docs: changelog for macro expansion and the corrected corpus audit

Assisted-by: GLM 5.3 Flash
test(verify): register the campaign kernels in the ground-truth suites
2026-09-20 14:25:47 +02:00 · 2026-09-20 14:25:47 +02:00 · 2026-09-20 14:25:47 +02:00 · 2026-09-20 14:25:47 +02:00 · 2026-09-20 14:25:47 +02:00 · 2026-09-20 14:25:47 +02:00
248 changed files with 26672 additions and 3265 deletions
@@ -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 ./...
@@ -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 +226,162 @@ 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 $ENV{GITEA_TOKEN}},
+                q{-H}, q{Content-Type: application/octet-stream},
+                q{-X}, q{POST}, q{--data-binary}, qq{@$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);
+          '
@@ -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 +20,91 @@ 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: 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);
+          '
@@ -1,25 +1,13 @@
-# Metadata (always first, per repo convention)
 .idea/
 .zcode/
-.qwen/
-.mimocode/

-# Binaries
-/gasm
+# Build output
 /bin/
-*.exe
-
-# Test and coverage artefacts
+/gasm
 coverage.out
 *.test

-# Crash dumps
+# Crash dumps from the emulator runs
 core
 core.*
 *.core
-
-# Scratch / temporary work
-_scratch/
-
-# ZCode workspace
-.zcode
@@ -1,107 +1,134 @@
-# Contributing to gasm-devkit
+# Contributing

-Thanks for contributing to gasm-devkit.
+Contributions to **gasm-devkit** are governed by the Contributor terms
+below; submitting one means you accept them.
+
+## Contributor terms
+
+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 or later, the [just](https://github.com/casey/just)
-command runner, and a Linux host on amd64, arm64, riscv64 or loong64.
+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 suite, race detector, 80 % coverage gate
+just build
+just gates
 ```

 ## Workflow

-1. Branch from `development`; never commit directly to `main` (`main` is
-   release-only: merge from `development`, then tag).
-2. Commit with [Conventional Commits](https://www.conventionalcommits.org/):
-   `type(scope): description`: subject line only, imperative mood,
-   lowercase after the colon, no trailing dot.  Allowed types: `feat`,
-   `fix`, `docs`, `style`, `refactor`, `perf`, `test`, `chore`, `ci`,
-   `build`, `revert`.  The only line after the subject is the trailer:
-   `Assisted-by: <model-name>`.  No `Co-Authored-By`, no `Signed-off-by`,
-   no other trailers.
-3. Record every user-visible change in `CHANGELOG.md` under
-   `## [development]` (categories: Added, Changed, Fixed, Removed,
-   Security).
-4. Add or update tests; coverage must stay **at or above 80 %** (hard
-   gate, enforced by CI).
-5. Update the documentation when behaviour, flags or the public surface
-   change.
-6. Open a pull request against `development`.
+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`.

-Releases are cut by merging `development` into `main` and tagging `vX.Y.Z`;
-CI builds and publishes the binaries for all four architectures.
+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.

 ## Code style

-`gofmt` and `go vet` via `just fmt` / `just build`; both must pass with
-zero output; `go fix -diff ./...` must report nothing on touched packages.
+`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.

- Standard library only in production code; `golang.org/x/arch` is used
-  in tests only (round-trip decoding) and is never linked into the `gasm`
-  binary.
- No cgo, no C, no external toolchains at runtime.
- Explicit `if err != nil`; errors wrapped with
-  `fmt.Errorf("context: %w", err)`; no panics outside `main`.
- The parser, lexer and formatter are hand-written; the `arch` instruction
-  tables are generated only via `_gen/gen.go` (`just gen`), never edited.
+- `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.

-## Running a single test
+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
-go test -run TestVexGroundTruth ./asm/
-go test -run TestGroundTruthBasic ./verify/
-go test -run TestGOObjectLinkAndRun ./asm/
-go test -run TestFuzzWideCopy ./verify/
+## AI contribution policy
+
+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.
+
+- **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:
+
+  ```
+  Assisted-by: MODEL
  ```

-The interactive debugger (`gasm debug`) requires a compiled binary on
-`$PATH`; `go run` does not work for the traced child process.  Install
-first with `just install-bin`.
+  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.

-## CI (Gitea Actions)
+AI assists. It does not replace judgement.

-Workflows live in `.gitea/workflows/` and run on self-hosted runners:
+## Continuous integration
+
+Workflows live in `.gitea/workflows/` and run on the project's own runners:

 | Workflow | Trigger | What it does |
-|----------|---------|--------------|
-| Test | push / PR to `development` | gofmt check, `go vet`, `go test -race`, 80 % coverage gate |
-| Release | tag `v*` | cross-compiles binaries for linux/{amd64,arm64,riscv64,loong64} and publishes the Gitea release |
+|---|---|---|
+| 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.
-
-## AI Contribution Policy
-
-AI tools are welcome as productivity aids.  What matters is that
-contributions remain understandable, reviewable, and genuinely useful.
-
- **Disclose AI use.**  If you used AI to draft or generate any part of a
-  commit, issue, pull request, or code review, say so clearly.
- **Commit messages:** end every commit with exactly one trailer:
-  `Assisted-by: <model-name>` (e.g. `Assisted-by: GLM 5.3`).
- **Pull requests and issues:** attribute AI assistance in one trailing
-  line, e.g. `_Assisted-by: GLM 5.3_`.  Do not paste it into the PR
-  description as a section.
- **Take responsibility.**  You remain accountable for the accuracy,
-  completeness, and intent of everything you submit.
- **Review before marking ready.**  Read AI-generated diffs carefully, run
-  them locally, and add or update tests where appropriate.
- **Preferred models.**  Prefer open-weight models with transparent
-  training data: **GLM**, **DeepSeek**, and **MiMo**.
+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.

 ## Reporting bugs

-Open an issue at
-[sourcedock.dev/petrbalvin/gasm-devkit](https://sourcedock.dev/petrbalvin/gasm-devkit/issues)
-with the version (`gasm --version`), OS and architecture, the exact
-command, the full output, and the expected versus actual behaviour.
+Open an issue at `https://sourcedock.dev/petrbalvin/gasm-devkit/issues` with the
+version, the operating system and architecture, the exact command, the full output,
+and the expected against the actual behaviour.

-**Security issues:** email **opensource@petrbalvin.org** instead of opening
-a public issue.
+**Security issues do not go in the issue tracker.** Report them as
+[SECURITY.md](SECURITY.md) describes, to **opensource@petrbalvin.org**.
@@ -1,13 +1,65 @@
-# gasm-devkit
+# Plan 9 assembly tooling, inside and outside Go

-Developer tooling for **GAsm**, Go's built-in Plan 9 assembler.
+> **Warning: this is an experiment.** gasm-devkit 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-devkit is the
-missing toolkit: a single, self-contained binary, `gasm`, that brings proper
-developer tooling to Plan 9 assembly on amd64, arm64, riscv64 and loong64.
+**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-devkit is the missing toolkit: a single, self-contained
+binary, `gasm`, that serves both purposes.
+
+- **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.
+
+## 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)
+```
+
+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-devkit exists
+to give that syntax the tooling it deserves.

 ## Features

@@ -19,12 +71,14 @@ developer tooling to Plan 9 assembly on amd64, arm64, riscv64 and loong64.
  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 frame vs the `// func` signature),
-  `register-clobber` (Go ABI register liveness over the control-flow graph),
-  `stack-imbalance`, `abi0-register-args` and `unencodable-instruction`.
+  `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, linkable ELF objects (with DWARF5
-  debug sections) or the Go toolchain's own GOOBJ format, which `go build`
+  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.
@@ -36,7 +90,8 @@ developer tooling to Plan 9 assembly on amd64, arm64, riscv64 and loong64.
  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 with label-level coverage.
+  memory inspection, and headless script runs that report instruction and
+  label coverage.
 - **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
@@ -47,12 +102,11 @@ developer tooling to Plan 9 assembly on amd64, arm64, riscv64 and loong64.
  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.
- **Complete instruction coverage.**  The instruction tables are generated
-  from the Go toolchain's own assembler source, so the toolkit recognises
-  every mnemonic the real assembler accepts; `just gen` refreshes them.

 ### 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 |
@@ -63,27 +117,94 @@ developer tooling to Plan 9 assembly on amd64, arm64, riscv64 and loong64.
 "Common opcodes" are the instructions shared by every architecture (`RET`,
 `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.
+
+The same measurement runs over GOROOT's whole assembly corpus:
+`gasm audit-instructions --corpus` reports 136 of 433 attemptable files
+(31.4 %) assembling for every target architecture today (files named for
+other Go ports are counted but never attempted), with the top failure
+reasons per architecture; the number moves with every release.
+
+### Validation status
+
+**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.
+
+| 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 |
+
+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/...`.
+
+## 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: the
+  JIT's executable-memory mapping and the ptrace debugger layer are the
+  two pieces of porting work.  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-devkit/releases).
-From source (Go 1.27 or later):
+From source (Go 1.27.1):

 ```sh
 go install sourcedock.dev/petrbalvin/gasm-devkit/cmd/gasm@latest
 ```

-Or from a repository checkout, with the development version stamped:
+Or from a repository checkout:

 ```sh
-just install-bin
+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
@@ -119,7 +240,7 @@ 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      # which labels did execution reach?
+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
@@ -142,9 +263,9 @@ infers the target architecture from the file-name suffix
 ## Development

 ```sh
-just install      # download module dependencies
-just build        # go vet + gofmt check, zero errors and zero warnings
-just test         # full suite, race detector, 80 % coverage gate
+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
 ```
@@ -155,14 +276,17 @@ recipe.

 ## Documentation

- [docs/ARCHITECTURE.md](docs/ARCHITECTURE.md): components and data flow
 - [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/DEVELOPMENT.md](docs/DEVELOPMENT.md): development setup and recipes
- [docs/DECISIONS.md](docs/DECISIONS.md): deferred design decisions
 - [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)
@@ -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 |
+|---|---|
+| 0.34.0 | 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.
@@ -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
 }

@@ -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:
@@ -145,11 +145,38 @@ 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",
@@ -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
 }
@@ -4,6 +4,7 @@
 package asm

 import (
+	"encoding/binary"
 	"os"
 	"os/exec"
 	"path/filepath"
@@ -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 {
@@ -187,10 +187,32 @@ func arm64Prologue(fi arm64FrameInfo) []byte {
 	return a64WordsLE(ws...)
 }

-// arm64SubImmWords emits SUB $imm, SP, Rd: the immediate form when the value
-// fits the imm12 field (plain, or shifted left by 12 when it is a multiple
-// of 4096); otherwise the toolchain materialises it into REGTMP (R27) and
-// subtracts the register in the extended-register form.
+// 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)}
@@ -198,15 +220,21 @@ func arm64SubImmWords(imm uint32, rd uint32) []uint32 {
 	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
-// and REGTMP fallback ladder.
+// 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)}
@@ -214,12 +242,36 @@ func arm64AddImmWords(imm uint32, rd uint32) []uint32 {
 	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
 // deallocate the frame when present) followed by RET (BR LR).
@@ -237,11 +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)
 			)
-			ws = append(ws, arm64AddImmWords(uint32(fi.autosize), 31)...)
+			ws = append(ws, arm64RetAddWords(uint32(fi.autosize))...)
 		}
 	}
 	// RET: BR LR (0xd65f03c0)
@@ -258,22 +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
@@ -382,9 +444,12 @@ func arm64GuardBytes(fi arm64FrameInfo, blockStart int) []byte {
 		ws = append(ws, wordsOf(mov)...)
 		ml := len(mov) / 4
 		ws = append(ws, arm64DPExtWords(arm64OpSubs, 27, 31, 17)) // SUBS R17, RSP, R27
-		ws = append(ws, br(8+ml, a64CondLO))
+		// 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(8+ml+8, a64CondLS))
+		ws = append(ws, br(16+4*ml, a64CondLS))
 	}
 	return a64WordsLE(ws...)
 }
@@ -67,6 +67,9 @@ type spadjStep struct {
 // patch sites (for the file-level layout to resolve), the label table and the
 // stack-adjustment boundaries.
 func assemble(t *ast.Text, link *linkInfo) ([]byte, []sbPatch, map[string]int, []spadjStep, []LineEntry, error) {
+	if err := checkAdjspBalance(t); err != nil {
+		return nil, nil, nil, nil, nil, err
+	}
 	fi := computeFrame(t)
 	chain := jumpChain(t)
 	resolve := func(name string) string {
@@ -114,6 +117,11 @@ func assemble(t *ast.Text, link *linkInfo) ([]byte, []sbPatch, map[string]int, [
 			if !isJumpMnemonic(mnem) || mnem == "CALL" || long[i] {
 				continue
 			}
+			// A zero-operand jump parses; its arity is reported during
+			// emission (encodeJump), so the layout must not index Operands.
+			if len(s.Operands) != 1 {
+				continue
+			}
 			name, ok := labelName(s.Operands[0])
 			if !ok {
 				continue // reported during emission
@@ -137,29 +145,23 @@ func assemble(t *ast.Text, link *linkInfo) ([]byte, []sbPatch, map[string]int, [
 		}
 		if fi.splitClass == 2 && !guardJBlong {
 			// The underflow JB sits before the CMPQ; its displacement spans
-			// the rest of the guard plus the prologue and the body.
-			jbLen := 2
-			if guardJBlong {
-				jbLen = 6
-			}
-			rest := fi.guardLen(guardJBlong, guardJBElong) - (9 + 3 + 7 + jbLen)
+			// the rest of the guard plus the prologue and the body.  The JB
+			// is still the short form this branch tests (relaxing it is this
+			// branch's job), so guardLen is taken with a short JB and the
+			// subtraction drops the prefix and the JB's own 2 bytes.
+			rest := fi.guardLen(false, guardJBElong) - (9 + 3 + 7 + 2)
 			if !fits8(int64(rest + len(fi.prologue) + bodyLen)) {
 				guardJBlong = true
 				changed = true
 			}
 		}
 		// The morestack JMP returns to the function start, so its
-		// displacement is the negated distance from its own end.
-		if !moreJMPlong {
-			jmpLen := 2
-			if moreJMPlong {
-				jmpLen = 5
-			}
-			if !fits8(-int64(guard + len(fi.prologue) + bodyLen + 5 + jmpLen)) {
+		// displacement is the negated distance from its own end; while it is
+		// still short, its own length is 2 bytes.
+		if !moreJMPlong && !fits8(-int64(guard+len(fi.prologue)+bodyLen+5+2)) {
 			moreJMPlong = true
 			changed = true
 		}
-		}
 		if !changed {
 			break
 		}
@@ -181,7 +183,15 @@ func assemble(t *ast.Text, link *linkInfo) ([]byte, []sbPatch, map[string]int, [
 	var out []byte
 	var patches []sbPatch
 	if fi.needSplit {
-		guard, tlsPatch := buildGuard(fi, int32(len(fi.prologue)+bodyLen), int32(fi.guardLen(guardJBlong, guardJBElong)-(9+3+7+2)+len(fi.prologue)+bodyLen))
+		// The JBE ends the guard, so its displacement is the prologue plus
+		// the body; the underflow JB additionally spans the trailing CMPQ and
+		// JBE, whose combined length is guardLen minus the prefix and the
+		// JB's own length (2 short, 6 long).
+		jbLen := 2
+		if guardJBlong {
+			jbLen = 6
+		}
+		guard, tlsPatch := buildGuard(fi, int32(len(fi.prologue)+bodyLen), int32(fi.guardLen(guardJBlong, guardJBElong)-(9+3+7+jbLen)+len(fi.prologue)+bodyLen))
 		out = append(out, guard...)
 		patches = append(patches, tlsPatch)
 	}
@@ -196,6 +206,14 @@ func assemble(t *ast.Text, link *linkInfo) ([]byte, []sbPatch, map[string]int, [
 			spadjStep{guardLen + len(fi.prologue), 8 + fi.size},
 		)
 	}
+	// frameBase is the SP delta the prologue leaves: 8 for the saved base
+	// pointer plus the frame, 0 frameless.  bodyDelta tracks the ADJSP
+	// statements' straight-line sum, so a mid-body step's value is the
+	// frame base plus what the body has opened so far.
+	frameBase, bodyDelta := 0, 0
+	if fi.useFP {
+		frameBase = 8 + fi.size
+	}
 	pos := guardLen + len(fi.prologue)
 	for i, stmt := range t.Body {
 		s, ok := stmt.(*ast.Instr)
@@ -223,6 +241,16 @@ func assemble(t *ast.Text, link *linkInfo) ([]byte, []sbPatch, map[string]int, [
 				ps[k].kind = RelCall
 			}
 		}
+		if strings.ToUpper(s.Mnemonic.Text) == "ADJSP" && len(s.Operands) == 1 && s.Operands[0].Imm.HasVal {
+			// The statement shifted SP mid-body: record the new running
+			// delta as the value in effect from just past the instruction.
+			v := s.Operands[0].Imm.Val
+			if s.Operands[0].Imm.Neg {
+				v = -v
+			}
+			bodyDelta += int(v)
+			steps = append(steps, spadjStep{pos + len(code), frameBase + bodyDelta})
+		}
 		patches = append(patches, ps...)
 		lines = append(lines, LineEntry{Offset: pos, Line: s.Pos().Line})
 		out = append(out, code...)
@@ -337,7 +365,21 @@ func computeFrame(t *ast.Text) frameInfo {
 	if t.Frame != nil && t.Frame.Imm.HasVal {
 		fi.size = int(t.Frame.Imm.Val)
 	}
-	if fi.size > 0 {
+	if fi.size == 0 && hasCall(t) {
+		// The toolchain gives a frameless function containing a CALL an
+		// 8-byte frame for the pushed base pointer: the prologue saves BP
+		// with no stack adjustment, every RET pops it back, FP references
+		// pass one extra slot, and the virtual SP is the hardware SP.
+		fi.size = 8
+		fi.useFP = true
+		// The push is the frame: the saved BP sits at SP+0 and the
+		// return address at SP+8, so arguments begin at SP+16.  Unlike
+		// a SUBQ frame, the 8-byte size must not be added again.
+		fi.fpAdjust = 16
+		fi.spAdjust = 0
+		fi.prologue = []byte{0x55, 0x48, 0x89, 0xE5} // PUSHQ BP; MOVQ SP, BP
+		fi.epilogue = []byte{0x5D}                   // POPQ BP
+	} else if fi.size > 0 {
 		fi.useFP = true
 		fi.fpAdjust = int64(fi.size) + 16 // frame + saved BP + return address
 		fi.spAdjust = int64(fi.size)
@@ -391,6 +433,40 @@ func hasCall(t *ast.Text) bool {
 	return false
 }

+// checkAdjspBalance mirrors the toolchain's push/pop walk: every ADJSP
+// shifts SP away from the entry state and every RET must see the shifts
+// closed.  The assembler's own prologue and epilogue contribute matching
+// deltas on both sides, so the statements' straight-line sum must be zero
+// at each RET; branches do not reset the walk, which runs over the program
+// list in source order.  go tool asm reports an offender as "unbalanced
+// PUSH/POP" (verified against ADJSP $16 before a RET, accepted as a
+// $16/$-16 pair, per-RET rather than per-function).
+func checkAdjspBalance(t *ast.Text) error {
+	delta := 0
+	for _, stmt := range t.Body {
+		in, ok := stmt.(*ast.Instr)
+		if !ok {
+			continue
+		}
+		switch strings.ToUpper(in.Mnemonic.Text) {
+		case "ADJSP":
+			if len(in.Operands) != 1 || !in.Operands[0].Imm.HasVal {
+				continue // reported during emission
+			}
+			v := in.Operands[0].Imm.Val
+			if in.Operands[0].Imm.Neg {
+				v = -v
+			}
+			delta += int(v)
+		case "RET":
+			if delta != 0 {
+				return fmt.Errorf("unbalanced PUSH/POP")
+			}
+		}
+	}
+	return nil
+}
+
 // guardLen returns the byte length of the stack-split guard prefix.  The
 // final conditional branch (JBE, and JB in the big class) is 2 bytes in the
 // short form and 6 in the long form.
@@ -415,16 +491,6 @@ func (fi frameInfo) guardLen(jbLong, jbeLong bool) int {
 	}
 }

-// moreLen returns the byte length of the trailing morestack block: the CALL
-// (always rel32) plus the JMP back to the function start.
-func moreLen(jmpLong bool) int {
-	jmp := 2
-	if jmpLong {
-		jmp = 5
-	}
-	return 5 + jmp
-}
-
 // buildGuard emits the stack-split guard prefix.  jbeDisp and jbDisp are the
 // already-computed displacements of the conditional branches that jump to the
 // morestack block (unused in classes without them).  The TLS load carries a
@@ -518,6 +584,13 @@ func instrSize(s *ast.Instr, fi frameInfo, long bool, link *linkInfo) (int, erro
 		if (mnem == "CALL" || mnem == "JMP") && isSBCall(s) {
 			return 5, nil // opcode + rel32, always the long form
 		}
+		if (mnem == "CALL" || mnem == "JMP") && indirectJumpTarget(s) {
+			code, err := encodeIndirectJump(s, mnem)
+			if err != nil {
+				return 0, err
+			}
+			return len(code), nil
+		}
 		return jumpSize(mnem, long), nil
 	}
 	code, _, err := encodeInstr(s, 0, nil, fi, false, nil, link)
@@ -585,6 +658,15 @@ func encodeInstr(s *ast.Instr, pc int, offsets map[string]int, fi frameInfo, lon
 			}
 			return append(prefix, code...), ps, nil
 		}
+		if (mnem == "CALL" || mnem == "JMP") && indirectJumpTarget(s) {
+			// JMP/CALL through a register or memory: no relocation and no
+			// label to resolve, the operand fully determines the bytes.
+			code, err = encodeIndirectJump(s, mnem)
+			if err != nil {
+				return nil, nil, err
+			}
+			return append(prefix, code...), nil, nil
+		}
 		code, err = encodeJump(s, mnem, pc+len(prefix), offsets, long, resolve)
 	} else {
 		code, ps, err = encodeNormal(s, fi, link)
@@ -706,6 +788,44 @@ func labelName(op *ast.Operand) (string, bool) {
 	return "", false
 }

+// indirectJumpTarget reports whether the JMP/CALL operand addresses a
+// register or a memory location rather than a label or a static symbol.
+// A bare identifier is a register when the register table knows the name and
+// a label otherwise, which is exactly how the parser cannot distinguish them.
+func indirectJumpTarget(s *ast.Instr) bool {
+	if len(s.Operands) != 1 || s.Operands[0].Kind != ast.OpAddr {
+		return false
+	}
+	a := s.Operands[0].Addr
+	if a.Base != "" || a.Index != "" {
+		return true
+	}
+	if a.Sym != nil && a.Sym.Pseudo == "" && a.Sym.Name != "" {
+		if _, ok := ParseReg(a.Sym.Name); ok {
+			return true
+		}
+	}
+	return false
+}
+
+// encodeIndirectJump assembles a JMP/CALL through a register or memory
+// operand, which carries no relocation and no label to resolve.
+func encodeIndirectJump(s *ast.Instr, mnem string) ([]byte, error) {
+	ops := make([]Operand, len(s.Operands))
+	for i, op := range s.Operands {
+		o, err := operandFromAST(op, 8, frameInfo{}, nil)
+		if err != nil {
+			return nil, err
+		}
+		ops[i] = o
+	}
+	e := &enc{}
+	if err := e.encodeIndirectBranch(mnem, ops); err != nil {
+		return nil, err
+	}
+	return e.out, nil
+}
+
 // spReg is the hardware stack pointer used to realise FP/SP pseudo-operands.
 var spReg = Reg{idx: 4, size: 8}

@@ -160,6 +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.
@@ -202,8 +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, `
@@ -353,6 +404,19 @@ TEXT ·pf(SB), NOSPLIT, $0
 	}
 }

+// 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
@@ -375,3 +439,132 @@ func TestSubSPEncodings(t *testing.T) {
 		}
 	}
 }
+
+// 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))
+	}
+}
@@ -43,7 +43,9 @@ const (
 	stInfoShift = 4

 	rX8664PC32 = 2
-	rX8664TPOFF32 = 20
+	// 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.
@@ -219,7 +221,7 @@ func (img *Image) ELFObject() ([]byte, error) {
 		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[:]...)
 		}
@@ -228,15 +230,32 @@ func (img *Image) ELFObject() ([]byte, error) {
 	shstrOff := len(out)
 	out = append(out, stSections.bytes()...)

-	// DWARF debug sections (no relocations, the linker resolves DWARF fixups).
+	// 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, "gasm.s", symIdx, dwAlign)
+	dw := appendDWARFSections(&out, img, dwarfSourceName(img), symIdx, dwAlign, cfiAMD64)
+	dwarfStart := 0 // section index of .debug_abbrev, set when DWARF is present
 	if dw != nil {
-		nSections += 4 // .debug_abbrev, .debug_info, .debug_line, .debug_line_str
+		// 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)
@@ -267,14 +286,37 @@ func (img *Image) ELFObject() ([]byte, error) {
 	}
 	putSh(".shstrtab", shtStrtab, 0, shstrOff, len(stSections.bytes()), 0, 0, 1, 0)

-	// DWARF section headers.
+	// 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)
+			}
 		}
 	}

@@ -12,43 +12,74 @@ import (
 // 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.
+// 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
+	// Abbrev 1: DW_TAG_compile_unit.
 	b = append(b, 1)                 // abbreviation code
-	b = append(b, 0x11)     // DW_TAG_compile_unit
+	b = appendUleb(b, dwTagCompUnit) // DW_TAG_compile_unit
 	b = append(b, 1)                 // DW_CHILDREN_yes
-	b = appendUleb(b, 0x1b) // DW_AT_low_pc
-	b = appendUleb(b, 0x01) // DW_FORM_addr
-	b = appendUleb(b, 0x29) // DW_AT_high_pc
-	b = appendUleb(b, 0x07) // DW_FORM_data8
-	b = appendUleb(b, 0x10) // DW_AT_stmt_list
-	b = appendUleb(b, 0x25) // DW_FORM_sec_offset
-	b = appendUleb(b, 0x01) // DW_AT_name
-	b = appendUleb(b, 0x08) // DW_FORM_string
-	b = appendUleb(b, 0)    // end of attributes
+	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
+	// Abbrev 2: DW_TAG_subprogram.
 	b = append(b, 2)                 // abbreviation code
-	b = append(b, 0x2e)     // DW_TAG_subprogram
+	b = appendUleb(b, dwTagSubprog)  // DW_TAG_subprogram
 	b = append(b, 0)                 // DW_CHILDREN_no
-	b = appendUleb(b, 0x03) // DW_AT_name
-	b = appendUleb(b, 0x08) // DW_FORM_string
-	b = appendUleb(b, 0x11) // DW_AT_low_pc
-	b = appendUleb(b, 0x01) // DW_FORM_addr
-	b = appendUleb(b, 0x29) // DW_AT_high_pc
-	b = appendUleb(b, 0x07) // DW_FORM_data8
-	b = appendUleb(b, 0x3f) // DW_AT_frame_base
-	b = appendUleb(b, 0x18) // DW_FORM_exprloc
-	b = appendUleb(b, 0x3b) // DW_AT_decl_file
-	b = appendUleb(b, 0x0b) // DW_FORM_data1
-	b = appendUleb(b, 0x37) // DW_AT_decl_line
-	b = appendUleb(b, 0x0b) // DW_FORM_data1
-	b = appendUleb(b, 0x63) // DW_AT_external
-	b = appendUleb(b, 0x0b) // DW_FORM_flag
-	b = appendUleb(b, 0)    // end of attributes
+	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)
@@ -68,6 +99,9 @@ type dwarfSections struct {
 	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 {
@@ -76,8 +110,9 @@ type dwarfReloc struct {
 	addend int64
 }

-// emitDWARF generates complete DWARF5 sections for the image.
-func emitDWARF(img *Image, srcFile string) *dwarfSections {
+// 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()

@@ -86,19 +121,21 @@ func emitDWARF(img *Image, srcFile string) *dwarfSections {
 	lineStr.add(srcFile)
 	ds.debugLineStr = lineStr.bytes()

-	// Build .debug_line.
-	ds.debugLine = dwarfBuildLineSection(img, ds)
+	// 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)
+	ds.debugFrame = dwarfBuildFrameSection(img, cfi, ds)
 	return ds
 }

-// dwarfBuildLineSection builds a complete .debug_line section.
-func dwarfBuildLineSection(img *Image, ds *dwarfSections) []byte {
+// 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

@@ -120,15 +157,25 @@ func dwarfBuildLineSection(img *Image, ds *dwarfSections) []byte {
 	// Standard opcode lengths (opcode 1..opcode_base-1).
 	b = append(b, 0, 1, 1, 1, 1, 0, 0, 0, 1, 0)

-	// Directory table (DWARF5 format).
-	b = append(b, 0) // one directory entry (index 0 = empty)
-	// File table.
-	b = appendUleb(b, 1) // file count
-	// File 1: name index into .debug_line_str, dir index, time, size.
-	b = appendUleb(b, 0) // name (index 0 in line_str)
-	b = appendUleb(b, 0) // directory index
-	b = appendUleb(b, 0) // last modification time
-	b = appendUleb(b, 0) // file size
+	// 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)

@@ -176,8 +223,11 @@ func dwarfBuildLineSection(img *Image, ds *dwarfSections) []byte {

 	// Patch unit_length.
 	le.PutUint32(b[headerStart:], uint32(len(b)-headerStart-4))
-	// Patch header_length.
-	le.PutUint32(b[headerStart+6:], uint32(headerEnd-headerStart-10))
+	// 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
 }

@@ -195,14 +245,16 @@ func dwarfBuildInfoSection(img *Image, srcFile string, ds *dwarfSections) []byte

 	// DW_TAG_compile_unit (abbrev 1).
 	b = append(b, 1) // abbreviation code
-	// DW_AT_low_pc: address of .text start.
-	infoRelocBase := len(b)
+	// 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(infoRelocBase),
+			off:  uint64(len(b) - 8),
 			name: img.Funcs[0].Name,
-		addend: 0,
 		})
+	}
 	// DW_AT_high_pc: size of .text.
 	b = le.AppendUint64(b, uint64(len(img.Code)))
 	// DW_AT_stmt_list: offset into .debug_line (0).
@@ -229,8 +281,9 @@ func dwarfBuildInfoSection(img *Image, srcFile string, ds *dwarfSections) []byte
 		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: file index 1.
-		b = append(b, 1)
+		// 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.
@@ -253,15 +306,44 @@ 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 {
-	return binary.AppendVarint(b, v)
+	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) []byte {
+func dwarfBuildFrameSection(img *Image, cfi cfiArch, ds *dwarfSections) []byte {
 	var b []byte
 	le := binary.LittleEndian

@@ -273,10 +355,10 @@ func dwarfBuildFrameSection(img *Image) []byte {
 	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, 16)              // return address register (LR on arm64, RIP on amd64)
+	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, 31) // register: SP (RSP=7 on amd64, SP=31 on arm64)
+	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.
@@ -287,7 +369,12 @@ func dwarfBuildFrameSection(img *Image) []byte {
 		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 (relocated by linker).
+		// 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))
@@ -3,6 +3,17 @@

 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 {
@@ -11,14 +22,22 @@ type dwarfELFSections struct {
 	lineOff, lineSize       int
 	lineStrOff, lineStrSize int
 	frameOff, frameSize     int
-	// Relocations for .debug_info address references.
+	// .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.
+	// 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
+	off    uint64 // offset within the target section
 	sym    int    // symbol index in .symtab
 	addend int64
 }
@@ -29,8 +48,9 @@ type elfDwarfReloc struct {
 //
 // 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.
-func appendDWARFSections(out *[]byte, img *Image, srcFile string, symIdx map[string]int, align func(int)) *dwarfELFSections {
+// 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 {
@@ -45,7 +65,7 @@ func appendDWARFSections(out *[]byte, img *Image, srcFile string, symIdx map[str
 		}
 		nameToIdx[name] = idx
 	}
-	ds := emitDWARF(img, srcFile)
+	ds := emitDWARF(img, srcFile, cfi)
 	if ds == nil || len(ds.debugAbbrev) == 0 {
 		return nil
 	}
@@ -68,15 +88,11 @@ func appendDWARFSections(out *[]byte, img *Image, srcFile string, symIdx map[str
 	align(1)
 	result.lineOff = len(*out)
 	result.lineSize = len(ds.debugLine)
-	lineBase := len(*out)
 	*out = append(*out, ds.debugLine...)
-
-	// Patch .debug_line relocations: replace placeholder addresses with
-	// actual .text offsets via symbol lookup.
 	for _, dr := range ds.lineRelocs {
 		if idx, ok := nameToIdx[dr.name]; ok {
 			result.lineRelocs = append(result.lineRelocs, elfDwarfReloc{
-				off:    uint64(lineBase) + dr.off,
+				off:    dr.off,
 				sym:    idx,
 				addend: dr.addend,
 			})
@@ -87,33 +103,77 @@ func appendDWARFSections(out *[]byte, img *Image, srcFile string, symIdx map[str
 	align(1)
 	result.infoOff = len(*out)
 	result.infoSize = len(ds.debugInfo)
-	infoBase := len(*out)
 	*out = append(*out, ds.debugInfo...)
-
-	// .debug_frame
-	if len(ds.debugFrame) > 0 {
-		align(1)
-		result.frameOff = len(*out)
-		result.frameSize = len(ds.debugFrame)
-		*out = append(*out, ds.debugFrame...)
-	}
-
-	// Patch .debug_info relocations.
 	for _, dr := range ds.infoRelocs {
 		if idx, ok := nameToIdx[dr.name]; ok {
 			result.infoRelocs = append(result.infoRelocs, elfDwarfReloc{
-				off:    uint64(infoBase) + dr.off,
+				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",
 }
@@ -4,11 +4,313 @@
 package asm

 import (
+	"bytes"
+	"encoding/binary"
 	"testing"

 	"sourcedock.dev/petrbalvin/gasm-devkit/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
@@ -27,7 +329,7 @@ TEXT ·add(SB), NOSPLIT, $0-24
 		t.Fatalf("assemble: %v", err)
 	}

-	ds := emitDWARF(img, "test_amd64.s")
+	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 {
@@ -68,14 +370,3 @@ TEXT ·add(SB), NOSPLIT, $0-24
 		t.Fatal("no .debug_info relocations")
 	}
 }
-
-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])
-	}
-}
@@ -12,6 +12,7 @@ import (
 	"path/filepath"
 	"testing"

+	"sourcedock.dev/petrbalvin/gasm-devkit/ast"
 	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
 )

@@ -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)
@@ -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 +609,144 @@ 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)
+	}
 }
@@ -81,10 +81,15 @@ 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         → R_AARCH64_ADD_ABS_LO12_NC
-	//   LDR/STR X   → R_AARCH64_LDST64_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
@@ -97,28 +102,34 @@ 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)
 			}
-			var typ uint32
-			switch {
-			case r.Kind == RelArm64Branch:
-				typ = rArm64Call26
-			case r.Kind == RelArm64LDST64 && r.Off%4 == 4:
-				typ = rArm64Ldst64Lo12NC
-			case r.Kind == RelArm64Addr && r.Off%4 == 4:
-				typ = rArm64AddAbsLo12NC
-			default:
-				typ = rArm64PrelPgHi21
-			}
+			switch r.Kind {
+			case RelArm64Branch:
 				relas = append(relas, elfRela{
-				off:    uint64(fn.Offset + r.Off),
-				typ:    typ,
-				sym:    idx,
-				addend: r.Addend,
+					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)
+			}
 		}
 	}

@@ -193,15 +204,32 @@ func (img *Image) ELFAARCH64Object() ([]byte, error) {
 	shstrOff := len(out)
 	out = append(out, stSections.bytes()...)

