feat(arch): add the extended-instruction layer with SVE arithmetic

Assisted-by: GLM 5.3 Flash

.gitea/workflows/freebsd.yml

@@ -0,0 +1,42 @@
+# FreeBSD compile gates. Dispatched by hand, never on a push.
+#
+# The debugger's ptrace surface and the JIT substrate are the two
+# FreeBSD-portable layers the tree carries; the forge has no FreeBSD runner,
+# so they can only be compile-gated, and three foreign-GOOS builds of the
+# whole module are minutes of one-core work the push pipeline's budget cannot
+# carry. The push pipeline stays fast and light; this workflow is the
+# deliberate run, before a release or after touching the ported layers.
+# Running the ptrace suite itself needs real FreeBSD hardware.
+#
+# A dispatched workflow takes no concurrency block: it is one deliberate run.
+name: FreeBSD build
+
+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:
+  build:
+    runs-on: fedora
+    timeout-minutes: 10
+    steps:
+      - uses: actions/checkout@v7
+
+      - uses: actions/setup-go@v6
+        with:
+          # The module is the source of truth for the version, so it cannot drift.
+          go-version-file: go.mod
+          cache: true
+
+      - name: FreeBSD build (amd64)
+        run: GOOS=freebsd GOARCH=amd64 go build ./...
+
+      - name: FreeBSD build (arm64)
+        run: GOOS=freebsd GOARCH=arm64 go build ./...
+
+      - name: FreeBSD build (riscv64)
+        run: GOOS=freebsd GOARCH=riscv64 go build ./...

.gitea/workflows/release.yml

@@ -342,7 +342,10 @@ jobs:
               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},
+                # The @ must not sit inside a qq{} string: there it starts an
+                # array interpolation and the upload body collapses to empty,
+                # which Gitea stores as a 201-created zero-byte attachment.
+                q{-X}, q{POST}, q{--data-binary}, q{@} . $path,
                 qq{$ENV{GITEA_SERVER_URL}/api/v1/repos/$ENV{GITEA_REPOSITORY}/releases/$id/assets?name=$name});
               open(my $curl, q{-|}, @cmd) or die qq{curl: $!};
               my $code = <$curl>;

.gitea/workflows/test.yml

@@ -6,6 +6,11 @@
 # everything runs in one job. Extra jobs would duplicate the checkout, the Go setup and
 # the dependency download three times without buying any parallelism.
 #
+# The budget is part of the contract: a push run is fast and light, about two minutes,
+# and nothing that cannot run natively on the runner belongs here. The FreeBSD compile
+# gates live in freebsd.yml behind workflow_dispatch for that reason; the GOOBJ link
+# parity campaign is an opt-in local verification (just link-parity).
+#
 # 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
@@ -46,10 +51,9 @@ jobs:
           go-version-file: go.mod
           cache: true
 
-      - 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
+      # No Install Perl step: the fedora image carries perl (verified by run
+      # 76: the install degraded into a package upgrade costing ~50 s), and a
+      # dnf on the push path is network work the budget does not need.
 
       # 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.
@@ -82,7 +86,11 @@ jobs:
         # 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/...
+        # -short skips the deliberate-run categories inside the suites (the live
+        # ptrace sessions above all): they need the machine to themselves and a
+        # starved single-core runner turns each into a timeout the budget cannot
+        # carry. The local `just test` gate runs everything, in full.
+        run: go test -short -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
@@ -90,12 +98,12 @@ jobs:
         # 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/...
+        # -short skips the debugger's live ptrace sessions, the deliberate-run
+        # category the runner cannot starve-proof. The live go-tool-asm oracle
+        # comparison (TestGroundTruth in ./verify/...) runs inside the coverage
+        # sweep above; it is not re-run as its own step, because every second on
+        # this box is budget.
+        run: go test -short -count=1 -timeout 10m ./cmd/... ./debug/...
 
       - name: Coverage floor
         run: |

CHANGELOG.md

@@ -1,6 +1,6 @@
 # Changelog
 
-All notable changes to gasm-devkit are documented here.
+All notable changes to gasm-sdk are documented here.
 
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
@@ -9,6 +9,178 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ### Added
 
