
+		"VADDWODVW": {0xE046 << 15, false}, "VADDWODQV": {0xE047 << 15, false},
+		"VSUBWODHB": {0xE048 << 15, false}, "VSUBWODWH": {0xE049 << 15, false},
+		"VSUBWODVW": {0xE04A << 15, false}, "VSUBWODQV": {0xE04B << 15, false},
+		"VSUBWEVHBU": {0xE060 << 15, false}, "VSUBWEVWHU": {0xE061 << 15, false},
+		"VSUBWEVVWU": {0xE062 << 15, false}, "VSUBWEVQVU": {0xE063 << 15, false},
+		"VADDWEVHBU": {0xE05C << 15, false}, "VADDWEVWHU": {0xE05D << 15, false},
+		"VADDWEVVWU": {0xE05E << 15, false}, "VADDWEVQVU": {0xE05F << 15, false},
+		"VADDWODHBU": {0xE064 << 15, false}, "VADDWODWHU": {0xE065 << 15, false},
+		"VADDWODVWU": {0xE066 << 15, false}, "VADDWODQVU": {0xE067 << 15, false},
+		"VSUBWODHBU": {0xE068 << 15, false}, "VSUBWODWHU": {0xE069 << 15, false},
+		"VSUBWODVWU": {0xE06A << 15, false}, "VSUBWODQVU": {0xE06B << 15, false},
+		"VSHUFH": {0xE2F5 << 15, false}, "VSHUFW": {0xE2F6 << 15, false},
+		"VSHUFV": {0xE2F7 << 15, false},
+		"XVSLLB": {0xE9D0 << 15, true}, "XVSLLH": {0xE9D1 << 15, true},
+		"XVSLLW": {0xE9D2 << 15, true}, "XVSLLV": {0xE9D3 << 15, true},
+		"XVSRLB": {0xE9D4 << 15, true}, "XVSRLH": {0xE9D5 << 15, true},
+		"XVSRLW": {0xE9D6 << 15, true}, "XVSRLV": {0xE9D7 << 15, true},
+		"XVSRAB": {0xE9D8 << 15, true}, "XVSRAH": {0xE9D9 << 15, true},
+		"XVSRAW": {0xE9DA << 15, true}, "XVSRAV": {0xE9DB << 15, true},
+		"XVROTRB": {0xE9DC << 15, true}, "XVROTRH": {0xE9DD << 15, true},
+		"XVROTRW": {0xE9DE << 15, true}, "XVROTRV": {0xE9DF << 15, true},
+		"XVILVLB": {0xEA34 << 15, true}, "XVILVLH": {0xEA35 << 15, true},
+		"XVILVLW": {0xEA36 << 15, true}, "XVILVLV": {0xEA37 << 15, true},
+		"XVILVHB": {0xEA38 << 15, true}, "XVILVHH": {0xEA39 << 15, true},
+		"XVILVHW": {0xEA3A << 15, true}, "XVILVHV": {0xEA3B << 15, true},
+		"XVBITCLRB": {0xEA18 << 15, true}, "XVBITCLRH": {0xEA19 << 15, true},
+		"XVBITCLRW": {0xEA1A << 15, true}, "XVBITCLRV": {0xEA1B << 15, true},
+		"XVBITSETB": {0xEA1C << 15, true}, "XVBITSETH": {0xEA1D << 15, true},
+		"XVBITSETW": {0xEA1E << 15, true}, "XVBITSETV": {0xEA1F << 15, true},
+		"XVBITREVB": {0xEA20 << 15, true}, "XVBITREVH": {0xEA21 << 15, true},
+		"XVBITREVW": {0xEA22 << 15, true}, "XVBITREVV": {0xEA23 << 15, true},
+		"XVORV": {0xEA4D << 15, true}, "XVNORV": {0xEA4F << 15, true},
+		"XVANDNV": {0xEA50 << 15, true}, "XVORNV": {0xEA51 << 15, true},
+		"XVSEQH": {0xE801 << 15, true}, "XVSEQW": {0xE802 << 15, true},
+		"XVSLTB": {0xE80C << 15, true}, "XVSLTH": {0xE80D << 15, true},
+		"XVSLTW": {0xE80E << 15, true}, "XVSLTV": {0xE80F << 15, true},
+		"XVSLTBU": {0xE810 << 15, true}, "XVSLTHU": {0xE811 << 15, true},
+		"XVSLTWU": {0xE812 << 15, true}, "XVSLTVU": {0xE813 << 15, true},
+		"XVADDWEVHB": {0xE83C << 15, true}, "XVADDWEVWH": {0xE83D << 15, true},
+		"XVADDWEVVW": {0xE83E << 15, true}, "XVADDWEVQV": {0xE83F << 15, true},
+		"XVSUBWEVHB": {0xE840 << 15, true}, "XVSUBWEVWH": {0xE841 << 15, true},
+		"XVSUBWEVVW": {0xE842 << 15, true}, "XVSUBWEVQV": {0xE843 << 15, true},
+		"XVADDWODHB": {0xE844 << 15, true}, "XVADDWODWH": {0xE845 << 15, true},
+		"XVADDWODVW": {0xE846 << 15, true}, "XVADDWODQV": {0xE847 << 15, true},
+		"XVSUBWODHB": {0xE848 << 15, true}, "XVSUBWODWH": {0xE849 << 15, true},
+		"XVSUBWODVW": {0xE84A << 15, true}, "XVSUBWODQV": {0xE84B << 15, true},
+		"XVADDWEVHBU": {0xE85C << 15, true}, "XVADDWEVWHU": {0xE85D << 15, true},
+		"XVADDWEVVWU": {0xE85E << 15, true}, "XVADDWEVQVU": {0xE85F << 15, true},
+		"XVSUBWEVHBU": {0xE860 << 15, true}, "XVSUBWEVWHU": {0xE861 << 15, true},
+		"XVSUBWEVVWU": {0xE862 << 15, true}, "XVSUBWEVQVU": {0xE863 << 15, true},
+		"XVADDWODHBU": {0xE864 << 15, true}, "XVADDWODWHU": {0xE865 << 15, true},
+		"XVADDWODVWU": {0xE866 << 15, true}, "XVADDWODQVU": {0xE867 << 15, true},
+		"XVSUBWODHBU": {0xE868 << 15, true}, "XVSUBWODWHU": {0xE869 << 15, true},
+		"XVSUBWODVWU": {0xE86A << 15, true}, "XVSUBWODQVU": {0xE86B << 15, true},
+		"XVSHUFH": {0xEAF5 << 15, true}, "XVSHUFW": {0xEAF6 << 15, true},
+		"XVSHUFV": {0xEAF7 << 15, true},
+	}
+	for _, tab := range []map[string]l64Vec3Enc{addsub, muldiv, vecmisc} {
+		for m, e := range tab {
+			if _, dup := vec3[m]; dup {
+				panic("loong64: duplicate vector mnemonic " + m)
+			}
+			vec3[m] = e
+		}
+	}
+	for m, e := range vec3 {
+		l64InstrTable[m] = l64Enc{format: l64Fvvv, op: e.op}
+		l64VecBank[m] = e.lasx
+	}
+
+	// Immediate forms: INSTR $imm, vj, vd (or INSTR $imm, vd).  The immediate
+	// range, bias and field mask are the ones the toolchain encodes: vandi.b
+	// stores the raw 8-bit constant, vsrari.b stores imm+8 (lane-width
+	// bias), the si5 compares store 5-bit two's-complement values and vseqi.d
+	// a 7-bit field the toolchain range-checks down to si5.
+	// The mnemonics that also have a register form (the shifts, the bit
+	// clear/set/rev families, VSEQ and the logic immediates) keep their
+	// three-register entry in l64InstrTable; the dispatcher picks the
+	// immediate opcode from l64VecImmInfo by operand kind, so the immediate
+	// entries must not overwrite the table.
+	vecImm := map[string]l64VecImmEnc{
+		"VANDB":  {0xE7A0 << 15, false, 0, 255, 0, 0xFF},
+		"XVANDB": {0xEFA0 << 15, true, 0, 255, 0, 0xFF},
+		"VORB":   {0xE7A8 << 15, false, 0, 255, 0, 0xFF},
+		"XVORB":  {0xEFA8 << 15, true, 0, 255, 0, 0xFF},
+		"VXORB":  {0xE7B0 << 15, false, 0, 255, 0, 0xFF},
+		"XVXORB": {0xEFB0 << 15, true, 0, 255, 0, 0xFF},
+		"VNORB":  {0xE7B8 << 15, false, 0, 255, 0, 0xFF},
+		"XVNORB": {0xEFB8 << 15, true, 0, 255, 0, 0xFF},
+		"VSEQB":  {0xE500 << 15, false, -16, 15, 0, 0x1F},
+		"XVSEQB": {0xE900 << 15, true, -16, 15, 0, 0x1F},
+		// vseqi.h/w accept the same si5 window as vseqi.b; vseqi.d carries a
+		// 7-bit field, but the toolchain range-checks it down to si5 as well
+		// (GOARCH=loong64 go tool asm rejects VSEQV $32 and VSEQV $-64).
+		"VSEQH":  {0xE501 << 15, false, -16, 15, 0, 0x1F},
+		"XVSEQH": {0xED01 << 15, true, -16, 15, 0, 0x1F},
+		"VSEQW":  {0xE502 << 15, false, -16, 15, 0, 0x1F},
+		"XVSEQW": {0xED02 << 15, true, -16, 15, 0, 0x1F},
+		"VSEQV":  {0xE503 << 15, false, -16, 15, 0, 0x7F},
+		"XVSEQV": {0xE903 << 15, true, -16, 15, 0, 0x7F},
+		// vslti compares against a signed (or, in the U spellings, unsigned)
+		// si5/ui5 constant.
+		"VSLTB":   {0xE50C << 15, false, -16, 15, 0, 0x1F},
+		"XVSLTB":  {0xED0C << 15, true, -16, 15, 0, 0x1F},
+		"VSLTH":   {0xE50D << 15, false, -16, 15, 0, 0x1F},
+		"XVSLTH":  {0xED0D << 15, true, -16, 15, 0, 0x1F},
+		"VSLTW":   {0xE50E << 15, false, -16, 15, 0, 0x1F},
+		"XVSLTW":  {0xED0E << 15, true, -16, 15, 0, 0x1F},
+		"VSLTV":   {0xE50F << 15, false, -16, 15, 0, 0x1F},
+		"XVSLTV":  {0xED0F << 15, true, -16, 15, 0, 0x1F},
+		"VSLTBU":  {0xE510 << 15, false, 0, 31, 0, 0x1F},
+		"XVSLTBU": {0xED10 << 15, true, 0, 31, 0, 0x1F},
+		"VSLTHU":  {0xE511 << 15, false, 0, 31, 0, 0x1F},
+		"XVSLTHU": {0xED11 << 15, true, 0, 31, 0, 0x1F},
+		"VSLTWU":  {0xE512 << 15, false, 0, 31, 0, 0x1F},
+		"XVSLTWU": {0xED12 << 15, true, 0, 31, 0, 0x1F},
+		"VSLTVU":  {0xE513 << 15, false, 0, 31, 0, 0x1F},
+		"XVSLTVU": {0xED13 << 15, true, 0, 31, 0, 0x1F},
+		// vaddi/vsubi take ui5 constants for every width on this toolchain
+		// (VADDVU $32 is rejected by the oracle although the field is ui8).
+		"VADDBU":  {0xE514 << 15, false, 0, 31, 0, 0x1F},
+		"XVADDBU": {0xED14 << 15, true, 0, 31, 0, 0x1F},
+		"VADDHU":  {0xE515 << 15, false, 0, 31, 0, 0x1F},
+		"XVADDHU": {0xED15 << 15, true, 0, 31, 0, 0x1F},
+		"VADDWU":  {0xE516 << 15, false, 0, 31, 0, 0x1F},
+		"XVADDWU": {0xED16 << 15, true, 0, 31, 0, 0x1F},
+		"VADDVU":  {0xE517 << 15, false, 0, 31, 0, 0x1F},
+		"XVADDVU": {0xED17 << 15, true, 0, 31, 0, 0x1F},
+		"VSUBBU":  {0xE518 << 15, false, 0, 31, 0, 0x1F},
+		"XVSUBBU": {0xED18 << 15, true, 0, 31, 0, 0x1F},
+		"VSUBHU":  {0xE519 << 15, false, 0, 31, 0, 0x1F},
+		"XVSUBHU": {0xED19 << 15, true, 0, 31, 0, 0x1F},
+		"VSUBWU":  {0xE51A << 15, false, 0, 31, 0, 0x1F},
+		"XVSUBWU": {0xED1A << 15, true, 0, 31, 0, 0x1F},
+		"VSUBVU":  {0xE51B << 15, false, 0, 31, 0, 0x1F},
+		"XVSUBVU": {0xED1B << 15, true, 0, 31, 0, 0x1F},
+		// The shift/rotate immediates ride in a width-sized field whose upper
+		// bits carry the lane-width code: vslli.b stores ui3 at [12:0] with
+		// bits [14:13] inside the opcode, vslli.h ui4 under a 4 bit mask, and
+		// the .w/.d spellings a raw ui5/ui6.
+		"VSLLB":  {0x732C2000, false, 0, 7, 0, 0x7},
+		"XVSLLB": {0x772C2000, true, 0, 7, 0, 0x7},
+		"VSLLH":  {0x732C4000, false, 0, 15, 0, 0xF},
+		"XVSLLH": {0x772C4000, true, 0, 15, 0, 0xF},
+		"VSLLW":  {0xE659 << 15, false, 0, 31, 0, 0x1F},
+		"XVSLLW": {0xEE59 << 15, true, 0, 31, 0, 0x1F},
+		"VSLLV":  {0xE65A << 15, false, 0, 63, 0, 0x3F},
+		"XVSLLV": {0xEE5A << 15, true, 0, 63, 0, 0x3F},
+		"VSRLB":  {0x73302000, false, 0, 7, 0, 0x7},
+		"XVSRLB": {0x77302000, true, 0, 7, 0, 0x7},
+		"VSRLH":  {0x73304000, false, 0, 15, 0, 0xF},
+		"XVSRLH": {0x77304000, true, 0, 15, 0, 0xF},
+		"VSRLW":  {0xE661 << 15, false, 0, 31, 0, 0x1F},
+		"XVSRLW": {0xEE61 << 15, true, 0, 31, 0, 0x1F},
+		"VSRLV":  {0xE662 << 15, false, 0, 63, 0, 0x3F},
+		"XVSRLV": {0xEE62 << 15, true, 0, 63, 0, 0x3F},
+		// vsrari/vrotri bias the field so the lane-width code rides above the
+		// shift amount (.b adds 8, .h 16, .w 32; .d is a raw ui6).
+		"VSRAB":   {0xE668 << 15, false, 0, 7, 8, 0x1F},
+		"XVSRAB":  {0xEE68 << 15, true, 0, 7, 8, 0x1F},
+		"VSRAH":   {0x73344000, false, 0, 15, 0, 0xF},
+		"XVSRAH":  {0x77344000, true, 0, 15, 0, 0xF},
+		"VSRAW":   {0xE669 << 15, false, 0, 31, 0, 0x1F},
+		"XVSRAW":  {0xEE69 << 15, true, 0, 31, 0, 0x1F},
+		"VSRAV":   {0xE66A << 15, false, 0, 63, 0, 0x3F},
+		"XVSRAV":  {0xEE6A << 15, true, 0, 63, 0, 0x3F},
+		"VROTRB":  {0x72A02000, false, 0, 7, 0, 0x7},
+		"XVROTRB": {0x76A02000, true, 0, 7, 0, 0x7},
+		"VROTRH":  {0x72A04000, false, 0, 15, 0, 0xF},
+		"XVROTRH": {0x76A04000, true, 0, 15, 0, 0xF},
+		"VROTRW":  {0xE541 << 15, false, 0, 31, 0, 0x1F},
+		"XVROTRW": {0xED41 << 15, true, 0, 31, 0, 0x1F},
+		"VROTRV":  {0xE542 << 15, false, 0, 63, 0, 0x3F},
+		"XVROTRV": {0xED42 << 15, true, 0, 63, 0, 0x3F},
+		// vbitclri/vbitseti/vbitrevi follow the same width-coded layout.
+		"VBITCLRB":  {0x73102000, false, 0, 7, 0, 0x7},
+		"XVBITCLRB": {0x77102000, true, 0, 7, 0, 0x7},
+		"VBITCLRH":  {0x73104000, false, 0, 15, 0, 0xF},
+		"XVBITCLRH": {0x77104000, true, 0, 15, 0, 0xF},
+		"VBITCLRW":  {0xE621 << 15, false, 0, 31, 0, 0x1F},
+		"XVBITCLRW": {0xEE21 << 15, true, 0, 31, 0, 0x1F},
+		"VBITCLRV":  {0xE622 << 15, false, 0, 63, 0, 0x3F},
+		"XVBITCLRV": {0xEE22 << 15, true, 0, 63, 0, 0x3F},
+		"VBITSETB":  {0x73142000, false, 0, 7, 0, 0x7},
+		"XVBITSETB": {0x77142000, true, 0, 7, 0, 0x7},
+		"VBITSETH":  {0x73144000, false, 0, 15, 0, 0xF},
+		"XVBITSETH": {0x77144000, true, 0, 15, 0, 0xF},
+		"VBITSETW":  {0xE629 << 15, false, 0, 31, 0, 0x1F},
+		"XVBITSETW": {0xEE29 << 15, true, 0, 31, 0, 0x1F},
+		"VBITSETV":  {0xE62A << 15, false, 0, 63, 0, 0x3F},
+		"XVBITSETV": {0xEE2A << 15, true, 0, 63, 0, 0x3F},
+		"VBITREVB":  {0x73182000, false, 0, 7, 0, 0x7},
+		"XVBITREVB": {0x77182000, true, 0, 7, 0, 0x7},
+		"VBITREVH":  {0x73184000, false, 0, 15, 0, 0xF},
+		"XVBITREVH": {0x77184000, true, 0, 15, 0, 0xF},
+		"VBITREVW":  {0xE631 << 15, false, 0, 31, 0, 0x1F},
+		"XVBITREVW": {0xEE31 << 15, true, 0, 31, 0, 0x1F},
+		"VBITREVV":  {0xE632 << 15, false, 0, 63, 0, 0x3F},
+		"XVBITREVV": {0xEE32 << 15, true, 0, 63, 0, 0x3F},
+		// The 4-bit-select shuffles and the byte-extract/insert permutations
+		// take ui8 (the .d shuffle ui4 range-checked to 0..15 by the
+		// toolchain) packing both position nibbles.
+		"VSHUF4IB":   {0xE720 << 15, false, 0, 255, 0, 0xFF},
+		"XVSHUF4IB":  {0xEF20 << 15, true, 0, 255, 0, 0xFF},
+		"VSHUF4IH":   {0xE728 << 15, false, 0, 255, 0, 0xFF},
+		"XVSHUF4IH":  {0xEF28 << 15, true, 0, 255, 0, 0xFF},
+		"VSHUF4IW":   {0xE730 << 15, false, 0, 255, 0, 0xFF},
+		"XVSHUF4IW":  {0xEF30 << 15, true, 0, 255, 0, 0xFF},
+		"VSHUF4IV":   {0xE738 << 15, false, 0, 15, 0, 0xFF},
+		"XVSHUF4IV":  {0xEF38 << 15, true, 0, 15, 0, 0xFF},
+		"VPERMIW":    {0xE7C8 << 15, false, 0, 255, 0, 0xFF},
+		"XVPERMIW":   {0xEFC8 << 15, true, 0, 255, 0, 0xFF},
+		"XVPERMIV":   {0xEFD0 << 15, true, 0, 255, 0, 0xFF},
+		"XVPERMIQ":   {0xEFD8 << 15, true, 0, 255, 0, 0xFF},
+		"VEXTRINSB":  {0xE718 << 15, false, 0, 255, 0, 0xFF},
+		"XVEXTRINSB": {0xEF18 << 15, true, 0, 255, 0, 0xFF},
+		"VEXTRINSH":  {0xE710 << 15, false, 0, 255, 0, 0xFF},
+		"XVEXTRINSH": {0xEF10 << 15, true, 0, 255, 0, 0xFF},
+		"VEXTRINSW":  {0xE708 << 15, false, 0, 255, 0, 0xFF},
+		"XVEXTRINSW": {0xEF08 << 15, true, 0, 255, 0, 0xFF},
+		"VEXTRINSV":  {0xE700 << 15, false, 0, 255, 0, 0xFF},
+		"XVEXTRINSV": {0xEF00 << 15, true, 0, 255, 0, 0xFF},
+	}
+	for m, e := range vecImm {
+		l64VecImmInfo[m] = e
+		l64VecBank[m] = e.lasx
+	}
+
+	// Vector-to-condition flag: INSTR vj, FCCn (vsetnez.v, vsetanyeqz.*,
+	// vsetallnez.*): the sub-op rides in the rk field.
+	vecCf := map[string]uint32{
+		"VSETNEV": 0xE539<<15 | 7<<10, "XVSETNEV": 0xED39<<15 | 7<<10,
+		"VSETANYEQB": 0xE539<<15 | 8<<10, "XVSETANYEQB": 0xED39<<15 | 8<<10,
+		"VSETANYEQV": 0xE539<<15 | 11<<10, "XVSETANYEQV": 0xED39<<15 | 11<<10,
+		"VSETALLNEV": 0xE539<<15 | 15<<10, "XVSETALLNEV": 0xED39<<15 | 15<<10,
+		"VSETEQV": 0xE539<<15 | 6<<10, "XVSETEQV": 0xED39<<15 | 6<<10,
+		"VSETANYEQH": 0xE539<<15 | 9<<10, "XVSETANYEQH": 0xED39<<15 | 9<<10,
+		"VSETANYEQW": 0xE539<<15 | 10<<10, "XVSETANYEQW": 0xED39<<15 | 10<<10,
+		"VSETALLNEB": 0xE539<<15 | 12<<10, "XVSETALLNEB": 0xED39<<15 | 12<<10,
+		"VSETALLNEH": 0xE539<<15 | 13<<10, "XVSETALLNEH": 0xED39<<15 | 13<<10,
+		"VSETALLNEW": 0xE539<<15 | 14<<10, "XVSETALLNEW": 0xED39<<15 | 14<<10,
+	}
+	for m, op := range vecCf {
+		l64InstrTable[m] = l64Enc{format: l64Fvcf, op: op}
+		l64VecBank[m] = strings.HasPrefix(m, "XV")
+	}
+
+	// Lane popcount and the two-operand vector FP/unary spellings: INSTR vj,
+	// vd (the 2R layout with the opcode extending over the unused vk field;
+	// the low byte of each constant is the instruction's own sub-op).
+	vec2r := map[string]l64Vec3Enc{
+		"VPCNTV": {0x1CA70B << 10, false}, "XVPCNTV": {0x1DA70B << 10, true},
+	}
+	// The rest of the lane popcounts, the vector negations and the vector FP
+	// unary conversions (loong64enc1.s).
+	vec2rMore := map[string]l64Vec3Enc{
+		"VPCNTB": {0x1CA708 << 10, false}, "VPCNTH": {0x1CA709 << 10, false},
+		"VPCNTW": {0x1CA70A << 10, false},
+		"VNEGB":  {0x1CA70C << 10, false}, "VNEGH": {0x1CA70D << 10, false},
+		"VNEGW": {0x1CA70E << 10, false}, "VNEGV": {0x1CA70F << 10, false},
+		"VFCLASSF": {0x1CA735 << 10, false}, "VFCLASSD": {0x1CA736 << 10, false},
+		"VFSQRTF": {0x1CA739 << 10, false}, "VFSQRTD": {0x1CA73A << 10, false},
+		"VFRECIPF": {0x1CA73D << 10, false}, "VFRECIPD": {0x1CA73E << 10, false},
+		"VFRSQRTF": {0x1CA741 << 10, false}, "VFRSQRTD": {0x1CA742 << 10, false},
+		"VFRINTF": {0x1CA74D << 10, false}, "VFRINTD": {0x1CA74E << 10, false},
+		"VFRINTRMF": {0x1CA751 << 10, false}, "VFRINTRMD": {0x1CA752 << 10, false},
+		"VFRINTRPF": {0x1CA755 << 10, false}, "VFRINTRPD": {0x1CA756 << 10, false},
+		"VFRINTRZF": {0x1CA759 << 10, false}, "VFRINTRZD": {0x1CA75A << 10, false},
+		"VFRINTRNEF": {0x1CA75D << 10, false}, "VFRINTRNED": {0x1CA75E << 10, false},
+		"XVPCNTB": {0x1DA708 << 10, true}, "XVPCNTH": {0x1DA709 << 10, true},
+		"XVPCNTW": {0x1DA70A << 10, true},
+		"XVNEGB":  {0x1DA70C << 10, true}, "XVNEGH": {0x1DA70D << 10, true},
+		"XVNEGW": {0x1DA70E << 10, true}, "XVNEGV": {0x1DA70F << 10, true},
+		"XVFCLASSF": {0x1DA735 << 10, true}, "XVFCLASSD": {0x1DA736 << 10, true},
+		"XVFSQRTF": {0x1DA739 << 10, true}, "XVFSQRTD": {0x1DA73A << 10, true},
+		"XVFRECIPF": {0x1DA73D << 10, true}, "XVFRECIPD": {0x1DA73E << 10, true},
+		"XVFRSQRTF": {0x1DA741 << 10, true}, "XVFRSQRTD": {0x1DA742 << 10, true},
+		"XVFRINTF": {0x1DA74D << 10, true}, "XVFRINTD": {0x1DA74E << 10, true},
+		"XVFRINTRMF": {0x1DA751 << 10, true}, "XVFRINTRMD": {0x1DA752 << 10, true},
+		"XVFRINTRPF": {0x1DA755 << 10, true}, "XVFRINTRPD": {0x1DA756 << 10, true},
+		"XVFRINTRZF": {0x1DA759 << 10, true}, "XVFRINTRZD": {0x1DA75A << 10, true},
+		"XVFRINTRNEF": {0x1DA75D << 10, true}, "XVFRINTRNED": {0x1DA75E << 10, true},
+	}
+	maps.Copy(vec2r, vec2rMore)
+	for m, e := range vec2r {
+		l64InstrTable[m] = l64Enc{format: l64Frr, op: e.op}
+		l64VecBank[m] = e.lasx
+		l64Vec2R[m] = true
+	}
+
+	// The four-register byte shuffle: INSTR va, vk, vj, vd (the operand the
+	// table reads in each field position, va at bits [19:15]).
+	vec4r := map[string]l64Vec3Enc{
+		"VSHUFB": {0x0D50 << 16, false}, "XVSHUFB": {0x0D60 << 16, true},
+	}
+	for m, e := range vec4r {
+		l64InstrTable[m] = l64Enc{format: l64Fvvvv, op: e.op}
+		l64VecBank[m] = e.lasx
+		l64Vec4R[m] = true
+	}
 }
 
 // l64FpMovTable maps (mnemonic, from-class, to-class) to the 2R opcode of the
-// register move between the integer and floating-point register banks — the
+// register move between the integer and floating-point register banks, the
 // MOVW/MOVV specials the Go assembler accepts.
 var l64FpMovTable = map[string]uint32{
 	"MOVV.R.F":    0x452a << 10, // movgr2fr.d

asm/loong64_encode_test.go

@@ -8,8 +8,8 @@ import (
 	"encoding/binary"
 	"testing"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/ast"
-	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+	"sourcedock.dev/petrbalvin/gasm-sdk/ast"
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
 )
 
 // firstTextLOONG64 parses assembly source and returns the first TEXT body.
@@ -263,6 +263,20 @@ func TestLOONG64_regNames(t *testing.T) {
 			t.Errorf("loong64RegNum(%q) = %d, want %d", name, got, want)
 		}
 	}
+	// The X/V spellings name the LSX/LASX vector banks, a register class of
+	// their own: the oracle (GOARCH=loong64 go tool asm) rejects `BEQZ X0`
+	// with "unrecognized instruction" while assembling `VADDV V0, V1, V2`
+	// and `XVADDV X0, X1, X2`, so loong64RegNum stays strict and the vector
+	// operands resolve through loong64VecRegNum only.
+	vecCases := map[string]int{
+		"V0": 0, "V31": 31, "X0": 0, "X31": 31,
+		"R4": -1, "F0": -1, "FCC0": -1, "V32": -1, "X32": -1, "V": -1, "X": -1,
+	}
+	for name, want := range vecCases {
+		if got := loong64VecRegNum(name); got != want {
+			t.Errorf("loong64VecRegNum(%q) = %d, want %d", name, got, want)
+		}
+	}
 }
 
 func TestLOONG64_bytesEqualGroundTruth(t *testing.T) {
@@ -291,3 +305,529 @@ done:
 		t.Errorf("code = % x\nwant % x", code, want)
 	}
 }
+
+// TestLOONG64IndirectBranch pins the indirect branch encodings: JMP (Rj) and
+// JAL (Rj) lower to jirl, and the raw JIRL spelling encodes the written
+// offset (the Go loong64 assembler deletes raw JIRL instructions entirely,
+// so this form is a gasm-only superset with faithful semantics).
+func TestLOONG64IndirectBranch(t *testing.T) {
+	fn := firstTextLOONG64(t, `#include "textflag.h"
+TEXT ·f(SB), NOSPLIT, $0-0
+	JMP (R4)
+	JIRL R0, R4, 8
+	RET
+`)
+	code := assembleLOONG64Helper(t, fn)
+	wantWords(t, code,
+		0x4C000080, // jirl r0, r4, 0
+		0x4C002080, // jirl r0, r4, 8
+		0x4C000020, // jirl r0, r1, 0 (RET)
+	)
+
+	// JAL (R5) links, so the toolchain gives the function its autosize-8
+	// prologue and epilogue around the call and the closing RET.
+	fn = firstTextLOONG64(t, `#include "textflag.h"
+TEXT ·f(SB), NOSPLIT, $0-0
+	JAL (R5)
+	RET
+`)
+	code = assembleLOONG64Helper(t, fn)
+	wantWords(t, code,
+		0x29FFE061, // st.d r1, -8(r3) (prologue saves RA below the new SP)
+		0x02FFE063, // addi.d r3, r3, -8 (prologue opens the frame)
+		0x29C00061, // st.d r1, 0(r3) (prologue saves RA at SP)
+		0x4C0000A1, // jirl r1, r5, 0
+		0x28C00061, // ld.d r1, 0(r3) (epilogue restores RA)
+		0x02C02063, // addi.d r3, r3, 8
+		0x4C000020, // jirl r0, r1, 0 (RET)
+	)
+}
+
+// TestLOONG64_vector pins the LSX/LASX slice against words read off
+// GOARCH=loong64 go tool asm (cross-checked against the toolchain's own
+// loong64enc1.s): the three-register forms, the immediate forms with their
+// biases, the vector-to-condition forms, lane popcount, the FP conversion,
+// FSEL and the VMOVQ move family.
+func TestLOONG64_vector(t *testing.T) {
+	t.Run("three-register and immediate forms", func(t *testing.T) {
+		fn := firstTextLOONG64(t, `#include "textflag.h"
+TEXT ·v(SB), NOSPLIT, $0
+	VADDV V1, V2, V3
+	VADDW V1, V2, V3
+	VADDV V2, V1
+	VANDV V1, V2
+	VXORV V1, V2, V3
+	VSEQB V1, V2, V3
+	VSEQV V1, V2, V3
+	VSRAB V1, V2, V3
+	VROTRW V1, V2, V3
+	VANDB $0, V2, V3
+	VANDB $255, V2
+	VSEQB $3, V2, V3
+	VSEQV $15, V2, V3
+	VSEQV $-15, V2, V3
+	VSRAB $7, V1, V2
+	VROTRW $16, V1, V2
+	VPCNTV V1, V2
+	XVADDV X1, X2, X3
+	XVXORV X1, X2, X3
+	XVSEQB X1, X2, X3
+	XVPCNTV X1, X2
+	RET
+`)
+		code := assembleLOONG64Helper(t, fn)
+		wantWords(t, code,
+			0x700B8443, // vadd.v v3, v2, v1
+			0x700B0443, // vadd.w
+			0x700B8821, // vadd.v v1, v1, v2 (two-operand form)
+			0x71260442, // vand.v v2, v2, v1
+			0x71270443, // vxor.v
+			0x70000443, // vseq.b
+			0x70018443, // vseq.d
+			0x70EC0443, // vsra.b
+			0x70EF0443, // vrotr.w
+			0x73D00043, // vandi.b v3, v2, 0
+			0x73D3FC42, // vandi.b v2, v2, 255 (two-operand form)
+			0x72800C43, // vseqi.b v3, v2, 3
+			0x7281BC43, // vseqi.d v3, v2, 15
+			0x7281C443, // vseqi.d v3, v2, -15 (7-bit two's complement)
+			0x73343C22, // vsrai.b v2, v1, 7 (encoded as 7+8)
+			0x72A0C022, // vrotri.w v2, v1, 16
+			0x729C2C22, // vpcnt.d v2, v1
+			0x740B8443, // xvadd.d x3, x2, x1
+			0x75270443, // xvxor.d
+			0x74000443, // xvseq.b
+			0x769C2C22, // xvpcnt.d x2, x1
+			0x4C000020,
+		)
+	})
+
+	t.Run("vector-to-condition", func(t *testing.T) {
+		fn := firstTextLOONG64(t, `#include "textflag.h"
+TEXT ·v(SB), NOSPLIT, $0
+	VSETNEV V1, FCC0
+	VSETANYEQB V1, FCC0
+	VSETANYEQV V2, FCC0
+	VSETALLNEV V0, FCC0
+	XVSETNEV X1, FCC0
+	XVSETALLNEV X1, FCC0
+	RET
+`)
+		code := assembleLOONG64Helper(t, fn)
+		wantWords(t, code,
+			0x729C9C20, // vsetnez.d fcc0, v1
+			0x729CA020, // vsetanyeqz.b
+			0x729CAC40, // vsetanyeqz.d
+			0x729CBC00, // vsetallnez.d
+			0x769C9C20, // xvsetnez.d
+			0x769CBC20, // xvsetallnez.d
+			0x4C000020,
+		)
+	})
+
+	t.Run("FP convert and FSEL", func(t *testing.T) {
+		fn := firstTextLOONG64(t, `#include "textflag.h"
+TEXT ·v(SB), NOSPLIT, $0
+	FFINTDV F0, F1
+	FSEL FCC0, F3, F4, F3
+	FSEL FCC1, F1, F2
+	RET
+`)
+		code := assembleLOONG64Helper(t, fn)
+		wantWords(t, code,
+			0x011D2801, // ffint.d.v f1, f0
+			0x0D000C83, // fsel f3, f4, f3, fcc0
+			0x0D008442, // fsel f2, f2, f1, fcc1
+			0x4C000020,
+		)
+	})
+
+	t.Run("VMOVQ move family", func(t *testing.T) {
+		fn := firstTextLOONG64(t, `#include "textflag.h"
+TEXT ·v(SB), NOSPLIT, $0
+	VMOVQ V1, V9
+	VMOVQ (R4), V2
+	VMOVQ 16(R4), V2
+	VMOVQ V0, (R4)
+	VMOVQ V0, 32(R4)
+	VMOVQ (R4)(R7), V3
+	VMOVQ V3, (R4)(R7)
+	VMOVQ R6, V0.B16
+	VMOVQ R6, V12.W4
+	VMOVQ (R4), V4.W4
+	XVMOVQ X3, X7
+	XVMOVQ (R4), X2
+	XVMOVQ X0, (R4)
+	XVMOVQ (R4)(R7), X4
+	XVMOVQ X0, (R4)(R7)
+	XVMOVQ R6, X0.B32
+	RET
+`)
+		code := assembleLOONG64Helper(t, fn)
+		wantWords(t, code,
+			0x732D0029, // vori.b v9, v1, 0 (register move)
+			0x2C000082, // vld v2, r4, 0
+			0x2C004082, // vld v2, r4, 16
+			0x2C400080, // vst v0, r4, 0
+			0x2C408080, // vst v0, r4, 32
+			0x38401C83, // vldx v3, r4, r7
+			0x38441C83, // vstx v3, r4, r7
+			0x729F00C0, // vreplgr2vr.b v0, r6
+			0x729F08CC, // vreplgr2vr.w v12, r6
+			0x30200084, // vldrepl.w v4, r4, 0
+			0x772D0067, // xvori.b x7, x3, 0
+			0x2C800082, // xvld x2, r4, 0
+			0x2CC00080, // xvst x0, r4, 0
+			0x38481C84, // xvldx x4, r4, r7
+			0x384C1C80, // xvstx x0, r4, r7
+			0x769F00C0, // xvreplgr2vr.b x0, r6
+			0x4C000020,
+		)
+	})
+
+	t.Run("element extract and insert", func(t *testing.T) {
+		fn := firstTextLOONG64(t, `#include "textflag.h"
+TEXT ·v(SB), NOSPLIT, $0
+	VMOVQ V0.V[0], R10
+	VMOVQ V6.V[1], R8
+	VMOVQ R9, V1.V[0]
+	XVMOVQ X0.V[0], R10
+	XVMOVQ X5.W[7], R7
+	XVMOVQ R4, X7.V[3]
+	RET
+`)
+		code := assembleLOONG64Helper(t, fn)
+		wantWords(t, code,
+			0x72EFF00A, // vpickve2gr.d r10, v0, 0
+			0x72EFF4C8, // vpickve2gr.d r8, v6, 1
+			0x72EBF121, // vinsgr2vr.d v1, r9, 0
+			0x76EFE00A, // xvpickve2gr.d r10, x0, 0
+			0x76EFDCA7, // xvpickve2gr.w r7, x5, 7
+			0x76EBEC87, // xvinsgr2vr.d x7, r4, 3
+			0x4C000020,
+		)
+	})
+
+	// The integer and FP add/subtract families with their saturating pairs
+	// and immediate spellings (loong64enc1.s words).
+	t.Run("add and subtract families", func(t *testing.T) {
+		fn := firstTextLOONG64(t, `#include "textflag.h"
+TEXT ·v(SB), NOSPLIT, $0
+	VADDB V1, V2, V3
+	VADDF V1, V2, V3
+	VADDD V1, V2, V3
+	VSUBD V1, V2, V3
+	VSADDV V1, V2, V3
+	VSSUBVU V1, V2, V3
+	VADDBU $1, V2, V1
+	VADDBU $1, V2
+	VSUBVU $31, V2
+	XVSADDV X3, X2, X1
+	XVSUBD X1, X2, X3
+	RET
+`)
+		code := assembleLOONG64Helper(t, fn)
+		wantWords(t, code,
+			0x700A0443, // vadd.b
+			0x71308443, // vadd.f
+			0x71310443, // vadd.d
+			0x71330443, // vsub.d
+			0x70478443, // vsadd.v
+			0x704D8443, // vssub.u.d
+			0x728A0441, // vaddi.bu v1, v2, 1
+			0x728A0442, // vaddi.bu v2, v2, 1 (two-operand form)
+			0x728DFC42, // vsubi.du v2, v2, 31 (two-operand form)
+			0x74478C41, // xvsadd.d x1, x2, x3
+			0x75330443, // xvsub.d x3, x2, x1
+			0x4C000020,
+		)
+	})
+
+	// The multiply, divide and accumulate families.
+	t.Run("multiply and divide families", func(t *testing.T) {
+		fn := firstTextLOONG64(t, `#include "textflag.h"
+TEXT ·v(SB), NOSPLIT, $0
+	VMULV V1, V2, V3
+	VMUHHU V1, V2, V3
+	VDIVBU V1, V2, V3
+	VMODV V1, V2, V3
+	VMADDB V1, V2, V3
+	VMSUBV V1, V2, V3
+	VMULWEVHB V1, V2, V3
+	VMULWODQV V1, V2, V3
+	VMADDWEVHBUB V1, V2, V3
+	XVDIVD X1, X2, X3
+	RET
+`)
+		code := assembleLOONG64Helper(t, fn)
+		wantWords(t, code,
+			0x70858443, // vmul.v
+			0x70888443, // vmuh.u.d
+			0x70E40443, // vdiv.u.b
+			0x70E38443, // vmod.d
+			0x70A80443, // vmadd.b
+			0x70AB8443, // vmsub.d
+			0x70900443, // vmulwev.h.b
+			0x70938443, // vmulwod.q.d
+			0x70BC0443, // vmaddwev.h.bu.b
+			0x753B0443, // xvdiv.d
+			0x4C000020,
+		)
+	})
+
+	// The shift, bit and interleave families in register and immediate
+	// spellings, with the width-coded shift immediates.
+	t.Run("shift, bit and interleave families", func(t *testing.T) {
+		fn := firstTextLOONG64(t, `#include "textflag.h"
+TEXT ·v(SB), NOSPLIT, $0
+	VSLLV V1, V2, V3
+	VROTRB V1, V2, V3
+	VBITCLRV V1, V2, V3
+	VBITSETW V1, V2, V3
+	VBITREVV V1, V2, V3
+	VILVLB V1, V2, V3
+	VILVHV V1, V2, V3
+	VSLLB $7, V1, V2
+	VSLLB $5, V1
+	VSRLH $15, V1, V2
+	VSRAW $31, V1, V2
+	VSRAV $63, V1, V2
+	VROTRV $63, V1, V2
+	VBITCLRB $7, V2, V3
+	VBITREVV $63, V2, V3
+	VSEQH $-16, V2, V3
+	VSLTB $1, V2, V3
+	VSLTHU $31, V2, V3
+	XVILVLV X3, X2, X1
+	XVSLLB $7, X2, X1
+	XVSRAV $63, X2, X1
+	XVBITREVV $63, X2, X1
+	RET
+`)
+		code := assembleLOONG64Helper(t, fn)
+		wantWords(t, code,
+			0x70E98443, // vsll.d
+			0x70EE0443, // vrotr.b
+			0x710D8443, // vbitclr.d
+			0x710F0443, // vbitset.w
+			0x71118443, // vbitrev.d
+			0x711A0443, // vilvl.b
+			0x711D8443, // vilvh.d
+			0x732C3C22, // vslli.b v2, v1, 7
+			0x732C3421, // vslli.b v1, v1, 5 (two-operand form)
+			0x73307C22, // vsrli.h v2, v1, 15
+			0x7334FC22, // vsrai.w v2, v1, 31
+			0x7335FC22, // vsrai.d v2, v1, 63
+			0x72A1FC22, // vrotri.d v2, v1, 63
+			0x73103C43, // vbitclri.b v3, v2, 7
+			0x7319FC43, // vbitrevi.d v3, v2, 63
+			0x7280C043, // vseqi.h v3, v2, -16
+			0x72860443, // vslti.b v3, v2, 1
+			0x7288FC43, // vslti.hu v3, v2, 31
+			0x751B8C41, // xvilvl.d x1, x2, x3
+			0x772C3C41, // xvslli.b x1, x2, 7
+			0x7735FC41, // xvsrai.d x1, x2, 63
+			0x7719FC41, // xvbitrevi.d x1, x2, 63
+			0x4C000020,
+		)
+	})
+
+	// The shuffle, select and permutation families, including the
+	// four-register byte shuffle.
+	t.Run("shuffle and permutation families", func(t *testing.T) {
+		fn := firstTextLOONG64(t, `#include "textflag.h"
+TEXT ·v(SB), NOSPLIT, $0
+	VSHUFH V1, V2, V3
+	VSHUFW V1, V2, V3
+	VSHUFV V1, V2, V3
+	VSHUFB V1, V2, V3, V4
+	XVSHUFB X1, X2, X3, X4
+	VSHUF4IB $255, V2, V1
+	VSHUF4IV $15, V2, V1
+	XVSHUF4IV $15, X1, X2
+	VEXTRINSB $0x18, V1, V2
+	XVEXTRINSV $0x81, X1, X2
+	VPERMIW $0x1B, V1, V2
+	XVPERMIQ $0x4B, X1, X2
+	RET
+`)
+		code := assembleLOONG64Helper(t, fn)
+		wantWords(t, code,
+			0x717A8443, // vshuf.h
+			0x717B0443, // vshuf.w
+			0x717B8443, // vshuf.d
+			0x0D508864, // vshuf.b v4, v3, v2, v1
+			0x0D608864, // xvshuf.b
+			0x7393FC41, // vshuf4i.b v1, v2, 255
+			0x739C3C41, // vshuf4i.d v1, v2, 15
+			0x779C3C22, // xvshuf4i.d x2, x1, 15
+			0x738C6022, // vextrins.b v2, v1, 0x18
+			0x77820422, // xvextrins.d x2, x1, 0x81
+			0x73E46C22, // vpermi.w v2, v1, 0x1b
+			0x77ED2C22, // xvpermi.q x2, x1, 0x4b
+			0x4C000020,
+		)
+	})
+
+	// The vector FP families, the unary spellings, the compare-to-flag
+	// additions and the scalar int/float conversions.
+	t.Run("FP and conversion families", func(t *testing.T) {
+		fn := firstTextLOONG64(t, `#include "textflag.h"
+TEXT ·v(SB), NOSPLIT, $0
+	VADDF V1, V2, V3
+	VMULF V1, V2, V3
+	VFCLASSD V1, V2
+	VFSQRTF V1, V2
+	VFRECIPD V1, V2
+	VFRSQRTF V1, V2
+	VFRINTF V1, V2
+	VFRINTRNED V1, V2
+	VNEGB V1, V2
+	VPCNTB V1, V2
+	XVNEGV X2, X1
+	XVPCNTW X3, X2
+	XVFRINTRNEF X1, X2
+	VSETEQV V1, FCC0
+	VSETANYEQH V1, FCC0
+	VSETALLNEB V1, FCC0
+	XVSETALLNEW X1, FCC0
+	FFINTFW F0, F1
+	FTINTVD F0, F1
+	RET
+`)
+		code := assembleLOONG64Helper(t, fn)
+		wantWords(t, code,
+			0x71308443, // vfadd.s
+			0x71388443, // vfmul.s
+			0x729CD822, // vfclass.d
+			0x729CE422, // vfsqrt.s
+			0x729CF822, // vfrecip.d
+			0x729D0422, // vfrsqrt.s
+			0x729D3422, // vfrint.s
+			0x729D7822, // vfrintne.s
+			0x729C3022, // vneg.b
+			0x729C2022, // vpcnt.b
+			0x769C3C41, // xvneg.d x1, x2
+			0x769C2862, // xvpcnt.w x2, x3
+			0x769D7422, // xvfrintne.s x2, x1
+			0x729C9820, // vseteqz.d fcc0, v1
+			0x729CA420, // vsetanyeqz.h
+			0x729CB020, // vsetallnez.b
+			0x769CB820, // xvsetallnez.w
+			0x011D1001, // ffint.s.w f1, f0
+			0x011B2801, // ftint.l.d f1, f0
+			0x4C000020,
+		)
+	})
+}
+
+// TestLOONG64_vectorErrors pins the register-class and range diagnostics of
+// the vector slice; each shape is rejected by the oracle as well
+// (GOARCH=loong64 go tool asm).
+func TestLOONG64_vectorErrors(t *testing.T) {
+	cases := []string{
+		// Integer registers in vector positions.
+		`TEXT ·e(SB), NOSPLIT, $0
+	VADDV R4, R5, R6
+	RET
+`,
+		// Crossed banks: LSX spellings take V, LASX spellings X.
+		`TEXT ·e(SB), NOSPLIT, $0
+	VADDV X1, X2, X3
+	RET
+`,
+		`TEXT ·e(SB), NOSPLIT, $0
+	XVADDV V1, V2, V3
+	RET
+`,
+		// The LASX bank has no .b/.h element forms.
+		`TEXT ·e(SB), NOSPLIT, $0
+	XVMOVQ R4, X2.B[0]
+	RET
+`,
+		// Immediate ranges.
+		`TEXT ·e(SB), NOSPLIT, $0
+	VANDB $256, V2
+	RET
+`,
+		`TEXT ·e(SB), NOSPLIT, $0
+	VSEQB $16, V2, V3
+	RET
+`,
+		`TEXT ·e(SB), NOSPLIT, $0
+	VROTRW $32, V1, V2
+	RET
+`,
+		`TEXT ·e(SB), NOSPLIT, $0
+	VADDVU $32, V2
+	RET
+`,
+		`TEXT ·e(SB), NOSPLIT, $0
+	VSEQV $32, V2, V3
+	RET
+`,
+		`TEXT ·e(SB), NOSPLIT, $0
+	VSHUF4IV $16, V2, V1
+	RET
+`,
+		`TEXT ·e(SB), NOSPLIT, $0
+	VEXTRINSB $256, V1, V2
+	RET
+`,
+		`TEXT ·e(SB), NOSPLIT, $0
+	VSLTV $-17, V2, V3
+	RET
+`,
+		// VSHUFB wants four vector registers.
+		`TEXT ·e(SB), NOSPLIT, $0
+	VSHUFB V1, V2, V3
+	RET
+`,
+		// The FCC forms still refuse vector registers.
+		`TEXT ·e(SB), NOSPLIT, $0
+	VSETEQV V1, V2
+	RET
+`,
+		// VSET* wants an FCC flag, not a vector register.
+		`TEXT ·e(SB), NOSPLIT, $0
+	VSETNEV V1, V2
+	RET
+`,
+	}
+	for i, src := range cases {
+		fn := firstTextLOONG64(t, src)
+		if _, _, _, _, _, err := assembleLOONG64(fn); err == nil {
+			t.Errorf("case %d: expected an error, got none", i)
+		}
+	}
+}
+
+// TestLOONG64_dbarAtomics pins the _dbar (acquire/release) AMO variants.
+// The oracle words come from GOARCH=loong64 go tool objdump of kernels
+// assembled with go tool asm, and match the toolchain's loong64enc1.s.
+func TestLOONG64_dbarAtomics(t *testing.T) {
+	fn := firstTextLOONG64(t, `#include "textflag.h"
+TEXT ·atoms(SB), NOSPLIT, $0
+	AMADDDBW R14, (R13), R12
+	AMADDDBV R14, (R13), R12
+	AMANDDBW R5, (R4), R6
+	AMANDDBV R5, (R4), R6
+	AMORDBW R5, (R4), R0
+	AMORDBV R5, (R4), R6
+	AMSWAPDBW R5, (R4), R6
+	AMCASDBV R6, (R4), R5
+	RET
+`)
+	code := assembleLOONG64Helper(t, fn)
+	wantWords(t, code,
+		0x386A39AC, // amadd_db.w r12, r13, r14
+		0x386AB9AC, // amadd_db.d
+		0x386B1486, // amand_db.w r6, r4, r5
+		0x386B9486, // amand_db.d
+		0x386C1480, // amor_db.w r0, r4, r5
+		0x386C9486, // amor_db.d
+		0x38691486, // amswap_db.w
+		0x385B9885, // amcas_db.w
+		0x4C000020,
+	)
+}

asm/loong64_frame.go

@@ -6,7 +6,7 @@ package asm
 import (
 	"strings"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/ast"
+	"sourcedock.dev/petrbalvin/gasm-sdk/ast"
 )
 
 // Loong64 frame mapping, matching the Go toolchain's loong64 backend.
@@ -20,14 +20,20 @@ import (
 // (the toolchain aligns frames with `if autosize&4 != 0 { autosize += 4 }`).
 // A leaf function (no calls) with a zero frame gets no prologue at all.
 //
-// Prologue (autosize > 0), byte-identical to the toolchain:
+// Prologue (autosize > 0, small), byte-identical to the toolchain:
 //
 //	MOVV R1, -autosize(R3)   // save LR below the new SP (traceback-safe)
 //	ADDV $-autosize, R3      // open the frame
 //	MOVV R1, 0(R3)           // save LR again at SP (signal-safety)
 //
+// Large frames (autosize past the 12-bit offset or immediate ranges) expand
+// the store and the adjust through REGTMP (R30) exactly as the toolchain's
+// assembler does: the store via the rounding LU12IW split, the adjust via
+// the floor LU12IW/ORI split.
+//
 // Epilogue: MOVV 0(R3), R1; ADDV $autosize, R3 (non-leaf only for the LR
-// restore); the RET's jirl r0, r1, 0 follows.
+// restore; the adjust materialised when the immediate does not fit); the
+// RET's jirl r0, r1, 0 follows.
 