-	// DWARF debug sections.
+	// 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, "gasm.s", symIdx, dwAlign)
+	dw := appendDWARFSections(&out, img, dwarfSourceName(img), symIdx, dwAlign, cfiARM64)
+	dwarfStart := 0 // section index of .debug_abbrev, set when DWARF is present
 	if dw != nil {
-		nSections += 4
+		// 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)
@@ -230,13 +258,37 @@ func (img *Image) ELFAARCH64Object() ([]byte, error) {
 		putSh(".rela.text", shtRela, 0, relaOff, 24*len(relas), secSymtab, secText, 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)
+			}
 		}
 	}

@@ -6,6 +6,7 @@ package asm
 import (
 	"bytes"
 	"debug/elf"
+	"encoding/binary"
 	"testing"

 	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
@@ -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
@@ -13,6 +13,12 @@ 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)
@@ -187,9 +193,25 @@ func (img *Image) ELFLOONG64Object() ([]byte, error) {
 			out = append(out, 0)
 		}
 	}
-	dw := appendDWARFSections(&out, img, "gasm.s", symIdx, dwAlign)
+	dw := appendDWARFSections(&out, img, dwarfSourceName(img), symIdx, dwAlign, cfiLOONG64)
+	dwarfStart := 0 // section index of .debug_abbrev, set when DWARF is present
 	if dw != nil {
-		nSections += 4
+		// 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)
@@ -218,13 +240,37 @@ func (img *Image) ELFLOONG64Object() ([]byte, error) {
 		putSh(".rela.text", shtRela, 0, relaOff, 24*len(relas), secSymtab, secText, 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)
+			}
 		}
 	}

@@ -237,7 +283,7 @@ 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)
@@ -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")
@@ -13,8 +13,13 @@ 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
@@ -83,9 +88,14 @@ 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
@@ -99,21 +109,20 @@ 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)
 			}
@@ -196,9 +205,25 @@ func (img *Image) ELFRISCVObject() ([]byte, error) {
 			out = append(out, 0)
 		}
 	}
-	dw := appendDWARFSections(&out, img, "gasm.s", symIdx, dwAlign)
+	dw := appendDWARFSections(&out, img, dwarfSourceName(img), symIdx, dwAlign, cfiRISCV64)
+	dwarfStart := 0 // section index of .debug_abbrev, set when DWARF is present
 	if dw != nil {
-		nSections += 4
+		// 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)
@@ -227,13 +252,37 @@ func (img *Image) ELFRISCVObject() ([]byte, error) {
 		putSh(".rela.text", shtRela, 0, relaOff, 24*len(relas), secSymtab, secText, 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)
+			}
 		}
 	}

@@ -246,7 +295,7 @@ 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)
@@ -18,7 +18,14 @@ func Encodable(mnemonic string) bool {

 	// Fixed-name instructions (no size suffix).
 	switch upper {
-	case "RET", "NOP", "CALL", "JMP":
+	case "RET", "NOP", "CALL", "JMP",
+		"POPFQ", "PUSHFQ", "INT", "LDMXCSR", "STMXCSR", "CMPSD", "SHA256RNDS2",
+		// The literal-data pseudo-ops, the accepted-and-ignored END and the
+		// SP adjust.
+		"BYTE", "WORD", "LONG", "QUAD", "END", "ADJSP":
+		return true
+	}
+	if _, ok := noOperandTable[upper]; ok {
 		return true
 	}
 	if _, ok := condCode(upper); ok {
@@ -31,30 +38,50 @@ func Encodable(mnemonic string) bool {
 		return false
 	}
 	if isVex(base) || isEvex(base) || isKOp(base) || isGather(base) || isScatter(base) ||
-		base == "KMOVW" || base == "KMOVQ" {
+		base == "KMOVW" || base == "KMOVQ" || base == "KMOVB" || base == "KMOVD" {
 		return true
 	}

 	// CMOV carries size then condition (CMOVLGT); SET carries the condition
-	// alone (SETNE).
+	// 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.
+	// 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.
@@ -69,20 +96,32 @@ func Encodable(mnemonic string) bool {
 		}
 	}
 	switch base2 {
-	case "MOV",
-		"ADD", "SUB", "AND", "OR", "XOR", "CMP",
+	case "MOV", "MOVD",
+		"ADD", "SUB", "AND", "OR", "XOR", "CMP", "ADC", "SBB",
 		"TEST",
 		"LEA",
-		"INC", "DEC", "NEG", "NOT",
-		"SHL", "SHR", "SAR",
+		"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",
-		"MOVOU", "MOVO", "MOVUPS", "MOVAPS", "MOVUPD", "MOVAPD", "MOVSD", "MOVSS":
+		"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
@@ -40,14 +40,69 @@ 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)
+	}

 	// 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 +112,8 @@ 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) ||
+		base == "KMOVW" || base == "KMOVQ" || base == "KMOVB" || base == "KMOVD" {
 		return e.encodeVec(base, ops, sfx)
 	}
 	if sfx.any() {
@@ -91,36 +147,81 @@ func (e *enc) encode(mnem string, ops []Operand) error {
 	if m, ok := sseBinTable[base]; ok {
 		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)
+		return e.encodePushPop(ops, size, false)
 	case "BSF", "BSR", "LZCNT", "TZCNT", "POPCNT":
 		return e.encodeCount(base, ops, size)
 	case "BSWAP":
 		return e.encodeBswap(ops, size)
 	case "PREFETCHNTA", "PREFETCHT0", "PREFETCHT1", "PREFETCHT2":
 		return e.encodePrefetch(base, ops)
-	case "MOVBLZX", "MOVBQZX", "MOVWLZX", "MOVWQZX", "MOVWLSX", "MOVLQSX":
+	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 "MOVOU", "MOVO", "MOVOA", "MOVUPS", "MOVAPS", "MOVUPD", "MOVAPD", "MOVSD", "MOVSS":
 		return e.encodeSSEMove(sseMoveTable[base], ops)
 	}
 	return fmt.Errorf("unsupported instruction %q", mnem)
@@ -150,6 +251,73 @@ var prefetchVariant = map[string]int{
 	"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
+}
+
+// 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
+}
+
 // splitSize separates a trailing B/W/L/Q size suffix from the mnemonic.
 func splitSize(upper string) (base string, size int) {
 	if upper == "" {
@@ -179,7 +347,7 @@ 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)
 		}
@@ -335,6 +503,12 @@ func setMem(i *instr, regField int, m Mem) error {
 // 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
@@ -149,6 +149,45 @@ func TestPushPop(t *testing.T) {
 	checkSyntax(t, "push rbx", "PUSHQ", BX)
 	checkSyntax(t, "pop r12", "POPQ", Reg{idx: 12, size: 8})
 	checkSyntax(t, "push 0x5", "PUSHQ", Imm(5))
+	// The W spelling carries the 0x66 operand-size prefix, byte for byte
+	// with go tool asm; the L and B spellings are illegal in 64-bit mode
+	// there and rejected here rather than silently widened.
+	cases := []struct {
+		name string
+		mnem string
+		ops  []Operand
+		want string
+	}{
+		{"PUSHW AX", "PUSHW", []Operand{AX}, "6650"},
+		{"POPW AX", "POPW", []Operand{AX}, "6658"},
+		{"PUSHW $5", "PUSHW", []Operand{Imm(5)}, "666a05"},
+		{"PUSHW (AX)", "PUSHW", []Operand{Ptr(AX, 0, 2)}, "66ff30"},
+		{"PUSHQ AX", "PUSHQ", []Operand{AX}, "50"},
+	}
+	for _, c := range cases {
+		code, err := Encode(c.mnem, c.ops...)
+		if err != nil {
+			t.Errorf("%s: %v", c.name, err)
+			continue
+		}
+		if got := fmt.Sprintf("%x", code); got != c.want {
+			t.Errorf("%s: bytes %s, want %s", c.name, got, c.want)
+		}
+	}
+	for _, c := range []struct {
+		name string
+		mnem string
+		ops  []Operand
+	}{
+		{"PUSHL AX", "PUSHL", []Operand{AX}},
+		{"PUSHL R8", "PUSHL", []Operand{Reg{idx: 8, size: 8}}},
+		{"POPL BX", "POPL", []Operand{BX}},
+		{"PUSHB AX", "PUSHB", []Operand{AX}},
+	} {
+		if _, err := Encode(c.mnem, c.ops...); err == nil {
+			t.Errorf("%s: expected an error, got none", c.name)
+		}
+	}
 }

 func TestUnary(t *testing.T) {
@@ -161,10 +200,60 @@ func TestUnary(t *testing.T) {
 func TestShift(t *testing.T) {
 	checkSyntax(t, "shl rdx, 0x2", "SHLQ", Imm(2), DX)
 	checkSyntax(t, "shl rdx, cl", "SHLQ", CL, DX)
+	checkSyntax(t, "shl rdx, cl", "SHLQ", CX, DX)
 	checkSyntax(t, "shl rdx, 0x1", "SHLQ", Imm(1), DX)
 	checkSyntax(t, "sar rcx, 0x1f", "SARQ", Imm(31), CX)
 }

+// TestDoubleShift pins the three-operand SHL/SHR form, which encodes as
+// SHLD/SHRD: go tool asm accepts it for SHL/SHR at W/L/Q widths and rejects
+// it for SAR, SAL, the rotates and the B width.  The byte pins mirror the
+// oracle's objdump output (48 0f a4 fe 0d for the first case, and so on).
+func TestDoubleShift(t *testing.T) {
+	cases := []struct {
+		name string
+		mnem string
+		ops  []Operand
+		want string // hex encoding
+	}{
+		{"SHLQ imm", "SHLQ", []Operand{Imm(0x0d), DI, SI}, "480fa4fe0d"},
+		{"SHLQ CX high regs", "SHLQ", []Operand{CX, Reg{idx: 8, size: 8}, Reg{idx: 9, size: 8}}, "4d0fa5c1"},
+		{"SHRQ imm", "SHRQ", []Operand{Imm(1), AX, CX}, "480facc101"},
+		{"SHLW imm", "SHLW", []Operand{Imm(1), AX, CX}, "660fa4c101"},
+		{"SHRD CL", "SHRQ", []Operand{CL, AX, CX}, "480fadc1"},
+		{"SHLD imm high regs", "SHLQ", []Operand{Imm(2), Reg{idx: 10, size: 8}, Reg{idx: 11, size: 8}}, "4d0fa4d302"},
+		{"SHRD imm max", "SHRQ", []Operand{Imm(63), Reg{idx: 9, size: 8}, Reg{idx: 15, size: 8}}, "4d0faccf3f"},
+	}
+	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: bytes %s, want %s", c.name, got, c.want)
+		}
+	}
+	// Rejected forms: the oracle rejects every one of these.
+	rejected := []struct {
+		name string
+		mnem string
+		ops  []Operand
+	}{
+		{"SARQ three operands", "SARQ", []Operand{Imm(1), AX, CX}},
+		{"SALQ three operands", "SALQ", []Operand{Imm(1), AX, CX}},
+		{"ROLQ three operands", "ROLQ", []Operand{Imm(1), AX, CX}},
+		{"SHLB three operands", "SHLB", []Operand{Imm(1), AL, CL}},
+		{"SHRQ memory source", "SHRQ", []Operand{Imm(1), Ptr(AX, 0, 8), CX}},
+		{"SHRQ ECX count", "SHRQ", []Operand{Reg{idx: 1, size: 4}, AX, CX}},
+	}
+	for _, c := range rejected {
+		if _, err := Encode(c.mnem, c.ops...); err == nil {
+			t.Errorf("%s: Encode succeeded, want rejection", c.name)
+		}
+	}
+}
+
 func TestImul(t *testing.T) {
 	checkSyntax(t, "imul rdx, rcx", "IMULQ", CX, DX)
 	checkSyntax(t, "imul edx, edx, 0x3", "IMULL", Imm(3), DX, DX)
@@ -181,6 +270,39 @@ func TestControl(t *testing.T) {
 	checkOp(t, x86asm.JBE, "JLS", Imm(0))
 }

+// TestIndirectControlFlow pins the indirect JMP/CALL forms: FF /4 for JMP and
+// FF /2 for CALL through a register or memory.  A REX appears only for the
+// extended registers, never REX.W: the branch operand size is fixed at 64
+// bits in long mode.
+func TestIndirectControlFlow(t *testing.T) {
+	cases := []struct {
+		name string
+		mnem string
+		ops  []Operand
+		want string
+	}{
+		{"JMP AX", "JMP", []Operand{AX}, "ffe0"},
+		{"CALL AX", "CALL", []Operand{AX}, "ffd0"},
+		{"JMP (BX)", "JMP", []Operand{Ptr(BX, 0, 8)}, "ff23"},
+		{"CALL (BX)", "CALL", []Operand{Ptr(BX, 0, 8)}, "ff13"},
+		{"JMP 8(BX)", "JMP", []Operand{Ptr(BX, 8, 8)}, "ff6308"},
+		{"CALL -16(BX)", "CALL", []Operand{Ptr(BX, -16, 8)}, "ff53f0"},
+		{"JMP R8", "JMP", []Operand{Reg{idx: 8, size: 2}}, "41ffe0"},
+		{"CALL R9", "CALL", []Operand{Reg{idx: 9, size: 2}}, "41ffd1"},
+		{"JMP R15", "JMP", []Operand{Reg{idx: 15, size: 2}}, "41ffe7"},
+	}
+	for _, c := range cases {
+		code, err := Encode(c.mnem, c.ops...)
+		if err != nil {
+			t.Errorf("%s: %v", c.name, err)
+			continue
+		}
+		if got := fmt.Sprintf("%x", code); got != c.want {
+			t.Errorf("%s: got %s, want %s", c.name, got, c.want)
+		}
+	}
+}
+
 // TestSSEMoveGroundTruth checks the legacy (non-VEX) SSE moves byte for byte
 // against the Go assembler.  wantOp is the decoder's name, which differs from
 // the Plan 9 spelling for the octa moves (MOVOU = MOVDQU, MOVO = MOVDQA).
@@ -205,6 +327,12 @@ func TestSSEMoveGroundTruth(t *testing.T) {
 		{"MOVSD (SI),X1", "MOVSD", []Operand{Ptr(SI, 0, 8), vreg(t, "X1")}, "f20f100e", "MOVSD_XMM"},
 		{"MOVSD X1,X2", "MOVSD", []Operand{vreg(t, "X1"), vreg(t, "X2")}, "f20f10d1", "MOVSD_XMM"},
 		{"MOVSS X3,(DI)", "MOVSS", []Operand{vreg(t, "X3"), Ptr(DI, 0, 4)}, "f30f111f", "MOVSS"},
+		// Static-symbol (SB) references: the GOROOT crypto kernels load and
+		// store octa constants by name (MOVOU bswapMask<>+0(SB), X0).
+		{"MOVOU sym,X0", "MOVOU", []Operand{sbMem{size: 16, name: "bswapMask"}, vreg(t, "X0")}, "f30f6f0500000000", "MOVDQU"},
+		{"MOVOU X0,sym+8", "MOVOU", []Operand{vreg(t, "X0"), sbMem{size: 16, name: "bswapMask", addend: 8}}, "f30f7f0500000000", "MOVDQU"},
+		{"MOVO sym,X1", "MOVO", []Operand{sbMem{size: 16, name: "gcmPoly"}, vreg(t, "X1")}, "660f6f0d00000000", "MOVDQA"},
+		{"MOVO X2,sym", "MOVO", []Operand{vreg(t, "X2"), sbMem{size: 16, name: "gcmPoly"}}, "660f7f1500000000", "MOVDQA"},
 	}
 	for _, c := range cases {
 		code, err := Encode(c.mnem, c.ops...)
@@ -230,7 +358,7 @@ func TestSSEMoveGroundTruth(t *testing.T) {
 // TestGoFlacScalarTail encodes the scalar tail of an analyze kernel to confirm
 // the encoder handles a realistic instruction sequence.
 func TestGoFlacScalarTail(t *testing.T) {
-	// MOVQ swin_base+0(FP), SI — modelled as MOVQ disp(reg), reg.
+	// MOVQ swin_base+0(FP), SI; modelled as MOVQ disp(reg), reg.
 	checkSyntax(t, "mov rsi, qword ptr [rax+0x10]", "MOVQ", Ptr(AX, 0x10, 8), SI)
 	checkSyntax(t, "lea r9, ptr [rsi+4*rbx]", "LEAQ", Idx(SI, BX, 4, 0, 8), Reg{idx: 9, size: 8})
 	checkSyntax(t, "and r10, -0x8", "ANDQ", Imm(-8), Reg{idx: 10, size: 8})
@@ -285,6 +413,18 @@ func TestScalarGroundTruth(t *testing.T) {
 		{"MOVBQZX AL,R8", "MOVBQZX", []Operand{AL, r8}, "4c0fb6c0", "MOVZX"},
 		{"MOVWLZX AX,CX", "MOVWLZX", []Operand{AX, CX}, "0fb7c8", "MOVZX"},
 		{"MOVWQZX AX,R8", "MOVWQZX", []Operand{AX, r8}, "4c0fb7c0", "MOVZX"},
+		// The width pairs the toolchain accepts and GOROOT uses; bytes
+		// pinned from go tool asm (see testdata/verify/widen_amd64.s).
+		{"MOVBWZX (BX),R11W", "MOVBWZX", []Operand{Ptr(BX, 0, 1), Reg{idx: 11, size: 2}}, "66440fb61b", "MOVZX"},
+		{"MOVBWSX (BX),R11W", "MOVBWSX", []Operand{Ptr(BX, 0, 1), Reg{idx: 11, size: 2}}, "66440fbe1b", "MOVSX"},
+		{"MOVBLSX (BX),AX", "MOVBLSX", []Operand{Ptr(BX, 0, 1), AX}, "0fbe03", "MOVSX"},
+		{"MOVBQSX (BX),R8", "MOVBQSX", []Operand{Ptr(BX, 0, 1), r8}, "4c0fbe03", "MOVSX"},
+		{"MOVWQSX (BX),R9", "MOVWQSX", []Operand{Ptr(BX, 0, 2), r9}, "4c0fbf0b", "MOVSX"},
+		// A long to quad zero-extend is a plain 32-bit move.
+		{"MOVLQZX (BX),DX", "MOVLQZX", []Operand{Ptr(BX, 0, 4), DX}, "8b13", "MOV"},
+		{"MOVLQZX AX,DX", "MOVLQZX", []Operand{AX, DX}, "8bd0", "MOV"},
+		{"PMOVMSKB X1,AX", "PMOVMSKB", []Operand{vreg(t, "X1"), AX}, "660fd7c1", "PMOVMSKB"},
+		{"PMOVMSKB X11,CX", "PMOVMSKB", []Operand{vreg(t, "X11"), CX}, "66410fd7cb", "PMOVMSKB"},
 		{"CVTSL2SD R8,X13", "CVTSL2SD", []Operand{r8, vreg(t, "X13")}, "f2450f2ae8", "CVTSI2SD"},
 		{"CVTSL2SD AX,X0", "CVTSL2SD", []Operand{AX, vreg(t, "X0")}, "f20f2ac0", "CVTSI2SD"},
 		{"CVTSQ2SD R8,X13", "CVTSQ2SD", []Operand{r8, vreg(t, "X13")}, "f24d0f2ae8", "CVTSI2SD"},
@@ -364,6 +504,346 @@ func TestScalarErrors(t *testing.T) {
 	}
 }

+// TestImmediateOutOfRange pins the go-tool-asm parity of the immediate and
+// displacement spans: a scalar immediate must fit a signed or unsigned 32-bit
+// word (only MOVQ reg, $imm takes the full int64), a scalar shift count must
+// be an unsigned byte, and a displacement must fit int32.  Every rejected
+// shape here is rejected by `go tool asm` too; every accepted one encodes the
+// same bytes.
+func TestImmediateOutOfRange(t *testing.T) {
+	cases := []struct {
+		name string
+		mnem string
+		ops  []Operand
+	}{
+		{"SHLQ count 300", "SHLQ", []Operand{Imm(300), AX}},
+		{"SHLQ count -1", "SHLQ", []Operand{Imm(-1), AX}},
+		{"SHLW count 256", "SHLW", []Operand{Imm(256), DX}},
+		{"SHLB count 300", "SHLB", []Operand{Imm(300), BL}},
+		{"MOVL imm32+", "MOVL", []Operand{Imm(4294967296), AX}},
+		{"MOVL imm32-", "MOVL", []Operand{Imm(-2147483649), AX}},
+		{"MOVW imm32+", "MOVW", []Operand{Imm(4294967296), AX}},
+		{"MOVB imm32+", "MOVB", []Operand{Imm(4294967296), AL}},
+		{"ADDB imm32+", "ADDB", []Operand{Imm(4294967296), AL}},
+		{"ADDL imm32+", "ADDL", []Operand{Imm(4294967296), AX}},
+		{"ADDQ imm32+", "ADDQ", []Operand{Imm(8589934592), AX}},
+		{"CMPQ imm32+", "CMPQ", []Operand{AX, Imm(4294967296)}},
+		{"CMPQ imm32-", "CMPQ", []Operand{AX, Imm(-2147483649)}},
+		{"TESTL imm32+", "TESTL", []Operand{Imm(4294967296), AX}},
+		{"IMUL3L imm32+", "IMUL3L", []Operand{Imm(4294967296), CX, DX}},
+		{"PUSHQ imm32+", "PUSHQ", []Operand{Imm(4294967296)}},
+		{"MOVQ mem imm32+", "MOVQ", []Operand{Imm(4294967296), Ptr(AX, 0, 8)}},
+		{"disp32+", "MOVQ", []Operand{Ptr(AX, 4294967296, 8), BX}},
+		{"disp32+ max", "MOVQ", []Operand{Ptr(AX, 2147483648, 8), BX}},
+		{"disp32-", "MOVQ", []Operand{Ptr(AX, -2147483649, 8), BX}},
+		{"VEX disp32+", "VMOVDQU", []Operand{Ptr(AX, 4294967296, 32), vreg(t, "Y1")}},
+		{"EVEX disp32+", "VMOVDQU32", []Operand{Ptr(AX, 4294967296, 64), vreg(t, "Z1")}},
+	}
+	for _, c := range cases {
+		if _, err := Encode(c.mnem, c.ops...); err == nil {
+			t.Errorf("%s: expected an error, got none", c.name)
+		}
+	}
+}
+
+// TestImmediateTruncation pins the toolchain-matching truncations inside the
+// accepted 32-bit span: the narrower fields take the low bits silently, byte
+// for byte with `go tool asm` (which rejects none of these).
+func TestImmediateTruncation(t *testing.T) {
+	cases := []struct {
+		name string
+		mnem string
+		ops  []Operand
+		want string
+	}{
+		{"ADDB $256,BL", "ADDB", []Operand{Imm(256), BL}, "80c300"},
+		{"ADDB $1000,BL", "ADDB", []Operand{Imm(1000), BL}, "80c3e8"},
+		{"MOVB $256,AL", "MOVB", []Operand{Imm(256), AL}, "b000"},
+		{"MOVB $-129,AL", "MOVB", []Operand{Imm(-129), AL}, "b07f"},
+		{"MOVW $65536,AX", "MOVW", []Operand{Imm(65536), AX}, "66b80000"},
+		{"MOVW $65535,AX", "MOVW", []Operand{Imm(65535), AX}, "66b8ffff"},
+		{"MOVW $-32769,AX", "MOVW", []Operand{Imm(-32769), AX}, "66b8ff7f"},
+		{"MOVL $4294967295,AX", "MOVL", []Operand{Imm(4294967295), AX}, "b8ffffffff"},
+		{"ADDQ $4294967295,AX", "ADDQ", []Operand{Imm(4294967295), AX}, "4805ffffffff"},
+		{"CMPB BL,$255", "CMPB", []Operand{BL, Imm(255)}, "80fbff"},
+		{"CMPQ AX,$4294967295", "CMPQ", []Operand{AX, Imm(4294967295)}, "483dffffffff"},
+		{"MOVQ $4294967295,0(AX)", "MOVQ", []Operand{Imm(4294967295), Ptr(AX, 0, 8)}, "48c700ffffffff"},
+		{"SHLQ $255,AX", "SHLQ", []Operand{Imm(255), AX}, "48c1e0ff"},
+		{"SHLQ $0,AX", "SHLQ", []Operand{Imm(0), AX}, "48c1e000"},
+		// The one form beyond the 32-bit span: the imm64 MOVQ register move.
+		{"MOVQ $4294967296,AX", "MOVQ", []Operand{Imm(4294967296), AX}, "48b80000000001000000"},
+		{"MOVQ disp32 max", "MOVQ", []Operand{Ptr(AX, 2147483647, 8), BX}, "488b98ffffff7f"},
+	}
+	for _, c := range cases {
+		code, err := Encode(c.mnem, c.ops...)
+		if err != nil {
+			t.Errorf("%s: %v", c.name, err)
+			continue
+		}
+		if got := fmt.Sprintf("%x", code); got != c.want {
+			t.Errorf("%s: bytes %s, want %s", c.name, got, c.want)
+		}
+	}
+}
+
+// TestEncodableCmovSize pins the linter contract for CMOVcc: Encodable must
+// reject the spellings Encode rejects, so a mnemonic like CMOVBGT (no size
+// letter) is not reported as encodable.
+func TestEncodableCmovSize(t *testing.T) {
+	for _, m := range []string{"CMOVBGT", "CMOVXEQ", "CMOVB", "CMOV", "CMOVWXX"} {
+		if Encodable(m) {
+			t.Errorf("Encodable(%q) = true, want false", m)
+		}
+	}
+	for _, m := range []string{"CMOVLGT", "CMOVQGT", "CMOVWLS", "CMOVLEQ"} {
+		if !Encodable(m) {
+			t.Errorf("Encodable(%q) = false, want true", m)
+		}
+	}
+}
+
+// TestCarryShiftMulGroundTruth pins the carry-flag ALU family (ADC/SBB with
+// their accumulator immediate forms), the rotate family, MUL/DIV/IDIV and the
+// bit-test family byte for byte against go tool asm (see
+// testdata/verify/scalar_amd64.s).
+func TestCarryShiftMulGroundTruth(t *testing.T) {
+	cases := []struct {
+		name string
+		mnem string
+		ops  []Operand
+		want string
+	}{
+		{"ADCQ AX,BX", "ADCQ", []Operand{AX, BX}, "4811c3"},
+		{"ADCL AX,BX", "ADCL", []Operand{AX, BX}, "11c3"},
+		{"ADCB AL,BL", "ADCB", []Operand{AL, BL}, "10c3"},
+		{"ADCW AX,BX", "ADCW", []Operand{AX, BX}, "6611c3"},
+		{"SBBQ AX,BX", "SBBQ", []Operand{AX, BX}, "4819c3"},
+		{"ADCQ $5,BX", "ADCQ", []Operand{Imm(5), BX}, "4883d305"},
+		{"ADCQ $300,BX", "ADCQ", []Operand{Imm(300), BX}, "4881d32c010000"},
+		{"ADCQ $300,AX", "ADCQ", []Operand{Imm(300), AX}, "48152c010000"},
+		{"ADCB $5,AL", "ADCB", []Operand{Imm(5), AL}, "1405"},
+		{"SBBQ $300,AX", "SBBQ", []Operand{Imm(300), AX}, "481d2c010000"},
+		{"ADCQ AX,(BX)", "ADCQ", []Operand{AX, Ptr(BX, 0, 8)}, "481103"},
+		{"ROLQ $3,AX", "ROLQ", []Operand{Imm(3), AX}, "48c1c003"},
+		{"ROLL CX,BX", "ROLL", []Operand{CL, BX}, "d3c3"},
+		{"RORQ CL,AX", "RORQ", []Operand{CL, AX}, "48d3c8"},
+		{"RCRQ $1,BX", "RCRQ", []Operand{Imm(1), BX}, "48d1db"},
+		{"RCLQ $3,AX", "RCLQ", []Operand{Imm(3), AX}, "48c1d003"},
+		{"RORB CL,BL", "RORB", []Operand{CL, BL}, "d2cb"},
+		{"SALQ $2,AX", "SALQ", []Operand{Imm(2), AX}, "48c1e002"},
+		{"ROLW $1,AX", "ROLW", []Operand{Imm(1), AX}, "66d1c0"},
+		{"MULQ CX", "MULQ", []Operand{CX}, "48f7e1"},
+		{"MULL CX", "MULL", []Operand{CX}, "f7e1"},
+		{"MULB CL", "MULB", []Operand{CL}, "f6e1"},
+		{"DIVL CX", "DIVL", []Operand{CX}, "f7f1"},
+		{"IDIVQ CX", "IDIVQ", []Operand{CX}, "48f7f9"},
+		{"MULW CX", "MULW", []Operand{CX}, "66f7e1"},
+		{"BTQ AX,DX", "BTQ", []Operand{AX, DX}, "480fa3c2"},
+		{"BTL AX,DX", "BTL", []Operand{AX, DX}, "0fa3c2"},
+		{"BTW AX,DX", "BTW", []Operand{AX, DX}, "660fa3c2"},
+		{"BTQ $3,BX", "BTQ", []Operand{Imm(3), BX}, "480fbae303"},
+		{"BTQ $3,(AX)", "BTQ", []Operand{Imm(3), Ptr(AX, 0, 8)}, "480fba2003"},
+		{"BTSQ $5,BX", "BTSQ", []Operand{Imm(5), BX}, "480fbaeb05"},
+		{"BTCQ AX,BX", "BTCQ", []Operand{AX, BX}, "480fbbc3"},
+		{"BTRQ $7,BX", "BTRQ", []Operand{Imm(7), BX}, "480fbaf307"},
+	}
+	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 := fmt.Sprintf("%x", code); got != c.want {
+			t.Errorf("%s = %s, want %s", c.name, got, c.want)
+		}
+	}
+	// The bit-test immediate is an unsigned bit index with the negative
+	// spelling accepted, the shuffle convention: BTQ $300 must be rejected.
+	if _, err := Encode("BTQ", Imm(300), AX); err == nil {
+		t.Errorf("BTQ $300: expected an error, got none")
+	}
+}
+
+// TestAtomicSystemGroundTruth pins the exchange/compare-exchange/accumulate
+// family, the string primitives, the flag and system instructions, the MXCSR
+// pair, the scalar float-to-int conversions and the x87 FMOVD byte for byte
+// against go tool asm (see testdata/verify/atomics_amd64.s and
+// testdata/verify/system_amd64.s).
+func TestAtomicSystemGroundTruth(t *testing.T) {
+	r8 := Reg{idx: 8, size: 8}
+	cases := []struct {
+		name string
+		mnem string
+		ops  []Operand
+		want string
+	}{
+		{"XCHGQ AX,BX", "XCHGQ", []Operand{AX, BX}, "4893"},
+		{"XCHGQ BX,AX", "XCHGQ", []Operand{BX, AX}, "4893"},
+		{"XCHGL AX,BX", "XCHGL", []Operand{AX, BX}, "93"},
+		{"XCHGB AL,BL", "XCHGB", []Operand{AL, BL}, "86c3"},
+		{"XCHGW AX,BX", "XCHGW", []Operand{AX, BX}, "6693"},
+		{"XCHGQ R8,R9", "XCHGQ", []Operand{r8, Reg{idx: 9, size: 8}}, "4d87c1"},
+		{"XCHGQ BX,(AX)", "XCHGQ", []Operand{BX, Ptr(AX, 0, 8)}, "488718"},
+		{"XCHGQ (AX),BX", "XCHGQ", []Operand{Ptr(AX, 0, 8), BX}, "488718"},
+		{"XCHGQ AX,(BX)", "XCHGQ", []Operand{AX, Ptr(BX, 0, 8)}, "488703"},
+		{"CMPXCHGL AX,BX", "CMPXCHGL", []Operand{AX, BX}, "0fb1c3"},
+		{"CMPXCHGQ AX,(BX)", "CMPXCHGQ", []Operand{AX, Ptr(BX, 0, 8)}, "480fb103"},
+		{"CMPXCHGB AL,(BX)", "CMPXCHGB", []Operand{AL, Ptr(BX, 0, 1)}, "0fb003"},
+		{"CMPXCHGW AX,BX", "CMPXCHGW", []Operand{AX, BX}, "660fb1c3"},
+		{"XADDL AX,BX", "XADDL", []Operand{AX, BX}, "0fc1c3"},
+		{"XADDQ AX,(BX)", "XADDQ", []Operand{AX, Ptr(BX, 0, 8)}, "480fc103"},
+		{"XADDB AL,(BX)", "XADDB", []Operand{AL, Ptr(BX, 0, 1)}, "0fc003"},
+		{"XADDW AX,BX", "XADDW", []Operand{AX, BX}, "660fc1c3"},
+		{"ADCXL AX,CX", "ADCXL", []Operand{AX, CX}, "660f38f6c8"},
+		{"ADCXQ AX,CX", "ADCXQ", []Operand{AX, CX}, "66480f38f6c8"},
+		{"ADOXL AX,CX", "ADOXL", []Operand{AX, CX}, "f30f38f6c8"},
+		{"ADOXQ AX,CX", "ADOXQ", []Operand{AX, CX}, "f3480f38f6c8"},
+		{"CRC32B AX,CX", "CRC32B", []Operand{AX, CX}, "f20f38f0c8"},
+		{"CRC32W AX,CX", "CRC32W", []Operand{AX, CX}, "66f20f38f1c8"},
+		{"CRC32L AX,CX", "CRC32L", []Operand{AX, CX}, "f20f38f1c8"},
+		{"CRC32Q AX,CX", "CRC32Q", []Operand{AX, CX}, "f2480f38f1c8"},
+		{"CRC32L (AX),CX", "CRC32L", []Operand{Ptr(AX, 0, 4), CX}, "f20f38f108"},
+		{"MOVSQ", "MOVSQ", []Operand{}, "48a5"},
+		{"MOVSL", "MOVSL", []Operand{}, "a5"},
+		{"MOVSB", "MOVSB", []Operand{}, "a4"},
+		{"MOVSW", "MOVSW", []Operand{}, "66a5"},
+		{"STOSB", "STOSB", []Operand{}, "aa"},
+		{"STOSQ", "STOSQ", []Operand{}, "48ab"},
+		{"STOSL", "STOSL", []Operand{}, "ab"},
+		{"STOSW", "STOSW", []Operand{}, "66ab"},
+		{"CLD", "CLD", []Operand{}, "fc"},
+		{"STD", "STD", []Operand{}, "fd"},
+		{"POPFQ", "POPFQ", []Operand{}, "9d"},
+		{"PUSHFQ", "PUSHFQ", []Operand{}, "9c"},
+		{"CPUID", "CPUID", []Operand{}, "0fa2"},
+		{"RDTSC", "RDTSC", []Operand{}, "0f31"},
+		{"RDTSCP", "RDTSCP", []Operand{}, "0f01f9"},
+		{"SYSCALL", "SYSCALL", []Operand{}, "0f05"},
+		{"XGETBV", "XGETBV", []Operand{}, "0f01d0"},
+		{"PAUSE", "PAUSE", []Operand{}, "f390"},
+		{"LFENCE", "LFENCE", []Operand{}, "0faee8"},
+		{"MFENCE", "MFENCE", []Operand{}, "0faef0"},
+		{"SFENCE", "SFENCE", []Operand{}, "0faef8"},
+		{"UNDEF", "UNDEF", []Operand{}, "0f0b"},
+		{"INT $3", "INT", []Operand{Imm(3)}, "cd03"},
+		{"LDMXCSR (AX)", "LDMXCSR", []Operand{Ptr(AX, 0, 4)}, "0fae10"},
+		{"STMXCSR (AX)", "STMXCSR", []Operand{Ptr(AX, 0, 4)}, "0fae18"},
+		{"CVTSD2SL X0,AX", "CVTSD2SL", []Operand{vreg(t, "X0"), AX}, "f20f2dc0"},
+		{"CVTTSD2SQ X0,AX", "CVTTSD2SQ", []Operand{vreg(t, "X0"), AX}, "f2480f2cc0"},
+		{"CVTTSD2SL X0,AX", "CVTTSD2SL", []Operand{vreg(t, "X0"), AX}, "f20f2cc0"},
+		{"CVTSS2SQ X0,AX", "CVTSS2SQ", []Operand{vreg(t, "X0"), AX}, "f3480f2dc0"},
+		{"FMOVD (AX),F0", "FMOVD", []Operand{Ptr(AX, 0, 8), vreg(t, "F0")}, "dd00"},
+		{"FMOVD F0,(AX)", "FMOVD", []Operand{vreg(t, "F0"), Ptr(AX, 0, 8)}, "dd10"},
+		{"FMOVD F0,F1", "FMOVD", []Operand{vreg(t, "F0"), vreg(t, "F1")}, "ddd1"},
+		{"MOVD AX,X0", "MOVD", []Operand{AX, vreg(t, "X0")}, "66480f6ec0"},
+		{"MOVD X0,AX", "MOVD", []Operand{vreg(t, "X0"), AX}, "66480f7ec0"},
+		{"MOVD X0,X1", "MOVD", []Operand{vreg(t, "X0"), vreg(t, "X1")}, "f30f7ec8"},
+		{"MOVD (AX),X0", "MOVD", []Operand{Ptr(AX, 0, 8), vreg(t, "X0")}, "f30f7e00"},
+		{"MOVD X0,(AX)", "MOVD", []Operand{vreg(t, "X0"), Ptr(AX, 0, 8)}, "660fd600"},
+	}
+	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 := fmt.Sprintf("%x", code); got != c.want {
+			t.Errorf("%s = %s, want %s", c.name, got, c.want)
+		}
+	}
+	// LDMXCSR/STMXCSR take a memory operand only.
+	if _, err := Encode("LDMXCSR", AX); err == nil {
+		t.Errorf("LDMXCSR AX: expected an error, got none")
+	}
+}
+
+// TestSSEGapsGroundTruth pins the legacy SSE gap families: the scalar
+// compare and square root, the Plan 9 packed spellings, the imm8-controlled
+// shuffles, the lane extracts and inserts, the packed integer shifts and the
+// AES/SHA round instructions, byte for byte against go tool asm (see
+// testdata/verify/crypto_amd64.s and testdata/verify/sse_amd64.s).
+func TestSSEGapsGroundTruth(t *testing.T) {
+	cases := []struct {
+		name string
+		mnem string
+		ops  []Operand
+		want string
+	}{
+		{"ANDNPD X0,X1", "ANDNPD", []Operand{vreg(t, "X0"), vreg(t, "X1")}, "660f55c8"},
+		{"ANDNPS X0,X1", "ANDNPS", []Operand{vreg(t, "X0"), vreg(t, "X1")}, "0f55c8"},
+		{"COMISD X0,X1", "COMISD", []Operand{vreg(t, "X0"), vreg(t, "X1")}, "660f2fc8"},
+		{"SQRTSD X0,X1", "SQRTSD", []Operand{vreg(t, "X0"), vreg(t, "X1")}, "f20f51c8"},
+		{"PSHUFL $3,X0,X1", "PSHUFL", []Operand{Imm(3), vreg(t, "X0"), vreg(t, "X1")}, "660f70c803"},
+		{"PALIGNR $2,X0,X1", "PALIGNR", []Operand{Imm(2), vreg(t, "X0"), vreg(t, "X1")}, "660f3a0fc802"},
+		{"PBLENDW $3,X0,X1", "PBLENDW", []Operand{Imm(3), vreg(t, "X0"), vreg(t, "X1")}, "660f3a0ec803"},
+		{"PCMPESTRI $1,X0,X1", "PCMPESTRI", []Operand{Imm(1), vreg(t, "X0"), vreg(t, "X1")}, "660f3a61c801"},
+		{"PCLMULQDQ $0,X0,X1", "PCLMULQDQ", []Operand{Imm(0), vreg(t, "X0"), vreg(t, "X1")}, "660f3a44c800"},
+		{"PCLMULQDQ $0,(AX),X1", "PCLMULQDQ", []Operand{Imm(0), Ptr(AX, 0, 16), vreg(t, "X1")}, "660f3a440800"},
+		{"PEXTRB $1,X0,AX", "PEXTRB", []Operand{Imm(1), vreg(t, "X0"), AX}, "660f3a14c001"},
+		{"PEXTRD $1,X0,AX", "PEXTRD", []Operand{Imm(1), vreg(t, "X0"), AX}, "660f3a16c001"},
+		{"PEXTRQ $1,X0,AX", "PEXTRQ", []Operand{Imm(1), vreg(t, "X0"), AX}, "66480f3a16c001"},
+		{"PEXTRW $1,X0,AX", "PEXTRW", []Operand{Imm(1), vreg(t, "X0"), AX}, "660fc5c001"},
+		{"PEXTRW $1,X0,(AX)", "PEXTRW", []Operand{Imm(1), vreg(t, "X0"), Ptr(AX, 0, 2)}, "660f3a150001"},
+		{"PINSRB $1,AX,X0", "PINSRB", []Operand{Imm(1), AX, vreg(t, "X0")}, "660f3a20c001"},
+		{"PINSRD $1,AX,X0", "PINSRD", []Operand{Imm(1), AX, vreg(t, "X0")}, "660f3a22c001"},
+		{"PINSRQ $1,AX,X0", "PINSRQ", []Operand{Imm(1), AX, vreg(t, "X0")}, "66480f3a22c001"},
+		{"PINSRW $1,AX,X0", "PINSRW", []Operand{Imm(1), AX, vreg(t, "X0")}, "660fc4c001"},
+		{"PINSRW $1,(AX),X0", "PINSRW", []Operand{Imm(1), Ptr(AX, 0, 2), vreg(t, "X0")}, "660fc40001"},
+		{"PSLLL $2,X0", "PSLLL", []Operand{Imm(2), vreg(t, "X0")}, "660f72f002"},
+		{"PSRAL $2,X0", "PSRAL", []Operand{Imm(2), vreg(t, "X0")}, "660f72e002"},
+		{"PSRLL $2,X0", "PSRLL", []Operand{Imm(2), vreg(t, "X0")}, "660f72d002"},
+		{"PSRLQ $2,X0", "PSRLQ", []Operand{Imm(2), vreg(t, "X0")}, "660f73d002"},
+		{"PSLLQ $2,X0", "PSLLQ", []Operand{Imm(2), vreg(t, "X0")}, "660f73f002"},
+		{"PSLLW $2,X0", "PSLLW", []Operand{Imm(2), vreg(t, "X0")}, "660f71f002"},
+		{"PSRLW $2,X0", "PSRLW", []Operand{Imm(2), vreg(t, "X0")}, "660f71d002"},
+		{"PSRAW $2,X0", "PSRAW", []Operand{Imm(2), vreg(t, "X0")}, "660f71e002"},
+		{"PSLLDQ $2,X0", "PSLLDQ", []Operand{Imm(2), vreg(t, "X0")}, "660f73f802"},
+		{"PSRLDQ $2,X0", "PSRLDQ", []Operand{Imm(2), vreg(t, "X0")}, "660f73d802"},
+		{"PSLLL X0,X1", "PSLLL", []Operand{vreg(t, "X0"), vreg(t, "X1")}, "660ff2c8"},
+		{"PSRLQ X0,X1", "PSRLQ", []Operand{vreg(t, "X0"), vreg(t, "X1")}, "660fd3c8"},
+		{"PSLLL (AX),X1", "PSLLL", []Operand{Ptr(AX, 0, 16), vreg(t, "X1")}, "660ff208"},
+		{"PSUBL X0,X1", "PSUBL", []Operand{vreg(t, "X0"), vreg(t, "X1")}, "660ffac8"},
+		{"PADDL X0,X1", "PADDL", []Operand{vreg(t, "X0"), vreg(t, "X1")}, "660ffec8"},
+		{"PCMPEQL X0,X1", "PCMPEQL", []Operand{vreg(t, "X0"), vreg(t, "X1")}, "660f76c8"},
+		{"PUNPCKLBW X0,X1", "PUNPCKLBW", []Operand{vreg(t, "X0"), vreg(t, "X1")}, "660f60c8"},
+		{"MOVOA X0,X1", "MOVOA", []Operand{vreg(t, "X0"), vreg(t, "X1")}, "660f6fc8"},
+		{"MOVOA (AX),X1", "MOVOA", []Operand{Ptr(AX, 0, 16), vreg(t, "X1")}, "660f6f08"},
+		{"MOVOA X0,(AX)", "MOVOA", []Operand{vreg(t, "X0"), Ptr(AX, 0, 16)}, "660f7f00"},
+		{"AESIMC X0,X1", "AESIMC", []Operand{vreg(t, "X0"), vreg(t, "X1")}, "660f38dbc8"},
+		{"AESIMC (AX),X1", "AESIMC", []Operand{Ptr(AX, 0, 16), vreg(t, "X1")}, "660f38db08"},
+		{"AESENC X0,X1", "AESENC", []Operand{vreg(t, "X0"), vreg(t, "X1")}, "660f38dcc8"},
+		{"AESENCLAST X0,X1", "AESENCLAST", []Operand{vreg(t, "X0"), vreg(t, "X1")}, "660f38ddc8"},
+		{"AESDEC X0,X1", "AESDEC", []Operand{vreg(t, "X0"), vreg(t, "X1")}, "660f38dec8"},
+		{"AESDECLAST X0,X1", "AESDECLAST", []Operand{vreg(t, "X0"), vreg(t, "X1")}, "660f38dfc8"},
+		{"AESKEYGENASSIST $0,X0,X1", "AESKEYGENASSIST", []Operand{Imm(0), vreg(t, "X0"), vreg(t, "X1")}, "660f3adfc800"},
+		{"SHA1MSG1 X0,X1", "SHA1MSG1", []Operand{vreg(t, "X0"), vreg(t, "X1")}, "0f38c9c8"},
+		{"SHA1MSG2 X0,X1", "SHA1MSG2", []Operand{vreg(t, "X0"), vreg(t, "X1")}, "0f38cac8"},
+		{"SHA1NEXTE X0,X1", "SHA1NEXTE", []Operand{vreg(t, "X0"), vreg(t, "X1")}, "0f38c8c8"},
+		{"SHA1RNDS4 $0,X0,X1", "SHA1RNDS4", []Operand{Imm(0), vreg(t, "X0"), vreg(t, "X1")}, "0f3accc800"},
+		{"SHA256MSG1 X0,X1", "SHA256MSG1", []Operand{vreg(t, "X0"), vreg(t, "X1")}, "0f38ccc8"},
+		{"SHA256MSG2 X0,X1", "SHA256MSG2", []Operand{vreg(t, "X0"), vreg(t, "X1")}, "0f38cdc8"},
+		{"SHA256RNDS2 X0,X1,X2", "SHA256RNDS2", []Operand{vreg(t, "X0"), vreg(t, "X1"), vreg(t, "X2")}, "0f38cbd1"},
+	}
+	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 := fmt.Sprintf("%x", code); got != c.want {
+			t.Errorf("%s = %s, want %s", c.name, got, c.want)
+		}
+	}
+	// SHA256RNDS2's first operand must be the literal X0.
+	if _, err := Encode("SHA256RNDS2", vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")); err == nil {
+		t.Errorf("SHA256RNDS2 X1,...: expected an error, got none")
+	}
+	// PSLLDQ has no variable-count form.
+	if _, err := Encode("PSLLDQ", vreg(t, "X0"), vreg(t, "X1")); err == nil {
+		t.Errorf("PSLLDQ X0,X1: expected an error, got none")
+	}
+}
+
 // TestSSEBinGroundTruth checks the legacy packed/scalar binary family
 // byte for byte (no prefix / 66 / F2 / F3 variants).
 func TestSSEBinGroundTruth(t *testing.T) {
@@ -421,7 +901,7 @@ func TestSSEShuffleGroundTruth(t *testing.T) {
 }