+- **The FreeBSD port of the debugger.**  `gasm debug` runs on FreeBSD on
+  amd64, arm64 and riscv64 with the same interactive surface as on Linux:
+  breakpoints, hardware watchpoints (x86 debug registers, the arm64 debug
+  register file), single-stepping, register and memory access, all behind
+  the kernel's own ptrace requests, with tracee memory through `PT_IO` and
+  stop reports through `PT_LWPINFO`.  The JIT substrate maps executable
+  memory through `golang.org/x/sys/unix`, so `verify` builds on FreeBSD
+  too.  The pipeline compile-gates all three architectures; live
+  validation awaits a FreeBSD machine.
+- **Workspace-wide navigation in the language server.**  `gasm lsp` indexes
+  the `.s` files under the workspace root beyond the documents the editor
+  has open, so go-to-definition, find references and workspace symbol search
+  reach files that were never opened.  An open buffer always shadows its
+  disk copy, and watched-file events together with a per-query freshness
+  check keep the index current.
+- **Quick fixes for the textflag include and the argument area.**  The
+  `missing-textflag-include` warning offers to add the include after the
+  last one in the file, and the `abi-argsize` warning offers to set the
+  TEXT argument area to the size the `// func` signature implies, computed
+  by the new `lint.ExpectedArgSize`.
+- **Eleven new lint rules over the directives, the data section and the
+  sharpest addressing edges.**  `missing-argsize` flags a TEXT that
+  declares no argument area its `// func` signature implies;
+  `noframe-frame-size` a NOFRAME with a positive frame;
+  `unnamed-fp-reference` a nameless `0(FP)`, which both assemblers
+  reject; `hardware-sp-addressing` a negative offset off the hardware SP
+  rather than the virtual frame; `vex-sse-mixing` a kernel that mixes VEX
+  and legacy SSE spellings and pays the transition penalty;
+  `unnamed-result` a `ret+N(FP)` the signature names; `data-width`,
+  `data-value-overflow`, `data-string-width`, `data-without-globl` and
+  `data-exceeds-globl` police the DATA width against its value type and
+  the GLOBL size behind it.  `missing-ret` now also flags a function
+  whose tail can fall off its end even though a RET sits somewhere in the
+  body, and `invalid-textflag` also reports a flag misplaced between TEXT
+  and GLOBL.
+- **Hover documentation and wider completions in the language server.**
+  Hovering a directive or pseudo-operation (TEXT, DATA, GLOBL, PCALIGN,
+  FUNCDATA, PCDATA, the BYTE family) shows its grammar and rules in
+  preference to the empty instruction-table entry, and completion offers
+  those names beside the instruction set.  The `missing-argsize` warning
+  carries a quick fix that declares the argument area the signature
+  implies (`$0` becomes `$0-16`).
+- **The GOOBJ link parity gate.**  A regression test assembles a kernel
+  per architecture through `gasm asm --format goobj`, substitutes the
+  object into a real `go build`'s package archive, proves the archive
+  carries it byte for byte and re-links with cmd/link on amd64, arm64,
+  riscv64 and loong64.  The binaries run, natively on amd64 and under
+  qemu-user on arm64 and riscv64, and their output must match the
+  toolchain-built baseline; loong64 is link-only.  The pipeline installs
+  qemu-user and runs the gate on every push.
+- **The amd64 and loong64 encoders close four more corpus files.**  The
+  whole-tree measure moves to 272 of 322 (84.5 %): amd64 gains the
+  one-operand IMUL, the SSE compare family (CMPPD, CMPPS, CMPSS), RETFL,
+  the LOOP family, the MMX register bank with its bank-crossing moves,
+  MOVNTDQ, the CR and DR register moves, PUSH and POP of FS and GS, the
+  `(TLS)` pseudo-base, the wait and cache controls (CLWB, CLDEMOTE,
+  TPAUSE, UMONITOR, UMWAIT, RDPID, ENDBR64), indirect branches with the
+  star spelling (`JMP *(R12)(R13*4)`), `RET sym(SB)` as the tail jump,
+  the colon shift spelling (`SHLL CX, R11:AX`) and the EVEX and VEX forms
+  of the rounds, AES key assist, string compares, extracts, blends and
+  permutes; loong64 gains the acquire and release pair (LLACQ, SCREL,
+  with the vector widths), the VMOVQ and XVMOVQ lane forms and the BYTE
+  literal-data escape hatch the other architectures already take.
+
+### Changed
+
+- **The corpus audit assembles like the build.**  A file's `//go:build`
+  constraint decides which target architectures attempt it: cpu_x86.s is
+  an x86 build alone, and the msan and goexperiment.runtimesecret trees
+  are compiled by no supported build, so they leave the measured set
+  instead of failing it.  The headline now reads "assemble for every
+  applicable target": every real-code GOROOT assembly file, the tree
+  without testdata, assembles for all four architectures (250 of 250,
+  100 %); over the whole tree including testdata the measure is 272 of
+  322 (84.5 %).
+- **The module moves to `sourcedock.dev/petrbalvin/gasm-sdk`.**  The
+  repository and the module rename together with the product, now the
+  GAsm Software Development Kit.  Fresh installs become
+  `go install sourcedock.dev/petrbalvin/gasm-sdk/cmd/gasm@latest`, and
+  installs pinned to the old `gasm-sdk` path stop resolving once the
+  repository takes the new name: reinstall from the new path.  The
+  binary stays `gasm`.
+
+### Fixed
+
+- **Rename edits land in their own documents.**  A rename collected the
+  ranges of every reference across the open documents but applied them all
+  to the document that started it, so renaming a symbol used in a second
+  file moved that file's text into the first.  Each edit now applies to the
+  document it was collected in.
+- **Negative numeric PC-relative jumps.**  `JMP -3(PC)`, the shape the
+  runtime's exit loops write (sys_linux_amd64.s, sys_netbsd_amd64.s),
+  resolved to nothing: only the forward forms counted.  A negative count
+  now walks the same instruction statements backwards, labels excluded,
+  byte-identical with the toolchain.
+- **The arm64 move-wide family reads its immediate as an unsigned
+  pattern.**  `MOVK $(40000<<48)` folds to a negative int64 and was
+  rejected; the toolchain picks the 16-bit lane from the 64-bit bit
+  pattern, so the encoder now does the same, and a zero immediate is
+  rejected where the toolchain rejects it.
+- **NOPTR data emits its own symbol kind.**  A GLOBL with NOPTR was
+  emitted as plain SDATA, the kind the linker holds to its Go type
+  information requirement, so every gasm object carrying runtime-shaped
+  data (`GLOBL ·x(SB), NOPTR, ...`) died in cmd/link with "missing Go
+  type information".  NOPTR data is now SNOPTRDATA, the toolchain's kind
+  for pointer-free globals, and RODATA still wins where both flags
+  appear, exactly as the toolchain chooses.  The three end-to-end link
+  tests that should have caught this substituted the gasm object after
+  the package archive was already packed, so they passed vacuously; they
+  now substitute inside the archive, prove the substitution byte for
+  byte and re-link.
+- **Latent encoder divergences against the toolchain, found by the
+  whole-file differential harness.**  On loong64, `MOVx $off(reg)` lost
+  its base register, SCQ swapped its operand fields, the vector lane
+  inserts did not scale offsets by the element width and two immediate
+  opcodes were mistyped; on amd64, NOP with operands encoded 0x90 where
+  the toolchain emits nothing at all, the VEX gather length bit ignored
+  the VSIB index width, four permute and extract families were pushed
+  into EVEX where the toolchain stays VEX, and a reference to a static
+  symbol no GLOBL defines failed where the toolchain defers it to the
+  linker as an external relocation.
+- **Seven front-end defects found by fuzzing.**  The formatter swallowed
+  the statement after a leading block comment (`/* head */ MOVQ AX, BX`
+  formatted to the comment alone), dropped trailing tokens after an
+  `#include` header, and trimmed the trailing whitespace inside a string
+  literal; the parser peeled only one label of a stacked pair (`a: b:`
+  parsed differently than it formatted); deeply nested parentheses in an
+  immediate overflowed the stack through the constant folder, which now
+  stops at a bounded depth and falls back to the ordinary operand paths;
+  macro expansion is linear in the invocation count instead of
+  quadratic; and amplifying macros (the billion-laughs shape) stop at a
+  work budget sized by the line and the macro table, reported as a
+  diagnostic instead of running for hours.
+- **The debugger handles the edges its tests now reach.**  Memory reads
+  and writes cover exactly the requested bytes, so a request ending in
+  the last page of a mapping no longer fails on the unmapped page behind
+  it; disassembly shrinks its instruction window at a mapping's end
+  instead of failing; a debuggee that dies before signalling readiness
+  writes a failure notice the launcher reads, so launch fails fast with
+  the reason instead of after the whole poll (and the poll no longer
+  waits on the child, which could consume the SIGSTOP park and hang the
+  launch); amd64 watchpoints acknowledge the sticky DR6 hit bits and
+  clear the address register on release; the breakpoint listing is
+  ordered by address so its numbering is stable; the REPL rejects bad
+  counts, sizes and watchpoint types instead of silently guessing,
+  reports stops on signals during stepping, and a breakpoint-class trap
+  that matches no breakpoint and leaves the PC in place surfaces instead
+  of spinning the continue loop and the coverage run forever.
+
+## [0.35.0] - 2026-09-22
+
+### Added
+
+- **The go_asm.h generator.**  `gasm asm` generates the package's go_asm.h
+  itself when an assembly file includes it: the Go files beside the source
+  are type-checked for the target architecture and the constants and field
+  offsets become assembler defines, so package-context files assemble with
+  no compiler and no `go build` in the loop.  `-GOOS` selects the
+  type-checking GOOS for GOOS-specific files, and the corpus audit derives
+  the GOOS from the file name.
+- **ELF data relocations on arm64, riscv64 and loong64.**  `gasm asm
+  --format elf` emits `.rela.data` for symbol-valued DATA initialisers on
+  every architecture (amd64 carried them already), so standalone ELF
+  objects link on all four targets.
+- **Corpus failure listing.**  `gasm audit-instructions --corpus --list`
+  prints every failing file with its failure reason, per architecture,
+  instead of one representative file per reason.
+- **DATA with symbol values and relaxed symbol spellings.**  DATA
+  initialisers accept `$symbol(SB)` values, laid down as an absolute
+  relocation at the data field (GOOBJ on all four architectures and ELF
+  on all four as of this release), and U+2215 is accepted inside symbol
+  package paths.
 - **Macro expansion and include splicing.**  `gasm asm`, `gasm diff` and
   `gasm audit-instructions` now preprocess assembly the way the
   toolchain does: object and parameterised `#define` macros expand at
@@ -19,7 +191,8 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   (`$(32-7)`, `$~63`, `(index*4)(base)`) fold at parse.  Expansion
   happens only on the assembly path: `gasm lint`, `gasm fmt` and the
   language server keep reading the raw file.
-- **The GOROOT instruction wave, part 1.**  The encoder now covers the
+- **Encoder coverage: the instruction families GOROOT's real code
+  uses.**  The encoder now covers the
   instruction families GOROOT's real code uses that gasm lacked,
   byte-verified against `go tool asm`: on amd64 the carry ALU, the
   atomics (CMPXCHG, XADD, XCHG), AES-NI, SHA-1/256, PCLMULQDQ, CRC32,
@@ -35,7 +208,8 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   Also fixed on the way: arm64 `CASD`/`CASW` lacked an opcode bit, and
   riscv64 `VSETVLI` with an immediate length now canonicalises to
   `vsetivli` as the toolchain does.
-- **The GOROOT instruction wave, part 2.**  The encoder gains the
+- **Encoder coverage: quad-register AVX-512 and floating-point
+  immediates.**  The encoder gains the
   quad-register AVX-512 families (4FMAPS, 4FNMADD, 4VNNIW, VP4DPWSSD,
   VP4DPWSSDS) with the register list riding the inverted V'VVVV field,
   floating-point immediates on the SSE scalar moves and arithmetic
@@ -47,21 +221,77 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   ranges, index-only VSIB memory operands and bare trailing immediates;
   macro substitution reaches parameters used with element suffixes
   (`A.S4`), and `;` separates statements in plain files.
-- **`gasm asm -GOOS`.**  The go_asm.h generator type-checks per target
-  GOOS, so the darwin-only and windows-only runtime files assemble with
-  their own defines; the corpus audit derives the GOOS from the file
-  name.  DATA initialisers accept `$symbol(SB)` values (an absolute
-  relocation at the data field, GOOBJ on all four architectures and ELF
-  on amd64, arm64, riscv64 and loong64), and U+2215 is accepted inside
-  symbol package paths.
-- **The corpus audit measures honestly.**  Files named for Go ports gasm
-  does not target (arm, 386, s390x, ...) are no longer attempted for the
-  four supported architectures (no supported build compiles them), and
-  the headline rate is reported over attemptable files: 136 of 433 on
-  the full corpus (31.4 %), 135 of 383 on real code (35.2 %), from the
-  127 that the previous release measured.  The probe battery that
-  decides encodability gained the operand shapes the new families use.
--
+- **Per-architecture reference pages.**  [docs/asm/](docs/asm/README.md)
+  gains AMD64, ARM64, RISCV64 and LOONG64: the register files and the
+  roles the ABI fixes, addressing, operand order with every special form,
+  constants and materialisation, alignment, fences and the relocations
+  each target emits.  An instruction inventory appendix per architecture
+  is generated from the toolchain's own tables by `just gen`, and the
+  regenerated tables recognise 147 more mnemonics than the previous
+  release carried (arm64 107, riscv64 31, loong64 9).
+- **The Plan 9 assembly language reference.**  [docs/asm/](docs/asm/README.md)
+  opens the complete language reference with its common core: the lexicon,
+  statement structure and constant expressions, the operand grammar with
+  the pseudo-registers and symbol naming, the directives and the function
+  flag vocabulary, preprocessing with `#define` and `#include`, and the
+  Go-embedded layer (ABI0, prototypes, `go_asm.h`, `funcdata.h` and the
+  runtime contract).  Every claim is verified against `go tool asm` of
+  Go 1.27.1 and gasm's differential tests; the per-architecture pages and
+  generated instruction appendices follow.
+- **GOOBJ format specification.**  [docs/GOOBJ.md](docs/GOOBJ.md)
+  documents the Go object file format in full: both containers, the 96
+  byte header and all 19 blocks, every structure with its byte
+  offsets, symbol kinds and flag bits, all 106 relocation types with
+  the weak variants, aux symbols, the FuncInfo payload, the pc-value
+  table encoding, the content hashes and the builtin table, all
+  verified byte for byte against objects produced by Go 1.27.1's own
+  tools.
+
+### Changed
+
+- **The corpus audit measures like a build.**  Files named for a Go port
+  gasm does not target (arm, 386, s390x, ...) are never attempted, because
+  no supported build compiles them; the GOOS comes from the file name; and
+  each target's go_asm.h is generated on the fly.  The headline is reported
+  over attemptable files: 291 of 353 on the full corpus (82.4 %) assemble
+  for every target architecture and 295 of 303 on real code (97.4 %),
+  against 108 of 627 over all files (17.2 %) that the previous release
+  measured.
+
+### Fixed
+
+- **The operand forms GOROOT writes.**  Numeric PC-relative jumps
+  (`JEQ 2(PC)`, the park loop `JMP 0(PC)`) resolve with the toolchain's
+  own instruction counting and fold jump-to-jump chains exactly as its
+  branch optimiser does; symbol immediates (`MOVQ $sym(SB), AX`)
+  assemble to the toolchain's RIP-relative LEA with an R_PCREL
+  relocation; negated constant expressions in operands (`ADJSP
+  $-(REGS - 8)`, the shape the cgo ABI macros write) fold; the immediate
+  multiply (`IMULQ $1000000000, AX`) encodes with the toolchain's
+  0x69/0x6B selection; the TLS access pair assembles as the toolchain's
+  one-instruction form (the bare `MOVQ TLS, r` load nops out and
+  `off(r)(TLS*1)` folds to the segment-prefixed absolute whose disp32
+  carries the R_TLSLE relocation, per-GOOS); arm64 accepts the
+  bare-register indirect branch (`BL R9` beside `BL (R9)`, both BLR) and
+  the zero-immediate store (`MOVD $0, mem` through the zero register,
+  rejecting non-zero immediates as the toolchain does); `PCALIGN` now
+  aligns on amd64, padding with the toolchain's greedy
+  single-instruction NOPs; the segment-absolute forms (`MOVQ 0x30(GS),
+  AX` and the store direction) and the absolute crash-store
+  (`MOVL $0xf1, 0xf1`) encode; and `gasm asm` predefines the
+  `GOARCH_<arch>` and `GOOS_<goos>` macros the go command passes to
+  `go tool asm`, so GOROOT headers' `#ifdef GOARCH_amd64` platform
+  blocks (`go_tls.h`'s `get_tls` and friends) select as intended.  The
+  GOROOT corpus measure moves to 291 of 353 files assembling for every
+  target architecture (82.4 %), 97.4 % of the real-code corpus, from
+  70.8 % and 82.2 %.
+- **Tool corrections across the pipeline.**  The formatter keeps square
+  brackets in SIMD operands, statement separators and canonical macro
+  bodies; the linter drops false positives on shift counts, SETcc
+  spellings and ABIInternal references; the lexer treats a trailing
+  carriage return as a line end so comment text stays idempotent; and
+  arm64 rejects bare BTI with a diagnostic while accepting the full
+  family.
 