 // loong64FrameInfo holds the frame layout derived from a TEXT directive.
 type loong64FrameInfo struct {
@@ -36,6 +42,11 @@ type loong64FrameInfo struct {
 	args     int  // the declared -argsize
 	noSplit  bool // the NOSPLIT flag
 	leaf     bool // no call instructions in the body
+
+	// Stack-split guard state: like amd64 and arm64, a leaf function with a
+	// small autosize is auto-marked NOSPLIT by the toolchain.
+	needSplit  bool
+	splitClass int // 0: <=StackSmall, 1: <=StackBig, 2: >StackBig
 }
 
 // loong64ComputeFrame derives the frame layout for a TEXT function.
@@ -59,9 +70,157 @@ func loong64ComputeFrame(t *ast.Text) loong64FrameInfo {
 		// A zero-frame non-leaf function still opens an 8-byte frame for LR.
 		fi.autosize = 8
 	}
+	switch {
+	case fi.noSplit:
+	case fi.autosize < stackSmall && fi.leaf:
+		// Auto-NOSPLIT, as the toolchain's leaf mark concludes.
+	default:
+		fi.needSplit = true
+		switch {
+		case fi.autosize <= stackSmall:
+			fi.splitClass = 0
+		case fi.autosize <= stackBig:
+			fi.splitClass = 1
+		default:
+			fi.splitClass = 2
+		}
+	}
 	return fi
 }
 
+// loong64GuardLen returns the byte length of the stack-split guard prefix
+// (zero when the function needs no guard).  The big class materialises two
+// constants through R30; each materialisation shrinks by one word when the
+// constant's low 12 bits are zero.
+func loong64GuardLen(fi loong64FrameInfo) int {
+	if !fi.needSplit {
+		return 0
+	}
+	off := int64(fi.autosize - stackSmall)
+	switch fi.splitClass {
+	case 0:
+		return 12
+	case 1:
+		if off <= 2048 {
+			return 16 // ADDV $-off fits the signed 12-bit immediate
+		}
+		return 24 // MOVV + LU12IW + ORI + ADDV + SGTU + BEQ
+	default:
+		// MOVV + [mat] + SGTU + BNE + [mat] + ADDV + SGTU + BEQ
+		return (6 + loong64MatLen(off) + loong64MatLen(-off)) * 4
+	}
+}
+
+// loong64MatLen reports the word count of materialising v in R30: a value
+// with a zero high part needs only the ORI (the toolchain's MOVW $v, R30),
+// one with a zero low part only the LU12IW.
+func loong64MatLen(v int64) int {
+	if v>>12 == 0 || v&0xFFF == 0 {
+		return 1
+	}
+	return 2
+}
+
+// loong64MatWords appends the words that materialise v in R30, splitting it
+// as v>>12 plus the zero-extended low 12 bits.
+func loong64MatWords(ws []uint32, v int64) []uint32 {
+	hi := v >> 12
+	lo := v & 0xFFF
+	if hi == 0 {
+		return append(ws, l64irr(l64OriOp, int(v), 0, 30))
+	}
+	ws = append(ws, l64ir(l64Lu12iwOp, int(hi), 30))
+	if lo != 0 {
+		ws = append(ws, l64irr(l64OriOp, int(lo), 30, 30))
+	}
+	return ws
+}
+
+// The LU12IW and ORI opcode bases (2RI20 and 2RI12 formats); the ORI reads
+// and writes rd itself.
+const (
+	l64Lu12iwOp = 0x0a << 25
+	l64OriOp    = 0x0e << 22
+)
+
+// loong64Imm12 reports whether v fits a signed 12-bit immediate.
+func loong64Imm12(v int64) bool { return v >= -2048 && v <= 2047 }
+
+// loong64GuardBytes emits the stack-split guard prefix.  blockStart is the
+// function-relative address of the morestack call at the end of the function;
+// branch displacements are in instructions and are computed from each
+// branch's own position.
+func loong64GuardBytes(fi loong64FrameInfo, blockStart int) []byte {
+	// MOVV 16(g), R20 (g.stackguard0), g = R22.
+	ws := []uint32{l64irr(l64loadStoreTable["MOVV"].ld, 16, 22, 20)}
+	off := int64(fi.autosize - stackSmall)
+	// beq appends BEQ R20, blockStart from the branch's own position.
+	beq := func() {
+		ws = append(ws, loong64Beqz(20, int32((blockStart-len(ws)*4)>>2)))
+	}
+	switch fi.splitClass {
+	case 0:
+		// SGTU SP, R20, R20; BEQ R20, more
+		ws = append(ws, l64rrr(l64DualTable["SGTU"].rrr, 3, 20, 20))
+		beq()
+	case 1:
+		ws = append(ws, loong64MediumWords(off)...)
+		ws = append(ws, l64rrr(l64DualTable["SGTU"].rrr, 24, 20, 20))
+		beq()
+	default:
+		// SGTU $off, SP, R24 catches the SP underflow a huge frame would
+		// cause; BNE jumps to morestack in that case.
+		ws = append(ws, loong64MatWords(nil, off)...)
+		ws = append(ws, l64rrr(l64DualTable["SGTU"].rrr, 30, 3, 24))
+		ws = append(ws, loong64Bnez(24, int32((blockStart-len(ws)*4)>>2)))
+		ws = append(ws, loong64MatWords(nil, -off)...)
+		ws = append(ws, l64rrr(l64DualTable["ADDV"].rrr, 30, 3, 24))
+		ws = append(ws, l64rrr(l64DualTable["SGTU"].rrr, 24, 20, 20))
+		beq()
+	}
+	return l64WordsLE(ws...)
+}
+
+// loong64MediumWords emits the medium-class stack check for offset off: the
+// ADDV immediate when it fits, otherwise the same sequence with the constant
+// materialised in R30.
+func loong64MediumWords(off int64) []uint32 {
+	if off <= 2048 {
+		return []uint32{l64irr(l64DualTable["ADDV"].imm, int(-off), 3, 24)}
+	}
+	ws := loong64MatWords(nil, -off)
+	return append(ws, l64rrr(l64DualTable["ADDV"].rrr, 30, 3, 24))
+}
+
+// loong64Beqz/loong64Bnez build the 21-bit conditional branches against R0
+// that the toolchain emits for its guard compares.
+func loong64Beqz(rj int, dispInstr int32) uint32 {
+	return l64ir21(l64branch21Table["BEQZ"], int(dispInstr), rj)
+}
+
+func loong64Bnez(rj int, dispInstr int32) uint32 {
+	return l64ir21(l64branch21Table["BNEZ"], int(dispInstr), rj)
+}
+
+// loong64MoreStackBlock emits the trailing block: MOVV R1, R31 (save LR, the
+// toolchain's OR R1, R0, R31 expansion), BL runtime.morestack_noctxt, B back
+// to the function entry.
+func loong64MoreStackBlock(blockStart int) ([]byte, Reloc) {
+	ws := []uint32{
+		l64rrr(l64DualTable["OR"].rrr, 0, 1, 31), // MOVV R1, R31 (OR R1, R0, R31)
+		l64bbl(l64jumpTable["BL"], 0),            // BL, patched by the linker
+	}
+	disp := (-(blockStart + 8)) >> 2
+	ws = append(ws, l64bbl(l64jumpTable["B"], int(disp)))
+	reloc := Reloc{
+		Off:   blockStart + 4,
+		After: blockStart + 8,
+		Name:  "runtime\u00b7morestack_noctxt",
+		Kind:  RelLoong64Branch,
+	}
+	return l64WordsLE(ws...), reloc
+}
+
 // loong64IsLeaf reports whether a function contains no call instructions
 // (JAL/BL/CALL), matching the toolchain's LEAF mark, which drives the frame
 // and the epilogue shape.
@@ -79,17 +238,37 @@ func loong64IsLeaf(t *ast.Text) bool {
 	return true
 }
 
-// loong64Prologue returns the prologue bytes for a loong64 function.
+// loong64Prologue returns the prologue bytes for a loong64 function.  When
+// the LR store offset leaves the toolchain's 12-bit store range ([-2046,
+// 2045], BIG_12 = 2046) or the SP adjust immediate its 12-bit immediate
+// range, each switches to the R30 materialisation the assembler expands it
+// to: the store uses the rounding %hi/%lo split (LU12IW of (v+2048)>>12,
+// REGTMP += SP, store at the raw offset), the adjust the floor split
+// (LU12IW, ORI when the low part is non-zero, REGTMP += SP).
 func loong64Prologue(fi loong64FrameInfo) []byte {
 	if fi.autosize == 0 {
 		return nil
 	}
 	addiD := l64DualTable["ADDV"].imm
-	return l64WordsLE(
-		l64irr(l64loadStoreTable["MOVV"].st, -fi.autosize, 3, 1), // MOVV R1, -autosize(R3)
-		l64irr(addiD, -fi.autosize, 3, 3),                        // ADDV $-autosize, R3
-		l64irr(l64loadStoreTable["MOVV"].st, 0, 3, 1),            // MOVV R1, 0(R3)
-	)
+	var ws []uint32
+	storeBase := 3
+	if fi.autosize > 2046 {
+		// The store goes through REGTMP: LU12IW of the rounding split,
+		// REGTMP += SP, then the store at REGTMP with the truncated offset.
+		v := -int64(fi.autosize)
+		ws = append(ws, l64ir(l64Lu12iwOp, int((v+2048)>>12), 30))
+		ws = append(ws, l64rrr(l64DualTable["ADDV"].rrr, 3, 30, 30))
+		storeBase = 30
+	}
+	ws = append(ws, l64irr(l64loadStoreTable["MOVV"].st, -fi.autosize, storeBase, 1)) // MOVV R1, -autosize(base)
+	if loong64Imm12(-int64(fi.autosize)) {
+		ws = append(ws, l64irr(addiD, -fi.autosize, 3, 3)) // ADDV $-autosize, R3
+	} else {
+		ws = append(ws, loong64MatWords(nil, -int64(fi.autosize))...)
+		ws = append(ws, l64rrr(l64DualTable["ADDV"].rrr, 30, 3, 3))
+	}
+	ws = append(ws, l64irr(l64loadStoreTable["MOVV"].st, 0, 3, 1)) // MOVV R1, 0(R3)
+	return l64WordsLE(ws...)
 }
 
 // loong64Return returns the bytes for a RET: the epilogue (restore LR and
@@ -98,17 +277,49 @@ func loong64Return(fi loong64FrameInfo) []byte {
 	var ws []uint32
 	if fi.autosize != 0 {
 		if !fi.leaf {
-			// MOVV 0(R3), R1 — restore the link register.
+			// MOVV 0(R3), R1, restore the link register.
 			ws = append(ws, l64irr(l64loadStoreTable["MOVV"].ld, 0, 3, 1))
 		}
-		// ADDV $autosize, R3 — close the frame.
-		ws = append(ws, l64irr(l64DualTable["ADDV"].imm, fi.autosize, 3, 3))
+		// ADDV $autosize, R3, close the frame (materialised when the
+		// immediate does not fit).
+		if loong64Imm12(int64(fi.autosize)) {
+			ws = append(ws, l64irr(l64DualTable["ADDV"].imm, fi.autosize, 3, 3))
+		} else {
+			ws = append(ws, loong64MatWords(nil, int64(fi.autosize))...)
+			ws = append(ws, l64rrr(l64DualTable["ADDV"].rrr, 30, 3, 3))
+		}
 	}
-	// jirl r0, r1, 0 — return.
+	// jirl r0, r1, 0, return.
 	ws = append(ws, l64irr16(l64branchTable["JIRL"], 0, 1, 0))
 	return l64WordsLE(ws...)
 }
 
+// loong64StoreWords reports the prologue word count of the LR store, and
+// loong64AdjustWords the word count of an SP adjust of v: the immediate
+// forms when they fit, otherwise the R30 materialisation sequences.
+func loong64StoreWords(autosize int) int {
+	if autosize > 2046 {
+		return 3
+	}
+	return 1
+}
+
+func loong64AdjustWords(v int64) int {
+	if loong64Imm12(v) {
+		return 1
+	}
+	return loong64MatLen(v) + 1
+}
+
+// loong64EpilogueWords reports the epilogue word count the RET expands to.
+func loong64EpilogueWords(fi loong64FrameInfo) int {
+	n := loong64AdjustWords(int64(fi.autosize))
+	if !fi.leaf {
+		n++
+	}
+	return n
+}
+
 // loong64ResolvePseudo translates a pseudo-register memory reference into a
 // hardware base register and offset.  x+N(FP) → (N + autosize + 8)(SP);
 // x-N(SP) → (autosize - N)(SP).  Returns base = -1 for an unresolvable

asm/loong64_more_test.go

@@ -7,7 +7,7 @@ import (
 	"bytes"
 	"testing"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
 )
 
 // TestLOONG64_sys exercises the no-operand system instructions and the
@@ -232,7 +232,10 @@ DATA ·table+0(SB)/8, $42
 }
 
 // TestLOONG64_errors checks the encoder's error paths: undefined labels,
-// invalid register operands and operand-count mismatches.
+// invalid register operands and operand-count mismatches.  The X0 and
+// AMADDW cases follow the oracle: GOARCH=loong64 go tool asm rejects
+// `BEQZ X0` (the X bank is not an integer register) and the two-register
+// `AMADDW R4, R5` (the AM* family is strictly `val, (addr), result`).
 func TestLOONG64_errors(t *testing.T) {
 	cases := []string{
 		`TEXT ·e(SB), NOSPLIT, $0
@@ -280,7 +283,7 @@ done:
 // TestLOONG64_pcsp checks the stack-adjustment table of a framed function:
 // the prologue raises the SP delta by autosize (in effect from the third
 // instruction) and the RET's epilogue restores it to zero, with the pc deltas
-// in MinLC (4) units — byte-identical to `go tool asm`.
+// in MinLC (4) units; byte-identical to `go tool asm`.
 func TestLOONG64_pcsp(t *testing.T) {
 	cases := []struct {
 		name string
@@ -347,21 +350,94 @@ TEXT ·sb(SB), NOSPLIT, $0
 	}
 }
 
-// TestLOONG64_movImmToFp checks the immediate-to-FP move forms.
+// TestLOONG64_movImmToFp checks the immediate-to-FP move: MOVW $c, Fd is the
+// only spelling the toolchain accepts, expanding to ori (or addi.w for the
+// negative span) into R30 plus movgr2fr.w.  The pinned words are the
+// toolchain's own bytes; the other widths and out-of-range constants are
+// illegal combinations there and are diagnosed here.
 func TestLOONG64_movImmToFp(t *testing.T) {
 	fn := firstTextLOONG64(t, `#include "textflag.h"
 TEXT ·fpmov(SB), NOSPLIT, $0
-	MOVV $0x1, F0
+	MOVW $0x1, F0
 	MOVW $0x2, F4
+	MOVW $-1, F4
 	RET
 `)
 	code := assembleLOONG64Helper(t, fn)
 	want := []byte{
-		0x00, 0x04, 0x80, 0x03, // ori f0, r0, 1
-		0x04, 0x08, 0x80, 0x03, // ori f4, r0, 2
+		0x1e, 0x04, 0x80, 0x03, // ori r30, r0, 1
+		0xc0, 0xa7, 0x14, 0x01, // movgr2fr.w f0, r30
+		0x1e, 0x08, 0x80, 0x03, // ori r30, r0, 2
+		0xc4, 0xa7, 0x14, 0x01, // movgr2fr.w f4, r30
+		0x1e, 0xfc, 0xbf, 0x02, // addi.w r30, r0, -1
+		0xc4, 0xa7, 0x14, 0x01, // movgr2fr.w f4, r30
 		0x20, 0x00, 0x00, 0x4c, // jirl r0, r1, 0
 	}
 	if !bytes.Equal(code, want) {
 		t.Errorf("code = % x\nwant % x", code, want)
 	}
 }
+
+// TestLOONG64_movImmToFpErrors checks the immediate-to-FP diagnostics: the
+// widths the toolchain rejects as illegal combinations, and constants beyond
+// the 12-bit ori/addi.w span (the toolchain never materialises a wider
+// constant on this path).
+func TestLOONG64_movImmToFpErrors(t *testing.T) {
+	cases := []string{
+		"MOVV $1, F0",
+		"MOVF $2, F4",
+		"MOVD $2, F4",
+		"MOVW $100000, F1",
+		"MOVW $-2049, F1",
+		"MOVW $4096, F1",
+	}
+	for _, src := range cases {
+		fn := firstTextLOONG64(t, "#include \"textflag.h\"\nTEXT ·e(SB), NOSPLIT, $0\n\t"+src+"\n\tRET\n")
+		if _, _, _, _, _, err := assembleLOONG64(fn); err == nil {
+			t.Errorf("%s: expected an error, got none", src)
+		}
+	}
+}
+
+// TestLOONG64_branch16Unsigned pins the unsigned two-operand branches: with
+// one register BLTU/BGEU keep the register-register form against R0 (never
+// taken), the toolchain's encoding, where a beqz would test the wrong
+// condition; the three-operand forms are unchanged.
+func TestLOONG64_branch16Unsigned(t *testing.T) {
+	fn := firstTextLOONG64(t, `#include "textflag.h"
+TEXT ·u(SB), NOSPLIT, $0
+	BLTU R4, done
+	BGEU R5, done
+	BLTU R6, R7, done
+	BGEU R8, R9, done
+done:
+	RET
+`)
+	code := assembleLOONG64Helper(t, fn)
+	wantWords(t, code,
+		0x68001080, // bltu r4, r0, +4
+		0x6C000CA0, // bgeu r5, r0, +3
+		0x680008C7, // bltu r6, r7, +2
+		0x6C000509, // bgeu r8, r9, +1
+		0x4C000020, // jirl r0, r1, 0
+	)
+}
+
+// TestLOONG64_bitFieldRange checks the BSTRINS/BSTRPICK bit-number
+// validation, mirroring the toolchain's "illegal bit number" rule: 0..31 for
+// the .w forms, 0..63 for the .d forms, and lsb <= msb.
+func TestLOONG64_bitFieldRange(t *testing.T) {
+	cases := []string{
+		"BSTRINSW $32, R4, $0, R5",
+		"BSTRPICKW $31, R4, $32, R5",
+		"BSTRINSV $64, R4, $0, R5",
+		"BSTRPICKV $3, R4, $4, R5",
+		"BSTRINSW $-1, R4, $0, R5",
+	}
+	for _, src := range cases {
+		fn := firstTextLOONG64(t, "#include \"textflag.h\"\nTEXT ·e(SB), NOSPLIT, $0\n\t"+src+"\n\tRET\n")
+		if _, _, _, _, _, err := assembleLOONG64(fn); err == nil {
+			t.Errorf("%s: expected an error, got none", src)
+		}
+	}
+}

asm/loong64_reloc_test.go

@@ -0,0 +1,42 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package asm
+
+import (
+	"testing"
+
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
+)
+
+// TestLOONG64RelocOffsetsIncludePrologue pins the function-relative
+// relocation offsets of a framed loong64 function: the offsets used to
+// exclude the prologue, so every relocation landed on a prologue
+// instruction in the GOOBJ/ELF output.
+func TestLOONG64RelocOffsetsIncludePrologue(t *testing.T) {
+	f, errs := parser.Parse("k_loong64.s", "TEXT \u00b7f(SB), $16-0\n"+
+		"\tMOVV $gdata(SB), R4\n"+
+		"\tRET\n"+
+		"GLOBL gdata(SB), $8\n")
+	if len(errs) > 0 {
+		t.Fatalf("parse: %v", errs)
+	}
+	img, err := AssembleFileLOONG64(f)
+	if err != nil {
+		t.Fatalf("assemble: %v", err)
+	}
+	fn := img.Funcs[0]
+
+	// Layout: 12-byte prologue (autosize 32), pcalau12i+addi.d (12, 16),
+	// epilogue with RET.
+	if len(fn.Relocs) != 2 {
+		t.Fatalf("relocs = %d, want 2", len(fn.Relocs))
+	}
+	hi, lo := fn.Relocs[0], fn.Relocs[1]
+	if hi.Kind != RelLoong64AddrHi || hi.Off != 12 || hi.After != 12 {
+		t.Errorf("hi reloc = {off %d after %d kind %d}, want {off 12 after 12 kind RelLoong64AddrHi}", hi.Off, hi.After, hi.Kind)
+	}
+	if lo.Kind != RelLoong64AddrLo || lo.Off != 16 || lo.After != 16 {
+		t.Errorf("lo reloc = {off %d after %d kind %d}, want {off 16 after 16 kind RelLoong64AddrLo}", lo.Off, lo.After, lo.Kind)
+	}
+}

asm/operand.go

@@ -14,6 +14,51 @@ type Imm int64
 
 func (Imm) isOperand() {}
 
+// RegList is a bracketed register range, [Z0-Z3]: the four-register source
+// of the 4FMAPS and 4VNNIW families.  The EVEX emit path carries the list's
+// low register through the inverted 5-bit V'VVVV field; the three higher
+// registers are implied by the instruction, so only the pair travels here.
+type RegList struct {
+	Lo Reg
+	Hi Reg // implied by the encoding; Lo.idx+3 by construction
+}
+
+func (RegList) isOperand() {}
+
+// FloatImm is a floating-point immediate ($-1.0).  The SSE mnemonics whose
+// encoding takes an XMM/memory source at that position rewrite it as a read
+// from a read-only pool constant ($f64.<hex> or $f32.<hex>), the toolchain's
+// own behaviour; every other instruction rejects it.
+type FloatImm struct {
+	Text string // the numeric text as written, sign excluded
+	Neg  bool   // a leading minus
+}
+
+func (FloatImm) isOperand() {}
+
+// TLSMem is a thread-local access, the source form off(base)(TLS*1) with the
+// base dropped: the toolchain's one-instruction TLS rewrite assembles it as
+// the segment-prefixed absolute whose disp32 carries an R_TLS_LE patch site
+// (the linker fills the TLS slot offset).
+type TLSMem struct {
+	Disp int64
+	Size int
+	Seg  byte // the segment override: FS (0x64) or GS (0x65) on windows
+}
+
+func (TLSMem) isOperand() {}
+
+// SegAbs is a segment-absolute access, 0x30(GS): the segment override
+// prefixes a disp32 absolute reference with no relocation.  The base
+// register spellings GS and FS produce it.
+type SegAbs struct {
+	Disp int64
+	Size int
+	Seg  byte // 0x64 FS, 0x65 GS
+}
+
+func (SegAbs) isOperand() {}
+
 // Mem is a memory operand of the form disp(base)(index*scale).
 type Mem struct {
 	Base     Reg
@@ -23,6 +68,7 @@ type Mem struct {
 	Size     int // operand width in bytes
 	HasBase  bool
 	HasIndex bool
+	Seg      byte // segment override prefix (0x64 FS, 0x65 GS); 0 = none
 }
 
 func (Mem) isOperand() {}
@@ -37,11 +83,6 @@ func Idx(base, index Reg, scale int, disp int64, size int) Mem {
 	return Mem{Base: base, Index: index, Scale: scale, Disp: disp, Size: size, HasBase: true, HasIndex: true}
 }
 
-// Rip builds a RIP-relative memory operand (RIP)+disp.
-func Rip(disp int64, size int) Mem {
-	return Mem{Disp: disp, Size: size}
-}
-
 // sbMem is a memory operand that references a static (SB) symbol.  It encodes
 // as a RIP-relative reference with a placeholder displacement; the encoder
 // records a patch site so the file-level layout can fill in the true rel32
@@ -53,3 +94,16 @@ type sbMem struct {
 }
 
 func (sbMem) isOperand() {}
+
+// isX86Mem reports whether the operand is an amd64 memory reference: a base
+// or indexed Mem, or an SB-relative sbMem.  Encoders that gate on "memory in
+// this position" must accept both; the r/m emitters distinguish the two
+// themselves.
+func isX86Mem(o Operand) bool {
+	switch o.(type) {
+	case Mem, sbMem:
+		return true
+	default:
+		return false
+	}
+}

asm/reg.go

@@ -7,36 +7,39 @@
 // by round-tripping through golang.org/x/arch's decoder in the tests.
 package asm
 
+import "maps"
+
 import "strings"
 
 // Reg is an x86-64 register.  In Plan 9 assembly the classic names (AX, BX, …)
-// are size-agnostic — the instruction suffix (MOVQ vs MOVL) fixes the width —
+// are size-agnostic, the instruction suffix (MOVQ vs MOVL) fixes the width
 // so the encoder keys off the register's index and lets the mnemonic supply the
 // size.  The high flag marks the legacy high-byte registers AH/CH/DH/BH, which
-// occupy indices 4–7 yet take no REX prefix, unlike SPL/BPL/SIL/DIL that share
+// occupy indices 4-7 yet take no REX prefix, unlike SPL/BPL/SIL/DIL that share
 // those indices but require one.  The mask flag marks the AVX-512 opmask
-// registers K0–K7.
+// registers K0-K7, the fp flag the x87 stack registers F0-F7.
 type Reg struct {
 	idx  int
 	size int  // informational width implied by the name; the mnemonic decides
 	high bool // AH/CH/DH/BH
-	mask bool // K0–K7 opmask register
+	mask bool // K0-K7 opmask register
+	fp   bool // F0-F7 x87 stack register
 }
 
-// Index returns the register number (0–15 for GPRs, 0–31 for vectors).
+// Index returns the register number (0-15 for GPRs, 0-31 for vectors).
 func (r Reg) Index() int { return r.idx }
 
 // Size returns the width in bytes implied by the register's name.
 func (r Reg) Size() int { return r.size }
 
-// IsMask reports whether r is an AVX-512 opmask register (K0–K7).
+// IsMask reports whether r is an AVX-512 opmask register (K0-K7).
 func (r Reg) IsMask() bool { return r.mask }
 
 func (r Reg) isOperand() {}
 
 // needsREX reports whether this register forces a REX prefix at the given
-// operand size: the extended registers R8–R15 always do, and at byte size the
-// low registers SPL/BPL/SIL/DIL (indices 4–7, not high) do as well.
+// operand size: the extended registers R8-R15 always do, and at byte size the
+// low registers SPL/BPL/SIL/DIL (indices 4-7, not high) do as well.
 func (r Reg) needsREX(opSize int) bool {
 	if r.idx >= 8 {
 		return true
@@ -63,28 +66,28 @@ var (
 	CX = Reg{idx: 1, size: 2}
 	DX = Reg{idx: 2, size: 2}
 	BX = Reg{idx: 3, size: 2}
-	SP = Reg{idx: 4, size: 2}
-	BP = Reg{idx: 5, size: 2}
+	_  = Reg{idx: 4, size: 2}
+	_  = Reg{idx: 5, size: 2}
 	SI = Reg{idx: 6, size: 2}
 	DI = Reg{idx: 7, size: 2}
 
-	EAX = Reg{idx: 0, size: 4}
-	ECX = Reg{idx: 1, size: 4}
-	EDX = Reg{idx: 2, size: 4}
-	EBX = Reg{idx: 3, size: 4}
-	ESP = Reg{idx: 4, size: 4}
-	EBP = Reg{idx: 5, size: 4}
-	ESI = Reg{idx: 6, size: 4}
-	EDI = Reg{idx: 7, size: 4}
+	_ = Reg{idx: 0, size: 4}
+	_ = Reg{idx: 1, size: 4}
+	_ = Reg{idx: 2, size: 4}
+	_ = Reg{idx: 3, size: 4}
+	_ = Reg{idx: 4, size: 4}
+	_ = Reg{idx: 5, size: 4}
+	_ = Reg{idx: 6, size: 4}
+	_ = Reg{idx: 7, size: 4}
 
-	RAX = Reg{idx: 0, size: 8}
-	RCX = Reg{idx: 1, size: 8}
-	RDX = Reg{idx: 2, size: 8}
-	RBX = Reg{idx: 3, size: 8}
-	RSP = Reg{idx: 4, size: 8}
-	RBP = Reg{idx: 5, size: 8}
-	RSI = Reg{idx: 6, size: 8}
-	RDI = Reg{idx: 7, size: 8}
+	_ = Reg{idx: 0, size: 8}
+	_ = Reg{idx: 1, size: 8}
+	_ = Reg{idx: 2, size: 8}
+	_ = Reg{idx: 3, size: 8}
+	_ = Reg{idx: 4, size: 8}
+	_ = Reg{idx: 5, size: 8}
+	_ = Reg{idx: 6, size: 8}
+	_ = Reg{idx: 7, size: 8}
 )
 
 // regByName maps an assembly register name (case-insensitive) to a Reg.
@@ -121,19 +124,17 @@ func buildRegByName() map[string]Reg {
 	}
 
 	// 8-bit: AL..BH, SPL..DIL, R8B..R15B.
-	for n, r := range map[string]Reg{
+	maps.Copy(m, map[string]Reg{
 		"AL": AL, "CL": CL, "DL": DL, "BL": BL,
 		"AH": AH, "CH": CH, "DH": DH, "BH": BH,
 		"SPL": SPL, "BPL": BPL, "SIL": SIL, "DIL": DIL,
-	} {
-		m[n] = r
-	}
+	})
 	for i := 8; i <= 15; i++ {
 		m["R"+itoa(i)+"B"] = Reg{idx: i, size: 1}
 	}
 
 	// Vector: X0..X31 (128-bit, size 16), Y0..Y31 (256-bit, size 32),
-	// Z0..Z31 (512-bit, size 64).  Indices 16–31 are only encodable in EVEX
+	// Z0..Z31 (512-bit, size 64).  Indices 16-31 are only encodable in EVEX
 	// (AVX-512) instructions; the encoder validates that through its tables.
 	for i := 0; i <= 31; i++ {
 		m["X"+itoa(i)] = Reg{idx: i, size: 16}
@@ -144,6 +145,10 @@ func buildRegByName() map[string]Reg {
 	for i := 0; i <= 7; i++ {
 		m["K"+itoa(i)] = Reg{idx: i, size: 8, mask: true}
 	}
+	// x87 stack: F0..F7.
+	for i := 0; i <= 7; i++ {
+		m["F"+itoa(i)] = Reg{idx: i, size: 8, fp: true}
+	}
 	return m
 }
 

asm/riscv_encode.go

@@ -63,9 +63,9 @@ func riscvRegNum(name string) int {
 		return 24
 	case "X25", "S9":
 		return 25
-	case "X26", "S10":
+	case "X26", "S10", "CTXT":
 		return 26
-	case "X27", "S11":
+	case "X27", "S11", "g":
 		return 27
 	case "X28", "T3":
 		return 28
@@ -141,10 +141,36 @@ func riscvRegNum(name string) int {
 	case "F31", "FT11":
 		return 31
 	default:
+		// Vector registers V0-V31 (the "V" extension).  They share the
+		// register numbering with the integer file: a bare number 0-31.
+		if len(name) >= 2 && name[0] == 'V' {
+			if n, ok := parseRegDigits(name[1:], 31); ok {
+				return n
+			}
+		}
 		return -1
 	}
 }
 
+// parseRegDigits parses a decimal register suffix and reports whether it is
+// within [0, max].
+func parseRegDigits(digits string, max int) (int, bool) {
+	if digits == "" {
+		return 0, false
+	}
+	n := 0
+	for i := 0; i < len(digits); i++ {
+		if digits[i] < '0' || digits[i] > '9' {
+			return 0, false
+		}
+		n = n*10 + int(digits[i]-'0')
+		if n > max {
+			return 0, false
+		}
+	}
+	return n, true
+}
+
 // RISC-V instruction encoding parameters.
 type riscvEnc struct {
 	opcode uint32 // bits [6:0]
@@ -154,7 +180,7 @@ type riscvEnc struct {
 
 // riscvInstrTable maps RISC-V mnemonics to their encoding.
 var riscvInstrTable = map[string]riscvEnc{
-	// RV64I — R-type arithmetic/logic.
+	// RV64I, R-type arithmetic/logic.
 	"ADD":  {0x33, 0x0, 0x00},
 	"SUB":  {0x33, 0x0, 0x20},
 	"SLL":  {0x33, 0x1, 0x00},
@@ -165,20 +191,20 @@ var riscvInstrTable = map[string]riscvEnc{
 	"SRA":  {0x33, 0x5, 0x20},
 	"OR":   {0x33, 0x6, 0x00},
 	"AND":  {0x33, 0x7, 0x00},
-	// RV64I — 32-bit variants (W suffix).
+	// RV64I, 32-bit variants (W suffix).
 	"ADDW": {0x3B, 0x0, 0x00},
 	"SUBW": {0x3B, 0x0, 0x20},
 	"SLLW": {0x3B, 0x1, 0x00},
 	"SRLW": {0x3B, 0x5, 0x00},
 	"SRAW": {0x3B, 0x5, 0x20},
-	// RV64I — I-type shift-immediate (shamt in rs2 field).
+	// RV64I, I-type shift-immediate (shamt in rs2 field).
 	"SLLI":  {0x13, 0x1, 0x00},
 	"SRLI":  {0x13, 0x5, 0x00},
 	"SRAI":  {0x13, 0x5, 0x20},
 	"SLLIW": {0x1B, 0x1, 0x00},
 	"SRLIW": {0x1B, 0x5, 0x00},
 	"SRAIW": {0x1B, 0x5, 0x20},
-	// RV64M — multiply/divide.
+	// RV64M, multiply/divide.
 	"MUL":    {0x33, 0x0, 0x01},
 	"MULH":   {0x33, 0x1, 0x01},
 	"MULHSU": {0x33, 0x2, 0x01},
@@ -187,13 +213,16 @@ var riscvInstrTable = map[string]riscvEnc{
 	"DIVU":   {0x33, 0x5, 0x01},
 	"REM":    {0x33, 0x6, 0x01},
 	"REMU":   {0x33, 0x7, 0x01},
-	// RV64M — 32-bit variants.
+	// RV64M, 32-bit variants.
 	"MULW":  {0x3B, 0x0, 0x01},
 	"DIVW":  {0x3B, 0x4, 0x01},
 	"DIVUW": {0x3B, 0x5, 0x01},
 	"REMW":  {0x3B, 0x6, 0x01},
 	"REMUW": {0x3B, 0x7, 0x01},
-	// RV64I — I-type arithmetic.
+	// Zicond conditional zeroing.
+	"CZEROEQZ": {0x33, 0x5, 0x07},
+	"CZERONEZ": {0x33, 0x7, 0x07},
+	// RV64I, I-type arithmetic.
 	"ADDI":  {0x13, 0x0, 0x00},
 	"ADDIW": {0x1B, 0x0, 0x00},
 	"SLTI":  {0x13, 0x2, 0x00},
@@ -221,38 +250,49 @@ var riscvInstrTable = map[string]riscvEnc{
 	"BGE":  {0x63, 0x5, 0x00},
 	"BLTU": {0x63, 0x6, 0x00},
 	"BGEU": {0x63, 0x7, 0x00},
+	// The swapped-spelling comparison forms: encoded as BLT/BGE/BLTU/BGEU
+	// with the register operands swapped.
+	"BGT":  {0x63, 0x4, 0x00},
+	"BLE":  {0x63, 0x5, 0x00},
+	"BGTU": {0x63, 0x6, 0x00},
+	"BLEU": {0x63, 0x7, 0x00},
 	// U-type.
 	"LUI":   {0x37, 0x0, 0x00},
 	"AUIPC": {0x17, 0x0, 0x00},
 	// System.
-	"ECALL":  {0x73, 0x0, 0x00},
-	"EBREAK": {0x73, 0x0, 0x00},
-	"FENCE":  {0x0F, 0x0, 0x00},
-	// JALR — indirect jump/call (I-type).
+	"ECALL":     {0x73, 0x0, 0x00},
+	"EBREAK":    {0x73, 0x0, 0x00},
+	"FENCE":     {0x0F, 0x0, 0x00},
+	"FENCE.TSO": {0x0F, 0x0, 0x00},
+	"PAUSE":     {0x0F, 0x0, 0x00},
+	// JALR, indirect jump/call (I-type).
 	"JALR": {0x67, 0x0, 0x00},
 
-	// RV64A — atomics (AMO opcode 0x2F).
-	// funct3: 0x2 = word, 0x3 = doubleword. funct5 in bits [31:27].
-	"AMOSWAPW": {0x2F, 0x2, 0x01 << 2},
-	"AMOSWAPD": {0x2F, 0x3, 0x01 << 2},
-	"AMOADDW":  {0x2F, 0x2, 0x00 << 2},
-	"AMOADDD":  {0x2F, 0x3, 0x00 << 2},
-	"AMOANDW":  {0x2F, 0x2, 0x0C << 2},
-	"AMOANDD":  {0x2F, 0x3, 0x0C << 2},
-	"AMOORW":   {0x2F, 0x2, 0x06 << 2},
-	"AMOORD":   {0x2F, 0x3, 0x06 << 2},
-	"AMOXORW":  {0x2F, 0x2, 0x04 << 2},
-	"AMOXORD":  {0x2F, 0x3, 0x04 << 2},
-	"AMOMAXW":  {0x2F, 0x2, 0x14 << 2},
-	"AMOMAXD":  {0x2F, 0x3, 0x14 << 2},
-	"AMOMINW":  {0x2F, 0x2, 0x10 << 2},
-	"AMOMIND":  {0x2F, 0x3, 0x10 << 2},
-	"AMOMAXUW": {0x2F, 0x2, 0x1C << 2},
-	"AMOMAXUD": {0x2F, 0x3, 0x1C << 2},
-	"AMOMINUW": {0x2F, 0x2, 0x18 << 2},
-	"AMOMINUD": {0x2F, 0x3, 0x18 << 2},
+	// RV64A, atomics (AMO opcode 0x2F).
+	// funct3: 0x2 = word, 0x3 = doubleword.  The stored funct7 is the full
+	// 7-bit field: funct5 in the upper five bits and the aq/rl ordering bits in
+	// the lower two, exactly as the toolchain writes them: every AMO sets both
+	// aq and rl (funct7 |= 3).
+	"AMOSWAPW": {0x2F, 0x2, 0x01<<2 | 0x3},
+	"AMOSWAPD": {0x2F, 0x3, 0x01<<2 | 0x3},
+	"AMOADDW":  {0x2F, 0x2, 0x00<<2 | 0x3},
+	"AMOADDD":  {0x2F, 0x3, 0x00<<2 | 0x3},
+	"AMOANDW":  {0x2F, 0x2, 0x0C<<2 | 0x3},
+	"AMOANDD":  {0x2F, 0x3, 0x0C<<2 | 0x3},
+	"AMOORW":   {0x2F, 0x2, 0x08<<2 | 0x3},
+	"AMOORD":   {0x2F, 0x3, 0x08<<2 | 0x3},
+	"AMOXORW":  {0x2F, 0x2, 0x04<<2 | 0x3},
+	"AMOXORD":  {0x2F, 0x3, 0x04<<2 | 0x3},
+	"AMOMAXW":  {0x2F, 0x2, 0x14<<2 | 0x3},
+	"AMOMAXD":  {0x2F, 0x3, 0x14<<2 | 0x3},
+	"AMOMINW":  {0x2F, 0x2, 0x10<<2 | 0x3},
+	"AMOMIND":  {0x2F, 0x3, 0x10<<2 | 0x3},
+	"AMOMAXUW": {0x2F, 0x2, 0x1C<<2 | 0x3},
+	"AMOMAXUD": {0x2F, 0x3, 0x1C<<2 | 0x3},
+	"AMOMINUW": {0x2F, 0x2, 0x18<<2 | 0x3},
+	"AMOMINUD": {0x2F, 0x3, 0x18<<2 | 0x3},
 
-	// RV64F/D — floating-point arithmetic.
+	// RV64F/D, floating-point arithmetic.
 	"FADDS":  {0x53, 0x0, 0x00},
 	"FSUBS":  {0x53, 0x0, 0x04},
 	"FMULS":  {0x53, 0x0, 0x08},
@@ -273,14 +313,25 @@ var riscvInstrTable = map[string]riscvEnc{
 	"FMAXS": {0x53, 0x1, 0x14},
 	"FMIND": {0x53, 0x0, 0x15},
 	"FMAXD": {0x53, 0x1, 0x15},
+	// FP sign injection (double): rs2 carries the sign source.
+	"FSGNJD":  {0x53, 0x0, 0x11},
+	"FSGNJS":  {0x53, 0x0, 0x10},
+	"FSGNJX":  {0x53, 0x0, 0x14},
+	"FSGNJXD": {0x53, 0x0, 0x15},
+	"FSGNJXS": {0x53, 0x0, 0x14},
+	"FSGNJND": {0x53, 0x1, 0x11},
+	"FSGNJNS": {0x53, 0x1, 0x10},
+	"FSGNJNX": {0x53, 0x1, 0x14},
 
-	// RV64A — load-reserved / store-conditional (funct5 0x02 / 0x03).
-	"LRW": {0x2F, 0x2, 0x02 << 2},
-	"LRD": {0x2F, 0x3, 0x02 << 2},
-	"SCW": {0x2F, 0x2, 0x03 << 2},
-	"SCD": {0x2F, 0x3, 0x03 << 2},
+	// RV64A, load-reserved / store-conditional (funct5 0x02 / 0x03).
+	// The toolchain gives LR acquire ordering (aq = 1) and SC release
+	// ordering (rl = 1).
+	"LRW": {0x2F, 0x2, 0x02<<2 | 0x2},
+	"LRD": {0x2F, 0x3, 0x02<<2 | 0x2},
+	"SCW": {0x2F, 0x2, 0x03<<2 | 0x1},
+	"SCD": {0x2F, 0x3, 0x03<<2 | 0x1},
 
-	// FP compare — result in integer register (funct7 0x50/0x51).
+	// FP compare, result in integer register (funct7 0x50/0x51).
 	"FEQS": {0x53, 0x2, 0x50},
 	"FLTS": {0x53, 0x1, 0x50},
 	"FLES": {0x53, 0x0, 0x50},
@@ -296,11 +347,11 @@ func riscvRType(enc riscvEnc, rd, rs1, rs2 int) uint32 {
 }
 
 // riscvAMOType encodes an atomic (AMO) instruction.
-// Layout: funct5 | aq | rl | rs2 | rs1 | funct3 | rd | opcode.
-// The funct5 is stored in the upper bits of enc.funct7 (shifted left by 2).
+// Layout: funct7 | rs2 | rs1 | funct3 | rd | opcode, where funct7 carries the
+// funct5 in its upper five bits and the aq/rl ordering bits in the lower two
+// (the table stores the full field, so the word needs no reassembly).
 func riscvAMOType(enc riscvEnc, rd, rs1, rs2 int) uint32 {
-	funct5 := enc.funct7 >> 2 // extract funct5 from the stored value
-	return (funct5 << 27) | (uint32(rs2) << 20) | (uint32(rs1) << 15) |
+	return (enc.funct7 << 25) | (uint32(rs2) << 20) | (uint32(rs1) << 15) |
 		(enc.funct3 << 12) | (uint32(rd) << 7) | enc.opcode
 }
 
@@ -328,6 +379,8 @@ var riscvCvtTable = map[string]riscvCvtEnc{
 	"FCVTSWU": {0x68, 0x1, 0x53}, // uint32 → float32
 	"FCVTSL":  {0x68, 0x2, 0x53}, // int64 → float32
 	"FCVTSLU": {0x68, 0x3, 0x53}, // uint64 → float32
+	"FCLASSS": {0x70, 0x0, 0x53}, // classify float32 → GPR mask
+	"FCLASSD": {0x70, 0x0, 0x53}, // classify float64 → GPR mask
 	"FCVTDW":  {0x69, 0x0, 0x53}, // int32 → float64
 	"FCVTDWU": {0x69, 0x1, 0x53}, // uint32 → float64
 	"FCVTDL":  {0x69, 0x2, 0x53}, // int64 → float64
@@ -340,6 +393,10 @@ var riscvCvtTable = map[string]riscvCvtEnc{
 	"FMVDX": {0x79, 0x0, 0x53}, // int64 → float64 (bit move)
 	"FMVXW": {0x70, 0x0, 0x53}, // float32 → int32 (bit move)
 	"FMVWX": {0x78, 0x0, 0x53}, // int32 → float32 (bit move)
+	// The toolchain's W/D suffix spellings of the same moves.
+	"FMVXS": {0x70, 0x0, 0x53},
+	"FMVFS": {0x78, 0x0, 0x53},
+	"FMVSX": {0x79, 0x0, 0x53},
 }
 
 // riscvCvtType encodes an FP conversion instruction.
@@ -371,7 +428,7 @@ var riscvFmaTable = map[string]riscvFmaEnc{
 // riscvFmaType encodes an R4-type fused multiply-add instruction.
 func riscvFmaType(enc riscvFmaEnc, rd, rs1, rs2, rs3 int) uint32 {
 	return (uint32(rs3) << 27) | (enc.fmt << 25) | (uint32(rs2) << 20) |
-		(uint32(rs1) << 15) | (0x0 << 12) /* rm=dynamic */ | (uint32(rd) << 7) | enc.opcode
+		(uint32(rs1) << 15) | (0x0 << 12) /* rm=RNE */ | (uint32(rd) << 7) | enc.opcode
 }
 
 // CSR (Control and Status Register) instructions.
@@ -439,13 +496,78 @@ func riscvJType(rd int, offset int32) uint32 {
 		0x6F // JAL opcode
 }
 
+// ---- RVV ("V" extension) encoding helpers ----
+
+// The OP-V major opcode and its funct3 subclasses.
+const (
+	riscvOpV = 0x57 // the vector operation opcode (also OPcfg for vset*)
+	// funct3 values: 0 OPIVV, 1 OPFVV, 2 OPMVV, 3 OPIVI, 4 OPIVX,
+	// 5 OPFVF, 6 OPMVX, 7 vsetvli.
+	riscvVf3VV  = 0x0 // vector-vector
+	riscvVf3MV  = 0x2 // vector mask
+	riscvVf3VI  = 0x3 // vector-immediate
+	riscvVf3VX  = 0x4 // vector-scalar
+	riscvVf3Cfg = 0x7 // vsetvli
+)
+
+// riscvVType composes the vsetvli/vsetivli vtype immediate: the register
+// group multiplier in [2:0], the selected element width in [5:3] and the
+// tail-agnostic and mask-agnostic policies in bits 6 and 7.
+func riscvVType(vsew, vlmul, vta, vma int) int {
+	return vlmul | vsew<<3 | vta<<6 | vma<<7
+}
+
+// riscvVSetEnc encodes VSETVLI and VSETIVLI: imm[31:20] = vtype, rs1 = the
+// avl register or 5-bit uimm, rd = the destination.  Both carry funct3 7; a
+// vsetivli is distinguished by bits [31:30] set in the immediate (the 0xC00
+// the toolchain writes above its 10-bit vtype).
+func riscvVSetEnc(vsetivli bool, avl, vtype, rd int) uint32 {
+	imm := vtype & 0x3FF
+	if vsetivli {
+		imm |= 0xC00
+	}
+	return uint32(imm)<<20 | uint32(avl&0x1F)<<15 | uint32(riscvVf3Cfg)<<12 |
+		uint32(rd)<<7 | riscvOpV
+}
+
+// riscvVLSType encodes a vector load or store: the full 32-bit word with the
+// segment count in bits [31:29], the addressing mode in bits [28:26], the
+// unmasked bit at 25 and the width in funct3.  width follows the load
+// convention (0 = 8-bit, 5 = 16-bit, 6 = 32-bit, 7 = 64-bit).
+func riscvVLSType(op uint32, nf, mop, width int, rs2 int32, rs1, rd int) uint32 {
+	return uint32(nf&0x7)<<29 | uint32(mop&0x7)<<26 | 1<<25 |
+		uint32(rs2)<<20 | uint32(rs1)<<15 | uint32(width&0x7)<<12 |
+		uint32(rd)<<7 | op
+}
+
+// riscvVVInstr encodes an OP-V instruction with the six-bit operation code in
+// funct7's upper bits, bit 25 as the unmasked flag and the three registers in
+// the standard positions.  vs1 may name an integer register for the *VX forms
+// (the scalar sits in the rs1 field) or an immediate for the *VI forms.
+func riscvVVInstr(funct6, funct3 int, vs1 int32, vs2, vd int) uint32 {
+	return uint32(funct6&0x3F)<<26 | 1<<25 | uint32(vs1)<<15 |
+		uint32(funct3)<<12 | uint32(vs2)<<20 | uint32(vd)<<7 | riscvOpV
+}
+
+// riscvVUnaryInstr encodes a one-vector-operand OP-V instruction whose fixed
+// fields live where the second source register would be: rs1Field and vs2 are
+// written verbatim (the oracle writes fixed non-zero constants there for some
+// instructions, such as 0x11 in the rs1 field of vmfirst.m and vid.v).
+func riscvVUnaryInstr(funct6, funct3 int, rs1Field int32, vs2, vd int) uint32 {
+	return uint32(funct6&0x3F)<<26 | 1<<25 | uint32(vs2&0x1F)<<20 |
+		uint32(rs1Field&0x1F)<<15 | uint32(funct3&0x7)<<12 | uint32(vd&0x1F)<<7 | riscvOpV
+}
+
+// riscvSegNF maps a segment count to the 3-bit nf field (count - 1).
+func riscvSegNF(n int) int32 { return int32(n - 1) }
+
 // ---- RVC (compressed) encoding helpers ----
 
 // isRVCIntReg reports whether a register number can be encoded in the 3-bit
-// prime register field used by compressed instructions (x8–x15).
+// prime register field used by compressed instructions (x8-x15).
 func isRVCIntReg(r int) bool { return r >= 8 && r <= 15 }
 
-// rvcReg3 returns the 3-bit encoding for registers x8–x15 (0–7).
+// rvcReg3 returns the 3-bit encoding for registers x8-x15 (0-7).
 func rvcReg3(r int) uint32 { return uint32(r - 8) }
 
 // rvcCR encodes a CR-type (register) compressed instruction.
@@ -455,7 +577,7 @@ func rvcCR(funct4, rd, rs2 uint32) uint16 {
 }
 
 // rvcCI encodes a CI-type (immediate) compressed instruction.
-// Used for C.ADDI, C.LI, C.LUI, C.ADDIW — linear 6-bit immediate.
+// Used for C.ADDI, C.LI, C.LUI, C.ADDIW, linear 6-bit immediate.
 func rvcCI(funct3, rd uint32, imm uint32) uint16 {
 	return uint16((funct3 << 13) | ((imm>>5)&1)<<12 | (rd << 7) | (imm&0x1F)<<2 | 0x1)
 }
@@ -520,11 +642,13 @@ func rvcCL(funct3, rd, rs1 uint32, imm uint32) uint16 {
 