 // TestMOVQXMMGroundTruth pins the SSE2 packed-quadword move encodings:
-// loads and register moves on F3 0F 7E, stores on 66 0F D6 — the forms
+// loads and register moves on F3 0F 7E, stores on 66 0F D6; the forms
 // the GPR-move fallback silently corrupted.
 func TestMOVQXMMGroundTruth(t *testing.T) {
 	cases := []struct {
@@ -445,3 +925,142 @@ func TestMOVQXMMGroundTruth(t *testing.T) {
 		}
 	}
 }
+
+// TestPrefixStatements pins LOCK, REP and REPN.  go tool asm encodes each as
+// a standalone one-byte instruction with a PC of its own (F0, F3, F2), not a
+// prefix field merged into the following instruction, and it validates
+// nothing about the pairing (LOCK before NOP assembles).  The prefixed
+// atomic and string shapes are the bytes the runtime's own kernels need.
+func TestPrefixStatements(t *testing.T) {
+	cases := []struct {
+		name string
+		mnem string
+		ops  []Operand
+		want string
+	}{
+		{"LOCK", "LOCK", nil, "f0"},
+		{"REP", "REP", nil, "f3"},
+		{"REPN", "REPN", nil, "f2"},
+		// LOCK; CMPXCHGQ AX, (BX)
+		{"LOCK CMPXCHGQ", "CMPXCHGQ", []Operand{AX, Ptr(BX, 0, 8)}, "480fb103"},
+		// REP; MOVSQ
+		{"REP MOVSQ", "MOVSQ", nil, "48a5"},
+		// REPN; MOVSB
+		{"REPN MOVSB", "MOVSB", nil, "a4"},
+	}
+	for _, c := range cases {
+		code, err := Encode(c.mnem, c.ops...)
+		if err != nil {
+			t.Errorf("%s: %v", c.name, err)
+			continue
+		}
+		if got := fmt.Sprintf("%x", code); got != c.want {
+			t.Errorf("%s = %s, want %s", c.name, got, c.want)
+		}
+	}
+	// The prefix statements take no operands, as the toolchain reports for
+	// LOCK AX.
+	if _, err := Encode("LOCK", AX); err == nil {
+		t.Error("LOCK AX assembled, want an error")
+	}
+	if _, err := Encode("REP", Imm(1)); err == nil {
+		t.Error("REP $1 assembled, want an error")
+	}
+}
+
+// TestDataEmission pins BYTE, WORD, LONG and QUAD: the immediate lands in
+// the text stream as 1, 2, 4 or 8 little-endian bytes with no opcode
+// lookup, truncated to the width rather than range-checked (go tool asm
+// emits FF for BYTE $0x1FF and 45 23 for WORD $0x12345, both silently).
+func TestDataEmission(t *testing.T) {
+	cases := []struct {
+		name string
+		mnem string
+		imm  Imm
+		want string
+	}{
+		{"BYTE", "BYTE", 0x0f, "0f"},
+		{"BYTE negative", "BYTE", -1, "ff"},
+		{"BYTE truncated", "BYTE", 0x1ff, "ff"},
+		{"WORD", "WORD", 0x1234, "3412"},
+		{"WORD negative", "WORD", -1, "ffff"},
+		{"WORD truncated", "WORD", 0x12345, "4523"},
+		{"LONG", "LONG", 0x11223344, "44332211"},
+		{"LONG negative", "LONG", -1, "ffffffff"},
+		{"QUAD", "QUAD", 0x1122334455667788, "8877665544332211"},
+		{"QUAD negative", "QUAD", -2, "feffffffffffffff"},
+	}
+	for _, c := range cases {
+		code, err := Encode(c.mnem, c.imm)
+		if err != nil {
+			t.Errorf("%s: %v", c.name, err)
+			continue
+		}
+		if got := fmt.Sprintf("%x", code); got != c.want {
+			t.Errorf("%s = %s, want %s", c.name, got, c.want)
+		}
+	}
+	// Exactly one immediate: the toolchain rejects BYTE $1, $2, $3, and a
+	// register or a missing operand is no immediate at all.
+	if _, err := Encode("BYTE"); err == nil {
+		t.Error("BYTE with no operand assembled, want an error")
+	}
+	if _, err := Encode("BYTE", Imm(1), Imm(2)); err == nil {
+		t.Error("BYTE $1, $2 assembled, want an error")
+	}
+	if _, err := Encode("WORD", AX); err == nil {
+		t.Error("WORD AX assembled, want an error")
+	}
+}
+
+// TestEndIgnored pins END: go tool asm drops the statement entirely, so it
+// encodes to zero bytes and takes any operands without complaint (the
+// toolchain accepts END $0 and END AX alike).
+func TestEndIgnored(t *testing.T) {
+	for _, ops := range [][]Operand{nil, {Imm(0)}, {AX}} {
+		code, err := Encode("END", ops...)
+		if err != nil {
+			t.Errorf("END: %v", err)
+			continue
+		}
+		if len(code) != 0 {
+			t.Errorf("END = %x, want no bytes", code)
+		}
+	}
+}
+
+// TestAdjsp pins ADJSP: a positive immediate is SUBQ $imm, SP, a negative
+// one ADDQ $-imm, SP, in the imm8 or imm32 form the magnitude picks; $0
+// has no encoding (go tool asm refuses ADJSP $0 outright).
+func TestAdjsp(t *testing.T) {
+	cases := []struct {
+		name string
+		imm  Imm
+		want string
+	}{
+		{"imm8", 112, "4883ec70"},
+		{"imm8 negative", -112, "4883c470"},
+		{"imm32", 200, "4881ecc8000000"},
+		{"imm32 negative", -200, "4881c4c8000000"},
+		{"small", 8, "4883ec08"},
+	}
+	for _, c := range cases {
+		code, err := Encode("ADJSP", c.imm)
+		if err != nil {
+			t.Errorf("%s: %v", c.name, err)
+			continue
+		}
+		if got := fmt.Sprintf("%x", code); got != c.want {
+			t.Errorf("ADJSP %d = %s, want %s", int64(c.imm), got, c.want)
+		}
+	}
+	if _, err := Encode("ADJSP", Imm(0)); err == nil {
+		t.Error("ADJSP $0 assembled, want an error")
+	}
+	if _, err := Encode("ADJSP"); err == nil {
+		t.Error("ADJSP with no operand assembled, want an error")
+	}
+	if _, err := Encode("ADJSP", AX); err == nil {
+		t.Error("ADJSP AX assembled, want an error")
+	}
+}
@@ -182,8 +182,17 @@ var evexTable = map[string]evexSpec{
 	// EVEX.66.0F38, permutes (NDS form).
 	"VPERMB":   {2, 0x8D, 0, 1, -1, vexNDS3, [3]int{16, 32, 64}},
 	"VPERMW":   {2, 0x8D, 1, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VPERMI2B": {2, 0x75, 0, 1, -1, vexNDS3, [3]int{16, 32, 64}},
 	"VPERMI2D": {2, 0x76, 0, 1, -1, vexNDS3, [3]int{16, 32, 64}},
 	"VPERMI2Q": {2, 0x76, 1, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	// EVEX.66.0F38, population count (reg=dst, rm=src; W selects byte/word
+	// against dword/qword).
+	"VPOPCNTB": {2, 0x54, 0, 1, -1, vexRM, [3]int{16, 32, 64}},
+	"VPOPCNTD": {2, 0x55, 0, 1, -1, vexRM, [3]int{16, 32, 64}},
+	"VPOPCNTQ": {2, 0x55, 1, 1, -1, vexRM, [3]int{16, 32, 64}},
+	// EVEX.66.0F.W1, the qword spelling of the packed OR (VPORQ has no VEX
+	// form in the Go assembler: it always encodes through EVEX).
+	"VPORQ":     {1, 0xEB, 1, 1, -1, vexNDS3, [3]int{16, 32, 64}},
 	"VPERMT2D":  {2, 0x7E, 0, 1, -1, vexNDS3, [3]int{16, 32, 64}},
 	"VPERMT2Q":  {2, 0x7E, 1, 1, -1, vexNDS3, [3]int{16, 32, 64}},
 	"VPERMT2PD": {2, 0x7F, 1, 1, -1, vexNDS3, [3]int{16, 32, 64}},
@@ -509,7 +518,8 @@ var evexBcastTable = map[string]evexBcastSpec{
 }

 // evexMoveSpec describes an EVEX move (load and store opcodes, like the VEX
-// move table).
+// move table).  vecOK and xmmOnly mirror the VEX twin's operand rules: a
+// scalar move (vecOK false, xmmOnly true) takes XMM↔memory operands only.
 type evexMoveSpec struct {
 	mapSel  int
 	pp      int
@@ -517,32 +527,34 @@ type evexMoveSpec struct {
 	store   byte // vector → r/m
 	w       int
 	n       [3]int
+	vecOK   bool // the non-memory operand may be a vector register
+	xmmOnly bool // wider than XMM registers are rejected
 }

 // evexMoveTable maps an upper-case EVEX move mnemonic to its encoding.
 var evexMoveTable = map[string]evexMoveSpec{
 	// EVEX.128/256/512.F3.0F.W0, unaligned integer move.
-	"VMOVDQU32": {1, 2, 0x6F, 0x7F, 0, [3]int{16, 32, 64}},
+	"VMOVDQU32": {1, 2, 0x6F, 0x7F, 0, [3]int{16, 32, 64}, true, false},
 	// EVEX.128/256/512.F3.0F.W1, unaligned qword move.
-	"VMOVDQU64": {1, 2, 0x6F, 0x7F, 1, [3]int{16, 32, 64}},
+	"VMOVDQU64": {1, 2, 0x6F, 0x7F, 1, [3]int{16, 32, 64}, true, false},
 	// EVEX.128/256/512.F2.0F.W0, unaligned byte move (byte/word moves use the
 	// F2 prefix, dword/qword moves F3; the element size only changes the tuple
 	// semantics).
-	"VMOVDQU8": {1, 3, 0x6F, 0x7F, 0, [3]int{16, 32, 64}},
+	"VMOVDQU8": {1, 3, 0x6F, 0x7F, 0, [3]int{16, 32, 64}, true, false},
 	// EVEX.128/256/512.F2.0F.W1, unaligned word move (shares the qword
 	// encoding).
-	"VMOVDQU16": {1, 3, 0x6F, 0x7F, 1, [3]int{16, 32, 64}},
+	"VMOVDQU16": {1, 3, 0x6F, 0x7F, 1, [3]int{16, 32, 64}, true, false},
 	// EVEX.128/256/512.66.0F.W1, unaligned packed double move.
-	"VMOVUPD": {1, 1, 0x10, 0x11, 1, [3]int{16, 32, 64}},
+	"VMOVUPD": {1, 1, 0x10, 0x11, 1, [3]int{16, 32, 64}, true, false},
 	// EVEX.128/256/512, aligned packed moves.
-	"VMOVAPS": {1, 0, 0x28, 0x29, 0, [3]int{16, 32, 64}},
-	"VMOVAPD": {1, 1, 0x28, 0x29, 1, [3]int{16, 32, 64}},
+	"VMOVAPS": {1, 0, 0x28, 0x29, 0, [3]int{16, 32, 64}, true, false},
+	"VMOVAPD": {1, 1, 0x28, 0x29, 1, [3]int{16, 32, 64}, true, false},
 	// EVEX.128/256/512.66.0F, aligned integer moves.
-	"VMOVDQA32": {1, 1, 0x6F, 0x7F, 0, [3]int{16, 32, 64}},
-	"VMOVDQA64": {1, 1, 0x6F, 0x7F, 1, [3]int{16, 32, 64}},
+	"VMOVDQA32": {1, 1, 0x6F, 0x7F, 0, [3]int{16, 32, 64}, true, false},
+	"VMOVDQA64": {1, 1, 0x6F, 0x7F, 1, [3]int{16, 32, 64}, true, false},
 	// EVEX.128.F3.0F.W0, scalar single move, memory operands (the
 	// three-operand register form is not supported).
-	"VMOVSS": {1, 2, 0x10, 0x11, 0, [3]int{4, 4, 4}},
+	"VMOVSS": {1, 2, 0x10, 0x11, 0, [3]int{4, 4, 4}, false, true},
 }

 // isEvex reports whether the mnemonic has an EVEX encoding we handle.
@@ -1022,6 +1034,12 @@ func (e *enc) encodeEvexMove(mnem string, ms evexMoveSpec, ops []Operand, mask i
 	var rm Operand
 	switch {
 	case srcIsVec && dstIsVec:
+		// A store-form reg-reg move, the layout the Go assembler uses; a
+		// scalar move has no two-register form at all (the register form
+		// takes three operands), matching the VEX twin's vecOK rule.
+		if !ms.vecOK {
+			return fmt.Errorf("%s does not take two vector registers", mnem)
+		}
 		reg, rm = srcReg, dst
 	case srcIsVec:
 		if !memOperand(dst) {
@@ -1037,6 +1055,12 @@ func (e *enc) encodeEvexMove(mnem string, ms evexMoveSpec, ops []Operand, mask i
 	default:
 		return fmt.Errorf("%s needs a vector register operand", mnem)
 	}
+	// The scalar move is 128-bit only, so the register the length follows
+	// must be an XMM (the VEX twin's xmmOnly rule; EVEX also reaches ZMM,
+	// hence the inequality rather than a YMM test).
+	if ms.xmmOnly && reg.size != 16 {
+		return fmt.Errorf("%s operates on XMM registers only", mnem)
+	}
 	spec := evexSpec{mapSel: ms.mapSel, opcode: op, w: ms.w, pp: ms.pp, opdigit: -1, n: ms.n}
 	return e.emitEvexFields(spec, reg.vecLenBit(), reg.idx, -1, rm, mask, sfx)
 }
@@ -1171,9 +1195,6 @@ func (e *enc) emitEvexFields(spec evexSpec, ll, regIdx, vvvvIdx int, rm Operand,
 		if r.idx&16 != 0 {
 			xBar = 0
 		}
-		if r.idx&16 != 0 {
-			xBar = 0
-		}
 	case Mem:
 		var err error
 		modrm, sib, disp, xBar, bBar, err = memComponentsEvex(regIdx&7, r, spec.n[ll])
@@ -1232,6 +1253,11 @@ func (e *enc) emitEvexFields(spec evexSpec, ll, regIdx, vvvvIdx int, rm Operand,
 func memComponentsEvex(regField int, m Mem, n int) (modrm, sib int, disp []byte, xBar, bBar int, err error) {
 	sib = -1
 	xBar, bBar = 1, 1 // inverted bits: 1 = no extension
+	// The disp32 fallback bounds the displacement by int32, and the
+	// compressed disp8 form reaches at most ±127×64, well inside it.
+	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)
+	}
 	if !m.HasBase && !m.HasIndex {
 		return regField<<3 | 0x05, -1, le32(m.Disp), 1, 1, nil // RIP-relative
 	}
@@ -1418,18 +1444,20 @@ var evexKOperand = map[string]bool{
 }

 // kmovSpec describes a KMOV width: the opcode depends on the operand
-// direction, kk (k/mem → K is 90, k → k uses the same), kmem (K → mem),
-// gprk (GPR/mem → K), kgpr (K → GPR), and the GPR forms carry a mandatory
-// prefix and W for the wider widths.
+// direction, kk (k → k), kmem (k → mem), gprk (GPR/mem → k) and kgpr
+// (k → GPR).  Each direction group carries its own mandatory prefix and W:
+// the k-destination/source forms share one pair, the GPR forms another.
 type kmovSpec struct {
 	kk, kmem, gprk, kgpr byte
-	gprPP                int
-	w                    int
+	kPP, kW              int // prefix and VEX.W for the k forms
+	gprPP, gprW          int // prefix and VEX.W for the GPR forms
 }

 var kmovTable = map[string]kmovSpec{
-	"KMOVW": {0x90, 0x91, 0x92, 0x93, 0, 0},
-	"KMOVQ": {0x90, 0x91, 0x92, 0x93, 3, 1},
+	"KMOVW": {0x90, 0x91, 0x92, 0x93, 0, 0, 0, 0},
+	"KMOVB": {0x90, 0x91, 0x92, 0x93, 1, 0, 1, 0},
+	"KMOVD": {0x90, 0x91, 0x92, 0x93, 1, 1, 3, 0},
+	"KMOVQ": {0x90, 0x91, 0x92, 0x93, 0, 1, 3, 1},
 }

 // encodeKmov encodes a KMOV width, selecting the opcode by direction.
@@ -1443,14 +1471,14 @@ func (e *enc) encodeKmov(upper string, ops []Operand) error {
 	dstReg, dstIsReg := dst.(Reg)
 	srcK := srcIsReg && srcReg.mask
 	dstK := dstIsReg && dstReg.mask
-	spec := vexSpec{mapSel: 1, w: ks.w, pp: 0, opdigit: -1}
 	switch {
 	case srcK && dstK:
-		spec.opcode = ks.kk // k ← k: reg = dst, rm = src
+		// k ← k: reg = dst, rm = src.
+		spec := vexSpec{mapSel: 1, opcode: ks.kk, w: ks.kW, pp: ks.kPP, opdigit: -1}
 		return e.emitVexFields(spec, 0, dstReg.idx&7, 0, 15, src)
 	case srcK && dstIsReg:
-		spec.opcode = ks.kgpr // GPR ← k: reg = dst, rm = src
-		spec.pp = ks.gprPP
+		// GPR ← k: reg = dst, rm = src.
+		spec := vexSpec{mapSel: 1, opcode: ks.kgpr, w: ks.gprW, pp: ks.gprPP, opdigit: -1}
 		rBit := 0
 		if dstReg.idx >= 8 {
 			rBit = 1
@@ -1460,11 +1488,17 @@ func (e *enc) encodeKmov(upper string, ops []Operand) error {
 		if _, ok := dst.(Mem); !ok {
 			return fmt.Errorf("%s: invalid destination operand", upper)
 		}
-		spec.opcode = ks.kmem // mem ← k: reg = src, rm = dst
+		// mem ← k: reg = src, rm = dst.
+		spec := vexSpec{mapSel: 1, opcode: ks.kmem, w: ks.kW, pp: ks.kPP, opdigit: -1}
 		return e.emitVexFields(spec, 0, srcReg.idx&7, 0, 15, dst)
 	case dstK:
-		spec.opcode = ks.gprk // k ← GPR/mem: reg = dst, rm = src
-		spec.pp = ks.gprPP
+		// k ← GPR: reg = dst, rm = src.  A memory source shares the k ← k
+		// opcode and prefix group (the ykmovb layout the Go assembler uses).
+		opcode, w, pp := ks.gprk, ks.gprW, ks.gprPP
+		if memOperand(src) {
+			opcode, w, pp = ks.kk, ks.kW, ks.kPP
+		}
+		spec := vexSpec{mapSel: 1, opcode: opcode, w: w, pp: pp, opdigit: -1}
 		return e.emitVexFields(spec, 0, dstReg.idx&7, 0, 15, src)
 	}
 	return fmt.Errorf("%s requires a K register operand", upper)
@@ -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"},
@@ -399,8 +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) {
@@ -502,9 +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) {
@@ -675,6 +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 {
@@ -68,11 +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
 )

@@ -287,7 +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,
 		})
 	}
@@ -317,7 +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))
@@ -40,8 +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)
@@ -50,7 +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
@@ -145,40 +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
+	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
 }

@@ -192,12 +222,16 @@ 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
 	}
@@ -247,7 +281,7 @@ 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.
@@ -256,6 +290,8 @@ func parseGOOBJ(data []byte) (*goobjFile, error) {
 	f := &goobjFile{
 		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
@@ -306,22 +342,26 @@ 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 before, after, ok := strings.Cut(full, "."); ok {
 		return before, after
@@ -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)
@@ -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) {
@@ -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)
+	}
+}
@@ -18,19 +18,40 @@ import (
 // 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) {
+	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, 4
+			return relocArm64LDST64, 8
 		default:
-			return relocArm64Addr, 4
+			return relocArm64Addr, 8
 		}
 	})
 }
@@ -6,6 +6,7 @@ package asm
 import (
 	"bytes"
 	"encoding/hex"
+	"strings"
 	"testing"

 	"sourcedock.dev/petrbalvin/gasm-devkit/ast"
@@ -27,6 +28,12 @@ func TestStackGuardBytes(t *testing.T) {
 			"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",
@@ -145,6 +152,45 @@ func TestStackGuardBytesARM64(t *testing.T) {
 	}
 }

+// 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.
@@ -19,8 +19,8 @@ import (
 	"sourcedock.dev/petrbalvin/gasm-devkit/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"
@@ -81,7 +81,7 @@ 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"
@@ -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]
@@ -25,6 +25,10 @@ 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.
@@ -81,12 +85,9 @@ func (fl *FuncLayout) LineAt(offset int) int {
 	return 0
 }

-// RelocKind 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 (
@@ -96,7 +97,6 @@ const (
 	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 pair)
@@ -106,6 +106,13 @@ const (
 )

 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.
 	Off      int
 	After    int
 	Name     string
@@ -150,7 +157,7 @@ func AssembleFile(f *ast.File) (*Image, error) {
 	}
 	link := &linkInfo{symbols: known, allowExternal: true}

-	img := &Image{Symbols: map[string]int{}}
+	img := &Image{Symbols: map[string]int{}, SourcePath: f.Path}
 	textOff := map[string]int{}
 	type asmFunc struct {
 		name    string
@@ -267,7 +274,7 @@ func AssembleFileRISCV(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 {
@@ -339,7 +346,7 @@ 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 {
@@ -440,33 +447,37 @@ type dataSym struct {
 }

 // 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" {
+		gd, ok := d.(*ast.Globl)
+		if !ok {
 			continue
 		}
-			name := dd.Name.Name
+		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 dd.Size != nil && dd.Size.Imm.HasVal {
-				size = int(dd.Size.Imm.Val)
+		if gd.Size != nil && gd.Size.Imm.HasVal {
+			size = int(gd.Size.Imm.Val)
 		}
 		index[name] = len(syms)
 		ds := dataSym{
 			name:   name,
-				pkg:    dd.Name.Pkg,
+			pkg:    gd.Name.Pkg,
 			buf:    make([]byte, size),
 			size:   size,
-				static: dd.Name.Static,
+			static: gd.Name.Static,
 		}
-			for _, f := range dd.Flags {
+		for _, f := range gd.Flags {
 			switch f {
 			case "RODATA":
 				ds.rodata = true
@@ -487,8 +498,12 @@ func collectData(f *ast.File) ([]dataSym, error) {
 			}
 		}
 		syms = append(syms, ds)
-
-		case *ast.Data:
+	}
+	for _, d := range f.Decls {
+		dd, ok := d.(*ast.Data)
+		if !ok {
+			continue
+		}
 		if dd.Name == nil || dd.Name.Pseudo != "SB" {
 			continue
 		}
@@ -518,7 +533,6 @@ func collectData(f *ast.File) ([]dataSym, error) {
 			buf[off+int64(j)] = byte(v >> (8 * j))
 		}
 	}
-	}
 	return syms, nil
 }

@@ -10,8 +10,8 @@ import (
 	"sourcedock.dev/petrbalvin/gasm-devkit/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", `
@@ -30,7 +30,10 @@ 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"
+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
@@ -103,7 +106,12 @@ 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:], 31)
 	}
@@ -148,6 +156,19 @@ func loong64RegSpecial(digits 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.
@@ -199,7 +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)
 }
@@ -245,7 +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
@@ -253,6 +276,8 @@ 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
 )

 // l64Enc is one instruction's encoding: its bit layout (format) and the
@@ -275,10 +300,68 @@ 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{}
+
+// 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.
 	rrr := map[string]uint32{
@@ -358,6 +441,10 @@ 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,
 	}
 	for m, op := range rr {
 		l64InstrTable[m] = l64Enc{format: l64Frr, op: op}
@@ -414,12 +501,14 @@ 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}
@@ -453,6 +542,10 @@ func init() {
 	l64InstrTable["PRELD"] = l64Enc{format: l64Fpreld, op: 0x0ab << 22}

 	// 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,
@@ -470,10 +563,84 @@ 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,
 	}
 	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},
+	}
+	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, vsrai.b stores imm+8 (byte-lane bias),
+	// vseqi.b and vseqi.d store 5-bit and 7-bit two's-complement values.
+	// The mnemonics that also have a register form (VSEQB, VSEQV, VSRAB,
+	// VROTRW) 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},
+		"VSEQB":  {0xE500 << 15, false, -16, 15, 0, 0x1F},
+		"XVSEQB": {0xE900 << 15, true, -16, 15, 0, 0x1F},
+		"VSEQV":  {0xE503 << 15, false, -64, 63, 0, 0x7F},
+		"XVSEQV": {0xE903 << 15, true, -64, 63, 0, 0x7F},
+		"VSRAB":  {0xE668 << 15, false, 0, 7, 8, 0x1F},
+		"VROTRW": {0xE541 << 15, false, 0, 31, 0, 0x1F},
+	}
+	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,
+	}
+	for m, op := range vecCf {
+		l64InstrTable[m] = l64Enc{format: l64Fvcf, op: op}
+		l64VecBank[m] = strings.HasPrefix(m, "XV")
+	}
+
+	// Lane popcount: INSTR vj, vd (the 2R layout with the opcode extending
+	// over the unused vk field).
+	vec2r := map[string]l64Vec3Enc{
+		"VPCNTV": {0x1CA70B << 10, false}, "XVPCNTV": {0x1DA70B << 10, true},
+	}
+	for m, e := range vec2r {
+		l64InstrTable[m] = l64Enc{format: l64Frr, op: e.op}
+		l64VecBank[m] = e.lasx
+		l64Vec2R[m] = true
+	}
 }

 // l64FpMovTable maps (mnemonic, from-class, to-class) to the 2R opcode of the
@@ -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 +305,287 @@ 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,
+		)
+	})
+}
+
+// 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
+`,
+		// 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,
+	)
+}
@@ -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)
+		}
+	}
+}
@@ -48,3 +48,16 @@ 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
+	}
+}
@@ -17,12 +17,13 @@ import "strings"
 // 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
 // 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
+	fp   bool // F0-F7 x87 stack register
 }

 // Index returns the register number (0-15 for GPRs, 0-31 for vectors).
@@ -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
 }

@@ -65,7 +65,7 @@ func riscvRegNum(name string) int {
 		return 25
 	case "X26", "S10":
 		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]
@@ -193,6 +219,9 @@ var riscvInstrTable = map[string]riscvEnc{
 	"DIVUW": {0x3B, 0x5, 0x01},
 	"REMW":  {0x3B, 0x6, 0x01},
 	"REMUW": {0x3B, 0x7, 0x01},
+	// Zicond conditional zeroing.
+	"CZEROEQZ": {0x33, 0x5, 0x07},
+	"CZERONEZ": {0x33, 0x7, 0x07},
 	// RV64I, I-type arithmetic.
 	"ADDI":  {0x13, 0x0, 0x00},
 	"ADDIW": {0x1B, 0x0, 0x00},
@@ -221,6 +250,12 @@ 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},
@@ -228,29 +263,34 @@ var riscvInstrTable = map[string]riscvEnc{
 	"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},
+	// 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.
 	"FADDS":  {0x53, 0x0, 0x00},
@@ -273,12 +313,23 @@ 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},
+	// 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).
 	"FEQS": {0x53, 0x2, 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.
@@ -439,6 +496,71 @@ 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
@@ -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))
@@ -5,6 +5,9 @@ package asm

 import (
 	"bytes"
+	"encoding/binary"
+	"encoding/hex"
+	"strings"
 	"testing"

 	"sourcedock.dev/petrbalvin/gasm-devkit/ast"
@@ -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"
@@ -761,3 +839,371 @@ sub:
 		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)
+}
+
+// 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 (the
+// toolchain materialises such constants via SLLI expansion, which this
+// assembler does not implement).
+func TestRISCV_Imm64Rejected(t *testing.T) {
+	cases := []string{
+		"MOV $0x123456789, X10",
+		"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)
+	}
+}
+
+// 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
+	)
+}
@@ -4,6 +4,7 @@
 package asm

 import (
+	"fmt"
 	"strings"

 	"sourcedock.dev/petrbalvin/gasm-devkit/ast"
@@ -93,12 +94,19 @@ func riscvIsLeaf(t *ast.Text) bool {
 				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
 			}
 		}
@@ -234,33 +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
 	}
+	adj := int32(-fi.autosize)
+	if fits12(adj) {
 		// SD (4 bytes) + ADDI/C.ADDI (2 or 4 bytes).
-	return 4 + riscvSPAdjustLen(int32(-fi.autosize))
+		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
 	}
+	adj := int32(fi.autosize)
+	if fits12(adj) {
 		// C.LDSP (2 bytes) + ADDI/C.ADDI (2 or 4 bytes).
-	return 2 + riscvSPAdjustLen(int32(fi.autosize))
+		return 2 + len(riscvSPAdjust(adj))
 	}
-
-func riscvSPAdjustLen(imm int32) int {
-	if imm != 0 && imm%16 == 0 && imm >= -512 && imm <= 511 {
-		return 2
-	}
-	if riscvFitsCAddi(imm) {
-		return 2
-	}
-	return 4
+	return 2 + len(riscvAddToSP(adj))
 }

 // riscvResolvePseudo translates a pseudo-register memory reference into a
@@ -286,19 +297,21 @@ func riscvResolvePseudo(sym *ast.Symbol, fi riscvFrameInfo) (base int, off int32
 // 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 {
-	_, reloc := riscvGuard(fi)
-	_ = reloc
-	return len(riscvGuardBytes(fi))
+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) {
+func riscvGuard(fi riscvFrameInfo) ([]byte, Reloc, error) {
 	if !fi.needSplit {
-		return nil, Reloc{}
+		return nil, Reloc{}, nil
 	}
 	// MOV 16(g), X6 (g.stackguard0), g = X27.
 	out := wordLE(riscvIType(riscvEnc{0x03, 0x3, 0x00}, 6, 27, 16))
@@ -310,14 +323,14 @@ func riscvGuard(fi riscvFrameInfo) ([]byte, Reloc) {
 	var reloc Reloc
 	switch fi.splitClass {
 	case 0:
-		// BLTU X6, SP, done (+8: over the CALL and the JMP back)
+		// 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 (+8)
+		// 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))...)
@@ -335,7 +348,10 @@ func riscvGuard(fi riscvFrameInfo) ([]byte, Reloc) {
 		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 {
-			addi = 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))...)
@@ -344,11 +360,5 @@ func riscvGuard(fi riscvFrameInfo) ([]byte, Reloc) {
 		out = append(out, wordLE(riscvJType(5, 0))...)
 		out = append(out, jalBack()...)
 	}
-	return out, reloc
-}
-
-// riscvGuardBytes emits the guard prefix bytes alone (sizing helper).
-func riscvGuardBytes(fi riscvFrameInfo) []byte {
-	g, _ := riscvGuard(fi)
-	return g
+	return out, reloc, nil
 }
@@ -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).
@@ -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"
@@ -41,7 +41,7 @@ const (
 	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).
+	// 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 +52,15 @@ 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
 )

 // vexSpec describes one VEX instruction's encoding parameters.
@@ -125,6 +134,12 @@ var vexTable = map[string]vexSpec{
 	"VMAXSS": {1, 0x5F, 0, 2, -1, vexNDS3},
 	// 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,
 	// no vvvv).
@@ -192,6 +207,31 @@ var vexTable = map[string]vexSpec{

 	// 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},
+	"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).
 	"KTESTW": {1, 0x99, 0, 0, -1, vexRM},
@@ -200,6 +240,14 @@ var vexTable = map[string]vexSpec{
 	// rm=scalar memory; SD is 256-bit only).
 	"VBROADCASTSS": {2, 0x18, 0, 1, -1, vexRM},
 	"VBROADCASTSD": {2, 0x19, 0, 1, -1, vexRM},
+	// 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},
@@ -290,6 +338,8 @@ type vexMoveSpec struct {
 var vexMoveTable = map[string]vexMoveSpec{
 	// 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, 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.
@@ -324,6 +374,14 @@ func (e *enc) encodeVex(mnemUpper string, ops []Operand) error {
 			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)
 	}
@@ -356,6 +414,14 @@ 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 vexRMRev:
+		return e.encodeVexRMRev(spec, ops)
 	}
 	return fmt.Errorf("unhandled VEX form for %s", mnemUpper)
 }
@@ -607,6 +673,83 @@ 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")
+	}
+	return e.emitVexFields(spec, 0, dstReg.idx&7, 0, 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
+}
+
+// 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
@@ -19,6 +19,20 @@ 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,
+}
+
 // 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 +51,15 @@ 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 +194,7 @@ 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 +205,38 @@ 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", ""},
 		// 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 +270,7 @@ 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 +287,7 @@ 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 +308,12 @@ 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 +340,12 @@ 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
 		}
@@ -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
@@ -37,21 +37,38 @@ import (
 // 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 [amd64|arm64|riscv64|loong64]", `
+	fs := newCommand("audit-instructions", "gasm audit-instructions [--corpus [dir]] [-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), known-but-unencodable names (the backlog) and go-
-only names (feature gaps).  The Go side is probed black-box with a battery
-of bare mnemonics, so the audit tracks whatever toolchain `+"`go env GOROOT`"+`
-provides.
+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.
 `)
+	corpus := fs.Bool("corpus", false, "assemble a corpus of .s files and report pass rates and failure reasons")
+	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)
+	}
 	archName := "amd64"
 	switch n := len(fs.Args()); {
 	case n > 1:
-		return fmt.Errorf("audit-instructions takes at most one architecture argument")
+		return &usageError{fmt.Errorf("audit-instructions takes at most one architecture argument")}
 	case n == 1:
 		archName = strings.ToLower(fs.Arg(0))
 	}
@@ -97,7 +114,7 @@ provides.