 ## [0.34.0] - 2026-09-20
 

CONTRIBUTING.md

@@ -1,6 +1,6 @@
 # Contributing
 
-Contributions to **gasm-devkit** are governed by the Contributor terms
+Contributions to **gasm-sdk** are governed by the Contributor terms
 below; submitting one means you accept them.
 
 ## Contributor terms
@@ -29,8 +29,8 @@ 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
+git clone https://sourcedock.dev/petrbalvin/gasm-sdk.git
+cd gasm-sdk
 just build
 just gates
 ```
@@ -126,7 +126,7 @@ tag, where it would double the time and the memory a shared runner cannot spare.
 
 ## Reporting bugs
 
-Open an issue at `https://sourcedock.dev/petrbalvin/gasm-devkit/issues` with the
+Open an issue at `https://sourcedock.dev/petrbalvin/gasm-sdk/issues` with the
 version, the operating system and architecture, the exact command, the full output,
 and the expected against the actual behaviour.
 

README.md

@@ -1,6 +1,6 @@
-# Plan 9 assembly tooling, inside and outside Go
+# GAsm: Software Development Kit for Plan 9 Assembly
 
-> **Warning: this is an experiment.** gasm-devkit is under active
+> **Warning: this is an experiment.** gasm-sdk is under active
 > development and is not stable. The version is 0.x.x: commands, flags,
 > output formats and behaviour can change without warning at any time.
 > A 1.0.0 release is light years away. Nothing in this document is a
@@ -13,7 +13,7 @@
 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
+statement. gasm-sdk is the missing toolkit: a single, self-contained
 binary, `gasm`, that serves both purposes.
 
 - **Help develop Plan 9 assembly.** Formatting, linting, disassembly,
@@ -58,7 +58,7 @@ 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
+compliment a language for machine code can earn. gasm-sdk exists
 to give that syntax the tooling it deserves.
 
 ## Features
@@ -81,7 +81,10 @@ to give that syntax the tooling it deserves.
   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.
+  toolchain's, so split functions link too.  The assembler preprocesses
+  like the toolchain (`#define`, `#include` with `-I`, `#ifdef`), generates
+  `go_asm.h` from the package's Go files, and carries `PCALIGN`, the
+  `LOCK`/`REP` prefixes and the literal-data pseudo-ops.
 - **Disassembler.**  `gasm dis` lists a `.s` file's functions at their real
   offsets after assembling, or disassembles raw bytes from a file or stdin.
 - **Dynamic verification.**  `gasm verify` JIT-loads assembled functions into
@@ -91,13 +94,16 @@ to give that syntax the tooling it deserves.
 - **Debugger.**  `gasm debug` is a source-level ptrace debugger with
   breakpoints (optionally conditional), hardware watchpoints, register and
   memory inspection, and headless script runs that report instruction and
-  label coverage.
+  label coverage; it runs on Linux (all four architectures) and FreeBSD
+  (amd64, arm64, riscv64).
 - **Language server.**  `gasm lsp` serves completion, hover, document symbols,
   push and pull diagnostics, semantic-token highlighting, go-to-definition,
   find references, rename, formatting, inlay hints, code actions, signature
   help, document highlights, workspace symbol search, #include document
   links and folding ranges over stdio; definition, references and rename
-  work across every open document.
+  work across every open document and the indexed workspace files beyond
+  them, and the quick fixes add a missing textflag.h include and set the
+  argument area from the // func signature.
 - **Comparators and audits.**  `gasm diff` compares the machine code of two
   assembly files byte-for-byte, `gasm profile` shows basic-block structure,
   `gasm audit-instructions` diffs the encoder against the installed toolchain,
@@ -110,9 +116,9 @@ Four architectures, the four that matter in practice:
 | Architecture | GOARCH      | File suffix  | Instructions recognised                     |
 |--------------|-------------|--------------|---------------------------------------------|
 | AMD64        | `amd64`     | `_amd64.s`   | 1600 + common opcodes + traditional aliases |
-| ARM64        | `arm64`     | `_arm64.s`   | 538 + common opcodes                        |
-| RISC-V       | `riscv64`   | `_riscv64.s` | 961 + common opcodes                        |
-| LoongArch    | `loong64`   | `_loong64.s` | 799 + common opcodes                        |
+| ARM64        | `arm64`     | `_arm64.s`   | 645 + common opcodes                        |
+| RISC-V       | `riscv64`   | `_riscv64.s` | 992 + common opcodes                        |
+| LoongArch    | `loong64`   | `_loong64.s` | 808 + common opcodes                        |
 
 "Common opcodes" are the instructions shared by every architecture (`RET`,
 `JMP`, `NOP`, `CALL`, `TEXT`, `FUNCDATA`, `PCDATA`, ...).  AMD64 additionally
@@ -124,10 +130,13 @@ 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.
+`gasm audit-instructions --corpus` reports every real-code GOROOT assembly
+file (the tree without testdata) assembling for every target its build
+admits: 250 of 250, 100 %. Over the whole tree including testdata the
+measure is 271 of 322 attemptable (84.2 %); files named for other Go ports
+are counted but never attempted, and `//go:build` constraints decide which
+targets attempt a file at all, exactly as the build does. The number moves
+with every release.
 
 ### Validation status
 
@@ -142,7 +151,7 @@ actually been executed.
 |---|---|---|
 | 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 |
+| Debugger: ptrace tracing, breakpoints, watchpoints, coverage | native hardware | emulation cannot run ptrace; the layer compiles and its architecture-neutral units run under `go test ./...`, nothing more. FreeBSD (amd64, arm64, riscv64) is in the same position: the port compiles behind the cross-build gate and its integration test is ready, but no FreeBSD machine has executed it |
 
 Consequences, stated plainly. An emulator is a model of a CPU, not the
 CPU: instruction semantics are implemented in software and can differ
@@ -157,6 +166,39 @@ 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/...`.
 
+## The documentation goal
+
+The toolkit is the primary goal. The secondary one is documentation: a
+specification of the Plan 9 assembly language and of the GOOBJ object
+format that is 100 % complete, detailed enough to implement against,
+and written to a professional standard. These are the two subjects this
+project works with every day, and they are the two for which no usable
+documentation exists.
+
+Go documents the language on a single page, "A Quick Guide to Go's
+Assembler", which carries no section for loong64, one of the four
+architectures gasm supports, and covers a fraction of what each
+assembler accepts. What exists beyond it lives as comments inside the
+toolchain's internal source: per-architecture reference manuals for
+arm64, ppc64, riscv64 and loong64, written for the toolchain's own
+maintainers rather than for an outside reader, and none at all for
+amd64. GOOBJ fares worst of all. The format that `go build` consumes
+has no specification anywhere: it is described by a comment in an
+internal package, it is not a stable interface, and it can change with
+any toolchain release.
+
+The gap is therefore filled the only way it can be filled: by reverse
+engineering the toolchain itself, the same work the encoders already
+perform. Most of the documentation can come from nowhere else, and it
+is written as that knowledge is produced during development. It is
+verified the way the code is verified: an encoding documented here is
+one that differential tests against `go tool asm` confirm
+byte-for-byte, and a format field documented here is one the linker
+demonstrably reads. The work has begun: [docs/GOOBJ.md](docs/GOOBJ.md)
+specifies the object file format completely, and
+[docs/asm/README.md](docs/asm/README.md) opens the language reference
+with its common core. The per-architecture pages follow.
+
 ## Direction
 
 The plan, in the order it is being worked:
@@ -179,9 +221,11 @@ The plan, in the order it is being worked:
   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.
+  architectures and is where the binary builds.  FreeBSD follows on amd64,
+  arm64 and riscv64: the JIT's executable-memory mapping and the ptrace
+  debugger layer are ported (the debugger's live validation awaits a
+  FreeBSD machine, as the validation status states).  Other unix systems
+  may follow those two.
 - **Four architectures, no more.**  amd64, arm64, riscv64 and loong64.
   No others are planned.
 
@@ -189,11 +233,11 @@ The plan, in the order it is being worked:
 
 Prebuilt binaries for linux/amd64, linux/arm64, linux/riscv64 and
 linux/loong64 are on the
-[releases page](https://sourcedock.dev/petrbalvin/gasm-devkit/releases).
+[releases page](https://sourcedock.dev/petrbalvin/gasm-sdk/releases).
 From source (Go 1.27.1):
 