 // rvcCS encodes a register-relative compressed store (op=00 quadrant): C.SW
 // (funct3=6), C.SD (funct3=7) or C.FSD (funct3=5).  imm is the full byte
-// offset; the immediate bits are extracted per the RISC-V CS format.
+// offset; the immediate bits are extracted per the RISC-V CS format, with the
+// same five-bit patterns as the load side ({5,4,3,7,6} and {5,4,3,2,6},
+// matching the toolchain's encodeCS).
 func rvcCS(funct3, rs2, rs1 uint32, imm uint32) uint16 {
-	pattern := []int{5, 3, 7, 6}
+	pattern := []int{5, 4, 3, 7, 6}
 	if funct3 == 0x6 {
-		pattern = []int{5, 3, 2, 6}
+		pattern = []int{5, 4, 3, 2, 6}
 	}
 	packed := encodeRVCPattern(imm, pattern)
 	return uint16((funct3 << 13) | ((packed>>2)&0x7)<<10 | (rs1 << 7) | ((packed & 0x3) << 5) | (rs2 << 2))

asm/riscv_encode_test.go

@@ -5,10 +5,13 @@ package asm
 
 import (
 	"bytes"
+	"encoding/binary"
+	"encoding/hex"
+	"strings"
 	"testing"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/ast"
-	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+	"sourcedock.dev/petrbalvin/gasm-sdk/ast"
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
 )
 
 // firstTextRISCV parses assembly source and returns the first TEXT function body.
@@ -30,7 +33,7 @@ func firstTextRISCV(t *testing.T, src string) *ast.Text {
 // assembleRISCVHelper assembles one TEXT function and returns its code bytes.
 func assembleRISCVHelper(t *testing.T, fn *ast.Text) []byte {
 	t.Helper()
-	code, _, _, _, _, err := assembleRISCV(fn)
+	code, _, _, _, _, _, err := assembleRISCV(fn)
 	if err != nil {
 		t.Fatalf("assemble: %v", err)
 	}
@@ -289,7 +292,7 @@ TEXT ·cmp(SB), NOSPLIT, $0
 }
 
 func TestRISCV_forwardBranch(t *testing.T) {
-	// Forward label reference — must not fail.
+	// Forward label reference; must not fail.
 	fn := firstTextRISCV(t, `#include "textflag.h"
 TEXT ·fwd(SB), NOSPLIT, $0
 	ADDI $1, X10, X10
@@ -691,6 +694,81 @@ DATA answer<>+0(SB)/8, $42
 	}
 }
 
+// TestRISCV_RVC_StorePatterns pins the register-relative compressed store
+// encodings for offsets with immediate bits 4 and 5 set, byte-identical to
+// the toolchain's encodeCS (patterns {5,4,3,7,6} and {5,4,3,2,6}).
+// Regression: the store-side patterns dropped imm[4], so every such store
+// silently encoded the wrong address while the loads stayed correct.
+func TestRISCV_RVC_StorePatterns(t *testing.T) {
+	fn := firstTextRISCV(t, `#include "textflag.h"
+TEXT ·csstores(SB), NOSPLIT, $0
+	SD  X9, 24(X8)
+	SW  X10, 16(X11)
+	FSD F8, 40(X12)
+	LD  24(X8), X9
+	LW  16(X11), X10
+	FLD 40(X12), F8
+	RET
+`)
+	code := assembleRISCVHelper(t, fn)
+	want := []byte{
+		0x04, 0xec, // c.sd x9, 24(x8)
+		0x88, 0xc9, // c.sw x10, 16(x11)
+		0x00, 0xb6, // c.fsd f8, 40(x12)
+		0x04, 0x6c, // c.ld x9, 24(x8)
+		0x88, 0x49, // c.lw x10, 16(x11)
+		0x00, 0x36, // c.fld f8, 40(x12)
+		0x67, 0x80, 0x00, 0x00, // jalr x0, 0(x1)
+	}
+	if !bytes.Equal(code, want) {
+		t.Errorf("code = % x\nwant % x", code, want)
+	}
+}
+
+// TestRISCV_FENCE pins the FENCE encoding: the toolchain expands the bare
+// mnemonic to fence iorw, iorw (0x0FF0000F), not fence 0,0.
+func TestRISCV_FENCE(t *testing.T) {
+	fn := firstTextRISCV(t, `#include "textflag.h"
+TEXT ·fence(SB), NOSPLIT, $0
+	FENCE
+	RET
+`)
+	code := assembleRISCVHelper(t, fn)
+	want := []byte{
+		0x0f, 0x00, 0xf0, 0x0f, // fence iorw, iorw
+		0x67, 0x80, 0x00, 0x00, // jalr x0, 0(x1)
+	}
+	if !bytes.Equal(code, want) {
+		t.Errorf("code = % x\nwant % x", code, want)
+	}
+}
+
+// TestRISCV_RVC_WidthSpellings pins the compression of the GOROOT width
+// spellings: MOVW and MOVD lower to their base load/store and compress
+// exactly like LW/SW/FLD/FSD would (the toolchain compresses these shapes;
+// before the normalisation they stayed 4 bytes).
+func TestRISCV_RVC_WidthSpellings(t *testing.T) {
+	fn := firstTextRISCV(t, `#include "textflag.h"
+TEXT ·widths(SB), NOSPLIT, $0-16
+	MOVW w+0(FP), X9
+	MOVW X9, v+4(FP)
+	MOVD d+0(FP), F8
+	MOVD F8, r+8(FP)
+	RET
+`)
+	code := assembleRISCVHelper(t, fn)
+	want := []byte{
+		0xa2, 0x44, // c.lwsp x9, 8
+		0x26, 0xc6, // c.swsp x9, 12
+		0x22, 0x24, // c.fldsp f8, 8
+		0x22, 0xa8, // c.fsdsp f8, 16
+		0x67, 0x80, 0x00, 0x00, // jalr x0, 0(x1)
+	}
+	if !bytes.Equal(code, want) {
+		t.Errorf("code = % x\nwant % x", code, want)
+	}
+}
+
 func TestRISCV_system_instrs(t *testing.T) {
 	// Test FENCE, ECALL, EBREAK encoding.
 	fn := firstTextRISCV(t, `#include "textflag.h"
@@ -707,16 +785,140 @@ TEXT ·sys(SB), NOSPLIT, $0
 	}
 }
 
-func TestRISCV_MOV_sym_FP_error(t *testing.T) {
-	// MOV $sym(FP), rd should return an error (unsupported).
+func TestRISCV_MOV_sym_FP(t *testing.T) {
+	// MOV $sym(FP), rd lowers to the frame-adjusted ADDI against SP: the
+	// toolchain's argframe spelling.  A zero frame leaves the offset at the
+	// 8-byte link slot, compressed to C.ADDI4SPN.
 	fn := firstTextRISCV(t, `#include "textflag.h"
-TEXT ·badfp(SB), NOSPLIT, $0
+TEXT ·argfp(SB), NOSPLIT, $0
 	MOV $arg(FP), X10
 	RET
 `)
-	_, _, _, _, _, err := assembleRISCV(fn)
-	if err == nil {
-		t.Error("expected error for MOV $arg(FP), got nil")
+	code, _, _, _, _, _, err := assembleRISCV(fn)
+	if err != nil {
+		t.Fatalf("assemble: %v", err)
+	}
+	// prologue (0: leaf, zero frame) + C.ADDI4SPN (2) + RET (4) = 6
+	want := []byte{0x28, 0x00, 0x67, 0x80, 0x00, 0x00}
+	if string(code) != string(want) {
+		t.Errorf("got % x, want % x", code, want)
+	}
+}
+
+func TestRISCV_Bookkeeping(t *testing.T) {
+	// FUNCDATA and PCDATA contribute no bytes; UNDEF is the toolchain's
+	// ebreak, compressed to C.EBREAK under RVC.
+	fn := firstTextRISCV(t, `#include "textflag.h"
+TEXT ·book(SB), NOSPLIT, $0-8
+	FUNCDATA $0, marks<>(SB)
+	PCDATA $1, $1
+	UNDEF
+	MOV $1, X10
+	MOV X10, ret+0(FP)
+	RET
+`)
+	code, _, _, _, _, _, err := assembleRISCV(fn)
+	if err != nil {
+		t.Fatalf("assemble: %v", err)
+	}
+	// C.EBREAK (2) + C.LI X10, 1 (2) + C.SWSP (2) + RET (4) = 10: the
+	// FUNCDATA and PCDATA statements contribute nothing.
+	want := []byte{0x02, 0x90, 0x05, 0x45, 0x2a, 0xe4, 0x67, 0x80, 0x00, 0x00}
+	if string(code) != string(want) {
+		t.Errorf("got % x, want % x", code, want)
+	}
+}
+
+func TestRISCV_JMPPCRel(t *testing.T) {
+	// JMP N(PC): the displacement tracks the instruction N source slots
+	// away in the final layout (0 the jump itself, negative backwards).
+	fn := firstTextRISCV(t, `#include "textflag.h"
+TEXT ·slots(SB), NOSPLIT, $0-0
+	JMP 2(PC)
+	MOV $1, X11
+	MOV $2, X12
+	MOV X12, X11
+	JMP -3(PC)
+	RET
+`)
+	code, _, _, _, _, _, err := assembleRISCV(fn)
+	if err != nil {
+		t.Fatalf("assemble: %v", err)
+	}
+	// JMP 2(PC) lands on the C.MV six bytes ahead; JMP -3(PC) lands back on
+	// the first C.LI, six bytes behind.
+	want := []byte{
+		0x6f, 0x00, 0x60, 0x00, // JAL X0, 6
+		0x85, 0x45, // C.LI X11, 1
+		0x09, 0x46, // C.LI X12, 2
+		0xb2, 0x85, // C.MV X11, X12
+		0x6f, 0xf0, 0xbf, 0xff, // JAL X0, -6
+		0x67, 0x80, 0x00, 0x00, // RET
+	}
+	if string(code) != string(want) {
+		t.Errorf("got % x, want % x", code, want)
+	}
+}
+
+func TestRISCV_MOVWideImm(t *testing.T) {
+	// Shift-sequence constants compress like the toolchain's expansion.
+	fn := firstTextRISCV(t, `#include "textflag.h"
+TEXT ·wide(SB), NOSPLIT, $0-0
+	MOV $0x8000000000000000, X5
+	MOV $0x100000000, X5
+	MOV $0x000fffffffffffda, X5
+	RET
+`)
+	code, _, _, _, _, _, err := assembleRISCV(fn)
+	if err != nil {
+		t.Fatalf("assemble: %v", err)
+	}
+	// C.LI -1, C.SLLI 63; C.LI 1, C.SLLI 32; C.LI -19, C.SLLI 13, SRLI 12.
+	want := []byte{
+		0xfd, 0x52, 0xfe, 0x12,
+		0x85, 0x42, 0x82, 0x12,
+		0xb5, 0x52, 0xb6, 0x02, 0x93, 0xd2, 0xc2, 0x00,
+		0x67, 0x80, 0x00, 0x00,
+	}
+	if string(code) != string(want) {
+		t.Errorf("got % x, want % x", code, want)
+	}
+}
+
+func TestRISCV_MOVImmPool(t *testing.T) {
+	// A constant outside the shift shapes loads from the pooled $i64 data
+	// symbol via AUIPC+LD, named like the toolchain's pool.
+	src := `#include "textflag.h"
+TEXT ·pool(SB), NOSPLIT, $0-8
+	MOV $0x0101010101010101, X16
+	MOV X16, ret+0(FP)
+	RET
+`
+	f, errs := parser.Parse("pool_riscv64.s", src)
+	if len(errs) > 0 {
+		t.Fatalf("parse: %v", errs)
+	}
+	img, err := AssembleFileRISCV(f)
+	if err != nil {
+		t.Fatalf("AssembleFileRISCV: %v", err)
+	}
+	// AUIPC X16, 0 + LD X16, 0(X16): the relocation pair carries the symbol.
+	wantCode := []byte{0x17, 0x08, 0x00, 0x00, 0x03, 0x38, 0x08, 0x00}
+	if string(img.Code[0:8]) != string(wantCode) {
+		t.Errorf("pool load: got % x", img.Code[0:8])
+	}
+	var lit *DataSymbol
+	for i := range img.DataSyms {
+		if img.DataSyms[i].Name == "$i64.0101010101010101" {
+			lit = &img.DataSyms[i]
+		}
+	}
+	if lit == nil {
+		t.Fatalf("pool symbol missing: %v", img.DataSyms)
+	}
+	wantData := []byte{0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01}
+	if string(img.Data[lit.Offset:lit.Offset+8]) != string(wantData) {
+		t.Errorf("pool bytes: got % x", img.Data[lit.Offset:lit.Offset+8])
 	}
 }
 
@@ -727,7 +929,7 @@ TEXT ·calltest(SB), NOSPLIT, $0
 	CALL ext(SB)
 	RET
 `)
-	code, _, relocs, _, _, err := assembleRISCV(fn)
+	code, _, relocs, _, _, _, err := assembleRISCV(fn)
 	if err != nil {
 		t.Fatalf("assemble: %v", err)
 	}
@@ -756,8 +958,385 @@ TEXT ·calllocal(SB), NOSPLIT, $0
 sub:
 	RET
 `)
-	_, _, _, _, _, err := assembleRISCV(fn)
+	_, _, _, _, _, _, err := assembleRISCV(fn)
 	if err == nil {
 		t.Error("expected error for CALL to local label, got nil")
 	}
 }
+
+// TestRISCVIndirectBranch pins the indirect branch encodings: JMP (X5) is the
+// toolchain's JALR X0, 0(X5), and the trampoline form JALR rd, offset(rs1)
+// takes its destination from the first operand (regression: the base
+// register was once read as the destination, silently jumping to X0).
+func TestRISCVIndirectBranch(t *testing.T) {
+	fn := firstTextRISCV(t, `#include "textflag.h"
+TEXT ·f(SB), NOSPLIT, $0-0
+	JMP (X5)
+	JALR X0, 0(X6)
+	JALR X28, 0(X9)
+	RET
+`)
+	code := assembleRISCVHelper(t, fn)
+	wantWords(t, code,
+		0x00028067, // jalr x0, 5(x0), 0
+		0x00030067, // jalr x0, 6(x0), 0
+		0x00048e67, // jalr x28, 9(x0), 0
+		0x00008067, // jalr x0, 1(x0), 0 (RET)
+	)
+}
+
+// encodeOneInstrRISCV encodes a single parsed instruction against a synthetic
+// offsets map, the smallest honest harness for the branch-range diagnostics:
+// the spans are far larger than any source a test would want to spell out.
+func encodeOneInstrRISCV(t *testing.T, src string, pc int, offsets map[string]int) ([]byte, error) {
+	t.Helper()
+	fn := firstTextRISCV(t, "#include \"textflag.h\"\n"+src)
+	instr := fn.Body[0].(*ast.Instr)
+	return encodeRISCVInstr(instr, pc, offsets, riscvFrameInfo{}, nil, nil, nil)
+}
+
+// TestRISCVBranchJumpRange checks that displacements beyond the B-type span
+// [-4096, 4094] and the J-type span [-1048576, 1048574] are diagnosed instead
+// of wrapping silently to a wrong target.
+func TestRISCVBranchJumpRange(t *testing.T) {
+	cases := []struct {
+		name string
+		src  string
+		off  int // the target's function-relative offset (pc 0)
+		ok   bool
+	}{
+		{"branch max", "BEQ X10, X11, tgt\nRET\n", 4094, true},
+		{"branch past max", "BEQ X10, X11, tgt\nRET\n", 4096, false},
+		{"branch back max", "BEQ X10, X11, tgt\nRET\n", -4096, true},
+		{"branch back past max", "BEQ X10, X11, tgt\nRET\n", -4098, false},
+		{"branchz past max", "BEQZ X10, tgt\nRET\n", 4096, false},
+		{"jump max", "JMP tgt\nRET\n", 1048574, true},
+		{"jump past max", "JMP tgt\nRET\n", 1048576, false},
+		{"jump back max", "JMP tgt\nRET\n", -1048576, true},
+		{"jump back past max", "JMP tgt\nRET\n", -1048578, false},
+		{"jal past max", "JAL tgt\nRET\n", 1048576, false},
+	}
+	for _, c := range cases {
+		t.Run(c.name, func(t *testing.T) {
+			_, err := encodeOneInstrRISCV(t, "TEXT ·f(SB), NOSPLIT, $0\n\t"+c.src, 0, map[string]int{"tgt": c.off})
+			if c.ok && err != nil {
+				t.Fatalf("unexpected error: %v", err)
+			}
+			if !c.ok && err == nil {
+				t.Fatal("expected an out-of-range diagnostic, got none")
+			}
+		})
+	}
+}
+
+// TestRISCVBranchFarBody drives the relaxation pass through the full
+// assembler: a forward branch over a body larger than the B-type span is
+// rewritten as an inverted branch over an inserted JMP, the same layout the
+// toolchain produces, instead of wrapping to a wrong target.
+func TestRISCVBranchFarBody(t *testing.T) {
+	var sb strings.Builder
+	sb.WriteString("#include \"textflag.h\"\nTEXT ·far(SB), NOSPLIT, $0\n\tBEQ X10, X11, done\n")
+	for range 1100 {
+		sb.WriteString("\tADD X10, X11, X12\n")
+	}
+	sb.WriteString("done:\n\tRET\n")
+	fn := firstTextRISCV(t, sb.String())
+	out, _, _, _, _, _, err := assembleRISCV(fn)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	// The relaxed branch at offset 0 targets the inserted JMP at 4 (bne
+	// x10, x11, +4); the JMP at 4 carries the far forward displacement.
+	wantBranch := wordLE(riscvBType(riscvEnc{0x63, 0x1, 0x00}, 10, 11, 4))
+	if !bytes.Equal(out[0:4], wantBranch) {
+		t.Errorf("relaxed branch = %x, want %x", out[0:4], wantBranch)
+	}
+	// done sits after 1100 ADDs: 4 + 4400, i.e. offset 4404 from the JMP at 4.
+	wantJmp := wordLE(riscvJType(0, 4404))
+	if !bytes.Equal(out[4:8], wantJmp) {
+		t.Errorf("inserted JMP = %x, want %x", out[4:8], wantJmp)
+	}
+}
+
+// TestRISCV_CSRRange checks the CSR address range: the 12-bit field is
+// diagnosed rather than masked, so CSRRW $4096 does not silently address
+// CSR 0.
+func TestRISCV_CSRRange(t *testing.T) {
+	fn := firstTextRISCV(t, `#include "textflag.h"
+TEXT ·csrhi(SB), NOSPLIT, $0
+	CSRRW $4096, X10, X11
+	RET
+`)
+	if _, _, _, _, _, _, err := assembleRISCV(fn); err == nil {
+		t.Error("expected an out-of-range error for CSR $4096, got none")
+	}
+	fn = firstTextRISCV(t, `#include "textflag.h"
+TEXT ·csrmax(SB), NOSPLIT, $0
+	CSRRW $4095, X10, X11
+	RET
+`)
+	if _, _, _, _, _, _, err := assembleRISCV(fn); err != nil {
+		t.Errorf("CSR $4095 must assemble: %v", err)
+	}
+}
+
+// TestRISCV_Imm64Rejected checks that immediates outside the signed 32-bit
+// span are diagnosed instead of silently truncated to their low 32 bits for
+// the I-type arithmetic; the MOV forms materialise the wide constant instead
+// (shift sequence or pooled load), like the toolchain.
+func TestRISCV_Imm64Rejected(t *testing.T) {
+	cases := []string{
+		"ADDI $0x100000000, X10, X11",
+		"ANDI $-0x800000001, X10, X11",
+		"SUB $0x100000000, X10, X11",
+	}
+	for _, src := range cases {
+		fn := firstTextRISCV(t, "#include \"textflag.h\"\nTEXT ·wide(SB), NOSPLIT, $0\n\t"+src+"\n\tRET\n")
+		if _, _, _, _, _, _, err := assembleRISCV(fn); err == nil {
+			t.Errorf("%s: expected an out-of-range error, got none", src)
+		}
+	}
+	// The full signed 32-bit span still assembles, including the SUB form
+	// whose negated immediate only just fits.
+	fn := firstTextRISCV(t, `#include "textflag.h"
+TEXT ·edge(SB), NOSPLIT, $0
+	MOV  $2147483647, X10
+	MOV  $-2147483648, X11
+	SUB  $0x80000000, X12, X13
+	RET
+`)
+	if _, _, _, _, _, _, err := assembleRISCV(fn); err != nil {
+		t.Errorf("int32-span immediates must assemble: %v", err)
+	}
+	// Beyond the span the MOV forms materialise the constant like the
+	// toolchain instead of diagnosing it.
+	fn = firstTextRISCV(t, `#include "textflag.h"
+TEXT ·pool(SB), NOSPLIT, $0
+	MOV $0x123456789, X10
+	RET
+`)
+	if _, _, _, _, _, _, err := assembleRISCV(fn); err != nil {
+		t.Errorf("MOV with a 64-bit immediate must assemble: %v", err)
+	}
+}
+
+// riscvWants decodes code as little-endian words and pins each one; the
+// expected values below were read off GOARCH=riscv64 go tool objdump of
+// kernels assembled with go tool asm (the toolchain's riscv64.s testdata
+// cross-checks the same words).
+func riscvWants(t *testing.T, code []byte, want ...uint32) {
+	t.Helper()
+	got := make([]uint32, 0, len(code)/4)
+	for i := 0; i+4 <= len(code); i += 4 {
+		got = append(got, binary.LittleEndian.Uint32(code[i:]))
+	}
+	if len(got) < len(want) {
+		t.Fatalf("word count = %d, want %d\ncode: % x", len(got), len(want), code)
+	}
+	// The RET (JALR) ends the sequence; only the pinned prefix is compared.
+	for i := range want {
+		if got[i] != want[i] {
+			t.Errorf("word %d = %08x, want %08x", i, got[i], want[i])
+		}
+	}
+}
+
+// riscvWantsHex pins the exact hex encoding of a function's instruction
+// bytes, including any 2-byte compressed instructions in the stream; the
+// expected strings were read off GOARCH=riscv64 go tool objdump of kernels
+// assembled with go tool asm (the toolchain's riscv64.s testdata
+// cross-checks the same words).
+func riscvWantsHex(t *testing.T, code []byte, wantHex string) {
+	t.Helper()
+	got := hex.EncodeToString(code)
+	if got != wantHex {
+		t.Errorf("code = %s, want %s", got, wantHex)
+	}
+}
+
+// TestRISCV_extendedPseudos pins the toolchain-synthesised instructions:
+// ANDN/ORN (XORI + AND/OR through the destination or TMP), the five-word
+// MIN/MAX expansion, the four-word rotate, ROR's compressed reverse shift
+// (C.SLLI when rd == rs1, both non-zero, 1 <= sll <= 63), the identical-
+// input MIN/MAX fold to C.MV, FABSD (FSGNJX.D), SEQZ and RDTIME (csrrs with
+// the time CSR).
+func TestRISCV_extendedPseudos(t *testing.T) {
+	t.Run("logic and minmax", func(t *testing.T) {
+		fn := firstTextRISCV(t, `#include "textflag.h"
+TEXT ·l(SB), NOSPLIT, $0
+	ANDN X19, X20, X21
+	ANDN X19, X20
+	ORN X20, X19
+	MAX X26, X28, X29
+	MIN X29, X30, X5
+	MAX X5, X5
+	MAX X5, X5, X6
+	SEQZ X5, X6
+	NEG X5, X6
+	NOT X5
+	RDTIME X5
+	RET
+`)
+		code := assembleRISCVHelper(t, fn)
+		// Words 0-10 up to the folded C.MV pair (halfwords 96 82 and 16 83),
+		// then SEQZ, NEG, NOT and RDTIME.
+		riscvWantsHex(t, code,
+			"93caf9ffb37a5a01"+"93cff9ff337afa01"+"934ffaffb3e9f901"+
+				"b32fae01b30ff041b34eae01b3fedf01b34ede01"+
+				"b3afee01b30ff041b342df01b3f25f00b3425f00"+
+				"9682"+"1683"+
+				"13b31200"+"33035040"+"93c2f2ff"+"f32210c0"+"67800000")
+	})
+
+	t.Run("rotate", func(t *testing.T) {
+		fn := firstTextRISCV(t, `#include "textflag.h"
+TEXT ·r(SB), NOSPLIT, $0
+	ROR X10, X11, X12
+	ROR X10, X11
+	ROR $63, X11
+	RORIW $31, X13, X14
+	RORIW $1, X14, X15
+	RORIW $3, X14
+	RORW X15, X16, X17
+	RORW $31, X13
+	RET
+`)
+		code := assembleRISCVHelper(t, fn)
+		// The third ROR carries the compressed C.SLLI (05 86) in mid-stream.
+		riscvWantsHex(t, code,
+			"b30fa040b39ff50133d6a50033e6cf00"+
+				"b30fa040b39ff501b3d5a500b3e5bf00"+
+				"93dff5038605b3e5bf00"+
+				"9bdff6011b97160033e7ef00"+
+				"9b5f17009b17f701b3e7ff00"+
+				"9b5f37001b17d70133e7ef00"+
+				"b30ff040bb1ff801bb58f800b3e81f01"+
+				"9bdff6019b961600b3e6df00"+"67800000")
+	})
+
+	t.Run("fp and branches", func(t *testing.T) {
+		fn := firstTextRISCV(t, `#include "textflag.h"
+TEXT ·f(SB), NOSPLIT, $0
+	FABSD F1, F2
+	FSGNJD F1, F0, F2
+	FMADDD F1, F2, F3, F4
+	FMSUBD F1, F2, F3, F4
+	FNMSUBD F1, F2, F3, F4
+	BGT X5, X6, tgt
+	BLE X5, X6, tgt
+	BGTU X5, X6, tgt
+	BLEU X5, X6, tgt
+tgt:
+	RDTIME X5
+	RET
+`)
+		code := assembleRISCVHelper(t, fn)
+		riscvWantsHex(t, code,
+			"53a11022"+"53011022"+"4382201a4782201a4b82201a"+
+				"63485300635653006364530063725300"+ // blt/bge/bltu/bgeu x6, x5
+				"f32210c0"+"67800000")
+	})
+}
+
+// TestRISCV_amoWords pins the full AMO family: every AMO carries aq and rl
+// (funct7 |= 3), LR is acquire (funct7 |= 2) and SC release (funct7 |= 1),
+// exactly as GOARCH=riscv64 go tool asm encodes them.
+func TestRISCV_amoWords(t *testing.T) {
+	fn := firstTextRISCV(t, `#include "textflag.h"
+TEXT ·amo(SB), NOSPLIT, $0
+	AMOSWAPW X5, (X6), X7
+	AMOSWAPD X5, (X6), X7
+	AMOADDW X5, (X6), X7
+	AMOADDD X5, (X6), X7
+	AMOANDW X5, (X6), X7
+	AMOANDD X5, (X6), X7
+	AMOORW X5, (X6), X7
+	AMOORD X5, (X6), X7
+	AMOXORW X5, (X6), X7
+	AMOXORD X5, (X6), X7
+	AMOMAXW X5, (X6), X7
+	AMOMAXD X5, (X6), X7
+	AMOMAXUW X5, (X6), X7
+	AMOMAXUD X5, (X6), X7
+	AMOMINUW X5, (X6), X7
+	AMOMINUD X5, (X6), X7
+	LRW (X5), X6
+	LRD (X5), X6
+	SCW X5, (X6), X7
+	SCD X5, (X6), X7
+	RET
+`)
+	code := assembleRISCVHelper(t, fn)
+	riscvWants(t, code,
+		0x0E5323AF, // amoswap.w
+		0x0E5333AF, // amoswap.d
+		0x065323AF, // amoaddd.w
+		0x065333AF, // amoadd.d
+		0x665323AF, // amoand.w
+		0x665333AF, // amoand.d
+		0x465323AF, // amoor.w
+		0x465333AF, // amoor.d
+		0x265323AF, // amoxor.w
+		0x265333AF, // amoxor.d
+		0xA65323AF, // amomax.w
+		0xA65333AF, // amomax.d
+		0xE65323AF, // amomaxu.w
+		0xE65333AF, // amomaxu.d
+		0xC65323AF, // amominu.w
+		0xC65333AF, // amominu.d
+		0x1402A32F, // lr.w (aq)
+		0x1402B32F, // lr.d
+		0x1A5323AF, // sc.w (rl)
+		0x1A5333AF, // sc.d
+	)
+}
+
+// TestRISCV_vectorWords pins the RVV slice and the VSET* encodings.  The
+// toolchain canonicalises an immediate avl to vsetivli even under the
+// VSETVLI spelling (`VSETVLI $15` and `VSETIVLI $15` come out byte-
+// identical), which is what the 0xC00 bit of the first word carries.
+func TestRISCV_vectorWords(t *testing.T) {
+	fn := firstTextRISCV(t, `#include "textflag.h"
+TEXT ·v(SB), NOSPLIT, $0
+	VSETVLI X5, E8, M8, TA, MA, X6
+	VSETIVLI $4, E32, M1, TA, MA, X0
+	VSETVLI $15, E32, M1, TA, MA, X12
+	VADDVV V1, V2, V3
+	VADDVX X12, V12, V12
+	VXORVV V8, V16, V24
+	VMSEQVX X12, V8, V0
+	VMSNEVV V8, V16, V0
+	VSLLVI $8, V28, V30
+	VSRLVI $25, V29, V29
+	VFIRSTM V0, X6
+	VIDV V12
+	VMV4RV V8, V24
+	VLE8V (X10), V8
+	VSE8V V24, (X10)
+	VSE32V V9, (X11)
+	VLSSEG4E32V (X14), X0, V0
+	VLSSEG8E32V (X10), X0, V4
+	RET
+`)
+	code := assembleRISCVHelper(t, fn)
+	riscvWants(t, code,
+		0x0C32F357, // vsetvli x6, x5, vtype 0xc3 (E8, M8, TA, MA)
+		0xCD027057, // vsetivli x0, 4
+		0xCD07F657, // vsetivli x12, 15: VSETVLI $15 canonicalises to the same word
+		0x022081D7, // vadd.vv v3, v2, v1
+		0x02C64657, // vadd.vx v12, v12, x12
+		0x2F040C57, // vxor.vv v24, v16, v8
+		0x62864057, // vmseq.vx v0, v8, x12
+		0x67040057, // vmsne.vv v0, v16, v8
+		0x97C43F57, // vsll.vi v30, v28, 8
+		0xA3DCBED7, // vsrl.vi v29, v29, 25
+		0x4208A357, // vmfirst.m x6, v0
+		0x5208A657, // vid.v v12
+		0x9E81BC57, // vmv4r.v v24, v8
+		0x02050407, // vle8.v v8, (x10)
+		0x02050C27, // vse8.v v24, (x10)
+		0x0205E4A7, // vse32.v v9, (x11)
+		0x6A076007, // vlsseg4e32.v v0, (x14), x0
+		0xEA056207, // vlsseg8e32.v v4, (x10), x0
+	)
+}

asm/riscv_frame.go

@@ -4,9 +4,10 @@
 package asm
 
 import (
+	"fmt"
 	"strings"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/ast"
+	"sourcedock.dev/petrbalvin/gasm-sdk/ast"
 )
 
 // RISC-V frame mapping, matching the Go toolchain's riscv64 backend.
@@ -32,6 +33,13 @@ import (
 // riscvFrameInfo holds the frame layout derived from a TEXT directive.
 type riscvFrameInfo struct {
 	autosize int // the real SP adjustment (locals + saved LR)
+
+	// Stack-split guard state: the toolchain emits the check for every
+	// non-NOSPLIT function whose autosize is nonzero (a zero autosize is
+	// "effectively NOSPLIT"); unlike amd64 and arm64 there is no leaf
+	// auto-NOSPLIT.
+	needSplit  bool
+	splitClass int // 0: <=StackSmall, 1: <=StackBig, 2: >StackBig
 }
 
 // riscvComputeFrame derives the frame layout for a TEXT function.
@@ -40,11 +48,34 @@ func riscvComputeFrame(t *ast.Text) riscvFrameInfo {
 	if frame != 0 || !riscvIsLeaf(t) {
 		// FixedFrameSize = 8: space for the saved link register.  A
 		// zero-frame non-leaf function still opens an 8-byte frame for LR.
-		return riscvFrameInfo{autosize: frame + 8}
+		autosize := frame + 8
+		fi := riscvFrameInfo{autosize: autosize}
+		if !hasNoSplitFlag(t) {
+			fi.needSplit = true
+			switch {
+			case autosize <= stackSmall:
+				fi.splitClass = 0
+			case autosize <= stackBig:
+				fi.splitClass = 1
+			default:
+				fi.splitClass = 2
+			}
+		}
+		return fi
 	}
 	return riscvFrameInfo{}
 }
 
+// hasNoSplitFlag reports whether the TEXT directive carries NOSPLIT.
+func hasNoSplitFlag(t *ast.Text) bool {
+	for _, f := range t.Flags {
+		if strings.EqualFold(f, "NOSPLIT") {
+			return true
+		}
+	}
+	return false
+}
+
 // riscvIsLeaf reports whether a function contains no call instructions.
 // CALL always links; JAL/JALR link only when their destination register is
 // the link register (X1), matching cmd/internal/obj/riscv's containsCall.
@@ -58,17 +89,24 @@ func riscvIsLeaf(t *ast.Text) bool {
 		case "CALL":
 			return false
 		case "JAL":
-			// JAL rd, target — a call only when rd is the link register.
+			// JAL rd, target, a call only when rd is the link register.
 			if len(in.Operands) >= 2 && regFromOperand(in.Operands[0]) == 1 {
 				return false
 			}
 		case "JALR":
-			// JALR rs1, rd — a call when rd is X1; JALR offset(rs1) always
-			// links to X1.
+			// JALR rd, offset(rs1) links when the destination register (the
+			// first operand) is X1; JALR rs1, rd links when the second
+			// register is X1; JALR offset(rs1) always links to X1.
 			if len(in.Operands) == 1 {
 				return false
 			}
-			if len(in.Operands) >= 2 && regFromOperand(in.Operands[1]) == 1 {
+			if isMemOperand(in.Operands[1]) {
+				if regFromOperand(in.Operands[0]) == 1 {
+					return false
+				}
+				continue
+			}
+			if regFromOperand(in.Operands[1]) == 1 {
 				return false
 			}
 		}
@@ -84,28 +122,99 @@ func riscvPrologue(fi riscvFrameInfo) []byte {
 		return nil
 	}
 	var out []byte
-	// MOV LR, -autosize(SP) — SD X1, -autosize(X2).  The negative offset is
-	// not compressible to C.SDSP (unsigned), so it stays 4 bytes.
-	out = append(out, wordLE(riscvSType(riscvEnc{0x23, 0x3, 0x00}, 2, 1, int32(-fi.autosize)))...)
-	// ADDI $-autosize, SP, SP — open the frame (C.ADDI when it fits).
-	out = append(out, riscvSPAdjust(int32(-fi.autosize))...)
-	// MOV LR, 0(SP) — SD X1, 0(X2) → C.SDSP X1, 0.
+	// MOV LR, -autosize(SP), SD X1, -autosize(X2).  The negative offset is
+	// not compressible to C.SDSP (unsigned), so it stays 4 bytes.  Beyond
+	// the imm12 range the toolchain materialises the address in X31.
+	if fits12(int32(-fi.autosize)) {
+		out = append(out, wordLE(riscvSType(riscvEnc{0x23, 0x3, 0x00}, 2, 1, int32(-fi.autosize)))...)
+	} else {
+		out = append(out, riscvAddressInX31(int32(-fi.autosize))...)
+		lo := int32(-fi.autosize) - (splitHi(int32(-fi.autosize)) << 12)
+		out = append(out, wordLE(riscvSType(riscvEnc{0x23, 0x3, 0x00}, 31, 1, lo))...)
+	}
+	// ADDI $-autosize, SP, SP, open the frame (C.ADDI when it fits; X31
+	// materialisation beyond imm12).
+	if fits12(int32(-fi.autosize)) {
+		out = append(out, riscvSPAdjust(int32(-fi.autosize))...)
+	} else {
+		out = append(out, riscvAddToSP(int32(-fi.autosize))...)
+	}
+	// MOV LR, 0(SP), SD X1, 0(X2) → C.SDSP X1, 0.
 	c := rvcSSP(0x7, 1, 0)
 	out = append(out, byte(c), byte(c>>8))
 	return out
 }
 
+func fits12(v int32) bool { return v >= -2048 && v <= 2047 }
+
+// splitHi returns the LUI half of the hi/lo split of v (what remains is the
+// sign-extended 12-bit low part).
+func splitHi(v int32) int32 {
+	_, high := splitRISCV32Imm(v)
+	return high
+}
+
+// riscvAddressInX31 materialises hi(v) into X31 against the stack pointer,
+// matching the toolchain's large-frame addressing: C.LUI (or LUI) X31, hi;
+// C.ADD (or ADD) X31, SP.
+func riscvAddressInX31(v int32) []byte {
+	return riscvAddressInX31WithBase(v, 2)
+}
+
+// riscvAddressInX31WithBase materialises hi(v) into X31 against an arbitrary
+// base register: LUI (or C.LUI) X31, hi; C.ADD X31, rs1.  The CR rs2 field
+// carries the full 5-bit register, so the compressed form is always
+// available.
+func riscvAddressInX31WithBase(v int32, rs1 int) []byte {
+	hi := splitHi(v)
+	var out []byte
+	if hi >= -32 && hi <= 31 {
+		c := rvcCI(0x3, 31, uint32(hi)&0x3F)
+		out = append(out, byte(c), byte(c>>8))
+	} else {
+		out = append(out, wordLE(riscvUType(riscvEnc{0x37, 0x0, 0x00}, 31, hi<<12))...)
+	}
+	c := rvcCR(0x9, 31, uint32(rs1))
+	return append(out, byte(c), byte(c>>8))
+}
+
+// riscvAddToSP adds v to SP through X31 for the values imm12 cannot carry:
+// C.LUI X31, hi; C.ADDIW X31, lo; C.ADD SP, X31 (the toolchain's form).
+func riscvAddToSP(v int32) []byte {
+	hi := splitHi(v)
+	lo := v - (hi << 12)
+	var out []byte
+	if hi >= -32 && hi <= 31 {
+		c := rvcCI(0x3, 31, uint32(hi)&0x3F)
+		out = append(out, byte(c), byte(c>>8))
+	} else {
+		out = append(out, wordLE(riscvUType(riscvEnc{0x37, 0x0, 0x00}, 31, hi<<12))...)
+	}
+	if lo >= -32 && lo <= 31 {
+		c := rvcCI(0x1, 31, uint32(lo)&0x3F)
+		out = append(out, byte(c), byte(c>>8))
+	} else {
+		out = append(out, wordLE(riscvIType(riscvEnc{0x1b, 0x0, 0x00}, 31, 31, lo))...)
+	}
+	c := rvcCR(0x9, 2, 31)
+	return append(out, byte(c), byte(c>>8))
+}
+
 // riscvReturn returns the bytes for a RET: the epilogue (restore LR and
 // deallocate the frame when present) followed by the uncompressed JALR X0,
 // 0(X1) the toolchain emits for RET (it never compresses RET to C.JR).
 func riscvReturn(fi riscvFrameInfo) []byte {
 	var out []byte
 	if fi.autosize != 0 {
-		// MOV 0(SP), LR — LD X1, 0(X2) → C.LDSP X1, 0.
+		// MOV 0(SP), LR, LD X1, 0(X2) → C.LDSP X1, 0.
 		c := rvcLSP(0x3, 1, 0)
 		out = append(out, byte(c), byte(c>>8))
-		// ADDI $autosize, SP, SP — close the frame (C.ADDI when it fits).
-		out = append(out, riscvSPAdjust(int32(fi.autosize))...)
+		// ADDI $autosize, SP, SP, close the frame (C.ADDI when it fits).
+		if fits12(int32(fi.autosize)) {
+			out = append(out, riscvSPAdjust(int32(fi.autosize))...)
+		} else {
+			out = append(out, riscvAddToSP(int32(fi.autosize))...)
+		}
 	}
 	// JALR X0, 0(X1).
 	return append(out, wordLE(riscvIType(riscvEnc{0x67, 0x0, 0x00}, 0, 1, 0))...)
@@ -133,33 +242,36 @@ func riscvFitsCAddi(imm int32) bool {
 }
 
 // riscvPrologueSpadjPC returns the function-relative byte offset where the
-// prologue has finished decrementing SP (the delta becomes autosize).
+// prologue has finished decrementing SP (the delta becomes autosize).  It is
+// computed from the same expansion functions the prologue emits, so the
+// large-frame X31 materialisations are counted: C.LUI + C.ADD before the SD,
+// C.LUI + ADDIW + C.ADD for the SP adjust.
 func riscvPrologueSpadjPC(fi riscvFrameInfo) int {
 	if fi.autosize == 0 {
 		return 0
 	}
-	// SD (4 bytes) + ADDI/C.ADDI (2 or 4 bytes).
-	return 4 + riscvSPAdjustLen(int32(-fi.autosize))
+	adj := int32(-fi.autosize)
+	if fits12(adj) {
+		// SD (4 bytes) + ADDI/C.ADDI (2 or 4 bytes).
+		return 4 + len(riscvSPAdjust(adj))
+	}
+	return len(riscvAddressInX31(adj)) + 4 + len(riscvAddToSP(adj))
 }
 
 // riscvReturnEpilogueLen returns the byte length of the RET's epilogue up to
-// (but not including) the final JALR — the point where SP is restored.
+// (but not including) the final JALR, the point where SP is restored.  The
+// small frame closes with C.LDSP + ADDI/C.ADDI; the large frame materialises
+// the adjustment through X31 (C.LUI + ADDIW + C.ADD).
 func riscvReturnEpilogueLen(fi riscvFrameInfo) int {
 	if fi.autosize == 0 {
 		return 0
 	}
-	// C.LDSP (2 bytes) + ADDI/C.ADDI (2 or 4 bytes).
-	return 2 + riscvSPAdjustLen(int32(fi.autosize))
-}
-
-func riscvSPAdjustLen(imm int32) int {
-	if imm != 0 && imm%16 == 0 && imm >= -512 && imm <= 511 {
-		return 2
+	adj := int32(fi.autosize)
+	if fits12(adj) {
+		// C.LDSP (2 bytes) + ADDI/C.ADDI (2 or 4 bytes).
+		return 2 + len(riscvSPAdjust(adj))
 	}
-	if riscvFitsCAddi(imm) {
-		return 2
-	}
-	return 4
+	return 2 + len(riscvAddToSP(adj))
 }
 
 // riscvResolvePseudo translates a pseudo-register memory reference into a
@@ -180,3 +292,73 @@ func riscvResolvePseudo(sym *ast.Symbol, fi riscvFrameInfo) (base int, off int32
 	}
 	return -1, 0
 }
+
+// riscvGuardLen returns the byte length of the stack-split guard prefix
+// including the inline morestack call (zero when the function needs no
+// guard).  Unlike amd64 and arm64, the toolchain places the morestack call
+// between the guard and the body: the guard branches forward over it.
+func riscvGuardLen(fi riscvFrameInfo) (int, error) {
+	g, _, err := riscvGuard(fi)
+	if err != nil {
+		return 0, err
+	}
+	return len(g), nil
+}
+
+// riscvGuard emits the stack-split guard prefix with the inline morestack
+// call: the branch skips forward over JAL X5 and JAL X0 straight into the
+// body; the JAL X5 carries the R_RISCV_JAL relocation.  All offsets are
+// relative to the guard itself, which sits at function offset 0.
+func riscvGuard(fi riscvFrameInfo) ([]byte, Reloc, error) {
+	if !fi.needSplit {
+		return nil, Reloc{}, nil
+	}
+	// MOV 16(g), X6 (g.stackguard0), g = X27.
+	out := wordLE(riscvIType(riscvEnc{0x03, 0x3, 0x00}, 6, 27, 16))
+	jalBack := func() []byte {
+		// JAL X0 back to the function start: it sits right after the JAL X5,
+		// so its displacement is minus the current offset.
+		return wordLE(riscvJType(0, int32(-len(out))))
+	}
+	var reloc Reloc
+	switch fi.splitClass {
+	case 0:
+		// BLTU X6, SP, done (+12: over the CALL and the JMP back)
+		out = append(out, wordLE(riscvBType(riscvEnc{0x63, 0x06, 0x00}, 6, 2, 12))...)
+		call := len(out)
+		reloc = Reloc{Off: call, After: call + 4, Name: "runtime\u00b7morestack_noctxt", Kind: RelRISCVJal}
+		out = append(out, wordLE(riscvJType(5, 0))...)
+		out = append(out, jalBack()...)
+	case 1:
+		// ADDI $-(framesize-StackSmall), SP, X7; BLTU X6, X7, done (+12)
+		off := int32(fi.autosize - stackSmall)
+		out = append(out, wordLE(riscvIType(riscvEnc{0x13, 0x0, 0x00}, 7, 2, -off))...)
+		out = append(out, wordLE(riscvBType(riscvEnc{0x63, 0x06, 0x00}, 6, 7, 12))...)
+		call := len(out)
+		reloc = Reloc{Off: call, After: call + 4, Name: "runtime\u00b7morestack_noctxt", Kind: RelRISCVJal}
+		out = append(out, wordLE(riscvJType(5, 0))...)
+		out = append(out, jalBack()...)
+	default:
+		// MOV $(framesize-StackSmall), X7; BLTU SP, X7, call;
+		// ADD $-(framesize-StackSmall), SP, X7; BLTU X6, X7, call
+		off := int32(fi.autosize - stackSmall)
+		mov := encodeRISCVLoadImm(7, off)
+		out = append(out, mov...)
+		addiLen := riscvItypeImmediateSize("ADDI", -off)
+		out = append(out, wordLE(riscvBType(riscvEnc{0x63, 0x06, 0x00}, 2, 7, int32(addiLen+8)))...)
+		addi, err := encodeRISCVItypeImmediate("ADDI", riscvEnc{0x13, 0x0, 0x00}, 7, 2, -off)
+		if err != nil {
+			// The ADDI expansion failed: the SP adjustment this class
+			// depends on is not emittable, and silently dropping it would
+			// corrupt every stack reference in the body.
+			return nil, Reloc{}, fmt.Errorf("stack-split guard: %w", err)
+		}
+		out = append(out, addi...)
+		out = append(out, wordLE(riscvBType(riscvEnc{0x63, 0x06, 0x00}, 6, 7, 12))...)
+		call := len(out)
+		reloc = Reloc{Off: call, After: call + 4, Name: "runtime\u00b7morestack_noctxt", Kind: RelRISCVJal}
+		out = append(out, wordLE(riscvJType(5, 0))...)
+		out = append(out, jalBack()...)
+	}
+	return out, reloc, nil
+}

asm/riscv_frame_test.go

@@ -6,7 +6,7 @@ package asm
 import (
 	"testing"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
 )
 
 // TestRISCVFrameSpadjAndLines checks that a framed function records its
@@ -65,6 +65,47 @@ TEXT ·framed(SB), NOSPLIT, $16-16
 	}
 }
 
+// TestRISCVFrameSpadjLargeFrame checks the stack-adjustment boundaries of a
+// frame past the imm12 range: the prologue materialises the LR-store address
+// and the SP adjustment through X31 (C.LUI + C.ADD + SD, then C.LUI + ADDIW +
+// C.ADD), so the SP boundary lands at PC 16, and the RET closes with
+// C.LDSP plus the same X31 adjustment, 10 bytes.  Regression: both helpers
+// assumed the small-frame prologue and reported 8 and 6.
+func TestRISCVFrameSpadjLargeFrame(t *testing.T) {
+	f, errs := parser.Parse("bigframe_riscv64.s", `#include "textflag.h"
+
+TEXT ·big(SB), NOSPLIT, $9000-8
+	MOV a+0(FP), X10
+	RET
+`)
+	if len(errs) > 0 {
+		t.Fatalf("parse: %v", errs)
+	}
+	img, err := AssembleFileRISCV(f)
+	if err != nil {
+		t.Fatalf("AssembleFileRISCV: %v", err)
+	}
+	fn := img.Funcs[0]
+
+	// autosize = 9008.  Prologue: C.LUI X31 + C.ADD X31,SP (4) + SD (4) +
+	// C.LUI X31 + ADDIW X31 + C.ADD SP,X31 (8) = 16 bytes to the SP boundary;
+	// C.SDSP X1 (2) follows, so the body starts at 18.
+	wantSpadj := []SpadjStep{{PC: 16, Value: 9008}, {PC: 36, Value: 0}}
+	if len(fn.Spadj) != len(wantSpadj) {
+		t.Fatalf("spadj = %v, want %v", fn.Spadj, wantSpadj)
+	}
+	for i := range wantSpadj {
+		if fn.Spadj[i] != wantSpadj[i] {
+			t.Errorf("spadj[%d] = %v, want %v", i, fn.Spadj[i], wantSpadj[i])
+		}
+	}
+	// The FP load materialises its 9016-byte offset through X31 as well
+	// (8 bytes), then RET's epilogue (C.LDSP + X31 adjust = 10) plus JALR.
+	if fn.Size != 18+8+14 {
+		t.Errorf("size = %d, want %d", fn.Size, 18+8+14)
+	}
+}
+
 // TestRISCVRegAliases checks the Go ABI register aliases that the toolchain
 // defines: LR is the link register (X1) and TMP is the assembler scratch
 // register (X31/T6).

asm/riscv_goobj_test.go

@@ -13,7 +13,7 @@ import (
 	"strings"
 	"testing"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
 )
 