 	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 recognized: %d\n", len(shared)+len(superset), countTrue(goKnown))
+	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 {
@@ -124,7 +141,7 @@ func auditArch(name string) (arch.Arch, error) {
 	case "loong64", "loong":
 		return arch.LOONG64, nil
 	}
-	return arch.Unknown, fmt.Errorf("unknown architecture %q: want amd64, arm64, riscv64 or loong64", name)
+	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.
@@ -205,6 +222,13 @@ func probeGoAsm(goarch string, names []string) (map[string]bool, error) {
 	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 {
@@ -244,18 +268,62 @@ func probeShapes(a arch.Arch) []string {
 			// 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",
 		}
 	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",
+			"",
 		}
 	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", "$65536, R4", "R4, R5, p2", "p2(SB)",
+			// AMO atomics: destination, base, source.
+			"R5, (R4), R6", "X5, (X4), X6",
+			// Segment stores take the first vector register aligned
+			// to the segment count, as the toolchain requires.
+			"(X5), X6, V0, V8", "(X5), X6, V0", "(X5), X0, V4",
+			// The LSX and LASX banks share the 5-bit numbering with F.
+			"V1, V2, V3", "X1, X2, X3", "V1, V2", "X1, X2", "V1", "X1",
+			// The vector compare-to-flag forms land in an FCC register.
+			"V1, FCC0", "X1, FCC0",
+			"",
 		}
 	}
 	return nil
@@ -305,3 +373,255 @@ func gasmAssembles(a arch.Arch, name, shape string) bool {
 func sanitize(name string) string {
 	return strings.NewReplacer(".", "_", "$", "_").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
+}
+
+func (t *corpusTally) fail(path, reason string) {
+	t.reasons[reason]++
+	if t.example[reason] == "" {
+		t.example[reason] = path
+	}
+}
+
+// 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) 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 and stays unresolvable on purpose.
+	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)
+	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
+	otherPort int // files named for another Go port: never attempted
+	full      int // files that assembled for every 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", "ppc64", "ppc64le", "riscv", "riscv64",
+	"s390x", "wasm",
+}
+
+// otherPortFile reports whether the file's name carries a Go-architecture
+// suffix gasm does not support.
+func otherPortFile(path string) bool {
+	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
+}
+
+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 name allows assembles it.
+	full, generic, otherPort := 0, 0, 0
+
+	for _, path := range files {
+		src, err := readSource(path)
+		if err != nil {
+			return nil, err
+		}
+		f, errs := parser.ParseWithOptions(path, src, parser.Options{Expand: true, IncludeDirs: dirs})
+
+		var wanted []int // indexes into targets
+		if a := arch.FromFilename(path); a != arch.Unknown {
+			for i, tg := range targets {
+				if tg.a == a {
+					wanted = append(wanted, i)
+				}
+			}
+		} else if otherPortFile(path) {
+			// A file named for a Go port gasm does not support (arm,
+			// 386, s390x, ...) 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.
+			otherPort++
+		} else {
+			generic++
+			for i := range targets {
+				wanted = append(wanted, i)
+			}
+		}
+
+		ok := true
+		for _, i := range wanted {
+			tg, t := targets[i], tallies[i]
+			t.attempted++
+			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)
+			}
+			if err != nil {
+				ok = false
+				t.fail(path, corpusReason(err))
+				continue
+			}
+			t.assembled++
+		}
+		if ok && len(wanted) > 0 {
+			full++
+		}
+	}
+
+	return &corpusStats{
+		root:      root,
+		files:     len(files),
+		generic:   generic,
+		otherPort: otherPort,
+		full:      full,
+		targets:   targets,
+		tallies:   tallies,
+	}, nil
+}
+
+// printCorpusStats renders the corpus audit report.
+func printCorpusStats(s *corpusStats) {
+	fmt.Printf("corpus %s: %d files (%d generic, attempted for all architectures; %d named for other Go ports, never attempted)\n", s.root, s.files, s.generic, s.otherPort)
+	// The rate is over the files a supported build would attempt: the
+	// other ports' files sit in the count for completeness but can never
+	// assemble, so counting them in the denominator would report the gap
+	// of architectures gasm deliberately does not target.
+	attemptable := max(s.files-s.otherPort, 1)
+	fmt.Printf("  assemble for every target architecture: %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])
+		}
+	}
+}
+
+// corpusReason buckets an assembly or parse failure for the histogram.
+func corpusReason(err error) string {
+	msg := err.Error()
+	switch {
+	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
+}
@@ -24,9 +24,10 @@ function in a traced subprocess (ptrace), then provides a REPL for
 single-stepping, breakpoints, register and memory inspection.

 REPL commands:
-  break <label|addr> [if <reg> <op> <val>]
+  break <label|addr|line> [if <reg> <op> <val|reg|*addr>]
                       set a breakpoint, optionally conditional on a
-                       register comparison (reg-reg or reg-immediate)
+                       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)
@@ -53,7 +54,7 @@ REPL commands:
 	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")
+	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) ---
@@ -231,6 +232,13 @@ REPL commands:
 			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
@@ -10,6 +10,7 @@ package main
 import (
 	"bytes"
 	"encoding/json"
+	"errors"
 	"flag"
 	"fmt"
 	"io"
@@ -18,6 +19,7 @@ import (
 	"os/exec"
 	"path/filepath"
 	"runtime"
+	"runtime/debug"
 	"slices"
 	"sort"
 	"strconv"
@@ -36,9 +38,32 @@ import (
 	"sourcedock.dev/petrbalvin/gasm-devkit/verify"
 )

-// version is the release version, stamped at build time via
-// -ldflags "-X main.version=…" (defaulting to the current release).
-var version = "0.33.0"
+// version reports the release the toolchain recorded for this build: the
+// tag on a tag, a pseudo-version below one, and (devel) outside version
+// control. Nothing is injected; the recorded value cannot go stale.
+func version() string {
+	bi, ok := debug.ReadBuildInfo()
+	if !ok || bi.Main.Version == "" {
+		return "(devel)"
+	}
+	return bi.Main.Version
+}
+
+// usageError marks an error the caller's arguments caused, which exits 2
+// instead of the 1 a runtime failure gets.
+type usageError struct{ err error }
+
+func (e *usageError) Error() string { return e.err.Error() }
+func (e *usageError) Unwrap() error { return e.err }
+
+// exitCodeFor maps an error onto the process exit status: 2 for a usage
+// error, 1 for anything else.
+func exitCodeFor(err error) int {
+	if _, ok := errors.AsType[*usageError](err); ok {
+		return 2
+	}
+	return 1
+}

 func main() {
 	if len(os.Args) < 2 {
@@ -69,12 +94,12 @@ func main() {
 	case "audit-instructions":
 		if err := cmdAuditInstructions(os.Args[2:]); err != nil {
 			fmt.Fprintln(os.Stderr, err)
-			os.Exit(1)
+			os.Exit(exitCodeFor(err))
 		}
 	case "scaffold":
 		if err := cmdScaffold(os.Args[2:]); err != nil {
 			fmt.Fprintln(os.Stderr, err)
-			os.Exit(1)
+			os.Exit(exitCodeFor(err))
 		}
 	case "lsp":
 		os.Exit(cmdLSP(os.Args[2:]))
@@ -88,22 +113,22 @@ func main() {
 	}
 }

-// cmdVersion prints the release version.
+// cmdVersion prints the recorded version.
 func cmdVersion() int {
-	fmt.Printf("gasm %s\n", version)
+	fmt.Printf("gasm %s\n", version())
 	return 0
 }

-// ANSI color helpers for terminal output.
+// ANSI colour helpers for terminal output.
 const (
-	colorReset  = "\033[0m"
-	colorBold   = "\033[1m"
-	colorCyan   = "\033[36m"
-	colorYellow = "\033[33m"
-	colorGray   = "\033[90m"
+	colourReset  = "\033[0m"
+	colourBold   = "\033[1m"
+	colourCyan   = "\033[36m"
+	colourYellow = "\033[33m"
+	colourGrey   = "\033[90m"
 )

-// isTTY reports whether the writer is a terminal (for color output).
+// isTTY reports whether the writer is a terminal (for colour output).
 func isTTY(w io.Writer) bool {
 	if f, ok := w.(*os.File); ok {
 		stat, _ := f.Stat()
@@ -114,12 +139,12 @@ func isTTY(w io.Writer) bool {

 func usage(w io.Writer) {
 	useColor := isTTY(w)
-	bold, cyan, yellow, gray, reset := "", "", "", "", ""
+	bold, cyan, yellow, grey, reset := "", "", "", "", ""
 	if useColor {
-		bold, cyan, yellow, gray, reset = colorBold, colorCyan, colorYellow, colorGray, colorReset
+		bold, cyan, yellow, grey, reset = colourBold, colourCyan, colourYellow, colourGrey, colourReset
 	}

-	fmt.Fprintf(w, "%sgasm %s%s: developer tooling for Go's Plan 9 assembler (GAsm)%s\n\n", bold, version, reset, reset)
+	fmt.Fprintf(w, "%sgasm %s%s: developer tooling for Go's Plan 9 assembler (GAsm)%s\n\n", bold, version(), reset, reset)
 	fmt.Fprintf(w, "gasm bundles a lexer, parser, formatter, linter, standalone assembler and\n")
 	fmt.Fprintf(w, "language server for Plan 9 assembly into one self-contained binary.\n\n")

@@ -145,12 +170,12 @@ func usage(w io.Writer) {
 		{"version", "print the version (same as --version)"},
 	}
 	for _, c := range commands {
-		fmt.Fprintf(w, "  %s%-10s%s %s%s%s\n", cyan, c.name, reset, gray, c.desc, reset)
+		fmt.Fprintf(w, "  %s%-10s%s %s%s%s\n", cyan, c.name, reset, grey, c.desc, reset)
 	}

 	fmt.Fprintf(w, "\n%sFlags:%s\n", yellow, reset)
-	fmt.Fprintf(w, "  %s-h, --help%s      %sshow this help%s\n", cyan, reset, gray, reset)
-	fmt.Fprintf(w, "  %s-V, --version%s   %sprint the version%s\n", cyan, reset, gray, reset)
+	fmt.Fprintf(w, "  %s-h, --help%s      %sshow this help%s\n", cyan, reset, grey, reset)
+	fmt.Fprintf(w, "  %s-V, --version%s   %sprint the version%s\n", cyan, reset, grey, reset)

 	fmt.Fprintf(w, "\nRun \"gasm <command> -h\" for a command's usage and flags.\n\n")

@@ -164,7 +189,7 @@ func usage(w io.Writer) {
 	}
 	for _, e := range examples {
 		if e.desc != "" {
-			fmt.Fprintf(w, "  %s%s%s  %s%s%s\n", cyan, e.cmd, reset, gray, e.desc, reset)
+			fmt.Fprintf(w, "  %s%s%s  %s%s%s\n", cyan, e.cmd, reset, grey, e.desc, reset)
 		} else {
 			fmt.Fprintf(w, "  %s%s%s\n", cyan, e.cmd, reset)
 		}
@@ -215,6 +240,16 @@ func readSource(path string) (string, error) {
 	return string(b), err
 }

+// includeDirs collects repeatable -I flags: the directories searched for
+// #include files during macro expansion and include splicing.
+type includeDirs []string
+
+func (d *includeDirs) String() string { return strings.Join(*d, ",") }
+func (d *includeDirs) Set(v string) error {
+	*d = append(*d, v)
+	return nil
+}
+
 func cmdTokens(args []string) int {
 	fs := newCommand("tokens", "gasm tokens <file>", `
 Print the lexical token stream of FILE: position, token kind and text, one
@@ -442,7 +477,7 @@ hover, document symbols, diagnostics and semantic-token highlighting.
 `)
 	fs.Parse(args)
 	srv := lsp.New(os.Stdin, os.Stdout)
-	srv.SetVersion(version)
+	srv.SetVersion(version())
 	if err := srv.Run(); err != nil {
 		fmt.Fprintln(os.Stderr, "gasm lsp:", err)
 		return 1
@@ -451,7 +486,7 @@ hover, document symbols, diagnostics and semantic-token highlighting.
 }

 func cmdAsm(args []string) int {
-	fs := newCommand("asm", "gasm asm [--format raw|elf|goobj] [-p pkg] [-o out] <file>", `
+	fs := newCommand("asm", "gasm asm [--format raw|elf|goobj] [-I dir] [-p pkg] [-GOARCH arch] [-o out] <file>", `
 	Assemble FILE without the Go toolchain: every TEXT function is encoded to
 	machine code and printed as a hex dump.  Supported architectures: amd64
 	(including VEX/AVX2 and EVEX/AVX-512), arm64 (AArch64 integer, FP,
@@ -461,27 +496,49 @@ func cmdAsm(args []string) int {
 With -o the output is written to a file instead.  The --format flag selects
 what is written: raw (the default) concatenates the functions and the data
 section into one self-consistent image; elf emits a relocatable object
-(.text/.data sections, a symbol table and one PC32 relocation per
-static-symbol reference) that links with the system toolchain; goobj emits
-the Go toolchain's own object format, which cmd/link consumes directly (it
-requires -p, the package path, and the installed Go toolchain).
+(.text/.data sections, a symbol table and one relocation per static-symbol
+reference, in the architecture's own form: R_X86_64_PC32 on amd64,
+R_AARCH64_*, R_RISCV_* or R_LARCH_* on the others) that links with the
+system toolchain; goobj emits the Go toolchain's own object format, which
+cmd/link consumes directly (it requires -p, the package path, and the
+installed Go toolchain: the object preamble is captured from go tool asm
+and the format version from go version).
 `)
 	out := fs.String("o", "", "write the output to this file")
 	format := fs.String("format", "raw", "output format: raw (concatenated image), elf or goobj (Go object)")
 	pkg := fs.String("p", "", "package path for --format goobj (qualifies the exported symbols)")
+	archName := fs.String("GOARCH", "", "target architecture: amd64, arm64, riscv64 or loong64 (overrides the file-name suffix)")
+	var dirs includeDirs
+	fs.Var(&dirs, "I", "directory to search for #include files (may be repeated)")
 	fs.Parse(args)
 	if fs.NArg() != 1 {
-		fmt.Fprintln(os.Stderr, "usage: gasm asm [--format raw|elf|goobj] [-p pkg] [-o out] <file>")
+		fmt.Fprintln(os.Stderr, "usage: gasm asm [--format raw|elf|goobj] [-I dir] [-p pkg] [-GOARCH arch] [-o out] <file>")
+		return 2
+	}
+	// The format is validated before anything else, so a bogus value exits 2
+	// with or without -o instead of silently dumping the hex of a raw image.
+	switch *format {
+	case "raw", "elf", "goobj":
+	default:
+		fmt.Fprintf(os.Stderr, "gasm asm: unknown format %q (want raw, elf or goobj)\n", *format)
 		return 2
 	}
 	path := fs.Arg(0)
 	targetArch := arch.FromFilename(path)
+	if *archName != "" {
+		a, err := auditArch(*archName)
+		if err != nil {
+			fmt.Fprintf(os.Stderr, "gasm asm: %v\n", err)
+			return 2
+		}
+		targetArch = a
+	}
 	src, err := readSource(path)
 	if err != nil {
 		fmt.Fprintln(os.Stderr, "gasm:", err)
 		return 1
 	}
-	f, errs := parser.Parse(path, src)
+	f, errs := parser.ParseWithOptions(path, src, parser.Options{Expand: true, IncludeDirs: dirs})
 	for _, e := range errs {
 		fmt.Fprintf(os.Stderr, "%s: %v\n", path, e)
 	}
@@ -494,10 +551,15 @@ requires -p, the package path, and the installed Go toolchain).
 		fmt.Fprintf(os.Stderr, "%s: %v\n", path, err)
 		return 1
 	}
-	if len(img.Funcs) == 0 {
-		fmt.Fprintln(os.Stderr, "gasm asm: no assemblable TEXT functions found")
+	if len(img.Funcs) == 0 && len(img.Data) == 0 {
+		// A file with neither code nor data assembles to nothing, which is
+		// almost always a wrong architecture rather than an intent.
+		fmt.Fprintln(os.Stderr, "gasm asm: no assemblable TEXT functions or GLOBL data found")
 		return 1
 	}
+	// Without -o the hex dump on stdout is the output; with -o the file is,
+	// and the dump is skipped, as the -o help text promises.
+	if *out == "" {
 		for _, fn := range img.Funcs {
 			code := img.Code[fn.Offset : fn.Offset+fn.Size]
 			fmt.Printf("%s: %d bytes\n", fn.Name, fn.Size)
@@ -532,6 +594,7 @@ requires -p, the package path, and the installed Go toolchain).
 				fmt.Println()
 			}
 		}
+	}
 	if *out != "" {
 		var obj []byte
 		var err error
@@ -567,9 +630,6 @@ requires -p, the package path, and the installed Go toolchain).
 				obj, err = img.GOObject(*pkg, path)
 			}
 			kind = "Go object"
-		default:
-			fmt.Fprintf(os.Stderr, "gasm asm: unknown format %q (want raw, elf or goobj)\n", *format)
-			return 2
 		}
 		if err != nil {
 			fmt.Fprintln(os.Stderr, "gasm asm:", err)
@@ -586,7 +646,7 @@ requires -p, the package path, and the installed Go toolchain).

 // cmdDiff compares the machine code of two assembly files.
 func cmdDiff(args []string) int {
-	set := newCommand("diff", "gasm diff <file1.s> <file2.s>", `
+	set := newCommand("diff", "gasm diff [-GOARCH arch] [-I dir] <file1.s> <file2.s>", `
 Compare the machine code produced by assembling two files.
 Shows which functions differ and the byte-level differences.
 Useful for verifying that two implementations produce identical code,
@@ -596,12 +656,24 @@ Use --map to compare functions whose names differ between the files,
 e.g. --map wideCopyAVX2=wideCopyAVX512 pairs the two regardless of suffix.
 `)
 	mapSpec := set.String("map", "", "comma-separated old=new pairs to match functions with different names")
+	archName := set.String("GOARCH", "", "target architecture for both files: amd64, arm64, riscv64 or loong64")
+	var dirs includeDirs
+	set.Var(&dirs, "I", "directory to search for #include files (may be repeated)")
 	set.Parse(args)
 	if set.NArg() != 2 {
-		fmt.Fprintln(os.Stderr, "usage: gasm diff <file1.s> <file2.s>")
+		fmt.Fprintln(os.Stderr, "usage: gasm diff [-GOARCH arch] [-I dir] <file1.s> <file2.s>")
 		return 2
 	}
 	path1, path2 := set.Arg(0), set.Arg(1)
+	forced := arch.Unknown
+	if *archName != "" {
+		a, err := auditArch(*archName)
+		if err != nil {
+			fmt.Fprintf(os.Stderr, "gasm diff: %v\n", err)
+			return 2
+		}
+		forced = a
+	}

 	// Parse the name mapping (file1 name → file2 name).
 	nameMap := make(map[string]string)
@@ -617,12 +689,12 @@ e.g. --map wideCopyAVX2=wideCopyAVX512 pairs the two regardless of suffix.
 	}

 	// Assemble both files.
-	img1, err := assemblePath(path1)
+	img1, err := assemblePath(path1, forced, dirs)
 	if err != nil {
 		fmt.Fprintf(os.Stderr, "gasm diff: %s: %v\n", path1, err)
 		return 1
 	}
-	img2, err := assemblePath(path2)
+	img2, err := assemblePath(path2, forced, dirs)
 	if err != nil {
 		fmt.Fprintf(os.Stderr, "gasm diff: %s: %v\n", path2, err)
 		return 1
@@ -697,20 +769,26 @@ func assembleFile(targetArch arch.Arch, f *ast.File) (*asm.Image, error) {
 	}
 }

-// assemblePath reads, parses and assembles a file (used by cmdDiff).
-func assemblePath(path string) (*asm.Image, error) {
+// assemblePath reads, preprocesses, parses and assembles a file (used by
+// cmdDiff).  A non-Unknown forced architecture overrides the file-name
+// suffix.
+func assemblePath(path string, forced arch.Arch, dirs includeDirs) (*asm.Image, error) {
 	src, err := readSource(path)
 	if err != nil {
 		return nil, err
 	}
-	f, errs := parser.Parse(path, src)
+	f, errs := parser.ParseWithOptions(path, src, parser.Options{Expand: true, IncludeDirs: dirs})
 	for _, e := range errs {
 		fmt.Fprintf(os.Stderr, "%s: %v\n", path, e)
 	}
 	if len(errs) > 0 {
 		return nil, fmt.Errorf("parse errors")
 	}
-	return assembleFile(arch.FromFilename(path), f)
+	target := forced
+	if target == arch.Unknown {
+		target = arch.FromFilename(path)
+	}
+	return assembleFile(target, f)
 }

 // printByteDiff shows the first few byte differences between two code blocks.
@@ -733,8 +811,8 @@ func cmdProfile(args []string) int {
 	flagSet := newCommand("profile", "gasm profile <file.s>", `
 Show the basic-block structure of functions in an assembly file.
 Lists each function's labels, their offsets, and the block boundaries.
-This is the static structure; for runtime execution counts, use
-gasm verify --fuzz which exercises the code paths.
+This is the static structure; for runtime execution counts use
+gasm debug --cover, and for input coverage gasm verify --fuzz.
 `)
 	flagSet.Parse(args)
 	if flagSet.NArg() != 1 {
@@ -878,11 +956,40 @@ func compareGroundTruth(img *asm.Image, gt map[string][]byte) (matched, total, d
 				goCmp[j] = 0
 			}
 		}
-		if bytes.Equal(gasmCmp, goCmp) {
+		// The toolchain pads text symbols to 16-byte boundaries with
+		// zeros, so a function whose size is not a multiple of 16
+		// carries trailing zeros in the ground truth that are not part
+		// of the encoding.  Compare up to the shorter side and require
+		// the remainder of whichever is longer to be zero, so padding
+		// never masks a real difference.
+		cmpLen := min(len(gasmCmp), len(goCmp))
+		equal := bytes.Equal(gasmCmp[:cmpLen], goCmp[:cmpLen])
+		if equal {
+			for _, b := range gasmCmp[cmpLen:] {
+				if b != 0 {
+					equal = false
+					break
+				}
+			}
+		}
+		if equal {
+			for _, b := range goCmp[cmpLen:] {
+				if b != 0 {
+					equal = false
+					break
+				}
+			}
+		}
+		if equal {
 			matched++
-			if len(fn.Relocs) > 0 {
+			switch {
+			case len(fn.Relocs) > 0 && len(goCmp) > cmpLen:
+				fmt.Printf("  %s: MATCH (%d bytes, %d relocs masked, %d padding)\n", fn.Name, fn.Size, len(fn.Relocs), len(goCmp)-cmpLen)
+			case len(fn.Relocs) > 0:
 				fmt.Printf("  %s: MATCH (%d bytes, %d relocs masked)\n", fn.Name, fn.Size, len(fn.Relocs))
-			} else {
+			case len(goCmp) > cmpLen:
+				fmt.Printf("  %s: MATCH (%d bytes, %d padding)\n", fn.Name, fn.Size, len(goCmp)-cmpLen)
+			default:
 				fmt.Printf("  %s: MATCH (%d bytes)\n", fn.Name, fn.Size)
 			}
 		} else {
@@ -930,8 +1037,8 @@ that tolerate nil pointers and zero lengths in their arguments.
 With -abi, each function is called with sentinel values in the registers
 the Go ABI fixes across calls (the frame pointer and the goroutine
 pointer) plus a canary below SP; violations are reported.  JIT-based
-checks run when the host matches the file's architecture (all but
-loong64, which is ground-truth only for now).
+checks run when the host matches the file's architecture, on all four
+architectures.

 With -fuzz, each function with a // func signature is differentially fuzzed
 against the go-tool-asm version in a subprocess (so a crash on a partial
@@ -944,7 +1051,8 @@ With -profile, the static basic-block structure is listed for each function.

 With -call, a single function is invoked with user-supplied buffers (-buf)
 instead of the smoke/abi/fuzz sweeps.  Useful for partial functions (e.g.
-decoders) that crash on random input but should succeed on valid data.
+decoders) that crash on random input but should succeed on valid data.  The
+function named must be NOSPLIT: a function with a stack frame is refused.

 With -save-corpus (and -fuzz), every input that crashes or mismatches is
 written to the directory as replayable JSON.  -replay re-runs saved
@@ -973,20 +1081,16 @@ each entry reproduces.
 	path := set.Arg(0)
 	targetArch := arch.FromFilename(path)
 	// JIT execution runs when the host CPU matches the kernel's
-	// architecture, except loong64: its trampoline is implemented but not
-	// yet validated against real hardware (the Go runtime cannot start
-	// under the available loong64 emulators), so those kernels take the
-	// toolchain-comparison path.
-	if targetArch != hostArch() || targetArch == arch.LOONG64 {
-		// No JIT on this host: ground truth and profile remain available.
-		// (loong64 is ground-truth-only everywhere for now: its trampoline
-		// is implemented but not yet validated against real hardware.)
+	// architecture; every trampoline is validated end to end under
+	// qemu-user emulation (the loong64 one included, via the raw-address
+	// leave handoff).
+	if targetArch != hostArch() {
+		// No JIT on this host: ground truth and profile remain available for
+		// every architecture, because cmdVerifyNonJIT assembles and compares
+		// against the toolchain without executing anything.
 		switch targetArch {
-		case arch.RISCV, arch.LOONG64, arch.ARM64:
+		case arch.AMD64, arch.RISCV, arch.LOONG64, arch.ARM64:
 			return cmdVerifyNonJIT(path, targetArch, *groundTruth, *profile)
-		case arch.AMD64:
-			fmt.Fprintln(os.Stderr, "gasm verify: JIT-based checks need an amd64 host; use --ground-truth here")
-			return 1
 		default:
 			fmt.Fprintln(os.Stderr, "gasm verify: unsupported architecture")
 			return 1
@@ -13,6 +13,9 @@ import (
 	"strings"
 	"syscall"
 	"testing"
+
+	"sourcedock.dev/petrbalvin/gasm-devkit/arch"
+	"sourcedock.dev/petrbalvin/gasm-devkit/asm"
 )

 const clean = "#include \"textflag.h\"\n" +
@@ -219,8 +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)
 	}
 }

@@ -241,6 +245,96 @@ func TestCmdArgErrors(t *testing.T) {
 	}
 }

+// 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.
@@ -274,7 +368,7 @@ func TestVerifySmokeCrashIsolation(t *testing.T) {
 	}
 	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)
+			t.Fatalf("verify died from %v; the crash was not isolated:\n%s", ws.Signal(), out)
 		}
 	}
 	if !strings.Contains(string(out), "CRASH") {
@@ -292,3 +386,77 @@ func TestSweepCheckLines(t *testing.T) {
 		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)
+	}
+}
@@ -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), " ")
+}
@@ -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)
+	}
+}
@@ -44,7 +44,7 @@ bodies, place the file in the kernel's package, and run it in CI.
 		rest = rest[1:]
 	}
 	if len(rest) != 1 {
-		return fmt.Errorf("usage: gasm scaffold differential <file.s>")
+		return &usageError{fmt.Errorf("usage: gasm scaffold differential <file.s>")}
 	}
 	path := rest[0]
 	src, err := os.ReadFile(path)
@@ -9,11 +9,11 @@ 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
@@ -32,8 +32,12 @@ type Condition struct {
 	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 {
+// 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
@@ -48,9 +52,16 @@ func (c *Condition) Eval(regs *Regs) bool {
 		}
 		expected = v
 	case c.MemAddr != 0:
-		// Register-memory comparison, requires a Session, not available here.
-		// Fall back to treating as constant (the caller should resolve).
-		expected = c.Value
+		// 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:
 		expected = c.Value
 	}
@@ -72,6 +83,18 @@ func (c *Condition) Eval(regs *Regs) bool {
 	}
 }

+// 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
@@ -83,6 +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)
@@ -100,13 +135,12 @@ func (bm *Breakpoints) SetWithCond(addr uint64, label string, cond *Condition) (
 	if err != nil {
 		return nil, err
 	}
-	orig := byte(word)
-	// Patch with the breakpoint instruction, preserving the rest of the word.
-	mask := uint64(0)
-	for range breakpointInsn {
-		mask = (mask << 8) | 0xFF
+	orig := make([]byte, len(breakpointInsn))
+	for i := range orig {
+		orig[i] = byte(word >> (8 * i))
 	}
-	patched := (word &^ mask) | breakpointWord(breakpointInsn)
+	// 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
 	}
@@ -134,26 +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.WriteString(fmt.Sprintf("  %d: %s at %#x [%s, %d hits]%s\n", i, label, bp.Addr, status, bp.hits, cond))
 	}
 	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)
 	}
+	if !bm.restore(addr, bp) {
 		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
+		return fmt.Errorf("debug: restore breakpoint at %#x failed, word is %#x", addr, word)
 	}
 	delete(bm.bps, addr)
 	return nil
@@ -185,43 +233,54 @@ 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 (PC-adjust matches
-// a breakpoint address), restores the original byte, rewinds PC, and
+// 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 a breakpoint trap, PC points past the breakpoint instruction.
+	// 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 PC to re-execute the original instruction.
+	// 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().
@@ -234,11 +293,7 @@ func (bm *Breakpoints) Reinsert(addr uint64) error {
 	if err != nil {
 		return err
 	}
-	mask := uint64(0)
-	for range breakpointInsn {
-		mask = (mask << 8) | 0xFF
-	}
-	patched := (word &^ mask) | breakpointWord(breakpointInsn)
+	patched := (word &^ breakpointMask()) | breakpointWord(breakpointInsn)
 	return bm.t.Poke(addr, patched)
 }

@@ -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)
+		}
+	}
+}
@@ -6,7 +6,6 @@
 package debug

 import (
-	"strings"
 	"testing"
 )

@@ -48,7 +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)
@@ -56,65 +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},
+// 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
 	}

-	tests := []struct {
-		offset int
-		want   int
-	}{
-		{0, 5},
-		{1, 5},
-		{4, 5},
-		{5, 6},
-		{7, 6},
-		{10, 7},
-		{12, 7},
-		{15, 8},
-		{20, 8},
+	eq := Condition{Reg: "rax", Op: "==", MemAddr: 0x5000}
+	if !eq.Eval(regs, mem) {
+		t.Error("register-memory comparison with matching word should hold")
 	}
-
-	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)
+	ne := Condition{Reg: "rax", Op: "!=", MemAddr: 0x5000}
+	if ne.Eval(regs, mem) {
+		t.Error("register-memory comparison with mismatching word should not hold")
 	}
+	bad := Condition{Reg: "rax", Op: "==", MemAddr: 0x6000}
+	if !bad.Eval(regs, mem) {
+		t.Error("unreadable memory must not block the breakpoint")
 	}
-
-	// 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)
-		}
+	noReader := Condition{Reg: "rax", Op: "==", MemAddr: 0x5000}
+	if !noReader.Eval(regs, nil) {
+		t.Error("missing memory reader must not block the breakpoint")
 	}
 }

@@ -141,139 +108,6 @@ 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{} // per-session slots start free

@@ -323,3 +157,19 @@ func TestWatchpointSlotTracking(t *testing.T) {
 		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)
+	}
+}
@@ -46,3 +46,11 @@ func (s *Session) DisassembleN(addr uint64, n int) string {
 	}
 	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"
+}
@@ -7,6 +7,7 @@ package debug

 import (
 	"fmt"
+	"strings"

 	"sourcedock.dev/petrbalvin/gasm-devkit/arch"
 	"sourcedock.dev/petrbalvin/gasm-devkit/disasm"
@@ -44,3 +45,15 @@ func (s *Session) DisassembleN(addr uint64, n int) string {
 	}
 	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
+}
@@ -7,6 +7,7 @@ package debug

 import (
 	"fmt"
+	"strings"

 	"sourcedock.dev/petrbalvin/gasm-devkit/arch"
 	"sourcedock.dev/petrbalvin/gasm-devkit/disasm"
@@ -44,3 +45,16 @@ func (s *Session) DisassembleN(addr uint64, n int) string {
 	}
 	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
+}
@@ -7,6 +7,7 @@ package debug

 import (
 	"fmt"
+	"strings"

 	"sourcedock.dev/petrbalvin/gasm-devkit/arch"
 	"sourcedock.dev/petrbalvin/gasm-devkit/disasm"
@@ -44,3 +45,17 @@ func (s *Session) DisassembleN(addr uint64, n int) string {
 	}
 	return result
 }
+
+// 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
+}
@@ -73,9 +73,29 @@ func decodeRflags(f uint64) string {
 	return flags[:len(flags)-1]
 }

-// archReturnAddr reads the return address from the stack (amd64 ABI0 convention).
+// archReturnAddr reads the return address of the current frame (amd64
+// ABI0 convention).  A function that contains a CALL (or has a frame) is
+// assembled with the prologue PUSHQ BP; MOVQ SP, BP, so mid-function the
+// word at SP is the saved caller BP, a stack address, and the return
+// address sits further up.  Walk the stack from SP and take the first word
+// that lies in an executable mapping: stack and data words never do, a
+// return address always does.
 func archReturnAddr(s *Session, regs *Regs) (uint64, error) {
-	return s.Peek(regs.GetSP())
+	ranges := execRanges(s.pid)
+	for off := uint64(0); off < 512; off += 8 {
+		word, err := s.Peek(regs.RSP + off)
+		if err != nil {
+			break
+		}
+		for _, r := range ranges {
+			if word >= r.lo && word < r.hi {
+				return word, nil
+			}
+		}
+	}
+	// No mapping available or nothing code-like on the stack: fall back to
+	// the raw entry convention, [SP] before any push.
+	return s.Peek(regs.RSP)
 }

 // archSPLabel returns the SP register name for display.
@@ -5,7 +5,10 @@

 package debug

-import "fmt"
+import (
+	"encoding/binary"
+	"fmt"
+)

 func printRegs(regs *Regs, codeBase, funcOff uint64) {
 	fmt.Printf("  PC  = %#016x  (func+%#x)\n", regs.PC, regs.PC-codeBase-funcOff)
@@ -31,8 +34,8 @@ func printRegs(regs *Regs, codeBase, funcOff uint64) {
 func printVectorRegs(v *VectorRegs) {
 	fmt.Println("\n  Vector registers (V0-V31):")
 	for i := 0; i < 32; i += 2 {
-		fmt.Printf("  V%-2d  = %016x%016x\n", i, v.V[i][8], v.V[i][0])
-		fmt.Printf("  V%-2d  = %016x%016x\n", i+1, v.V[i+1][8], v.V[i+1][0])
+		fmt.Printf("  V%-2d  = %016x%016x\n", i, binary.LittleEndian.Uint64(v.V[i][8:16]), binary.LittleEndian.Uint64(v.V[i][0:8]))
+		fmt.Printf("  V%-2d  = %016x%016x\n", i+1, binary.LittleEndian.Uint64(v.V[i+1][8:16]), binary.LittleEndian.Uint64(v.V[i+1][0:8]))
 	}
 }