 ```sh
-go install sourcedock.dev/petrbalvin/gasm-devkit/cmd/gasm@latest
+go install sourcedock.dev/petrbalvin/gasm-sdk/cmd/gasm@latest
 ```
 
 Or from a repository checkout:
@@ -282,6 +326,8 @@ recipe.
   ~/.local/share/man (MANDIR overrides); `just uninstall-man` removes
   them
 - [docs/ARCHITECTURE.md](docs/ARCHITECTURE.md): components and data flow
+- [docs/GOOBJ.md](docs/GOOBJ.md): the GOOBJ object file format specification
+- [docs/asm/](docs/asm/README.md): the Plan 9 assembly language reference
 - [docs/DEVELOPMENT.md](docs/DEVELOPMENT.md): development setup and recipes
 - [CHANGELOG.md](CHANGELOG.md): release history
 

SECURITY.md

@@ -7,7 +7,7 @@ releases do not receive them.
 
 | Version | Supported |
 |---|---|
-| 0.34.0 | yes |
+| 0.35.0 | yes |
 | older releases | no |
 
 ## Reporting a vulnerability

_gen/gen.go

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

arch/amd64_gen.go

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

arch/arm64_ext.go

@@ -0,0 +1,624 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+// This file carries the extended-instruction layer: instructions the Go
+// toolchain does not know at all, described as data and validated against
+// golden vectors from the Arm Architecture Reference Manual rather than
+// against the toolchain.  It sits beside the generated tables, never inside
+// them: arch/arm64_gen.go stays untouched, and Extensions returns the layer
+// per architecture so a later amd64 table attaches through the same door.
+//
+// The first entry is the arm64 SVE and SVE2 integer add/subtract/multiply
+// family (twenty-three forms over four word shapes).  The encodings are
+// transcribed from the manual and cross-checked against the GNU assembler's
+// and LLVM's published encodings; the golden vectors in arm64_ext_test.go pin
+// the bytes.
+
+package arch
+
+import "fmt"
+
+// ExtOperandKind classifies one operand of an extended instruction.
+type ExtOperandKind uint8
+
+// Operand kinds.
+const (
+	ExtZReg ExtOperandKind = iota // scalable vector register Z0-Z31
+	ExtPReg                       // predicate register P0-P15
+	ExtImm                        // immediate
+)
+
+// String returns a short label for the kind.
+func (k ExtOperandKind) String() string {
+	switch k {
+	case ExtZReg:
+		return "scalable vector register"
+	case ExtPReg:
+		return "predicate register"
+	case ExtImm:
+		return "immediate"
+	default:
+		return "operand"
+	}
+}
+
+// ExtArrangement is the element-size suffix a scalable vector operand
+// carries: .B, .H, .S, .D or .Q.  ExtArrNone means the operand is written
+// bare, which the SVE forms in this layer reject.
+type ExtArrangement uint8
+
+// Arrangements, widest last.
+const (
+	ExtArrNone ExtArrangement = iota
+	ExtArrB                   // 8-bit elements
+	ExtArrH                   // 16-bit elements
+	ExtArrS                   // 32-bit elements
+	ExtArrD                   // 64-bit elements
+	ExtArrQ                   // 128-bit elements
+)
+
+// String returns the assembler suffix, with the leading dot.
+func (a ExtArrangement) String() string {
+	switch a {
+	case ExtArrB:
+		return ".B"
+	case ExtArrH:
+		return ".H"
+	case ExtArrS:
+		return ".S"
+	case ExtArrD:
+		return ".D"
+	case ExtArrQ:
+		return ".Q"
+	default:
+		return ""
+	}
+}
+
+// Width returns the byte width of one element under the arrangement.
+func (a ExtArrangement) Width() int {
+	switch a {
+	case ExtArrB:
+		return 1
+	case ExtArrH:
+		return 2
+	case ExtArrS:
+		return 4
+	case ExtArrD:
+		return 8
+	case ExtArrQ:
+		return 16
+	default:
+		return 0
+	}
+}
+
+// sizeBits maps the arrangement onto the two-bit size field the integer SVE
+// classes carry at bits 23..22: 00=B, 01=H, 10=S, 11=D.  ok is false for the
+// arrangements no such class accepts (.Q and the bare spelling).
+func (a ExtArrangement) sizeBits() (uint32, bool) {
+	switch a {
+	case ExtArrB, ExtArrH, ExtArrS, ExtArrD:
+		return uint32(a) - 1, true
+	default:
+		return 0, false
+	}
+}
+
+// ExtQualifier is the predicate qualifier spelled after the slash.
+type ExtQualifier uint8
+
+// Predicate qualifiers.
+const (
+	ExtQualNone    ExtQualifier = iota // bare Pn (non-predicating position)
+	ExtQualMerging                     // /M, inactive lanes keep the destination
+	ExtQualZeroing                     // /Z, inactive lanes become zero
+)
+
+// String returns the assembler spelling, with the leading slash.
+func (q ExtQualifier) String() string {
+	switch q {
+	case ExtQualMerging:
+		return "/M"
+	case ExtQualZeroing:
+		return "/Z"
+	default:
+		return ""
+	}
+}
+
+// ExtOperand is one operand of an extended instruction, already resolved to
+// its pieces: a register with its arrangement and qualifier, or an immediate
+// with its optional left shift.  The assembler's future hook constructs these
+// from the parsed statement; Encode consumes them.
+type ExtOperand struct {
+	Kind ExtOperandKind
+	Reg  int            // register number (Z: 0..31, P: 0..15)
+	Arr  ExtArrangement // element-size suffix; ExtArrNone when bare
+	Qual ExtQualifier   // predicate qualifier; ExtQualNone elsewhere
+	Imm  int64          // immediate value (ExtImm only)
+	// Shift carries the LSL amount an immediate form shifts the constant by
+	// before use (0 or 8 in the SVE add/subtract immediate class).  HasShift
+	// separates a spelled shift (validated as written) from an unshifted
+	// operand (the encoder may derive the sh bit from the value).
+	Shift    int
+	HasShift bool
+}
+
+// ExtVector builds a scalable vector operand, ADD Z1.S style.
+func ExtVector(reg int, arr ExtArrangement) ExtOperand {
+	return ExtOperand{Kind: ExtZReg, Reg: reg, Arr: arr}
+}
+
+// ExtPredicate builds a predicate operand with its qualifier, P0/M style.
+func ExtPredicate(reg int, qual ExtQualifier) ExtOperand {
+	return ExtOperand{Kind: ExtPReg, Reg: reg, Qual: qual}
+}
+
+// ExtImmediate builds an unshifted immediate operand.
+func ExtImmediate(v int64) ExtOperand {
+	return ExtOperand{Kind: ExtImm, Imm: v}
+}
+
+// ExtShiftedImmediate builds an immediate operand with a spelled LSL amount.
+func ExtShiftedImmediate(v int64, shift int) ExtOperand {
+	return ExtOperand{Kind: ExtImm, Imm: v, Shift: shift, HasShift: true}
+}
+
+// ExtField is one named field of the 32-bit encoding word: a bit offset from
+// the least significant end and the field's width.
+type ExtField struct {
+	Off   uint8
+	Width uint8
+}
+
+// extMask returns the field's bits as a mask.
+func extMask(f ExtField) uint32 {
+	return ^uint32(0) >> (32 - f.Width)
+}
+
+// extSet ORs v into the field of word.
+func extSet(word uint32, f ExtField, v uint32) uint32 {
+	return word | (v&extMask(f))<<f.Off
+}
+
+// The fields the SVE integer classes use.  The 5-bit register fields are
+// named after their role in the three-vector class; the predicated class
+// reuses extFieldRn for its Zm operand and extFieldPg for the governing
+// predicate, which that class narrows to three bits (P0-P7).
+var (
+	extFieldRd   = ExtField{0, 5}  // destination (Zd or Zdn)
+	extFieldRn   = ExtField{5, 5}  // first source (Zn, or Zm in the predicated class)
+	extFieldRm   = ExtField{16, 5} // second source (Zm in the three-vector class)
+	extFieldPg   = ExtField{10, 3} // governing predicate P0-P7 (predicated class)
+	extFieldImm8 = ExtField{5, 8}  // the immediate, bits 12..5
+	extFieldSh   = ExtField{13, 1} // the shift flag: 1 means LSL #8
+	extSizeBHSD  = ExtField{22, 2} // element-size field of every class here, bits 23..22
+)
+