 // TestGOObjectRISCVCallReloc checks that CALL sym(SB) emits a single JAL
@@ -74,6 +74,89 @@ DATA callee<>+0(SB)/8, $42
 		t.Error("ELF object missing R_RISCV_JAL relocation")
 	}
 }
+
+// TestELFRISCVPCRELLO12Anchor checks the psABI's LO12 pairing rule: the
+// R_RISCV_PCREL_LO12_I/S relocation must reference a symbol whose value is
+// the AUIPC site of its HI20 partner (psABI §8.4.9; cmd/link generates one
+// local text symbol per AUIPC for exactly this).  The emitter pairs each
+// HI20 (against the target symbol) with a LO12 against the .text section
+// symbol whose addend is the AUIPC's section-relative offset, so S + A is
+// the AUIPC address.
+func TestELFRISCVPCRELLO12Anchor(t *testing.T) {
+	f, errs := parser.Parse("k_riscv64.s", `
+#include "textflag.h"
+
+TEXT ·sb(SB), NOSPLIT, $0-0
+	MOV $answer<>(SB), X10
+	MOV answer<>(SB), X11
+	MOV X12, answer<>(SB)
+	RET
+
+GLOBL answer<>(SB), RODATA, $8
+DATA answer<>+0(SB)/8, $42
+`)
+	if len(errs) > 0 {
+		t.Fatalf("parse: %v", errs)
+	}
+	img, err := AssembleFileRISCV(f)
+	if err != nil {
+		t.Fatalf("AssembleFileRISCV: %v", err)
+	}
+	obj, err := img.ELFRISCVObject()
+	if err != nil {
+		t.Fatalf("ELFRISCVObject: %v", err)
+	}
+	ef, err := elf.NewFile(bytes.NewReader(obj))
+	if err != nil {
+		t.Fatalf("parse ELF: %v", err)
+	}
+	defer ef.Close()
+	if flags := binary.LittleEndian.Uint32(obj[48:]); flags != efRISCVFloatAbiDouble {
+		t.Errorf("e_flags = %#x, want %#x (EF_RISCV_FLOAT_ABI_DOUBLE)", flags, efRISCVFloatAbiDouble)
+	}
+	rela := ef.Section(".rela.text")
+	if rela == nil {
+		t.Fatal("missing .rela.text")
+	}
+	b, err := rela.Data()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(b) != 6*24 {
+		t.Fatalf(".rela.text holds %d entries, want six (three HI20/LO12 pairs)", len(b)/24)
+	}
+	le := binary.LittleEndian
+	wantLo := []uint32{rRISCVPCRELLO12I, rRISCVPCRELLO12I, rRISCVPCRELLO12S}
+	for p := range 3 {
+		auipc := 8 * p
+		hi := b[p*2*24:]
+		lo := b[(p*2+1)*24:]
+		if off := le.Uint64(hi[0:]); off != uint64(auipc) {
+			t.Errorf("pair %d: HI20 r_offset = %d, want %d (the AUIPC)", p, off, auipc)
+		}
+		if typ := uint32(le.Uint64(hi[8:])); typ != rRISCVPCRELHI20 {
+			t.Errorf("pair %d: HI20 type = %d, want %d", p, typ, rRISCVPCRELHI20)
+		}
+		if sym := int(le.Uint64(hi[8:]) >> 32); sym == 0 || sym == 1 {
+			t.Errorf("pair %d: HI20 against symbol %d, want the target", p, sym)
+		}
+		if off := le.Uint64(lo[0:]); off != uint64(auipc+4) {
+			t.Errorf("pair %d: LO12 r_offset = %d, want %d", p, off, auipc+4)
+		}
+		if typ := uint32(le.Uint64(lo[8:])); typ != wantLo[p] {
+			t.Errorf("pair %d: LO12 type = %d, want %d", p, typ, wantLo[p])
+		}
+		// The LO12 must denote the AUIPC site: the .text section symbol
+		// (index 1) plus the AUIPC's section-relative offset as addend.
+		if sym := int(le.Uint64(lo[8:]) >> 32); sym != 1 {
+			t.Errorf("pair %d: LO12 against symbol %d, want 1 (the .text section symbol)", p, sym)
+		}
+		if add := int64(le.Uint64(lo[16:])); add != int64(auipc) {
+			t.Errorf("pair %d: LO12 addend = %d, want %d (S + A = the AUIPC address)", p, add, auipc)
+		}
+	}
+}
+
 func TestGOObjectRISCVStructure(t *testing.T) {
 	f, errs := parser.Parse("k_riscv64.s", `
 #include "textflag.h"
@@ -141,7 +224,7 @@ DATA answer<>+0(SB)/8, $42
 	first := int(le.Uint32(relocIdx[4*(4+4):]))
 	wantType := []uint16{relocRISCVPcrelItype, relocRISCVPcrelItype, relocRISCVPcrelStype}
 	wantOffAbs := []int{0, 8, 16}
-	for i := 0; i < 3; i++ {
+	for i := range 3 {
 		e := relocs[(first+i)*23:]
 		if int32(le.Uint32(e[0:])) != int32(wantOffAbs[i]) || e[4] != 8 || le.Uint16(e[5:]) != wantType[i] ||
 			le.Uint32(e[15:]) != pkgIdxSelf || le.Uint32(e[19:]) != 0 {
@@ -221,7 +304,7 @@ func main() {
 		t.Fatalf("baseline build: %v\n%s", err, buildLog)
 	}
 	var pkgArch, work, linkLine, asmObj string
-	for _, line := range strings.Split(string(buildLog), "\n") {
+	for line := range strings.SplitSeq(string(buildLog), "\n") {
 		switch {
 		case strings.HasPrefix(line, "WORK="):
 			work = strings.TrimPrefix(line, "WORK=")
@@ -279,7 +362,7 @@ func main() {
 	newArch := filepath.Join(dir, "pkg.a")
 	args := []string{"tool", "pack", "c", newArch}
 	seen := map[string]bool{}
-	for _, m := range strings.Fields(string(listOut)) {
+	for m := range strings.FieldsSeq(string(listOut)) {
 		if seen[m] {
 			continue
 		}

asm/vex.go

@@ -36,12 +36,12 @@ const (
 	vexNDS3Imm
 	// vexExtract is the lane-extract form `OP $imm, ysrc, xdst`: ModRM.reg =
 	// ysrc (op1), ModRM.rm = xdst or memory (op2), imm8 = op0.  The YMM
-	// source lives in the reg field, the destination in r/m — the PEXTR-style
+	// source lives in the reg field, the destination in r/m, the PEXTR-style
 	// layout.  VEXTRACTI128 and VEXTRACTF128 use this shape.
 	vexExtract
 	// vexRMRev is the reversed two-operand form `OP src, dst` with the source
-	// in ModRM.reg and the destination in r/m — the layout of the EVEX
-	// narrowing stores (VPMOVDW, VPMOVQD).
+	// in ModRM.reg and the destination in r/m, the layout of the EVEX
+	// narrowing stores (VPMOVDW, VPMOVQD) and of the non-temporal VMOVNTDQ.
 	vexRMRev
 	// vexRMSrcLen is the two-operand conversion form `OP src, dst` whose
 	// vector length follows the source: the packed-double → dword
@@ -52,6 +52,30 @@ const (
 	vexRMSrcLen
 	// vexZero is the no-operand form (VZEROUPPER).
 	vexZero
+	// vexZeroAll is the no-operand form that zeroes the full upper state
+	// (VZEROALL, the L = 1 twin of VZEROUPPER).
+	vexZeroAll
+	// vexNDS3GPR is the three-operand NDS form over general-purpose
+	// registers (ANDN, MULX): reg = dst, vvvv = src1, rm = src2, L = 0.
+	vexNDS3GPR
+	// vexImmRMGPR is the immediate form over general-purpose registers
+	// (RORX): reg = dst, rm = src, imm8 = op0, L = 0.
+	vexImmRMGPR
+	// vexRMOpGPR is the two-operand /digit form over general-purpose
+	// registers (BLSI, BLSMSK, BLSR): ModRM.reg = /digit, ModRM.rm = src
+	// (op0), VEX.vvvv = dst (op1), L = 0.
+	vexRMOpGPR
+	// vexCountGPR is the three-operand count form over general-purpose
+	// registers (SHLX, SHRX, SARX, BEXTR, BZHI): the first operand rides
+	// VEX.vvvv and the second is r/m, the opposite pairing of the ANDN
+	// family, with reg = dst (op2), L = 0.
+	vexCountGPR
+	// vexExtractGPR is the lane-extract-to-GPR form `OP $imm, xsrc, GPR/mem
+	// dst`: ModRM.reg = xsrc (op1), ModRM.rm = destination (op2), imm8 =
+	// op0, the VPEXTRB/W/D/Q layout.  EVEX only; the destination never
+	// carries a vector length, so the register the L'L field follows is the
+	// XMM source.
+	vexExtractGPR
 )
 
 // vexSpec describes one VEX instruction's encoding parameters.
@@ -68,7 +92,7 @@ type vexSpec struct {
 // incrementally; every entry is covered by a byte-for-byte ground-truth test
 // against the Go assembler.
 var vexTable = map[string]vexSpec{
-	// VEX.128/256.66.0F.WIG — integer arithmetic / logic / compare.
+	// VEX.128/256.66.0F.WIG, integer arithmetic / logic / compare.
 	"VPADDD":      {1, 0xFE, 0, 1, -1, vexNDS3},
 	"VPADDQ":      {1, 0xD4, 0, 1, -1, vexNDS3},
 	"VPSUBD":      {1, 0xFA, 0, 1, -1, vexNDS3},
@@ -82,7 +106,7 @@ var vexTable = map[string]vexSpec{
 	"VPUNPCKHDQ":  {1, 0x6A, 0, 1, -1, vexNDS3},
 	"VPUNPCKLQDQ": {1, 0x6C, 0, 1, -1, vexNDS3},
 	"VPACKSSDW":   {1, 0x6B, 0, 1, -1, vexNDS3},
-	// VEX.256.66.0F38.W0 — dword permute (three-operand NDS form).
+	// VEX.256.66.0F38.W0, dword permute (three-operand NDS form).
 	"VPERMD": {2, 0x36, 0, 1, -1, vexNDS3},
 	// VEX.128/256.66.0F38.WIG.
 	"VPMULLD":  {2, 0x40, 0, 1, -1, vexNDS3},
@@ -90,14 +114,14 @@ var vexTable = map[string]vexSpec{
 	"VPSHUFB":  {2, 0x00, 0, 1, -1, vexNDS3},
 	"VPCMPGTQ": {2, 0x37, 0, 1, -1, vexNDS3},
 
-	// VEX.128/256.66.0F.WIG — packed double-precision arithmetic / logic.
+	// VEX.128/256.66.0F.WIG, packed double-precision arithmetic / logic.
 	"VADDPD": {1, 0x58, 0, 1, -1, vexNDS3},
 	"VMULPD": {1, 0x59, 0, 1, -1, vexNDS3},
 	"VSUBPD": {1, 0x5C, 0, 1, -1, vexNDS3},
 	"VDIVPD": {1, 0x5E, 0, 1, -1, vexNDS3},
 	"VMINPD": {1, 0x5D, 0, 1, -1, vexNDS3},
 	"VMAXPD": {1, 0x5F, 0, 1, -1, vexNDS3},
-	// VEX.128/256.0F.WIG — packed single-precision arithmetic.
+	// VEX.128/256.0F.WIG, packed single-precision arithmetic.
 	"VADDPS":    {1, 0x58, 0, 0, -1, vexNDS3},
 	"VMULPS":    {1, 0x59, 0, 0, -1, vexNDS3},
 	"VSUBPS":    {1, 0x5C, 0, 0, -1, vexNDS3},
@@ -107,7 +131,7 @@ var vexTable = map[string]vexSpec{
 	"VXORPD":    {1, 0x57, 0, 1, -1, vexNDS3},
 	"VUNPCKHPD": {1, 0x15, 0, 1, -1, vexNDS3},
 	"VUNPCKLPD": {1, 0x14, 0, 1, -1, vexNDS3},
-	// VEX.128.F2.0F.WIG — scalar double-precision arithmetic (the packed
+	// VEX.128.F2.0F.WIG, scalar double-precision arithmetic (the packed
 	// opcodes with an F2 pp).
 	"VADDSD": {1, 0x58, 0, 3, -1, vexNDS3},
 	"VSUBSD": {1, 0x5C, 0, 3, -1, vexNDS3},
@@ -115,7 +139,7 @@ var vexTable = map[string]vexSpec{
 	"VDIVSD": {1, 0x5E, 0, 3, -1, vexNDS3},
 	"VMINSD": {1, 0x5D, 0, 3, -1, vexNDS3},
 	"VMAXSD": {1, 0x5F, 0, 3, -1, vexNDS3},
-	// VEX.128.F3.0F.WIG — scalar single-precision arithmetic (the packed
+	// VEX.128.F3.0F.WIG, scalar single-precision arithmetic (the packed
 	// opcodes with an F3 pp).
 	"VADDSS": {1, 0x58, 0, 2, -1, vexNDS3},
 	"VSUBSS": {1, 0x5C, 0, 2, -1, vexNDS3},
@@ -123,10 +147,16 @@ var vexTable = map[string]vexSpec{
 	"VDIVSS": {1, 0x5E, 0, 2, -1, vexNDS3},
 	"VMINSS": {1, 0x5D, 0, 2, -1, vexNDS3},
 	"VMAXSS": {1, 0x5F, 0, 2, -1, vexNDS3},
-	// VEX.128/256.66.0F38.W1 — fused multiply-add (NDS form).
+	// VEX.128/256.66.0F38.W1, fused multiply-add (NDS form).
 	"VFMADD231PD": {2, 0xB8, 1, 1, -1, vexNDS3},
+	// Scalar fused multiply-add (NDS form).  The Go assembler carries the
+	// same 66 prefix as the packed forms on every FMA row, and W1 on the
+	// double-precision spellings, so SD shares PD's prefix/W pair and the
+	// scalar width rides on the W bit.
+	"VFMADD213SD":  {2, 0xA9, 1, 1, -1, vexNDS3},
+	"VFNMADD231SD": {2, 0xBD, 1, 1, -1, vexNDS3},
 
-	// VEX.128/256.66.0F38.WIG — sign/zero extend and broadcast (reg=dst, rm=src,
+	// VEX.128/256.66.0F38.WIG, sign/zero extend and broadcast (reg=dst, rm=src,
 	// no vvvv).
 	"VPMOVSXWD":    {2, 0x23, 0, 1, -1, vexRM},
 	"VPMOVSXDQ":    {2, 0x25, 0, 1, -1, vexRM},
@@ -141,70 +171,131 @@ var vexTable = map[string]vexSpec{
 	"VPMOVZXWQ":    {2, 0x34, 0, 1, -1, vexRM},
 	"VPBROADCASTD": {2, 0x58, 0, 1, -1, vexRM},
 	"VPBROADCASTQ": {2, 0x59, 0, 1, -1, vexRM},
-	// VEX.128/256.F3.0F.WIG — signed dword to packed double conversion
+	"VPBROADCASTB": {2, 0x78, 0, 1, -1, vexRM},
+	"VPBROADCASTW": {2, 0x79, 0, 1, -1, vexRM},
+	// VEX.128/256.F3.0F.WIG, signed dword to packed double conversion
 	// (reg=dst, rm=src, no vvvv; the length follows the destination).
 	"VCVTDQ2PD": {1, 0xE6, 0, 2, -1, vexRM},
-	// VEX.128/256.0F.WIG — signed dword to packed single conversion
+	// VEX.128/256.0F.WIG, signed dword to packed single conversion
 	// (reg=dst, rm=src, no vvvv, no mandatory prefix).
 	"VCVTDQ2PS": {1, 0x5B, 0, 0, -1, vexRM},
-	// VEX.128/256.0F.WIG — packed single to packed double conversion
+	// VEX.128/256.0F.WIG, packed single to packed double conversion
 	// (reg=dst, rm=src; the destination is the wide operand and sets the
 	// length).  Intel's maps prescribe the F3 prefix here (VEX.pp = 10), but
 	// the Go assembler emits the instruction with pp = 00, and gasm follows
-	// the Go assembler's bytes — its machine code is the oracle, not the
+	// the Go assembler's bytes, its machine code is the oracle, not the
 	// manual.
 	"VCVTPS2PD": {1, 0x5A, 0, 0, -1, vexRM},
-	// VEX.128.F2.0F.WIG — duplicate the low double of each 128-bit lane
+	// VEX.128.F2.0F.WIG, duplicate the low double of each 128-bit lane
 	// (reg=dst, rm=src, no vvvv; the length follows the destination).
 	"VMOVDDUP": {1, 0x12, 0, 3, -1, vexRM},
-	// VEX.128/256.66.0F.WIG — move mask to a GPR (reg=gpr dst, rm=vec src).
+	// VEX.128/256.66.0F.WIG, move mask to a GPR (reg=gpr dst, rm=vec src).
 	"VPMOVMSKB": {1, 0xD7, 0, 1, -1, vexRM},
 	"VMOVMSKPS": {1, 0x50, 0, 0, -1, vexRM}, // no 66 prefix (that would be VMOVMSKPD)
 
-	// VEX.128/256.66.0F.WIG — immediate shifts (opdigit selects the shift).
+	// VEX.128/256.66.0F.WIG, immediate shifts (opdigit selects the shift).
 	"VPSLLD": {1, 0x72, 0, 1, 6, vexShiftImm},
 	"VPSRAD": {1, 0x72, 0, 1, 4, vexShiftImm},
 	"VPSRLD": {1, 0x72, 0, 1, 2, vexShiftImm},
 	"VPSRLQ": {1, 0x73, 0, 1, 2, vexShiftImm},
 	"VPSLLQ": {1, 0x73, 0, 1, 6, vexShiftImm},
 
-	// VEX.128/256.66.0F.WIG — immediate shuffle (reg=dst, rm=src, imm8).
+	// VEX.128/256.66.0F.WIG, immediate shuffle (reg=dst, rm=src, imm8).
 	"VPSHUFD": {1, 0x70, 0, 1, -1, vexImmRM},
-	// VEX.256.66.0F3A.W1 — qword permute (reg=dst, rm=src, imm8).
+	// VEX.256.66.0F3A.W1, qword permute (reg=dst, rm=src, imm8).
 	"VPERMQ": {3, 0x00, 1, 1, -1, vexImmRM},
 
-	// VEX.128/256.66.0F.WIG — two-source shuffle (reg=dst, vvvv=src1, rm=src2,
+	// VEX.128/256.66.0F.WIG, two-source shuffle (reg=dst, vvvv=src1, rm=src2,
 	// imm8).
 	"VSHUFPD": {1, 0xC6, 0, 1, -1, vexNDS3Imm},
-	// VEX.256.66.0F3A.W0 — permute / insert (same shape; VINSERTI128's rm is
+	// VEX.256.66.0F3A.W0, permute / insert (same shape; VINSERTI128's rm is
 	// the XMM or memory source).
 	"VPERM2I128":  {3, 0x46, 0, 1, -1, vexNDS3Imm},
 	"VINSERTI128": {3, 0x38, 0, 1, -1, vexNDS3Imm},
 
-	// VEX.256.66.0F3A.W0 — lane extract (reg=YMM src, rm=XMM/memory dst, imm8).
+	// VEX.256.66.0F3A.W0, lane extract (reg=YMM src, rm=XMM/memory dst, imm8).
 	"VEXTRACTI128": {3, 0x39, 0, 1, -1, vexExtract},
 	"VEXTRACTF128": {3, 0x19, 0, 1, -1, vexExtract},
-	// VEX.128/256.66.0F3A.W0 — half-precision convert back ($imm, src, dst:
-	// reg=src, rm=XMM/memory dst, imm8 — the extract layout).
+	// VEX.128/256.66.0F3A.W0, half-precision convert back ($imm, src, dst:
+	// reg=src, rm=XMM/memory dst, imm8, the extract layout).
 	"VCVTPS2PH": {3, 0x1D, 0, 1, -1, vexExtract},
 
-	// VEX.128.0F.W0 — no operands.
+	// VEX.128.0F.W0, no operands.
 	"VZEROUPPER": {1, 0x77, 0, 0, -1, vexZero},
+	// VEX.256.0F.W0, zero all vector registers (the L = 1 twin).
+	"VZEROALL": {1, 0x77, 0, 0, -1, vexZeroAll},
+	// VEX.128/256.66.0F38, byte shuffle shifts and the packed byte compare.
+	"VPSLLDQ":  {1, 0x73, 0, 1, 7, vexShiftImm},
+	"VPSRLDQ":  {1, 0x73, 0, 1, 3, vexShiftImm},
+	"VPCMPEQB": {1, 0x74, 0, 1, -1, vexNDS3},
+	// VEX.128/256.0F.WIG, packed single XOR (NDS form).
+	"VXORPS": {1, 0x57, 0, 0, -1, vexNDS3},
+	// VEX.256.66.0F3A.W0, two-source permutes and blends with an imm8 control.
+	"VPERM2F128": {3, 0x06, 0, 1, -1, vexNDS3Imm},
+	"VPBLENDD":   {3, 0x02, 0, 1, -1, vexNDS3Imm},
+	// VEX.128/256.66.0F3A.WIG, byte align (NDS + imm8); the ZMM spelling
+	// falls through to the EVEX table.
+	"VPALIGNR": {3, 0x0F, 0, 1, -1, vexNDS3Imm},
+	// VEX.128/256.66.0F3A.W0, carry-less multiply ($imm, src2, src1, dst).
+	"VPCLMULQDQ": {3, 0x44, 0, 1, -1, vexNDS3Imm},
+	// VEX.128/256.66.0F3A.W1, GF(2^8) affine transform (NDS + imm8).
+	"VGF2P8AFFINEQB": {3, 0xCE, 1, 1, -1, vexNDS3Imm},
+	// BMI1/BMI2 general-register VEX forms (see vexNDS3GPR/vexImmRMGPR).
+	"ANDNL": {2, 0xF2, 0, 0, -1, vexNDS3GPR},
+	"ANDNQ": {2, 0xF2, 1, 0, -1, vexNDS3GPR},
+	"MULXL": {2, 0xF6, 0, 3, -1, vexNDS3GPR},
+	"MULXQ": {2, 0xF6, 1, 3, -1, vexNDS3GPR},
+	// VEX.NDS.LZ.0F38, the BMI2 three-operand bit ops: BEXTR and BZHI
+	// share the F7/F5 opcodes across W, the variable shifts carry their
+	// direction in the prefix (SHLX 66, SHRX F2, SARX F3) and PDEP/PEXT
+	// in F2/F3.
+	"BEXTRL": {2, 0xF7, 0, 0, -1, vexCountGPR},
+	"BEXTRQ": {2, 0xF7, 1, 0, -1, vexCountGPR},
+	"BZHIL":  {2, 0xF5, 0, 0, -1, vexCountGPR},
+	"BZHIQ":  {2, 0xF5, 1, 0, -1, vexCountGPR},
+	"SARXL":  {2, 0xF7, 0, 2, -1, vexCountGPR},
+	"SARXQ":  {2, 0xF7, 1, 2, -1, vexCountGPR},
+	"SHLXL":  {2, 0xF7, 0, 1, -1, vexCountGPR},
+	"SHLXQ":  {2, 0xF7, 1, 1, -1, vexCountGPR},
+	"SHRXL":  {2, 0xF7, 0, 3, -1, vexCountGPR},
+	"SHRXQ":  {2, 0xF7, 1, 3, -1, vexCountGPR},
+	"PDEPL":  {2, 0xF5, 0, 3, -1, vexNDS3GPR},
+	"PDEPQ":  {2, 0xF5, 1, 3, -1, vexNDS3GPR},
+	"PEXTL":  {2, 0xF5, 0, 2, -1, vexNDS3GPR},
+	"PEXTQ":  {2, 0xF5, 1, 2, -1, vexNDS3GPR},
+	// VEX.LZ.0F38.W, the BMI1 unary bit ops (src, dst: ModRM.reg = /digit,
+	// rm = src, vvvv = dst).
+	"BLSIL":   {2, 0xF3, 0, 0, 3, vexRMOpGPR},
+	"BLSIQ":   {2, 0xF3, 1, 0, 3, vexRMOpGPR},
+	"BLSMSKL": {2, 0xF3, 0, 0, 2, vexRMOpGPR},
+	"BLSMSKQ": {2, 0xF3, 1, 0, 2, vexRMOpGPR},
+	"BLSRL":   {2, 0xF3, 0, 0, 1, vexRMOpGPR},
+	"BLSRQ":   {2, 0xF3, 1, 0, 1, vexRMOpGPR},
+	"RORXL":   {3, 0xF0, 0, 3, -1, vexImmRMGPR},
+	"RORXQ":   {3, 0xF0, 1, 3, -1, vexImmRMGPR},
 
-	// VEX.128.0F.W0 — mask-register test (KTESTW k1, k2: reg = dst, rm = src).
+	// VEX.128.0F.W0, mask-register test (KTESTW k1, k2: reg = dst, rm = src).
 	"KTESTW": {1, 0x99, 0, 0, -1, vexRM},
 
-	// VEX.66.0F38.W0 — broadcast a single/double to all lanes (reg=dst,
+	// VEX.66.0F38.W0, broadcast a single/double to all lanes (reg=dst,
 	// rm=scalar memory; SD is 256-bit only).
 	"VBROADCASTSS": {2, 0x18, 0, 1, -1, vexRM},
 	"VBROADCASTSD": {2, 0x19, 0, 1, -1, vexRM},
-	// VEX.66.0F38.W0 — half-precision convert (reg=dst, rm=half-width
+	// VEX.256.66.0F38.W0, broadcast a 128-bit lane into both halves of a
+	// YMM (the encoder rejects an XMM destination, as go tool asm does).
+	"VBROADCASTI128": {2, 0x5A, 0, 1, -1, vexRM},
+	// VEX.128/256.66.0F.WIG, non-temporal store (vector source in reg,
+	// memory destination in rm).
+	"VMOVNTDQ": {1, 0xE7, 0, 1, -1, vexRMRev},
+	// VEX.128/256.66.0F38.W0, test (reg=dst, rm=src, no vvvv).
+	"VPTEST": {2, 0x17, 0, 1, -1, vexRM},
+	// VEX.66.0F38.W0, half-precision convert (reg=dst, rm=half-width
 	// source).
 	"VCVTPH2PS": {2, 0x13, 0, 1, -1, vexRM},
-	// VEX.F3.0F.WIG — replicate even/odd singles (reg=dst, rm=src).
+	// VEX.F3.0F.WIG, replicate even/odd singles (reg=dst, rm=src).
 	"VMOVSLDUP": {1, 0x12, 0, 2, -1, vexRM},
 	"VMOVSHDUP": {1, 0x16, 0, 2, -1, vexRM},
-	// VEX.66.0F.WIG — packed double to packed single conversion, the X/Y
+	// VEX.66.0F.WIG, packed double to packed single conversion, the X/Y
 	// spellings: the destination is always XMM and the spelling fixes the
 	// source length (X = 128, Y = 256).
 	"VCVTPD2PSX": {1, 0x5A, 0, 1, -1, vexRMSrcLen},
@@ -228,18 +319,140 @@ var vexTable = map[string]vexSpec{
 	"VCVTSI2SSL": {1, 0x2A, 0, 2, -1, vexNDS3},
 	"VCVTSI2SSQ": {1, 0x2A, 1, 2, -1, vexNDS3},
 
-	// VEX.128/256.66.0F.WIG — word shifts (opdigit selects the shift).
+	// VEX.128/256.66.0F.WIG, word shifts (opdigit selects the shift).
 	"VPSRLW": {1, 0x71, 0, 1, 2, vexShiftImm},
 	"VPSRAW": {1, 0x71, 0, 1, 4, vexShiftImm},
 	"VPSLLW": {1, 0x71, 0, 1, 6, vexShiftImm},
 
-	// VEX.F2.0F — packed double to packed dword conversions, truncating and
+	// VEX.F2.0F, packed double to packed dword conversions, truncating and
 	// non-truncating.  The destination is always XMM; the X/Y spellings fix
-	// the source length (XMM/YMM), and VEX.L follows it — see vexSrcLen.
+	// the source length (XMM/YMM), and VEX.L follows it, see vexSrcLen.
 	"VCVTPD2DQX":  {1, 0xE6, 0, 3, -1, vexRMSrcLen},
 	"VCVTPD2DQY":  {1, 0xE6, 0, 3, -1, vexRMSrcLen},
 	"VCVTTPD2DQX": {1, 0xE6, 0, 1, -1, vexRMSrcLen},
 	"VCVTTPD2DQY": {1, 0xE6, 0, 1, -1, vexRMSrcLen},
+
+	// --- the VEX forms the avx512enc corpus exercises alongside the EVEX
+	// spellings, read off the toolchain opcode tables ---
+	"VAESDEC":           {2, 0xDE, 0, 1, -1, vexNDS3},
+	"VAESDECLAST":       {2, 0xDF, 0, 1, -1, vexNDS3},
+	"VAESENC":           {2, 0xDC, 0, 1, -1, vexNDS3},
+	"VAESENCLAST":       {2, 0xDD, 0, 1, -1, vexNDS3},
+	"VANDNPD":           {1, 0x55, 0, 1, -1, vexNDS3},
+	"VANDPD":            {1, 0x54, 0, 1, -1, vexNDS3},
+	"VCOMISD":           {1, 0x2F, 0, 1, -1, vexRM},
+	"VCVTSD2SS":         {1, 0x5A, 0, 3, -1, vexNDS3},
+	"VCVTSS2SD":         {1, 0x5A, 0, 2, -1, vexNDS3},
+	"VFMADD132PD":       {2, 0x98, 1, 1, -1, vexNDS3},
+	"VFMADD132PS":       {2, 0x98, 0, 1, -1, vexNDS3},
+	"VFMADD132SD":       {2, 0x99, 1, 1, -1, vexNDS3},
+	"VFMADD132SS":       {2, 0x99, 0, 1, -1, vexNDS3},
+	"VFMADD213PD":       {2, 0xA8, 1, 1, -1, vexNDS3},
+	"VFMADD213PS":       {2, 0xA8, 0, 1, -1, vexNDS3},
+	"VFMADD213SS":       {2, 0xA9, 0, 1, -1, vexNDS3},
+	"VFMADD231PS":       {2, 0xB8, 0, 1, -1, vexNDS3},
+	"VFMADD231SD":       {2, 0xB9, 1, 1, -1, vexNDS3},
+	"VFMADD231SS":       {2, 0xB9, 0, 1, -1, vexNDS3},
+	"VFMADDSUB132PD":    {2, 0x96, 1, 1, -1, vexNDS3},
+	"VFMADDSUB132PS":    {2, 0x96, 0, 1, -1, vexNDS3},
+	"VFMADDSUB213PD":    {2, 0xA6, 1, 1, -1, vexNDS3},
+	"VFMADDSUB213PS":    {2, 0xA6, 0, 1, -1, vexNDS3},
+	"VFMADDSUB231PD":    {2, 0xB6, 1, 1, -1, vexNDS3},
+	"VFMADDSUB231PS":    {2, 0xB6, 0, 1, -1, vexNDS3},
+	"VFMSUB132PD":       {2, 0x9A, 1, 1, -1, vexNDS3},
+	"VFMSUB132PS":       {2, 0x9A, 0, 1, -1, vexNDS3},
+	"VFMSUB132SD":       {2, 0x9B, 1, 1, -1, vexNDS3},
+	"VFMSUB132SS":       {2, 0x9B, 0, 1, -1, vexNDS3},
+	"VFMSUB213PD":       {2, 0xAA, 1, 1, -1, vexNDS3},
+	"VFMSUB213PS":       {2, 0xAA, 0, 1, -1, vexNDS3},
+	"VFMSUB213SD":       {2, 0xAB, 1, 1, -1, vexNDS3},
+	"VFMSUB213SS":       {2, 0xAB, 0, 1, -1, vexNDS3},
+	"VFMSUB231PD":       {2, 0xBA, 1, 1, -1, vexNDS3},
+	"VFMSUB231PS":       {2, 0xBA, 0, 1, -1, vexNDS3},
+	"VFMSUB231SD":       {2, 0xBB, 1, 1, -1, vexNDS3},
+	"VFMSUB231SS":       {2, 0xBB, 0, 1, -1, vexNDS3},
+	"VFMSUBADD132PD":    {2, 0x97, 1, 1, -1, vexNDS3},
+	"VFMSUBADD132PS":    {2, 0x97, 0, 1, -1, vexNDS3},
+	"VFMSUBADD213PD":    {2, 0xA7, 1, 1, -1, vexNDS3},
+	"VFMSUBADD213PS":    {2, 0xA7, 0, 1, -1, vexNDS3},
+	"VFMSUBADD231PD":    {2, 0xB7, 1, 1, -1, vexNDS3},
+	"VFMSUBADD231PS":    {2, 0xB7, 0, 1, -1, vexNDS3},
+	"VFNMADD132PD":      {2, 0x9C, 1, 1, -1, vexNDS3},
+	"VFNMADD132PS":      {2, 0x9C, 0, 1, -1, vexNDS3},
+	"VFNMADD132SD":      {2, 0x9D, 1, 1, -1, vexNDS3},
+	"VFNMADD132SS":      {2, 0x9D, 0, 1, -1, vexNDS3},
+	"VFNMADD213PD":      {2, 0xAC, 1, 1, -1, vexNDS3},
+	"VFNMADD213PS":      {2, 0xAC, 0, 1, -1, vexNDS3},
+	"VFNMADD213SD":      {2, 0xAD, 1, 1, -1, vexNDS3},
+	"VFNMADD213SS":      {2, 0xAD, 0, 1, -1, vexNDS3},
+	"VFNMADD231PD":      {2, 0xBC, 1, 1, -1, vexNDS3},
+	"VFNMADD231PS":      {2, 0xBC, 0, 1, -1, vexNDS3},
+	"VFNMADD231SS":      {2, 0xBD, 0, 1, -1, vexNDS3},
+	"VFNMSUB132PD":      {2, 0x9E, 1, 1, -1, vexNDS3},
+	"VFNMSUB132PS":      {2, 0x9E, 0, 1, -1, vexNDS3},
+	"VFNMSUB132SD":      {2, 0x9F, 1, 1, -1, vexNDS3},
+	"VFNMSUB132SS":      {2, 0x9F, 0, 1, -1, vexNDS3},
+	"VFNMSUB213PD":      {2, 0xAE, 1, 1, -1, vexNDS3},
+	"VFNMSUB213PS":      {2, 0xAE, 0, 1, -1, vexNDS3},
+	"VFNMSUB213SD":      {2, 0xAF, 1, 1, -1, vexNDS3},
+	"VFNMSUB213SS":      {2, 0xAF, 0, 1, -1, vexNDS3},
+	"VFNMSUB231PD":      {2, 0xBE, 1, 1, -1, vexNDS3},
+	"VFNMSUB231PS":      {2, 0xBE, 0, 1, -1, vexNDS3},
+	"VFNMSUB231SD":      {2, 0xBF, 1, 1, -1, vexNDS3},
+	"VFNMSUB231SS":      {2, 0xBF, 0, 1, -1, vexNDS3},
+	"VGF2P8AFFINEINVQB": {3, 0xCF, 1, 1, -1, vexNDS3Imm},
+	"VGF2P8MULB":        {2, 0xCF, 0, 1, -1, vexNDS3},
+	"VMOVNTDQA":         {2, 0x2A, 0, 1, -1, vexRM},
+	"VMOVNTPD":          {1, 0x2B, 0, 1, -1, vexRMRev},
+	"VORPD":             {1, 0x56, 0, 1, -1, vexNDS3},
+	"VPADDSB":           {1, 0xEC, 0, 1, -1, vexNDS3},
+	"VPADDSW":           {1, 0xED, 0, 1, -1, vexNDS3},
+	"VPADDUSB":          {1, 0xDC, 0, 1, -1, vexNDS3},
+	"VPADDUSW":          {1, 0xDD, 0, 1, -1, vexNDS3},
+	"VPCMPEQQ":          {2, 0x29, 0, 1, -1, vexNDS3},
+	"VPCMPEQW":          {1, 0x75, 0, 1, -1, vexNDS3},
+	"VPCMPGTB":          {1, 0x64, 0, 1, -1, vexNDS3},
+	"VPCMPGTD":          {1, 0x66, 0, 1, -1, vexNDS3},
+	"VPCMPGTW":          {1, 0x65, 0, 1, -1, vexNDS3},
+	"VPERMPS":           {2, 0x16, 0, 1, -1, vexNDS3},
+	"VPEXTRB":           {3, 0x14, 0, 1, -1, vexExtract},
+	"VPEXTRD":           {3, 0x16, 0, 1, -1, vexExtract},
+	"VPEXTRQ":           {3, 0x16, 1, 1, -1, vexExtract},
+	"VPINSRD":           {3, 0x22, 0, 1, -1, vexNDS3Imm},
+	"VPINSRQ":           {3, 0x22, 1, 1, -1, vexNDS3Imm},
+	"VPMULHRSW":         {2, 0x0B, 0, 1, -1, vexNDS3},
+	"VPMULHW":           {1, 0xE5, 0, 1, -1, vexNDS3},
+	"VPMULUDQ":          {1, 0xF4, 0, 1, -1, vexNDS3},
+	"VPSADBW":           {1, 0xF6, 0, 1, -1, vexNDS3},
+	"VPSUBSB":           {1, 0xE8, 0, 1, -1, vexNDS3},
+	"VPSUBSW":           {1, 0xE9, 0, 1, -1, vexNDS3},
+	"VPSUBUSB":          {1, 0xD8, 0, 1, -1, vexNDS3},
+	"VPSUBUSW":          {1, 0xD9, 0, 1, -1, vexNDS3},
+	"VPUNPCKHBW":        {1, 0x68, 0, 1, -1, vexNDS3},
+	"VPUNPCKHQDQ":       {1, 0x6D, 0, 1, -1, vexNDS3},
+	"VPUNPCKHWD":        {1, 0x69, 0, 1, -1, vexNDS3},
+	"VPUNPCKLBW":        {1, 0x60, 0, 1, -1, vexNDS3},
+	"VPUNPCKLWD":        {1, 0x61, 0, 1, -1, vexNDS3},
+	"VSQRTPD":           {1, 0x51, 0, 1, -1, vexRM},
+	"VSQRTSD":           {1, 0x51, 0, 3, -1, vexNDS3},
+	"VSQRTSS":           {1, 0x51, 0, 2, -1, vexNDS3},
+	"VUCOMISD":          {1, 0x2E, 0, 1, -1, vexRM},
+
+	// VEX.0F.WIG, the plain-prefix single/double arithmetic and unpack
+	// spellings (no 66 prefix; WIG, so W = 0).
+	"VANDNPS":   {1, 0x55, 0, 0, -1, vexNDS3},
+	"VANDPS":    {1, 0x54, 0, 0, -1, vexNDS3},
+	"VORPS":     {1, 0x56, 0, 0, -1, vexNDS3},
+	"VUNPCKLPS": {1, 0x14, 0, 0, -1, vexNDS3},
+	"VUNPCKHPS": {1, 0x15, 0, 0, -1, vexNDS3},
+	"VSQRTPS":   {1, 0x51, 0, 0, -1, vexRM},
+	"VMOVNTPS":  {1, 0x2B, 0, 0, -1, vexRMRev},
+	// VEX.128.66.0F, the scalar and packed compare forms.
+	"VCOMISS":  {1, 0x2F, 0, 1, -1, vexRM},
+	"VUCOMISS": {1, 0x2E, 0, 0, -1, vexRM},
+	// VEX.128.0F.F3/F2.W0, the high/low word shuffles ($imm, src, dst).
+	"VPSHUFHW": {1, 0x70, 0, 2, -1, vexImmRM},
+	"VPSHUFLW": {1, 0x70, 0, 3, -1, vexImmRM},
 }
 
 // vexSrcLen maps a source-length conversion mnemonic (the X/Y spellings of
@@ -255,7 +468,7 @@ var vexSrcLen = map[string]int{
 	"VCVTPD2PSY":  1,
 }
 
-// vexVarShift maps the shift mnemonics to their variable-count opcode — the
+// vexVarShift maps the shift mnemonics to their variable-count opcode, the
 // form whose count comes from an XMM register or memory (VPSRLQ X0, Y8, Y8),
 // an ordinary NDS encoding rather than the /digit immediate form above.
 var vexVarShift = map[string]byte{
@@ -286,20 +499,22 @@ type vexMoveSpec struct {
 
 // vexMoveTable maps an upper-case move mnemonic to its encoding.
 var vexMoveTable = map[string]vexMoveSpec{
-	// VEX.128/256.F3.0F.WIG — unaligned integer move.
+	// VEX.128/256.F3.0F.WIG, unaligned integer move.
 	"VMOVDQU": {1, 2, 0x6F, 0x7F, 0, 0, 0, 0, true, false, false},
-	// VEX.128/256.66.0F.WIG — unaligned packed double move.
+	// VEX.128/256.66.0F.WIG, aligned integer move.
+	"VMOVDQA": {1, 1, 0x6F, 0x7F, 0, 0, 0, 0, true, false, false},
+	// VEX.128/256.66.0F.WIG, unaligned packed double move.
 	"VMOVUPD": {1, 1, 0x10, 0x11, 0, 0, 0, 0, true, false, false},
-	// VEX.128.66.0F.W0 — 32-bit GPR/memory ↔ XMM.
+	// VEX.128.66.0F.W0, 32-bit GPR/memory ↔ XMM.
 	"VMOVD": {1, 1, 0x6E, 0x7E, 0, 0, 0, 0, false, true, true},
-	// VMOVQ — 66 6E W1 (r/m→xmm), 66 7E W1 (xmm→r/m), 66 D6 W0 (xmm→xmm).
+	// VMOVQ, 66 6E W1 (r/m→xmm), 66 7E W1 (xmm→r/m), 66 D6 W0 (xmm→xmm).
 	"VMOVQ": {1, 1, 0x6E, 0x7E, 1, 1, 0xD6, 0, true, true, true},
-	// VEX.128.F2.0F.WIG — scalar double move, memory operands only (the
+	// VEX.128.F2.0F.WIG, scalar double move, memory operands only (the
 	// register form takes three operands and is not supported yet).
 	"VMOVSD": {1, 3, 0x10, 0x11, 0, 0, 0, 0, false, false, true},
-	// VEX.128.F3.0F.WIG — scalar single move, memory operands only.
+	// VEX.128.F3.0F.WIG, scalar single move, memory operands only.
 	"VMOVSS": {1, 2, 0x10, 0x11, 0, 0, 0, 0, false, false, true},
-	// VEX.128/256 — aligned packed moves.
+	// VEX.128/256, aligned packed moves.
 	"VMOVAPS": {1, 0, 0x28, 0x29, 0, 0, 0, 0, true, false, false},
 	"VMOVAPD": {1, 1, 0x28, 0x29, 0, 0, 0, 0, true, false, false},
 }
@@ -315,13 +530,21 @@ func isVex(mnemUpper string) bool {
 
 // encodeVex encodes a VEX instruction with operands in Plan 9 order.
 func (e *enc) encodeVex(mnemUpper string, ops []Operand) error {
-	// Vector register indices 16–31 exist only in EVEX encodings; fail
+	// Vector register indices 16-31 exist only in EVEX encodings; fail
 	// loudly rather than silently truncating the index.
 	for _, op := range ops {
 		if r, ok := op.(Reg); ok && r.isVec() && r.idx >= 16 {
 			return fmt.Errorf("%s: vector register index %d needs an EVEX (AVX-512) instruction", mnemUpper, r.idx)
 		}
 	}
+	// VBROADCASTI128 broadcasts a 128-bit lane into a 256-bit destination
+	// only; an XMM destination is rejected exactly as go tool asm does.
+	if mnemUpper == "VBROADCASTI128" {
+		dstReg, ok := ops[len(ops)-1].(Reg)
+		if len(ops) != 2 || !ok || dstReg.size != 32 {
+			return fmt.Errorf("VBROADCASTI128 requires a YMM destination")
+		}
+	}
 	if ms, ok := vexMoveTable[mnemUpper]; ok {
 		return e.encodeVexMove(mnemUpper, ms, ops)
 	}
@@ -354,6 +577,18 @@ func (e *enc) encodeVex(mnemUpper string, ops []Operand) error {
 		return e.encodeVexRMSrcLen(mnemUpper, spec, ops)
 	case vexZero:
 		return e.encodeVexZero(mnemUpper, spec, ops)
+	case vexZeroAll:
+		return e.encodeVexZeroAll(mnemUpper, spec, ops)
+	case vexNDS3GPR:
+		return e.encodeVexNDS3GPR(spec, ops)
+	case vexImmRMGPR:
+		return e.encodeVexImmRMGPR(spec, ops)
+	case vexRMOpGPR:
+		return e.encodeVexRMOpGPR(spec, ops)
+	case vexCountGPR:
+		return e.encodeVexCountGPR(spec, ops)
+	case vexRMRev:
+		return e.encodeVexRMRev(spec, ops)
 	}
 	return fmt.Errorf("unhandled VEX form for %s", mnemUpper)
 }
@@ -418,7 +653,7 @@ func (e *enc) encodeVexRM(spec vexSpec, ops []Operand) error {
 }
 
 // encodeVexRMSrcLen encodes a length-narrowing conversion: OP src, dst with
-// the destination always XMM and the VEX.L bit following the source — fixed
+// the destination always XMM and the VEX.L bit following the source, fixed
 // by the mnemonic's spelling (VCVTPD2DQX = 128, VCVTPD2DQY = 256) even when
 // the source is memory.
 func (e *enc) encodeVexRMSrcLen(mnem string, spec vexSpec, ops []Operand) error {
@@ -455,9 +690,10 @@ func (e *enc) encodeVexShiftImm(spec vexSpec, ops []Operand) error {
 	if !ok {
 		return fmt.Errorf("shift count must be an immediate")
 	}
-	srcReg, ok := src.(Reg)
-	if !ok || !srcReg.isVec() {
-		return fmt.Errorf("shift source must be a vector register")
+	// The count source is a vector register or memory; the VEX length
+	// follows the destination register either way.
+	if !vecOrMem(src) {
+		return fmt.Errorf("shift source must be a vector register or memory")
 	}
 	dstReg, ok := dst.(Reg)
 	if !ok || !dstReg.isVec() {
@@ -465,7 +701,7 @@ func (e *enc) encodeVexShiftImm(spec vexSpec, ops []Operand) error {
 	}
 
 	vvvvBar := 15 - (dstReg.idx & 15)
-	if err := e.emitVexFields(spec, dstReg.vecLenBit(), spec.opdigit, 0, vvvvBar, srcReg); err != nil {
+	if err := e.emitVexFields(spec, dstReg.vecLenBit(), spec.opdigit, 0, vvvvBar, src); err != nil {
 		return err
 	}
 	immByte, err := imm8(int64(immVal))
@@ -605,6 +841,129 @@ func (e *enc) encodeVexZero(mnem string, spec vexSpec, ops []Operand) error {
 	return nil
 }
 