@@ -0,0 +1,427 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+//go:build linux && amd64
+
+package debug
+
+import (
+	"bytes"
+	"fmt"
+	"io"
+	"os"
+	"path/filepath"
+	"runtime"
+	"strings"
+	"testing"
+	"time"
+	"unsafe"
+
+	"sourcedock.dev/petrbalvin/gasm-devkit/asm"
+	"sourcedock.dev/petrbalvin/gasm-devkit/verify"
+)
+
+// Integration tests beyond the basic entry breakpoint: hardware watchpoints,
+// conditional breakpoints, next/finish over a CALL, faulting kernels and the
+// xstate vector-register readout.  All drive a real ptrace session, so they
+// run on amd64 hosts only.
+
+// writeKernel writes an assembly source to a temporary file with the
+// architecture suffix the assembler dispatcher expects.
+func writeKernel(t *testing.T, src string) string {
+	t.Helper()
+	path := filepath.Join(t.TempDir(), "kernel_amd64.s")
+	if err := os.WriteFile(path, []byte(src), 0o644); err != nil {
+		t.Fatalf("write kernel: %v", err)
+	}
+	return path
+}
+
+// launchKernel launches a session for the kernel source and returns the
+// session, its breakpoint manager and the function layout.
+func launchKernel(t *testing.T, bin, path, funcName string, args []byte) (*Session, *Breakpoints, asm.FuncLayout) {
+	t.Helper()
+	k, err := verify.Load(path)
+	if err != nil {
+		t.Fatalf("Load: %v", err)
+	}
+	t.Cleanup(k.Close)
+	fl, err := k.Func(funcName)
+	if err != nil {
+		t.Fatalf("Func: %v", err)
+	}
+	if len(args) < fl.Args {
+		padded := make([]byte, fl.Args)
+		copy(padded, args)
+		args = padded
+	}
+	sess, err := Launch(bin, path, funcName, args)
+	if err != nil {
+		t.Fatalf("Launch: %v", err)
+	}
+	t.Cleanup(sess.Kill)
+	bm := NewBreakpoints(sess)
+	return sess, bm, fl
+}
+
+// runToEntry resumes the freshly launched debuggee until the breakpoint at
+// the function entry traps, mirroring the REPL continue loop: the debuggee
+// SIGSTOPs twice (launch barrier and entry barrier) before entering the JIT
+// call.
+func runToEntry(t *testing.T, sess *Session, bm *Breakpoints, entry uint64) {
+	t.Helper()
+	for range 50 {
+		for _, bp := range bm.All() {
+			bm.Reinsert(bp.Addr)
+		}
+		if err := sess.Continue(); err != nil {
+			t.Fatalf("Continue: %v", err)
+		}
+		if sess.Exited() {
+			t.Fatal("debuggee exited before the entry breakpoint trapped")
+		}
+		regs, err := sess.GetRegs()
+		if err != nil {
+			t.Fatalf("GetRegs: %v", err)
+		}
+		if bm.HandleTrap(&regs) != nil {
+			return
+		}
+	}
+	t.Fatal("no entry breakpoint trap after 50 resumes")
+}
+
+// captureStdout runs fn with os.Stdout redirected to a pipe and returns
+// what it printed (the REPL writes its reports to stdout).
+func captureStdout(t *testing.T, fn func()) string {
+	t.Helper()
+	r, w, err := os.Pipe()
+	if err != nil {
+		t.Fatalf("pipe: %v", err)
+	}
+	old := os.Stdout
+	os.Stdout = w
+	done := make(chan string, 1)
+	go func() {
+		b, _ := io.ReadAll(r)
+		done <- string(b)
+	}()
+	defer func() { os.Stdout = old }()
+	fn()
+	w.Close()
+	return <-done
+}
+
+// TestWatchpointArmRunHit proves the debug-register offsets: the watchpoint
+// must fire on the store, with si_addr naming the watched address.  The
+// kernel writes its return value to ret+0(FP), which is the 8-byte word
+// right above the stack pointer at entry.
+func TestWatchpointArmRunHit(t *testing.T) {
+	runtime.LockOSThread()
+	defer runtime.UnlockOSThread()
+	bin := buildGasm(t)
+
+	const kernel = `#include "textflag.h"
+
+// func wpret() int64
+TEXT ·wpret(SB), NOSPLIT, $0-8
+	MOVQ $0x5a5a5a5a5a5a5a5a, AX
+	MOVQ AX, ret+0(FP)
+	RET
+`
+	path := writeKernel(t, kernel)
+	sess, bm, fl := launchKernel(t, bin, path, "wpret", nil)
+
+	entry := sess.CodeBase() + uint64(fl.Offset)
+	if _, err := bm.Set(entry, "entry"); err != nil {
+		t.Fatalf("Set: %v", err)
+	}
+	runToEntry(t, sess, bm, entry)
+
+	regs, err := sess.GetRegs()
+	if err != nil {
+		t.Fatalf("GetRegs: %v", err)
+	}
+	watched := regs.RSP + 8 // ret+0(FP): the store target
+
+	slot := sess.FindFreeWatchpointSlot()
+	if slot < 0 {
+		t.Fatal("no free watchpoint slot")
+	}
+	if err := sess.SetWatchpoint(slot, watched, WatchWrite, 8); err != nil {
+		t.Fatalf("SetWatchpoint: %v (wrong debug-register offsets?)", err)
+	}
+
+	if err := sess.Continue(); err != nil {
+		t.Fatalf("Continue: %v", err)
+	}
+	reason, addr := sess.StopInfo()
+	if reason != StopWatchpoint {
+		t.Fatalf("stop reason = %v, want StopWatchpoint (DR0-DR3/DR7 offsets are wrong)", reason)
+	}
+	if addr != watched {
+		t.Fatalf("watchpoint address = %#x, want %#x", addr, watched)
+	}
+
+	// The watched word holds the stored value: x86 data breakpoints are
+	// reported with the access complete.
+	if word, err := sess.Peek(watched); err != nil || word != 0x5a5a5a5a5a5a5a5a {
+		t.Errorf("watched word = %#x (err %v), want 0x5a5a5a5a5a5a5a5a", word, err)
+	}
+	if err := sess.ClearWatchpoint(slot); err != nil {
+		t.Fatalf("ClearWatchpoint: %v", err)
+	}
+}
+
+// TestConditionalBreakpointFalseThenTrue proves the false-condition path:
+// the breakpoint steps over the original instruction, re-arms itself and
+// keeps running silently, and the true condition stops exactly once with the
+// register in the expected state.
+func TestConditionalBreakpointFalseThenTrue(t *testing.T) {
+	runtime.LockOSThread()
+	defer runtime.UnlockOSThread()
+	bin := buildGasm(t)
+
+	const kernel = `#include "textflag.h"
+
+// func countdown(n int64) int64
+TEXT ·countdown(SB), NOSPLIT, $0-16
+	MOVQ n+0(FP), CX
+loop:
+	DECQ CX
+	CMPQ CX, $0
+	JNE loop
+	MOVQ CX, ret+8(FP)
+	RET
+`
+	path := writeKernel(t, kernel)
+	sess, bm, fl := launchKernel(t, bin, path, "countdown", []byte{8})
+
+	loopAddr := sess.CodeBase() + uint64(fl.Offset) + uint64(fl.Labels["loop"])
+	// The length of the breakpointed instruction, from a disassembly taken
+	// before the INT3 is patched in.
+	_, insnLen, err := sess.Disassemble(loopAddr)
+	if err != nil || insnLen <= 0 {
+		t.Fatalf("Disassemble at %#x: len=%d err=%v", loopAddr, insnLen, err)
+	}
+	cond := &Condition{Reg: "rcx", Op: "==", Value: 1}
+	bp, err := bm.SetWithCond(loopAddr, "loop", cond)
+	if err != nil {
+		t.Fatalf("SetWithCond: %v", err)
+	}
+
+	hits := 0
+	exited := false
+	for range 200 {
+		for _, b := range bm.All() {
+			bm.Reinsert(b.Addr)
+		}
+		if err := sess.Continue(); err != nil {
+			exited = true
+			break // the debuggee finished
+		}
+		if sess.Exited() {
+			exited = true
+			break
+		}
+		if sig := sess.LastSignal(); sig != 0 {
+			t.Fatalf("unexpected signal stop %v", sig)
+		}
+		regs, err := sess.GetRegs()
+		if err != nil {
+			t.Fatalf("GetRegs: %v", err)
+		}
+		if hit := bm.HandleTrap(&regs); hit != nil {
+			hits++
+			if regs.RCX != 1 {
+				t.Fatalf("hit with RCX=%d, want 1", regs.RCX)
+			}
+			// Park after the instruction, as the REPL does.
+			if err := sess.Step(); err != nil {
+				t.Fatalf("Step: %v", err)
+			}
+		} else {
+			// A false evaluation must leave the debuggee past the whole
+			// original instruction: a PC inside it (trapAddr+1 on amd64)
+			// means the resume happens mid-instruction.
+			fresh, err := sess.GetRegs()
+			if err != nil {
+				t.Fatalf("GetRegs: %v", err)
+			}
+			if fresh.RIP > loopAddr && fresh.RIP < loopAddr+uint64(insnLen) {
+				t.Fatalf("false evaluation left the PC at %#x, inside the %d-byte instruction at %#x",
+					fresh.RIP, insnLen, loopAddr)
+			}
+		}
+	}
+	if hits != 1 {
+		t.Fatalf("conditional breakpoint hit %d times, want exactly 1 (false evaluations must run through silently)", hits)
+	}
+	if bp.Hits() != 1 {
+		t.Errorf("bp.Hits() = %d, want 1", bp.Hits())
+	}
+	if !exited || !sess.Exited() {
+		t.Fatal("debuggee did not run to completion after the conditional hit")
+	}
+}
+
+// TestNextAndFinishOverCall proves next and finish evaluate the trap with
+// registers fetched after the stop: next lands exactly on the instruction
+// after the CALL, and finish stops exactly on the return address.
+func TestNextAndFinishOverCall(t *testing.T) {
+	runtime.LockOSThread()
+	defer runtime.UnlockOSThread()
+	bin := buildGasm(t)
+
+	const kernel = `#include "textflag.h"
+
+// func caller(x int64) int64
+// The argument travels in AX: FP argument slots of CALL-bearing functions
+// are an assembler concern outside this test's scope.
+TEXT ·caller(SB), NOSPLIT, $0-16
+	MOVQ $5, AX
+	CALL ·bump(SB)
+aftercall:
+	MOVQ AX, ret+8(FP)
+	RET
+
+// func bump(x int64) int64
+TEXT ·bump(SB), NOSPLIT, $0-0
+	ADDQ $3, AX
+	RET
+`
+	path := writeKernel(t, kernel)
+
+	// next: step the prologue and the constant load (3 instructions), then
+	// step over the CALL and check the landing address and RAX.
+	sess, bm, fl := launchKernel(t, bin, path, "caller", nil)
+	entry := sess.CodeBase() + uint64(fl.Offset)
+	if _, err := bm.Set(entry, "entry"); err != nil {
+		t.Fatalf("Set: %v", err)
+	}
+	runToEntry(t, sess, bm, entry)
+	afterOff := uint64(fl.Labels["aftercall"])
+
+	out := captureStdout(t, func() {
+		REPL(sess, bm, sess.CodeBase(), fl.Offset, fl.Size, fl.Args, nil, nil,
+			strings.NewReader("step 3\nnext\nregs\nquit\n"))
+	})
+	if !strings.Contains(out, fmt.Sprintf("func+%#x", afterOff)) {
+		t.Errorf("next did not land on the instruction after the CALL (func+%#x); output:\n%s", afterOff, out)
+	}
+	if !strings.Contains(out, "RAX = 0x0000000000000008") {
+		t.Errorf("callee did not run exactly once under next (want RAX=8); output:\n%s", out)
+	}
+
+	// finish: run to the return address read off the stack at entry.
+	sess2, bm2, fl2 := launchKernel(t, bin, path, "caller", nil)
+	entry2 := sess2.CodeBase() + uint64(fl2.Offset)
+	if _, err := bm2.Set(entry2, "entry"); err != nil {
+		t.Fatalf("Set: %v", err)
+	}
+	runToEntry(t, sess2, bm2, entry2)
+	regs, err := sess2.GetRegs()
+	if err != nil {
+		t.Fatalf("GetRegs: %v", err)
+	}
+	retAddr, err := sess2.Peek(regs.RSP)
+	if err != nil {
+		t.Fatalf("Peek return address: %v", err)
+	}
+
+	out2 := captureStdout(t, func() {
+		REPL(sess2, bm2, sess2.CodeBase(), fl2.Offset, fl2.Size, fl2.Args, nil, nil,
+			strings.NewReader("step 1\nfinish\nquit\n"))
+	})
+	want := fmt.Sprintf("finished, now at %#x\n", retAddr)
+	if !strings.Contains(out2, want) {
+		t.Errorf("finish stopped at the wrong PC; want %q in output:\n%s", want, out2)
+	}
+}
+
+// TestSignalStopSurfaced proves a faulting kernel surfaces as a reported
+// stop instead of an infinite fault loop.  A regression here hangs, so a
+// watchdog fails the run rather than letting CI stall.
+func TestSignalStopSurfaced(t *testing.T) {
+	runtime.LockOSThread()
+	defer runtime.UnlockOSThread()
+	bin := buildGasm(t)
+
+	const kernel = `#include "textflag.h"
+
+// func crash() int64
+TEXT ·crash(SB), NOSPLIT, $0-8
+	XORQ AX, AX
+	MOVQ (AX), AX
+	MOVQ AX, ret+0(FP)
+	RET
+`
+	path := writeKernel(t, kernel)
+	sess, bm, _ := launchKernel(t, bin, path, "crash", nil)
+
+	timer := time.AfterFunc(time.Minute, func() {
+		panic("watchdog: the debugger hung on the faulting kernel instead of reporting the signal stop")
+	})
+	defer timer.Stop()
+
+	out := captureStdout(t, func() {
+		REPL(sess, bm, sess.CodeBase(), 0, 0, 0, nil, nil,
+			strings.NewReader("continue\nquit\n"))
+	})
+	if !strings.Contains(out, "stopped on signal") {
+		t.Errorf("SIGSEGV did not surface as a reported stop; output:\n%s", out)
+	}
+	if !sess.Exited() {
+		t.Error("debuggee should be killed by quit after the signal stop")
+	}
+}
+
+// TestGetVectorRegsXState proves the NT_X86_XSTATE readout: the request
+// succeeds on a normal process and the XMM halves agree with
+// PTRACE_GETFPREGS.
+func TestGetVectorRegsXState(t *testing.T) {
+	// The FPRegs layout must mirror the kernel's user_fpregs_struct
+	// exactly: PTRACE_GETFPREGS fills all 512 bytes, so a short struct
+	// overflows the caller's memory.
+	if got := unsafe.Sizeof(FPRegs{}); got != 512 {
+		t.Fatalf("sizeof(FPRegs) = %d, want 512", got)
+	}
+	if got := unsafe.Offsetof(FPRegs{}.XMM); got != 160 {
+		t.Fatalf("offsetof(FPRegs.XMM) = %d, want 160", got)
+	}
+
+	runtime.LockOSThread()
+	defer runtime.UnlockOSThread()
+	bin := buildGasm(t)
+
+	const kernel = `#include "textflag.h"
+
+// func vprobe() int64
+TEXT ·vprobe(SB), NOSPLIT, $0-8
+	MOVQ $1, AX
+	MOVQ AX, ret+0(FP)
+	RET
+`
+	path := writeKernel(t, kernel)
+	sess, bm, fl := launchKernel(t, bin, path, "vprobe", nil)
+
+	entry := sess.CodeBase() + uint64(fl.Offset)
+	if _, err := bm.Set(entry, "entry"); err != nil {
+		t.Fatalf("Set: %v", err)
+	}
+	runToEntry(t, sess, bm, entry)
+
+	v, err := sess.GetVectorRegs()
+	if err != nil {
+		t.Fatalf("GetVectorRegs: %v", err)
+	}
+	fp, err := sess.GetFPRegs()
+	if err != nil {
+		t.Fatalf("GetFPRegs: %v", err)
+	}
+	for i := range 16 {
+		if !bytes.Equal(v.YMM[i][:16], fp.XMM[i][:]) {
+			t.Errorf("YMM%d low half %x, want the FPRegs XMM half %x", i, v.YMM[i][:16], fp.XMM[i][:])
+		}
+	}
+}
@@ -23,7 +23,13 @@ type Session struct {
 	stopped  bool
 	exited   bool
 	codeBase uint64   // base address of the JIT code in the debuggee
-	wpSlots  [16]bool // hardware watchpoint slots in use (DR0-DR3, arm64 BADVR0-15)
+	tmpDir   string   // scratch directory of the session, removed on Kill
+	wpSlots  [16]bool // hardware watchpoint slots in use (DR0-DR3, arm64 DBGWVR0-15)
+	// lastSignal holds the signal of the most recent stop when that stop
+	// was a genuine signal-delivery-stop the caller must see (a fault such
+	// as SIGSEGV, SIGBUS, SIGFPE or SIGILL); 0 for breakpoint traps,
+	// single-steps, SIGSTOP and suppressed runtime signals.
+	lastSignal syscall.Signal
 }

 // Launch starts the debuggee subprocess (gasm debug --target ...) and
@@ -78,7 +84,7 @@ func LaunchWithBuffers(gasmBin, asmPath, funcName string, args []byte, bufSpec s
 		return nil, nil, fmt.Errorf("debug: start debuggee: %w", err)
 	}

-	s := &Session{pid: cmd.Process.Pid, cmd: cmd}
+	s := &Session{pid: cmd.Process.Pid, cmd: cmd, tmpDir: tmpDir}

 	readyFile := filepath.Join(tmpDir, "ready")
 	for range 500 {
@@ -125,29 +131,17 @@ func LaunchWithBuffers(gasmBin, asmPath, funcName string, args []byte, bufSpec s
 	return s, bufAddrs, nil
 }

-// wait waits for the debuggee to stop and returns the wait status.
-func (s *Session) wait() error {
-	var ws syscall.WaitStatus
-	_, err := syscall.Wait4(s.pid, &ws, 0, nil)
-	if err != nil {
-		return err
-	}
-	if ws.Exited() {
-		s.exited = true
-		return fmt.Errorf("debuggee exited with status %d", ws.ExitStatus())
-	}
-	s.stopped = true
-	return nil
-}
-
 // waitStopped consumes ptrace-stop events until one the debugger cares
-// about arrives: SIGTRAP (a breakpoint or a completed single-step) or the
-// debuggee's own SIGSTOP.  A Go tracee's runtime raises SIGURG for
-// asynchronous preemption, and every signal on a traced thread surfaces as
-// a signal-delivery-stop, so those are suppressed and the tracee resumed
-// without them.  Runtime noise is why a single wait can return in the
-// middle of runtime code and a resume can then fail: the event stream must
-// be drained by the tracer.
+// about arrives: SIGTRAP (a breakpoint or a completed single-step), the
+// debuggee's own SIGSTOP, or a genuine signal-delivery-stop.  A Go tracee's
+// runtime raises SIGURG for asynchronous preemption, and every signal on a
+// traced thread surfaces as a signal-delivery-stop, so SIGURG is suppressed
+// and the tracee resumed without it.  Every other signal (SIGSEGV, SIGBUS,
+// SIGFPE, SIGILL, ...) is returned to the caller: resuming with signal 0
+// would restart the faulting instruction and fault forever, so a faulting
+// kernel must surface as a stop the caller reports.  Runtime noise is also
+// why a single wait can return in the middle of runtime code and a resume
+// can then fail: the event stream must be drained by the tracer.
 func (s *Session) waitStopped() (syscall.Signal, error) {
 	for {
 		var ws syscall.WaitStatus
@@ -165,10 +159,12 @@ func (s *Session) waitStopped() (syscall.Signal, error) {
 		switch sig := ws.StopSignal(); sig {
 		case syscall.SIGTRAP, syscall.SIGSTOP:
 			s.stopped = true
+			s.lastSignal = 0
 			return sig, nil
-		default:
-			// Runtime noise (SIGURG preemption and friends): resume the
-			// tracee without delivering the signal.
+		case syscall.SIGURG:
+			// Go runtime asynchronous preemption: resume the tracee
+			// without delivering the signal.
+			s.lastSignal = 0
 			if _, _, errno := syscall.Syscall6(
 				syscall.SYS_PTRACE,
 				uintptr(syscall.PTRACE_CONT),
@@ -177,10 +173,22 @@ func (s *Session) waitStopped() (syscall.Signal, error) {
 			); errno != 0 {
 				return 0, fmt.Errorf("debug: PTRACE_CONT: %w", errno)
 			}
+		default:
+			// A genuine signal-delivery-stop.  Report it; the caller
+			// decides how to proceed.
+			s.stopped = true
+			s.lastSignal = sig
+			return sig, nil
 		}
 	}
 }

+// LastSignal returns the signal of the most recent stop when that stop was
+// a genuine signal-delivery-stop (a fault such as SIGSEGV, SIGFPE, SIGILL
+// or SIGBUS), and 0 for breakpoint traps, single-steps, SIGSTOP and
+// suppressed runtime signals.
+func (s *Session) LastSignal() syscall.Signal { return s.lastSignal }
+
 // Peek reads a word (8 bytes) from the debuggee's memory at addr.
 func (s *Session) Peek(addr uint64) (uint64, error) {
 	mem, err := os.OpenFile(fmt.Sprintf("/proc/%d/mem", s.pid), os.O_RDONLY, 0)
@@ -294,7 +302,8 @@ func (s *Session) Pid() int { return s.pid }
 // CodeBase returns the base address of the JIT code in the debuggee.
 func (s *Session) CodeBase() uint64 { return s.codeBase }

-// Kill terminates the debuggee.
+// Kill terminates the debuggee and removes the session's scratch
+// directory, so a successful session leaves no gasm-debug-* debris behind.
 func (s *Session) Kill() {
 	if !s.exited {
 		syscall.Kill(s.pid, syscall.SIGKILL)
@@ -304,6 +313,35 @@ func (s *Session) Kill() {
 	if s.cmd != nil && s.cmd.Process != nil {
 		s.cmd.Wait()
 	}
+	if s.tmpDir != "" {
+		os.RemoveAll(s.tmpDir)
+		s.tmpDir = ""
+	}
+}
+
+// execRange is one executable mapping of the debuggee.
+type execRange struct {
+	lo, hi uint64
+}
+
+// execRanges parses the debuggee's executable mappings from /proc/pid/maps.
+func execRanges(pid int) []execRange {
+	data, err := os.ReadFile(fmt.Sprintf("/proc/%d/maps", pid))
+	if err != nil {
+		return nil
+	}
+	var out []execRange
+	for line := range strings.SplitSeq(string(data), "\n") {
+		fields := strings.Fields(line)
+		if len(fields) < 2 || !strings.Contains(fields[1], "x") {
+			continue
+		}
+		var lo, hi uint64
+		if _, err := fmt.Sscanf(fields[0], "%x-%x", &lo, &hi); err == nil {
+			out = append(out, execRange{lo, hi})
+		}
+	}
+	return out
 }

 // findRWXMapping reads /proc/pid/maps and returns the base address of the
@@ -6,6 +6,7 @@
 package debug

 import (
+	"encoding/binary"
 	"fmt"
 	"syscall"
 	"unsafe"
@@ -44,20 +45,24 @@ func (s *Session) SetRegs(regs *Regs) error {
 	return nil
 }

-// FPRegs holds the x87 FPU and SSE (XMM) register state from PTRACE_GETFPREGS.
+// FPRegs holds the x87 FPU and SSE (XMM) register state from
+// PTRACE_GETFPREGS.  The layout is the kernel's struct user_fpregs_struct
+// (sys/user.h), the FXSAVE image: 512 bytes with XMM0-15 at offset 160.
+// The i387 fcs/ds segment fields do not exist in the 64-bit layout.  The
+// size matters: the copy fills all 512 bytes, so a short or misaligned
+// struct makes PTRACE_GETFPREGS overflow the caller's memory.
 type FPRegs struct {
 	FCW       uint16
 	FSW       uint16
-	FTW       byte
+	FTW       uint16
 	FOP       uint16
 	FIP       uint64
-	FCS       uint16
 	FDP       uint64
-	FDS       uint16
 	MXCSR     uint32
 	MXCSRMask uint32
 	ST        [8][16]byte  // x87 stack (10 bytes per reg, padded to 16)
-	XMM       [16][16]byte // XMM0-15
+	XMM       [16][16]byte // XMM0-15, struct offset 160
+	Reserved  [96]byte     // FXSAVE padding, to the full 512 bytes
 }

 // GetFPRegs retrieves the FPU/SSE register state of the stopped debuggee.
@@ -82,16 +87,68 @@ type VectorRegs struct {
 	YMM [16][32]byte // YMM0-15 (full 256-bit values)
 }

-// GetVectorRegs retrieves the YMM registers via PTRACE_GETREGSET + XSAVE.
+// NT_X86_XSTATE (0x202), the xsave extended-state regset
+// (include/uapi/linux/elf.h).
+const ntX86XState = 0x202
+
+// Layout of the buffer PTRACE_GETREGSET returns for NT_X86_XSTATE: the
+// 512-byte legacy fxsave image (x87 state in 0-159, XMM0-15 in 160-511),
+// then the 64-byte xsave header whose first 8 bytes are xstate_bv, then one
+// component per set feature bit, each 64-byte aligned.  The YMM high halves
+// are the first extended component, at offset 576; that offset is fixed by
+// the ISA on AVX-capable x86-64.  XFEATURE_MASK_YMM is bit 2 of xstate_bv
+// (arch/x86/include/asm/fpu/types.h); the high halves are zero when the bit
+// is clear.
+const (
+	xsaveXMMOffset    = 160
+	xsaveXMMSize      = 256
+	xsaveHeaderOffset = 512
+	xsaveBVOffset     = xsaveHeaderOffset
+	ymmOffset         = xsaveHeaderOffset + 64 // 576
+	ymmSize           = 256                    // 16 registers, 16 bytes each
+	xfeatureMaskYMM   = 1 << 2
+	xstateMaxBuffer   = 4096 // CPUID(0xD).xsave_size is far below this
+)
+
+// GetVectorRegs retrieves the YMM registers via PTRACE_GETREGSET on
+// NT_X86_XSTATE.  The low (XMM) halves always come from the legacy image;
+// the high halves are copied only when xstate_bv reports the YMM feature,
+// and read as zero otherwise.  When the regset request fails the FP image
+// still provides correct XMM halves, so that is the fallback.
 func (s *Session) GetVectorRegs() (VectorRegs, error) {
 	var v VectorRegs
+	buf := make([]byte, xstateMaxBuffer)
+	iovec := syscall.Iovec{
+		Base: &buf[0],
+		Len:  uint64(len(buf)),
+	}
+	_, _, errno := syscall.Syscall6(
+		syscall.SYS_PTRACE,
+		uintptr(syscall.PTRACE_GETREGSET),
+		uintptr(s.pid),
+		uintptr(ntX86XState),
+		uintptr(unsafe.Pointer(&iovec)),
+		0, 0,
+	)
+	if errno != 0 {
 		fp, err := s.GetFPRegs()
 		if err != nil {
 			return v, err
 		}
 		for i := range 16 {
-		for j := range 16 {
-			v.YMM[i][j] = fp.XMM[i][j]
+			copy(v.YMM[i][:16], fp.XMM[i][:])
+		}
+		return v, nil
+	}
+	n := int(iovec.Len)
+	for i := range 16 {
+		copy(v.YMM[i][:16], buf[xsaveXMMOffset+16*i:xsaveXMMOffset+16*i+16])
+	}
+	if n >= ymmOffset+ymmSize {
+		if binary.LittleEndian.Uint64(buf[xsaveBVOffset:xsaveBVOffset+8])&xfeatureMaskYMM != 0 {
+			for i := range 16 {
+				copy(v.YMM[i][16:], buf[ymmOffset+16*i:ymmOffset+16*i+16])
+			}
 		}
 	}
 	return v, nil
@@ -91,5 +91,9 @@ func (r *Regs) RegValue(name string) (uint64, bool) {
 // breakpointInsn is the software breakpoint instruction.
 var breakpointInsn = []byte{0xCC} // INT3

-// breakpointPCAdjust is how far PC is past the breakpoint instruction after a trap.
+// breakpointPCAdjust is how far PC is past the breakpoint instruction after
+// a trap.  x86-64 reports the #DB for INT3 with RIP on the byte after the
+// INT3 (Intel SDM vol 3, "Debug Exceptions"), so the trap address is
+// PC-1.  The other supported architectures leave the PC on the trap
+// instruction and use 0 there.
 const breakpointPCAdjust = 1
@@ -130,5 +130,11 @@ func (r *Regs) RegValue(name string) (uint64, bool) {
 // breakpointInsn is the software breakpoint instruction (BRK #0).
 var breakpointInsn = []byte{0x00, 0x00, 0x20, 0xD4} // BRK #0

-// breakpointPCAdjust is how far PC is past the breakpoint instruction after a trap.
-const breakpointPCAdjust = 4
+// breakpointPCAdjust is how far PC is past the breakpoint instruction after
+// a trap: 0, because the arm64 kernel delivers the BRK SIGTRAP with the PC
+// still on the BRK.  do_el0_brk64 calls send_user_sigtrap, which uses
+// instruction_pointer(regs) unmodified (arch/arm64/kernel/debug-monitors.c);
+// only the kernel-internal skip paths advance the PC.  GDB history agrees:
+// decr_pc_after_break on aarch64 Linux is 0 (the +4 variant was a QEMU bug,
+// sourceware PR 17280).
+const breakpointPCAdjust = 0
@@ -126,5 +126,9 @@ func (r *Regs) RegValue(name string) (uint64, bool) {
 // breakpointInsn is the software breakpoint instruction (BRK $0).
 var breakpointInsn = []byte{0x05, 0x00, 0x2a, 0x00} // break 0

-// breakpointPCAdjust is how far PC is past the breakpoint instruction after a trap.
-const breakpointPCAdjust = 4
+// breakpointPCAdjust is how far PC is past the breakpoint instruction after
+// a trap: 0, because the kernel delivers the break SIGTRAP with csr_era
+// still on the break instruction.  do_bp passes regs->csr_era straight to
+// force_sig_fault(SIGTRAP, TRAP_BRKPT, ...) and never adjusts era on the
+// signal path (arch/loongarch/kernel/traps.c).
+const breakpointPCAdjust = 0
@@ -126,5 +126,9 @@ func (r *Regs) RegValue(name string) (uint64, bool) {
 // breakpointInsn is the software breakpoint instruction (EBREAK).
 var breakpointInsn = []byte{0x73, 0x00, 0x10, 0x00} // ebreak

-// breakpointPCAdjust is how far PC is past the breakpoint instruction after a trap.
-const breakpointPCAdjust = 4
+// breakpointPCAdjust is how far PC is past the breakpoint instruction after
+// a trap: 0, because the kernel delivers the EBREAK SIGTRAP with sepc still
+// on the ebreak.  handle_break passes regs->epc straight to
+// force_sig_fault(SIGTRAP, TRAP_BRKPT, ...) and only the kernel-internal
+// WARN/CFI paths advance epc (arch/riscv/kernel/traps.c).
+const breakpointPCAdjust = 0
@@ -7,9 +7,10 @@ package debug

 import (
 	"bufio"
+	"cmp"
 	"fmt"
 	"io"
-	"sort"
+	"slices"
 	"strconv"
 	"strings"
 )
@@ -32,7 +33,7 @@ func REPL(s *Session, bm *Breakpoints, codeBase uint64, funcOffset, funcSize, ar
 	entryAddr := codeBase + uint64(funcOffset)

 	fmt.Printf("stopped at function entry: %#x (%d bytes)\n", entryAddr, funcSize)
-	fmt.Println("commands: break <label|addr> | step [n] | continue | disas [n] | regs | where | x <addr> [len] | w <addr> <val...> | labels | quit")
+	fmt.Println("commands: break <label|addr|line> | step [n] | continue | disas [n] | regs | where | x <addr> [len] | w <addr> <val...> | labels | quit")

 	scanner := bufio.NewScanner(in)

@@ -93,7 +94,7 @@ func REPL(s *Session, bm *Breakpoints, codeBase uint64, funcOffset, funcSize, ar
 			regs, _ := s.GetRegs()
 			pc := regs.GetPC()
 			text, instLen, _ := s.Disassemble(pc)
-			if strings.HasPrefix(strings.ToLower(text), "call") || strings.HasPrefix(strings.ToLower(text), "bl") {
+			if isCallInsn(text) {
 				afterAddr := pc + uint64(instLen)
 				_, err := bm.Set(afterAddr, "(next)")
 				if err != nil {
@@ -108,6 +109,21 @@ func REPL(s *Session, bm *Breakpoints, codeBase uint64, funcOffset, funcSize, ar
 					bm.Clear(afterAddr)
 					continue
 				}
+				if s.Exited() {
+					bm.Clear(afterAddr)
+					fmt.Println("debuggee exited")
+					continue
+				}
+				if sig := s.LastSignal(); sig != 0 {
+					bm.Clear(afterAddr)
+					regs, _ := s.GetRegs()
+					fmt.Printf("stopped on signal %v at %#x\n", sig, regs.GetPC())
+					continue
+				}
+				// Fetch the registers after the stop: the trap must be
+				// evaluated against the real PC, not the pre-Continue
+				// snapshot, and a stale SetRegs would clobber live state.
+				regs, _ = s.GetRegs()
 				bm.HandleTrap(&regs)
 				bm.Clear(afterAddr)
 			} else {
@@ -143,9 +159,21 @@ func REPL(s *Session, bm *Breakpoints, codeBase uint64, funcOffset, funcSize, ar
 				bm.Clear(retAddr)
 				continue
 			}
-			if !s.Exited() {
-				bm.HandleTrap(&regs)
+			if s.Exited() {
+				bm.Clear(retAddr)
+				fmt.Println("debuggee exited")
+				continue
 			}
+			if sig := s.LastSignal(); sig != 0 {
+				bm.Clear(retAddr)
+				regs, _ := s.GetRegs()
+				fmt.Printf("stopped on signal %v at %#x\n", sig, regs.GetPC())
+				continue
+			}
+			// Fetch the registers after the stop, as the continue case
+			// does: HandleTrap must see the PC the trap left behind.
+			regs, _ = s.GetRegs()
+			bm.HandleTrap(&regs)
 			bm.Clear(retAddr)
 			if s.Exited() {
 				fmt.Println("debuggee exited")
@@ -171,6 +199,15 @@ func REPL(s *Session, bm *Breakpoints, codeBase uint64, funcOffset, funcSize, ar
 					fmt.Println("debuggee exited")
 					break
 				}
+				if sig := s.LastSignal(); sig != 0 {
+					// A genuine signal-delivery-stop (a fault): report it
+					// and return to the prompt.  Continuing would restart
+					// the faulting instruction and fault forever.
+					regs, _ := s.GetRegs()
+					fmt.Printf("stopped on signal %v at %#x (func+%#x)\n",
+						sig, regs.GetPC(), regs.GetPC()-codeBase-uint64(funcOffset))
+					break
+				}
 				reason, wpAddr := s.StopInfo()
 				if reason == StopWatchpoint {
 					fmt.Printf("watchpoint hit at %#x\n", wpAddr)
@@ -196,7 +233,7 @@ func REPL(s *Session, bm *Breakpoints, codeBase uint64, funcOffset, funcSize, ar

 		case "break", "b":
 			if len(parts) < 2 {
-				fmt.Println("usage: break <label|addr|line> [if <reg> <op> <val>]")
+				fmt.Println("usage: break <label|addr|line> [if <reg> <op> <val|reg|*addr>]")
 				continue
 			}
 			var addr uint64
@@ -221,13 +258,27 @@ func REPL(s *Session, bm *Breakpoints, codeBase uint64, funcOffset, funcSize, ar
 				reg := strings.ToLower(parts[3])
 				op := parts[4]
 				operand := parts[5]
-				if val, err := strconv.ParseUint(operand, 0, 64); err == nil {
+				switch {
+				case strings.HasPrefix(operand, "*"):
+					// Memory operand: compare against the 8-byte word at
+					// the address, resolved in the debuggee when the
+					// breakpoint is evaluated.
+					addr, err := strconv.ParseUint(strings.TrimPrefix(operand, "*"), 0, 64)
+					if err != nil {
+						fmt.Printf("invalid memory operand: %s\n", operand)
+						continue
+					}
+					cond = &Condition{Reg: reg, Op: op, MemAddr: addr}
+				default:
+					val, err := strconv.ParseUint(operand, 0, 64)
+					if err == nil {
 						cond = &Condition{Reg: reg, Op: op, Value: val}
 					} else {
 						cond = &Condition{Reg: reg, Op: op, Reg2: strings.ToLower(operand)}
 					}
+				}
 			} else if len(parts) >= 4 && parts[2] == "if" {
-				fmt.Println("usage: break <label|addr> if <reg> <op> <value|reg>")
+				fmt.Println("usage: break <label|addr|line> if <reg> <op> <value|reg|*addr>")
 				continue
 			}
 			bp, err := bm.SetWithCond(addr, label, cond)
@@ -237,7 +288,7 @@ func REPL(s *Session, bm *Breakpoints, codeBase uint64, funcOffset, funcSize, ar
 			}
 			condStr := ""
 			if cond != nil {
-				condStr = fmt.Sprintf(" if %s %s %#x", cond.Reg, cond.Op, cond.Value)
+				condStr = " if " + cond.String()
 			}
 			fmt.Printf("breakpoint set: %s at %#x (func+%#x)%s\n", bp.Label, bp.Addr, bp.Addr-codeBase-uint64(funcOffset), condStr)

@@ -277,7 +328,11 @@ func REPL(s *Session, bm *Breakpoints, codeBase uint64, funcOffset, funcSize, ar
 				addr, _ = resolveAddr(parts[1], codeBase, uint64(funcOffset), labels)
 			}
 			if len(parts) > 2 {
-				length, _ = strconv.Atoi(parts[2])
+				// A malformed or non-positive length would panic
+				// ReadMemory's make; fall back to the default instead.
+				if n, err := strconv.Atoi(parts[2]); err == nil && n > 0 {
+					length = n
+				}
 			}
 			mem, err := s.ReadMemory(addr, length)
 			if err != nil {
@@ -336,10 +391,8 @@ func REPL(s *Session, bm *Breakpoints, codeBase uint64, funcOffset, funcSize, ar
 			}

 		case "labels", "l":
-			sorted := make([]Label, len(labels))
-			copy(sorted, labels)
-			sort.Slice(sorted, func(i, j int) bool { return sorted[i].Offset < sorted[j].Offset })
-			for _, l := range sorted {
+			slices.SortFunc(labels, func(a, b Label) int { return cmp.Compare(a.Offset, b.Offset) })
+			for _, l := range labels {
 				fmt.Printf("  func+%#04x  %s\n", l.Offset, l.Name)
 			}

@@ -369,7 +422,10 @@ func REPL(s *Session, bm *Breakpoints, codeBase uint64, funcOffset, funcSize, ar
 			fmt.Println()

 		case "help", "h", "?":
-			fmt.Printf(`  break <label|addr> [if <reg> <op> <val>]  set a breakpoint
+			fmt.Printf(`  break <label|addr|line> [if <reg> <op> <val|reg|*addr>]
+                       set a breakpoint, optionally conditional on a
+                       register compared to a constant, a register, or the
+                       8-byte word at *addr
  delete <label|addr>  remove a breakpoint
  info break           list all breakpoints
  watch <addr> [r|w] [size]  set a hardware watchpoint (write by default)
@@ -463,8 +519,8 @@ func REPL(s *Session, bm *Breakpoints, codeBase uint64, funcOffset, funcSize, ar
 		case "unwatch":
 			if len(parts) >= 2 {
 				slot, err := strconv.Atoi(parts[1])
-				if err != nil || slot < 0 || slot > 3 {
-					fmt.Println("usage: unwatch [<slot>]")
+				if err != nil || slot < 0 || slot >= maxWatchpoints() {
+					fmt.Printf("usage: unwatch [<slot 0-%d>]\n", maxWatchpoints()-1)
 					continue
 				}
 				if err := s.ClearWatchpoint(slot); err != nil {
@@ -6,6 +6,7 @@
 package debug

 import (
+	"encoding/binary"
 	"syscall"
 	"unsafe"
 )
@@ -61,8 +62,15 @@ func (s *Session) StopInfo() (StopReason, uint64) {
 	case trapBRKPT:
 		return StopBreakpoint, 0
 	case trapHWBRKPT:
-		addr := *(*uint64)(unsafe.Add(unsafe.Pointer(&info), 16))
-		return StopWatchpoint, addr
+		// si_addr sits at struct offset 16 (12 bytes of signo/errno/code
+		// plus 4 bytes of union alignment).  The siginfo buffer is only
+		// 4-byte aligned, so the address is read byte-wise to keep the
+		// load aligned on riscv64 and loong64.  What si_addr names is
+		// architecture-specific (the data address on arm64, the
+		// instruction pointer on x86), so the per-architecture
+		// archWatchpointAddr resolves it to the watched address.
+		addr := binary.LittleEndian.Uint64(info._pad[4:12])
+		return StopWatchpoint, archWatchpointAddr(s, addr)
 	default:
 		return StopSingleStep, 0
 	}
@@ -28,15 +28,7 @@ func RunTarget(asmPath, funcName, argsFile, tmpDir string) error {
 		return fmt.Errorf("debug target: parse: %v", errs[0])
 	}

-	var img *asm.Image
-	switch "arm64" {
-	case "arm64":
-		img, err = asm.AssembleFileARM64(file)
-	case "riscv64":
-		img, err = asm.AssembleFileRISCV(file)
-	case "loong64":
-		img, err = asm.AssembleFileLOONG64(file)
-	}
+	img, err := asm.AssembleFileARM64(file)
 	if err != nil {
 		return fmt.Errorf("debug target: assemble: %w", err)
 	}
@@ -28,15 +28,7 @@ func RunTarget(asmPath, funcName, argsFile, tmpDir string) error {
 		return fmt.Errorf("debug target: parse: %v", errs[0])
 	}

-	var img *asm.Image
-	switch "loong64" {
-	case "arm64":
-		img, err = asm.AssembleFileARM64(file)
-	case "riscv64":
-		img, err = asm.AssembleFileRISCV(file)
-	case "loong64":
-		img, err = asm.AssembleFileLOONG64(file)
-	}
+	img, err := asm.AssembleFileLOONG64(file)
 	if err != nil {
 		return fmt.Errorf("debug target: assemble: %w", err)
 	}
@@ -12,10 +12,11 @@ type tracer interface {
 	Peek(addr uint64) (uint64, error)
 	Poke(addr uint64, val uint64) error
 	SetRegs(regs *Regs) error
+	Step() error
 	Pid() int
 }

-// mockTracer records Peek/Poke calls and provides fake register state.
+// mockTracer records Peek/Poke/Step calls and provides fake register state.
 type mockTracer struct {
 	mem   map[uint64]byte
 	peeks []uint64
@@ -23,6 +24,7 @@ type mockTracer struct {
 		addr uint64
 		val  uint64
 	}
+	steps int
 	regs  *Regs
 }

@@ -57,4 +59,8 @@ func (m *mockTracer) SetRegs(regs *Regs) error {
 	m.regs = regs
 	return nil
 }
+func (m *mockTracer) Step() error {
+	m.steps++
+	return nil
+}
 func (m *mockTracer) Pid() int { return 42 }
@@ -8,10 +8,48 @@ package debug
 import (
 	"fmt"
 	"syscall"
+	"unsafe"
 )

 // Hardware watchpoint support via x86-64 debug registers (DR0-DR3, DR7).