+// ExtForm enumerates the operand shapes the extension layer defines, in Plan
+// 9 order (sources first, destination last).  A destructive SVE operand is
+// written once, in destination position: the encoding carries no second copy.
+type ExtForm uint8
+
+// Operand shapes.
+const (
+	// ExtFormVectors is the unpredicated three-vector form, the SVE integer
+	// add/subtract (unpredicated) class: ADD Z0.S, Z1.S, Z2.S computes
+	// Z0 = Z1 + Z2.  Operands: Zn, Zm, Zd.
+	ExtFormVectors ExtForm = iota
+	// ExtFormPredicated is the governed destructive form, the SVE integer
+	// add/subtract vectors (predicated) class: ADD Z1.S, P0/M, Z0.S computes
+	// Z0 = Z0 + Z1 for the active lanes.  Operands: Zm, Pg/M, Zdn.  The
+	// governing predicate is a 3-bit field, so only P0-P7 encode here, and
+	// the class takes the merging qualifier alone: a zeroing form would need
+	// a MOVPRFX expansion, which one data word cannot carry.
+	ExtFormPredicated
+	// ExtFormImmediate is the add/subtract immediate form, the SVE integer
+	// add/subtract (immediate) class: ADD $255, Z0.S computes
+	// Z0 = Z0 + 255.  Operands: imm{, LSL #8}, Zdn.  The constant is an
+	// unsigned imm8, optionally shifted left by 8 bits; a bare multiple of
+	// 256 (up to 65280) derives the shift, the spelling the GNU assembler
+	// canonicalises too.  .B takes no shift.
+	ExtFormImmediate
+	// ExtFormSignedImmediate is the signed immediate form of the SVE integer
+	// multiply (immediate) class: MUL $-128, Z0.B computes Z0 = Z0 * -128.
+	// Operands: simm8, Zdn.  No shift exists in this class.
+	ExtFormSignedImmediate
+)
+
+// Arity returns the operand count the form takes.
+func (f ExtForm) Arity() int {
+	switch f {
+	case ExtFormVectors, ExtFormPredicated:
+		return 3
+	case ExtFormImmediate, ExtFormSignedImmediate:
+		return 2
+	default:
+		return 0
+	}
+}
+
+// Kinds returns the operand kind each position of the form wants, in the
+// order the operands arrive.  The registry uses the list to pick the most
+// specific rejection when every matching form refuses an operand list.
+func (f ExtForm) Kinds() []ExtOperandKind {
+	switch f {
+	case ExtFormVectors:
+		return []ExtOperandKind{ExtZReg, ExtZReg, ExtZReg}
+	case ExtFormPredicated:
+		return []ExtOperandKind{ExtZReg, ExtPReg, ExtZReg}
+	case ExtFormImmediate, ExtFormSignedImmediate:
+		return []ExtOperandKind{ExtImm, ExtZReg}
+	default:
+		return nil
+	}
+}
+
+// String returns a short label for the form, for diagnostics.
+func (f ExtForm) String() string {
+	switch f {
+	case ExtFormVectors:
+		return "unpredicated vectors"
+	case ExtFormPredicated:
+		return "predicated (merging)"
+	case ExtFormImmediate:
+		return "unsigned immediate"
+	case ExtFormSignedImmediate:
+		return "signed immediate"
+	default:
+		return "unknown form"
+	}
+}
+
+// ExtFeature names the architecture feature an extended instruction belongs
+// to.  The field is metadata: the assembler offers every instruction it
+// registers, and a feature check is the caller's decision, not the encoder's.
+type ExtFeature string
+
+// The features the arm64 layer covers.
+const (
+	ExtFeatureSVE  ExtFeature = "sve"
+	ExtFeatureSVE2 ExtFeature = "sve2"
+)
+
+// ExtInstr is one extended instruction: the metadata a lookup needs and the
+// encoding as data.  Word holds the fixed bits of the 32-bit encoding with
+// every operand field and the size field zero; the form says which fields the
+// operands fill; the size field receives the arrangement's bits at encode
+// time.  Ref names the manual entry the encoding is transcribed from, the
+// golden source in place of a toolchain oracle.
+type ExtInstr struct {
+	Name    string     // upper-case mnemonic
+	Summary string     // one line of hover documentation
+	Word    uint32     // fixed encoding bits, operand fields zero
+	Form    ExtForm    // operand shape
+	Size    ExtField   // element-size field the arrangement fills
+	Feature ExtFeature // sve or sve2
+	Ref     string     // the ARM ARM entry the encoding comes from
+}
+
+// Encode assembles the operands into the 4 little-endian bytes of the
+// instruction word.  The operand kinds, register ranges, arrangements and
+// immediate ranges are validated against the form; an operand the class
+// cannot carry is an error, never a silent mis-encoding.
+func (in ExtInstr) Encode(ops []ExtOperand) ([]byte, error) {
+	if len(ops) != in.Form.Arity() {
+		return nil, fmt.Errorf("%s: the %s form takes %d operands, got %d",
+			in.Name, in.Form, in.Form.Arity(), len(ops))
+	}
+	switch in.Form {
+	case ExtFormVectors:
+		return in.encodeVectors(ops)
+	case ExtFormPredicated:
+		return in.encodePredicated(ops)
+	case ExtFormImmediate:
+		return in.encodeImmediate(ops)
+	case ExtFormSignedImmediate:
+		return in.encodeSignedImmediate(ops)
+	default:
+		return nil, fmt.Errorf("%s: unknown form %d", in.Name, in.Form)
+	}
+}
+
+// encodeVectors fills the unpredicated three-vector form: Zn, Zm, Zd, all
+// under one required arrangement.
+func (in ExtInstr) encodeVectors(ops []ExtOperand) ([]byte, error) {
+	for i, op := range ops {
+		if op.Kind != ExtZReg {
+			return nil, fmt.Errorf("%s: operand %d wants a scalable vector register, got %s",
+				in.Name, i+1, op.Kind)
+		}
+		if op.Reg < 0 || op.Reg > 31 {
+			return nil, fmt.Errorf("%s: operand %d is Z%d, outside Z0-Z31", in.Name, i+1, op.Reg)
+		}
+	}
+	arr, err := in.sharedArrangement(ops)
+	if err != nil {
+		return nil, err
+	}
+	size, ok := arr.sizeBits()
+	if !ok {
+		return nil, fmt.Errorf("%s: arrangement %s has no size encoding in this class", in.Name, arr)
+	}
+	word := in.Word
+	word = extSet(word, extFieldRn, uint32(ops[0].Reg))
+	word = extSet(word, extFieldRm, uint32(ops[1].Reg))
+	word = extSet(word, extFieldRd, uint32(ops[2].Reg))
+	word = extSet(word, in.Size, size)
+	return extWordLE(word), nil
+}
+
+// encodePredicated fills the governed destructive form: Zm, Pg/M, Zdn.  The
+// predicate is a 3-bit field, the merging qualifier alone, and carries no
+// arrangement suffix in this class.
+func (in ExtInstr) encodePredicated(ops []ExtOperand) ([]byte, error) {
+	zm, pg, zdn := ops[0], ops[1], ops[2]
+	if zm.Kind != ExtZReg {
+		return nil, fmt.Errorf("%s: operand 1 wants a scalable vector register, got %s",
+			in.Name, zm.Kind)
+	}
+	if zm.Reg < 0 || zm.Reg > 31 {
+		return nil, fmt.Errorf("%s: operand 1 is Z%d, outside Z0-Z31", in.Name, zm.Reg)
+	}
+	if pg.Kind != ExtPReg {
+		return nil, fmt.Errorf("%s: operand 2 wants a predicate register, got %s",
+			in.Name, pg.Kind)
+	}
+	if pg.Reg < 0 || pg.Reg > 7 {
+		return nil, fmt.Errorf("%s: operand 2 is P%d, outside P0-P7 in this class", in.Name, pg.Reg)
+	}
+	if pg.Qual != ExtQualMerging {
+		return nil, fmt.Errorf("%s: operand 2 wants the merging qualifier /M, got %q",
+			in.Name, pg.Qual)
+	}
+	if pg.Arr != ExtArrNone {
+		return nil, fmt.Errorf("%s: the governing predicate carries no arrangement suffix, got %s",
+			in.Name, pg.Arr)
+	}
+	if zdn.Kind != ExtZReg {
+		return nil, fmt.Errorf("%s: operand 3 wants a scalable vector register, got %s",
+			in.Name, zdn.Kind)
+	}
+	if zdn.Reg < 0 || zdn.Reg > 31 {
+		return nil, fmt.Errorf("%s: operand 3 is Z%d, outside Z0-Z31", in.Name, zdn.Reg)
+	}
+	if zm.Arr != zdn.Arr {
+		return nil, fmt.Errorf("%s: operands 1 and 3 carry arrangements %s and %s, they must match",
+			in.Name, zm.Arr, zdn.Arr)
+	}
+	size, ok := zdn.Arr.sizeBits()
+	if !ok {
+		return nil, fmt.Errorf("%s: arrangement %s has no size encoding in this class", in.Name, zdn.Arr)
+	}
+	word := in.Word