+// encodeVexZeroAll encodes a no-operand instruction (VZEROALL), the L = 1
+// twin of VZEROUPPER.
+func (e *enc) encodeVexZeroAll(mnem string, spec vexSpec, ops []Operand) error {
+	if len(ops) != 0 {
+		return fmt.Errorf("%s expects no operands, got %d", mnem, len(ops))
+	}
+	// 2-byte VEX: R̄ = 1, v̄vvv = 1111 (unused), L = 1.
+	e.out = append(e.out, 0xC5, byte(1<<7|15<<3|1<<2|spec.pp), spec.opcode)
+	return nil
+}
+
+// encodeVexNDS3GPR encodes the three-operand NDS form over general-purpose
+// registers (ANDN, MULX): OP src2, src1, dst with reg = dst, vvvv = src1,
+// rm = src2 and L = 0.
+func (e *enc) encodeVexNDS3GPR(spec vexSpec, ops []Operand) error {
+	if len(ops) != 3 {
+		return fmt.Errorf("VEX NDS instruction expects 3 operands, got %d", len(ops))
+	}
+	src2, src1, dst := ops[0], ops[1], ops[2]
+	dstReg, ok := dst.(Reg)
+	if !ok || dstReg.isVec() {
+		return fmt.Errorf("VEX destination must be a general-purpose register")
+	}
+	vvvvReg, ok := src1.(Reg)
+	if !ok || vvvvReg.isVec() {
+		return fmt.Errorf("VEX vvvv operand must be a general-purpose register")
+	}
+	rBit := 0
+	if dstReg.idx >= 8 {
+		rBit = 1
+	}
+	return e.emitVexFields(spec, 0, dstReg.idx&7, rBit, 15-(vvvvReg.idx&15), src2)
+}
+
+// encodeVexImmRMGPR encodes the immediate form over general-purpose
+// registers (RORX): OP $imm, src, dst with reg = dst, rm = src, L = 0.
+func (e *enc) encodeVexImmRMGPR(spec vexSpec, ops []Operand) error {
+	if len(ops) != 3 {
+		return fmt.Errorf("instruction expects 3 operands ($imm, src, dst), got %d", len(ops))
+	}
+	imm, src, dst := ops[0], ops[1], ops[2]
+	immVal, ok := imm.(Imm)
+	if !ok {
+		return fmt.Errorf("shift control must be an immediate")
+	}
+	dstReg, ok := dst.(Reg)
+	if !ok || dstReg.isVec() {
+		return fmt.Errorf("VEX destination must be a general-purpose register")
+	}
+	immByte, err := imm8(int64(immVal))
+	if err != nil {
+		return err
+	}
+	if err := e.emitVexFields(spec, 0, dstReg.idx&7, 0, 15, src); err != nil {
+		return err
+	}
+	e.out = append(e.out, immByte)
+	return nil
+}
+
+// encodeVexRMOpGPR encodes the two-operand /digit form over general-purpose
+// registers (BLSI, BLSMSK, BLSR): OP src, dst with ModRM.reg = /digit,
+// ModRM.rm = src and VEX.vvvv = dst.
+func (e *enc) encodeVexRMOpGPR(spec vexSpec, ops []Operand) error {
+	if len(ops) != 2 {
+		return fmt.Errorf("instruction expects 2 operands (src, dst), got %d", len(ops))
+	}
+	src, dst := ops[0], ops[1]
+	dstReg, ok := dst.(Reg)
+	if !ok || dstReg.isVec() {
+		return fmt.Errorf("VEX destination must be a general-purpose register")
+	}
+	return e.emitVexFields(spec, 0, spec.opdigit, 0, 15-(dstReg.idx&15), src)
+}
+
+// encodeVexCountGPR encodes the three-operand count form over general-purpose
+// registers (SHLX, SHRX, SARX, BEXTR, BZHI): OP src, count, dst with
+// VEX.vvvv = src (op0), ModRM.rm = count (op1), ModRM.reg = dst (op2).
+func (e *enc) encodeVexCountGPR(spec vexSpec, ops []Operand) error {
+	if len(ops) != 3 {
+		return fmt.Errorf("VEX count instruction expects 3 operands, got %d", len(ops))
+	}
+	src, count, dst := ops[0], ops[1], ops[2]
+	dstReg, ok := dst.(Reg)
+	if !ok || dstReg.isVec() {
+		return fmt.Errorf("VEX destination must be a general-purpose register")
+	}
+	countReg, ok := count.(Reg)
+	if !ok || countReg.isVec() {
+		return fmt.Errorf("VEX count operand must be a general-purpose register")
+	}
+	srcReg, ok := src.(Reg)
+	if !ok || srcReg.isVec() {
+		return fmt.Errorf("VEX count source must be a general-purpose register")
+	}
+	rBit := 0
+	if dstReg.idx >= 8 {
+		rBit = 1
+	}
+	return e.emitVexFields(spec, 0, dstReg.idx&7, rBit, 15-(srcReg.idx&15), count)
+}
+
+// encodeVexRMRev encodes the reversed two-operand form: OP src, dst with the
+// vector source in ModRM.reg and the memory destination in r/m (VMOVNTDQ,
+// a store with no register-destination form).
+func (e *enc) encodeVexRMRev(spec vexSpec, ops []Operand) error {
+	if len(ops) != 2 {
+		return fmt.Errorf("store expects 2 operands, got %d", len(ops))
+	}
+	srcReg, ok := ops[0].(Reg)
+	if !ok || !srcReg.isVec() {
+		return fmt.Errorf("store source must be a vector register")
+	}
+	if !memOperand(ops[1]) {
+		return fmt.Errorf("store destination must be memory")
+	}
+	rBit := 0
+	if srcReg.idx >= 8 {
+		rBit = 1
+	}
+	return e.emitVexFields(spec, srcReg.vecLenBit(), srcReg.idx&7, rBit, 15, ops[1])
+}
+
 // encodeVexMove encodes a two-operand move (VMOVDQU, VMOVUPD, VMOVD, VMOVQ,
 // VMOVSD), picking the direction-specific opcode and VEX.W.  A vector→vector
 // move uses the store-form layout (reg = source, rm = destination), matching

asm/vex_test.go

@@ -19,6 +19,65 @@ func vreg(t *testing.T, name string) Reg {
 	return r
 }
 
+// x86asmUnrecognised lists the VEX mnemonics whose machine code the
+// golang.org/x/arch decoder cannot resolve; their bytes are verified against
+// go tool asm in the ground-truth tests instead.
+var x86asmUnrecognised = map[string]bool{
+	"ANDNL":        true,
+	"ANDNQ":        true,
+	"MULXL":        true,
+	"MULXQ":        true,
+	"RORXL":        true,
+	"RORXQ":        true,
+	"VFMADD213SD":  true,
+	"VFNMADD231SD": true,
+	// The scalar FMA spellings the decoder's tables lack entirely.
+	"VFMADD132SD":  true,
+	"VFMADD132SS":  true,
+	"VFMADD213SS":  true,
+	"VFMADD231SD":  true,
+	"VFMADD231SS":  true,
+	"VFMSUB132SD":  true,
+	"VFMSUB132SS":  true,
+	"VFMSUB213SD":  true,
+	"VFMSUB213SS":  true,
+	"VFMSUB231SD":  true,
+	"VFMSUB231SS":  true,
+	"VFNMADD132SD": true,
+	"VFNMADD132SS": true,
+	"VFNMADD213SD": true,
+	"VFNMADD213SS": true,
+	"VFNMADD231SS": true,
+	"VFNMSUB132SD": true,
+	"VFNMSUB132SS": true,
+	"VFNMSUB213SD": true,
+	"VFNMSUB213SS": true,
+	"VFNMSUB231SD": true,
+	"VFNMSUB231SS": true,
+	// The BMI1 unary bit ops the decoder's AVX tables lack.
+	"BLSIL":   true,
+	"BLSIQ":   true,
+	"BLSMSKL": true,
+	"BLSMSKQ": true,
+	"BLSRL":   true,
+	"BLSRQ":   true,
+	// The BMI2 bit ops whose W1/LZ rows the decoder misses.
+	"BEXTRL": true,
+	"BEXTRQ": true,
+	"BZHIL":  true,
+	"BZHIQ":  true,
+	"PDEPL":  true,
+	"PDEPQ":  true,
+	"PEXTL":  true,
+	"PEXTQ":  true,
+	"SARXL":  true,
+	"SARXQ":  true,
+	"SHLXL":  true,
+	"SHLXQ":  true,
+	"SHRXL":  true,
+	"SHRXQ":  true,
+}
+
 // TestVexNDS3 encodes `mnem Y0, Y1, Y2` for every three-operand NDS
 // instruction and verifies it round-trips through the x86 decoder to the same
 // mnemonic.  A wrong opcode/map/pp surfaces as a different decoded instruction.
@@ -37,8 +96,15 @@ func TestVexNDS3(t *testing.T) {
 			t.Errorf("%s: Encode: %v", mnem, err)
 			continue
 		}
+		// The x86 decoder's table lacks a handful of rows the Go assembler
+		// emits (the scalar 213/231 FMA spellings among them); those are
+		// pinned byte for byte against go tool asm in TestVexGroundTruth
+		// instead of round-tripped here.
 		inst, err := x86asm.Decode(code, 64)
 		if err != nil {
+			if strings.Contains(err.Error(), "unrecognized instruction") && x86asmUnrecognised[mnem] {
+				continue
+			}
 			t.Errorf("%s: Decode(% x): %v", mnem, err, code)
 			continue
 		}
@@ -173,7 +239,7 @@ func TestVexGroundTruth(t *testing.T) {
 		{"VPMULLD Y1,Y2,Y3", "VPMULLD", []Operand{vreg(t, "Y1"), vreg(t, "Y2"), vreg(t, "Y3")}, "c4e26d40d9", ""},
 		{"VPUNPCKLDQ Y4,Y3,Y5", "VPUNPCKLDQ", []Operand{vreg(t, "Y4"), vreg(t, "Y3"), vreg(t, "Y5")}, "c5e562ec", ""},
 		{"VPERMD Y1,Y2,Y3", "VPERMD", []Operand{vreg(t, "Y1"), vreg(t, "Y2"), vreg(t, "Y3")}, "c4e26d36d9", ""},
-		// Floating point (packed and scalar) and FMA — same NDS form, the pp
+		// Floating point (packed and scalar) and FMA; same NDS form, the pp
 		// bits and map select the operation.
 		{"VADDPD Y9,Y8,Y8", "VADDPD", []Operand{vreg(t, "Y9"), vreg(t, "Y8"), vreg(t, "Y8")}, "c4413d58c1", ""},
 		{"VADDPD X1,X2,X3", "VADDPD", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "c5e958d9", ""},
@@ -184,6 +250,50 @@ func TestVexGroundTruth(t *testing.T) {
 		{"VMULSD X0,X1,X1", "VMULSD", []Operand{vreg(t, "X0"), vreg(t, "X1"), vreg(t, "X1")}, "c5f359c8", ""},
 		{"VFMADD231PD Y14,Y12,Y8", "VFMADD231PD", []Operand{vreg(t, "Y14"), vreg(t, "Y12"), vreg(t, "Y8")}, "c4429db8c6", ""},
 		{"VFMADD231PD (DI),Y12,Y8", "VFMADD231PD", []Operand{Ptr(DI, 0, 32), vreg(t, "Y12"), vreg(t, "Y8")}, "c4629db807", ""},
+		{"VFMADD213SD X0,X1,X2", "VFMADD213SD", []Operand{vreg(t, "X0"), vreg(t, "X1"), vreg(t, "X2")}, "c4e2f1a9d0", ""},
+		{"VFNMADD231SD X0,X1,X2", "VFNMADD231SD", []Operand{vreg(t, "X0"), vreg(t, "X1"), vreg(t, "X2")}, "c4e2f1bdd0", ""},
+		// Packed single XOR and byte compare (NDS form).
+		{"VXORPS Y0,Y1,Y2", "VXORPS", []Operand{vreg(t, "Y0"), vreg(t, "Y1"), vreg(t, "Y2")}, "c5f457d0", ""},
+		{"VPCMPEQB Y0,Y1,Y2", "VPCMPEQB", []Operand{vreg(t, "Y0"), vreg(t, "Y1"), vreg(t, "Y2")}, "c5f574d0", ""},
+		// Octa byte shifts (vvvv carries the destination).
+		{"VPSLLDQ $2,X0,X1", "VPSLLDQ", []Operand{Imm(2), vreg(t, "X0"), vreg(t, "X1")}, "c5f173f802", ""},
+		{"VPSRLDQ $2,Y0,Y1", "VPSRLDQ", []Operand{Imm(2), vreg(t, "Y0"), vreg(t, "Y1")}, "c5f573d802", ""},
+		// Two-source shuffle, blend and carry-less multiply (NDS + imm8).
+		{"VPERM2F128 $3,Y0,Y1,Y2", "VPERM2F128", []Operand{Imm(3), vreg(t, "Y0"), vreg(t, "Y1"), vreg(t, "Y2")}, "c4e37506d003", ""},
+		{"VPBLENDD $3,X0,X1,X2", "VPBLENDD", []Operand{Imm(3), vreg(t, "X0"), vreg(t, "X1"), vreg(t, "X2")}, "c4e37102d003", ""},
+		{"VPBLENDD $3,Y0,Y1,Y2", "VPBLENDD", []Operand{Imm(3), vreg(t, "Y0"), vreg(t, "Y1"), vreg(t, "Y2")}, "c4e37502d003", ""},
+		{"VPCLMULQDQ $0,X0,X1,X2", "VPCLMULQDQ", []Operand{Imm(0), vreg(t, "X0"), vreg(t, "X1"), vreg(t, "X2")}, "c4e37144d000", ""},
+		{"VGF2P8AFFINEQB $0,X0,X1,X2", "VGF2P8AFFINEQB", []Operand{Imm(0), vreg(t, "X0"), vreg(t, "X1"), vreg(t, "X2")}, "c4e3f1ced000", ""},
+		// Two-operand test and the non-temporal and broadcast stores.
+		{"VPTEST X0,X1", "VPTEST", []Operand{vreg(t, "X0"), vreg(t, "X1")}, "c4e27917c8", ""},
+		{"VPTEST Y0,Y1", "VPTEST", []Operand{vreg(t, "Y0"), vreg(t, "Y1")}, "c4e27d17c8", ""},
+		{"VMOVNTDQ Y0,(AX)", "VMOVNTDQ", []Operand{vreg(t, "Y0"), Ptr(AX, 0, 32)}, "c5fde700", ""},
+		{"VMOVNTDQ X0,(AX)", "VMOVNTDQ", []Operand{vreg(t, "X0"), Ptr(AX, 0, 16)}, "c5f9e700", ""},
+		{"VBROADCASTI128 (AX),Y1", "VBROADCASTI128", []Operand{Ptr(AX, 0, 16), vreg(t, "Y1")}, "c4e27d5a08", ""},
+		// Aligned integer move and the full zeroing form.
+		{"VMOVDQA X0,X1", "VMOVDQA", []Operand{vreg(t, "X0"), vreg(t, "X1")}, "c5f97fc1", ""},
+		{"VMOVDQA (AX),X1", "VMOVDQA", []Operand{Ptr(AX, 0, 16), vreg(t, "X1")}, "c5f96f08", ""},
+		{"VMOVDQA Y0,Y1", "VMOVDQA", []Operand{vreg(t, "Y0"), vreg(t, "Y1")}, "c5fd7fc1", ""},
+		{"VZEROALL", "VZEROALL", []Operand{}, "c5fc77", ""},
+		// BMI1/BMI2 general-register VEX forms.
+		{"ANDNL AX,BX,CX", "ANDNL", []Operand{AX, BX, CX}, "c4e260f2c8", ""},
+		{"ANDNQ AX,BX,CX", "ANDNQ", []Operand{AX, BX, CX}, "c4e2e0f2c8", ""},
+		{"MULXL AX,BX,CX", "MULXL", []Operand{AX, BX, CX}, "c4e263f6c8", ""},
+		{"MULXQ AX,BX,CX", "MULXQ", []Operand{AX, BX, CX}, "c4e2e3f6c8", ""},
+		{"RORXL $3,AX,CX", "RORXL", []Operand{Imm(3), AX, CX}, "c4e37bf0c803", ""},
+		{"RORXQ $3,AX,CX", "RORXQ", []Operand{Imm(3), AX, CX}, "c4e3fbf0c803", ""},
+		// BMI2 variable shifts and bit ops (three general registers).
+		{"SHLXL AX,CX,R15", "SHLXL", []Operand{AX, CX, vreg(t, "R15")}, "c46279f7f9", ""},
+		{"SHRXQ R8,DX,AX", "SHRXQ", []Operand{vreg(t, "R8"), DX, AX}, "c4e2bbf7c2", ""},
+		{"SARXQ AX,DX,R9", "SARXQ", []Operand{AX, DX, vreg(t, "R9")}, "c462faf7ca", ""},
+		{"BEXTRL AX,CX,R15", "BEXTRL", []Operand{AX, CX, vreg(t, "R15")}, "c46278f7f9", ""},
+		{"BZHIQ AX,CX,R15", "BZHIQ", []Operand{AX, CX, vreg(t, "R15")}, "c462f8f5f9", ""},
+		{"PDEPQ AX,CX,R15", "PDEPQ", []Operand{AX, CX, vreg(t, "R15")}, "c462f3f5f8", ""},
+		{"PEXTQ AX,CX,R15", "PEXTQ", []Operand{AX, CX, vreg(t, "R15")}, "c462f2f5f8", ""},
+		// BMI1 unary bit ops (src, dst: /digit in ModRM.reg, dst in vvvv).
+		{"BLSIL AX,CX", "BLSIL", []Operand{AX, CX}, "c4e270f3d8", ""},
+		{"BLSRQ AX,CX", "BLSRQ", []Operand{AX, CX}, "c4e2f0f3c8", ""},
+		{"BLSMSKQ AX,CX", "BLSMSKQ", []Operand{AX, CX}, "c4e2f0f3d0", ""},
 		// Two-operand reg/rm form (v̄vvv must be 1111).
 		{"VPMOVSXDQ X0,Y4", "VPMOVSXDQ", []Operand{vreg(t, "X0"), vreg(t, "Y4")}, "c4e27d25e0", ""},
 		{"VPMOVSXWD (SI),Y0", "VPMOVSXWD", []Operand{Ptr(SI, 0, 8), vreg(t, "Y0")}, "c4e27d2306", ""},
@@ -217,7 +327,7 @@ func TestVexGroundTruth(t *testing.T) {
 		{"VEXTRACTI128 $1,Y8,X9", "VEXTRACTI128", []Operand{Imm(1), vreg(t, "Y8"), vreg(t, "X9")}, "c4437d39c101", ""},
 		{"VEXTRACTI128 $1,Y8,(DI)", "VEXTRACTI128", []Operand{Imm(1), vreg(t, "Y8"), Ptr(DI, 0, 16)}, "c4637d390701", ""},
 		{"VEXTRACTF128 $1,Y8,X9", "VEXTRACTF128", []Operand{Imm(1), vreg(t, "Y8"), vreg(t, "X9")}, "c4437d19c101", ""},
-		// Moves — each direction picks its own opcode and VEX.W.
+		// Moves; each direction picks its own opcode and VEX.W.
 		{"VMOVDQU (SI),Y1", "VMOVDQU", []Operand{Ptr(SI, 0, 32), vreg(t, "Y1")}, "c5fe6f0e", ""},
 		{"VMOVDQU Y3,(DI)", "VMOVDQU", []Operand{vreg(t, "Y3"), Ptr(DI, 0, 32)}, "c5fe7f1f", ""},
 		{"VMOVDQU X1,X2", "VMOVDQU", []Operand{vreg(t, "X1"), vreg(t, "X2")}, "c5fa7fca", ""},
@@ -234,7 +344,7 @@ func TestVexGroundTruth(t *testing.T) {
 		{"VMOVD AX,X0", "VMOVD", []Operand{AX, vreg(t, "X0")}, "c5f96ec0", ""},
 		{"VMOVSD (SI),X8", "VMOVSD", []Operand{Ptr(SI, 0, 8), vreg(t, "X8")}, "c57b1006", ""},
 		{"VMOVSD X8,(SI)", "VMOVSD", []Operand{vreg(t, "X8"), Ptr(SI, 0, 8)}, "c57b1106", ""},
-		// Packed double arithmetic and unpack — the NDS form, the opcode
+		// Packed double arithmetic and unpack; the NDS form, the opcode
 		// selects the operation.
 		{"VSUBPD Y1,Y2,Y3", "VSUBPD", []Operand{vreg(t, "Y1"), vreg(t, "Y2"), vreg(t, "Y3")}, "c5ed5cd9", ""},
 		{"VDIVPD X1,X2,X3", "VDIVPD", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "c5e95ed9", ""},
@@ -255,12 +365,12 @@ func TestVexGroundTruth(t *testing.T) {
 		{"VMINSS X6,X7,X8", "VMINSS", []Operand{vreg(t, "X6"), vreg(t, "X7"), vreg(t, "X8")}, "c5425dc6", ""},
 		{"VMAXSS X1,X2,X3", "VMAXSS", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "c5ea5fd9", ""},
 		{"VADDSD 8(AX),X1,X2", "VADDSD", []Operand{Ptr(AX, 8, 8), vreg(t, "X1"), vreg(t, "X2")}, "c5f3585008", ""},
-		// VMOVDDUP — duplicate the low double (reg=dst, rm=src, F2 pp).
+		// VMOVDDUP; duplicate the low double (reg=dst, rm=src, F2 pp).
 		{"VMOVDDUP X1,X2", "VMOVDDUP", []Operand{vreg(t, "X1"), vreg(t, "X2")}, "c5fb12d1", ""},
 		{"VMOVDDUP Y1,Y2", "VMOVDDUP", []Operand{vreg(t, "Y1"), vreg(t, "Y2")}, "c5ff12d1", ""},
 		{"VMOVDDUP 8(AX),X1", "VMOVDDUP", []Operand{Ptr(AX, 8, 8), vreg(t, "X1")}, "c5fb124808", ""},
 		// Conversions: DQ→PS (no prefix), PS→PD (Go emits it without the F3
-		// prefix — see the table comment), DQ→PD.
+		// prefix; see the table comment), DQ→PD.
 		{"VCVTDQ2PS X1,X2", "VCVTDQ2PS", []Operand{vreg(t, "X1"), vreg(t, "X2")}, "c5f85bd1", ""},
 		{"VCVTDQ2PS Y3,Y4", "VCVTDQ2PS", []Operand{vreg(t, "Y3"), vreg(t, "Y4")}, "c5fc5be3", ""},
 		{"VCVTPS2PD X1,X2", "VCVTPS2PD", []Operand{vreg(t, "X1"), vreg(t, "X2")}, "c5f85ad1", ""},
@@ -287,6 +397,12 @@ func TestVexGroundTruth(t *testing.T) {
 		}
 		inst, err := x86asm.Decode(code, 64)
 		if err != nil {
+			// The decoder's AVX/BMI table lacks a few rows the Go
+			// assembler emits (the GPR VEX forms and the scalar FMA
+			// spellings); their bytes are the ground truth here.
+			if x86asmUnrecognised[c.mnem] {
+				continue
+			}
 			t.Errorf("%s: Decode(% x): %v", c.name, code, err)
 			continue
 		}

ast/ast.go

@@ -8,7 +8,7 @@
 // to the arch package; the AST records syntax only.
 package ast
 
-import "sourcedock.dev/petrbalvin/gasm-devkit/token"
+import "sourcedock.dev/petrbalvin/gasm-sdk/token"
 
 // File is the parsed representation of one .s source file.
 type File struct {
@@ -17,7 +17,7 @@ type File struct {
 	Orphans []Stmt // labels/instructions seen before any TEXT directive
 	// Macros holds the names introduced by #define directives in this file.
 	// The linter uses it to avoid flagging macro invocations as unknown
-	// instructions (macro expansion itself is out of scope — see the docs).
+	// instructions (macro expansion itself is out of scope, see the docs).
 	Macros map[string]bool
 }
 
@@ -114,6 +114,7 @@ type Symbol struct {
 	Pkg    string // package prefix before the middle dot ("" = current package)
 	Name   string // identifier without the middle dot or <>
 	Static bool   // the <> marker is present
+	ABI    string // the <NAME> ABI marker, e.g. ABIInternal ("" when absent)
 	Pseudo string // FP, SP, SB or PC ("" for a bare name)
 	Offset int64
 	HasOff bool
@@ -123,10 +124,8 @@ type Symbol struct {
 // OpKind classifies an operand syntactically.
 type OpKind int
 
-// Operand kinds.
 const (
-	OpInvalid   OpKind = iota
-	OpImmediate        // $value
+	OpImmediate = iota // $value
 	OpAddr             // register, memory reference, symbol or label
 )
 
@@ -152,11 +151,21 @@ type Immediate struct {
 // Address is a non-immediate operand: a register, a memory reference, a symbol
 // reference or a label.  Fields are populated best-effort from the syntax.
 type Address struct {
-	Sym    *Symbol // name reference (bare ident, or name+off(pseudo))
-	Base   string  // base register, from (base)
-	Index  string  // index register, from (index*scale)
-	Scale  int     // index scale; 0 when absent
-	Offset int64   // leading displacement, from off(base)
-	HasOff bool    // a leading displacement is present
-	Shift  string  // verbatim arm64 shift suffix, e.g. "<<2"
+	Sym    *Symbol   // name reference (bare ident, or name+off(pseudo))
+	Base   string    // base register, from (base)
+	Index  string    // index register, from (index*scale)
+	Scale  int       // index scale; 0 when absent
+	Offset int64     // leading displacement, from off(base)
+	HasOff bool      // a leading displacement is present
+	Shift  string    // verbatim arm64 shift suffix, e.g. "<< 2"
+	Range  *RegRange // bracketed register range; nil for every other form
+}
+
+// RegRange is a bracketed register range, [Z0-Z3]: the amd64 spelling of
+// the four-register source of the 4FMAPS/4VNNIW families.  Lo and Hi carry
+// the verbatim register spellings; the range is inclusive at both ends.
+type RegRange struct {
+	Lo  string
+	Hi  string
+	Pos token.Position
 }

ast/ast_test.go

@@ -6,7 +6,7 @@ package ast
 import (
 	"testing"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/token"
+	"sourcedock.dev/petrbalvin/gasm-sdk/token"
 )
 
 func pos(line, col int) token.Position { return token.Position{Line: line, Column: col} }
@@ -54,11 +54,11 @@ func TestStmtPositions(t *testing.T) {
 // TestInterfaces confirms the node types satisfy their interfaces, so callers
 // can range over Decls and Stmts.
 func TestInterfaces(t *testing.T) {
-	var decls []Decl = []Decl{&Include{}, &Preproc{}, &Text{}, &Globl{}, &Data{}}
+	var decls = []Decl{&Include{}, &Preproc{}, &Text{}, &Globl{}, &Data{}}
 	if len(decls) != 5 {
 		t.Fatal("decl interface set")
 	}
-	var stmts []Stmt = []Stmt{&Label{}, &Instr{}}
+	var stmts = []Stmt{&Label{}, &Instr{}}
 	if len(stmts) != 2 {
 		t.Fatal("stmt interface set")
 	}

cmd/gasm/asmhdr.go

@@ -0,0 +1,367 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package main
+
+import (
+	"errors"
+	"fmt"
+	"go/ast"
+	"go/build"
+	"go/constant"
+	"go/parser"
+	"go/token"
+	"go/types"
+	"os"
+	"path/filepath"
+	"regexp"
+	"strings"
+
+	"sourcedock.dev/petrbalvin/gasm-sdk/arch"
+)
+
+// go_asm.h is the header the Go compiler writes for every package that
+// carries assembly (the compiler's -asmhdr output): "#define const_NAME
+// value" for each package constant, and for each named struct type
+// "#define TYPE__size size" plus one "#define TYPE_field offset" per field.
+// GOROOT assembly includes it, and a standalone assembler has no compiler
+// to have produced it, so gasm generates the equivalent itself: the package
+// the .s file lives in is parsed and type-checked here, with the target
+// architecture's own sizes, and the same defines are written out.  The
+// type-checking GOOS is selected by the caller: a GOOS-specific file
+// (sys_darwin_arm64.s) needs its platform's defines, which a header from
+// the ambient GOOS silently omits.
+//
+// The emitter mirrors cmd/compile's dumpasmhdr exactly: constants come out
+// as "const_NAME", struct entries as "NAME__size" followed by the fields in
+// declaration order, blank names are skipped, and float and complex
+// constants are omitted (the assembler carries integers, bools and strings
+// only).  Aliases to structs are emitted, generic types are not: they have
+// no fixed size.  A define the assembly references but this header does not
+// carry surfaces later as the assembler's own "undefined" diagnostic naming
+// the define, which is the honest failure.
+
+// goAsmInclude matches the #include "go_asm.h" directive, tolerant of
+// whitespace, so the wiring knows which files need a generated header
+// before the preprocessor runs and would report the header as missing.
+var goAsmInclude = regexp.MustCompile(`(?m)^\s*#\s*include\s+"go_asm\.h"`)
+
+// needsGoAsmHeader reports whether src includes go_asm.h.
+func needsGoAsmHeader(src string) bool {
+	return goAsmInclude.MatchString(src)
+}
+
+// goAsmHeaderResolved reports whether the include of go_asm.h from a file in
+// asmDir already resolves: to a header in the package directory itself, or
+// in one of the -I directories, the way the preprocessor searches.  Only an
+// unresolved include is generated for; a header someone placed by hand is
+// the tool the author chose, and it also wins the preprocessor's own search
+// order, so generating a second copy would be dead weight at best.
+func goAsmHeaderResolved(asmDir string, dirs []string) bool {
+	candidates := []string{filepath.Join(asmDir, "go_asm.h")}
+	for _, d := range dirs {
+		candidates = append(candidates, filepath.Join(d, "go_asm.h"))
+	}
+	for _, candidate := range candidates {
+		if st, err := os.Stat(candidate); err == nil && !st.IsDir() {
+			return true
+		}
+	}
+	return false
+}
+
+// generateGoAsmHeader type-checks the Go package in pkgDir for goos and
+// goarch, writes its go_asm.h equivalent into dir, and returns dir.  An
+// empty goos means the ambient one.  The caller owns the directory and its
+// removal.
+func generateGoAsmHeader(pkgDir, goos, goarch, dir string) (string, error) {
+	if goos == "" {
+		goos = build.Default.GOOS
+	}
+	imp := newSourceImporter(goos, goarch)
+	if imp.sizes == nil {
+		return "", fmt.Errorf("go_asm.h: unknown GOARCH %q", goarch)
+	}
+	bp, err := imp.ctxt.ImportDir(pkgDir, 0)
+	if err != nil {
+		return "", fmt.Errorf("go_asm.h for GOARCH %s in %s: %w", goarch, pkgDir, err)
+	}
+	files, errs := imp.parse(bp)
+	if len(errs) > 0 {
+		return "", fmt.Errorf("go_asm.h for GOARCH %s in %s: %s", goarch, pkgDir, errorList(errs))
+	}
+	_, info, errs := imp.checkPackage(bp, files)
+	if len(errs) > 0 {
+		return "", fmt.Errorf("go_asm.h for GOARCH %s in %s: package does not type-check: %s", goarch, pkgDir, errorList(errs))
+	}
+
+	var b strings.Builder
+	fmt.Fprintf(&b, "// generated by gasm from package %s (GOOS %s, GOARCH %s)\n\n", bp.Name, goos, goarch)
+	// Files in the build's own order and declarations in source order: the
+	// same walk the compiler's reader makes, so the header reads the same
+	// way the toolchain's does.  Order carries no meaning to the assembler
+	// (defines form a table), only to a human diffing against one.
+	for _, f := range files {
+		for _, decl := range f.Decls {
+			gd, ok := decl.(*ast.GenDecl)
+			if !ok {
+				continue
+			}
+			for _, spec := range gd.Specs {
+				switch gd.Tok {
+				case token.CONST:
+					vs, ok := spec.(*ast.ValueSpec)
+					if !ok {
+						continue
+					}
+					for _, name := range vs.Names {
+						emitConst(&b, info.Defs[name], name.Name)
+					}
+				case token.TYPE:
+					ts, ok := spec.(*ast.TypeSpec)
+					if !ok {
+						continue
+					}
+					emitStruct(&b, imp.sizes, info.Defs[ts.Name], ts.Name.Name)
+				}
+			}
+		}
+	}
+
+	if err := os.MkdirAll(dir, 0o755); err != nil {
+		return "", fmt.Errorf("go_asm.h for GOARCH %s in %s: %w", goarch, pkgDir, err)
+	}
+	out := filepath.Join(dir, "go_asm.h")
+	if err := os.WriteFile(out, []byte(b.String()), 0o644); err != nil {
+		return "", fmt.Errorf("go_asm.h for GOARCH %s in %s: %w", goarch, pkgDir, err)
+	}
+	return dir, nil
+}
+
+// emitConst writes one const define, skipping what the toolchain skips:
+// blank names, and float and complex values the assembler has no syntax for.
+func emitConst(b *strings.Builder, obj types.Object, name string) {
+	c, ok := obj.(*types.Const)
+	if !ok || name == "_" {
+		return
+	}
+	switch c.Val().Kind() {
+	case constant.Float, constant.Complex, constant.Unknown:
+		return
+	}
+	fmt.Fprintf(b, "#define const_%s %s\n", name, c.Val().ExactString())
+}
+
+// emitStruct writes one named struct type's size and field offsets,
+// skipping what the toolchain skips: blank names, non-struct types, and
+// generic types, whose size depends on their instantiation.
+func emitStruct(b *strings.Builder, sizes types.Sizes, obj types.Object, name string) {
+	tn, ok := obj.(*types.TypeName)
+	if !ok || name == "_" {
+		return
+	}
+	t := types.Unalias(tn.Type())
+	// Generic types are spelled *types.Named with a type-parameter list;
+	// a plain struct type or an instantiated one carries none.
+	if named, ok := t.(*types.Named); ok && named.TypeParams().Len() > 0 {
+		return
+	}
+	st, ok := t.Underlying().(*types.Struct)
+	if !ok {
+		return
+	}
+	fmt.Fprintf(b, "#define %s__size %d\n", name, sizes.Sizeof(t))
+	fields := make([]*types.Var, st.NumFields())
+	for i := range st.NumFields() {
+		fields[i] = st.Field(i)
+	}
+	for i, off := range sizes.Offsetsof(fields) {
+		fld := fields[i]
+		if fld.Name() == "_" {
+			continue
+		}
+		fmt.Fprintf(b, "#define %s_%s %d\n", name, fld.Name(), off)
+	}
+}
+
+// errorList renders at most three errors, enough to say what is wrong
+// without burying the diagnostic the caller actually reads.
+func errorList(errs []error) string {
+	if len(errs) > 3 {
+		errs = errs[:3]
+	}
+	msgs := make([]string, len(errs))
+	for i, err := range errs {
+		msgs[i] = err.Error()
+	}
+	return strings.Join(msgs, "; ")
+}
+
+// sourceImporter type-checks imported packages from source with the target
+// architecture's sizes.  go/importer's "source" importer pins the host
+// GOARCH, which would lay out imported types (internal/cpu, internal/abi)
+// for the wrong target on a cross-architecture header, so the recursion is
+// carried here with one build context and one sizes instance per
+// architecture.
+type sourceImporter struct {
+	fset  *token.FileSet
+	ctxt  *build.Context
+	sizes types.Sizes
+	pkgs  map[string]*types.Package
+}
+
+// newSourceImporter returns the importer for one target GOOS and GOARCH.
+// Cgo is disabled so the file set is deterministic and independent of the
+// host's C toolchain: cgo-tagged files drop out of the build exactly as
+// they do from a CGO_ENABLED=0 build, whose assembly is what gasm targets.
+func newSourceImporter(goos, goarch string) *sourceImporter {
+	ctxt := new(build.Context)
+	*ctxt = build.Default
+	ctxt.GOOS = goos
+	ctxt.GOARCH = goarch
+	ctxt.CgoEnabled = false
+	return &sourceImporter{
+		fset:  token.NewFileSet(),
+		ctxt:  ctxt,
+		sizes: types.SizesFor("gc", goarch),
+		pkgs:  map[string]*types.Package{},
+	}
+}
+
+// Import type-checks one imported package and memoises it.  "unsafe" must
+// resolve to go/types' own package, never to the source in GOROOT/src/unsafe:
+// the source declares Sizeof and Offsetof as ordinary functions over
+// ArbitraryType, and checking against that signature rejects half the
+// unsafe arithmetic the gc compiler accepts, which is exactly the divergence
+// srcimporter guards against the same way.
+func (im *sourceImporter) Import(path string) (*types.Package, error) {
+	if path == "unsafe" {
+		return types.Unsafe, nil
+	}
+	if p, ok := im.pkgs[path]; ok {
+		return p, nil
+	}
+	bp, err := im.ctxt.Import(path, "", 0)
+	if err != nil {
+		return nil, err
+	}
+	files, errs := im.parse(bp)
+	if len(errs) > 0 {
+		return nil, errors.New(errorList(errs))
+	}
+	pkg, _, _ := im.checkPackage(bp, files)
+	im.pkgs[path] = pkg
+	return pkg, nil
+}
+
+// parse reads the build package's Go files.  Import-level failures (no Go
+// files for the target, unreadable files) come back as errors, and the
+// type-check decides the rest.
+func (im *sourceImporter) parse(bp *build.Package) ([]*ast.File, []error) {
+	if len(bp.GoFiles) == 0 {
+		return nil, []error{fmt.Errorf("no Go source files for GOOS=%s GOARCH=%s", im.ctxt.GOOS, im.ctxt.GOARCH)}
+	}
+	var (
+		files []*ast.File
+		errs  []error
+	)
+	for _, name := range bp.GoFiles {
+		f, err := parser.ParseFile(im.fset, filepath.Join(bp.Dir, name), nil, parser.SkipObjectResolution)
+		if err != nil {
+			errs = append(errs, err)
+			continue
+		}
+		files = append(files, f)
+	}
+	return files, errs
+}
+
+// checkPackage type-checks one package's files with the importer's sizes,
+// recording every error: a header from a package that does not type-check
+// could silently mis-state an offset, so the caller refuses the header
+// rather than trusting it.  The returned Defs map backs the root package's
+// emission walk; imports only need the checked package itself.
+func (im *sourceImporter) checkPackage(bp *build.Package, files []*ast.File) (*types.Package, *types.Info, []error) {
+	var errs []error
+	conf := &types.Config{
+		Importer: im,
+		Sizes:    im.sizes,
+		Error:    func(err error) { errs = append(errs, err) },
+	}
+	info := &types.Info{Defs: map[*ast.Ident]types.Object{}}
+	pkg, _ := conf.Check(bp.ImportPath, im.fset, files, info)
+	return pkg, info, errs
+}
+
+// asmhdrCache generates one go_asm.h per package directory and target
+// architecture under one temp root, for callers that assemble many files
+// (the corpus audit).  Failures are cached too: a package that does not
+// type-check must not be re-checked once per file.
+type asmhdrCache struct {
+	root string
+	dirs map[string]string // "pkgDir\x00goos\x00goarch" -> directory holding go_asm.h
+	errs map[string]error
+}
+
+func newAsmhdrCache() (*asmhdrCache, error) {
+	root, err := os.MkdirTemp("", "gasm-asmhdr")
+	if err != nil {
+		return nil, err
+	}
+	return &asmhdrCache{root: root, dirs: map[string]string{}, errs: map[string]error{}}, nil
+}
+
+// dirFor returns the directory holding the generated go_asm.h for pkgDir
+// under goos and goarch, generating it on first use.  An empty goos means
+// the ambient one, resolved here so that one package cannot generate twice
+// under an explicit and an implicit spelling of the same GOOS.
+func (c *asmhdrCache) dirFor(pkgDir, goos, goarch string) (string, error) {
+	if goos == "" {
+		goos = build.Default.GOOS
+	}
+	key := pkgDir + "\x00" + goos + "\x00" + goarch
+	if dir, ok := c.dirs[key]; ok {
+		return dir, nil
+	}
+	if err, ok := c.errs[key]; ok {
+		return "", err
+	}
+	dir := filepath.Join(c.root, fmt.Sprintf("h%d_%s_%s", len(c.dirs), goos, goarch))
+	if _, err := generateGoAsmHeader(pkgDir, goos, goarch, dir); err != nil {
+		c.errs[key] = err
+		return "", err
+	}
+	c.dirs[key] = dir
+	return dir, nil
+}
+
+// close removes the temp root.
+func (c *asmhdrCache) close() { os.RemoveAll(c.root) }
+
+// ensureGoAsmHeader prepares the include directory a file that includes
+// go_asm.h needs: the generated header for the package in path's directory,
+// for the file's target GOOS and architecture.  It reports a usage error
+// when the architecture cannot be determined, and passes through the
+// generator's diagnostics, which name the package.
+func ensureGoAsmHeader(path string, target arch.Arch, goos string, cache *asmhdrCache) (string, func(), error) {
+	if path == "-" {
+		return "", nil, errors.New("cannot generate go_asm.h for standard input (no package directory)")
+	}
+	if target == arch.Unknown {
+		return "", nil, errors.New("a file that includes go_asm.h needs a target architecture: name the file _<arch>.s or pass -GOARCH")
+	}
+	if cache != nil {
+		dir, err := cache.dirFor(filepath.Dir(path), goos, goarchName(target))
+		return dir, func() {}, err
+	}
+	root, err := os.MkdirTemp("", "gasm-asmhdr")
+	if err != nil {
+		return "", nil, err
+	}
+	dir, err := generateGoAsmHeader(filepath.Dir(path), goos, goarchName(target), root)
+	if err != nil {
+		os.RemoveAll(root)
+		return "", nil, err
+	}
+	return dir, func() { os.RemoveAll(root) }, nil
+}

cmd/gasm/asmhdr_test.go

@@ -0,0 +1,476 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package main
+
+import (
+	"os"
+	"os/exec"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+// writePkg lays out a minimal Go package in a temp directory.
+func writePkg(t *testing.T, files map[string]string) string {
+	t.Helper()
+	dir := t.TempDir()
+	for name, src := range files {
+		if err := os.WriteFile(filepath.Join(dir, name), []byte(src), 0o644); err != nil {
+			t.Fatal(err)
+		}
+	}
+	return dir
+}
+
+// generateFor generates the header for dir and returns its text.  An empty
+// goos means the ambient one.
+func generateFor(t *testing.T, dir, goos, goarch string) string {
+	t.Helper()
+	hdrDir, err := generateGoAsmHeader(dir, goos, goarch, t.TempDir())
+	if err != nil {
+		t.Fatalf("generateGoAsmHeader(%q, %s, %s): %v", dir, goos, goarch, err)
+	}
+	b, err := os.ReadFile(filepath.Join(hdrDir, "go_asm.h"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	return string(b)
+}
+
+func TestGenerateGoAsmHeaderShape(t *testing.T) {
+	dir := writePkg(t, map[string]string{"sample.go": `package sample
+
+const bufSize = 1024
+
+const (
+	a = iota * 8
+	b
+	c
+)
+
+const (
+	strConst = "hello"
+	boolConst = true
+	floatConst = 1.5
+	_ = "the blank identifier is skipped"
+)
+
+const shift = 1 << 20
+
+type reader struct {
+	r    int64
+	w    int64
+	_    [4]byte
+	name string
+}
+
+type scalar int
+
+type aliased struct {
+	k uint32
+	v uint32
+}
+
+type alias = aliased
+`})
+	hdr := generateFor(t, dir, "", "amd64")
+	want := []string{
+		"#define const_bufSize 1024",
+		// iota resolves through go/types, one define per name.
+		"#define const_a 0",
+		"#define const_b 8",
+		"#define const_c 16",
+		`#define const_strConst "hello"`,
+		"#define const_boolConst true",
+		// Floats are the toolchain's own skip, as are blank names.
+		"#define const_shift 1048576",
+		// The blank field still occupies its bytes: the pad after w runs to
+		// the string's 8-byte alignment.
+		"#define reader__size 40",
+		"#define reader_r 0",
+		"#define reader_w 8",
+		"#define reader_name 24",
+		// Non-struct named types carry no defines; aliases to structs do.
+		"#define aliased__size 8",
+		"#define aliased_k 0",
+		"#define aliased_v 4",
+		"#define alias__size 8",
+		"#define alias_k 0",
+		"#define alias_v 4",
+	}
+	for _, w := range want {
+		if !strings.Contains(hdr, w+"\n") {
+			t.Errorf("header misses %q\ngot:\n%s", w, hdr)
+		}
+	}
+	for _, banned := range []string{"#define const_floatConst", "#define _ ", "#define scalar"} {
+		if strings.Contains(hdr, banned) {
+			t.Errorf("header must not carry %s\ngot:\n%s", banned, hdr)
+		}
+	}
+}
+
+func TestGenerateGoAsmHeaderPerArch(t *testing.T) {
+	dir := writePkg(t, map[string]string{
+		"common.go": `package perarch
+
+type layout struct {
+	a int32
+	p uintptr
+}
+`,
+		// The build-tagged file set is part of the contract: a per-arch
+		// package is exactly how internal/cpu declares its layouts.
+		"const_amd64.go": `//go:build amd64
+
+package perarch
+
+const flavour = 1
+`,
+		"const_arm64.go": `//go:build arm64
+
+package perarch
+
+const flavour = 2
+`,
+	})
+	amd64 := generateFor(t, dir, "", "amd64")
+	arm64 := generateFor(t, dir, "", "arm64")
+	if !strings.Contains(amd64, "#define const_flavour 1\n") {
+		t.Errorf("amd64 header misses const_flavour 1:\n%s", amd64)
+	}
+	if !strings.Contains(arm64, "#define const_flavour 2\n") {
+		t.Errorf("arm64 header misses const_flavour 2:\n%s", arm64)
+	}
+	if strings.Contains(arm64, "#define const_flavour 1\n") {
+		t.Errorf("arm64 header must not carry the amd64 file's value")
+	}
+	// SizesFor makes the layout the target's: uintptr is 4 bytes wide on
+	// 386 and 8 on amd64, which must move p and grow the struct.
+	if !strings.Contains(amd64, "#define layout__size 16\n") || !strings.Contains(amd64, "#define layout_p 8\n") {
+		t.Errorf("amd64 layout wrong:\n%s", amd64)
+	}
+	w386 := generateFor(t, dir, "", "386")
+	if !strings.Contains(w386, "#define layout__size 8\n") || !strings.Contains(w386, "#define layout_p 4\n") {
+		t.Errorf("386 layout wrong:\n%s", w386)
+	}
+}
+
+// TestGenerateGoAsmHeaderGOOS pins the GOOS half of the target: only the
+// platform's own files type-check into the header, which is why
+// sys_darwin_arm64.s cannot assemble against a linux-generated one.
+func TestGenerateGoAsmHeaderGOOS(t *testing.T) {
+	dir := writePkg(t, map[string]string{
+		"common.go": `package goosaware
+
+type shared struct {
+	a int32
+}
+`,
+		"plat_darwin.go": `//go:build darwin
+
+package goosaware
+
+type platform struct {
+	trampoline_numer int64
+}
+`,
+		"plat_windows.go": `//go:build windows
+
+package goosaware
+
+type platform struct {
+	callbackArgs__size int32
+}
+`,
+	})
+	darwin := generateFor(t, dir, "darwin", "arm64")
+	if !strings.Contains(darwin, "#define platform__size 8\n") || !strings.Contains(darwin, "#define platform_trampoline_numer 0\n") {
+		t.Errorf("darwin header misses the darwin layout:\n%s", darwin)
+	}
+	if strings.Contains(darwin, "callbackArgs") {
+		t.Errorf("darwin header must not carry the windows layout:\n%s", darwin)
+	}
+	windows := generateFor(t, dir, "windows", "arm64")
+	if !strings.Contains(windows, "#define platform_callbackArgs__size 0\n") {
+		t.Errorf("windows header misses the windows layout:\n%s", windows)
+	}
+	if strings.Contains(windows, "trampoline_numer") {
+		t.Errorf("windows header must not carry the darwin layout:\n%s", windows)
+	}
+	// The ambient GOOS is neither of the two, so only shared's defines are
+	// emitted; the shared type keeps its layout there.
+	ambient := generateFor(t, dir, "", "arm64")
+	if !strings.Contains(ambient, "#define shared__size 4\n") {
+		t.Errorf("ambient header misses the shared layout:\n%s", ambient)
+	}
+	if strings.Contains(ambient, "#define platform_") {
+		t.Errorf("ambient header must not carry either platform layout:\n%s", ambient)
+	}
+}
+
+func TestGoosFromFilename(t *testing.T) {
+	for path, want := range map[string]string{
+		"/x/sys_darwin_arm64.s":     "darwin",
+		"/x/sys_windows_arm64.s":    "windows",
+		"/x/asm_linux_amd64.s":      "linux",
+		"/x/rt0_darwin_arm64.s":     "darwin",
+		"/x/vgetrandom_zos_s390x.s": "zos",
+		"/x/rt0_js_wasm.s":          "js",
+		"/x/memmove_amd64.s":        "",
+		"/x/vlop_arm.s":             "",
+		"/x/stubs.s":                "",
+	} {
+		if got := goosFromFilename(path); got != want {
+			t.Errorf("goosFromFilename(%q) = %q, want %q", path, got, want)
+		}
+	}
+}
+
+func TestGenerateGoAsmHeaderErrors(t *testing.T) {
+	t.Run("type error", func(t *testing.T) {
+		dir := writePkg(t, map[string]string{"bad.go": `package bad
+
+const x = undefinedIdent
+`})
+		_, err := generateGoAsmHeader(dir, "", "amd64", t.TempDir())
+		if err == nil {
+			t.Fatal("generation must fail for a package that does not type-check")
+		}
+		if !strings.Contains(err.Error(), dir) {
+			t.Errorf("error must name the package directory: %v", err)
+		}
+		if !strings.Contains(err.Error(), "type-check") {
+			t.Errorf("error must say the package does not type-check: %v", err)
+		}
+	})
+	t.Run("no go files", func(t *testing.T) {
+		dir := t.TempDir()
+		_, err := generateGoAsmHeader(dir, "", "amd64", t.TempDir())
+		if err == nil {
+			t.Fatal("generation must fail without Go files")
+		}
+		if !strings.Contains(err.Error(), dir) {
+			t.Errorf("error must name the package directory: %v", err)
+		}
+	})
+}
+
+func TestNeedsGoAsmHeader(t *testing.T) {
+	yes := "#include \"go_asm.h\"\n#include \"textflag.h\"\n"
+	no := "#include \"textflag.h\"\n#include \"funcdata.h\"\n"
+	if !needsGoAsmHeader(yes) {
+		t.Error("needsGoAsmHeader(missing on a go_asm.h include)")
+	}
+	if needsGoAsmHeader(no) {
+		t.Error("needsGoAsmHeader claims other headers need generation")
+	}
+}
+
+func TestGoAsmHeaderResolved(t *testing.T) {
+	dir := t.TempDir()
+	if goAsmHeaderResolved(dir, nil) {
+		t.Error("resolved with no header anywhere")
+	}
+	other := t.TempDir()
+	if goAsmHeaderResolved(dir, []string{other}) {
+		t.Error("resolved with an empty -I directory")
+	}
+	if err := os.WriteFile(filepath.Join(dir, "go_asm.h"), nil, 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if !goAsmHeaderResolved(dir, nil) {
+		t.Error("not resolved with the header in the package directory")
+	}
+}
+
+func TestOtherGOOSFile(t *testing.T) {
+	for path, want := range map[string]bool{
+		"/x/sys_windows_amd64.s": true,
+		"/x/rt0_js_wasm.s":       true,
+		"/x/sys_darwin_arm64.s":  true,
+		"/x/sys_linux_amd64.s":   false,
+		"/x/time_linux_amd64.s":  false,
+		"/x/memmove_amd64.s":     false,
+		"/x/generic.s":           false,
+	} {
+		if got := otherGOOSFile(path); got != want {
+			t.Errorf("otherGOOSFile(%q) = %v, want %v", path, got, want)
+		}
+	}
+}
+
+// TestRunCorpusAuditGoAsm covers the audit wiring end to end: a package
+// beside its kernel, the kernel living off the generated defines, and the
+// histogram recording a generation failure as its own reason.
+func TestRunCorpusAuditGoAsm(t *testing.T) {
+	dir := t.TempDir()
+	write := func(name, src string) {
+		t.Helper()
+		if err := os.WriteFile(filepath.Join(dir, name), []byte(src), 0o644); err != nil {
+			t.Fatal(err)
+		}
+	}
+	write("pkg.go", `package corpus
+
+const pageSize = 4096
+
+type header struct {
+	magic uint64
+	flags uint64
+}
+`)
+	write("kern_amd64.s", "#include \"go_asm.h\"\nTEXT \xc2\xb7f(SB), NOSPLIT, $0-16\n\tMOVQ\t$const_pageSize, AX\n\tMOVQ\t$header__size, BX\n\tRET\n")
+	// The defines live in the file's own package; a kernel in a directory
+	// without Go files has no package to generate from.
+	if err := os.MkdirAll(filepath.Join(dir, "sub"), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	write(filepath.Join("sub", "lonely_arm64.s"), "#include \"go_asm.h\"\nTEXT \xc2\xb7g(SB), NOSPLIT, $0-0\n\tRET\n")
+
+	stats, err := runCorpusAudit(dir, nil)
+	if err != nil {
+		t.Fatalf("runCorpusAudit: %v", err)
+	}
+	get := func(name string) *corpusTally {
+		for i, tg := range stats.targets {
+			if tg.name == name {
+				return stats.tallies[i]
+			}
+		}
+		t.Fatalf("no tally for %s", name)
+		return nil
+	}
+	if a := get("amd64"); a.attempted != 1 || a.assembled != 1 {
+		t.Errorf("amd64 = %d/%d, want 1/1", a.assembled, a.attempted)
+	}
+	// lonely_arm64.s is an arm64 file whose package cannot be generated.
+	if a := get("arm64"); a.attempted != 1 || a.assembled != 0 {
+		t.Errorf("arm64 = %d/%d, want 0/1", a.assembled, a.attempted)
+	}
+	if r := get("arm64").reasons["go_asm.h generation failed"]; r != 1 {
+		t.Errorf("arm64 go_asm.h failure count = %d, want 1", r)
+	}
+}
+
+// TestRunCorpusAuditGOOS covers the filename-derived GOOS end to end: a
+// kernel whose name names darwin must have its header type-checked with
+// GOOS=darwin, so the darwin-only constant it offsets with is defined.  The
+// operand mirrors sys_darwin_arm64.s's trampoline, where a missing define
+// leaves an unexpanded symbol in the offset and fails.
+func TestRunCorpusAuditGOOS(t *testing.T) {
+	dir := t.TempDir()
+	write := func(name, src string) {
+		t.Helper()
+		if err := os.WriteFile(filepath.Join(dir, name), []byte(src), 0o644); err != nil {
+			t.Fatal(err)
+		}
+	}
+	write("pkg.go", "package corpus\n")
+	write("plat_darwin.go", "//go:build darwin\n\npackage corpus\n\nconst trampolineNumer = 8\n")
+	write("kern_darwin_arm64.s", "#include \"go_asm.h\"\n"+
+		"GLOBL timebase<>(SB), NOPTR, $16\n"+
+		"TEXT \xc2\xb7g(SB), NOSPLIT, $0-0\n"+
+		"\tMOVD\ttimebase<>+const_trampolineNumer(SB), R0\n"+
+		"\tRET\n")
+
+	stats, err := runCorpusAudit(dir, nil)
+	if err != nil {
+		t.Fatalf("runCorpusAudit: %v", err)
+	}
+	var arm *corpusTally
+	for i, tg := range stats.targets {
+		if tg.name == "arm64" {
+			arm = stats.tallies[i]
+		}
+	}
+	if arm == nil {
+		t.Fatal("no arm64 tally")
+	}
+	if arm.attempted != 1 || arm.assembled != 1 {
+		t.Errorf("arm64 = %d/%d, want 1/1; reasons: %v", arm.assembled, arm.attempted, arm.reasons)
+	}
+}
+
+// TestRunCorpusAuditBuildConstraint covers the //go:build classification end
+// to end: a generic-named file whose constraint admits one target is
+// attempted there alone (cpu_x86.s on amd64), and a file whose constraint
+// admits none of the four targets is never attempted (the msan and
+// goexperiment trees).
+func TestRunCorpusAuditBuildConstraint(t *testing.T) {
+	dir := t.TempDir()
+	write := func(name, src string) {
+		t.Helper()
+		if err := os.WriteFile(filepath.Join(dir, name), []byte(src), 0o644); err != nil {
+			t.Fatal(err)
+		}
+	}
+	write("x86.s", "//go:build 386 || amd64\n\nTEXT \xc2\xb7f(SB), NOSPLIT, $0\n\tRET\n")
+	write("racey.s", "//go:build race\n\nTEXT \xc2\xb7r(SB), NOSPLIT, $0\n\tRET\n")
+	write("plain.s", "TEXT \xc2\xb7p(SB), NOSPLIT, $0\n\tRET\n")
+
+	stats, err := runCorpusAudit(dir, nil)
+	if err != nil {
+		t.Fatalf("runCorpusAudit: %v", err)
+	}
+	tally := func(name string) *corpusTally {
+		for i, tg := range stats.targets {
+			if tg.name == name {
+				return stats.tallies[i]
+			}
+		}
+		t.Fatalf("no tally for %s", name)
+		return nil
+	}
+	if stats.narrowed != 1 || stats.excluded != 1 || stats.generic != 1 {
+		t.Errorf("buckets = narrowed %d, excluded %d, generic %d; want 1, 1, 1", stats.narrowed, stats.excluded, stats.generic)
+	}
+	if a := tally("amd64"); a.attempted != 2 || a.assembled != 2 {
+		t.Errorf("amd64 = %d/%d, want 2/2 (x86.s and plain.s)", a.assembled, a.attempted)
+	}
+	for _, name := range []string{"arm64", "riscv64", "loong64"} {
+		if a := tally(name); a.attempted != 1 || a.assembled != 1 {
+			t.Errorf("%s = %d/%d, want 1/1 (plain.s only)", name, a.assembled, a.attempted)
+		}
+	}
+	if stats.full != 2 {
+		t.Errorf("full = %d, want 2 (x86.s over its one target, plain.s over all four)", stats.full)
+	}
+}
+
+// TestGenerateGoAsmHeaderRuntime pins the generator against the real thing:
+// the runtime package of the ambient toolchain, whose header the toolchain's
+// own -asmhdr output was sampled from.  Skipped in short mode: it type-checks
+// the whole package.  The GOROOT comes from the go command itself, so the
+// test follows whatever toolchain the host provides.
+func TestGenerateGoAsmHeaderRuntime(t *testing.T) {
+	if testing.Short() {
+		t.Skip("type-checks the whole runtime package")
+	}
+	out, err := exec.Command("go", "env", "GOROOT").Output()
+	if err != nil {
+		t.Skipf("no Go toolchain: %v", err)
+	}
+	runtimeDir := filepath.Join(strings.TrimSpace(string(out)), "src", "runtime")
+	dir, err := generateGoAsmHeader(runtimeDir, "", "amd64", t.TempDir())
+	if err != nil {
+		t.Fatalf("generateGoAsmHeader(runtime): %v", err)
+	}
+	b, err := os.ReadFile(dir + "/go_asm.h")
+	if err != nil {
+		t.Fatal(err)
+	}
+	hdr := string(b)
+	for _, want := range []string{
+		"#define const_hashSize 8\n",
+		"#define const_avxSupported 1\n",
+		"#define const_pageSize 8192\n",
+		"#define g_stackguard0 16\n",
+		"#define m__size ",
+	} {
+		if !strings.Contains(hdr, want) {
+			t.Errorf("runtime header misses %q", want)
+		}
+	}
+}