+// The kernel translates PTRACE_POKEUSER/PEEKUSER offsets inside
+// [offsetof(struct user, u_debugreg[0]), u_debugreg[7]] to DR0-DR7
+// (arch/x86/kernel/ptrace.c, arch_ptrace).  sys/user.h places u_debugreg at
+// 0x350: DR0-DR3 are 0x350/0x358/0x360/0x368, DR6 (status) is 0x380 and
+// DR7 (control) is 0x388.  Offsets below 0x350 write user_regs_struct
+// fields (r15 at 0x0, r10 at 0x38), not debug registers.
+const (
+	drOffset = 0x350 // offsetof(struct user, u_debugreg[0]), DR0
+	dr6Off   = 0x380 // offsetof(struct user, u_debugreg[6]), DR6
+	dr7Off   = 0x388 // offsetof(struct user, u_debugreg[7]), DR7
+)
+
+// archWatchpointAddr resolves the address of the watchpoint that fired.
+// x86 delivers si_addr = the instruction pointer of the trapping access
+// (arch/x86/kernel/ptrace.c send_sigtrap passes regs->ip), so the watched
+// data address is recovered from DR6's slot bits (B0-B3, positive polarity
+// through PEEKUSER) and the matching DR0-DR3.
+func archWatchpointAddr(s *Session, siAddr uint64) uint64 {
+	dr6, err := ptracePeekUser(s.pid, dr6Off)
+	if err != nil {
+		return siAddr
+	}
+	for slot := range 4 {
+		if dr6&(1<<slot) != 0 {
+			addr, err := ptracePeekUser(s.pid, drOffset+uintptr(slot*8))
+			if err == nil && addr != 0 {
+				return addr
+			}
+		}
+	}
+	return siAddr
+}
+
+// maxWatchpoints reports the number of hardware watchpoint slots the
+// architecture provides: four address registers, DR0-DR3.
+func maxWatchpoints() int { return 4 }
+
 // WatchpointType selects what triggers the watchpoint.
 type WatchpointType int

@@ -62,23 +100,11 @@ func (s *Session) SetWatchpoint(slot int, addr uint64, typ WatchpointType, size
 		return fmt.Errorf("debug: watchpoint size must be 1, 2, 4, or 8")
 	}

-	var drAddr uintptr
-	switch slot {
-	case 0:
-		drAddr = 0x0
-	case 1:
-		drAddr = 0x8
-	case 2:
-		drAddr = 0x10
-	case 3:
-		drAddr = 0x18
-	}
-
-	if err := ptracePokeUser(s.pid, drAddr, addr); err != nil {
+	if err := ptracePokeUser(s.pid, drOffset+uintptr(slot*8), addr); err != nil {
 		return fmt.Errorf("debug: set DR%d: %w", slot, err)
 	}

-	dr7, err := ptracePeekUser(s.pid, 0x38)
+	dr7, err := ptracePeekUser(s.pid, dr7Off)
 	if err != nil {
 		return fmt.Errorf("debug: read DR7: %w", err)
 	}
@@ -90,7 +116,7 @@ func (s *Session) SetWatchpoint(slot int, addr uint64, typ WatchpointType, size
 	mask := ^((uint64(1) << (2 * slot)) | (uint64(3) << (16 + 4*slot)) | (uint64(3) << (18 + 4*slot)))
 	dr7 = (dr7 & mask) | enableBit | rwBits | lenField

-	if err := ptracePokeUser(s.pid, 0x38, dr7); err != nil {
+	if err := ptracePokeUser(s.pid, dr7Off, dr7); err != nil {
 		return fmt.Errorf("debug: set DR7: %w", err)
 	}
 	s.wpSlots[slot] = true
@@ -105,12 +131,12 @@ func (s *Session) ClearWatchpoint(slot int) error {
 	if !s.wpSlots[slot] {
 		return fmt.Errorf("debug: watchpoint slot %d is not in use", slot)
 	}
-	dr7, err := ptracePeekUser(s.pid, 0x38)
+	dr7, err := ptracePeekUser(s.pid, dr7Off)
 	if err != nil {
 		return err
 	}
 	dr7 &^= uint64(1) << (2 * slot)
-	if err := ptracePokeUser(s.pid, 0x38, dr7); err != nil {
+	if err := ptracePokeUser(s.pid, dr7Off, dr7); err != nil {
 		return err
 	}
 	s.wpSlots[slot] = false
@@ -119,7 +145,7 @@ func (s *Session) ClearWatchpoint(slot int) error {

 // ClearAllWatchpoints removes all hardware watchpoints.
 func (s *Session) ClearAllWatchpoints() error {
-	for slot := range 4 {
+	for slot := range maxWatchpoints() {
 		if s.wpSlots[slot] {
 			if err := s.ClearWatchpoint(slot); err != nil {
 				return err
@@ -146,16 +172,21 @@ func ptracePokeUser(pid int, offset uintptr, val uint64) error {
 }

 func ptracePeekUser(pid int, offset uintptr) (uint64, error) {
+	// x86 PEEKUSR writes the word to the user-space pointer in data
+	// (arch/x86/kernel/ptrace.c uses put_user); passing 0 there fails with
+	// EFAULT, so the word is read through a real address.
 	const ptracePeekuser = 3
-	val, _, errno := syscall.Syscall6(
+	var word uint64
+	_, _, errno := syscall.Syscall6(
 		syscall.SYS_PTRACE,
 		uintptr(ptracePeekuser),
 		uintptr(pid),
 		offset,
-		0, 0, 0,
+		uintptr(unsafe.Pointer(&word)),
+		0, 0,
 	)
 	if errno != 0 {
 		return 0, errno
 	}
-	return uint64(val), nil
+	return word, nil
 }
@@ -12,7 +12,7 @@ import (
 )

 // Hardware watchpoint support via arm64 debug registers (DBGWVR/DBGWCR).
-// Accessed via PTRACE_SETREGSET with NT_ARM_HW_BREAK.
+// Accessed via PTRACE_GETREGSET/SETREGSET with NT_ARM_HW_WATCH.

 // WatchpointType selects what triggers the watchpoint.
 type WatchpointType int
@@ -22,26 +22,35 @@ const (
 	WatchRead  WatchpointType = 3
 )

-const maxWatchpoints = 16
+// maxWatchpoints reports the number of hardware watchpoint slots the
+// architecture provides: DBGWVR0-DBGWCR15.
+func maxWatchpoints() int { return 16 }

-// hwBreakState mirrors the kernel's struct user_hwdebug_state.
-type hwBreakState struct {
+// hwWatchState mirrors the kernel's struct user_hwdebug_state.
+type hwWatchState struct {
 	DbgInfo uint32
 	_pad    [4]byte
-	DbgRegs [16]hwBreakReg
+	DbgRegs [16]hwWatchReg
 }

-type hwBreakReg struct {
+type hwWatchReg struct {
 	Addr uint64
 	Ctrl uint64
 }

-const (
-	ntArmHWBreak = 0x403 // NT_ARM_HW_BREAK
-)
+// ntArmHWWatch is NT_ARM_HW_WATCH (0x403), the watchpoint regset
+// (include/uapi/linux/elf.h; 0x402 is NT_ARM_HW_BREAK).  Watchpoints and
+// breakpoints live in different regsets with the same struct shape, so the
+// constant is named for what it arms to keep a future edit from arming
+// breakpoints instead.
+const ntArmHWWatch = 0x403
+
+// archWatchpointAddr resolves the address of the watchpoint that fired:
+// the arm64 kernel already reports the watched data address as si_addr.
+func archWatchpointAddr(s *Session, siAddr uint64) uint64 { return siAddr }

 func (s *Session) FindFreeWatchpointSlot() int {
-	for i := range maxWatchpoints {
+	for i := range maxWatchpoints() {
 		if !s.wpSlots[i] {
 			return i
 		}
@@ -50,7 +59,7 @@ func (s *Session) FindFreeWatchpointSlot() int {
 }

 func (s *Session) IsWatchpointSlotUsed(slot int) bool {
-	if slot < 0 || slot >= maxWatchpoints {
+	if slot < 0 || slot >= maxWatchpoints() {
 		return false
 	}
 	return s.wpSlots[slot]
@@ -58,27 +67,31 @@ func (s *Session) IsWatchpointSlotUsed(slot int) bool {

 // SetWatchpoint installs a hardware watchpoint on the given address.
 func (s *Session) SetWatchpoint(slot int, addr uint64, typ WatchpointType, size int) error {
-	if slot < 0 || slot >= maxWatchpoints {
-		return fmt.Errorf("debug: watchpoint slot must be 0-%d", maxWatchpoints-1)
+	if slot < 0 || slot >= maxWatchpoints() {
+		return fmt.Errorf("debug: watchpoint slot must be 0-%d", maxWatchpoints()-1)
 	}
 	if s.wpSlots[slot] {
 		return fmt.Errorf("debug: watchpoint slot %d already in use", slot)
 	}

-	state, err := s.getHWBreakState()
+	state, err := s.getHWWatchState()
 	if err != nil {
 		return fmt.Errorf("debug: read watchpoint state: %w", err)
 	}

-	if uint32(slot) >= state.DbgInfo {
-		return fmt.Errorf("debug: slot %d exceeds available watchpoints (%d)", slot, state.DbgInfo)
+	// MDSCR_EL1 packs (debug_arch << 8) | num_slots into dbg_info, so only
+	// the low byte counts slots.
+	if uint32(slot) >= state.DbgInfo&0xff {
+		return fmt.Errorf("debug: slot %d exceeds available watchpoints (%d)", slot, state.DbgInfo&0xff)
 	}

 	state.DbgRegs[slot].Addr = addr
+	// DBGWCR bits 3-4 select the access type: 01 load, 10 store, 11 either
+	// (ARM DDI 0487, DBGWCR<n>_EL1 watchpoint type field).
 	ctrl := uint64(1) // enable
 	switch typ {
 	case WatchWrite:
-		ctrl |= 1 << 3 // store only
+		ctrl |= 2 << 3 // store only
 	case WatchRead:
 		ctrl |= 3 << 3 // load+store
 	}
@@ -98,7 +111,7 @@ func (s *Session) SetWatchpoint(slot int, addr uint64, typ WatchpointType, size
 	ctrl |= bas << 5
 	state.DbgRegs[slot].Ctrl = ctrl

-	if err := s.setHWBreakState(state); err != nil {
+	if err := s.setHWWatchState(state); err != nil {
 		return fmt.Errorf("debug: set watchpoint: %w", err)
 	}

@@ -107,20 +120,20 @@ func (s *Session) SetWatchpoint(slot int, addr uint64, typ WatchpointType, size
 }

 func (s *Session) ClearWatchpoint(slot int) error {
-	if slot < 0 || slot >= maxWatchpoints {
-		return fmt.Errorf("debug: watchpoint slot must be 0-%d", maxWatchpoints-1)
+	if slot < 0 || slot >= maxWatchpoints() {
+		return fmt.Errorf("debug: watchpoint slot must be 0-%d", maxWatchpoints()-1)
 	}
 	if !s.wpSlots[slot] {
 		return fmt.Errorf("debug: watchpoint slot %d is not in use", slot)
 	}

-	state, err := s.getHWBreakState()
+	state, err := s.getHWWatchState()
 	if err != nil {
 		return err
 	}
 	state.DbgRegs[slot].Addr = 0
 	state.DbgRegs[slot].Ctrl = 0
-	if err := s.setHWBreakState(state); err != nil {
+	if err := s.setHWWatchState(state); err != nil {
 		return err
 	}
 	s.wpSlots[slot] = false
@@ -128,7 +141,7 @@ func (s *Session) ClearWatchpoint(slot int) error {
 }

 func (s *Session) ClearAllWatchpoints() error {
-	for slot := 0; slot < maxWatchpoints; slot++ {
+	for slot := range maxWatchpoints() {
 		if s.wpSlots[slot] {
 			if err := s.ClearWatchpoint(slot); err != nil {
 				return err
@@ -138,8 +151,8 @@ func (s *Session) ClearAllWatchpoints() error {
 	return nil
 }

-func (s *Session) getHWBreakState() (*hwBreakState, error) {
-	var state hwBreakState
+func (s *Session) getHWWatchState() (*hwWatchState, error) {
+	var state hwWatchState
 	iovec := syscall.Iovec{
 		Base: (*byte)(unsafe.Pointer(&state)),
 		Len:  uint64(unsafe.Sizeof(state)),
@@ -148,7 +161,7 @@ func (s *Session) getHWBreakState() (*hwBreakState, error) {
 		syscall.SYS_PTRACE,
 		uintptr(syscall.PTRACE_GETREGSET),
 		uintptr(s.pid),
-		uintptr(ntArmHWBreak),
+		uintptr(ntArmHWWatch),
 		uintptr(unsafe.Pointer(&iovec)),
 		0, 0,
 	)
@@ -158,7 +171,7 @@ func (s *Session) getHWBreakState() (*hwBreakState, error) {
 	return &state, nil
 }

-func (s *Session) setHWBreakState(state *hwBreakState) error {
+func (s *Session) setHWWatchState(state *hwWatchState) error {
 	iovec := syscall.Iovec{
 		Base: (*byte)(unsafe.Pointer(state)),
 		Len:  uint64(unsafe.Sizeof(*state)),
@@ -167,7 +180,7 @@ func (s *Session) setHWBreakState(state *hwBreakState) error {
 		syscall.SYS_PTRACE,
 		uintptr(syscall.PTRACE_SETREGSET),
 		uintptr(s.pid),
-		uintptr(ntArmHWBreak),
+		uintptr(ntArmHWWatch),
 		uintptr(unsafe.Pointer(&iovec)),
 		0, 0,
 	)
@@ -8,10 +8,47 @@ package debug
 import (
 	"fmt"
 	"syscall"
+	"unsafe"
 )

-// Hardware watchpoint support for LoongArch via debug registers.
-// Uses PTRACE_POKEUSER/PEEKUSER to access HW watchpoint registers.
+// Hardware watchpoint support via the NT_LOONGARCH_HW_WATCH regset.
+//
+// The kernel's PTRACE_POKEUSER on loong64 accepts only the user_pt_regs
+// indices 0-34 (GPRs, orig_a0, era, badv, per
+// arch/loongarch/include/uapi/asm/ptrace.h), so there is no debug-register
+// window to poke.  The real interface is PTRACE_GETREGSET/SETREGSET on
+// NT_LOONGARCH_HW_WATCH (0xa06, include/uapi/linux/elf.h) with struct
+// user_watch_state_v2 (arch/loongarch/include/uapi/asm/ptrace.h): a dbg_info
+// word followed by 14 slots of {addr u64, mask u64, ctrl u32, pad u32}.
+// hw_break_get puts the slot count in the low byte of dbg_info
+// (arch/loongarch/kernel/ptrace.c, ptrace_hbp_get_resource_info) and
+// hw_break_set ignores dbg_info, reading addr, mask and ctrl per slot.
+
+const ntLoongHWWatch = 0xa06
+
+// loongWatchState mirrors the kernel's struct user_watch_state_v2.
+type loongWatchState struct {
+	DbgInfo uint64
+	DbgRegs [14]loongWatchReg
+}
+
+type loongWatchReg struct {
+	Addr uint64
+	Mask uint64
+	Ctrl uint32
+	Pad  uint32
+}
+
+// Control word bit layout (arch/loongarch/include/asm/hw_breakpoint.h):
+// bits 1-4 privilege enables (CTRL_PLV3_ENABLE, 0x10, covers user mode),
+// bits 8-9 access type (LOAD 1<<0, STORE 1<<1), bits 10-11 length
+// (0=8 bytes, 1=4, 2=2, 3=1, inverted like the hardware FWP cfg).
+const (
+	loongCtrlPLV3Enable = 0x10
+	loongTypeLoad       = 1 << 8
+	loongTypeStore      = 2 << 8
+	loongLenShift       = 10
+)

 // WatchpointType selects what triggers the watchpoint.
 type WatchpointType int
@@ -21,10 +58,16 @@ const (
 	WatchRead  WatchpointType = 3
 )

-const maxWatchpoints = 4
+// maxWatchpoints reports the slot capacity of the regset struct; the number
+// the hardware actually provides is read from dbg_info at arm time.
+func maxWatchpoints() int { return len(loongWatchState{}.DbgRegs) }
+
+// archWatchpointAddr resolves the address of the watchpoint that fired:
+// the loongarch kernel already reports the accessed address as si_addr.
+func archWatchpointAddr(s *Session, siAddr uint64) uint64 { return siAddr }

 func (s *Session) FindFreeWatchpointSlot() int {
-	for i := range maxWatchpoints {
+	for i := range maxWatchpoints() {
 		if !s.wpSlots[i] {
 			return i
 		}
@@ -33,53 +76,56 @@ func (s *Session) FindFreeWatchpointSlot() int {
 }

 func (s *Session) IsWatchpointSlotUsed(slot int) bool {
-	if slot < 0 || slot >= maxWatchpoints {
+	if slot < 0 || slot >= maxWatchpoints() {
 		return false
 	}
 	return s.wpSlots[slot]
 }

-// SetWatchpoint installs a hardware watchpoint.
+// SetWatchpoint installs a hardware watchpoint on the given address.
 func (s *Session) SetWatchpoint(slot int, addr uint64, typ WatchpointType, size int) error {
-	if slot < 0 || slot >= maxWatchpoints {
-		return fmt.Errorf("debug: watchpoint slot must be 0-%d", maxWatchpoints-1)
+	if slot < 0 || slot >= maxWatchpoints() {
+		return fmt.Errorf("debug: watchpoint slot must be 0-%d", maxWatchpoints()-1)
 	}
 	if s.wpSlots[slot] {
 		return fmt.Errorf("debug: watchpoint slot %d already in use", slot)
 	}
-	if size != 1 && size != 2 && size != 4 && size != 8 {
+
+	var ctrlType uint32
+	switch typ {
+	case WatchWrite:
+		ctrlType = loongTypeStore
+	case WatchRead:
+		ctrlType = loongTypeLoad | loongTypeStore
+	}
+	var lenBits uint32
+	switch size {
+	case 1:
+		lenBits = 3
+	case 2:
+		lenBits = 2
+	case 4:
+		lenBits = 1
+	case 8:
+		lenBits = 0
+	default:
 		return fmt.Errorf("debug: watchpoint size must be 1, 2, 4, or 8")
 	}

-	// LoongArch debug registers: DBGWVR (watchpoint value) and DBGWCR (watchpoint control).
-	// Accessed via PTRACE_POKEUSER at architecture-specific offsets.
-	if err := ptracePokeUser(s.pid, uintptr(0x1000+slot*8), addr); err != nil {
-		return fmt.Errorf("debug: set watchpoint address: %w", err)
+	state, err := s.getLoongWatchState()
+	if err != nil {
+		return fmt.Errorf("debug: read watchpoint state: %w", err)
+	}
+	if uint64(slot) >= state.DbgInfo&0xff {
+		return fmt.Errorf("debug: slot %d exceeds available watchpoints (%d)", slot, state.DbgInfo&0xff)
 	}

-	// DBGWCR: enable + type + size.
-	var wcr uint64 = 1 // enable
-	switch typ {
-	case WatchWrite:
-		wcr |= 1 << 3 // store
-	case WatchRead:
-		wcr |= 3 << 3 // load+store
-	}
-	var sizeBits uint64
-	switch size {
-	case 1:
-		sizeBits = 0
-	case 2:
-		sizeBits = 1
-	case 4:
-		sizeBits = 2
-	case 8:
-		sizeBits = 3
-	}
-	wcr |= sizeBits << 5
+	state.DbgRegs[slot].Addr = addr
+	state.DbgRegs[slot].Mask = 0
+	state.DbgRegs[slot].Ctrl = loongCtrlPLV3Enable | ctrlType | lenBits<<loongLenShift

-	if err := ptracePokeUser(s.pid, uintptr(0x1001+slot*8), wcr); err != nil {
-		return fmt.Errorf("debug: set watchpoint control: %w", err)
+	if err := s.setLoongWatchState(state); err != nil {
+		return fmt.Errorf("debug: set watchpoint: %w", err)
 	}

 	s.wpSlots[slot] = true
@@ -87,14 +133,21 @@ func (s *Session) SetWatchpoint(slot int, addr uint64, typ WatchpointType, size
 }

 func (s *Session) ClearWatchpoint(slot int) error {
-	if slot < 0 || slot >= maxWatchpoints {
-		return fmt.Errorf("debug: watchpoint slot must be 0-%d", maxWatchpoints-1)
+	if slot < 0 || slot >= maxWatchpoints() {
+		return fmt.Errorf("debug: watchpoint slot must be 0-%d", maxWatchpoints()-1)
 	}
 	if !s.wpSlots[slot] {
 		return fmt.Errorf("debug: watchpoint slot %d is not in use", slot)
 	}

-	if err := ptracePokeUser(s.pid, uintptr(0x1001+slot*8), 0); err != nil {
+	state, err := s.getLoongWatchState()
+	if err != nil {
+		return err
+	}
+	state.DbgRegs[slot].Addr = 0
+	state.DbgRegs[slot].Mask = 0
+	state.DbgRegs[slot].Ctrl = 0
+	if err := s.setLoongWatchState(state); err != nil {
 		return err
 	}
 	s.wpSlots[slot] = false
@@ -102,7 +155,7 @@ func (s *Session) ClearWatchpoint(slot int) error {
 }

 func (s *Session) ClearAllWatchpoints() error {
-	for slot := 0; slot < maxWatchpoints; slot++ {
+	for slot := range maxWatchpoints() {
 		if s.wpSlots[slot] {
 			if err := s.ClearWatchpoint(slot); err != nil {
 				return err
@@ -112,14 +165,37 @@ func (s *Session) ClearAllWatchpoints() error {
 	return nil
 }

-func ptracePokeUser(pid int, offset uintptr, val uint64) error {
-	const ptracePokeuser = 6
+func (s *Session) getLoongWatchState() (*loongWatchState, error) {
+	var state loongWatchState
+	iovec := syscall.Iovec{
+		Base: (*byte)(unsafe.Pointer(&state)),
+		Len:  uint64(unsafe.Sizeof(state)),
+	}
 	_, _, errno := syscall.Syscall6(
 		syscall.SYS_PTRACE,
-		uintptr(ptracePokeuser),
-		uintptr(pid),
-		offset,
-		uintptr(val),
+		uintptr(syscall.PTRACE_GETREGSET),
+		uintptr(s.pid),
+		uintptr(ntLoongHWWatch),
+		uintptr(unsafe.Pointer(&iovec)),
+		0, 0,
+	)
+	if errno != 0 {
+		return nil, errno
+	}
+	return &state, nil
+}
+
+func (s *Session) setLoongWatchState(state *loongWatchState) error {
+	iovec := syscall.Iovec{
+		Base: (*byte)(unsafe.Pointer(state)),
+		Len:  uint64(unsafe.Sizeof(*state)),
+	}
+	_, _, errno := syscall.Syscall6(
+		syscall.SYS_PTRACE,
+		uintptr(syscall.PTRACE_SETREGSET),
+		uintptr(s.pid),
+		uintptr(ntLoongHWWatch),
+		uintptr(unsafe.Pointer(&iovec)),
 		0, 0,
 	)
 	if errno != 0 {
@@ -127,18 +203,3 @@ func ptracePokeUser(pid int, offset uintptr, val uint64) error {
 	}
 	return nil
 }
-
-func ptracePeekUser(pid int, offset uintptr) (uint64, error) {
-	const ptracePeekuser = 3
-	val, _, errno := syscall.Syscall6(
-		syscall.SYS_PTRACE,
-		uintptr(ptracePeekuser),
-		uintptr(pid),
-		offset,
-		0, 0, 0,
-	)
-	if errno != 0 {
-		return 0, errno
-	}
-	return uint64(val), nil
-}
@@ -7,11 +7,16 @@ package debug

 import (
 	"fmt"
-	"syscall"
 )

-// Hardware watchpoint support for RISC-V via Sdtrig trigger registers.
-// Uses PTRACE_POKEUSER/PEEKUSER to access debug registers.
+// Hardware watchpoints are not reachable through the riscv64 kernel ptrace
+// interface.  arch/riscv/kernel/ptrace.c forwards every POKEUSER/PEEKUSER to
+// the generic ptrace_request, and the riscv user_regset view contains only
+// the GPR, FP and vector regsets: there is no debug-register or trigger
+// regset, and offsets outside the view fail with EIO.  The Sdtrig CSRs
+// (tselect/tdata1/tdata2) are not exposed to ptrace either.  Until the
+// kernel grows a trigger regset, SetWatchpoint reports the fact instead of
+// poking a window that does not exist.

 // WatchpointType selects what triggers the watchpoint.
 type WatchpointType int
@@ -21,10 +26,18 @@ const (
 	WatchRead  WatchpointType = 3
 )

-const maxWatchpoints = 4
+// maxWatchpoints reports the number of hardware watchpoint slots the
+// architecture provides.  riscv64 exposes none via ptrace; the bound exists
+// so the slot bookkeeping stays consistent.
+func maxWatchpoints() int { return 4 }
+
+// archWatchpointAddr resolves the address of the watchpoint that fired.
+// Unreachable in practice (watchpoints cannot be armed), but si_addr names
+// the accessed address where the kernel does report one.
+func archWatchpointAddr(s *Session, siAddr uint64) uint64 { return siAddr }

 func (s *Session) FindFreeWatchpointSlot() int {
-	for i := range maxWatchpoints {
+	for i := range maxWatchpoints() {
 		if !s.wpSlots[i] {
 			return i
 		}
@@ -33,115 +46,26 @@ func (s *Session) FindFreeWatchpointSlot() int {
 }

 func (s *Session) IsWatchpointSlotUsed(slot int) bool {
-	if slot < 0 || slot >= maxWatchpoints {
+	if slot < 0 || slot >= maxWatchpoints() {
 		return false
 	}
 	return s.wpSlots[slot]
 }

-// SetWatchpoint installs a hardware watchpoint.
+// SetWatchpoint always fails: the riscv64 kernel ptrace interface has no
+// hardware-watchpoint access.
 func (s *Session) SetWatchpoint(slot int, addr uint64, typ WatchpointType, size int) error {
-	if slot < 0 || slot >= maxWatchpoints {
-		return fmt.Errorf("debug: watchpoint slot must be 0-%d", maxWatchpoints-1)
-	}
-	if s.wpSlots[slot] {
-		return fmt.Errorf("debug: watchpoint slot %d already in use", slot)
-	}
-	if size != 1 && size != 2 && size != 4 && size != 8 {
-		return fmt.Errorf("debug: watchpoint size must be 1, 2, 4, or 8")
-	}
-
-	// RISC-V trigger registers: tdata1 encodes type/control, tdata2 holds address.
-	// The exact encoding depends on the trigger implementation (Sdtrig).
-	// Use PTRACE_POKEUSER to write to the trigger CSRs via the kernel's
-	// debug register interface.
-	if err := ptracePokeUser(s.pid, uintptr(0x1000+slot*8), addr); err != nil {
-		return fmt.Errorf("debug: set watchpoint address: %w", err)
-	}
-
-	// tdata1: set match control.  Mode=2 (data match), select=0, action=1 (debug exception).
-	var tdata1 uint64 = 2 << 60 // type = match (2)
-	tdata1 |= 1 << 0            // action = enter debug mode
-	tdata1 |= 1 << 7            // store (write) trigger
-	if typ == WatchRead {
-		tdata1 |= 1 << 6 // load trigger
-	}
-	// Size encoding: 0=1byte, 1=2byte, 2=4byte, 3=8byte.
-	var sizeBits uint64
-	switch size {
-	case 1:
-		sizeBits = 0
-	case 2:
-		sizeBits = 1
-	case 4:
-		sizeBits = 2
-	case 8:
-		sizeBits = 3
-	}
-	tdata1 |= sizeBits << 16 // size field
-
-	if err := ptracePokeUser(s.pid, uintptr(0x1001+slot*8), tdata1); err != nil {
-		return fmt.Errorf("debug: set watchpoint control: %w", err)
-	}
-
-	s.wpSlots[slot] = true
-	return nil
+	return fmt.Errorf("debug: hardware watchpoints are not supported by the riscv64 kernel ptrace interface")
 }

+// ClearWatchpoint always fails: no watchpoint can ever be armed.
 func (s *Session) ClearWatchpoint(slot int) error {
-	if slot < 0 || slot >= maxWatchpoints {
-		return fmt.Errorf("debug: watchpoint slot must be 0-%d", maxWatchpoints-1)
+	if slot < 0 || slot >= maxWatchpoints() {
+		return fmt.Errorf("debug: watchpoint slot must be 0-%d", maxWatchpoints()-1)
 	}
-	if !s.wpSlots[slot] {
 	return fmt.Errorf("debug: watchpoint slot %d is not in use", slot)
 }

-	// Disable by clearing tdata1.
-	if err := ptracePokeUser(s.pid, uintptr(0x1001+slot*8), 0); err != nil {
-		return err
-	}
-	s.wpSlots[slot] = false
-	return nil
-}
-
 func (s *Session) ClearAllWatchpoints() error {
-	for slot := 0; slot < maxWatchpoints; slot++ {
-		if s.wpSlots[slot] {
-			if err := s.ClearWatchpoint(slot); err != nil {
-				return err
-			}
-		}
-	}
 	return nil
 }
-
-func ptracePokeUser(pid int, offset uintptr, val uint64) error {
-	const ptracePokeuser = 6
-	_, _, errno := syscall.Syscall6(
-		syscall.SYS_PTRACE,
-		uintptr(ptracePokeuser),
-		uintptr(pid),
-		offset,
-		uintptr(val),
-		0, 0,
-	)
-	if errno != 0 {
-		return errno
-	}
-	return nil
-}
-
-func ptracePeekUser(pid int, offset uintptr) (uint64, error) {
-	const ptracePeekuser = 3
-	val, _, errno := syscall.Syscall6(
-		syscall.SYS_PTRACE,
-		uintptr(ptracePeekuser),
-		uintptr(pid),
-		offset,
-		0, 0, 0,
-	)
-	if errno != 0 {
-		return 0, errno
-	}
-	return uint64(val), nil
-}
@@ -4,26 +4,30 @@ How gasm-devkit is put together and why.

 Repository: [sourcedock.dev/petrbalvin/gasm-devkit](https://sourcedock.dev/petrbalvin/gasm-devkit)

-## Design goals
+## Overview
+
+Three design goals shape everything below.

 1. **A real AST, not a grammar hack.**  The linter, analyser, assembler and
   language server all need to *reason* about assembly, not just colour it.
   So the centre of the toolkit is a hand-written lexer and a parser that
   produce a typed AST with source positions on every node.
 2. **Architecture as data, not code.**  Per-architecture differences (amd64,
-   arm64, riscv64, loong64) live in register and instruction *tables* (`arch`),
-   never in `if arch == …` branches scattered through the logic.  The
-   instruction tables are generated from the Go toolchain's own assembler
-   source (`just gen`), so adding or refreshing an architecture is a data
-   operation, not a coding one.
+   arm64, riscv64, loong64) live in register and instruction *tables* (`arch`)
+   and per-architecture encoders, rather than in `if arch == …` branches
+   threaded through the analysis; the arch tests that remain are dispatch and
+   policy points, such as which encoder a file name selects and which
+   registers the liveness pass audits.  The instruction tables are generated
+   from the Go toolchain's own assembler source (`just gen`), so refreshing an
+   architecture is a data operation, not a coding one.
 3. **Open integration surface.**  Everything the toolkit can do is reachable
   through two vendor-neutral interfaces: a CLI and an LSP server.  No editor
   owns the toolkit; the toolkit is offered to editors on standard terms.

-## Pipeline
+The components, and how data moves between them:

 ```mermaid
-graph TD
+flowchart TD
    SRC["source .s"] --> LEX["lexer<br/>token stream"]
    LEX --> PAR["parser<br/>AST + diagnostics"]
    LEX --> FMT["format<br/>re-space tokens"]
@@ -33,18 +37,55 @@ graph TD
    ARCH["arch tables<br/>amd64 / arm64 / riscv64 / loong64"] --> LINT
    ARCH --> LSP
    LINT --> LSP
+    PAR --> ASM["asm<br/>encoders, image, object emitters"]
+    ASM --> VER["verify<br/>JIT mapping, ABI checks, fuzzing"]
+    ASM --> DBG["debug<br/>ptrace session"]
+    VER --> DBG
+    DIS["disasm<br/>golang.org/x/arch"] --> DBG
    FMT --> CLI["gasm CLI"]
    LINT --> CLI
    PAR --> CLI
    LEX --> CLI
+    ASM --> CLI
+    VER --> CLI
+    DBG --> CLI
+    DIS --> CLI
    LSP --> EDITOR["any LSP editor"]
 ```

 The lexer is the shared foundation: the parser builds the AST from it, the
 formatter re-spaces its tokens directly, and the language server uses it for
-semantic highlighting.
+semantic highlighting.  The packages follow a dependency chain: static analysis
+builds only on the AST, the standalone assembler emits object code, and both
+the dynamic analysis and the debugger consume the execution substrate the
+assembler provides.

-## Components
+## Packages
+
+| Package | Responsibility |
+|---|---|
+| `token` | token kinds and positions |
+| `lexer` | hand-written scanner; permissive, and it never panics |
+| `ast` | the typed syntax tree: declarations, lines, operands |
+| `parser` | line-oriented parser producing the AST and its diagnostics |
+| `arch` | register and instruction tables for the four architectures |
+| `lint` | static checks over the AST |
+| `format` | canonical formatter over the token stream |
+| `lsp` | the language server |
+| `asm` | standalone assembler: encoders, image layout, object emitters |
+| `disasm` | disassembly backend over golang.org/x/arch |
+| `verify` | JIT execution, ABI checks, differential fuzzing |
+| `debug` | interactive ptrace debugger |
+| `cmd/gasm` | the CLI |
+| `_gen` | rebuilds the `arch` tables from the Go toolchain source |
+
+The boundaries matter as much as the responsibilities: `ast` records syntax
+only, and whether a name is a register or a label is left to `arch`, so the
+parser stays architecture-agnostic.  `asm` produces the machine code, `verify`
+and `debug` are the two packages that map it executable (read-execute in
+`verify`, read-write-execute in the debuggee), and `cmd/gasm` is the CLI, with
+the verify sweep orchestration and the audit, scaffold and unified-diff
+helpers beside its flags and output.

 ### `token` and `lexer`

@@ -84,14 +125,17 @@ Register files are generated programmatically (the regular `R8`-`R15`,
 `X0`-`X15`, `Y0`-`Y15`, `Z0`-`Z31`, `K0`-`K7` ranges) plus the irregularly
 named registers listed explicitly.  Instruction names are **generated from the
 Go toolchain's own assembler source** (`cmd/internal/obj/<arch>/anames.go`,
-plus the common opcodes and the per-architecture front-end aliases such as the
-arm64 `B`/`BL` branches and the `.P`/`.W` load-store addressing suffixes) by
-`just gen`, so the tables always match what the real assembler accepts.  Each
-mnemonic maps to a summary and an optional operand-count range; counts are
-recorded only where unambiguous (`-1` disables the operand-count lint for that
-instruction) so the linter stays silent rather than guess.  For architectures
-with highly variable operand forms (arm64, riscv64, loong64) only a few
-fixed-arity instructions (`RET`, `NOP`, `JMP`, `CALL`) carry counts at all.
+plus the common opcodes in `cmd/internal/obj/util.go`) by `just gen`, so the
+tables always match what the real assembler accepts.  The spellings the
+toolchain's tables do not carry are hand-maintained instead: the front-end
+alias lists in `arch/arm64.go`, `arch/amd64.go` and `arch/loong64.go` (the
+arm64 `B`/`BL` branches among them), and the arm64 `.P`/`.W` load-store suffix
+stripping in `arch/arch.go`.  Each mnemonic maps to a summary and an optional
+operand-count range; counts are recorded only where unambiguous (`-1`
+disables the operand-count lint for that instruction) so the linter stays
+silent rather than guess.  For architectures with highly variable operand
+forms (arm64, riscv64, loong64) `relaxCounts` clears those counts, leaving
+`RET` and `NOP` with a range (`RET` alone on riscv64).

 ### `lint`

@@ -134,7 +178,8 @@ Two deeper analyses sit on top of the AST:
 - **`unreachable-code`.**  Code after a `RET` and before the next label is
  dead.  The check is suppressed for any function whose reachability cannot be
  decided statically: those using PC-relative jumps (`JMP 2(PC)`),
-  register-indirect branches (`JALR`/`JR`/`JIRL`/`BR`/`BLR`), or living in a
+  register-indirect branches (`JALR`/`JR`/`JIRL`/`BR`/`BLR`, or a `JMP`/`CALL`
+  through a register or memory operand), or living in a
  file with `#ifdef` conditionals.  `UNDEF` is deliberately not a terminator:
  code after it is occasionally intentional metadata.
 - **`register-clobber` (register liveness).**  The linter builds the function's
@@ -202,20 +247,20 @@ them from the standard LSP legend, so no editor-specific grammar is needed.