+	word = extSet(word, extFieldRn, uint32(zm.Reg))
+	word = extSet(word, extFieldPg, uint32(pg.Reg))
+	word = extSet(word, extFieldRd, uint32(zdn.Reg))
+	word = extSet(word, in.Size, size)
+	return extWordLE(word), nil
+}
+
+// encodeImmediate fills the add/subtract immediate form: imm{, LSL #8}, Zdn.
+// The class encodes an unsigned imm8 with one shift bit, so a bare multiple
+// of 256 derives the shift the way the GNU assembler canonicalises it.
+func (in ExtInstr) encodeImmediate(ops []ExtOperand) ([]byte, error) {
+	imm, zdn := ops[0], ops[1]
+	imm8, sh, err := in.addSubImmediate(imm, zdn.Arr)
+	if err != nil {
+		return nil, err
+	}
+	word := in.Word
+	word = extSet(word, extFieldImm8, uint32(imm8))
+	if sh != 0 {
+		word = extSet(word, extFieldSh, 1)
+	}
+	word, err = in.setDestAndSize(word, zdn)
+	if err != nil {
+		return nil, err
+	}
+	return extWordLE(word), nil
+}
+
+// encodeSignedImmediate fills the multiply immediate form: simm8, Zdn, with
+// no shift bit in the class.
+func (in ExtInstr) encodeSignedImmediate(ops []ExtOperand) ([]byte, error) {
+	imm, zdn := ops[0], ops[1]
+	if imm.Kind != ExtImm {
+		return nil, fmt.Errorf("%s: operand 1 wants an immediate, got %s", in.Name, imm.Kind)
+	}
+	if imm.HasShift {
+		return nil, fmt.Errorf("%s: the signed immediate class takes no shift", in.Name)
+	}
+	if imm.Imm < -128 || imm.Imm > 127 {
+		return nil, fmt.Errorf("%s: immediate %d is outside the signed 8-bit range -128..127",
+			in.Name, imm.Imm)
+	}
+	word := in.Word
+	word = extSet(word, extFieldImm8, uint32(imm.Imm))
+	word, err := in.setDestAndSize(word, zdn)
+	if err != nil {
+		return nil, err
+	}
+	return extWordLE(word), nil
+}
+
+// addSubImmediate resolves the immediate operand of the add/subtract
+// immediate class into its imm8 and shift bit: a spelled shift is validated
+// as written, a bare multiple of 256 (on .H, .S or .D) derives one.
+func (in ExtInstr) addSubImmediate(op ExtOperand, arr ExtArrangement) (imm8, sh int, err error) {
+	if op.Kind != ExtImm {
+		return 0, 0, fmt.Errorf("%s: operand 1 wants an immediate, got %s", in.Name, op.Kind)
+	}
+	switch {
+	case op.HasShift:
+		if op.Shift != 0 && op.Shift != 8 {
+			return 0, 0, fmt.Errorf("%s: the shift amount must be 0 or 8, got %d", in.Name, op.Shift)
+		}
+		if arr == ExtArrB && op.Shift != 0 {
+			return 0, 0, fmt.Errorf("%s: arrangement .B takes no shift", in.Name)
+		}
+		if op.Imm < 0 || op.Imm > 255 {
+			return 0, 0, fmt.Errorf("%s: immediate %d is outside the unsigned 8-bit range 0..255",
+				in.Name, op.Imm)
+		}
+		return int(op.Imm), op.Shift, nil
+	case op.Imm >= 0 && op.Imm <= 255:
+		return int(op.Imm), 0, nil
+	case arr != ExtArrB && op.Imm >= 256 && op.Imm <= 255<<8 && op.Imm%256 == 0:
+		// A bare multiple of 256 rides the shift bit, 65280 = 255<<8 included.
+		return int(op.Imm / 256), 8, nil
+	default:
+		return 0, 0, fmt.Errorf("%s: immediate %d is not an unsigned imm8%s, nor a multiple of 256 the shift bit can carry",
+			in.Name, op.Imm, arr.shiftNote())
+	}
+}
+
+// shiftNote describes where a shifted constant is expressible, for the
+// immediate range error.
+func (arr ExtArrangement) shiftNote() string {
+	if arr == ExtArrB {
+		return " (and .B takes no shifted constant)"
+	}
+	return " (a multiple of 256 up to 65280 shifts)"
+}
+
+// setDestAndSize fills the destructive destination register and the size
+// field from the arrangement the vector carries.
+func (in ExtInstr) setDestAndSize(word uint32, zdn ExtOperand) (uint32, error) {
+	if zdn.Kind != ExtZReg {
+		return 0, fmt.Errorf("%s: operand 2 wants a scalable vector register, got %s",
+			in.Name, zdn.Kind)
+	}
+	if zdn.Reg < 0 || zdn.Reg > 31 {
+		return 0, fmt.Errorf("%s: operand 2 is Z%d, outside Z0-Z31", in.Name, zdn.Reg)
+	}
+	size, ok := zdn.Arr.sizeBits()
+	if !ok {
+		return 0, fmt.Errorf("%s: arrangement %s has no size encoding in this class", in.Name, zdn.Arr)
+	}
+	word = extSet(word, extFieldRd, uint32(zdn.Reg))
+	word = extSet(word, in.Size, size)
+	return word, nil
+}
+
+// sharedArrangement returns the one arrangement all vector operands carry, or
+// an error when any operand is bare or they disagree.
+func (in ExtInstr) sharedArrangement(ops []ExtOperand) (ExtArrangement, error) {
+	arr := ops[0].Arr
+	for i, op := range ops {
+		if op.Arr == ExtArrNone {
+			return 0, fmt.Errorf("%s: operand %d carries no arrangement suffix", in.Name, i+1)
+		}
+		if op.Arr != arr {
+			return 0, fmt.Errorf("%s: operand %d carries arrangement %s, want %s",
+				in.Name, i+1, op.Arr, arr)
+		}
+	}
+	return arr, nil
+}
+
+// extWordLE returns a 32-bit encoding word as 4 little-endian bytes.
+func extWordLE(w uint32) []byte {
+	return []byte{byte(w), byte(w >> 8), byte(w >> 16), byte(w >> 24)}
+}
+
+// --- the arm64 SVE/SVE2 table ------------------------------------------------
+
+// arm64Extensions is the extended instruction layer of arm64: the SVE and
+// SVE2 integer add/subtract/multiply family.  The Go toolchain knows none of
+// these; the encodings are transcribed from the ARM Architecture Reference
+// Manual (DDI 0487J, Part C, Chapter C8, the alphabetical list of SVE
+// instructions) and cross-checked against the GNU assembler's and LLVM's
+// published encodings.
+var arm64Extensions = []ExtInstr{
+	// Unpredicated three-vector forms: ADD Z0.S, Z1.S, Z2.S.
+	{Name: "ADD", Summary: "Add scalable vector elements, unpredicated",
+		Word: 0x04200000, Form: ExtFormVectors, Size: extSizeBHSD, Feature: ExtFeatureSVE,
+		Ref: "ARM DDI 0487J, C8.2 SVE instruction descriptions: ADD (vectors, unpredicated)"},
+	{Name: "SUB", Summary: "Subtract scalable vector elements, unpredicated",
+		Word: 0x04200400, Form: ExtFormVectors, Size: extSizeBHSD, Feature: ExtFeatureSVE,
+		Ref: "ARM DDI 0487J, C8.2 SVE instruction descriptions: SUB (vectors, unpredicated)"},
+	{Name: "SQADD", Summary: "Add signed saturating scalable vector elements, unpredicated",
+		Word: 0x04201000, Form: ExtFormVectors, Size: extSizeBHSD, Feature: ExtFeatureSVE,
+		Ref: "ARM DDI 0487J, C8.2 SVE instruction descriptions: SQADD (vectors, unpredicated)"},
+	{Name: "UQADD", Summary: "Add unsigned saturating scalable vector elements, unpredicated",
+		Word: 0x04201400, Form: ExtFormVectors, Size: extSizeBHSD, Feature: ExtFeatureSVE,
+		Ref: "ARM DDI 0487J, C8.2 SVE instruction descriptions: UQADD (vectors, unpredicated)"},
+	{Name: "SQSUB", Summary: "Subtract signed saturating scalable vector elements, unpredicated",
+		Word: 0x04201800, Form: ExtFormVectors, Size: extSizeBHSD, Feature: ExtFeatureSVE,
+		Ref: "ARM DDI 0487J, C8.2 SVE instruction descriptions: SQSUB (vectors, unpredicated)"},
+	{Name: "UQSUB", Summary: "Subtract unsigned saturating scalable vector elements, unpredicated",
+		Word: 0x04201c00, Form: ExtFormVectors, Size: extSizeBHSD, Feature: ExtFeatureSVE,
+		Ref: "ARM DDI 0487J, C8.2 SVE instruction descriptions: UQSUB (vectors, unpredicated)"},
+	{Name: "MUL", Summary: "Multiply scalable vector elements, unpredicated",
+		Word: 0x04206000, Form: ExtFormVectors, Size: extSizeBHSD, Feature: ExtFeatureSVE2,
+		Ref: "ARM DDI 0487J, C8.2 SVE instruction descriptions: MUL (vectors, unpredicated)"},
+	{Name: "SMULH", Summary: "Multiply signed scalable vector elements, keeping the high half, unpredicated",
+		Word: 0x04206800, Form: ExtFormVectors, Size: extSizeBHSD, Feature: ExtFeatureSVE2,
+		Ref: "ARM DDI 0487J, C8.2 SVE instruction descriptions: SMULH (vectors, unpredicated)"},
+	{Name: "UMULH", Summary: "Multiply unsigned scalable vector elements, keeping the high half, unpredicated",