cmd/gasm/audit.go

@@ -0,0 +1,912 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package main
+
+import (
+	"fmt"
+	"go/build/constraint"
+	"maps"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"regexp"
+	"runtime"
+	"slices"
+	"strconv"
+	"strings"
+
+	"sourcedock.dev/petrbalvin/gasm-sdk/arch"
+	"sourcedock.dev/petrbalvin/gasm-sdk/asm"
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
+)
+
+// cmdAuditInstructions cross-checks a gasm encoder against the Go toolchain's
+// own assembler, probed black-box: every mnemonic in the gasm table is offered
+// to go tool asm in its bare form, and a mnemonic counts as known to Go when
+// the error is anything but "unrecognized instruction" (a wrong-shape error
+// still proves the mnemonic exists in Go's tables).  The audit answers three
+// questions at a glance:
+//
+//   - which mnemonics gasm can encode that go tool asm does not know
+//     (superset encodings, usable only through the gasm goobj path);
+//   - which mnemonics the architecture table knows but the encoder cannot
+//     emit yet (the implementation backlog);
+//   - which mnemonics go tool asm knows that gasm cannot encode (feature
+//     gaps).
+//
+// The amd64 derived families (Jcc, CMOVcc, SETcc) exist on both sides by
+// construction and are excluded from the diff; the other architectures list
+// their conditional branches outright.
+func cmdAuditInstructions(args []string) error {
+	fs := newCommand("audit-instructions", "gasm audit-instructions [--corpus [dir]] [--list] [-I dir] [amd64|arm64|riscv64|loong64]", `
+Compare the gasm encoder for the given architecture (default amd64) against
+go tool asm and print the diff: superset encodings (gasm-only, shippable via
+gasm asm --format goobj) and known-but-unencodable names (the backlog).  The
+Go side is probed black-box one bare mnemonic at a time, so the audit tracks
+whatever toolchain `+"`go env GOROOT`"+` provides; the gasm side answers from
+the encoder table on amd64 and from trial assembly over a battery of operand
+shapes elsewhere.  Names go tool asm knows and gasm does not cannot be
+enumerated by probing, because Go's table is visible only through names
+already in the gasm table; the report closes with a note saying so.
+
+With --corpus the audit changes shape: it assembles every .s file under the
+given directory (default GOROOT/src) with the gasm encoder only, no
+toolchain probing.  A file whose name carries a recognisable _arch suffix is
+attempted for that architecture; a file without one is attempted for all
+four, exactly as a GOARCH build would compile it.  The report gives the
+per-architecture pass rates and the most common failure reasons, which drive
+the encodability backlog by frequency rather than by table order.  With
+-list the report also prints every failing file with its reason, per
+architecture.
+`)
+	corpus := fs.Bool("corpus", false, "assemble a corpus of .s files and report pass rates and failure reasons")
+	list := fs.Bool("list", false, "with --corpus, list every failing file with its reason, per architecture")
+	var dirs includeDirs
+	fs.Var(&dirs, "I", "directory to search for #include files (may be repeated)")
+	if err := fs.Parse(args); err != nil {
+		return err
+	}
+	if *corpus {
+		return cmdAuditCorpus(fs.Args(), dirs, *list)
+	}
+	archName := "amd64"
+	switch n := len(fs.Args()); {
+	case n > 1:
+		return &usageError{fmt.Errorf("audit-instructions takes at most one architecture argument")}
+	case n == 1:
+		archName = strings.ToLower(fs.Arg(0))
+	}
+	a, err := auditArch(archName)
+	if err != nil {
+		return err
+	}
+
+	tab := arch.ForArch(a)
+	var names []string
+	seen := map[string]bool{}
+	for _, in := range tab.Instructions() {
+		name := strings.ToUpper(in.Name)
+		if a == arch.AMD64 && derivedFamily(name) || seen[name] {
+			continue
+		}
+		seen[name] = true
+		names = append(names, name)
+	}
+
+	goKnown, err := probeGoAsm(goarchName(a), names)
+	if err != nil {
+		return err
+	}
+
+	var superset, backlog, shared []string
+	for _, name := range names {
+		switch {
+		case !gasmEncodable(a, name):
+			backlog = append(backlog, name)
+		case !goKnown[name]:
+			superset = append(superset, name)
+		default:
+			shared = append(shared, name)
+		}
+	}
+	// GO-ONLY is not enumerable by probing: Go's table is only visible
+	// through names we already know, so nothing can be reported there.
+
+	slices.Sort(superset)
+	slices.Sort(backlog)
+	slices.Sort(shared)
+
+	w := os.Stdout
+	fmt.Fprintf(w, "gasm table (%s, families excluded): %d mnemonics\n", archName, len(names))
+	fmt.Fprintf(w, "gasm encodable: %d   go tool asm recognised: %d\n", len(shared)+len(superset), countTrue(goKnown))
+	fmt.Fprintf(w, "shared: %d\n", len(shared))
+	fmt.Fprintf(w, "\nSuperset encodings (gasm-only; ship via gasm asm --format goobj):\n")
+	for _, n := range superset {
+		fmt.Fprintf(w, "  %s\n", n)
+	}
+	fmt.Fprintf(w, "\nKnown but not encodable (backlog):\n")
+	for _, n := range backlog {
+		fmt.Fprintf(w, "  %s\n", n)
+	}
+	fmt.Fprintf(w, "\nGo-only names cannot be enumerated by probing; extend the gasm\n")
+	fmt.Fprintf(w, "table from the Go release notes when a new instruction family ships.\n")
+	return nil
+}
+
+// auditArch resolves the audit's architecture argument.
+func auditArch(name string) (arch.Arch, error) {
+	switch strings.ToLower(name) {
+	case "amd64":
+		return arch.AMD64, nil
+	case "arm64":
+		return arch.ARM64, nil
+	case "riscv64", "riscv":
+		return arch.RISCV, nil
+	case "loong64", "loong":
+		return arch.LOONG64, nil
+	}
+	return arch.Unknown, &usageError{fmt.Errorf("unknown architecture %q: want amd64, arm64, riscv64 or loong64", name)}
+}
+
+// goarchName maps an arch identifier onto its GOARCH spelling.
+func goarchName(a arch.Arch) string {
+	switch a {
+	case arch.ARM64:
+		return "arm64"
+	case arch.RISCV:
+		return "riscv64"
+	case arch.LOONG64:
+		return "loong64"
+	}
+	return "amd64"
+}
+
+func countTrue(m map[string]bool) int {
+	n := 0
+	for _, v := range m {
+		if v {
+			n++
+		}
+	}
+	return n
+}
+
+// derivedFamily reports whether a mnemonic belongs to a family both
+// assemblers construct from condition codes rather than list exhaustively
+// (JEQ/CMOVLGT/SETNE and friends).  Such names never probe cleanly, so
+// including them in the diff would be noise.  amd64 only: the other
+// architectures list their conditional branches outright.
+func derivedFamily(name string) bool {
+	if strings.HasPrefix(name, "J") && name != "JMP" && name != "JMPQ" {
+		return true
+	}
+	if strings.HasPrefix(name, "CMOV") || strings.HasPrefix(name, "SET") {
+		return true
+	}
+	return false
+}
+
+var unrecognizedRe = regexp.MustCompile(`unrecognized instruction`)
+
+// probeGoAsm feeds every mnemonic to go tool asm in one generated file and
+// classifies the diagnostics.  "Unrecognized instruction" is a parse-stage
+// verdict on the mnemonic alone, so a single bare-instruction probe per
+// mnemonic decides recognition; the combined file still reports every line's
+// error even when others fail.
+func probeGoAsm(goarch string, names []string) (map[string]bool, error) {
+	dir, err := os.MkdirTemp("", "gasm-audit")
+	if err != nil {
+		return nil, err
+	}
+	defer os.RemoveAll(dir)
+
+	var sb strings.Builder
+	sb.WriteString("TEXT ·probe(SB), 4, $0\n\tRET\n")
+	lineMnemonic := map[int]string{}
+	line := 3
+	for _, name := range names {
+		fmt.Fprintf(&sb, "TEXT ·p%s%d(SB), 4, $0\n", sanitize(name), line)
+		sb.WriteString("\t" + name + "\n\tRET\n")
+		lineMnemonic[line+1] = name // the instruction line, after TEXT
+		line += 3
+	}
+	probePath := filepath.Join(dir, "probe.s")
+	if err := os.WriteFile(probePath, []byte(sb.String()), 0o644); err != nil {
+		return nil, err
+	}
+
+	toolDir, err := exec.Command("go", "env", "GOTOOLDIR").Output()
+	if err != nil {
+		return nil, fmt.Errorf("go env GOTOOLDIR: %w", err)
+	}
+	asmBin := filepath.Join(strings.TrimSpace(string(toolDir)), "asm")
+	if _, err := os.Stat(asmBin); err != nil {
+		return nil, fmt.Errorf("go tool asm not found at %s", asmBin)
+	}
+	cmd := exec.Command(asmBin, "-p", "probe", "-o", filepath.Join(dir, "probe.o"), probePath)
+	cmd.Env = append(os.Environ(), "GOARCH="+goarch, "GOOS="+runtime.GOOS)
+	out, _ := cmd.CombinedOutput()
+	// The expected failure mode is a non-zero exit with compiler diagnostics
+	// on stdout; empty output means the probe broke at the exec level (a
+	// killed child, a tool that would not start), and seeding every name as
+	// recognized on that silence would fake a clean audit.
+	if len(out) == 0 {
+		return nil, fmt.Errorf("go tool asm probe for GOARCH=%s produced no output", goarch)
+	}
+
+	result := map[string]bool{}
+	for _, name := range names {
+		result[name] = true // no news = the name parsed fine
+	}
+	reParse := regexp.MustCompile(`probe\.s:(\d+):`)
+	for l := range strings.SplitSeq(string(out), "\n") {
+		m := reParse.FindStringSubmatch(l)
+		if m == nil {
+			continue
+		}
+		lineNo, err := strconv.Atoi(m[1])
+		if err != nil {
+			continue
+		}
+		if name, ok := lineMnemonic[lineNo]; ok && unrecognizedRe.MatchString(l) {
+			result[name] = false
+		}
+	}
+	return result, nil
+}
+
+// probeShapes lists representative operand shapes for the encodability
+// probe.  The assemblers report an unknown mnemonic and a known mnemonic
+// with no supported form alike ("unsupported <arch> instruction"), so only
+// a shape that assembles cleanly counts, and the backlog over-approximates:
+// a name whose real forms the battery misses lands there.  amd64 keeps its
+// exact table-driven check.
+func probeShapes(a arch.Arch) []string {
+	switch a {
+	case arch.ARM64:
+		return []string{
+			"X0, X1, X2", "X0, X1", "X0", "$1, X0", "X0, (X1)", "(X0), X1",
+			"X0, (X1, 8)", "(SP), X0", "F0, F1, F2", "F0, F1", "F0",
+			"V0.B16, V1.B16, V2.B16", "p2", "X0, p2", "X0, X1, p2",
+			// The conditional select family spells the condition first
+			// and takes R register spellings.
+			"EQ, R0, R1, R2", "EQ, R0, R1", "EQ, R0",
+			"GE, F0, F1, F2", "NE, F0, F1, $0",
+			// Pairs, acquire/release and exclusive atomics, LSE-AL forms.
+			"(R0), R1", "R0, (R1)", "R1, (R2), R3", "(R2, R3), 8(R1)",
+			"8(R1), (R2, R3)", "R1, R2, (R3)", "(R0)",
+			// System operations and their register/operand names.
+			"$4, R1, p2", "$35943", "$1", "$1, SPSel", "SPSel, R0",
+			"IVAC, R0", "(R0), PLDL1KEEP", "R1, R2, R3, R4",
+			// SIMD element, structure and literal-pool forms.
+			"(R0), [V1.B16]", "[V1.B16], (R0)", "V13.S[0], R1",
+			"R1, V2.B[3]", "$4, V1.B16, V2.B16", "V1.B16, (R0)",
+			"(R0), V1.B16", "",
+			// The spellings GOROOT's own kernels use, from the
+			// differential kernels this table was proven against.
+			"R0, p2", "R0, R1", "F0, F1, F2, F3", "$4, V1.B16, V2.B16, V3.B16, V4.B16",
+			"(R0), [V0.B8, V1.B8, V2.B8, V3.B8]", "$1, $2, V1",
+			"R0, R1, p2", "p2, R1", "$1234, R1", "DCZID_EL0, R1",
+			"$0", "R1, $4, EQ", "$33, R1, $25, R2", "$4, R1, p2",
+			"$4, V1.B8, V2.B8, V3.B8", "$63, V1.D2, V2.D2, V3.D2",
+			"V1.B16, [V2.B16], V3.B16", "V1.B8, [V2.B16, V3.B16], V4.B8",
+			"$4, V1.B16, V2.B16, V3.B16", "$15, V1", "V1, V2, p2",
+			"R0, R1, $1, $4, p2",
+			// The landing-pad kind, the compiler's PCDATA
+			// bookkeeping and the four-operand bitfield
+			// insert/extract family, as the toolchain's own
+			// testdata spells them.
+			"C", "$1, $0", "$0, R1, $1, R2",
+		}
+	case arch.RISCV:
+		return []string{
+			"X5, X6, X7", "X5, X6", "X5", "$1, X5", "X5, (X6)", "$1, X5, X6",
+			"(X5), X6", "F0, F1, F2", "F0, F1", "p2", "X1, p2", "X0, p2",
+			"X5, X6, p2", "p2(SB)",
+			// AMO atomics: destination, base, source.
+			"R5, (R4), R6", "X5, (X4), X6",
+			// Segment stores take the first vector register aligned
+			// to the segment count, as the toolchain requires.
+			"(X5), X6, V0, V8", "(X5), X6, V0", "(X5), X0, V4",
+			// The FP multiply-add family takes four registers.
+			"F0, F1, F2, F3",
+			// The RVV slice: register, vector-register and vtype forms.
+			"V1, V2, V3", "V1, X5, V2", "V1", "V1, (X5)", "(X5), V1",
+			"$15, V1", "$15", "V1, V2", "V1, X5",
+			"X5, X6, p2", "R5, R6, p2",
+			"X5, E8, M8, TA, MA, X6", "$4, E32, M1, TA, MA, X1",
+			"(X5), X6, V1, V2",
+			// The CSR immediate forms the toolchain's testdata spells:
+			// immediate, CSR name, destination.
+			"$2, TIME, X5",
+			"",
+		}
+	case arch.LOONG64:
+		return []string{
+			"R4, R5, R6", "R4, R5", "R4", "$1, R4", "R4, (R5)", "(R4), R5",
+			"F0, F1, F2", "F0, F1", "p2", "R1, p2", "R4, p2",
+			"$1, R4, R5, R6", "$65536, R4", "R4, R5, p2", "p2(SB)",
+			// AMO atomics: destination, base, source.
+			"R5, (R4), R6", "X5, (X4), X6",
+			// Segment stores take the first vector register aligned
+			// to the segment count, as the toolchain requires.
+			"(X5), X6, V0, V8", "(X5), X6, V0", "(X5), X0, V4",
+			// The LSX and LASX banks share the 5-bit numbering with F.
+			"V1, V2, V3", "X1, X2, X3", "V1, V2", "X1, X2", "V1", "X1",
+			// The vector compare-to-flag forms land in an FCC register.
+			"V1, FCC0", "X1, FCC0",
+			// The compiler's bookkeeping pair and the raw spellings the
+			// toolchain's own testdata carries: JIRL rd, rj, offset (the
+			// form RET lowers to), the prefetch with a 32-bit address and
+			// hint, and the byte-shuffle quads.
+			"$1, $0", "R1, R5, 0", "0(R7), $5, $0", "(R7), $5, $0",
+			"V1, V2, V3, V4", "X1, X2, X3, X4",
+			"",
+		}
+	}
+	return nil
+}
+
+// gasmEncodable reports whether the gasm encoder for a can emit the
+// mnemonic, decided by trial assembly over the shape battery.
+func gasmEncodable(a arch.Arch, name string) bool {
+	switch a {
+	case arch.ARM64, arch.RISCV, arch.LOONG64:
+	default:
+		return asm.Encodable(name)
+	}
+	for _, shape := range probeShapes(a) {
+		if gasmAssembles(a, name, shape) {
+			return true
+		}
+	}
+	return false
+}
+
+// gasmAssembles reports whether a one-instruction probe file containing name
+// with the given operand shape assembles without error.
+func gasmAssembles(a arch.Arch, name, shape string) bool {
+	src := "TEXT ·p(SB), NOSPLIT, $0\n\t" + name
+	if shape != "" {
+		src += " " + shape
+	}
+	src += "\n\tRET\np2:\n\tRET\n"
+	f, errs := parser.Parse("probe.s", src)
+	if len(errs) > 0 {
+		return false
+	}
+	var err error
+	switch a {
+	case arch.ARM64:
+		_, err = asm.AssembleFileARM64(f)
+	case arch.RISCV:
+		_, err = asm.AssembleFileRISCV(f)
+	case arch.LOONG64:
+		_, err = asm.AssembleFileLOONG64(f)
+	}
+	return err == nil
+}
+
+// sanitize makes a mnemonic safe for use in a Go symbol name.
+func sanitize(name string) string {
+	return strings.NewReplacer(".", "_", "$", "_").Replace(name)
+}
+
+// --- corpus audit -----------------------------------------------------------
+
+// corpusTarget is one architecture row of the corpus report.
+type corpusTarget struct {
+	a    arch.Arch
+	name string
+}
+
+// corpusTally accumulates one architecture's attempts over the corpus.
+type corpusTally struct {
+	attempted int
+	assembled int
+	reasons   map[string]int    // failure reason → count
+	example   map[string]string // failure reason → one representative file
+	fails     []corpusFailure   // every failure, in file order, for --list
+}
+
+// corpusFailure is one failed attempt, recorded for the --list report.
+type corpusFailure struct {
+	path   string
+	reason string
+	detail string
+}
+
+func (t *corpusTally) fail(path string, err error) {
+	reason := corpusReason(err)
+	t.reasons[reason]++
+	if t.example[reason] == "" {
+		t.example[reason] = path
+	}
+	t.fails = append(t.fails, corpusFailure{path: path, reason: reason, detail: firstLine(err.Error())})
+}
+
+// cmdAuditCorpus implements audit-instructions --corpus.  The include
+// directories carry #include resolution over a corpus whose files refer to
+// headers such as GOROOT/pkg/include, the same -I a toolchain comparison
+// needs.
+func cmdAuditCorpus(args []string, dirs includeDirs, list bool) error {
+	if len(args) > 1 {
+		return &usageError{fmt.Errorf("audit-instructions --corpus takes at most one directory argument")}
+	}
+	root := ""
+	if len(args) == 1 {
+		root = args[0]
+	} else {
+		out, err := exec.Command("go", "env", "GOROOT").Output()
+		if err != nil {
+			return fmt.Errorf("locate GOROOT: %w", err)
+		}
+		root = filepath.Join(strings.TrimSpace(string(out)), "src")
+	}
+	// The toolchain's shipped headers (funcdata.h and friends) define the
+	// macros GOROOT files include; a corpus audit measures those files, so
+	// the header directory joins the search path automatically.  go_asm.h
+	// is compiler-generated per package, so it is not resolved from here:
+	// files that include it get one generated per target architecture,
+	// which runCorpusAudit arranges.
+	if out, err := exec.Command("go", "env", "GOROOT").Output(); err == nil {
+		pkgInclude := filepath.Join(strings.TrimSpace(string(out)), "pkg", "include")
+		if fi, err := os.Stat(pkgInclude); err == nil && fi.IsDir() {
+			seen := false
+			for _, d := range dirs {
+				if d == pkgInclude {
+					seen = true
+				}
+			}
+			if !seen {
+				dirs = append(dirs, pkgInclude)
+			}
+		}
+	}
+	stats, err := runCorpusAudit(root, dirs)
+	if err != nil {
+		return err
+	}
+	printCorpusStats(stats, list)
+	return nil
+}
+
+// corpusStats is the outcome of one corpus audit run.
+type corpusStats struct {
+	root      string
+	files     int
+	generic   int // files attempted for all four architectures
+	narrowed  int // files whose //go:build admits a proper subset of the four
+	excluded  int // files whose //go:build admits none of the four: never compiled
+	otherPort int // files named for another Go port: never attempted
+	full      int // files that assembled for every applicable target architecture
+	targets   []corpusTarget
+	tallies   []*corpusTally
+}
+
+// runCorpusAudit assembles every .s file under root and returns the stats.
+// goPortSuffixes lists every architecture the Go project ports to.  A file
+// named for one of them belongs to that port's build, not to the generic
+// set, even when gasm does not support the architecture.
+var goPortSuffixes = []string{
+	"386", "amd64", "arm", "arm64", "loong64", "mips", "mips64",
+	"mips64le", "mipsle", "mips64x", "mipsx", "ppc64", "ppc64le",
+	"ppc64x", "riscv", "riscv64", "s390x", "wasm",
+}
+
+// otherPortFile reports whether the file belongs to a build no supported
+// target ever compiles: either its name carries a Go-architecture suffix
+// gasm does not support, or, for a file with no architecture suffix at all,
+// it names another GOOS, which go/build drops from the file set
+// (rt0_js_wasm.s is a javascript build, not a generic one).
+func otherPortFile(path string) bool {
+	if otherGOOSFile(path) {
+		return true
+	}
+	base := path
+	if i := strings.LastIndexByte(base, '/'); i >= 0 {
+		base = base[i+1:]
+	}
+	for _, sfx := range goPortSuffixes {
+		if strings.HasSuffix(base, "_"+sfx+".s") {
+			return true
+		}
+	}
+	return false
+}
+
+// goOSNames are the GOOS values go/build recognises in file names.
+var goOSNames = map[string]bool{
+	"aix": true, "android": true, "darwin": true, "dragonfly": true,
+	"freebsd": true, "hurd": true, "illumos": true, "ios": true,
+	"js": true, "linux": true, "nacl": true, "netbsd": true,
+	"openbsd": true, "plan9": true, "solaris": true, "wasip1": true,
+	"windows": true, "zos": true,
+}
+
+// resolveGOOS validates a -GOOS flag value, mirroring the architecture
+// check's surface: a usage error naming what the tool accepts.
+func resolveGOOS(name string) (string, error) {
+	lower := strings.ToLower(name)
+	if goOSNames[lower] {
+		return lower, nil
+	}
+	return "", &usageError{fmt.Errorf("unknown GOOS %q: want one of %s", name, strings.Join(slices.Sorted(maps.Keys(goOSNames)), ", "))}
+}
+
+// goosFromFilename returns the GOOS the file's name carries, by go/build's
+// goodOSArchFile rule: the GOOS segment sits last, or last before the
+// architecture segment (sys_darwin_arm64.s, vlop_arm.s carries none).  An
+// empty result means the name names no GOOS and the ambient one applies.
+func goosFromFilename(path string) string {
+	base := path
+	if i := strings.LastIndexByte(base, '/'); i >= 0 {
+		base = base[i+1:]
+	}
+	base = strings.TrimSuffix(base, ".s")
+	// go/build ignores everything before the first underscore, so a GOOS
+	// segment is only ever looked for from there on.
+	i := strings.IndexByte(base, '_')
+	if i < 0 {
+		return ""
+	}
+	segs := strings.Split(base[i:], "_")
+	if n := len(segs); n >= 2 && goOSNames[segs[n-2]] && slices.Contains(goPortSuffixes, segs[n-1]) {
+		return segs[n-2]
+	}
+	if goOSNames[segs[len(segs)-1]] {
+		return segs[len(segs)-1]
+	}
+	return ""
+}
+
+// otherGOOSFile reports whether the file's name names a GOOS other than the
+// host's, by go/build's file-name rules.
+func otherGOOSFile(path string) bool {
+	base := path
+	if i := strings.LastIndexByte(base, '/'); i >= 0 {
+		base = base[i+1:]
+	}
+	for seg := range strings.SplitSeq(strings.TrimSuffix(base, ".s"), "_") {
+		if goOSNames[seg] && seg != runtime.GOOS {
+			return true
+		}
+	}
+	return false
+}
+
+// buildConstraint returns the file's leading //go:build expression, or nil
+// when the file carries none.  The constraint governs the same header block
+// go/build reads: blank lines and comments may precede it, and the first
+// line that is neither ends the block.  A constraint that does not parse
+// narrows nothing, so the file stays in the attempted set: the audit must
+// never exclude a file the toolchain would compile.
+func buildConstraint(src string) constraint.Expr {
+	for line := range strings.SplitSeq(src, "\n") {
+		t := strings.TrimSpace(line)
+		switch {
+		case t == "":
+			continue
+		case strings.HasPrefix(t, "//"):
+			if constraint.IsGoBuild(t) {
+				e, err := constraint.Parse(t)
+				if err != nil {
+					return nil
+				}
+				return e
+			}
+			continue
+		default:
+			return nil
+		}
+	}
+	return nil
+}
+
+// unixOS is go/build's unixOS set: the GOOSes the unix build tag admits.
+var unixOS = map[string]bool{
+	"aix": true, "android": true, "darwin": true, "dragonfly": true,
+	"freebsd": true, "hurd": true, "illumos": true, "ios": true,
+	"linux": true, "netbsd": true, "openbsd": true, "solaris": true,
+}
+
+// constraintTags answers the build tags a plain `go build` sets for a
+// target: the GOOS and GOARCH, gc, and unix on the unix-like GOOSes.  No
+// experiment, sanitiser or cgo tag is ever true: the audit models the
+// default build, and no GOROOT assembly file's constraint hinges on cgo.
+func constraintTags(goarch, goos string) func(string) bool {
+	return func(tag string) bool {
+		switch tag {
+		case goarch, goos, "gc":
+			return true
+		case "unix":
+			return unixOS[goos]
+		}
+		return false
+	}
+}
+
+func runCorpusAudit(root string, dirs includeDirs) (*corpusStats, error) {
+	files, err := asmFiles(root)
+	if err != nil {
+		return nil, err
+	}
+
+	targets := []corpusTarget{
+		{arch.AMD64, "amd64"},
+		{arch.ARM64, "arm64"},
+		{arch.RISCV, "riscv64"},
+		{arch.LOONG64, "loong64"},
+	}
+	tallies := make([]*corpusTally, len(targets))
+	for i := range tallies {
+		tallies[i] = &corpusTally{reasons: map[string]int{}, example: map[string]string{}}
+	}
+	// full is the north-star number: a file counts when every architecture
+	// its build admits assembles it.
+	full, generic, otherPort, narrowedCount, excluded := 0, 0, 0, 0, 0
+
+	// Header generation is created on first use, so a corpus with no
+	// go_asm.h includes never pays for a temp directory.
+	var hdr *asmhdrCache
+	defer func() {
+		if hdr != nil {
+			hdr.close()
+		}
+	}()
+
+	for _, path := range files {
+		src, err := readSource(path)
+		if err != nil {
+			return nil, err
+		}
+
+		// The GOOS the header generation type-checks under follows the
+		// file's name when the name carries one; the ambient GOOS is the
+		// honest guess otherwise (a build tag naming another GOOS is
+		// invisible to a file-name rule).
+		goos := goosFromFilename(path)
+
+		// The GOOS the header generation type-checks under follows the
+		// file's name when the name carries one; the ambient GOOS is the
+		// honest guess otherwise.
+		namedArch := arch.FromFilename(path)
+		var wanted []int // indexes into targets
+		other := false
+		switch {
+		case namedArch != arch.Unknown:
+			for i, tg := range targets {
+				if tg.a == namedArch {
+					wanted = append(wanted, i)
+				}
+			}
+		case otherPortFile(path):
+			// A file named for a Go port gasm does not support (arm,
+			// 386, s390x, ...) or for another GOOS is compiled by no
+			// supported-arch build, so it is neither generic nor a
+			// per-arch attempt: counting it as generic would make the
+			// headline unreachably low for reasons no supported target
+			// can fix.
+			other = true
+			otherPort++
+		default:
+			for i := range targets {
+				wanted = append(wanted, i)
+			}
+		}
+
+		// A //go:build constraint narrows the set of targets the file is
+		// assembled for, the way the go command compiles the file only for
+		// the targets the expression admits: cpu_x86.s belongs to the x86
+		// build alone, and a file whose constraint admits none of the four
+		// targets (the goexperiment.runtimesecret and msan trees) is
+		// compiled by no supported build.  The tags mirror what a plain
+		// `go build` sets: the GOOS and GOARCH, gc, and unix on the
+		// unix-like GOOSes; no experiment, sanitiser or cgo tag is ever
+		// true.  The GOOS is the file's own when the name carries one,
+		// else the ambient one.
+		goosForEval := goos
+		if goosForEval == "" {
+			goosForEval = runtime.GOOS
+		}
+		narrowed := false
+		if len(wanted) > 0 {
+			if ce := buildConstraint(src); ce != nil {
+				kept := make([]int, 0, len(wanted))
+				for _, i := range wanted {
+					tg := targets[i]
+					if ce.Eval(constraintTags(goarchName(tg.a), goosForEval)) {
+						kept = append(kept, i)
+					}
+				}
+				if len(kept) < len(wanted) {
+					narrowed = true
+				}
+				wanted = kept
+			}
+		}
+
+		switch {
+		case other:
+			// already tallied above
+		case len(wanted) == 0:
+			excluded++
+		case namedArch != arch.Unknown:
+			// a per-arch attempt over the constraint's subset
+		case narrowed:
+			narrowedCount++
+		default:
+			generic++
+		}
+
+		// A file that includes go_asm.h parses against a per-target header:
+		// the defines differ per architecture (internal/cpu's layout, for
+		// one) and per GOOS (sys_darwin_arm64.s's trampoline constants,
+		// for another), so the parse cannot be shared the way a
+		// header-free file's can.  A generation failure is a failure for
+		// every target, named for the package rather than a bare "include
+		// not found".  A header already resolvable in the package
+		// directory or the -I list is left alone.
+		if len(wanted) > 0 && needsGoAsmHeader(src) && !goAsmHeaderResolved(filepath.Dir(path), dirs) {
+			if hdr == nil {
+				if hdr, err = newAsmhdrCache(); err != nil {
+					return nil, err
+				}
+			}
+			pkgDir := filepath.Dir(path)
+			ok := true
+			for _, i := range wanted {
+				tg, t := targets[i], tallies[i]
+				t.attempted++
+				hdrDir, err := hdr.dirFor(pkgDir, goos, goarchName(tg.a))
+				if err != nil {
+					ok = false
+					t.fail(path, err)
+					continue
+				}
+				f, errs := parser.ParseWithOptions(path, src, parser.Options{
+					Expand:      true,
+					IncludeDirs: append(slices.Clone(dirs), hdrDir),
+					Predefines:  platformPredefinesFor(goarchName(tg.a), goos),
+				})
+				if len(errs) > 0 {
+					ok = false
+					t.fail(path, errs[0])
+					continue
+				}
+				if _, err := assembleFile(tg.a, f, goos); err != nil {
+					ok = false
+					t.fail(path, err)
+					continue
+				}
+				t.assembled++
+			}
+			if ok && len(wanted) > 0 {
+				full++
+			}
+			continue
+		}
+
+		ok := true
+		for _, i := range wanted {
+			tg, t := targets[i], tallies[i]
+			t.attempted++
+			// The parse carries the target's platform predefines, so it
+			// cannot be shared across targets the way a header-free file's
+			// could: a #ifdef GOARCH_arm block must be live on arm64 and
+			// dead everywhere else.
+			f, errs := parser.ParseWithOptions(path, src, parser.Options{
+				Expand:      true,
+				IncludeDirs: dirs,
+				Predefines:  platformPredefinesFor(goarchName(tg.a), goos),
+			})
+			var err error
+			if len(errs) > 0 {
+				err = errs[0] // a parse failure is a failure for every target
+			} else {
+				_, err = assembleFile(tg.a, f, goos)
+			}
+			if err != nil {
+				ok = false
+				t.fail(path, err)
+				continue
+			}
+			t.assembled++
+		}
+		if ok && len(wanted) > 0 {
+			full++
+		}
+	}
+
+	return &corpusStats{
+		root:      root,
+		files:     len(files),
+		generic:   generic,
+		narrowed:  narrowedCount,
+		excluded:  excluded,
+		otherPort: otherPort,
+		full:      full,
+		targets:   targets,
+		tallies:   tallies,
+	}, nil
+}
+
+// printCorpusStats renders the corpus audit report.
+func printCorpusStats(s *corpusStats, list bool) {
+	fmt.Printf("corpus %s: %d files (%d generic, attempted for all architectures; %d narrowed by //go:build; %d excluded by //go:build; %d named for other Go ports, never attempted)\n",
+		s.root, s.files, s.generic, s.narrowed, s.excluded, s.otherPort)
+	// The rate is over the files a supported build would attempt: the
+	// other ports' files and the ones no supported target compiles sit in
+	// the count for completeness but can never assemble, so counting them
+	// in the denominator would report the gap of platforms gasm
+	// deliberately does not target.
+	attemptable := max(s.files-s.otherPort-s.excluded, 1)
+	fmt.Printf("  assemble for every applicable target: %d of %d attemptable (%.1f%%)\n", s.full, attemptable, 100*float64(s.full)/float64(attemptable))
+	for i, tg := range s.targets {
+		t := s.tallies[i]
+		fmt.Printf("  %s: %d/%d attempted\n", tg.name, t.assembled, t.attempted)
+		for _, r := range topReasons(t) {
+			fmt.Printf("    %4d  %s\n", t.reasons[r], r)
+			fmt.Printf("         e.g. %s\n", t.example[r])
+		}
+		if !list {
+			continue
+		}
+		for _, f := range t.fails {
+			fmt.Printf("    FAIL %s\n", f.path)
+			fmt.Printf("         %s: %s\n", f.reason, f.detail)
+		}
+	}
+}
+
+// corpusReason buckets an assembly or parse failure for the histogram.
+func corpusReason(err error) string {
+	msg := err.Error()
+	switch {
+	case strings.Contains(msg, "go_asm.h for GOARCH"):
+		return "go_asm.h generation failed"
+	case strings.Contains(msg, "unsupported"), strings.Contains(msg, "cannot encode"):
+		return "instruction not encodable"
+	case strings.Contains(msg, "undefined label"):
+		return "undefined label"
+	case strings.Contains(msg, "undefined symbol"), strings.Contains(msg, "external symbol"), strings.Contains(msg, "file-level assembly"):
+		return "undefined symbol or external"
+	case strings.Contains(msg, "operand"), strings.Contains(msg, "operand form"):
+		return "unsupported operand form"
+	default:
+		return "other: " + firstLine(msg)
+	}
+}
+
+// topReasons returns at most five reasons, most frequent first.
+func topReasons(t *corpusTally) []string {
+	type kv struct {
+		k string
+		n int
+	}
+	var kvs []kv
+	for k, n := range t.reasons {
+		kvs = append(kvs, kv{k, n})
+	}
+	slices.SortFunc(kvs, func(a, b kv) int { return b.n - a.n })
+	if len(kvs) > 5 {
+		kvs = kvs[:5]
+	}
+	out := make([]string, len(kvs))
+	for i, kv := range kvs {
+		out[i] = kv.k
+	}
+	return out
+}
+
+// firstLine returns the first line of an error message, truncated.
+func firstLine(msg string) string {
+	if i := strings.IndexByte(msg, '\n'); i >= 0 {
+		msg = msg[:i]
+	}
+	if len(msg) > 80 {
+		msg = msg[:80]
+	}
+	return msg
+}

cmd/gasm/audit_test.go

@@ -0,0 +1,126 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package main
+
+import (
+	"runtime"
+	"testing"
+
+	"sourcedock.dev/petrbalvin/gasm-sdk/arch"
+)
+
+func TestDerivedFamily(t *testing.T) {
+	for _, n := range []string{"JEQ", "JLT", "JCC", "CMOVLGT", "SETNE", "SETA"} {
+		if !derivedFamily(n) {
+			t.Errorf("derivedFamily(%q) = false, want true", n)
+		}
+	}
+	for _, n := range []string{"JMP", "ADDQ", "VPGATHERDD", "MOVBE", "PSHUFB"} {
+		if derivedFamily(n) {
+			t.Errorf("derivedFamily(%q) = true, want false", n)
+		}
+	}
+}
+
+func TestSanitize(t *testing.T) {
+	if got := sanitize("VPCMP.UB"); got != "VPCMP_UB" {
+		t.Errorf("sanitize: got %q", got)
+	}
+}
+
+func TestAuditArch(t *testing.T) {
+	for in, want := range map[string]arch.Arch{
+		"amd64": arch.AMD64, "arm64": arch.ARM64,
+		"riscv64": arch.RISCV, "riscv": arch.RISCV,
+		"loong64": arch.LOONG64, "LOONG": arch.LOONG64,
+	} {
+		got, err := auditArch(in)
+		if err != nil || got != want {
+			t.Errorf("auditArch(%q) = %v, %v; want %v", in, got, err, want)
+		}
+	}
+	if _, err := auditArch("mips"); err == nil {
+		t.Error("auditArch(mips) must fail")
+	}
+}
+
+func TestGasmEncodable(t *testing.T) {
+	cases := []struct {
+		a   arch.Arch
+		yes string
+		no  string
+	}{
+		{arch.AMD64, "ADDQ", "NOSUCHMNEMONIC"},
+		{arch.ARM64, "ADD", "NOSUCHMNEMONIC"},
+		{arch.RISCV, "ADD", "NOSUCHMNEMONIC"},
+		{arch.LOONG64, "ADDV", "NOSUCHMNEMONIC"},
+	}
+	for _, c := range cases {
+		if !gasmEncodable(c.a, c.yes) {
+			t.Errorf("%s: %s should be encodable", c.a, c.yes)
+		}
+		if gasmEncodable(c.a, c.no) {
+			t.Errorf("%s: %s should not be encodable", c.a, c.no)
+		}
+	}
+}
+
+// TestBuildConstraint pins the //go:build reader: the constraint governs the
+// leading comment block, the first non-comment line ends it (a tag below a
+// #include governs nothing, exactly as go/build drops it), and a file
+// without one admits every target.
+func TestBuildConstraint(t *testing.T) {
+	admits := func(src, goarch, goos string) bool {
+		t.Helper()
+		e := buildConstraint(src)
+		if e == nil {
+			return true
+		}
+		return e.Eval(constraintTags(goarch, goos))
+	}
+	const ret = "TEXT \xc2\xb7f(SB), NOSPLIT, $0\n\tRET\n"
+	cases := []struct {
+		name         string
+		src          string
+		amd64, arm64 bool
+	}{
+		{"no constraint", ret, true, true},
+		{"x86 only", "//go:build 386 || amd64\n\n" + ret, true, false},
+		{"arm64 and linux", "//go:build arm64 && linux\n\n" + ret, false, true},
+		{"msan never", "//go:build msan\n\n" + ret, false, false},
+		{"experiment never", "//go:build goexperiment.runtimesecret\n\n" + ret, false, false},
+		{"below an include governs nothing", "#include \"textflag.h\"\n//go:build amd64\n" + ret, true, true},
+		{"unparsable narrows nothing", "//go:build (amd64\n" + ret, true, true},
+	}
+	for _, c := range cases {
+		t.Run(c.name, func(t *testing.T) {
+			if got := admits(c.src, "amd64", runtime.GOOS); got != c.amd64 {
+				t.Errorf("amd64 admission = %v, want %v", got, c.amd64)
+			}
+			if got := admits(c.src, "arm64", runtime.GOOS); got != c.arm64 {
+				t.Errorf("arm64 admission = %v, want %v", got, c.arm64)
+			}
+		})
+	}
+}
+
+// TestConstraintTags pins the tag set a plain `go build` sets: the GOOS and
+// GOARCH, gc, unix on the unix-like GOOSes; nothing else is ever true.
+func TestConstraintTags(t *testing.T) {
+	ok := constraintTags("amd64", "linux")
+	for _, tag := range []string{"amd64", "linux", "gc", "unix"} {
+		if !ok(tag) {
+			t.Errorf("tag %q = false, want true", tag)
+		}
+	}
+	for _, tag := range []string{"arm64", "freebsd", "darwin", "cgo", "race", "msan", "goexperiment.runtimesecret"} {
+		if ok(tag) {
+			t.Errorf("tag %q = true, want false", tag)
+		}
+	}
+	fb := constraintTags("arm64", "freebsd")
+	if !fb("unix") {
+		t.Error("unix on freebsd = false, want true")
+	}
+}

cmd/gasm/debug_linux_amd64.go

@@ -1,193 +0,0 @@
-// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
-// SPDX-License-Identifier: BSD-3-Clause
-
-//go:build linux && amd64
-
-package main
-
-import (
-	"fmt"
-	"os"
-	"sort"
-	"strings"
-
-	"sourcedock.dev/petrbalvin/gasm-devkit/debug"
-	"sourcedock.dev/petrbalvin/gasm-devkit/verify"
-)
-
-func cmdDebug(args []string) int {
-	fs := newCommand("debug", "gasm debug <file.s> --func <name>", `
-Interactive debugger for JIT-assembled amd64 functions.  Launches the
-function in a traced subprocess (ptrace), then provides a REPL for
-single-stepping, breakpoints, register and memory inspection.
-
-REPL commands:
-  break <label|addr>   set a breakpoint at a label or absolute address
-  step [n]             single-step n instructions (default 1)
-  continue             run until next breakpoint or exit
-  regs                 print general-purpose registers
-  x [addr] [len]       hex-dump memory (default: current PC, 64 bytes)
-  labels               list function labels and offsets
-  quit                 kill the debuggee and exit
-`)
-	target := fs.Bool("target", false, "") // hidden: debuggee subprocess mode
-	funcName := fs.String("func", "", "function to debug")
-	argsFile := fs.String("args", "", "file containing the ABI0 argument block")
-	bufSpec := fs.String("buf", "", "buffer specification: name:size:pattern[,name:size:pattern...] where pattern is zero, ones, seq, or hex")
-	fs.Parse(args)
-
-	// --- Debuggee mode (internal, spawned by the debugger) ---
-	if *target {
-		tmpDir := os.Getenv("GASM_DEBUG_TMP")
-		if tmpDir == "" || fs.NArg() < 1 || *funcName == "" || *argsFile == "" {
-			fmt.Fprintln(os.Stderr, "gasm debug --target: internal mode")
-			return 2
-		}
-		if err := debug.RunTarget(fs.Arg(0), *funcName, *argsFile, tmpDir); err != nil {
-			fmt.Fprintf(os.Stderr, "gasm debug: %v\n", err)
-			return 1
-		}
-		return 0
-	}
-
-	// --- Debugger mode (interactive REPL) ---
-	if fs.NArg() < 1 || *funcName == "" {
-		fmt.Fprintln(os.Stderr, "usage: gasm debug <file.s> --func <name>")
-		return 2
-	}
-	path := fs.Arg(0)
-
-	// Load the kernel to extract function metadata and labels.
-	k, err := verify.Load(path)
-	if err != nil {
-		fmt.Fprintf(os.Stderr, "gasm debug: %v\n", err)
-		return 1
-	}
-	defer k.Close()
-
-	fl, err := k.Func(*funcName)
-	if err != nil {
-		fmt.Fprintf(os.Stderr, "gasm debug: %v\n", err)
-		return 1
-	}
-
-	// Build the label list for the REPL.
-	var labels []debug.Label
-	for name, off := range fl.Labels {
-		labels = append(labels, debug.Label{Name: name, Offset: off})
-	}
-	sort.Slice(labels, func(i, j int) bool { return labels[i].Offset < labels[j].Offset })
-
-	// Launch the debuggee with the argument block.
-	var argBlock []byte
-	var bufAddrs []uint64
-	var sess *debug.Session
-	if *bufSpec != "" {
-		// Parse the function signature to determine argument layout.
-		src, err := readSource(path)
-		if err != nil {
-			fmt.Fprintf(os.Stderr, "gasm debug: %v\n", err)
-			return 1
-		}
-		sig, ok := verify.ExtractFuncSig(src, *funcName)
-		if !ok {
-			fmt.Fprintf(os.Stderr, "gasm debug: no // func signature found for %s\n", *funcName)
-			return 1
-		}
-		layout := verify.ArgLayout(sig)
-
-		// Parse the buffer spec to get buffer names.
-		bufNames := parseBufNames(*bufSpec)
-
-		// Allocate buffers in the debuggee.
-		argBlock = make([]byte, fl.Args)
-		sess, bufAddrs, err = debug.LaunchWithBuffers("", path, *funcName, argBlock, *bufSpec)
-		if err != nil {
-			fmt.Fprintf(os.Stderr, "gasm debug: %v\n", err)
-			return 1
-		}
-
-		// Construct the argument block with buffer pointers at the correct positions.
-		bufIdx := 0
-		for _, arg := range layout {
-			if !arg.IsPtr {
-				continue
-			}
-			// Find the buffer that matches this argument.
-			for i, name := range bufNames {
-				if i < len(bufAddrs) && (name == arg.Name || strings.HasPrefix(arg.Name, name)) {
-					addr := bufAddrs[i]
-					off := arg.Offset
-					if off+8 <= len(argBlock) {
-						argBlock[off] = byte(addr)
-						argBlock[off+1] = byte(addr >> 8)
-						argBlock[off+2] = byte(addr >> 16)
-						argBlock[off+3] = byte(addr >> 24)
-						argBlock[off+4] = byte(addr >> 32)
-						argBlock[off+5] = byte(addr >> 40)
-						argBlock[off+6] = byte(addr >> 48)
-						argBlock[off+7] = byte(addr >> 56)
-					}
-					// For slices, also set the length and capacity.
-					if strings.HasPrefix(arg.Typ, "[]") && off+24 <= len(argBlock) {
-						// Find the buffer size from the spec.
-						size := parseBufSize(*bufSpec, name)
-						// Length at offset+8, capacity at offset+16.
-						for j := 0; j < 8; j++ {
-							argBlock[off+8+j] = byte(size >> (j * 8))
-							argBlock[off+16+j] = byte(size >> (j * 8))
-						}
-					}
-					bufIdx++
-					break
-				}
-			}
-		}
-		_ = bufIdx
-	} else {
-		argBlock = make([]byte, fl.Args)
-		sess, err = debug.Launch("", path, *funcName, argBlock)
-		if err != nil {
-			fmt.Fprintf(os.Stderr, "gasm debug: %v\n", err)
-			return 1
-		}
-	}
-	defer sess.Kill()
-
-	bm := debug.NewBreakpoints(sess)
-	fmt.Printf("gasm debug: %s in %s (pid %d)\n", *funcName, path, sess.Pid())
-
-	// Convert the line table for the REPL.
-	var srcLines []debug.SourceLine
-	for _, le := range fl.Lines {
-		srcLines = append(srcLines, debug.SourceLine{Offset: le.Offset, Line: le.Line})
-	}
-	debug.REPL(sess, bm, sess.CodeBase(), fl.Offset, fl.Size, fl.Args, labels, srcLines)
-	return 0
-}
-
-// parseBufNames extracts buffer names from a buffer specification.
-// Format: name:size:pattern[,name:size:pattern...]
-func parseBufNames(spec string) []string {
-	var names []string
-	for _, part := range strings.Split(spec, ",") {
-		fields := strings.SplitN(part, ":", 3)
-		if len(fields) >= 1 && fields[0] != "" {
-			names = append(names, fields[0])
-		}
-	}
-	return names
-}
-
-// parseBufSize extracts the size of a named buffer from a buffer specification.
-func parseBufSize(spec, name string) int {
-	for _, part := range strings.Split(spec, ",") {
-		fields := strings.SplitN(part, ":", 3)
-		if len(fields) >= 2 && fields[0] == name {
-			var size int
-			fmt.Sscanf(fields[1], "%d", &size)
-			return size
-		}
-	}
-	return 0
-}

cmd/gasm/debug_other.go

@@ -1,7 +1,7 @@
 // Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
 // SPDX-License-Identifier: BSD-3-Clause
 