 ### `asm`

-The standalone assembler (Phase 2).  Its core is an amd64 instruction encoder:
+The standalone assembler.  Its core is an amd64 instruction encoder:
 a REX/ModR-M/SIB/displacement/immediate engine plus the scalar instruction set,
 with the Plan 9 operand order (source first) mapped onto the x86 encoding.
 Every encoding is validated by decoding it again with `golang.org/x/arch`, the
-one module dependency, used in tests only and never linked into the binary.
+one module dependency, which also backs the `gasm dis` listings.

-A **RISC-V encoder** (Phase 5, RV64IMAFDC + RVC compression) encodes the full
+A **RISC-V encoder** (RV64IMAFDC + RVC compression) encodes the full
 integer, atomic, float/double, FMA and CSR instruction sets with the MOV
 pseudo-instruction and SB/global symbol references (AUIPC pairs with
 R_RISCV_PCREL_HI20/LO12 relocations).  The encoder compresses eligible
 instructions to 16-bit RVC forms and is validated byte-for-byte against
 `GOARCH=riscv64 go tool asm`.

-A **LoongArch encoder** (Phase 5, LoongArch64) encodes the integer and
+A **LoongArch encoder** (LoongArch64) encodes the integer and
 floating-point instruction sets with the dual-form arithmetic mnemonics (3R
 vs 2RI12), the 16/21-bit branch families, the MOV pseudo-instruction and its
 constant materialisation (the dcon classification driving lu12i.w/ori/lu32i.d/
@@ -226,7 +271,7 @@ relocations).  Like the RISC-V encoder it is validated byte-for-byte against
 `GOARCH=loong64 go tool asm`, and its GOOBJ output is proven end-to-end by
 substituting it into a cross-compiled `go build` and linking with `cmd/link`.

-An **AArch64 encoder** (Phase 5, arm64) encodes the integer instruction set
+An **AArch64 encoder** (arm64) encodes the integer instruction set
 with the data-processing (shifted register and immediate forms), load/store
 (scaled unsigned immediate and unscaled9-bit immediate), conditional and
 unconditional branches, the MOV pseudo-instruction and its constant
@@ -262,11 +307,11 @@ registers are translated onto the hardware stack pointer: `x+N(FP)` becomes
 pointer is set up, with the matching Go prologue/epilogue generated, so the
 output is byte-identical to the Go assembler for these cases.  SIMD is handled
 by a VEX (AVX/AVX2) encoder (the two- and three-byte VEX prefixes with XMM/YMM
-registers) across eight operand forms: the three-operand NDS form, the
-two-operand reg/rm form, the immediate-shift form (plus the variable-count
-shifts, which share the NDS shape with the count in an XMM register or
-memory), the immediate shuffle form (`VPSHUFD`, `VPERMQ`), the
-three-operand-plus-immediate form (`VSHUFPD`,
+registers) over nine operand forms plus a dedicated move encoder: the
+three-operand NDS form, the two-operand reg/rm form, the immediate-shift form
+(plus the variable-count shifts, which share the NDS shape with the count in
+an XMM register or memory), the immediate shuffle form (`VPSHUFD`, `VPERMQ`),
+the three-operand-plus-immediate form (`VSHUFPD`,
 `VPERM2I128`, `VINSERTI128`), the lane-extract form (`VEXTRACTI128`,
 `VEXTRACTF128`, where the YMM source occupies the reg field and the XMM or
 memory destination r/m), the direction-sensitive moves (`VMOVDQU`, `VMOVUPD`,
@@ -311,10 +356,10 @@ b bit and the L'L rounding-control field (broadcast keeps the vector length
 and scales disp8 by the element size), and combine with the .Z zeroing
 suffix.  Every encoding is validated two ways: by
 round-trip decoding through `golang.org/x/arch`, and byte-for-byte against
-the machine code the real Go assembler emits, a comparison that holds for
-whole functions: all 27 functions of both kernels assemble to exactly the Go
-toolchain's bytes, the lone exception being the displacements of the
-static-constant loads, which the Go linker fills at link time.
+the machine code the real Go assembler emits; the parity suites carry that
+comparison over whole kernel files on all four architectures, with the
+relocation fields masked because the Go linker fills those displacements at
+link time.

 File-level assembly (`AssembleFile`) goes beyond single functions: it
 materialises the file's static symbols (`GLOBL`/`DATA`) in a data section
@@ -361,7 +406,7 @@ compiled packages it references.

 ### `verify`

-The dynamic-analysis substrate (Phase 3).  It JIT-loads assembled images into
+The dynamic-analysis substrate.  It JIT-loads assembled images into
 executable memory and invokes them directly, enabling differential testing,
 runtime ABI checks and coverage profiling.

@@ -381,10 +426,10 @@ every architecture too: `enterJITChecked` plants sentinels in the registers
 the Go ABI fixes across calls (amd64 `BP`/`R14`, arm64 `R29`/`R28`, riscv64
 `X27`, loong64 `R22`; the latter two keep no hardware frame pointer) and the
 raw return trampoline `leaveJITCheckedRaw` verifies them, restoring the
-saved registers before Go code resumes.  riscv64 is validated end to
-end under qemu-user emulation; arm64 shares the same stack convention and
-fix; loong64 stays ground-truth-only until hardware validation (see
-docs/DECISIONS.md).  `gasm verify` runs the JIT checks when the host
+saved registers before Go code resumes.  All three non-amd64 trampolines
+are validated end to end under qemu-user emulation, the loong64 one
+through its raw-address leave handoff.
+`gasm verify` runs the JIT checks when the host
 matches the kernel's architecture and the toolchain comparisons
 elsewhere.

@@ -397,10 +442,13 @@ The `gasm verify` CLI subcommand exposes this: it loads a file, reports the
 available functions and (with `-smoke`) calls each NOSPLIT function with zeroed
 arguments to confirm the trampoline round-trips.  The `-smoke` and `-abi`
 sweeps run in parallel and each inside a child process, so a function that
-faults is reported without ending the sweep.  `gasm verify --fuzz` combines
-ABI checks (sentinel registers, canary, stack bounds) with differential fuzz
-testing, comparing the JIT-assembled kernel against the portable Go reference
-bit-for-bit while verifying the ABI contract on every iteration.  When a fuzz
+faults is reported without ending the sweep; `-abi` is where the ABI check
+lives, fuzzing each function with sentinel values in the registers the Go ABI
+fixes across calls and a canary below `SP`, and reporting a violation on any
+iteration.  `gasm verify --fuzz` is the differential campaign instead: it
+JIT-loads the kernel and the `go tool asm` build of the same kernel and
+compares the output argument areas bit-for-bit, one child process per function
+so a crash on a partial function is reported rather than fatal.  When a fuzz
 iteration crashes or mismatches, `FuzzResult.CrashInput` stores the exact input
 for reproducibility.  `gasm verify --call <func> --buf name:size:pattern`
 invokes a single function with user-supplied buffers (patterns: zero, ones,
@@ -420,30 +468,97 @@ masked), reporting any encoding drift.
 The interactive debugger (all four architectures).  It launches the target
 function in a child process that maps the JIT code, calls
 `PTRACE_TRACEME`, and stops; the parent attaches via ptrace and controls
-execution.  Breakpoints are patched as INT3 bytes through `/proc/pid/mem`
-(PTRACE_PEEKTEXT is unreliable with Go's multi-threaded runtime).
+execution.  Breakpoints are patched through `/proc/pid/mem`: the one-byte
+`INT3` on amd64, the four-byte break instruction on the other three (arm64
+`BRK #0`, riscv64 `ebreak`, loong64 `break 0`).
 The child pins its goroutine to the OS thread with `runtime.LockOSThread`
 so the traced thread is the one executing JIT code.  The REPL provides
-single-step, register inspection (GPR + YMM/XMM via `PTRACE_GETFPREGS`),
+single-step, register inspection (the GPRs on every architecture; on amd64 the
+XMM set through `PTRACE_GETFPREGS` and the YMM set through `PTRACE_GETREGSET`
+on `NT_X86_XSTATE`; on the other three the FP/SIMD regset through
+`PTRACE_GETREGSET` on `NT_PRFPREG`),
 label resolution, named buffer allocation with pattern filling
 (`--buf name:size:pattern`: zero, ones, seq, or hex), and breakpoint
 management.  Breakpoints accept conditions
 (`break <label> if <reg> <op> <val>`, including register-against-register
-comparisons), and hardware watchpoints work on all four architectures.
+comparisons), and hardware watchpoints work on amd64 (the DR0-DR3 debug
+registers), arm64 (`NT_ARM_HW_WATCH`) and loong64 (`NT_LOONGARCH_HW_WATCH`);
+riscv64 reports that its kernel ptrace interface exposes no trigger regset.
+The ptrace path is validated at run time on amd64, where the session tests are
+built; arm64, riscv64 and loong64 compile and are covered by the
+architecture-neutral units (label and line tables, the breakpoint manager).
 For non-interactive use, `--script` runs REPL commands from a file (or
 stdin) and exits, `--timeout` kills the debuggee when a run hangs (the
 watchdog is armed before the ptrace attach, so a sandboxed debuggee cannot
 block it), and `--cover` runs to completion with a breakpoint on every
-label and reports which blocks executed.
+instruction and reports which instructions executed and how often.

-## Extension points
+### Extending the toolkit

 - **New architecture:** add an entry to the generator in `_gen`, run
  `just gen`, and add a `buildXXX()` register file plus a case in `ForArch`.
 - **New lint rule:** add a function in `lint` and a rule-code constant.
 - **New LSP feature:** add a method case in `dispatch` and a handler.

-The phases follow a dependency chain.  Phase 1 (static analysis) builds only on
-the AST; Phase 2 (the standalone assembler) emits object code; Phases 3
-(dynamic analysis) and 4 (the debugger) both consume the execution substrate
-that the assembler provides.
+## Data flow
+
+The main operation, assembling one file:
+
+```mermaid
+sequenceDiagram
+    participant User
+    participant CLI as gasm CLI
+    participant Parser as parser
+    participant Asm as asm
+    participant Go as go toolchain
+    User->>CLI: gasm asm --format goobj -p pkg -o k.o k_amd64.s
+    CLI->>Parser: Parse(path, src)
+    Parser-->>CLI: AST, diagnostics
+    CLI->>Asm: AssembleFile(AST)
+    Asm->>Asm: encode operands, settle label offsets, lay out data
+    Asm-->>CLI: Image, code and data and relocations
+    CLI->>Asm: GOObject(pkg, path)
+    Asm->>Go: go list -json -export, externals only
+    Go-->>Asm: package and symbol indices
+    Asm-->>CLI: Go object bytes
+    CLI-->>User: wrote N bytes to k.o
+```
+
+Errors are produced where the parse or the encoding fails and become values at
+the CLI boundary: the parser returns a diagnostic list and never aborts a file,
+`AssembleFile` returns an error, and `cmd/gasm` prints what it has to stderr
+and returns a non-zero exit code.  The formatter and the linter take different
+inputs from the assembler: `gasm fmt` re-spaces the token stream
+(`format.Source` lexes the source text itself) and `gasm lint` walks the parsed
+AST, so neither depends on an encoding.
+
+## State and lifetime
+
+- The analysis packages (`lexer`, `parser`, `format`, `lint`, `arch`) hold only
+  read-only lookup tables and no mutable state: every call allocates its own
+  tokens and AST, and any number of goroutines may read the `arch` tables.
+- A `verify.Kernel` owns one executable mapping, which `Close` releases.  The
+  JIT trampolines keep the Go stack pointer and the checked-call sentinels in
+  package globals, so a call is a process-wide, one-at-a-time operation.  The
+  `gasm verify` sweeps therefore run each function in a child process, which
+  contains a crash and keeps the globals unshared.
+- `lsp.Server` is long-lived: it runs a single read and dispatch loop over the
+  stream and touches its document store only from that loop, so one server
+  serves one connection.
+- A `debug.Session` owns a traced child process and pins its goroutine to the
+  forking OS thread, because ptrace requests must stay on that thread.
+
+## Dependencies
+
+- **`golang.org/x/arch`** (v0.30.0) is the one module dependency: it is the
+  disassembler backend (`gasm dis` and the debugger's listings).  The tests
+  additionally decode through it to validate the encodings.
+- **The Go toolchain**, as an oracle and never as a library: `go tool asm`
+  supplies the object preamble and the ground truth for `gasm verify
+  --ground-truth`, `go list -json -export` locates the archives of the packages
+  a GOOBJ object references, and `_gen` parses
+  `$GOROOT/src/cmd/internal/obj/<arch>/anames.go` to rebuild the tables.
+- **Linux process interfaces** for the dynamic work: `mmap` and `mprotect` for
+  the JIT mapping, ptrace with `/proc/pid/mem` for the debugger.  That is why
+  `verify` runs a JIT check only when the host architecture matches the
+  kernel's, and why `debug` is Linux-only.
@@ -1,215 +1,496 @@
-# CLI Reference
+# Command line

-Repository: [sourcedock.dev/petrbalvin/gasm-devkit](https://sourcedock.dev/petrbalvin/gasm-devkit)
+The reference below is taken from the program's own `--help`. If the two disagree, the
+program is right and this file is a defect.

-`gasm` is a single binary with subcommands. Run `gasm --help` for an
-overview, or `gasm <command> -h` for a command's usage and flags.
+The same reference is installed as man pages: `just install-man` puts gasm(1) and a page
+for every command except `version` (which gasm(1) itself documents) into
+~/.local/share/man (`MANDIR` overrides). A test in `cmd/gasm` keeps the two from
+drifting: it compares each page's flag set and SYNOPSIS line with the binary's own `-h`
+output, and gasm(1)'s COMMANDS list with the top-level help. The prose is not compared.

-## Global Flags
+## Synopsis

-| Flag | Description |
-|------|-------------|
-| `-h`, `--help` | Show help |
-| `-V`, `--version` | Print the version |
+```sh
+gasm [global flags] <command> [command flags] [arguments]
+```

-## `gasm tokens <file>`
+## Commands

-Print the lexical token stream of FILE: position, token kind, and text,
-one token per line. FILE may be `-` to read standard input.
+| Command | Purpose |
+|---|---|
+| `tokens` | print the lexical token stream |
+| `parse` | parse a file and report syntax errors |
+| `fmt` | canonicalise the formatting of `.s` files |
+| `lint` | run the static checks |
+| `asm` | assemble `.s` files to machine code |
+| `dis` | disassemble machine code or an assembled file |
+| `verify` | JIT-assemble and run the dynamic checks |
+| `debug` | interactive source-level debugger |
+| `diff` | compare the machine code of two `.s` files |
+| `profile` | show the basic-block structure of the functions |
+| `audit-instructions` | diff the encoder against the toolchain's name table |
+| `scaffold` | generate a differential test skeleton for a kernel |
+| `lsp` | run the language server over stdio |
+| `version` | print the version |

-## `gasm parse <file>`
+## tokens

-Parse FILE and report syntax errors on stderr. On success, prints how
-many declarations and TEXT functions the file contains.
+```text
+Usage: gasm tokens <file>
+```

-## `gasm fmt [-w|-l|-d] [path...]`
+Print the lexical token stream of FILE: position, token kind and text, one
+token per line. FILE may be `-` to read standard input.

-Canonicalise the formatting of Plan 9 assembly sources: indentation,
-operand spacing, per-function mnemonic alignment, and blank-line layout.
+```sh
+gasm tokens hello_amd64.s
+```

-| Flag | Description |
-|------|-------------|
-| `-w` | Write result to the source file (default: print to stdout) |
-| `-l` | List files whose formatting differs, one per line; write nothing |
-| `-d` | Print a unified diff of the canonical formatting instead |
+```text
+1:1	#	"#"
+1:2	IDENT	"include"
+1:10	STRING	"\"textflag.h\""
+```

-With no arguments, or with a directory argument, every `.s` file below
-it is reformatted in place and the names of changed files are listed
-(`go fmt` style). `.` and `_` directories are skipped.
+## parse

-## `gasm lint <file...>`
+```text
+Usage: gasm parse <file>
+```

-Run static checks and print diagnostics as
-`file:line:col: severity: message [code]`. Exit status is non-zero when
-an error-severity diagnostic is found.
+Parse FILE and report syntax errors on stderr. On success, print how many
+declarations and TEXT functions the file contains. FILE may be `-` to read
+standard input.

-| Flag | Description |
-|------|-------------|
-| `-disable` | Comma-separated rule codes to disable |
+```sh
+gasm parse hello_amd64.s
+```
+
+```text
+hello_amd64.s: OK, 2 declarations, 1 functions
+```
+
+## fmt
+
+```text
+Usage: gasm fmt [-w|-l|-d] [path...]
+```
+
+| Flag | Default | Effect |
+|---|---|---|
+| `-w` | off | write the result back to the source file |
+| `-l` | off | list the files whose formatting differs; write nothing |
+| `-d` | off | print a unified diff of the canonical formatting instead |
+
+`-l` and `-d` are mutually exclusive. With no arguments, or with a directory
+argument, every `.s` file below it is reformatted in place and the names of the
+changed files are listed, the way `go fmt` does; `.` and `_` directories are
+skipped. Explicit file arguments print to stdout unless `-w` is given.
+
+```sh
+gasm fmt -l kernel_amd64.s
+```
+
+Empty output means every file is formatted, which is the shape a CI check
+wants; `-d` shows what would change:
+
+```sh
+gasm fmt -d ugly_amd64.s
+```
+
+```text
+--- ugly_amd64.s
+++ ugly_amd64.s
+@@ -2,8 +2,8 @@
+
+ // func add(a, b int) int
+ TEXT ·add(SB), NOSPLIT, $0-24
+-        MOVQ a+0(FP), AX
+-        ADDQ b+8(FP), AX
+	MOVQ a+0(FP), AX
+	ADDQ b+8(FP), AX
+```
+
+## lint
+
+```text
+Usage: gasm lint <file...>
+```
+
+| Flag | Default | Effect |
+|---|---|---|
+| `-disable` | empty | comma-separated rule codes to disable |
+
+Diagnostics are printed as `file:line:col: severity: message [code]`. The exit
+status is non-zero when an error-severity diagnostic is found; warnings (the
+register-clobber audit, for example) do not affect it.

 Rules: `unknown-instruction`, `operand-count`, `undefined-label`,
 `duplicate-label`, `missing-ret`, `missing-textflag-include`,
 `abi-argsize`, `unreachable-code`, `register-clobber`,
 `funcdata-pcdata`, `unused-label`, `invalid-textflag`,
 `stack-imbalance`, `register-width-mismatch`, `abi0-register-args`,
-`nonportable-register-name` and `unencodable-instruction`.
+`nonportable-register-name`, `unencodable-instruction` and
+`reserved-register-write`.

-## `gasm asm [--format raw|elf|goobj] [-p pkg] [-o out] <file>`
+```sh
+gasm lint kernel_amd64.s
+```

-Assemble FILE to machine code (amd64, arm64, riscv64, loong64).
+## asm

-| Flag | Description |
-|------|-------------|
-| `--format` | Output format: `raw` (default), `elf`, `goobj` |
-| `-p` | Package path (required for `--format goobj`) |
-| `-o` | Write output to file (default: hex dump to stdout) |
+```text
+Usage: gasm asm [--format raw|elf|goobj] [-I dir] [-p pkg] [-GOARCH arch] [-o out] <file>
+```

-## `gasm dis [-a arch] <file>`
+| Flag | Default | Effect |
+|---|---|---|
+| `-format` | `raw` | output format: `raw` (concatenated image), `elf` or `goobj` (Go object) |
+| `-I` | empty | directory to search for `#include` files; may be repeated, searched in order after the source directory |
+| `-p` | empty | package path for `--format goobj`, qualifying the exported symbols |
+| `-GOARCH` | empty | target architecture: `amd64`, `arm64`, `riscv64` or `loong64`; overrides the file-name suffix |
+| `-o` | empty | write the output to this file instead of a hex dump on stdout |

-Disassemble machine code to instruction text (via `golang.org/x/arch`).
+Supported architectures: amd64 (VEX/AVX2 and EVEX/AVX-512 included), arm64,
+riscv64 (RV64IMAFDC and RVC) and loong64, taken from the file's `_arch.s`
+suffix or from `-GOARCH`, which is how files whose names carry no
+recognisable suffix (most of GOROOT's, for example `cpu_x86.s`) are
+assembled. `raw` concatenates the functions and the data section into one
+self-consistent image; `elf` emits a relocatable object that links with the
+system toolchain; `goobj` emits the Go toolchain's own object format, which
+`cmd/link` consumes directly, and is the one format that needs the toolchain
+installed: the object preamble is captured from `go tool asm` and the format
+version from `go version`. `raw` and `elf` need no toolchain at all.

-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).
+Assembly preprocessing matches the toolchain's: `#define` macros (object and
+parameterised) expand at the point of use, `#undef`, `#ifdef`, `#ifndef`,
+`#else` and `#endif` behave as in `go tool asm`, `;` separates statements,
+and `#include "file"` splices the named file in, resolved against the source
+directory and then each `-I` directory in order. `textflag.h` is the one
+header that is not spliced: gasm consumes its flag names natively.

-| Flag | Description |
-|------|-------------|
-| `-a` | Architecture for raw input without a `_arch.s` name |
+```sh
+gasm asm hello_amd64.s
+```

-## `gasm verify [flags] <file.s>`
+```text
+add: 16 bytes
+  0000: 48 8b 44 24 08 48 03 44 24 10 48 89 44 24 18 c3
+```

-Assemble FILE, map it into executable memory, and run dynamic checks.
+## dis

-| Flag | Description |
-|------|-------------|
-| `--ground-truth` | Compare machine code byte-for-byte against `go tool asm` |
-| `--fuzz` | Differential fuzz: JIT both gasm and go-tool-asm, compare outputs |
-| `-n` | Fuzz iterations per function (default: 1000) |
-| `--abi` | Run ABI-checking calls (sentinel registers + red zone) |
-| `--abi-n` | Number of ABI check iterations with varied inputs (default: 100) |
-| `--profile` | List basic-block structure per function |
-| `--smoke` | Call each NOSPLIT function with zeroed args |
-| `--call <func>` | Invoke a single function with `--buf` instead of the sweeps |
-| `--buf <spec>` | Buffer spec for `--call`: `name:size:pattern[,name:size:pattern]` |
-| `--args <spec>` | Scalar args for `--call`: `name=value[,name=value]` (decimal or `0x` hex) |
-| `--repeat <n>` | Number of times to repeat a `--call` invocation (default: 1) |
-| `--save-corpus <dir>` | With `--fuzz`: write each failing input to DIR as replayable JSON |
-| `--replay <dir>` | Re-run saved corpus entries (JSON in DIR), one child process per entry |
+```text
+Usage: gasm dis [-a arch] <file>
+```

-The `--fuzz` mode runs each function in a subprocess; a partial function
-(e.g. a decoder that faults on malformed input) is reported as
-`CRASH` without killing the parent. Use `--call` with `--buf` to invoke
-partial functions with valid data instead.
+| Flag | Default | Effect |
+|---|---|---|
+| `-a` | empty | architecture for raw input without a `_arch.s` name |

-The `--call` mode parses the `// func` signature, allocates the requested
-buffers (`zero`, `ones`, `seq`, or a hex blob), builds the ABI0 argument
-block with buffer pointers/lengths/capacities at the matching parameter
-offsets, and prints the arg block before and after the call, showing
-return values and any output written to the buffers.  Scalar parameters
-are supplied with `--args` (decimal, or `0x` hex) at their ABI0 offsets.
+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.
+With any other file, or `-` for standard input, the bytes are disassembled
+linearly and `-a` selects the architecture (amd64, arm64, riscv64 or loong64).

-The `--save-corpus` mode records the logical arguments (buffer contents and
-scalars, not raw pointers) of every failing fuzz input as JSON.  `--replay`
-rebuilds a live argument block from each entry and calls it in its own child
-process, reporting `OK`, `CRASH (reproduced)` or `FAIL` per entry and
-exiting non-zero when any entry fails.
+```sh
+gasm dis hello_amd64.s
+```

-## `gasm debug [--func <name>] [--buf spec] [--script file] <file.s>`
+```text
+add: 16 bytes
+  0000: 48 8b 44 24 08   mov rax, qword ptr [rsp+0x8]
+  0005: 48 03 44 24 10   add rax, qword ptr [rsp+0x10]
+  000a: 48 89 44 24 18   mov qword ptr [rsp+0x18], rax
+  000f: c3               ret
+```

-Interactive debugger for JIT-assembled functions (amd64, arm64, riscv64,
-loong64). Requires a compiled binary on `$PATH` (not `go run`).
+## verify

-| Flag | Description |
-|------|-------------|
-| `--func` | Function to debug (required) |
-| `--buf` | Buffer spec: `name:size:pattern[,name:size:pattern]` |
-| `--args <file>` | File containing the ABI0 argument block |
-| `--script <file>` | Run REPL commands from a file (one per line) and exit; `-` reads stdin |
-| `--timeout <dur>` | Kill the debuggee after this duration (e.g. `30s`); for headless `--script` runs |
-| `--cover` | Run to completion with a breakpoint on every instruction; report which executed, how often, and which labels were reached |
+```text
+Usage: gasm verify [-smoke] [-abi] [-fuzz] [-ground-truth] [-profile] [-call] <file.s>
+```
+
+| Flag | Default | Effect |
+|---|---|---|
+| `--ground-truth` | off | compare the machine code byte-for-byte against `go tool asm` |
+| `--fuzz` | off | differential fuzz against the `go tool asm` build |
+| `-n` | 1000 | fuzz iterations per function |
+| `--abi` | off | ABI-checking calls: sentinel registers and a red-zone canary |
+| `--abi-n` | 100 | ABI check iterations with varied inputs |
+| `--profile` | off | list the basic-block structure per function |
+| `--smoke` | off | call each NOSPLIT function with zeroed arguments |
+| `--call` | empty | invoke a single NOSPLIT function with `--buf` instead of the sweeps |
+| `--buf` | empty | buffer spec for `--call`: `name:size:pattern[,name:size:pattern]` |
+| `--args` | empty | scalar args for `--call`: `name=value[,name=value]` (decimal or `0x` hex) |
+| `--repeat` | 1 | number of times to repeat a `--call` invocation |
+| `--save-corpus` | empty | with `--fuzz`: write each failing input to this directory as replayable JSON |
+| `--replay` | empty | re-run saved corpus entries, one child process per entry |
+
+The JIT checks run when the host matches the file's architecture; the
+toolchain comparison works everywhere. `--fuzz`, `--smoke` and `--abi` run each
+function in its own child process, so a partial function that faults on random
+input is reported as `CRASH` instead of ending the sweep; `--call` with `--buf`
+invokes such a function with valid data. The function named by `--call` must be
+NOSPLIT: a function with a stack frame is refused with a diagnostic and exits 1.
+
+```sh
+gasm verify --ground-truth hello_amd64.s
+```
+
+```text
+hello_amd64.s: 1 functions JIT-loaded
+  add: MATCH (16 bytes)
+ground truth: 1/1 functions byte-identical
+  add: 16 bytes, args=24, frame=0 NOSPLIT
+```
+
+```sh
+gasm verify --call add --args a=2,b=3 hello_amd64.s
+```
+
+```text
+add: 16 bytes, args=24
+  signature: func add(a int, b int) int
+  scalars:
+    a = 2
+    b = 3
+  args before: 02 00 00 00 00 00 00 00 03 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 (24 bytes)
+  args after:  02 00 00 00 00 00 00 00 03 00 00 00 00 00 00 00 05 00 00 00 00 00 00 00 (24 bytes)
+  call 1: OK
+```
+
+## debug
+
+```text
+Usage: gasm debug <file.s> --func <name>
+```
+
+| Flag | Default | Effect |
+|---|---|---|
+| `-func` | empty | the function to debug, required |
+| `-buf` | empty | buffer spec: `name:size:pattern[,name:size:pattern]` (zero, ones, seq or hex) |
+| `-args` | empty | file containing the ABI0 argument block |
+| `-script` | empty | run REPL commands from a file, one per line, and exit; `-` reads stdin |
+| `-timeout` | 0 | kill the debuggee after this duration, for headless `-script` runs; a timeout exits 3 |
+| `-cover` | off | run to completion with a breakpoint on every instruction and report which executed |
+
+The debugger re-executes the binary it is running as (`os.Executable()`) for the
+traced child, so the child is the same `gasm`, whether it is installed on `$PATH`
+or run with `go run ./cmd/gasm`; nothing has to be installed first. Requires
+Linux (ptrace), and all four architectures are supported.

 REPL commands:

-| Command | Description |
-|---------|-------------|
-| `break <label\|addr> [if <reg> <op> <val>]` | Set a breakpoint, optionally conditional |
-| `delete <label\|addr>` | Remove a breakpoint |
-| `info break` | List all breakpoints |
-| `step [n]`, `s` | Single-step n instructions |
-| `next`, `n` | Step over CALL |
-| `finish`, `fin` | Run until the function returns |
-| `continue`, `c` | Run until breakpoint, watchpoint or exit |
-| `disas [n]`, `u` | Disassemble n instructions at PC |
-| `regs` | Print general-purpose + vector/FP 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 |
-| `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 or all watchpoints |
-| `labels`, `l` | List function labels and offsets |
-| `help`, `h`, `?` | Show command help |
-| `quit`, `q` | Kill the debuggee and exit |
+| Command | Effect |
+|---|---|
+| `break <label\|addr\|line> [if <reg> <op> <val\|reg\|*addr>]`, `b` | set a breakpoint; the condition compares a register with a constant, another register or the 8-byte word at `*addr` |
+| `delete <label\|addr>`, `d` | remove a breakpoint |
+| `info break`, `info breakpoints`, `info b` | list the breakpoints |
+| `step [n]`, `s` | single-step n instructions |
+| `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 the PC |
+| `regs` | print the general-purpose and vector/FP registers |
+| `where` | show the source line and the nearest label at the PC |
+| `stack` | show the stack near RSP, the return address and the ABI0 args |
+| `bt`, `backtrace` | backtrace: the current frame and the return address |
+| `x [addr] [len]` | hex-dump memory |
+| `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 of them |
+| `labels`, `l` | list the function's labels and offsets |
+| `help`, `h`, `?` | show the command help |
+| `quit`, `q` | kill the debuggee and exit |

-## `gasm diff [--map old=new,...] <file1.s> <file2.s>`
+```sh
+gasm debug --func add --cover hello_amd64.s
+```

-Compare the machine code produced by assembling two files. Shows which
-functions differ and the first few differing bytes. Useful for verifying
-that two implementations produce identical code, or for tracking encoding
-changes between Go assembler versions.
+## diff

-| Flag | Description |
-|------|-------------|
-| `--map` | Comma-separated `old=new` pairs to match functions with different names |
+```text
+Usage: gasm diff [-GOARCH arch] [-I dir] <file1.s> <file2.s>
+```

-Without `--map`, functions are paired by exact name. With `--map`, a
-function named `old` in the first file is compared against the function
-named `new` in the second file (e.g. `--map wideCopyAVX2=wideCopyAVX512`
-pairs AVX2 and AVX-512 variants regardless of suffix).
+| Flag | Default | Effect |
+|---|---|---|
+| `-GOARCH` | empty | target architecture for both files, overriding the file-name suffixes |
+| `-I` | empty | directory to search for `#include` files; may be repeated, searched in order after the source directory |
+| `-map` | empty | comma-separated `old=new` pairs to match functions with different names |

-## `gasm profile <file.s>`
+Functions are paired by exact name unless `--map` says otherwise, so
+`--map wideCopyAVX2=wideCopyAVX512` pairs two variants regardless of suffix.
+The exit status is non-zero when anything differs.

-Show the basic-block structure of functions in an assembly file. Lists
-each function's labels, their offsets, and the block boundaries. This is
-the static structure; for runtime execution counts, use `gasm verify
--fuzz` which exercises the code paths.
+```sh
+gasm diff hello_amd64.s hello_amd64.s
+```

-## `gasm audit-instructions [amd64|arm64|riscv64|loong64]`
+```text
+add: identical (16 bytes)
+all functions identical
+```

-Compare the gasm encoder for the given architecture (default amd64)
-against the installed `go tool asm` and print the diff: superset
-encodings (gasm-only spellings, shippable via `gasm asm --format goobj`),
-known-but-unencodable names (the encoder backlog) and go-only names
-(feature gaps). The Go side is probed black-box with a battery of operand
-shapes per mnemonic, so the audit tracks whatever toolchain
-`go env GOROOT` provides. On non-amd64 architectures the backlog is an
-over-approximation: a name counts as encodable only when a probe shape
-assembles cleanly, so a name whose real forms the battery misses lands
-in the backlog.
+## profile

-## `gasm scaffold differential <file.s>`
+```text
+Usage: gasm profile <file.s>
+```

-Print a differential test skeleton for every `// func` signature in
-FILE. The generated 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.
+Show the basic-block structure of each function: its labels, their offsets and
+the block boundaries. This is the static structure; for runtime execution
+counts use `gasm debug --cover`, and for input coverage `gasm verify --fuzz`.