+		Word: 0x04206c00, Form: ExtFormVectors, Size: extSizeBHSD, Feature: ExtFeatureSVE2,
+		Ref: "ARM DDI 0487J, C8.2 SVE instruction descriptions: UMULH (vectors, unpredicated)"},
+
+	// Governed destructive forms, merging: ADD Z1.S, P0/M, Z0.S.
+	{Name: "ADD", Summary: "Add scalable vector elements under a governing predicate, merging",
+		Word: 0x04000000, Form: ExtFormPredicated, Size: extSizeBHSD, Feature: ExtFeatureSVE,
+		Ref: "ARM DDI 0487J, C8.2 SVE instruction descriptions: ADD (vectors, predicated)"},
+	{Name: "SUB", Summary: "Subtract scalable vector elements under a governing predicate, merging",
+		Word: 0x04010000, Form: ExtFormPredicated, Size: extSizeBHSD, Feature: ExtFeatureSVE,
+		Ref: "ARM DDI 0487J, C8.2 SVE instruction descriptions: SUB (vectors, predicated)"},
+	{Name: "SUBR", Summary: "Reverse-subtract scalable vector elements under a governing predicate, merging",
+		Word: 0x04030000, Form: ExtFormPredicated, Size: extSizeBHSD, Feature: ExtFeatureSVE,
+		Ref: "ARM DDI 0487J, C8.2 SVE instruction descriptions: SUBR (vectors, predicated)"},
+	{Name: "MUL", Summary: "Multiply scalable vector elements under a governing predicate, merging",
+		Word: 0x04100000, Form: ExtFormPredicated, Size: extSizeBHSD, Feature: ExtFeatureSVE,
+		Ref: "ARM DDI 0487J, C8.2 SVE instruction descriptions: MUL (vectors, predicated)"},
+	{Name: "SMULH", Summary: "Multiply signed scalable vector elements, keeping the high half, under a governing predicate, merging",
+		Word: 0x04120000, Form: ExtFormPredicated, Size: extSizeBHSD, Feature: ExtFeatureSVE,
+		Ref: "ARM DDI 0487J, C8.2 SVE instruction descriptions: SMULH (vectors, predicated)"},
+	{Name: "UMULH", Summary: "Multiply unsigned scalable vector elements, keeping the high half, under a governing predicate, merging",
+		Word: 0x04130000, Form: ExtFormPredicated, Size: extSizeBHSD, Feature: ExtFeatureSVE,
+		Ref: "ARM DDI 0487J, C8.2 SVE instruction descriptions: UMULH (vectors, predicated)"},
+
+	// Immediate forms: ADD $255, Z0.S.
+	{Name: "ADD", Summary: "Add an unsigned immediate to scalable vector elements",
+		Word: 0x2520c000, Form: ExtFormImmediate, Size: extSizeBHSD, Feature: ExtFeatureSVE,
+		Ref: "ARM DDI 0487J, C8.2 SVE instruction descriptions: ADD (vectors, immediate)"},
+	{Name: "SUB", Summary: "Subtract an unsigned immediate from scalable vector elements",
+		Word: 0x2521c000, Form: ExtFormImmediate, Size: extSizeBHSD, Feature: ExtFeatureSVE,
+		Ref: "ARM DDI 0487J, C8.2 SVE instruction descriptions: SUB (vectors, immediate)"},
+	{Name: "SUBR", Summary: "Subtract scalable vector elements from an unsigned immediate",
+		Word: 0x2523c000, Form: ExtFormImmediate, Size: extSizeBHSD, Feature: ExtFeatureSVE,
+		Ref: "ARM DDI 0487J, C8.2 SVE instruction descriptions: SUBR (vectors, immediate)"},
+	{Name: "SQADD", Summary: "Add a signed saturating unsigned immediate to scalable vector elements",
+		Word: 0x2524c000, Form: ExtFormImmediate, Size: extSizeBHSD, Feature: ExtFeatureSVE,
+		Ref: "ARM DDI 0487J, C8.2 SVE instruction descriptions: SQADD (vectors, immediate)"},
+	{Name: "UQADD", Summary: "Add an unsigned saturating immediate to scalable vector elements",
+		Word: 0x2525c000, Form: ExtFormImmediate, Size: extSizeBHSD, Feature: ExtFeatureSVE,
+		Ref: "ARM DDI 0487J, C8.2 SVE instruction descriptions: UQADD (vectors, immediate)"},
+	{Name: "SQSUB", Summary: "Subtract an unsigned immediate from scalable vector elements with signed saturation",
+		Word: 0x2526c000, Form: ExtFormImmediate, Size: extSizeBHSD, Feature: ExtFeatureSVE,
+		Ref: "ARM DDI 0487J, C8.2 SVE instruction descriptions: SQSUB (vectors, immediate)"},
+	{Name: "UQSUB", Summary: "Subtract an unsigned immediate from scalable vector elements with unsigned saturation",
+		Word: 0x2527c000, Form: ExtFormImmediate, Size: extSizeBHSD, Feature: ExtFeatureSVE,
+		Ref: "ARM DDI 0487J, C8.2 SVE instruction descriptions: UQSUB (vectors, immediate)"},
+
+	// The signed immediate of the multiply class: MUL $-128, Z0.B.
+	{Name: "MUL", Summary: "Multiply scalable vector elements by a signed immediate",
+		Word: 0x2530c000, Form: ExtFormSignedImmediate, Size: extSizeBHSD, Feature: ExtFeatureSVE,
+		Ref: "ARM DDI 0487J, C8.2 SVE instruction descriptions: MUL (vectors, immediate)"},
+}
+
+// Extensions returns the extended-instruction layer registered for a, outside
+// the generated tables.  An architecture whose extended layer is not built
+// yet returns nothing: the mechanism is ordinary code, not a build tag, and
+// it simply offers no instruction where none is registered.
+func Extensions(a Arch) []ExtInstr {
+	switch a {
+	case ARM64:
+		return arm64Extensions
+	default:
+		return nil
+	}
+}

arch/arm64_ext_test.go

@@ -0,0 +1,400 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package arch
+
+import (
+	"encoding/hex"
+	"strings"
+	"testing"
+)
+
+// The SVE encodings have no toolchain oracle: go tool asm knows no SVE at
+// all.  The golden words below are therefore transcribed from the ARM
+// Architecture Reference Manual (DDI 0487J, Part C, Chapter C8, the
+// alphabetical list of SVE instructions) and cross-checked against two
+// independent implementations of the manual, the GNU assembler and LLVM:
+// the rows marked "GNU" match a vector in binutils-gdb's own
+// gas/testsuite/gas/aarch64/sve.d (assembled under -march=armv8-a+sve), the
+// rows marked "LLVM" match the Inst field assignments in
+// SVEInstrFormats.td's sve_int_bin_cons_arit_0, sve_int_bin_pred_arit_log,
+// sve_int_arith_imm0 and sve2_int_mul classes.  Every class is covered by at
+// least one vector of each source.
+func extInstruction(t *testing.T, mnem string, form ExtForm) ExtInstr {
+	t.Helper()
+	for _, in := range Extensions(ARM64) {
+		if in.Name == mnem && in.Form == form {
+			return in
+		}
+	}
+	t.Fatalf("no extended %s with the %s form", mnem, form)
+	return ExtInstr{}
+}
+
+func TestArm64ExtGoldenBytes(t *testing.T) {
+	for _, tt := range []struct {
+		name  string
+		mnem  string
+		form  ExtForm
+		ops   []ExtOperand
+		want  uint32
+		GNUas string // the matching binutils-gdb sve.d line, empty when the class evidence comes from LLVM alone
+	}{
+		// Unpredicated three-vector forms: Zn, Zm, Zd.
+		{"add z0.b, z0.b, z0.b", "ADD", ExtFormVectors,
+			[]ExtOperand{ExtVector(0, ExtArrB), ExtVector(0, ExtArrB), ExtVector(0, ExtArrB)},
+			0x04200000, "04200000 add z0.b, z0.b, z0.b"},
+		{"add z0.b, z0.b, z31.b", "ADD", ExtFormVectors,
+			[]ExtOperand{ExtVector(0, ExtArrB), ExtVector(31, ExtArrB), ExtVector(0, ExtArrB)},
+			0x043f0000, "043f0000 add z0.b, z0.b, z31.b"},
+		{"add z31.b, z0.b, z0.b", "ADD", ExtFormVectors,
+			[]ExtOperand{ExtVector(0, ExtArrB), ExtVector(0, ExtArrB), ExtVector(31, ExtArrB)},
+			0x0420001f, "0420001f add z31.b, z0.b, z0.b"},
+		{"add z0.b, z2.b, z0.b", "ADD", ExtFormVectors,
+			[]ExtOperand{ExtVector(2, ExtArrB), ExtVector(0, ExtArrB), ExtVector(0, ExtArrB)},
+			0x04200040, "04200040 add z0.b, z2.b, z0.b"},
+		{"add z0.h, z0.h, z0.h", "ADD", ExtFormVectors,
+			[]ExtOperand{ExtVector(0, ExtArrH), ExtVector(0, ExtArrH), ExtVector(0, ExtArrH)},
+			0x04600000, "04600000 add z0.h, z0.h, z0.h"},
+		{"add z0.s, z0.s, z0.s", "ADD", ExtFormVectors,
+			[]ExtOperand{ExtVector(0, ExtArrS), ExtVector(0, ExtArrS), ExtVector(0, ExtArrS)},
+			0x04a00000, "04a00000 add z0.s, z0.s, z0.s"},
+		{"add z0.d, z0.d, z0.d", "ADD", ExtFormVectors,
+			[]ExtOperand{ExtVector(0, ExtArrD), ExtVector(0, ExtArrD), ExtVector(0, ExtArrD)},
+			0x04e00000, "04e00000 add z0.d, z0.d, z0.d"},