-//go:build !(linux && amd64)
+//go:build !(linux || (freebsd && (amd64 || arm64 || riscv64)))
 
 package main
 
@@ -11,6 +11,6 @@ import (
 )
 
 func cmdDebug(args []string) int {
-	fmt.Fprintln(os.Stderr, "gasm debug: the interactive debugger requires linux/amd64 (ptrace)")
+	fmt.Fprintln(os.Stderr, "gasm debug: the interactive debugger requires Linux or FreeBSD (ptrace)")
 	return 1
 }

cmd/gasm/debug_ptrace.go

@@ -0,0 +1,336 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+//go:build linux || (freebsd && (amd64 || arm64 || riscv64))
+
+package main
+
+import (
+	"fmt"
+	"io"
+	"os"
+	"sort"
+	"strings"
+	"time"
+
+	"sourcedock.dev/petrbalvin/gasm-sdk/debug"
+	"sourcedock.dev/petrbalvin/gasm-sdk/verify"
+)
+
+func cmdDebug(args []string) int {
+	fs := newCommand("debug", "gasm debug <file.s> --func <name>", `
+Interactive debugger for JIT-assembled functions.  Launches the
+function in a traced subprocess (ptrace), then provides a REPL for
+single-stepping, breakpoints, register and memory inspection.
+
+REPL commands:
+  break <label|addr|line> [if <reg> <op> <val|reg|*addr>]
+                       set a breakpoint, optionally conditional on a
+                       comparison of one register against a constant,
+                       another register, or the 8-byte word at *addr
+  delete <label|addr>  remove a breakpoint
+  info break           list all breakpoints
+  step [n], s          single-step n instructions (default 1)
+  next, n              step over a CALL
+  finish, fin          run until the function returns
+  continue, c          run until a breakpoint, watchpoint or exit
+  disas [n], u         disassemble n instructions at PC
+  regs                 print general-purpose and vector registers
+  where                show source line and nearest label at PC
+  stack                show stack near RSP (return address + ABI0 args)
+  bt, backtrace        backtrace (current frame + return address)
+  x [addr] [len]       hex-dump memory (default: current PC, 64 bytes)
+  w <addr> <val...>    write bytes to memory
+  set <reg> <value>    set a register
+  watch <addr> [r|w] [size]
+                       set a hardware watchpoint (write by default)
+  unwatch [<slot>]     clear one watchpoint, or all without an argument
+  labels, l            list function labels and offsets
+  help, h, ?           show command help
+  quit, q              kill the debuggee and exit
+`)
+	funcName := fs.String("func", "", "function to debug")
+	argsFile := fs.String("args", "", "file containing the ABI0 argument block")
+	bufSpec := fs.String("buf", "", "buffer specification: name:size:pattern[,name:size:pattern...] where pattern is zero, ones, seq, or hex")
+	script := fs.String("script", "", "run REPL commands from a file (one per line) and exit; '-' reads stdin")
+	cover := fs.Bool("cover", false, "run to completion with a breakpoint on every instruction and report which executed and how often")
+	timeout := fs.Duration("timeout", 0, "kill the debuggee after this duration (e.g. 30s); for headless --script runs; a timeout exits 3")
+	fs.Parse(args)
+
+	// --- Debuggee mode (internal, spawned by the debugger) ---
+	if os.Getenv("GASM_DEBUG_TARGET") != "" {
+		tmpDir := os.Getenv("GASM_DEBUG_TMP")
+		if tmpDir == "" || fs.NArg() < 1 || *funcName == "" || *argsFile == "" {
+			fmt.Fprintln(os.Stderr, "gasm debug: internal debuggee mode")
+			return 2
+		}
+		if err := debug.RunTarget(fs.Arg(0), *funcName, *argsFile, tmpDir); err != nil {
+			fmt.Fprintf(os.Stderr, "gasm debug: %v\n", err)
+			return 1
+		}
+		return 0
+	}
+
+	// --- Debugger mode (interactive REPL) ---
+	if fs.NArg() < 1 || *funcName == "" {
+		fmt.Fprintln(os.Stderr, "usage: gasm debug <file.s> --func <name>")
+		return 2
+	}
+	path := fs.Arg(0)
+
+	// The watchdog is armed before anything can block: ptrace attach and a
+	// continued kernel loop both hang the run when the environment forbids
+	// tracing or the kernel loops forever, and neither is interruptible from
+	// the inside.
+	if *timeout > 0 {
+		go func() {
+			time.Sleep(*timeout)
+			fmt.Fprintf(os.Stderr, "gasm debug: timeout (%s), killing the debuggee\n", *timeout)
+			os.Exit(3)
+		}()
+	}
+
+	// Load the kernel to extract function metadata and labels.
+	k, err := verify.Load(path)
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "gasm debug: %v\n", err)
+		return 1
+	}
+	defer k.Close()
+
+	fl, err := k.Func(*funcName)
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "gasm debug: %v\n", err)
+		return 1
+	}
+
+	// Build the label list for the REPL.
+	var labels []debug.Label
+	for name, off := range fl.Labels {
+		labels = append(labels, debug.Label{Name: name, Offset: off})
+	}
+	sort.Slice(labels, func(i, j int) bool { return labels[i].Offset < labels[j].Offset })
+
+	// Launch the debuggee with the argument block.
+	var argBlock []byte
+	var bufAddrs []uint64
+	var sess *debug.Session
+	if *bufSpec != "" {
+		// Parse the function signature to determine argument layout.
+		src, err := readSource(path)
+		if err != nil {
+			fmt.Fprintf(os.Stderr, "gasm debug: %v\n", err)
+			return 1
+		}
+		sig, ok := verify.ExtractFuncSig(src, *funcName)
+		if !ok {
+			fmt.Fprintf(os.Stderr, "gasm debug: no // func signature found for %s\n", *funcName)
+			return 1
+		}
+		layout := verify.ArgLayout(sig)
+
+		// Parse the buffer spec to get buffer names.
+		bufNames := parseBufNames(*bufSpec)
+
+		// Allocate buffers in the debuggee.
+		argBlock = make([]byte, fl.Args)
+		sess, bufAddrs, err = debug.LaunchWithBuffers("", path, *funcName, argBlock, *bufSpec)
+		if err != nil {
+			fmt.Fprintf(os.Stderr, "gasm debug: %v\n", err)
+			return 1
+		}
+
+		// Construct the argument block with buffer pointers at the correct positions.
+		for _, arg := range layout {
+			if !arg.IsPtr {
+				continue
+			}
+			// Find the buffer that matches this argument.
+			for i, name := range bufNames {
+				if i < len(bufAddrs) && (name == arg.Name || strings.HasPrefix(arg.Name, name)) {
+					addr := bufAddrs[i]
+					off := arg.Offset
+					if off+8 <= len(argBlock) {
+						argBlock[off] = byte(addr)
+						argBlock[off+1] = byte(addr >> 8)
+						argBlock[off+2] = byte(addr >> 16)
+						argBlock[off+3] = byte(addr >> 24)
+						argBlock[off+4] = byte(addr >> 32)
+						argBlock[off+5] = byte(addr >> 40)
+						argBlock[off+6] = byte(addr >> 48)
+						argBlock[off+7] = byte(addr >> 56)
+					}
+					// For slices, also set the length and capacity.
+					if strings.HasPrefix(arg.Typ, "[]") && off+24 <= len(argBlock) {
+						// Find the buffer size from the spec.
+						size := parseBufSize(*bufSpec, name)
+						// Length at offset+8, capacity at offset+16.
+						for j := range 8 {
+							argBlock[off+8+j] = byte(size >> (j * 8))
+							argBlock[off+16+j] = byte(size >> (j * 8))
+						}
+					}
+					break
+				}
+			}
+		}
+	} else {
+		argBlock = make([]byte, fl.Args)
+		sess, err = debug.Launch("", path, *funcName, argBlock)
+		if err != nil {
+			fmt.Fprintf(os.Stderr, "gasm debug: %v\n", err)
+			return 1
+		}
+	}
+	defer sess.Kill()
+
+	bm := debug.NewBreakpoints(sess)
+	fmt.Printf("gasm debug: %s in %s (pid %d)\n", *funcName, path, sess.Pid())
+
+	// Convert the line table for the command loop.
+	var srcLines []debug.SourceLine
+	for _, le := range fl.Lines {
+		srcLines = append(srcLines, debug.SourceLine{Offset: le.Offset, Line: le.Line})
+	}
+
+	// Coverage mode: pre-register a breakpoint on every instruction (walked
+	// by length through the function body while the debuggee is stopped) and
+	// let the kernel run to completion.  Each trap counts a hit for that
+	// instruction, so the final report shows exactly which instructions
+	// executed and how often, with the label-level view derived from it.
+	// Expect the run to slow to ptrace speed: one trap per executed
+	// instruction.
+	if *cover {
+		base := sess.CodeBase() + uint64(fl.Offset)
+		type coverInstr struct {
+			off  uint64
+			text string
+		}
+		var instrs []coverInstr
+		for off := uint64(0); off < uint64(fl.Size); {
+			text, ln, err := sess.Disassemble(base + off)
+			if err != nil || ln == 0 {
+				break
+			}
+			instrs = append(instrs, coverInstr{off: off, text: text})
+			off += uint64(ln)
+		}
+		for _, in := range instrs {
+			if _, err := bm.SetWithCond(base+in.off, fmt.Sprintf("func+%#x", in.off), nil); err != nil {
+				fmt.Fprintf(os.Stderr, "gasm debug: cover: %v\n", err)
+				return 1
+			}
+		}
+		fmt.Printf("gasm debug: coverage run over %d instructions\n", len(instrs))
+		for {
+			for _, bp := range bm.All() {
+				bm.Reinsert(bp.Addr)
+			}
+			if err := sess.Continue(); err != nil {
+				break // debuggee finished or died
+			}
+			if sess.Exited() {
+				break
+			}
+			// A genuine signal-delivery-stop (a fault in the kernel): the
+			// run cannot make progress, because resuming would restart the
+			// faulting instruction and fault forever.  Report and stop.
+			if sig := sess.LastSignal(); sig != 0 {
+				fmt.Printf("gasm debug: cover: stopped on signal %v\n", sig)
+				break
+			}
+			regs, rerr := sess.GetRegs()
+			if rerr != nil {
+				break
+			}
+			// HandleTrap restores the original byte, rewinds PC and counts
+			// the hit on the breakpoint itself.  Single-step over the
+			// restored instruction so the reinsertion at the top of the
+			// loop cannot re-trap on the same breakpoint.
+			if bp := bm.HandleTrap(&regs); bp != nil {
+				if err := sess.Step(); err != nil {
+					break
+				}
+			}
+		}
+		hits := map[uint64]int{}
+		traps := 0
+		for _, bp := range bm.All() {
+			if n := bp.Hits(); n > 0 {
+				hits[bp.Addr-base] = n
+				traps += n
+			}
+		}
+		var hit []string
+		var missed []string
+		for _, l := range labels {
+			if hits[uint64(l.Offset)] > 0 {
+				hit = append(hit, l.Name)
+			} else {
+				missed = append(missed, l.Name)
+			}
+		}
+		sort.Strings(hit)
+		sort.Strings(missed)
+		fmt.Printf("coverage: %d/%d instructions executed (%d traps)\n", len(hits), len(instrs), traps)
+		fmt.Printf("coverage: %d/%d labels reached\n", len(hit), len(labels))
+		for _, l := range hit {
+			fmt.Printf("  covered   %s\n", l)
+		}
+		for _, l := range missed {
+			fmt.Printf("  MISSED    %s\n", l)
+		}
+		fmt.Println("executed instructions:")
+		for _, in := range instrs {
+			if n := hits[in.off]; n > 0 {
+				fmt.Printf("  func+%#04x  %4dx  %s\n", in.off, n, in.text)
+			}
+		}
+		return 0
+	}
+
+	// Headless mode: run the script through the normal command loop and
+	// exit.  The watchdog armed above covers launch, continue and step.
+	var in io.Reader = os.Stdin
+	if *script != "" {
+		if *script == "-" {
+			in = os.Stdin
+		} else {
+			f, err := os.Open(*script)
+			if err != nil {
+				fmt.Fprintf(os.Stderr, "gasm debug: %v\n", err)
+				return 1
+			}
+			defer f.Close()
+			in = f
+		}
+	}
+	debug.REPL(sess, bm, sess.CodeBase(), fl.Offset, fl.Size, fl.Args, labels, srcLines, in)
+	return 0
+}
+
+// parseBufNames extracts buffer names from a buffer specification.
+// Format: name:size:pattern[,name:size:pattern...]
+func parseBufNames(spec string) []string {
+	var names []string
+	for part := range strings.SplitSeq(spec, ",") {
+		fields := strings.SplitN(part, ":", 3)
+		if len(fields) >= 1 && fields[0] != "" {
+			names = append(names, fields[0])
+		}
+	}
+	return names
+}
+
+// parseBufSize extracts the size of a named buffer from a buffer specification.
+func parseBufSize(spec, name string) int {
+	for part := range strings.SplitSeq(spec, ",") {
+		fields := strings.SplitN(part, ":", 3)
+		if len(fields) >= 2 && fields[0] == name {
+			var size int
+			fmt.Sscanf(fields[1], "%d", &size)
+			return size
+		}
+	}
+	return 0
+}

cmd/gasm/dis.go

@@ -0,0 +1,148 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package main
+
+import (
+	"fmt"
+	"os"
+	"sort"
+	"strings"
+
+	"sourcedock.dev/petrbalvin/gasm-sdk/arch"
+	"sourcedock.dev/petrbalvin/gasm-sdk/disasm"
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
+)
+
+// cmdDis disassembles machine code: either a raw binary (standard input with
+// "-") whose architecture is given with -a, or a .s file, which is assembled
+// first so the listing shows the real function and label layout.
+func cmdDis(args []string) int {
+	fs := newCommand("dis", "gasm dis [-a arch] <file>", `
+Disassemble machine code to instruction text (via golang.org/x/arch).
+
+With a .s file, the file is assembled first and the listing follows the
+real layout: one block per TEXT function, local labels printed at their
+offsets.  The architecture comes from the file name suffix, or from -a.
+
+With any other file, or "-" for standard input, the bytes are disassembled
+linearly and -a selects the architecture (amd64, arm64, riscv64 or
+loong64).
+`)
+	archName := fs.String("a", "", "architecture for raw input: amd64, arm64, riscv64 or loong64")
+	fs.Parse(args)
+	if fs.NArg() != 1 {
+		fmt.Fprintln(os.Stderr, "usage: gasm dis [-a arch] <file>")
+		return 2
+	}
+	path := fs.Arg(0)
+	var target arch.Arch
+	if *archName != "" {
+		var err error
+		target, err = auditArch(*archName)
+		if err != nil {
+			fmt.Fprintf(os.Stderr, "gasm dis: %v\n", err)
+			return 2
+		}
+	}
+
+	if strings.HasSuffix(path, ".s") {
+		if target == arch.Unknown {
+			target = arch.FromFilename(path)
+		}
+		if target == arch.Unknown {
+			fmt.Fprintln(os.Stderr, "gasm dis: cannot infer the architecture from the file name; use -a")
+			return 2
+		}
+		return disSource(path, target)
+	}
+
+	if target == arch.Unknown {
+		fmt.Fprintln(os.Stderr, "gasm dis: raw input needs -a (amd64, arm64, riscv64 or loong64)")
+		return 2
+	}
+	src, err := readSource(path)
+	if err != nil {
+		fmt.Fprintln(os.Stderr, "gasm dis:", err)
+		return 1
+	}
+	printListing(target, []byte(src), 0, nil)
+	return 0
+}
+
+// disSource assembles a .s file and prints one listing block per function.
+func disSource(path string, target arch.Arch) int {
+	src, err := readSource(path)
+	if err != nil {
+		fmt.Fprintln(os.Stderr, "gasm dis:", err)
+		return 1
+	}
+	f, errs := parser.Parse(path, src)
+	for _, e := range errs {
+		fmt.Fprintf(os.Stderr, "%s: %v\n", path, e)
+	}
+	if len(errs) > 0 {
+		return 1
+	}
+	img, err := assembleFile(target, f, "")
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "gasm dis: %v\n", err)
+		return 1
+	}
+	if len(img.Funcs) == 0 {
+		fmt.Fprintln(os.Stderr, "gasm dis: no assemblable TEXT functions found")
+		return 1
+	}
+	for _, fn := range img.Funcs {
+		code := img.Code[fn.Offset : fn.Offset+fn.Size]
+		fmt.Printf("%s: %d bytes\n", fn.Name, fn.Size)
+		labels := make(map[int][]string, len(fn.Labels))
+		for name, off := range fn.Labels {
+			labels[off] = append(labels[off], name)
+		}
+		for off := range labels {
+			sort.Strings(labels[off])
+		}
+		printListing(target, code, uint64(fn.Offset), labels)
+	}
+	if len(img.Data) > 0 {
+		fmt.Printf("data: %d bytes at 0x%x\n", len(img.Data), len(img.Code))
+	}
+	return 0
+}
+
+// printListing decodes code linearly from offset base, printing label lines
+// (label name to offset within the block) as they are reached.
+func printListing(a arch.Arch, code []byte, base uint64, labels map[int][]string) {
+	pc := 0
+	for pc < len(code) {
+		for _, name := range labels[pc] {
+			fmt.Printf("%s:\n", name)
+		}
+		ins, err := disasm.Decode(a, code[pc:], base+uint64(pc))
+		if err != nil {
+			break
+		}
+		end := min(pc+ins.Len, len(code))
+		fmt.Printf("  %04x: %-16s %s\n", base+uint64(pc), hexBytes(code[pc:end]), ins.Text)
+		if ins.Len <= 0 {
+			break
+		}
+		pc += ins.Len
+	}
+}
+
+// hexBytes renders up to 8 bytes as contiguous hex.
+func hexBytes(b []byte) string {
+	var sb strings.Builder
+	for i, c := range b {
+		if i == 8 {
+			break
+		}
+		if i > 0 {
+			sb.WriteByte(' ')
+		}
+		fmt.Fprintf(&sb, "%02x", c)
+	}
+	return sb.String()
+}

cmd/gasm/main_test.go

@@ -7,9 +7,15 @@ import (
 	"bytes"
 	"io"
 	"os"
+	"os/exec"
 	"path/filepath"
+	"runtime"
 	"strings"
+	"syscall"
 	"testing"
+
+	"sourcedock.dev/petrbalvin/gasm-sdk/arch"
+	"sourcedock.dev/petrbalvin/gasm-sdk/asm"
 )
 
 const clean = "#include \"textflag.h\"\n" +
@@ -216,8 +222,9 @@ func TestCmdVersion(t *testing.T) {
 	if code != 0 {
 		t.Fatalf("code = %d", code)
 	}
-	if !strings.Contains(out, version) {
-		t.Errorf("version output %q does not mention %q", out, version)
+	got := version()
+	if !strings.Contains(out, got) {
+		t.Errorf("version output %q does not mention %q", out, got)
 	}
 }
 
@@ -237,3 +244,219 @@ func TestCmdArgErrors(t *testing.T) {
 		t.Errorf("cmdParse() code = %d, want 2", code)
 	}
 }
+
+// TestUsageExitCodes pins the exit-code contract for the commands whose main
+// dispatches on a returned error: a wrong argument set exits 2, the same as
+// the commands that count their arguments themselves, while a runtime
+// failure (an unreadable file) keeps exit 1.
+func TestUsageExitCodes(t *testing.T) {
+	for name, err := range map[string]error{
+		"audit-instructions extra argument": cmdAuditInstructions([]string{"amd64", "extra"}),
+		"audit-instructions unknown arch":   cmdAuditInstructions([]string{"mips"}),
+		"audit-instructions corpus extra":   cmdAuditInstructions([]string{"--corpus", "a", "b"}),
+		"scaffold no arguments":             cmdScaffold(nil),
+		"scaffold extra arguments":          cmdScaffold([]string{"differential", "a.s", "b.s"}),
+	} {
+		if err == nil {
+			t.Errorf("%s: expected an error", name)
+			continue
+		}
+		if code := exitCodeFor(err); code != 2 {
+			t.Errorf("%s: exit code = %d, want 2 (err: %v)", name, code, err)
+		}
+	}
+	if err := cmdScaffold([]string{"differential", "/nonexistent/file.s"}); err == nil {
+		t.Error("scaffold on a missing file should fail")
+	} else if code := exitCodeFor(err); code != 1 {
+		t.Errorf("scaffold on a missing file: exit code = %d, want 1", code)
+	}
+}
+
+// TestCmdAsmFormatValidation checks that an unknown --format exits 2 with
+// and without -o, instead of assembling and silently dumping a raw image.
+func TestCmdAsmFormatValidation(t *testing.T) {
+	path := writeTemp(t, "f_amd64.s", clean)
+	out := filepath.Join(t.TempDir(), "f.bin")
+	if _, _, code := capture(func() int { return cmdAsm([]string{"--format", "bogus", path}) }); code != 2 {
+		t.Errorf("asm --format bogus without -o: code = %d, want 2", code)
+	}
+	if _, _, code := capture(func() int { return cmdAsm([]string{"--format", "bogus", "-o", out, path}) }); code != 2 {
+		t.Errorf("asm --format bogus with -o: code = %d, want 2", code)
+	}
+}
+
+// TestCmdAsmOutputFile pins the documented -o behaviour: the output goes to
+// the file and stdout carries no hex dump; without -o the dump is the output.
+func TestCmdAsmOutputFile(t *testing.T) {
+	path := writeTemp(t, "f_amd64.s", clean)
+	out := filepath.Join(t.TempDir(), "f.bin")
+	stdout, _, code := capture(func() int { return cmdAsm([]string{"-o", out, path}) })
+	if code != 0 {
+		t.Fatalf("code = %d", code)
+	}
+	if strings.Contains(stdout, "0000:") {
+		t.Errorf("stdout carries a hex dump despite -o:\n%s", stdout)
+	}
+	if !strings.Contains(stdout, "wrote ") {
+		t.Errorf("stdout misses the wrote line:\n%s", stdout)
+	}
+	b, err := os.ReadFile(out)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(b) == 0 {
+		t.Error("the output file is empty")
+	}
+
+	stdout, _, code = capture(func() int { return cmdAsm([]string{path}) })
+	if code != 0 {
+		t.Fatalf("without -o: code = %d", code)
+	}
+	if !strings.Contains(stdout, "0000:") {
+		t.Errorf("without -o the hex dump is missing:\n%s", stdout)
+	}
+}
+
+// TestVerifyNonJITAMD64GroundTruth drives the cross-architecture
+// ground-truth path for an amd64 kernel: the path a host of any other
+// architecture takes, which must compare against the toolchain rather than
+// refuse to run.
+func TestVerifyNonJITAMD64GroundTruth(t *testing.T) {
+	if testing.Short() {
+		t.Skip("runs go tool asm")
+	}
+	path := writeTemp(t, "f_amd64.s", clean)
+	out, _, code := capture(func() int { return cmdVerifyNonJIT(path, arch.AMD64, true, false) })
+	if code != 0 {
+		t.Fatalf("code = %d (%s)", code, out)
+	}
+	if !strings.Contains(out, "1/1 matched") {
+		t.Errorf("output misses the matched report:\n%s", out)
+	}
+}
+
+// TestVerifySmokeCrashIsolation checks that a function faulting on its
+// zeroed smoke arguments is reported as CRASH by a child process instead of
+// killing `gasm verify` itself.
+func TestVerifySmokeCrashIsolation(t *testing.T) {
+	if testing.Short() {
+		t.Skip("builds the gasm binary")
+	}
+	if runtime.GOARCH != "amd64" {
+		t.Skip("amd64 JIT only")
+	}
+	bin := filepath.Join(t.TempDir(), "gasm")
+	if out, err := exec.Command("go", "build", "-o", bin, ".").CombinedOutput(); err != nil {
+		t.Fatalf("build gasm: %v\n%s", err, out)
+	}
+	src := filepath.Join(t.TempDir(), "crash_amd64.s")
+	kernel := "#include \"textflag.h\"\n" +
+		"\n" +
+		"// func Fault(x []byte) int\n" +
+		"TEXT ·Fault(SB), NOSPLIT, $0-32\n" +
+		"\tMOVQ\tx+0(FP), AX\n" +
+		"\tMOVQ\t(AX), AX // faults on the zeroed nil pointer\n" +
+		"\tMOVQ\tAX, ret+24(FP)\n" +
+		"\tRET\n"
+	if err := os.WriteFile(src, []byte(kernel), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	cmd := exec.Command(bin, "verify", "-smoke", src)
+	out, err := cmd.CombinedOutput()
+	if err == nil {
+		t.Fatalf("expected a failure report, got success:\n%s", out)
+	}
+	if exitErr, ok := err.(*exec.ExitError); ok {
+		if ws, ok := exitErr.Sys().(syscall.WaitStatus); ok && ws.Signaled() {
+			t.Fatalf("verify died from %v; the crash was not isolated:\n%s", ws.Signal(), out)
+		}
+	}
+	if !strings.Contains(string(out), "CRASH") {
+		t.Errorf("output does not report CRASH:\n%s", out)
+	}
+}
+
+func TestSweepCheckLines(t *testing.T) {
+	out := []byte("crash_amd64.s: 1 functions JIT-loaded\n" +
+		"  Fault: 21 bytes, args=32, frame=0 NOSPLIT\n" +
+		"    smoke: OK\n" +
+		"    abi: clean (10 varied inputs)\n")
+	want := "    smoke: OK\n    abi: clean (10 varied inputs)"
+	if got := sweepCheckLines(out); got != want {
+		t.Errorf("sweepCheckLines = %q, want %q", got, want)
+	}
+}
+
+// TestRunCorpusAudit drives the corpus audit over a small fixture tree: one
+// suffixed amd64 file, one suffixed arm64 file whose body is not arm64, one
+// generic file, and one file that does not parse.
+func TestRunCorpusAudit(t *testing.T) {
+	dir := t.TempDir()
+	write := func(name, src string) {
+		t.Helper()
+		if err := os.WriteFile(filepath.Join(dir, name), []byte(src), 0o644); err != nil {
+			t.Fatal(err)
+		}
+	}
+	write("good_amd64.s", "#include \"textflag.h\"\nTEXT ·add(SB), NOSPLIT, $0-0\n\tMOVQ AX, BX\n\tRET\n")
+	write("bad_arm64.s", "#include \"textflag.h\"\nTEXT ·f(SB), NOSPLIT, $0-0\n\tMOVQ AX, BX\n\tRET\n")
+	write("generic.s", "#include \"textflag.h\"\nTEXT ·g(SB), NOSPLIT, $0-0\n\tRET\n")
+	write("broken.s", "#include \"textflag.h\"\nTEXT ·b(SB), NOSPLIT, $0-0\n\tJMP nowhere\n\tRET\n")
+
+	stats, err := runCorpusAudit(dir, nil)
+	if err != nil {
+		t.Fatalf("runCorpusAudit: %v", err)
+	}
+	if stats.files != 4 {
+		t.Errorf("files = %d, want 4", stats.files)
+	}
+	if stats.generic != 2 {
+		t.Errorf("generic = %d, want 2 (generic.s and broken.s)", stats.generic)
+	}
+	// good_amd64 and generic.s assemble everywhere they are attempted.
+	if stats.full != 2 {
+		t.Errorf("full = %d, want 2", stats.full)
+	}
+	get := func(name string) *corpusTally {
+		for i, tg := range stats.targets {
+			if tg.name == name {
+				return stats.tallies[i]
+			}
+		}
+		t.Fatalf("no tally for %s", name)
+		return nil
+	}
+	// amd64: good_amd64 + generic.s + broken.s; the broken file fails to parse.
+	if a := get("amd64"); a.attempted != 3 || a.assembled != 2 {
+		t.Errorf("amd64 = %d/%d, want 2/3", a.assembled, a.attempted)
+	}
+	// arm64: bad_arm64 (MOVQ is not arm64) + generic.s + broken.s.
+	if a := get("arm64"); a.attempted != 3 || a.assembled != 1 {
+		t.Errorf("arm64 = %d/%d, want 1/3", a.assembled, a.attempted)
+	}
+	if r := get("amd64").reasons["instruction not encodable"]; r != 0 {
+		t.Errorf("amd64 unexpected unencodable reason: %d", r)
+	}
+	if r := get("arm64").reasons["instruction not encodable"]; r != 1 {
+		t.Errorf("arm64 unencodable reasons = %d, want 1", r)
+	}
+}
+
+// TestCompareGroundTruthPadding pins the padding-aware ground-truth
+// comparison: the toolchain pads text symbols to 16-byte boundaries, so
+// trailing zeros in the reference must not read as a mismatch, while any
+// non-zero tail still must.
+func TestCompareGroundTruthPadding(t *testing.T) {
+	code := []byte{0x48, 0x8b, 0x07, 0xc3} // 4 bytes, not a multiple of 16
+	img := &asm.Image{Code: code, Funcs: []asm.FuncLayout{{Name: "f", Offset: 0, Size: len(code)}}}
+	padded := append(append([]byte(nil), code...), 0, 0, 0)
+	matched, total, diffs := compareGroundTruth(img, map[string][]byte{"f": padded})
+	if matched != 1 || total != 1 || diffs != 0 {
+		t.Fatalf("zero padding should match: matched=%d total=%d diffs=%d", matched, total, diffs)
+	}
+	dirty := append(append([]byte(nil), code...), 0, 0x90, 0)
+	matched, _, diffs = compareGroundTruth(img, map[string][]byte{"f": dirty})
+	if matched != 0 || diffs != 1 {
+		t.Fatalf("non-zero padding must mismatch: matched=%d diffs=%d", matched, diffs)
+	}
+}

cmd/gasm/man_test.go

@@ -0,0 +1,241 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package main
+
+import (
+	"os"
+	"os/exec"
+	"path/filepath"
+	"regexp"
+	"strings"
+	"testing"
+)
+
+// TestManPagesTrackTheCLI builds the binary once, then compares every
+// command's live `-h` output with its docs/man/gasm-<command>.1 page: the
+// flag sets must agree both ways, and the page's SYNOPSIS line must carry
+// the command's usage line.  A flag or a usage change that skips the man
+// page fails here, so the pages cannot drift from the binary.
+func TestManPagesTrackTheCLI(t *testing.T) {
+	if testing.Short() {
+		t.Skip("builds the gasm binary")
+	}
+	bin := filepath.Join(t.TempDir(), "gasm")
+	if out, err := exec.Command("go", "build", "-o", bin, ".").CombinedOutput(); err != nil {
+		t.Fatalf("build gasm: %v\n%s", err, out)
+	}
+
+	for _, cmd := range []string{
+		"tokens", "parse", "fmt", "lint", "asm", "dis", "verify",
+		"debug", "diff", "profile", "audit-instructions", "scaffold", "lsp",
+	} {
+		t.Run(cmd, func(t *testing.T) {
+			raw, err := os.ReadFile(filepath.Join("..", "..", "docs", "man", "gasm-"+cmd+".1"))
+			if err != nil {
+				t.Fatalf("read man page: %v", err)
+			}
+			page := string(raw)
+
+			out, _ := exec.Command(bin, cmd, "-h").CombinedOutput()
+			help := string(out)
+
+			binFlags := helpFlags(help)
+			pageFlags := roffFlags(page)
+			for f := range binFlags {
+				if !pageFlags[f] {
+					t.Errorf("flag -%s is in the binary's help but missing from the man page", f)
+				}
+			}
+			for f := range pageFlags {
+				if !binFlags[f] {
+					t.Errorf("flag -%s is in the man page but the binary does not accept it", f)
+				}
+			}
+
+			want := helpUsage(help)
+			got := roffSynopsis(page)
+			if want != "" && got != want {
+				t.Errorf("SYNOPSIS drift:\n page: %s\nbinary: %s", got, want)
+			}
+		})
+	}
+}
+
+// TestManCommandsTrackHelp compares the gasm(1) COMMANDS list with the
+// top-level help output, so a subcommand added to the binary cannot miss
+// its man entry and a stale entry cannot outlive its command.
+func TestManCommandsTrackHelp(t *testing.T) {
+	if testing.Short() {
+		t.Skip("builds the gasm binary")
+	}
+	bin := filepath.Join(t.TempDir(), "gasm")
+	if out, err := exec.Command("go", "build", "-o", bin, ".").CombinedOutput(); err != nil {
+		t.Fatalf("build gasm: %v\n%s", err, out)
+	}
+
+	raw, err := os.ReadFile(filepath.Join("..", "..", "docs", "man", "gasm.1"))
+	if err != nil {
+		t.Fatalf("read man page: %v", err)
+	}
+
+	helpOut, err := exec.Command(bin, "--help").Output()
+	if err != nil {
+		t.Fatalf("gasm --help: %v", err)
+	}
+
+	binCmds := helpCommands(string(helpOut))
+	pageCmds := roffCommands(string(raw))
+	for c := range binCmds {
+		if !pageCmds[c] {
+			t.Errorf("command %q is in the binary's help but missing from gasm(1) COMMANDS", c)
+		}
+	}
+	for c := range pageCmds {
+		if !binCmds[c] {
+			t.Errorf("command %q is in gasm(1) COMMANDS but the binary does not list it", c)
+		}
+	}
+}
+
+// helpFlags extracts the flag names from a `gasm <cmd> -h` output.
+func helpFlags(help string) map[string]bool {
+	m := map[string]bool{}
+	inFlags := false
+	for line := range strings.SplitSeq(help, "\n") {
+		if strings.TrimRight(line, " \t") == "Flags:" {
+			inFlags = true
+			continue
+		}
+		if !inFlags {
+			continue
+		}
+		if !strings.HasPrefix(line, "  -") {
+			continue
+		}
+		token := strings.FieldsFunc(strings.TrimLeft(line, " "), func(r rune) bool {
+			return r == ' ' || r == '\t'
+		})
+		if len(token) == 0 {
+			continue
+		}
+		m[strings.TrimLeft(token[0], "-")] = true
+	}
+	return m
+}
+
+var roffEscape = regexp.MustCompile(`\\f[BIRP]`)
+
+// roffFlags extracts the flag names from a man page's OPTIONS section.
+func roffFlags(page string) map[string]bool {
+	m := map[string]bool{}
+	inOptions := false
+	for line := range strings.SplitSeq(page, "\n") {
+		if strings.HasPrefix(line, ".SH ") {
+			inOptions = strings.HasPrefix(line, ".SH OPTIONS")
+			continue
+		}
+		if !inOptions {
+			continue
+		}
+		// Flag entries are written as either `.B \-flag` or `\fB\-flag`.
+		var body string
+		switch {
+		case strings.HasPrefix(line, `.B \-`):
+			body = line[3:]
+		case strings.HasPrefix(line, `\fB\-`):
+			body = line[1:]
+		default:
+			continue
+		}
+		name := roffEscape.ReplaceAllString(body, "")
+		name = strings.ReplaceAll(name, `\-`, "-")
+		name = strings.TrimSpace(name)
+		if i := strings.IndexAny(name, " \t"); i >= 0 {
+			name = name[:i]
+		}
+		m[strings.TrimLeft(name, "-")] = true
+	}
+	return m
+}
+
+// helpCommands extracts the command names from the top-level help output's
+// Commands section.
+func helpCommands(help string) map[string]bool {
+	m := map[string]bool{}
+	inCmds := false
+	for line := range strings.SplitSeq(help, "\n") {
+		if strings.TrimSpace(line) == "Commands:" {
+			inCmds = true
+			continue
+		}
+		if !inCmds {
+			continue
+		}
+		t := strings.TrimSpace(line)
+		if t == "" {
+			break
+		}
+		name, _, _ := strings.Cut(t, " ")
+		m[name] = true
+	}
+	return m
+}
+
+// roffCommands extracts the command names from gasm(1)'s COMMANDS section,
+// where each entry is written as `.B gasm\-<name>(1)` or `.B gasm <name>`.
+func roffCommands(page string) map[string]bool {
+	m := map[string]bool{}
+	inCmds := false
+	for line := range strings.SplitSeq(page, "\n") {
+		if strings.HasPrefix(line, ".SH ") {
+			inCmds = strings.HasPrefix(line, ".SH COMMANDS")
+			continue
+		}
+		if !inCmds || !strings.HasPrefix(line, ".B gasm") {
+			continue
+		}
+		entry := strings.ReplaceAll(strings.TrimPrefix(line, ".B "), `\-`, "-")
+		entry = strings.TrimSuffix(entry, "(1)")
+		switch {
+		case strings.HasPrefix(entry, "gasm-"):
+			m[strings.TrimPrefix(entry, "gasm-")] = true
+		case strings.HasPrefix(entry, "gasm "):
+			m[strings.TrimPrefix(entry, "gasm ")] = true
+		}
+	}
+	return m
+}
+
+// helpUsage returns the command's usage line without the "Usage: " prefix.
+func helpUsage(help string) string {
+	for line := range strings.SplitSeq(help, "\n") {
+		if strings.HasPrefix(line, "Usage: ") {
+			return normaliseUsage(line[len("Usage: "):])
+		}
+	}
+	return ""
+}
+
+// roffSynopsis returns the page's SYNOPSIS usage line, unescaped.
+func roffSynopsis(page string) string {
+	inSyn := false
+	for line := range strings.SplitSeq(page, "\n") {
+		if strings.HasPrefix(line, ".SH ") {
+			inSyn = strings.HasPrefix(line, ".SH SYNOPSIS")
+			continue
+		}
+		if !inSyn || !strings.HasPrefix(line, ".B ") {
+			continue
+		}
+		return normaliseUsage(strings.ReplaceAll(line[3:], `\-`, "-"))
+	}
+	return ""
+}
+
+// normaliseUsage flattens whitespace and drops the roff font escapes so that
+// the binary's usage line and the page's SYNOPSIS line compare equal.
+func normaliseUsage(s string) string {
+	s = roffEscape.ReplaceAllString(s, "")
+	return strings.Join(strings.Fields(s), " ")
+}

cmd/gasm/preproc_test.go

@@ -0,0 +1,109 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package main
+
+import (
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+// writeTree writes a directory of files and returns its root.
+func writeTree(t *testing.T, files map[string]string) string {
+	t.Helper()
+	dir := t.TempDir()
+	for name, content := range files {
+		path := filepath.Join(dir, name)
+		if err := os.MkdirAll(filepath.Dir(path), 0o755); err != nil {
+			t.Fatal(err)
+		}
+		if err := os.WriteFile(path, []byte(content), 0o644); err != nil {
+			t.Fatal(err)
+		}
+	}
+	return dir
+}
+
+// TestAsmMacroAndIncludeEndToEnd drives `gasm asm` over a source with an
+// in-file parameterised macro and an include resolved through -I, and checks
+// the assembled bytes came from the expansion (the loop body counts six
+// increments, two per expanded iteration).
+func TestAsmMacroAndIncludeEndToEnd(t *testing.T) {
+	if testing.Short() {
+		t.Skip("runs the assembler end to end")
+	}
+	dir := writeTree(t, map[string]string{
+		"inc/consts.h": "#define NITER 3\n",
+		"main_amd64.s": "#include \"textflag.h\"\n" +
+			"#include \"consts.h\"\n" +
+			"#define STEP(r) ADDQ $1, r; ADDQ $1, r\n" +
+			"TEXT ·f(SB), NOSPLIT, $0-8\n" +
+			"\tXORQ AX, AX\n" +
+			"\tMOVQ $NITER, CX\n" +
+			"loop:\n" +
+			"\tSTEP(AX)\n" +
+			"\tDECQ CX\n" +
+			"\tJNZ loop\n" +
+			"\tMOVQ AX, ret+0(FP)\n" +
+			"\tRET\n",
+	})
+	stdout, stderr, code := capture(func() int {
+		return cmdAsm([]string{"-I", filepath.Join(dir, "inc"), "-GOARCH", "amd64", filepath.Join(dir, "main_amd64.s")})
+	})
+	if code != 0 {
+		t.Fatalf("gasm asm exited %d: %s%s", code, stdout, stderr)
+	}
+	// The macro expanded to two ADDQ $1 encodings in the static body; the
+	// iteration count lives in the runtime loop.
+	if n := strings.Count(stdout, "83 c0 01"); n != 2 {
+		t.Errorf("found %d ADDQ $1 encodings in the image, want 2:\n%s", n, stdout)
+	}
+}
+
+// TestAsmIncludeResolutionOrder pins the -I search order end to end: the
+// including file's directory wins over the -I directories.
+func TestAsmIncludeResolutionOrder(t *testing.T) {
+	if testing.Short() {
+		t.Skip("runs the assembler end to end")
+	}
+	dir := writeTree(t, map[string]string{
+		"src/main_amd64.s": "#include \"textflag.h\"\n" +
+			"#include \"vals.h\"\n" +
+			"TEXT ·f(SB), NOSPLIT, $0\n" +
+			"\tMOVQ $VAL, AX\n" +
+			"\tRET\n",
+		"src/vals.h":   "#define VAL 1\n",
+		"late/vals.h":  "#define VAL 2\n",
+		"early/vals.h": "#define VAL 3\n",
+	})
+	stdout, stderr, code := capture(func() int {
+		return cmdAsm([]string{"-I", filepath.Join(dir, "early"), "-I", filepath.Join(dir, "late"),
+			"-GOARCH", "amd64", filepath.Join(dir, "src", "main_amd64.s")})
+	})
+	if code != 0 {
+		t.Fatalf("gasm asm exited %d: %s%s", code, stdout, stderr)
+	}
+	// VAL came from src/vals.h, not from either -I directory: the image
+	// loads the immediate 1.
+	if !strings.Contains(stdout, "b8 01 00 00 00") {
+		t.Errorf("expected the source-directory VAL (immediate 1) in:\n%s", stdout)
+	}
+}
+
+// TestAsmMissingIncludeIsAnError pins the diagnostic for an include that
+// resolves nowhere on the assembly path.
+func TestAsmMissingIncludeIsAnError(t *testing.T) {
+	if testing.Short() {
+		t.Skip("runs the assembler end to end")
+	}
+	path := writeTemp(t, "main_amd64.s", "#include \"textflag.h\"\n#include \"nothere.h\"\nTEXT ·f(SB), NOSPLIT, $0\n\tRET\n")
+	_, stderr, code := capture(func() int { return cmdAsm([]string{"-GOARCH", "amd64", path}) })
+	if code == 0 {
+		t.Fatal("gasm asm accepted a file whose include resolves nowhere")
+	}
+	if !strings.Contains(stderr, `#include "nothere.h"`) {
+		t.Errorf("stderr does not name the failing include: %s", stderr)
+	}
+}

cmd/gasm/scaffold.go

@@ -0,0 +1,345 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package main
+
+import (
+	"fmt"
+	"go/ast"
+	"go/parser"
+	"go/token"
+	"os"
+	"strings"
+
+	gasmast "sourcedock.dev/petrbalvin/gasm-sdk/ast"
+	gasmparser "sourcedock.dev/petrbalvin/gasm-sdk/parser"
+)
+
+// cmdScaffold generates a differential test skeleton for every kernel in a
+// file: a Go test that seeds random states, drives both the assembly kernel
+// and a caller-provided portable reference, and compares the outputs
+// byte-for-byte.  The lesson this encodes: a pipeline-level fuzz cannot see
+// an unwired kernel, only a direct-call differential against the portable
+// specification can, so every kernel ships with one.
+//
+// The generated file follows two conventions the caller fills in:
+//   - the assembly symbols resolve because the test lives in the kernel's
+//     own package (the //go:noescape declarations reference them);
+//   - each kernel gets a <name>Portable Go function the author implements as
+//     the specification, and the test fails on the first divergent byte.
+func cmdScaffold(args []string) error {
+	fs := newCommand("scaffold", "gasm scaffold differential <file.s>", `
+Print a differential test skeleton for every // func signature in FILE.
+The test seeds random states, drives the kernel and a portable reference
+(<name>Portable), and compares outputs byte-for-byte.  Write the reference
+bodies, place the file in the kernel's package, and run it in CI.
+`)
+	if err := fs.Parse(args); err != nil {
+		return err
+	}
+	rest := fs.Args()
+	// The first positional word is the scaffold style; "differential" is the
+	// only one today.
+	if len(rest) > 0 && rest[0] == "differential" {
+		rest = rest[1:]
+	}
+	if len(rest) != 1 {
+		return &usageError{fmt.Errorf("usage: gasm scaffold differential <file.s>")}
+	}
+	path := rest[0]
+	src, err := os.ReadFile(path)
+	if err != nil {
+		return err
+	}
+	f, errs := gasmparser.Parse(path, string(src))
+	if len(errs) > 0 {
+		return fmt.Errorf("parse: %v", errs[0])
+	}
+
+	var out strings.Builder
+	out.WriteString(headerComment)
+	out.WriteString("package " + packageName + "\n\n")
+	out.WriteString("import (\n\t\"bytes\"\n\t\"math/rand\"\n\t\"testing\"\n)\n\n")
+	out.WriteString(generatedHelpers)
+
+	kernels := 0
+	for _, d := range f.Decls {
+		txt, ok := d.(*gasmast.Text)
+		if !ok {
+			continue
+		}
+		params, results, ok := parseSig(txt.Doc)
+		if !ok || len(params) == 0 {
+			continue
+		}
+		kernels++
+		name := txt.Name.Name
+		fmt.Fprintf(&out, "// %sPortable is the specification %s is pinned against:\n", name, name)
+		fmt.Fprintf(&out, "// fill in a straightforward implementation of the same contract.\n")
+		fmt.Fprintf(&out, "func %sPortable(%s) (%s) {\n\tpanic(\"implement the portable specification\")\n}\n\n", name, paramDecl(params), resultDecl(results))
+
+		fmt.Fprintf(&out, "func Test%sDifferential(t *testing.T) {\n", strings.ToUpper(name[:1])+name[1:])
+		fmt.Fprintf(&out, "\trng := rand.New(rand.NewSource(1))\n")
+		fmt.Fprintf(&out, "\tfor range 1000 {\n")
+		// Seed two independent argument sets per iteration: the kernel runs
+		// on set A, the portable reference on set B, so in-place writes
+		// through pointer/slice arguments cannot contaminate the other side.
+		var sliceNames []string
+		seen := map[string]bool{}
+		aArgs := make([]string, 0, len(params))
+		bArgs := make([]string, 0, len(params))
+		for _, p := range params {
+			a, b, slices := genParamSeed(&out, p, seen)
+			aArgs = append(aArgs, a)
+			bArgs = append(bArgs, b)
+			sliceNames = append(sliceNames, slices...)
+		}
+		fmt.Fprintf(&out, "\t\tgot := %s(%s)\n", name, strings.Join(aArgs, ", "))
+		fmt.Fprintf(&out, "\t\twant := %sPortable(%s)\n", name, strings.Join(bArgs, ", "))
+		fmt.Fprintf(&out, "\t\tif !bytes.Equal(outputBytes(got), outputBytes(want)) {\n")
+		fmt.Fprintf(&out, "\t\t\tt.Fatalf(\"kernel diverges from the portable spec (seed 1, deterministic)\")\n")
+		fmt.Fprintf(&out, "\t\t}\n")
+		for _, s := range sliceNames {
+			fmt.Fprintf(&out, "\t\tif !bytes.Equal(outputBytes(%sA), outputBytes(%sB)) {\n", s, s)
+			fmt.Fprintf(&out, "\t\t\tt.Fatalf(\"kernel mutated %%q differently (seed 1, deterministic)\", %q)\n", s)
+			fmt.Fprintf(&out, "\t\t}\n")
+		}
+		fmt.Fprintf(&out, "\t}\n}\n\n")
+	}
+	if kernels == 0 {
+		return fmt.Errorf("%s: no // func signatures found; add one doc comment per kernel", path)
+	}
+	os.Stdout.WriteString(out.String())
+	return nil
+}
+
+const packageName = "yourpkg"
+
+const headerComment = `// Code generated by gasm scaffold differential; EDIT THE PANICS.
+// Each Test*Differential drives the assembly kernel and its portable
+// reference over the same random states and compares the outputs.
+// Place this file in the kernel's own package so the symbols resolve.
+
+`
+
+// sigParam is one parsed // func parameter.
+type sigParam struct {
+	Names []string
+	Type  string
+}
+
+type sigResult struct {
+	Names []string
+	Type  string
+}
+
+// parseSig parses the // func signature of a doc comment.
+func parseSig(doc string) ([]sigParam, []sigResult, bool) {
+	var line string
+	for l := range strings.SplitSeq(doc, "\n") {
+		if t := strings.TrimSpace(l); strings.HasPrefix(t, "func ") {
+			line = t
+			break
+		}
+	}
+	if line == "" {
+		return nil, nil, false
+	}
+	fset := token.NewFileSet()
+	f, err := parser.ParseFile(fset, "sig.go", "package p\n"+line+" {}\n", 0)
+	if err != nil {
+		return nil, nil, false
+	}
+	fd, ok := f.Decls[0].(*ast.FuncDecl)
+	if !ok || fd.Type == nil {
+		return nil, nil, false
+	}
+	var params []sigParam
+	for _, field := range fd.Type.Params.List {
+		typ := exprString(field.Type)
+		if len(field.Names) == 0 {
+			params = append(params, sigParam{Names: []string{""}, Type: typ})
+			continue
+		}
+		// Shared names (`L, result *byte`) expand to one entry per name:
+		// every name is a separate argument at the call site.
+		for _, n := range field.Names {
+			params = append(params, sigParam{Names: []string{n.Name}, Type: typ})
+		}
+	}
+	var results []sigResult
+	if fd.Type.Results != nil {
+		for _, field := range fd.Type.Results.List {
+			results = append(results, sigResult{Names: identNames(field.Names), Type: exprString(field.Type)})
+		}
+	}
+	return params, results, true
+}
+
+func identNames(idents []*ast.Ident) []string {
+	var out []string
+	for _, id := range idents {
+		out = append(out, id.Name)
+	}
+	return out
+}
+
+func exprString(e ast.Expr) string {
+	switch t := e.(type) {
+	case *ast.Ident:
+		return t.Name
+	case *ast.StarExpr:
+		return "*" + exprString(t.X)
+	case *ast.SelectorExpr:
+		return exprString(t.X) + "." + t.Sel.Name
+	case *ast.ArrayType:
+		if t.Len == nil {
+			return "[]" + exprString(t.Elt)
+		}
+		return "[N]" + exprString(t.Elt)
+	}
+	return "interface{}"
+}
+
+// paramDecl renders a parameter list for the portable reference signature.
+func paramDecl(params []sigParam) string {
+	var parts []string
+	for _, p := range params {
+		if len(p.Names) == 0 {
+			parts = append(parts, p.Type)
+			continue
+		}
+		for _, n := range p.Names {
+			parts = append(parts, n+" "+p.Type)
+		}
+	}
+	return strings.Join(parts, ", ")
+}
+
+// resultDecl renders a result list; unnamed results keep bare types.
+func resultDecl(results []sigResult) string {
+	if len(results) == 0 {
+		return ""
+	}
+	var parts []string
+	for _, r := range results {
+		parts = append(parts, r.Type)
+	}
+	return strings.Join(parts, ", ")
+}
+
+// genParamSeed emits the seeding statements for one parameter and returns
+// the kernel-side (A) and reference-side (B) argument expressions, plus the
+// names of any slice variables written in place (compared after the calls).
+func genParamSeed(out *strings.Builder, p sigParam, seen map[string]bool) (aArg, bArg string, slices []string) {
+	name := p.Names[0]
+	elem := strings.TrimPrefix(p.Type, "*")
+	isSlice := strings.HasPrefix(p.Type, "[]")
+	if isSlice {
+		elem = strings.TrimPrefix(p.Type, "[]")
+	}
+	switch {
+	case isSlice:
+		v := uniqueName(seen, name)
+		fmt.Fprintf(out, "\t\t%sA := make([]%s, 1+rng.Intn(512))\n", v, elem)
+		fmt.Fprintf(out, "\t\t%sB := make([]%s, len(%sA))\n", v, elem, v)
+		fmt.Fprintf(out, "\t\tfor i := range %sA {\n", v)
+		fmt.Fprintf(out, "\t\t\tw%s := %s(rng.Intn(256))\n", v, goCast(elem))
+		fmt.Fprintf(out, "\t\t\t%sA[i] = w%s\n", v, v)
+		fmt.Fprintf(out, "\t\t\t%sB[i] = w%s\n", v, v)
+		fmt.Fprintf(out, "\t\t}\n")
+		return v, v, []string{v}
+	case strings.HasPrefix(p.Type, "*"):
+		v := uniqueName(seen, name)
+		fmt.Fprintf(out, "\t\tvar %sA, %sB %s\n", v, v, elem)
+		fmt.Fprintf(out, "\t\tw%s := %s(rng.Intn(256))\n", v, goCast(elem))
+		fmt.Fprintf(out, "\t\t%sA = w%s\n", v, v)
+		fmt.Fprintf(out, "\t\t%sB = w%s\n", v, v)
+		return "&" + v + "A", "&" + v + "B", nil
+	default:
+		v := uniqueName(seen, name)
+		fmt.Fprintf(out, "\t\tw%s := %s(rng.Intn(512))\n", v, goCast(""))
+		fmt.Fprintf(out, "\t\tvar %sA, %sB %s = w%s, w%s\n", v, v, p.Type, v, v)
+		return v + "A", v + "B", nil
+	}
+}
+
+// uniqueName de-duplicates seeded variable names when one kernel takes two
+// parameters of the same name (impossible in Go) or a name repeats across
+// kernels in one file.
+func uniqueName(seen map[string]bool, base string) string {
+	if base == "" {
+		base = "arg"
+	}
+	if !seen[base] {
+		seen[base] = true
+		return base
+	}
+	for i := 2; ; i++ {
+		cand := fmt.Sprintf("%s%d", base, i)
+		if !seen[cand] {
+			seen[cand] = true
+			return cand
+		}
+	}
+}
+
+// goCast returns the conversion turning rng.Intn into the element type.
+func goCast(elem string) string {
+	switch elem {
+	case "byte", "uint8":
+		return "byte"
+	case "int8":
+		return "int8"
+	case "uint16":
+		return "uint16"
+	case "int16":
+		return "int16"
+	case "uint32":
+		return "uint32"
+	case "int32":
+		return "int32"
+	case "uint64":
+		return "uint64"
+	default:
+		return "int"
+	}
+}
+
+// generatedHelpers is emitted into every generated test file: outputBytes
+// narrows returned slices and scalars to a byte form for the comparison.
+// It lives in the template, not in this binary, because only the generated
+// file ever calls it.
+const generatedHelpers = `// outputBytes narrows a returned slice or scalar to bytes for the
+// comparison; extend the switch when a kernel returns a wider type.
+func outputBytes(v any) []byte {
+	switch t := v.(type) {
+	case []byte:
+		return t
+	case []int32:
+		b := make([]byte, 4*len(t))
+		for i, x := range t {
+			b[i*4] = byte(x)
+			b[i*4+1] = byte(x >> 8)
+			b[i*4+2] = byte(x >> 16)
+			b[i*4+3] = byte(x >> 24)
+		}
+		return b
+	case []uint16:
+		b := make([]byte, 2*len(t))
+		for i, x := range t {
+			b[i*2] = byte(x)
+			b[i*2+1] = byte(x >> 8)
+		}
+		return b
+	case int:
+		b := make([]byte, 8)
+		for i := range 8 {
+			b[i] = byte(uint64(t) >> (8 * i))
+		}
+		return b
+	default:
+		return nil
+	}
+}
+`