-## `gasm lsp`
+```sh
+gasm profile hello_amd64.s
+```

-Run the language server over standard input/output (JSON-RPC 2.0 with
-Content-Length framing). Point an LSP-capable editor at the binary and
-associate it with `.s` files. The target architecture is inferred from
-the file-name suffix (`_amd64.s`, `_arm64.s`, `_riscv64.s`,
-`_loong64.s`).
+```text
+add: 16 bytes, args=24, frame=0 NOSPLIT
+  basic blocks: 1
+```

-Provides: completion, hover, document symbols, push and pull
-diagnostics, semantic tokens, go-to-definition, find references, rename,
-document formatting, inlay hints, code actions, signature help, document
-highlights, workspace symbol search, #include document links, and
-folding ranges for function bodies.
+## audit-instructions
+
+```text
+Usage: gasm audit-instructions [--corpus [dir]] [-I dir] [amd64|arm64|riscv64|loong64]
+```
+
+Compare the gasm encoder for the given architecture (default amd64) against the
+installed `go tool asm` and print the diff: superset encodings (gasm-only
+spellings, shippable via `gasm asm --format goobj`) and known-but-unencodable
+names (the encoder backlog). The Go side is probed black-box one bare mnemonic
+at a time, classified by the toolchain's diagnostic for an instruction it does
+not know, 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 on the other architectures. On non-amd64
+architectures the backlog is therefore an over-approximation: a name counts as
+encodable only when a probe shape assembles cleanly, so a name whose real forms
+the battery misses lands in the backlog. Names the toolchain knows and gasm does
+not cannot be enumerated by probing at all, because Go's table is visible only
+through names already in the gasm table; the report closes with a note saying
+so rather than listing them.
+
+```sh
+gasm audit-instructions amd64
+```
+
+```text
+gasm table (amd64, families excluded): 1542 mnemonics
+gasm encodable: 587   go tool asm recognised: 1542
+shared: 587
+...
+```
+
+With `--corpus` the audit changes shape: it assembles every `.s` file under
+DIR (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 headline number (files
+that assemble for every target architecture), the per-architecture pass rates
+and the most common failure reasons with one representative file each, which
+drive the encodability backlog by frequency rather than by table order. A run
+over GOROOT takes under a second.
+
+```sh
+gasm audit-instructions --corpus
+gasm audit-instructions --corpus "$(go env GOROOT)/src/crypto"
+```
+
+```text
+corpus /usr/local/go/src: 627 files (365 generic, attempted for all architectures)
+  assemble for every target architecture: 127 (20.3%)
+  amd64: 82/464 attempted
+     165  unsupported operand form
+         e.g. /usr/local/go/src/cmd/asm/internal/asm/testdata/386enc.s
+     109  instruction not encodable
+         e.g. /usr/local/go/src/cmd/asm/internal/asm/testdata/386.s
+...
+```
+
+## scaffold
+
+```text
+Usage: gasm scaffold differential <file.s>
+```
+
+Print a differential test skeleton for every `// func` signature in FILE. The
+generated test seeds random states, drives the kernel and a portable reference
+(`<name>Portable`), and compares the outputs byte-for-byte. Write the reference
+bodies, place the file in the kernel's package, and run it in CI.
+
+```sh
+gasm scaffold differential kernel_amd64.s > kernel_differential_test.go
+```
+
+## lsp
+
+```text
+Usage: gasm lsp
+```
+
+Run the language server over standard input/output, JSON-RPC 2.0 with
+`Content-Length` framing. Point an LSP-capable editor at the binary and
+associate it with `.s` files; the target architecture is inferred from the
+file-name suffix (`_amd64.s`, `_arm64.s`, `_riscv64.s`, `_loong64.s`).
+
+Provides: completion, hover, document symbols, push and pull diagnostics,
+semantic tokens, go-to-definition, find references, rename, document
+formatting, inlay hints, code actions, signature help, document highlights,
+workspace symbol search, #include document links, and folding ranges for
+function bodies. Definition, references and rename work across every open
+document.
+
+## version
+
+```text
+Usage: gasm version
+```
+
+Print the version the toolchain recorded for the build, the same string as
+`gasm --version`: the tag on a tagged checkout, a pseudo-version naming the
+commit below one, with `+dirty` appended on a dirty tree and `(devel)` outside
+version control.
+
+## Global flags
+
+| Flag | Default | Effect |
+|---|---|---|
+| `-h`, `--help` | off | print the usage |
+| `-V`, `--version` | off | print the version |
+
+## Exit codes
+
+| Code | Meaning |
+|---|---|
+| `0` | success |
+| `1` | a failure the program detected: a parse or assembly error, an error-severity lint diagnostic, a mismatch in `verify`, a file that cannot be read |
+| `2` | the arguments were wrong: a missing or extra argument, an unknown command or format, an invalid `--map` pair |
+| `3` | `debug --timeout` killed the debuggee |
+
+## Examples
+
+Assemble a kernel, check it, and run it:
+
+```sh
+gasm lint kernel_amd64.s
+gasm fmt -l kernel_amd64.s
+gasm asm -o kernel.bin kernel_amd64.s
+gasm verify --ground-truth kernel_amd64.s
+```
+
+Link the kernel into a Go program through the toolchain's own object format:
+
+```sh
+gasm asm --format goobj -p example.com/kernel -o kernel.o kernel_amd64.s
+```
+
+Find which labels a failing kernel reaches, headlessly:
+
+```sh
+gasm debug --func decodeBlockAVX2 --cover --timeout 30s kernel_amd64.s
+```
@@ -1,111 +0,0 @@
-# Deferred decisions
-
-Design decisions deliberately postponed, with enough context to pick them up
-again without re-deriving the analysis.  Each entry records what is deferred,
-why, the options on the table, and the trigger that should reopen it.
-
---
-
-## GOOBJ external (cross-package) symbol references
-
-**Status:** resolved (v0.29.0+, 2026-08-07).
-
-**Approach taken.**  Instead of parsing the compiler's iexport data (which
-would have required either `golang.org/x/tools` or an in-house parser), the
-resolver reads the **GOOBJ data directly** from the target package's `.a`
-archive.  The `.a` file contains a `_go_.o` member whose GOOBJ s is the
-same one gasm writes; the parser reuses the same layout (`blkSymdef`,
-`blkNonpkgdef`, the string table), so no new dependency was needed.
-
-**How it works.**
-
-1. `go list -json -export <pkg>` finds the target package's `.a` file.
-2. `extractGOOBJ` reads the ar archive, finds the `_go_.o` member, skips
-   the `"go object …\n!\n"` preamble and parses the GOOBJ header.
-3. `goobjFile.symbols()` walks `blkSymdef` and `blkNonpkgdef` in definition
-   order (the same order the linker uses) to build the symbol-to-index
-   mapping.
-4. `resolveExternalSymbols` wires the resolved `{PkgIdx, SymIdx}` into the
-   GOOBJ emission.
-
-The resolver is invoked automatically when `img.Externals` is non-empty; it
-runs `go list` as a subprocess (consistent with `toolchainObjectPreamble`
-which already calls `go tool asm`).  All symbol data is cached per package
-for the lifetime of the GOOBJ emission.
-
-## 2026-08-30 non-amd64 JIT execution trampolines
-
-**Status:** resolved for riscv64 (validated end to end under qemu-user)
-and arm64 (fix in place, consistent with the observed frame convention);
-open for loong64 until hardware validation.
-
-**Root cause (found 2026-08-31).**  The trampolines advanced SP past the
-leave-address slot after loading it, while the assembled kernels read
-their first argument at SP+8 per the frame convention (the amd64 path
-already kept SP on that slot).  Removing the advance fixed riscv64
-immediately (plain and checked ABI tests pass under qemu-user); the
-arm64 kernel's pre-fix trace showed exactly the same SP+8 reading.  The
-apparent arm64/loong64 "crashes in the JIT" turned out to be dominated
-by an unrelated instability: the Go 1.26 and 1.27 runtimes crash under
-qemu-user arm64 emulation (GC worker start, identical signature with the
-JIT tests skipped, both qemu 7.2 and 10.2), and the Go loong64 runtime
-does not start at all.  `gasm verify` therefore keeps loong64 kernels on
-the ground-truth path until hardware validation; the GOARCH-guarded
-tests (`verify/jit_arch_test.go`, `verify/abi_arch_test.go`) are the
-hardware validation entry point.
-
-**State.**  The per-architecture trampolines compile for all targets, the
-kernels they execute are byte-for-byte correct against `go tool asm`, and
-under `qemu-aarch64` the arm64 kernel demonstrably executes and stores its
-result correctly.  The failure is on the return path into Go code: arm64
-and loong64 take a SIGSEGV after the kernel's RET (the Go-side unwind
-through `leaveJIT` and its interposed ABIInternal wrapper is the suspect),
-and riscv64 returns cleanly but with an untouched result area.  amd64 is
-unaffected (the checked trampoline saves and restores BP/R14 and the flow
-is validated end to end).
-
-**Evidence harness.**  `verify/jit_arch_test.go` (plain call) and
-`verify/abi_arch_test.go` (checked call) are GOARCH-guarded tests; build
-the test binary per target (`GOARCH=arm64 go test -c -o v.test ./verify/`)
-and run it under `qemu-aarch64-static` from the `verify/` directory.  A
-minimal reproducer pattern lives in the qemu exploration notes: verify
-loads, the kernel executes, the fault follows the return.
-
-**Fix direction.**  Compare the amd64 checked trampoline (GLOBL/DATA raw
-address, explicit SP/BP/R14 save-restore) against the arm64/riscv64/
-loong64 `leaveJIT` unwind, in particular the interaction with the
-ABIInternal wrapper that `reflect.ValueOf(leaveJIT).Pointer()` returns.
-The plain-call path (no sentinels) fails the same way, so the checked
-path is not the variable.
-
---
-
-## 2026-08-29 tooling round
-
- `lint abi0-register-args`: flags kernels whose `// func` parameters are
-  never read from the FP frame.  Motivated by a real latent bug: kernels
-  reading arguments from registers pass every test while the autogenerated
-  `F.abi0` wrapper happens to leave the caller's register values intact, and
-  break on a toolchain upgrade.
- `lint nonportable-register-name`: the RAX/EAX register spellings are a gasm
-  extension; go tool asm rejects them, so files using them only link through
-  the gasm goobj path.
- `lint unencodable-instruction`: a mnemonic in the architecture table that
-  `asm.Encodable` rejects is flagged at edit time instead of failing at
-  assembly time.
- `audit-instructions`: black-box diff of the encoder against go tool asm.
-  As of this round the tables fully overlap on names; the audit exists to
-  catch drift in both directions (future supersets and future gaps).
- `scaffold differential`: generates the direct-call differential skeleton
-  (two independent seed sets, output and in-place buffer comparison) that a
-  pipeline-level fuzz can never replace.
- `verify --args`: scalar arguments for `-call`, closing the repro gap where
-  only buffers could be supplied.
- `debug --script/--timeout/--cover`: headless debugging with a watchdog
-  armed before the ptrace attach (untracing sandboxes hang the attach), and
-  label-level block coverage for the "did my test ever enter that branch"
-  question.
- Superset policy remains: gasm may accept spellings and encodings go tool
-  asm lacks, but such kernels ship only via `gasm asm --format goobj`; the
-  audit reports the superset surface.  The register-alias superset is warned
-  about by lint because the default `go build` path cannot consume it.
@@ -4,79 +4,98 @@ Repository: [sourcedock.dev/petrbalvin/gasm-devkit](https://sourcedock.dev/petrb

 ## Prerequisites

- **Go** 1.27+ with `toolchain go1.27.0`
+- **Go** 1.27.1, the exact version the `go` directive in `go.mod` declares
 - **just**, the command runner; every task below is a just recipe
- No external dependencies beyond the Go toolchain
+- **A C compiler** (`gcc`): `just race` runs the suite under the race detector,
+  which needs cgo
+- **Perl**: the `test`, `fmt-check`, `install-man` and `uninstall-man` recipes
+  are Perl programs
+- **`gzip`**: `install-man` compresses the man pages with it
+- A Linux host on amd64, arm64, riscv64 or loong64: `gasm debug` needs ptrace
+  and the JIT checks of `gasm verify` need executable memory
+- **`golang.org/x/arch`**, the one module dependency, which the Go toolchain
+  fetches; nothing else sits outside the standard library

-## Quick Start
+## Setup

 ```sh
 git clone https://sourcedock.dev/petrbalvin/gasm-devkit.git
 cd gasm-devkit
-just install   # go mod download
-just build     # go vet + gofmt, must pass with zero output
-just test      # full suite, race detector, 80 % coverage gate
+just build     # compile bin/gasm, zero errors and zero warnings
+just gates     # build, fmt-check, vet, test, race: the definition of done
 ```

-## Just Recipes
+## Recipes

-### `just install`
+Every recipe in the `justfile`, and what it does.

-`go mod download`.  The only module dependency, `golang.org/x/arch`, is
-used in tests only.
-
-### `just build`
-
-Runs `go vet ./...` and checks `gofmt -l .` produces no output. This is
-the minimum bar before any commit.
+| Recipe | What it does |
+|---|---|
+| `default` (bare `just`) | prints the recipe list (`@just --list`) |
+| `just build` | compiles `bin/gasm` with `CGO_ENABLED=0` and stripped symbols; zero errors and zero warnings |
+| `just test` | the test gate: the suite with `-count=1`, the coverage profile and the 80 % floor, then the CLI and debugger tests outside the profile |
+| `just race` | the same suite under the race detector; the expensive one, so it runs once, inside `gates` |
+| `just unit [packages] [run]` | fast, cached, scoped run for iterating: no race and no coverage, so an unchanged package reports instantly |
+| `just fuzz <target> <pkg> [fuzztime]` | time-boxed fuzz of one target; the package is required, because `go test -fuzz` refuses more than one |
+| `just bench [packages]` | benchmarks (`-benchmem -count=5`); on an idle machine only |
+| `just fmt` | formats the tree in place with `gofmt` |
+| `just fmt-check` | zero diff; prints nothing when everything is formatted, which is the shape the CI step wants |
+| `just vet` | both static gates: `go vet` and `go fix -diff` |
+| `just gates` | `build`, `fmt-check`, `vet`, `test` and `race`, in that order: the definition of done |
+| `just clean` | removes the build artefacts, `bin/` and `coverage.out` |
+| `just install` | builds, then copies the binary into `bindir` (`~/.local/bin`) |
+| `just uninstall` | removes the installed binary from `bindir` |
+| `just install-man` | installs the man pages under `docs/man` into `~/.local/share/man/man1` (`MANDIR` overrides), gzip-compressed; not a gate |
+| `just uninstall-man` | removes the installed man pages |
+| `just run` | runs the CLI with `go run -buildvcs=true`; the recipe takes no arguments, so flags go through the package instead |
+| `just dev` | the same as `run`; the project has no watcher to add |
+| `just gen` | regenerates the `arch` instruction tables from the Go toolchain source; not a gate |

 ### `just test`

 ```sh
-go test -race -count=1 ./...
+go test -count=1 -timeout 10m -coverprofile=coverage.out \
+    ./arch/... ./asm/... ./ast/... ./disasm/... ./format/... ./lexer/... \
+    ./lint/... ./lsp/... ./parser/... ./token/... ./verify/...
 ```

-Plus a coverage run over the ten analysable packages (arch, asm, ast,
-format, lexer, lint, lsp, parser, token, verify; `debug` and `cmd/gasm`
-need hardware or are CLI glue) and an `awk` gate that fails if total
-coverage is below 80 %.
-
-### `just fmt`
+The suite runs over the logic packages (`-count=1`, so no cached pass
+counts): arch, asm, ast, disasm, format, lexer, lint, lsp, parser,
+token, verify. `debug` traces a live process and `cmd/gasm` is thin CLI
+glue, so both sit outside the profile sweep, and a thin `cmd/` in it
+would drag the coverage total under the floor. Their tests still run, in
+a second invocation without a profile:

 ```sh
-gofmt -w .
+go test -count=1 -timeout 10m ./cmd/... ./debug/...
 ```

-Run after editing any Go source. The output must be idempotent.
+That covers the CLI's exit codes and the guard that compares the manual
+pages with the binary's own help, and the debugger's architecture-neutral
+units. The floor fails if the total is below 80 %.  CI runs the same two
+commands with the same ten-minute bound, so
+the number is the same everywhere.

-### `just run -- <args>`
-
-Runs the CLI via `go run` with the version string stamped:
+### `just run`

 ```sh
-just run -- lint kernel_amd64.s
-just run -- fmt -w kernel_amd64.s
-just run -- verify --ground-truth kernel_amd64.s
+just run
+go run -buildvcs=true ./cmd/gasm lint kernel_amd64.s
+go run -buildvcs=true ./cmd/gasm verify --ground-truth kernel_amd64.s
 ```

-### `just install-bin`
-
-Installs the `gasm` binary into `$GOBIN` with the release version
-embedded via `-ldflags "-X main.version=..."`.
+The flag on `go run` is there because it does not stamp the build otherwise,
+which `--version` would then report as `(devel)`.

 ### `just gen`

-Regenerates the architecture instruction tables in `arch/` by parsing
-the Go toolchain's own assembler source
+Regenerates the architecture instruction tables in `arch/` by parsing the Go
+toolchain's own assembler source
 (`$GOROOT/src/cmd/internal/obj/<arch>/anames.go`). Requires a Go
 installation. Output is committed, with no runtime dependency on the
 toolchain.

-### `just uninstall`
-
-Removes `coverage.out`, the `gasm` binary, and `*.test` artefacts.
-
-## Running Individual Tests
+## Running a single test

 ```sh
 go test -run TestVexGroundTruth ./asm/
@@ -85,32 +104,54 @@ go test -run TestGOObjectLinkAndRun ./asm/
 go test -run TestFuzzWideCopy ./verify/
 ```

-## Debugger Note
+Add `-v` for the sub-test names, and `-race` when the change touches
+concurrency. `-count=1` defeats the test cache when a result looks stale.

-`gasm debug` spawns a child process from the binary on `$PATH`. It does
-not work with `go run`; install first:
+## Coverage

 ```sh
-just install-bin
-gasm debug --func decodeBlockAVX2 path/to/kernel_amd64.s
+just test
+go tool cover -func=coverage.out
 ```

-## Project Layout
+The `total:` line is the number that matters, and it stays at 80 percent or
+more.

+## Debugging the build
+
+```sh
+go build -gcflags='-m' ./...    # inlining decisions
+go build -gcflags='-S' ./...    # what the compiler generated
+go tool asm -S kernel_amd64.s   # how the toolchain's assembler encodes a kernel
+gasm dis kernel_amd64.s         # what gasm makes of the same kernel
+gasm tokens kernel_amd64.s      # the token stream
+gasm profile kernel_amd64.s     # the basic blocks of each function
 ```
-cmd/gasm/           CLI entry point (subcommands)
-token/              Lexical token kinds and positions
-lexer/              Hand-written scanner
-ast/                Abstract syntax tree
-parser/             Line-oriented parser
-arch/               Register and instruction tables (generated)
-lint/               Static analysis rules
-format/             Canonical formatter
-lsp/                Language Server Protocol server
-asm/                Standalone assembler, encoder, object emitters
-verify/             JIT execution, differential testing, ABI checks
-debug/              Interactive ptrace debugger (all four architectures)
-_gen/               Instruction table generator
-testdata/           Test fixtures
-docs/               Architecture, development, CLI reference
-```
+
+`gasm verify --ground-truth` is the differential check that ties the two
+together: it compares gasm's bytes with `go tool asm`'s, with the relocation
+sites masked, so an encoding drift shows up as a byte difference rather than a
+crash later.
+
+## Continuous integration
+
+Workflows live in `.gitea/workflows/` and run on the project's own runners:
+Test on a push or pull request to `development`, race dispatched by hand, and
+the release on a `v*` tag.  They are written by hand rather than through
+`just`, but they enforce the same set of gates minus the race detector, which
+the shared runner cannot afford on a push; a green `just gates` locally is
+therefore the fastest way to a green pipeline.
+
+## Releases
+
+Releases are cut by merging `development` into `main` and tagging `vX.Y.Z`,
+which triggers the release workflow: it builds the portable Linux targets,
+takes the notes from the matching `CHANGELOG.md` section and uploads the
+assets.  `SECURITY.md` carries the supported-versions table, so that table
+moves with the release; the pipeline refuses a tag the policy does not name.
+
+The version is never injected. `gasm --version` prints what the
+toolchain recorded in the build information: the tag on a tagged
+checkout, a pseudo-version naming the commit below one, `+dirty` on a
+dirty tree, and `(devel)` outside version control. There is no
+`-ldflags "-X"` anywhere and no version constant in the source.
@@ -0,0 +1,77 @@
+.TH GASM-ASM 1 "2026-09-19" "gasm" "User Commands"
+.SH NAME
+gasm-asm \- assemble Plan 9 assembly without the Go toolchain
+.SH SYNOPSIS
+.B gasm asm [\-\-format raw|elf|goobj] [\-I dir] [\-p pkg] [\-GOARCH arch] [\-o out] <file>
+.SH DESCRIPTION
+Assemble FILE without the Go toolchain: every TEXT function is encoded
+to machine code and printed as a hex dump.  Supported architectures:
+amd64 (including VEX/AVX2 and EVEX/AVX-512), arm64 (AArch64 integer,
+FP, conditional select, CRC32 and MOV pseudo), riscv64 (RV64IMAFDC and
+RVC) and loong64 (LoongArch base ISA).
+.PP
+With
+.B \-o
+the output is written to a file instead.  The
+.B \-\-format
+flag selects what is written:
+.B raw
+(the default) concatenates the functions and the data section into one
+self-consistent image;
+.B elf
+emits a relocatable object (.text/.data sections, a symbol table and
+one relocation per static-symbol reference, in the architecture's own
+form: R_X86_64_PC32 on amd64, R_AARCH64_*, R_RISCV_* or R_LARCH_* on the
+others) that links with the
+system toolchain;
+.B goobj
+emits the Go toolchain's own object format, which cmd/link consumes
+directly (it requires
+.BR \-p ,
+the package path, and the installed Go toolchain: the object preamble is
+captured from
+.B go tool asm
+and the format version from
+.BR "go version" ).
+.PP
+.B raw
+and
+.B elf
+need no toolchain at all.
+.PP
+Framed functions receive the stack-split guard and the trailing
+morestack block, byte-identical to the toolchain's output, so split
+functions link too.
+.SH OPTIONS
+.TP
+.B \-\-format \fIraw|elf|goobj\fR
+Output format; the default is raw.
+.TP
+.B \-I \fIdir\fR
+Directory to search for #include files; may be repeated, searched in
+order after the source directory.
+.TP
+.B \-p \fIpkg\fR
+Package path for --format goobj, qualifying the exported symbols.
+.TP
+.B \-GOARCH \fIarch\fR
+Target architecture: amd64, arm64, riscv64 or loong64; overrides the
+file-name suffix, which is how the suffix-less majority of GOROOT's
+files (cpu_x86.s, stub.s, ...) become assemblable.
+.TP
+.B \-o \fIfile\fR
+Write the output to this file instead of a hex dump on stdout.
+.SH EXIT STATUS
+Exits 0 on success, 1 when parsing or assembly fails, and 2 on a usage
+error.
+.SH EXAMPLES
+.nf
+gasm asm \-o hello.bin hello_amd64.s                      raw image
+gasm asm \-\-format elf \-o k.o k_amd64.s                  linkable ELF object
+gasm asm \-\-format goobj \-p pkg/path \-o k.o k_amd64.s   Go object for go build
+gasm asm \-GOARCH amd64 cpu_x86.s                         arch override
+.fi
+.SH SEE ALSO
+.BR gasm (1),
+.BR gasm\-dis (1),
+.BR gasm\-verify (1)
@@ -0,0 +1,58 @@
+.TH GASM-AUDIT-INSTRUCTIONS 1 "2026-09-19" "gasm" "User Commands"
+.SH NAME
+gasm-audit-instructions \- diff the encoder against the Go toolchain, or measure a corpus
+.SH SYNOPSIS
+.B gasm audit\-instructions [\-\-corpus [\fIdir\fR]] [\-I dir] [amd64|arm64|riscv64|loong64]
+.SH DESCRIPTION
+Compare the gasm encoder for the given architecture (default amd64)
+against
+.B go tool asm
+and print the diff: superset encodings (gasm-only, shippable via
+.BR "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
+.B 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
+.B 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 rather than listing them.
+.PP
+With
+.BR \-\-corpus ,
+the audit changes shape: it assembles every
+.I .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 headline number (files that assemble
+for every target architecture), the per-architecture pass rates and the
+most common failure reasons, which drive the encodability backlog by
+frequency rather than by table order.  A run over GOROOT takes under a
+second.
+.SH OPTIONS
+.TP
+.B \-\-corpus [\fIdir\fR]
+Assemble a corpus of .s files and report pass rates and failure
+reasons.
+.TP
+.B \-I \fIdir\fR
+Directory to search for #include files; may be repeated, searched in
+order after the source directory.  A corpus run whose files include
+toolchain headers (such as GOROOT/pkg/include) needs it, the same -I a
+toolchain comparison takes.
+.SH EXIT STATUS
+The mnemonic-diff mode reports through its output and exits 0; a failed
+probe or an unknown architecture exits non-zero.
+.SH EXAMPLES
+.nf
+gasm audit\-instructions amd64
+gasm audit\-instructions \-\-corpus
+gasm audit\-instructions \-\-corpus "$(go env GOROOT)/src/crypto"
+.fi
+.SH SEE ALSO
+.BR gasm (1),
+.BR gasm\-asm (1)
@@ -0,0 +1,117 @@
+.TH GASM-DEBUG 1 "2026-09-19" "gasm" "User Commands"
+.SH NAME
+gasm-debug \- interactive source-level debugger for JIT-assembled functions
+.SH SYNOPSIS
+.B gasm debug <file.s> \-\-func <name>
+.SH DESCRIPTION
+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.
+.PP
+With
+.B \-\-script
+the REPL commands run from a file and the session ends: the headless
+mode CI and scripts use.
+.B \-\-cover
+runs to completion with a breakpoint on every instruction and reports
+which executed and how often, the label-level coverage view.
+.SH REPL COMMANDS
+.TP
+.B break \fIlabel|addr|line\fR [\fBif \fIreg op val|reg|*addr\fR], b
+Set a breakpoint at a label, an address or a source line number, optionally
+conditional on a register comparison: against a constant, against another
+register, or against the 8-byte word at
+.BR *addr .
+.TP
+.B delete \fIlabel|addr\fR, d
+Remove a breakpoint.
+.TP
+.B info break, info breakpoints, info b
+List all breakpoints.
+.TP
+.BR step " [" n ], " s
+Single-step n instructions; the default is 1.
+.TP
+.BR next ", " n
+Step over a CALL.
+.TP
+.BR finish ", " fin
+Run until the function returns.
+.TP
+.BR continue ", " c
+Run until a breakpoint, watchpoint or exit.
+.TP
+.BR disas " [" n ], " u
+Disassemble n instructions at PC.
+.TP
+.B regs
+Print general-purpose and vector registers.
+.TP
+.B where
+Show the source line and nearest label at PC.
+.TP
+.B stack
+Show the stack near RSP (return address and ABI0 args).
+.TP
+.BR bt ", " backtrace
+Backtrace: current frame plus return address.
+.TP
+.B x [\fIaddr\fR] [\fIlen\fR]
+Hex-dump memory; the defaults are the current PC and 64 bytes.
+.TP
+.B w \fIaddr val...\fR
+Write bytes to memory.
+.TP
+.B set \fIreg value\fR
+Set a register.
+.TP
+.B watch \fIaddr\fR [\fBr|w\fR] [\fIsize\fR]
+Set a hardware watchpoint; writes are watched by default.
+.TP
+.B unwatch [\fIslot\fR]
+Clear one watchpoint, or all without an argument.
+.TP
+.BR labels ", " l
+List function labels and offsets.
+.TP
+.BR help ", " h ", " ?
+Show command help.
+.TP
+.BR quit ", " q
+Kill the debuggee and exit.
+.SH OPTIONS
+.TP
+.B \-args \fIfile\fR
+File containing the ABI0 argument block.
+.TP
+.B \-buf \fIspec\fR
+Buffer specification: name:size:pattern[,name:size:pattern...] where
+pattern is zero, ones, seq, or hex.
+.TP
+.B \-cover
+Run to completion with a breakpoint on every instruction and report
+which executed and how often.
+.TP
+.B \-func \fIname\fR
+Function to debug.
+.TP
+.B \-script \fIfile\fR
+Run REPL commands from a file (one per line) and exit; - reads stdin.
+.TP
+.B \-timeout \fIduration\fR
+Kill the debuggee after this duration (e.g. 30s); for headless --script
+runs; a timeout exits 3.
+.SH EXIT STATUS
+Exits 0 when the scripted session completes, 1 when the debuggee crashes
+or a check fails, and 3 when
+.B \-\-timeout
+kills the debuggee; the debugger is Linux-only.
+.SH EXAMPLES
+.nf
+gasm debug \-\-func name k.s
+gasm debug \-\-func name \-\-script cmds.txt \-\-timeout 30s k.s
+gasm debug \-\-func name \-\-cover k.s
+.fi
+.SH SEE ALSO
+.BR gasm (1),
+.BR gasm\-verify (1)
@@ -0,0 +1,39 @@
+.TH GASM-DIFF 1 "2026-09-19" "gasm" "User Commands"
+.SH NAME
+gasm-diff \- compare the machine code of two assembly files
+.SH SYNOPSIS
+.B gasm diff [\-GOARCH arch] [\-I dir] <file1.s> <file2.s>
+.SH DESCRIPTION
+Compare the machine code produced by assembling two files.  Shows which
+functions differ and the byte-level differences.  Useful for verifying
+that two implementations produce identical code, or for tracking
+encoding changes between Go assembler versions.
+.PP
+Functions are paired by exact name unless
+.B \-\-map
+says otherwise, so
+.B \-\-map wideCopyAVX2=wideCopyAVX512
+pairs two variants regardless of suffix.
+.SH OPTIONS
+.TP
+.B \-GOARCH \fIarch\fR
+Target architecture for both files: amd64, arm64, riscv64 or loong64;
+overrides the file-name suffixes.
+.TP
+.B \-I \fIdir\fR
+Directory to search for #include files; may be repeated, searched in
+order after the source directory.
+.TP
+.B \-\-map \fIspec\fR
+Comma-separated old=new pairs to match functions with different names.
+.SH EXIT STATUS
+Exits 0 when every paired function is identical and 1 when anything
+differs; a usage error exits 2.
+.SH EXAMPLES
+.nf
+gasm diff hello_amd64.s hello_amd64.s
+gasm diff \-\-map wideCopyAVX2=wideCopyAVX512 avx2_amd64.s avx512_amd64.s
+.fi
+.SH SEE ALSO
+.BR gasm (1),
+.BR gasm\-asm (1)
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
petrbalvin	97dfaa7526	docs: changelog for macro expansion and the corrected corpus audit Test / test (push) Successful in 2m14s Details Assisted-by: GLM 5.3 Flash	2026-09-20 14:25:47 +02:00
petrbalvin	66aa4dbc8b	test(verify): register the campaign kernels in the ground-truth suites Assisted-by: GLM 5.3 Flash	2026-09-20 14:25:47 +02:00
petrbalvin	dce5d31462	feat(amd64): LOCK and REP prefixes, literal data pseudo-ops and ADJSP Assisted-by: GLM 5.3 Flash	2026-09-20 14:25:47 +02:00
petrbalvin	9dc3987e02	feat(riscv64,loong64): PCALIGN, branch relaxation and operand shapes Assisted-by: GLM 5.3 Flash	2026-09-20 14:25:47 +02:00
petrbalvin	9b238a525a	feat(arm64): wide immediates, SIMD compare and system operand forms Assisted-by: GLM 5.3 Flash	2026-09-20 14:25:47 +02:00
petrbalvin	ad82aac663	feat(parser): macro expansion, conditionals and include splicing with -I Assisted-by: GLM 5.3 Flash	2026-09-20 14:25:47 +02:00
petrbalvin	0629f5e2df	feat(arm64): assemble PCALIGN padding and BYTE literal bytes Test / test (push) Successful in 2m16s Details Assisted-by: GLM 5.3 Flash	2026-09-20 11:49:05 +02:00
petrbalvin	ecb203dcf5	fix(lexer): treat trailing CR as line end so comment text is idempotent Test / test (push) Successful in 2m13s Details Assisted-by: GLM 5.3 Flash	2026-09-20 11:40:39 +02:00
petrbalvin	6c672567f3	feat(amd64): assemble the double-shift and static-SB operand shapes Assisted-by: GLM 5.3 Flash	2026-09-20 11:40:39 +02:00
petrbalvin	cc6e416c59	fix(lint): exempt shift counts, SETcc and ABIInternal from false positives Assisted-by: GLM 5.3 Flash	2026-09-20 11:40:39 +02:00
petrbalvin	c66a47973a	fix(format): preserve square brackets in SIMD operands Test / test (push) Successful in 2m15s Details Assisted-by: GLM 5.3 Flash	2026-09-20 09:58:25 +02:00
petrbalvin	9629897202	docs: changelog and readme for the instruction wave and the honest corpus rate Test / test (push) Successful in 2m17s Details Assisted-by: GLM 5.3 Flash	2026-09-20 06:45:03 +02:00
petrbalvin	5399a8a724	feat(audit): probe the new operand shapes and measure attemptable files Assisted-by: GLM 5.3 Flash	2026-09-20 06:45:03 +02:00
petrbalvin	de5d9f358e	feat(riscv64,loong64): encode AMO atomics, vector slices and bit ops Assisted-by: GLM 5.3 Flash	2026-09-20 06:44:51 +02:00
petrbalvin	ca3fdce0e0	feat(arm64): encode pairs, atomics, crypto, system and NEON slices Assisted-by: GLM 5.3 Flash	2026-09-20 06:44:51 +02:00
petrbalvin	fc2d92eabd	feat(amd64): encode the GOROOT instruction families Assisted-by: GLM 5.3 Flash	2026-09-20 06:44:51 +02:00
petrbalvin	39d2e80145	ci(release): refuse empty assets and verify what the release serves Test / test (push) Successful in 2m10s Details Assisted-by: DeepSeek V4.1 Flash	2026-09-20 02:01:36 +02:00
petrbalvin	9f4f949c1f	chore: prepare release v0.34.0 Test / test (push) Successful in 2m11s Details Release / gates (push) Successful in 2m11s Details Release / build (amd64, linux) (push) Successful in 1m13s Details Release / build (arm64, linux) (push) Successful in 1m10s Details Release / build (loong64, linux) (push) Successful in 1m12s Details Release / build (riscv64, linux) (push) Successful in 1m33s Details Release / release (push) Successful in 58s Details	2026-09-20 01:44:23 +02:00
petrbalvin	f0d5238c47	docs: state the validation status and correct claims the material contradicts Assisted-by: DeepSeek V4.1 Flash	2026-09-20 01:40:51 +02:00
petrbalvin	2931bbd6b2	ci(release): refuse a tag the security policy does not name Assisted-by: DeepSeek V4.1 Flash	2026-09-20 01:40:51 +02:00
petrbalvin	63562a503a	test(justfile): run the CLI and debugger tests outside the coverage set Assisted-by: DeepSeek V4.1 Flash	2026-09-20 01:40:51 +02:00
petrbalvin	e836d6150d	docs: changelog entry for the loong64 JIT enablement Test / test (push) Successful in 2m7s Details Assisted-by: GLM 5.3	2026-09-20 00:57:02 +02:00
petrbalvin	8a51b060da	feat(cmd): enable loong64 JIT execution, all trampolines qemu-validated Assisted-by: GLM 5.3	2026-09-20 00:57:02 +02:00
petrbalvin	d3d47db727	test(verify): seed the arm64 ABI kernel arguments Assisted-by: GLM 5.3	2026-09-20 00:57:02 +02:00
petrbalvin	0758556b7d	docs: changelog entries for the parity round and corpus number Assisted-by: GLM 5.3	2026-09-20 00:38:24 +02:00
petrbalvin	ddb8440340	fix(cmd): padding-aware ground-truth comparison Assisted-by: GLM 5.3	2026-09-20 00:38:24 +02:00
petrbalvin	f15ff66fb1	fix(riscv64): accept the g spelling of the goroutine register Assisted-by: GLM 5.3	2026-09-20 00:38:24 +02:00
petrbalvin	187e4856d3	feat(amd64): encode the mixed-width extend family and PMOVMSKB Assisted-by: GLM 5.3	2026-09-20 00:38:24 +02:00
petrbalvin	d315a998ce	fix(arm64): store-exclusive operand order and large-frame parity Assisted-by: GLM 5.3	2026-09-20 00:38:24 +02:00
petrbalvin	a6f3828c02	docs: changelog entries for the review fixes Test / test (push) Successful in 2m4s Details Assisted-by: GLM 5.3	2026-09-19 23:49:27 +02:00
petrbalvin	e3b35bb817	style(testdata): canonical gasm formatting for the verify kernels Assisted-by: GLM 5.3	2026-09-19 23:49:27 +02:00
petrbalvin	eb0a89e58d	ci(release): state the version contract inline Assisted-by: GLM 5.3	2026-09-19 23:49:27 +02:00
petrbalvin	dd32d9e66e	chore(justfile): one-line install-man comment and long flag forms Assisted-by: GLM 5.3	2026-09-19 23:49:27 +02:00
petrbalvin	3a73acb20a	docs: drop process labels and refresh the architecture and manual pages Assisted-by: GLM 5.3	2026-09-19 23:49:27 +02:00
petrbalvin	7604a9443f	fix(cmd): usage exit codes, asm output file and cross-arch ground truth Assisted-by: GLM 5.3	2026-09-19 23:49:19 +02:00
petrbalvin	b3908fc43d	fix(lsp): parse-error survival, symbol ranges and UTF-16 positions Assisted-by: GLM 5.3	2026-09-19 23:49:19 +02:00
petrbalvin	eb8b0cd316	fix(lint): trailing-label CFG guard and the goroutine alias Assisted-by: GLM 5.3	2026-09-19 23:49:19 +02:00
petrbalvin	a8bfd54ed2	fix(debug): hardware watchpoints, signal stops and breakpoint restore Assisted-by: GLM 5.3	2026-09-19 23:49:19 +02:00
petrbalvin	375182ef1f	fix(verify): arm64 stack save, adaptive canary and host gating Assisted-by: GLM 5.3	2026-09-19 23:49:19 +02:00
petrbalvin	87b1081c53	fix(goobj): external package and symbol indices and arm64 pair relocations Assisted-by: GLM 5.3	2026-09-19 23:49:13 +02:00
petrbalvin	f3c8510a58	fix(elf): relocation records, DWARF tables and per-architecture frame data Assisted-by: GLM 5.3	2026-09-19 23:49:13 +02:00
petrbalvin	ebdf14939f	fix(loong64): FP immediates through R30 and unsigned branch forms Assisted-by: GLM 5.3	2026-09-19 23:49:13 +02:00
petrbalvin	79a2c16bac	fix(riscv64): compressed store offsets, FENCE and branch range checks Assisted-by: GLM 5.3	2026-09-19 23:49:13 +02:00
petrbalvin	401386956c	fix(arm64): encode shifts, divides and multiplies and align sizes with emission Assisted-by: GLM 5.3	2026-09-19 23:49:07 +02:00
petrbalvin	4258131a3a	fix(amd64): correct guard displacements, frameless FP offsets and immediate ranges Assisted-by: GLM 5.3	2026-09-19 23:49:07 +02:00
petrbalvin	94e09e8070	fix(format): preserve flag separators and normalise CRLF input Assisted-by: GLM 5.3	2026-09-19 23:48:47 +02:00
petrbalvin	7aefe6a42d	fix(parser): parse ABI markers and keep TEXT decls usable on errors Assisted-by: GLM 5.3	2026-09-19 23:48:47 +02:00
petrbalvin	ac1c05c793	fix(lexer): tokenise the flag separator and handle NUL and invalid UTF-8 Assisted-by: GLM 5.3	2026-09-19 23:48:47 +02:00
petrbalvin	93c47a312a	feat(docs): man pages for gasm and every command, guarded against CLI drift Test / test (push) Successful in 2m4s Details Assisted-by: GLM 5.3 Flash	2026-09-19 21:18:43 +02:00
petrbalvin	708d0a0a5e	docs: trim the changelog entries to user-visible deltas Assisted-by: GLM 5.3 Flash	2026-09-19 20:54:56 +02:00
petrbalvin	3c8f7cb411	test(format): pin the fuzz-found crashers as regression seeds Assisted-by: GLM 5.3 Flash	2026-09-19 20:48:51 +02:00
petrbalvin	7c5b7a1419	docs: add the changelog entries and the corpus number to the readme Assisted-by: GLM 5.3 Flash	2026-09-19 20:48:51 +02:00
petrbalvin	bc3f448738	feat(format): fuzz targets for the parser and formatter Assisted-by: GLM 5.3 Flash	2026-09-19 20:41:43 +02:00
petrbalvin	f37f183577	feat(riscv64): GOROOT instruction shapes, DATA order and offset expressions Assisted-by: GLM 5.3 Flash	2026-09-19 19:58:43 +02:00
petrbalvin	1e77e58250	feat(gasm): audit a .s corpus with audit-instructions --corpus Assisted-by: GLM 5.3 Flash	2026-09-19 19:27:30 +02:00
petrbalvin	1d0969ed64	feat(gasm): select the asm and diff architecture with -GOARCH Assisted-by: GLM 5.3 Flash	2026-09-19 19:20:47 +02:00
petrbalvin	23c001be51	feat(asm): encode indirect JMP and CALL on all four architectures Assisted-by: GLM 5.3 Flash	2026-09-19 19:17:07 +02:00
petrbalvin	96e81cc98d	docs: add the Plan 9 assembly case and real-use note to the README Test / test (push) Successful in 2m6s Details	2026-09-19 18:06:18 +02:00
petrbalvin	c834d98210	docs: bring the document set into the standard shape Test / test (push) Successful in 2m28s Details Assisted-by: GLM 5.3 Flash	2026-09-17 20:33:18 +02:00
petrbalvin	03d6d4da54	style: put the repository assembly in gasm fmt canonical form Assisted-by: GLM 5.3 Flash	2026-09-17 20:33:18 +02:00
petrbalvin	0b42ce7952	style: use one spelling for colour across the CLI Assisted-by: GLM 5.3 Flash	2026-09-17 20:33:18 +02:00
petrbalvin	288a64ccd2	ci: align the pipelines with the hand-written templates Assisted-by: GLM 5.3 Flash	2026-09-17 20:33:18 +02:00
petrbalvin	5fddfa704b	build: declare the exact toolchain and the canonical recipes Assisted-by: GLM 5.3 Flash	2026-09-17 20:33:18 +02:00
petrbalvin	a2bb5eeb4e	chore: drop the stale comment from the ignore list Assisted-by: GLM 5.3 Flash	2026-09-17 20:33:14 +02:00
petrbalvin	48449b7a7f	build: declare the go1.27.1 toolchain Assisted-by: GLM 5.3 Flash	2026-09-16 23:12:31 +02:00
petrbalvin	3de043c494	docs: add SECURITY.md and record the round in the CHANGELOG Assisted-by: GLM 5.3 Flash	2026-09-16 23:12:31 +02:00
petrbalvin	0078f7be5c	style: purge em dashes from the produced text Assisted-by: GLM 5.3 Flash	2026-09-16 23:12:31 +02:00
petrbalvin	6a7317d141	chore: trim the ignore list to the convention Assisted-by: GLM 5.3 Flash	2026-09-16 22:53:01 +02:00
petrbalvin	d08523caa5	docs: move the recipe and version descriptions with the behaviour Assisted-by: GLM 5.3 Flash	2026-09-16 22:53:01 +02:00
petrbalvin	20e4b8d9c4	ci: align the pipelines with the hand-written templates Assisted-by: GLM 5.3 Flash	2026-09-16 22:53:01 +02:00
petrbalvin	61f4247cef	refactor(gasm): report the toolchain-recorded version Assisted-by: GLM 5.3 Flash	2026-09-16 22:53:01 +02:00
petrbalvin	049872ddff	build: restore the canonical justfile recipe set Assisted-by: GLM 5.3 Flash	2026-09-16 22:53:01 +02:00
petrbalvin	3669f64ff6	build: install the gasm binary into the user-local bin directory Test / vet (push) Successful in 46s Details Test / test (push) Successful in 2m44s Details Test / build (push) Successful in 42s Details	2026-09-14 23:41:21 +02:00