cmd/gasm/unidiff.go

@@ -0,0 +1,140 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package main
+
+import (
+	"fmt"
+	"slices"
+	"strings"
+)
+
+// unifiedDiff renders a unified diff with three lines of context between the
+// two line slices, in the form `gofmt -d` prints.  An empty result means the
+// inputs are identical.
+func unifiedDiff(name string, a, b []string) string {
+	if slices.Equal(a, b) {
+		return ""
+	}
+	var out strings.Builder
+	fmt.Fprintf(&out, "--- %s\n+++ %s\n", name, name)
+
+	// Longest common subsequence over the lines (assembly files are small
+	// enough for the quadratic table).
+	n, m := len(a), len(b)
+	lcs := make([][]int, n+1)
+	for i := range lcs {
+		lcs[i] = make([]int, m+1)
+	}
+	for i := n - 1; i >= 0; i-- {
+		for j := m - 1; j >= 0; j-- {
+			if a[i] == b[j] {
+				lcs[i][j] = lcs[i+1][j+1] + 1
+			} else if lcs[i+1][j] >= lcs[i][j+1] {
+				lcs[i][j] = lcs[i+1][j]
+			} else {
+				lcs[i][j] = lcs[i][j+1]
+			}
+		}
+	}
+
+	// Walk the LCS once, assigning every op its absolute position in both
+	// files (1-based, the position an insertion sits before).
+	type op struct {
+		kind         byte // ' ', '-' or '+'
+		aLine, bLine int
+		text         string
+	}
+	var ops []op
+	aPos, bPos := 0, 0
+	emit := func(kind byte, text string) {
+		ops = append(ops, op{kind: kind, aLine: aPos + 1, bLine: bPos + 1, text: text})
+		switch kind {
+		case ' ':
+			aPos++
+			bPos++
+		case '-':
+			aPos++
+		case '+':
+			bPos++
+		}
+	}
+	i, j := 0, 0
+	for i < n && j < m {
+		switch {
+		case a[i] == b[j]:
+			emit(' ', a[i])
+			i++
+			j++
+		case lcs[i+1][j] >= lcs[i][j+1]:
+			emit('-', a[i])
+			i++
+		default:
+			emit('+', b[j])
+			j++
+		}
+	}
+	for ; i < n; i++ {
+		emit('-', a[i])
+	}
+	for ; j < m; j++ {
+		emit('+', b[j])
+	}
+
+	// Group the edits into hunks: consecutive changes separated by more than
+	// twice the context lines start a new hunk.
+	const context = 3
+	var changes []int
+	for k, o := range ops {
+		if o.kind != ' ' {
+			changes = append(changes, k)
+		}
+	}
+	for g := 0; g < len(changes); {
+		last := g
+		for last+1 < len(changes) && changes[last+1]-changes[last]-1 <= 2*context {
+			last++
+		}
+		lo := max(0, changes[g]-context)
+		hi := min(len(ops), changes[last]+1+context)
+		// The header numbers are the first line of each side actually shown:
+		// the first context, deletion or insertion line.  A hunk that shows
+		// no old lines is a pure insertion and reports the position it sits
+		// before (0 at the top of the file); the mirror rule holds for a
+		// pure deletion.
+		aStart := ops[lo].aLine - 1
+		bStart := ops[lo].bLine - 1
+		countA, countB := 0, 0
+		for _, o := range ops[lo:hi] {
+			switch o.kind {
+			case ' ':
+				countA++
+				countB++
+			case '-':
+				countA++
+			case '+':
+				countB++
+			}
+		}
+		for _, o := range ops[lo:hi] {
+			if o.kind != '+' {
+				aStart = o.aLine
+				break
+			}
+		}
+		for _, o := range ops[lo:hi] {
+			if o.kind != '-' {
+				bStart = o.bLine
+				break
+			}
+		}
+		fmt.Fprintf(&out, "@@ -%d,%d +%d,%d @@\n", aStart, countA, bStart, countB)
+		for _, o := range ops[lo:hi] {
+			out.WriteByte(o.kind)
+			out.WriteString(o.text)
+			out.WriteByte('\n')
+		}
+		g = last + 1
+	}
+	return out.String()
+}

cmd/gasm/unidiff_test.go

@@ -0,0 +1,94 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package main
+
+import (
+	"slices"
+	"strings"
+	"testing"
+)
+
+func lines(ss ...string) []string { return ss }
+
+func TestUnifiedDiffIdentical(t *testing.T) {
+	if got := unifiedDiff("f", lines("a", "b"), lines("a", "b")); got != "" {
+		t.Errorf("identical inputs produced %q, want empty", got)
+	}
+}
+
+func TestUnifiedDiffSingleChange(t *testing.T) {
+	a := lines("1", "2", "3", "4", "5", "6", "7", "8")
+	b := lines("1", "2", "3!", "4", "5", "6", "7", "8")
+	want := "--- f\n+++ f\n" +
+		"@@ -1,6 +1,6 @@\n" +
+		" 1\n 2\n-3\n+3!\n 4\n 5\n 6\n"
+	if got := unifiedDiff("f", a, b); got != want {
+		t.Errorf("diff = %q, want %q", got, want)
+	}
+}
+
+func TestUnifiedDiffInsertAtStart(t *testing.T) {
+	got := unifiedDiff("f", lines("x"), lines("new", "x"))
+	// The single existing line is shown as trailing context, so the hunk
+	// covers it.
+	want := "--- f\n+++ f\n@@ -1,1 +1,2 @@\n+new\n x\n"
+	if got != want {
+		t.Errorf("diff = %q, want %q", got, want)
+	}
+}
+
+func TestUnifiedDiffDeleteAtEnd(t *testing.T) {
+	got := unifiedDiff("f", lines("x", "y"), lines("x"))
+	want := "--- f\n+++ f\n@@ -1,2 +1,1 @@\n x\n-y\n"
+	if got != want {
+		t.Errorf("diff = %q, want %q", got, want)
+	}
+}
+
+func TestUnifiedDiffTwoHunks(t *testing.T) {
+	var a, b []string
+	for i := 1; i <= 20; i++ {
+		a = append(a, itoa(i))
+		b = append(b, itoa(i))
+	}
+	b[1] = "2!"
+	b[17] = "18!"
+	got := unifiedDiff("f", a, b)
+	if !strings.Contains(got, "@@ -1,5 +1,5 @@\n 1\n-2\n+2!\n 3\n 4\n 5\n") {
+		t.Errorf("first hunk wrong:\n%s", got)
+	}
+	if !strings.Contains(got, "@@ -15,6 +15,6 @@\n 15\n 16\n 17\n-18\n+18!\n 19\n 20\n") {
+		t.Errorf("second hunk wrong:\n%s", got)
+	}
+}
+
+// TestUnifiedDiffAdjacentHunks merges changes separated by exactly twice the
+// context into one hunk.
+func TestUnifiedDiffAdjacentHunks(t *testing.T) {
+	a := lines("1", "2", "3", "4", "5", "6", "7", "8")
+	b := slices.Clone(a)
+	b[0] = "1!"
+	b[7] = "8!"
+	got := unifiedDiff("f", a, b)
+	want := "--- f\n+++ f\n" +
+		"@@ -1,8 +1,8 @@\n" +
+		"-1\n+1!\n 2\n 3\n 4\n 5\n 6\n 7\n-8\n+8!\n"
+	if got != want {
+		t.Errorf("diff = %q, want %q", got, want)
+	}
+}
+
+func itoa(n int) string {
+	if n == 0 {
+		return "0"
+	}
+	var buf [4]byte
+	i := len(buf)
+	for n > 0 {
+		i--
+		buf[i] = byte('0' + n%10)
+		n /= 10
+	}
+	return string(buf[i:])
+}

debug/breakpoint.go

@@ -1,89 +1,101 @@
 // Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
 // SPDX-License-Identifier: BSD-3-Clause
 
+//go:build linux || (freebsd && (amd64 || arm64 || riscv64))
+
 package debug
 
+import "strings"
+
 import "fmt"
 
-// Breakpoint is one INT3 breakpoint in the debuggee.
+// Breakpoint is one software breakpoint in the debuggee.
 type Breakpoint struct {
 	Addr    uint64 // absolute address in the debuggee
 	Label   string // source label ("" for raw addresses)
-	Orig    byte   // original byte at Addr (restored on removal)
+	Orig    []byte // original bytes at Addr (restored on removal)
 	Enabled bool
 	Cond    *Condition // optional condition (nil = unconditional)
 	hits    int
 }
 
-// Condition is a simple register-comparison condition evaluated when a
-// breakpoint is hit.  Format: <reg> <op> <value>.
+// Condition is a register-comparison condition evaluated when a breakpoint
+// is hit.  Supports three forms:
+//   - register vs constant:  <reg> <op> <value>
+//   - register vs register:  <reg> <op> <reg2>
+//   - register vs memory:    <reg> <op> *<addr>
 type Condition struct {
-	Reg   string // register name (rax, rbx, rip, rsp, ...)
-	Op    string // comparison operator: ==, !=, <, >, <=, >=
-	Value uint64
+	Reg     string // register name (rax, rbx, rip, rsp, ...)
+	Op      string // comparison operator: ==, !=, <, >, <=, >=
+	Value   uint64 // constant value (when Reg2 == "" and MemAddr == 0)
+	Reg2    string // second register name (for register-register comparison)
+	MemAddr uint64 // memory address (for register-memory comparison, prefixed with *)
 }
 
-// Eval checks the condition against the current registers.
-func (c *Condition) Eval(regs *Regs) bool {
-	var actual uint64
-	switch c.Reg {
-	case "rax", "eax", "ax", "al":
-		actual = regs.RAX
-	case "rbx", "ebx", "bx", "bl":
-		actual = regs.RBX
-	case "rcx", "ecx", "cx", "cl":
-		actual = regs.RCX
-	case "rdx", "edx", "dx", "dl":
-		actual = regs.RDX
-	case "rsi", "esi", "si":
-		actual = regs.RSI
-	case "rdi", "edi", "di":
-		actual = regs.RDI
-	case "rbp", "ebp", "bp":
-		actual = regs.RBP
-	case "rsp", "esp", "sp":
-		actual = regs.RSP
-	case "r8":
-		actual = regs.R8
-	case "r9":
-		actual = regs.R9
-	case "r10":
-		actual = regs.R10
-	case "r11":
-		actual = regs.R11
-	case "r12":
-		actual = regs.R12
-	case "r13":
-		actual = regs.R13
-	case "r14":
-		actual = regs.R14
-	case "r15":
-		actual = regs.R15
-	case "rip", "eip":
-		actual = regs.RIP
+// Eval checks the condition against the current registers.  For the
+// register-memory form, mem reads an 8-byte little-endian word from the
+// debuggee; it may be nil when no reader is available.  Anything that cannot
+// be decided (unknown register or operator, unreadable memory) does not
+// block the breakpoint.
+func (c *Condition) Eval(regs *Regs, mem func(addr uint64) (uint64, bool)) bool {
+	actual, ok := regs.RegValue(c.Reg)
+	if !ok {
+		return true // unknown register, don't block
+	}
+	var expected uint64
+	switch {
+	case c.Reg2 != "":
+		// Register-register comparison.
+		v, ok := regs.RegValue(c.Reg2)
+		if !ok {
+			return true
+		}
+		expected = v
+	case c.MemAddr != 0:
+		// Register-memory comparison, resolved in the debuggee at
+		// evaluation time.
+		if mem == nil {
+			return true
+		}
+		v, ok := mem(c.MemAddr)
+		if !ok {
+			return true
+		}
+		expected = v
 	default:
-		return true // unknown register — don't block
+		expected = c.Value
 	}
 	switch c.Op {
 	case "==", "=":
-		return actual == c.Value
+		return actual == expected
 	case "!=":
-		return actual != c.Value
+		return actual != expected
 	case "<":
-		return actual < c.Value
+		return actual < expected
 	case ">":
-		return actual > c.Value
+		return actual > expected
 	case "<=":
-		return actual <= c.Value
+		return actual <= expected
 	case ">=":
-		return actual >= c.Value
+		return actual >= expected
 	default:
 		return true
 	}
 }
 
-// Breakpoints manages the set of breakpoints for a Session.
-// Breakpoints manages software breakpoints for a debuggee.
+// String renders the condition for display.
+func (c *Condition) String() string {
+	switch {
+	case c.Reg2 != "":
+		return fmt.Sprintf("%s %s %s", c.Reg, c.Op, c.Reg2)
+	case c.MemAddr != 0:
+		return fmt.Sprintf("%s %s *%#x", c.Reg, c.Op, c.MemAddr)
+	default:
+		return fmt.Sprintf("%s %s %#x", c.Reg, c.Op, c.Value)
+	}
+}
+
+// Breakpoints manages the software breakpoints of one Session.
 type Breakpoints struct {
 	t   tracer
 	bps map[uint64]*Breakpoint
@@ -94,6 +106,18 @@ func NewBreakpoints(t tracer) *Breakpoints {
 	return &Breakpoints{t: t, bps: make(map[uint64]*Breakpoint)}
 }
 
+// breakpointMask is the byte mask of the breakpoint instruction inside a
+// peeked word: the low len(breakpointInsn) bytes, because every supported
+// architecture is little-endian and patches the instruction at the lowest
+// address of the word.
+func breakpointMask() uint64 {
+	var mask uint64
+	for range breakpointInsn {
+		mask = (mask << 8) | 0xFF
+	}
+	return mask
+}
+
 // Set installs a breakpoint at addr (replaces any existing one).
 func (bm *Breakpoints) Set(addr uint64, label string) (*Breakpoint, error) {
 	return bm.SetWithCond(addr, label, nil)
@@ -106,14 +130,17 @@ func (bm *Breakpoints) SetWithCond(addr uint64, label string, cond *Condition) (
 		bp.Cond = cond
 		return bp, nil
 	}
-	// Read the original byte.
+	// Read the original bytes.
 	word, err := bm.t.Peek(addr)
 	if err != nil {
 		return nil, err
 	}
-	orig := byte(word)
-	// Patch with INT3 (0xCC), preserving the rest of the word.
-	patched := (word &^ 0xFF) | 0xCC
+	orig := make([]byte, len(breakpointInsn))
+	for i := range orig {
+		orig[i] = byte(word >> (8 * i))
+	}
+	// Patch with the breakpoint instruction, preserving the rest of the word.
+	patched := (word &^ breakpointMask()) | breakpointWord(breakpointInsn)
 	if err := bm.t.Poke(addr, patched); err != nil {
 		return nil, err
 	}
@@ -122,17 +149,12 @@ func (bm *Breakpoints) SetWithCond(addr uint64, label string, cond *Condition) (
 	return bp, nil
 }
 
-// Hits returns the number of times the breakpoint has been hit.
-func (bp *Breakpoint) Hits() int {
-	return bp.hits
-}
-
 // Info returns a formatted list of all breakpoints.
 func (bm *Breakpoints) Info() string {
 	if len(bm.bps) == 0 {
 		return "no breakpoints set\n"
 	}
-	result := ""
+	var result strings.Builder
 	i := 0
 	for _, bp := range bm.bps {
 		i++
@@ -146,26 +168,40 @@ func (bm *Breakpoints) Info() string {
 		}
 		cond := ""
 		if bp.Cond != nil {
-			cond = fmt.Sprintf(" if %s %s %#x", bp.Cond.Reg, bp.Cond.Op, bp.Cond.Value)
+			cond = " if " + bp.Cond.String()
 		}
-		result += fmt.Sprintf("  %d: %s at %#x [%s, %d hits]%s\n", i, label, bp.Addr, status, bp.hits, cond)
+		result.WriteString(fmt.Sprintf("  %d: %s at %#x [%s, %d hits]%s\n", i, label, bp.Addr, status, bp.hits, cond))
 	}
-	return result
+	return result.String()
 }
 
-// Clear removes the breakpoint at addr, restoring the original byte.
+// restore writes the saved original bytes back over the breakpoint
+// instruction, preserving the rest of the peeked word.  It reports whether
+// both the peek and the poke succeeded.
+func (bm *Breakpoints) restore(addr uint64, bp *Breakpoint) bool {
+	word, err := bm.t.Peek(addr)
+	if err != nil {
+		return false
+	}
+	orig := uint64(0)
+	for i, b := range bp.Orig {
+		orig |= uint64(b) << (8 * i)
+	}
+	return bm.t.Poke(addr, (word&^breakpointMask())|orig) == nil
+}
+
+// Clear removes the breakpoint at addr, restoring the original bytes.
 func (bm *Breakpoints) Clear(addr uint64) error {
 	bp, ok := bm.bps[addr]
 	if !ok {
 		return fmt.Errorf("debug: no breakpoint at %#x", addr)
 	}
-	word, err := bm.t.Peek(addr)
-	if err != nil {
-		return err
-	}
-	restored := (word &^ 0xFF) | uint64(bp.Orig)
-	if err := bm.t.Poke(addr, restored); err != nil {
-		return err
+	if !bm.restore(addr, bp) {
+		word, err := bm.t.Peek(addr)
+		if err != nil {
+			return err
+		}
+		return fmt.Errorf("debug: restore breakpoint at %#x failed, word is %#x", addr, word)
 	}
 	delete(bm.bps, addr)
 	return nil
@@ -196,41 +232,55 @@ func (bm *Breakpoints) All() []*Breakpoint {
 }
 
 // HandleTrap is called after the debuggee stops on SIGTRAP.  It checks
-// whether the trap was caused by one of our breakpoints (RIP-1 matches
-// a breakpoint address), restores the original byte, rewinds RIP, and
+// whether the trap was caused by one of our breakpoints (PC-adjust matches
+// a breakpoint address), restores the original bytes, rewinds PC, and
 // returns the breakpoint that was hit (or nil if it was a single-step).
+// Hits returns how many times the breakpoint has been hit.
+func (bp *Breakpoint) Hits() int { return bp.hits }
+
 func (bm *Breakpoints) HandleTrap(regs *Regs) *Breakpoint {
-	// After INT3, RIP points to the byte AFTER the 0xCC.
-	trapAddr := regs.RIP - 1
+	// On amd64 the kernel reports the trap with RIP past the INT3; on the
+	// other supported architectures the PC still stands on the trap
+	// instruction, which breakpointPCAdjust encodes per architecture.
+	trapAddr := regs.GetPC() - uint64(breakpointPCAdjust)
 	bp, ok := bm.bps[trapAddr]
 	if !ok || !bp.Enabled {
 		return nil // single-step trap or unknown
 	}
 	// Check the condition (if any).
-	if bp.Cond != nil && !bp.Cond.Eval(regs) {
-		// Condition not met — restore the byte but do NOT rewind RIP.
-		// The process continues from the next instruction (past the INT3).
-		word, err := bm.t.Peek(trapAddr)
-		if err == nil {
-			restored := (word &^ 0xFF) | uint64(bp.Orig)
-			bm.t.Poke(trapAddr, restored)
+	if bp.Cond != nil && !bp.Cond.Eval(regs, bm.peekValue) {
+		// Condition not met: step the original instruction and re-arm the
+		// breakpoint, leaving the debuggee stopped just past it, ready to
+		// resume silently.  The PC must be rewound first: on architectures
+		// that report the trap past the instruction (amd64) it would
+		// otherwise sit on the second byte of the replaced instruction.
+		if !bm.restore(trapAddr, bp) {
+			return nil
 		}
-		// RIP is already past the INT3 (trapAddr + 1). Don't rewind.
+		regs.SetPC(trapAddr)
+		if err := bm.t.SetRegs(regs); err != nil {
+			return nil
+		}
+		if err := bm.t.Step(); err != nil {
+			return nil
+		}
+		bm.Reinsert(trapAddr)
 		return nil
 	}
 	bp.hits++
-	// Restore the original byte.
-	word, err := bm.t.Peek(trapAddr)
-	if err == nil {
-		restored := (word &^ 0xFF) | uint64(bp.Orig)
-		bm.t.Poke(trapAddr, restored)
-	}
-	// Rewind RIP to re-execute the original instruction.
-	regs.RIP = trapAddr
+	// Restore the original bytes and rewind PC to re-execute them.
+	bm.restore(trapAddr, bp)
+	regs.SetPC(trapAddr)
 	bm.t.SetRegs(regs)
 	return bp
 }
 
+// peekValue adapts tracer.Peek to the Condition value reader.
+func (bm *Breakpoints) peekValue(addr uint64) (uint64, bool) {
+	v, err := bm.t.Peek(addr)
+	return v, err == nil
+}
+
 // Reinsert re-inserts the breakpoint at addr after a single-step past it.
 // Called after Step() when we want the breakpoint to fire again on the
 // next Continue().
@@ -243,6 +293,15 @@ func (bm *Breakpoints) Reinsert(addr uint64) error {
 	if err != nil {
 		return err
 	}
-	patched := (word &^ 0xFF) | 0xCC
+	patched := (word &^ breakpointMask()) | breakpointWord(breakpointInsn)
 	return bm.t.Poke(addr, patched)
 }
+
+// breakpointWord converts the breakpoint instruction bytes to a uint64.
+func breakpointWord(insn []byte) uint64 {
+	var w uint64
+	for i, b := range insn {
+		w |= uint64(b) << (i * 8)
+	}
+	return w
+}

debug/debug_arch_test.go

@@ -0,0 +1,265 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+//go:build linux
+
+package debug
+
+// Architecture-neutral tests: label and line tables, and the breakpoint
+// manager against the mock tracer.  These do not launch a debuggee, so they
+// build on every supported linux architecture.
+
+import (
+	"strings"
+	"testing"
+)
+
+func TestLineAt(t *testing.T) {
+	lines := []SourceLine{
+		{Offset: 0, Line: 5},
+		{Offset: 5, Line: 6},
+		{Offset: 10, Line: 7},
+		{Offset: 15, Line: 8},
+	}
+
+	tests := []struct {
+		offset int
+		want   int
+	}{
+		{0, 5},
+		{1, 5},
+		{4, 5},
+		{5, 6},
+		{7, 6},
+		{10, 7},
+		{12, 7},
+		{15, 8},
+		{20, 8},
+	}
+
+	for _, tt := range tests {
+		got := lineAt(lines, tt.offset)
+		if got != tt.want {
+			t.Errorf("lineAt(lines, %d) = %d, want %d", tt.offset, got, tt.want)
+		}
+	}
+
+	// Empty table.
+	if lineAt(nil, 5) != 0 {
+		t.Error("lineAt(nil, 5) should return 0")
+	}
+}
+
+func TestOffsetForLine(t *testing.T) {
+	lines := []SourceLine{
+		{Offset: 0, Line: 5},
+		{Offset: 5, Line: 6},
+		{Offset: 10, Line: 7},
+	}
+
+	tests := []struct {
+		line int
+		want int
+	}{
+		{5, 0},
+		{6, 5},
+		{7, 10},
+		{99, -1}, // not found
+		{0, -1},  // not found
+	}
+
+	for _, tt := range tests {
+		got := offsetForLine(lines, tt.line)
+		if got != tt.want {
+			t.Errorf("offsetForLine(lines, %d) = %d, want %d", tt.line, got, tt.want)
+		}
+	}
+}
+
+func TestNearestLabel(t *testing.T) {
+	labels := []Label{
+		{Name: "start", Offset: 0},
+		{Name: "loop", Offset: 10},
+		{Name: "done", Offset: 20},
+	}
+
+	tests := []struct {
+		offset int
+		want   string
+	}{
+		{0, "start"},
+		{5, "start"},
+		{10, "loop"},
+		{15, "loop"},
+		{20, "done"},
+		{25, "done"},
+	}
+
+	for _, tt := range tests {
+		got := nearestLabel(labels, tt.offset)
+		if got != tt.want {
+			t.Errorf("nearestLabel(labels, %d) = %q, want %q", tt.offset, got, tt.want)
+		}
+	}
+}
+
+func TestBreakpointsSetAndClear(t *testing.T) {
+	tr := newMockTracer()
+	bm := NewBreakpoints(tr)
+
+	// Set a breakpoint at address 0x1000.
+	bp, err := bm.Set(0x1000, "test")
+	if err != nil {
+		t.Fatalf("Set: %v", err)
+	}
+	if !bp.Enabled {
+		t.Error("breakpoint not enabled")
+	}
+	if bp.Label != "test" {
+		t.Errorf("label = %q, want test", bp.Label)
+	}
+
+	// Verify Peek was called.
+	if len(tr.peeks) != 1 || tr.peeks[0] != 0x1000 {
+		t.Errorf("peeks = %v, want [0x1000]", tr.peeks)
+	}
+
+	// Verify Poke wrote the breakpoint instruction's bytes.
+	if len(tr.pokes) != 1 || tr.pokes[0].addr != 0x1000 {
+		t.Errorf("pokes = %v", tr.pokes)
+	}
+	if got := tr.pokes[0].val & breakpointMask(); got != breakpointWord(breakpointInsn) {
+		t.Errorf("patched bytes %#x, want %#x", got, breakpointWord(breakpointInsn))
+	}
+
+	// At should find it.
+	if bm.At(0x1000) == nil {
+		t.Error("At(0x1000) returned nil")
+	}
+
+	// All should return it.
+	all := bm.All()
+	if len(all) != 1 {
+		t.Errorf("All() = %d breakpoints, want 1", len(all))
+	}
+
+	// Clear it.
+	if err := bm.Clear(0x1000); err != nil {
+		t.Fatalf("Clear: %v", err)
+	}
+	if bm.At(0x1000) != nil {
+		t.Error("At(0x1000) after Clear should be nil")
+	}
+}
+
+// TestBreakpointRestoreWidth proves the restore path writes back every
+// byte of the breakpoint instruction's width, not just the first byte: on
+// arm64, riscv64 and loong64 the instruction is four bytes, and restoring
+// one byte would leave three bytes of the trap instruction in place.
+func TestBreakpointRestoreWidth(t *testing.T) {
+	tr := newMockTracer()
+	bm := NewBreakpoints(tr)
+	tr.mem[0x3000] = 0x11
+	tr.mem[0x3001] = 0x22
+	tr.mem[0x3002] = 0x33
+	tr.mem[0x3003] = 0x44
+
+	if _, err := bm.Set(0x3000, "width"); err != nil {
+		t.Fatalf("Set: %v", err)
+	}
+	for i, b := range breakpointInsn {
+		if tr.mem[0x3000+uint64(i)] != b {
+			t.Fatalf("byte %d after Set = %#x, want the breakpoint byte %#x", i, tr.mem[0x3000+uint64(i)], b)
+		}
+	}
+	if len(bm.At(0x3000).Orig) != len(breakpointInsn) {
+		t.Fatalf("Orig holds %d bytes, want %d", len(bm.At(0x3000).Orig), len(breakpointInsn))
+	}
+
+	if err := bm.Clear(0x3000); err != nil {
+		t.Fatalf("Clear: %v", err)
+	}
+	want := []byte{0x11, 0x22, 0x33, 0x44}
+	for i, b := range want {
+		if tr.mem[0x3000+uint64(i)] != b {
+			t.Errorf("byte %d after Clear = %#x, want %#x (restore must cover the full instruction width)", i, tr.mem[0x3000+uint64(i)], b)
+		}
+	}
+}
+
+func TestBreakpointsSetWithCond(t *testing.T) {
+	tr := newMockTracer()
+	bm := NewBreakpoints(tr)
+
+	cond := &Condition{Reg: "rax", Op: "==", Value: 42}
+	bp, err := bm.SetWithCond(0x2000, "cond_test", cond)
+	if err != nil {
+		t.Fatalf("SetWithCond: %v", err)
+	}
+	if bp.Cond == nil || bp.Cond.Value != 42 {
+		t.Error("condition not set")
+	}
+
+	// Re-setting the same address should update the condition.
+	cond2 := &Condition{Reg: "rbx", Op: "<", Value: 100}
+	bp2, err := bm.SetWithCond(0x2000, "cond_test2", cond2)
+	if err != nil {
+		t.Fatalf("SetWithCond (update): %v", err)
+	}
+	if bp2.Cond.Value != 100 {
+		t.Error("condition not updated")
+	}
+	// Should have only 1 Peek (first Set), second is update (no Peek needed).
+	if len(tr.peeks) != 1 {
+		t.Errorf("expected 1 Peek, got %d", len(tr.peeks))
+	}
+}
+
+func TestBreakpointsClearAll(t *testing.T) {
+	tr := newMockTracer()
+	bm := NewBreakpoints(tr)
+
+	bm.Set(0x1000, "a")
+	bm.Set(0x2000, "b")
+	bm.Set(0x3000, "c")
+
+	if len(bm.All()) != 3 {
+		t.Fatalf("expected 3 breakpoints, got %d", len(bm.All()))
+	}
+
+	bm.ClearAll()
+	if len(bm.All()) != 0 {
+		t.Errorf("ClearAll: expected 0 breakpoints, got %d", len(bm.All()))
+	}
+}
+
+func TestBreakpointInfo(t *testing.T) {
+	tr := newMockTracer()
+	bm := NewBreakpoints(tr)
+	bm.Set(0x4000, "info_test")
+
+	info := bm.Info()
+	if info == "" {
+		t.Error("Info returned empty string")
+	}
+	if !strings.Contains(info, "info_test") {
+		t.Errorf("Info %q does not contain label", info)
+	}
+}
+
+// TestConditionString covers the display of all three condition forms.
+func TestConditionString(t *testing.T) {
+	tests := []struct {
+		cond Condition
+		want string
+	}{
+		{Condition{Reg: "rax", Op: "==", Value: 42}, "rax == 0x2a"},
+		{Condition{Reg: "rax", Op: "!=", Reg2: "rbx"}, "rax != rbx"},
+		{Condition{Reg: "rax", Op: "<", MemAddr: 0x5000}, "rax < *0x5000"},
+	}
+	for _, tt := range tests {
+		if got := tt.cond.String(); got != tt.want {
+			t.Errorf("Condition.String() = %q, want %q", got, tt.want)
+		}
+	}
+}

debug/debug_test.go

@@ -1,10 +1,11 @@
 // Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
 // SPDX-License-Identifier: BSD-3-Clause
 
+//go:build linux && amd64
+
 package debug
 
 import (
-	"strings"
 	"testing"
 )
 
@@ -46,7 +47,7 @@ func TestConditionEval(t *testing.T) {
 	}
 
 	for _, tt := range tests {
-		got := tt.cond.Eval(regs)
+		got := tt.cond.Eval(regs, nil)
 		if got != tt.want {
 			t.Errorf("Condition{%q %q %d}.Eval() = %v, want %v",
 				tt.cond.Reg, tt.cond.Op, tt.cond.Value, got, tt.want)
@@ -54,65 +55,33 @@ func TestConditionEval(t *testing.T) {
 	}
 }
 
-func TestLineAt(t *testing.T) {
-	lines := []SourceLine{
-		{Offset: 0, Line: 5},
-		{Offset: 5, Line: 6},
-		{Offset: 10, Line: 7},
-		{Offset: 15, Line: 8},
-	}
-
-	tests := []struct {
-		offset int
-		want   int
-	}{
-		{0, 5},
-		{1, 5},
-		{4, 5},
-		{5, 6},
-		{7, 6},
-		{10, 7},
-		{12, 7},
-		{15, 8},
-		{20, 8},
-	}
-
-	for _, tt := range tests {
-		got := lineAt(lines, tt.offset)
-		if got != tt.want {
-			t.Errorf("lineAt(lines, %d) = %d, want %d", tt.offset, got, tt.want)
+// TestConditionEvalMem covers the register-memory form: the value is read
+// through the supplied reader, and a missing or failing reader must not
+// block the breakpoint.
+func TestConditionEvalMem(t *testing.T) {
+	regs := &Regs{RAX: 7}
+	mem := func(addr uint64) (uint64, bool) {
+		if addr == 0x5000 {
+			return 7, true
 		}
+		return 0, false
 	}
 
-	// Empty table.
-	if lineAt(nil, 5) != 0 {
-		t.Error("lineAt(nil, 5) should return 0")
+	eq := Condition{Reg: "rax", Op: "==", MemAddr: 0x5000}
+	if !eq.Eval(regs, mem) {
+		t.Error("register-memory comparison with matching word should hold")
 	}
-}
-
-func TestOffsetForLine(t *testing.T) {
-	lines := []SourceLine{
-		{Offset: 0, Line: 5},
-		{Offset: 5, Line: 6},
-		{Offset: 10, Line: 7},
+	ne := Condition{Reg: "rax", Op: "!=", MemAddr: 0x5000}
+	if ne.Eval(regs, mem) {
+		t.Error("register-memory comparison with mismatching word should not hold")
 	}
-
-	tests := []struct {
-		line int
-		want int
-	}{
-		{5, 0},
-		{6, 5},
-		{7, 10},
-		{99, -1}, // not found
-		{0, -1},  // not found
+	bad := Condition{Reg: "rax", Op: "==", MemAddr: 0x6000}
+	if !bad.Eval(regs, mem) {
+		t.Error("unreadable memory must not block the breakpoint")
 	}
-
-	for _, tt := range tests {
-		got := offsetForLine(lines, tt.line)
-		if got != tt.want {
-			t.Errorf("offsetForLine(lines, %d) = %d, want %d", tt.line, got, tt.want)
-		}
+	noReader := Condition{Reg: "rax", Op: "==", MemAddr: 0x5000}
+	if !noReader.Eval(regs, nil) {
+		t.Error("missing memory reader must not block the breakpoint")
 	}
 }
 
@@ -139,144 +108,11 @@ func TestDecodeRflags(t *testing.T) {
 	}
 }
 
-func TestNearestLabel(t *testing.T) {
-	labels := []Label{
-		{Name: "start", Offset: 0},
-		{Name: "loop", Offset: 10},
-		{Name: "done", Offset: 20},
-	}
-
-	tests := []struct {
-		offset int
-		want   string
-	}{
-		{0, "start"},
-		{5, "start"},
-		{10, "loop"},
-		{15, "loop"},
-		{20, "done"},
-		{25, "done"},
-	}
-
-	for _, tt := range tests {
-		got := nearestLabel(labels, tt.offset)
-		if got != tt.want {
-			t.Errorf("nearestLabel(labels, %d) = %q, want %q", tt.offset, got, tt.want)
-		}
-	}
-}
-
-func TestBreakpointsSetAndClear(t *testing.T) {
-	tr := newMockTracer()
-	bm := NewBreakpoints(tr)
-
-	// Set a breakpoint at address 0x1000.
-	bp, err := bm.Set(0x1000, "test")
-	if err != nil {
-		t.Fatalf("Set: %v", err)
-	}
-	if !bp.Enabled {
-		t.Error("breakpoint not enabled")
-	}
-	if bp.Label != "test" {
-		t.Errorf("label = %q, want test", bp.Label)
-	}
-
-	// Verify Peek was called.
-	if len(tr.peeks) != 1 || tr.peeks[0] != 0x1000 {
-		t.Errorf("peeks = %v, want [0x1000]", tr.peeks)
-	}
-
-	// Verify Poke wrote INT3.
-	if len(tr.pokes) != 1 || tr.pokes[0].addr != 0x1000 {
-		t.Errorf("pokes = %v", tr.pokes)
-	}
-
-	// At should find it.
-	if bm.At(0x1000) == nil {
-		t.Error("At(0x1000) returned nil")
-	}
-
-	// All should return it.
-	all := bm.All()
-	if len(all) != 1 {
-		t.Errorf("All() = %d breakpoints, want 1", len(all))
-	}
-
-	// Clear it.
-	if err := bm.Clear(0x1000); err != nil {
-		t.Fatalf("Clear: %v", err)
-	}
-	if bm.At(0x1000) != nil {
-		t.Error("At(0x1000) after Clear should be nil")
-	}
-}
-
-func TestBreakpointsSetWithCond(t *testing.T) {
-	tr := newMockTracer()
-	bm := NewBreakpoints(tr)
-
-	cond := &Condition{Reg: "rax", Op: "==", Value: 42}
-	bp, err := bm.SetWithCond(0x2000, "cond_test", cond)
-	if err != nil {
-		t.Fatalf("SetWithCond: %v", err)
-	}
-	if bp.Cond == nil || bp.Cond.Value != 42 {
-		t.Error("condition not set")
-	}
-
-	// Re-setting the same address should update the condition.
-	cond2 := &Condition{Reg: "rbx", Op: "<", Value: 100}
-	bp2, err := bm.SetWithCond(0x2000, "cond_test2", cond2)
-	if err != nil {
-		t.Fatalf("SetWithCond (update): %v", err)
-	}
-	if bp2.Cond.Value != 100 {
-		t.Error("condition not updated")
-	}
-	// Should have only 1 Peek (first Set), second is update (no Peek needed).
-	if len(tr.peeks) != 1 {
-		t.Errorf("expected 1 Peek, got %d", len(tr.peeks))
-	}
-}
-
-func TestBreakpointsClearAll(t *testing.T) {
-	tr := newMockTracer()
-	bm := NewBreakpoints(tr)
-
-	bm.Set(0x1000, "a")
-	bm.Set(0x2000, "b")
-	bm.Set(0x3000, "c")
-
-	if len(bm.All()) != 3 {
-		t.Fatalf("expected 3 breakpoints, got %d", len(bm.All()))
-	}
-
-	bm.ClearAll()
-	if len(bm.All()) != 0 {
-		t.Errorf("ClearAll: expected 0 breakpoints, got %d", len(bm.All()))
-	}
-}
-
-func TestBreakpointInfo(t *testing.T) {
-	tr := newMockTracer()
-	bm := NewBreakpoints(tr)
-	bm.Set(0x4000, "info_test")
-
-	info := bm.Info()
-	if info == "" {
-		t.Error("Info returned empty string")
-	}
-	if !strings.Contains(info, "info_test") {
-		t.Errorf("Info %q does not contain label", info)
-	}
-}
-
 func TestWatchpointSlotTracking(t *testing.T) {
-	s := &Session{}
+	s := &Session{} // per-session slots start free
 
 	// All four slots are free initially.
-	for i := 0; i < 4; i++ {
+	for i := range 4 {
 		if s.IsWatchpointSlotUsed(i) {
 			t.Errorf("slot %d should be free initially", i)
 		}
@@ -314,10 +150,26 @@ func TestWatchpointSlotTracking(t *testing.T) {
 	}
 
 	// Mark all slots used: FindFreeWatchpointSlot returns -1.
-	for i := 0; i < 4; i++ {
+	for i := range 4 {
 		s.wpSlots[i] = true
 	}
 	if got := s.FindFreeWatchpointSlot(); got != -1 {
 		t.Errorf("FindFreeWatchpointSlot() with all slots used = %d, want -1", got)
 	}
 }
+
+// TestUnwatchSlotBound checks the bound the REPL parses against: it must
+// cover the architecture's whole slot range, not a hardcoded 0-3.
+func TestUnwatchSlotBound(t *testing.T) {
+	max := maxWatchpoints()
+	if max < 4 {
+		t.Fatalf("maxWatchpoints() = %d, want at least 4", max)
+	}
+	s := &Session{}
+	if s.IsWatchpointSlotUsed(max - 1) {
+		t.Errorf("slot %d should be free initially", max-1)
+	}
+	if s.IsWatchpointSlotUsed(max) {
+		t.Errorf("slot %d must be out of range", max)
+	}
+}

debug/disasm_freebsd_amd64.go

@@ -0,0 +1,56 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+//go:build freebsd && amd64
+
+package debug
+
+import (
+	"fmt"
+	"strings"
+
+	"sourcedock.dev/petrbalvin/gasm-sdk/arch"
+	"sourcedock.dev/petrbalvin/gasm-sdk/disasm"
+)
+
+// Disassemble decodes the instruction at the given address in the debuggee's
+// memory and returns its text representation and length in bytes.
+func (s *Session) Disassemble(addr uint64) (string, int, error) {
+	mem, err := s.ReadMemory(addr, 15)
+	if err != nil {
+		return "", 0, err
+	}
+	ins, err := disasm.Decode(arch.AMD64, mem, addr)
+	if err != nil {
+		return "", 0, err
+	}
+	return ins.Text, ins.Len, nil
+}
+
+// DisassembleN decodes up to n instructions starting at addr and returns
+// them as a formatted string with addresses and byte offsets.
+func (s *Session) DisassembleN(addr uint64, n int) string {
+	var result strings.Builder
+	pc := addr
+	for range n {
+		text, length, err := s.Disassemble(pc)
+		if err != nil {
+			result.WriteString(fmt.Sprintf("  %#08x: <error: %v>\n", pc, err))
+			break
+		}
+		result.WriteString(fmt.Sprintf("  %#08x: %s\n", pc, text))
+		if length == 0 {
+			length = 1
+		}
+		pc += uint64(length)
+	}
+	return result.String()
+}
+
+// isCallInsn reports whether disassembled text (x86asm.IntelSyntax) is a
+// call.  The first token must match exactly: a prefix test would also catch
+// unrelated mnemonics.
+func isCallInsn(text string) bool {
+	m, _, _ := strings.Cut(text, " ")
+	return strings.ToLower(m) == "call"
+}

debug/disasm_freebsd_arm64.go

@@ -0,0 +1,60 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+//go:build freebsd && arm64
+
+package debug
+
+import (
+	"fmt"
+	"strings"
+
+	"sourcedock.dev/petrbalvin/gasm-sdk/arch"
+	"sourcedock.dev/petrbalvin/gasm-sdk/disasm"
+)
+
+// Disassemble decodes the instruction at the given address in the debuggee's
+// memory and returns its text representation and length in bytes.
+func (s *Session) Disassemble(addr uint64) (string, int, error) {
+	mem, err := s.ReadMemory(addr, 4)
+	if err != nil {
+		return "", 0, err
+	}
+	ins, err := disasm.Decode(arch.ARM64, mem, addr)
+	if err != nil {
+		return "", 0, err
+	}
+	return ins.Text, ins.Len, nil
+}
+
+// DisassembleN decodes up to n instructions starting at addr and returns
+// them as a formatted string with addresses and byte offsets.
+func (s *Session) DisassembleN(addr uint64, n int) string {
+	var result strings.Builder
+	pc := addr
+	for range n {
+		text, length, err := s.Disassemble(pc)
+		if err != nil {
+			result.WriteString(fmt.Sprintf("  %#08x: <error: %v>\n", pc, err))
+			break
+		}
+		result.WriteString(fmt.Sprintf("  %#08x: %s\n", pc, text))
+		if length == 0 {
+			length = 1
+		}
+		pc += uint64(length)
+	}
+	return result.String()
+}
+
+// isCallInsn reports whether disassembled text (arm64asm.GoSyntax) is a
+// call.  GoSyntax renders bl as CALL; the native mnemonic is accepted too.
+// The first token must match exactly so branches never match.
+func isCallInsn(text string) bool {
+	m, _, _ := strings.Cut(text, " ")
+	switch strings.ToLower(m) {
+	case "call", "bl":
+		return true
+	}
+	return false
+}

debug/disasm_freebsd_riscv64.go

@@ -0,0 +1,62 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+//go:build freebsd && riscv64
+
+package debug
+
+import (
+	"fmt"
+	"strings"
+
+	"sourcedock.dev/petrbalvin/gasm-sdk/arch"
+	"sourcedock.dev/petrbalvin/gasm-sdk/disasm"
+)
+
+// Disassemble decodes the instruction at the given address in the debuggee's
+// memory and returns its text representation and length in bytes.
+func (s *Session) Disassemble(addr uint64) (string, int, error) {
+	mem, err := s.ReadMemory(addr, 4)
+	if err != nil {
+		return "", 0, err
+	}
+	ins, err := disasm.Decode(arch.RISCV, mem, addr)
+	if err != nil {
+		return "", 0, err
+	}
+	return ins.Text, ins.Len, nil
+}
+
+// DisassembleN decodes up to n instructions starting at addr and returns
+// them as a formatted string with addresses and byte offsets.
+func (s *Session) DisassembleN(addr uint64, n int) string {
+	var result strings.Builder
+	pc := addr
+	for range n {
+		text, length, err := s.Disassemble(pc)
+		if err != nil {
+			result.WriteString(fmt.Sprintf("  %#08x: <error: %v>\n", pc, err))
+			break
+		}
+		result.WriteString(fmt.Sprintf("  %#08x: %s\n", pc, text))
+		if length == 0 {
+			length = 1
+		}
+		pc += uint64(length)
+	}
+	return result.String()
+}
+
+// isCallInsn reports whether disassembled text (riscv64asm.GoSyntax) is a
+// call.  GoSyntax renders jal and jalr calls as CALL; the native mnemonics
+// are accepted too.  The first token must match exactly: a prefix test on
+// "bl" would catch branches on other architectures, and jalr as ret prints
+// RET, which must not be stepped over.
+func isCallInsn(text string) bool {
+	m, _, _ := strings.Cut(text, " ")
+	switch strings.ToLower(m) {
+	case "call", "jal", "jalr":
+		return true
+	}
+	return false
+}

debug/disasm_linux_amd64.go

@@ -7,46 +7,50 @@ package debug
 
 import (
 	"fmt"
+	"strings"
 
-	"golang.org/x/arch/x86/x86asm"
+	"sourcedock.dev/petrbalvin/gasm-sdk/arch"
+	"sourcedock.dev/petrbalvin/gasm-sdk/disasm"
 )
 
 // Disassemble decodes the instruction at the given address in the debuggee's
 // memory and returns its text representation and length in bytes.
 func (s *Session) Disassemble(addr uint64) (string, int, error) {
-	// Read up to 15 bytes (max x86 instruction length).
 	mem, err := s.ReadMemory(addr, 15)
 	if err != nil {
-		// Try a shorter read if we're near a page boundary.
-		mem, err = s.ReadMemory(addr, 1)
-		if err != nil {
-			return "", 0, err
-		}
+		return "", 0, err
 	}
-	inst, err := x86asm.Decode(mem, 64)
+	ins, err := disasm.Decode(arch.AMD64, mem, addr)
 	if err != nil {
-		return "???", 1, nil
+		return "", 0, err
 	}
-	text := x86asm.IntelSyntax(inst, addr, nil)
-	return text, inst.Len, nil
+	return ins.Text, ins.Len, nil
 }
 
 // DisassembleN decodes up to n instructions starting at addr and returns
 // them as a formatted string with addresses and byte offsets.
 func (s *Session) DisassembleN(addr uint64, n int) string {
-	var result string
+	var result strings.Builder
 	pc := addr
-	for i := 0; i < n; i++ {
+	for range n {
 		text, length, err := s.Disassemble(pc)
 		if err != nil {
-			result += fmt.Sprintf("  %#08x: <error: %v>\n", pc, err)
+			result.WriteString(fmt.Sprintf("  %#08x: <error: %v>\n", pc, err))
 			break
 		}
-		result += fmt.Sprintf("  %#08x: %s\n", pc, text)
+		result.WriteString(fmt.Sprintf("  %#08x: %s\n", pc, text))
 		if length == 0 {
 			length = 1
 		}
 		pc += uint64(length)
 	}
-	return result
+	return result.String()
+}
+
+// isCallInsn reports whether disassembled text (x86asm.IntelSyntax) is a
+// call.  The first token must match exactly: a prefix test would also catch
+// unrelated mnemonics.
+func isCallInsn(text string) bool {
+	m, _, _ := strings.Cut(text, " ")
+	return strings.ToLower(m) == "call"
 }

debug/disasm_linux_arm64.go

@@ -0,0 +1,59 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+//go:build linux && arm64
+
+package debug
+
+import (
+	"fmt"
+	"strings"
+
+	"sourcedock.dev/petrbalvin/gasm-sdk/arch"
+	"sourcedock.dev/petrbalvin/gasm-sdk/disasm"
+)
+
+// Disassemble decodes the instruction at the given address in the debuggee's
+// memory and returns its text representation and length in bytes.
+func (s *Session) Disassemble(addr uint64) (string, int, error) {
+	mem, err := s.ReadMemory(addr, 4)
+	if err != nil {
+		return "", 0, err
+	}
+	ins, err := disasm.Decode(arch.ARM64, mem, addr)
+	if err != nil {
+		return "", 0, err
+	}
+	return ins.Text, ins.Len, nil
+}
+
+// DisassembleN decodes up to n instructions starting at addr.
+func (s *Session) DisassembleN(addr uint64, n int) string {
+	var result string
+	pc := addr
+	for range n {
+		text, length, err := s.Disassemble(pc)
+		if err != nil {
+			result += fmt.Sprintf("  %#08x: <error: %v>\n", pc, err)
+			break
+		}
+		result += fmt.Sprintf("  %#08x: %s\n", pc, text)
+		if length == 0 {
+			length = 4
+		}
+		pc += uint64(length)
+	}
+	return result
+}
+
+// isCallInsn reports whether disassembled text (arm64asm.GoSyntax) is a
+// call.  GoSyntax renders bl as CALL; the native mnemonic is accepted too.
+// The first token must match exactly so branches never match.
+func isCallInsn(text string) bool {
+	m, _, _ := strings.Cut(text, " ")
+	switch strings.ToLower(m) {
+	case "call", "bl":
+		return true
+	}
+	return false
+}

debug/disasm_linux_loong64.go

@@ -0,0 +1,60 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+//go:build linux && loong64
+
+package debug
+
+import (
+	"fmt"
+	"strings"
+
+	"sourcedock.dev/petrbalvin/gasm-sdk/arch"
+	"sourcedock.dev/petrbalvin/gasm-sdk/disasm"
+)
+
+// Disassemble decodes the instruction at the given address in the debuggee's
+// memory and returns its text representation and length in bytes.
+func (s *Session) Disassemble(addr uint64) (string, int, error) {
+	mem, err := s.ReadMemory(addr, 4)
+	if err != nil {
+		return "", 0, err
+	}
+	ins, err := disasm.Decode(arch.LOONG64, mem, addr)
+	if err != nil {
+		return "", 0, err
+	}
+	return ins.Text, ins.Len, nil
+}
+
+// DisassembleN decodes up to n instructions starting at addr.
+func (s *Session) DisassembleN(addr uint64, n int) string {
+	var result string
+	pc := addr
+	for range n {
+		text, length, err := s.Disassemble(pc)
+		if err != nil {
+			result += fmt.Sprintf("  %#08x: <error: %v>\n", pc, err)
+			break
+		}
+		result += fmt.Sprintf("  %#08x: %s\n", pc, text)
+		if length == 0 {
+			length = 4
+		}
+		pc += uint64(length)
+	}
+	return result
+}
+
+// isCallInsn reports whether disassembled text (loong64asm.GoSyntax) is a
+// call.  GoSyntax renders bl and jirl calls as CALL (jirl returns print
+// RET); the native mnemonics are accepted too.  The first token must match
+// exactly: a "bl" prefix would catch bltz and other branches.
+func isCallInsn(text string) bool {
+	m, _, _ := strings.Cut(text, " ")
+	switch strings.ToLower(m) {
+	case "call", "bl", "jirl":
+		return true
+	}
+	return false
+}