+		{"sub z0.b, z0.b, z0.b", "SUB", ExtFormVectors,
+			[]ExtOperand{ExtVector(0, ExtArrB), ExtVector(0, ExtArrB), ExtVector(0, ExtArrB)},
+			0x04200400, "04200400 sub z0.b, z0.b, z0.b"},
+		{"sub z0.b, z0.b, z3.b", "SUB", ExtFormVectors,
+			[]ExtOperand{ExtVector(0, ExtArrB), ExtVector(3, ExtArrB), ExtVector(0, ExtArrB)},
+			0x04230400, "04230400 sub z0.b, z0.b, z3.b"},
+		{"sqadd z0.b, z0.b, z0.b", "SQADD", ExtFormVectors,
+			[]ExtOperand{ExtVector(0, ExtArrB), ExtVector(0, ExtArrB), ExtVector(0, ExtArrB)},
+			0x04201000, "04201000 sqadd z0.b, z0.b, z0.b"},
+		{"sqadd z0.b, z0.b, z3.b", "SQADD", ExtFormVectors,
+			[]ExtOperand{ExtVector(0, ExtArrB), ExtVector(3, ExtArrB), ExtVector(0, ExtArrB)},
+			0x04231000, "04231000 sqadd z0.b, z0.b, z3.b"},
+		{"sqadd z0.d, z0.d, z0.d", "SQADD", ExtFormVectors,
+			[]ExtOperand{ExtVector(0, ExtArrD), ExtVector(0, ExtArrD), ExtVector(0, ExtArrD)},
+			0x04e01000, "04e01000 sqadd z0.d, z0.d, z0.d"},
+		{"uqadd z0.b, z0.b, z0.b", "UQADD", ExtFormVectors,
+			[]ExtOperand{ExtVector(0, ExtArrB), ExtVector(0, ExtArrB), ExtVector(0, ExtArrB)},
+			0x04201400, "04201400 uqadd z0.b, z0.b, z0.b"},
+		{"sqsub z0.b, z0.b, z0.b", "SQSUB", ExtFormVectors,
+			[]ExtOperand{ExtVector(0, ExtArrB), ExtVector(0, ExtArrB), ExtVector(0, ExtArrB)},
+			0x04201800, "04201800 sqsub z0.b, z0.b, z0.b"},
+		{"sqsub z0.h, z0.h, z0.h", "SQSUB", ExtFormVectors,
+			[]ExtOperand{ExtVector(0, ExtArrH), ExtVector(0, ExtArrH), ExtVector(0, ExtArrH)},
+			0x04601800, "04601800 sqsub z0.h, z0.h, z0.h"},
+		{"uqsub z0.b, z0.b, z0.b", "UQSUB", ExtFormVectors,
+			[]ExtOperand{ExtVector(0, ExtArrB), ExtVector(0, ExtArrB), ExtVector(0, ExtArrB)},
+			0x04201c00, "04201c00 uqsub z0.b, z0.b, z0.b"},
+		{"mul z0.b, z0.b, z0.b (sve2)", "MUL", ExtFormVectors,
+			[]ExtOperand{ExtVector(0, ExtArrB), ExtVector(0, ExtArrB), ExtVector(0, ExtArrB)},
+			0x04206000, "04206000 mul z0.b, z0.b, z0.b"},
+		{"mul z17.b, z21.b, z27.b (sve2)", "MUL", ExtFormVectors,
+			[]ExtOperand{ExtVector(21, ExtArrB), ExtVector(27, ExtArrB), ExtVector(17, ExtArrB)},
+			0x043b62b1, "043b62b1 mul z17.b, z21.b, z27.b"},
+		{"mul z0.d, z0.d, z0.d (sve2)", "MUL", ExtFormVectors,
+			[]ExtOperand{ExtVector(0, ExtArrD), ExtVector(0, ExtArrD), ExtVector(0, ExtArrD)},
+			0x04e06000, "04e06000 mul z0.d, z0.d, z0.d"},
+		{"smulh z0.b, z0.b, z0.b (sve2)", "SMULH", ExtFormVectors,
+			[]ExtOperand{ExtVector(0, ExtArrB), ExtVector(0, ExtArrB), ExtVector(0, ExtArrB)},
+			0x04206800, "04206800 smulh z0.b, z0.b, z0.b"},
+		{"smulh z17.b, z21.b, z27.b (sve2)", "SMULH", ExtFormVectors,
+			[]ExtOperand{ExtVector(21, ExtArrB), ExtVector(27, ExtArrB), ExtVector(17, ExtArrB)},
+			0x043b6ab1, "043b6ab1 smulh z17.b, z21.b, z27.b"},
+		{"umulh z0.b, z0.b, z0.b (sve2)", "UMULH", ExtFormVectors,
+			[]ExtOperand{ExtVector(0, ExtArrB), ExtVector(0, ExtArrB), ExtVector(0, ExtArrB)},
+			0x04206c00, "04206c00 umulh z0.b, z0.b, z0.b"},
+		{"umulh z17.b, z21.b, z27.b (sve2)", "UMULH", ExtFormVectors,
+			[]ExtOperand{ExtVector(21, ExtArrB), ExtVector(27, ExtArrB), ExtVector(17, ExtArrB)},
+			0x043b6eb1, "043b6eb1 umulh z17.b, z21.b, z27.b"},
+
+		// Governed destructive forms, merging: Zm, Pg/M, Zdn.
+		{"add z0.b, p0/m, z0.b", "ADD", ExtFormPredicated,
+			[]ExtOperand{ExtVector(0, ExtArrB), ExtPredicate(0, ExtQualMerging), ExtVector(0, ExtArrB)},
+			0x04000000, "04000000 add z0.b, p0/m, z0.b, z0.b"},
+		{"add z0.b, p2/m, z0.b", "ADD", ExtFormPredicated,
+			[]ExtOperand{ExtVector(0, ExtArrB), ExtPredicate(2, ExtQualMerging), ExtVector(0, ExtArrB)},
+			0x04000800, "04000800 add z0.b, p2/m, z0.b, z0.b"},
+		{"add z0.b, p7/m, z0.b", "ADD", ExtFormPredicated,
+			[]ExtOperand{ExtVector(0, ExtArrB), ExtPredicate(7, ExtQualMerging), ExtVector(0, ExtArrB)},
+			0x04001c00, "04001c00 add z0.b, p7/m, z0.b, z0.b"},
+		{"add z31.b, p0/m, z31.b", "ADD", ExtFormPredicated,
+			[]ExtOperand{ExtVector(0, ExtArrB), ExtPredicate(0, ExtQualMerging), ExtVector(31, ExtArrB)},
+			0x0400001f, "0400001f add z31.b, p0/m, z31.b, z0.b"},
+		{"add z0.b, p0/m, z0.b (zm 31)", "ADD", ExtFormPredicated,
+			[]ExtOperand{ExtVector(31, ExtArrB), ExtPredicate(0, ExtQualMerging), ExtVector(0, ExtArrB)},
+			0x040003e0, "040003e0 add z0.b, p0/m, z0.b, z31.b"},
+		{"add z0.s, p0/m, z0.s", "ADD", ExtFormPredicated,
+			[]ExtOperand{ExtVector(0, ExtArrS), ExtPredicate(0, ExtQualMerging), ExtVector(0, ExtArrS)},
+			0x04800000, "04800000 add z0.s, p0/m, z0.s, z0.s"},
+		{"sub z0.b, p0/m, z0.b", "SUB", ExtFormPredicated,
+			[]ExtOperand{ExtVector(0, ExtArrB), ExtPredicate(0, ExtQualMerging), ExtVector(0, ExtArrB)},
+			0x04010000, "04010000 sub z0.b, p0/m, z0.b, z0.b"},
+		{"sub z0.b, p7/m, z0.b", "SUB", ExtFormPredicated,
+			[]ExtOperand{ExtVector(0, ExtArrB), ExtPredicate(7, ExtQualMerging), ExtVector(0, ExtArrB)},
+			0x04011c00, "04011c00 sub z0.b, p7/m, z0.b, z0.b"},
+		{"sub z3.b, p0/m, z3.b", "SUB", ExtFormPredicated,
+			[]ExtOperand{ExtVector(0, ExtArrB), ExtPredicate(0, ExtQualMerging), ExtVector(3, ExtArrB)},
+			0x04010003, "04010003 sub z3.b, p0/m, z3.b, z0.b"},
+		{"subr z0.b, p0/m, z0.b", "SUBR", ExtFormPredicated,
+			[]ExtOperand{ExtVector(0, ExtArrB), ExtPredicate(0, ExtQualMerging), ExtVector(0, ExtArrB)},
+			0x04030000, "04030000 subr z0.b, p0/m, z0.b, z0.b"},
+		{"subr z0.h, p0/m, z0.h", "SUBR", ExtFormPredicated,
+			[]ExtOperand{ExtVector(0, ExtArrH), ExtPredicate(0, ExtQualMerging), ExtVector(0, ExtArrH)},
+			0x04430000, "04430000 subr z0.h, p0/m, z0.h, z0.h"},
+		{"mul z0.b, p0/m, z0.b", "MUL", ExtFormPredicated,
+			[]ExtOperand{ExtVector(0, ExtArrB), ExtPredicate(0, ExtQualMerging), ExtVector(0, ExtArrB)},
+			0x04100000, "04100000 mul z0.b, p0/m, z0.b, z0.b"},
+		{"mul z0.b, p2/m, z0.b", "MUL", ExtFormPredicated,
+			[]ExtOperand{ExtVector(0, ExtArrB), ExtPredicate(2, ExtQualMerging), ExtVector(0, ExtArrB)},
+			0x04100800, "04100800 mul z0.b, p2/m, z0.b, z0.b"},
+		{"mul z0.b, p0/m, z0.b (zm 31)", "MUL", ExtFormPredicated,
+			[]ExtOperand{ExtVector(31, ExtArrB), ExtPredicate(0, ExtQualMerging), ExtVector(0, ExtArrB)},
+			0x041003e0, "041003e0 mul z0.b, p0/m, z0.b, z31.b"},
+		{"smulh z0.b, p0/m, z0.b", "SMULH", ExtFormPredicated,
+			[]ExtOperand{ExtVector(0, ExtArrB), ExtPredicate(0, ExtQualMerging), ExtVector(0, ExtArrB)},
+			0x04120000, "04120000 smulh z0.b, p0/m, z0.b, z0.b"},
+		{"smulh z0.b, p2/m, z0.b", "SMULH", ExtFormPredicated,
+			[]ExtOperand{ExtVector(0, ExtArrB), ExtPredicate(2, ExtQualMerging), ExtVector(0, ExtArrB)},
+			0x04120800, "04120800 smulh z0.b, p2/m, z0.b, z0.b"},
+		{"smulh z0.s, p0/m, z0.s", "SMULH", ExtFormPredicated,
+			[]ExtOperand{ExtVector(0, ExtArrS), ExtPredicate(0, ExtQualMerging), ExtVector(0, ExtArrS)},
+			0x04920000, "04920000 smulh z0.s, p0/m, z0.s, z0.s"},
+		{"umulh z0.b, p0/m, z0.b", "UMULH", ExtFormPredicated,
+			[]ExtOperand{ExtVector(0, ExtArrB), ExtPredicate(0, ExtQualMerging), ExtVector(0, ExtArrB)},
+			0x04130000, "04130000 umulh z0.b, p0/m, z0.b, z0.b"},
+
+		// Immediate forms: imm{, LSL #8}, Zdn.
+		{"add z0.b, z0.b, #0", "ADD", ExtFormImmediate,
+			[]ExtOperand{ExtImmediate(0), ExtVector(0, ExtArrB)},
+			0x2520c000, "2520c000 add z0.b, z0.b, #0"},
+		{"add z0.b, z0.b, #127", "ADD", ExtFormImmediate,
+			[]ExtOperand{ExtImmediate(127), ExtVector(0, ExtArrB)},
+			0x2520cfe0, "2520cfe0 add z0.b, z0.b, #127"},
+		{"add z0.h, z0.h, #0, lsl #8", "ADD", ExtFormImmediate,
+			[]ExtOperand{ExtShiftedImmediate(0, 8), ExtVector(0, ExtArrH)},
+			0x2560e000, "2560e000 add z0.h, z0.h, #0, lsl #8"},
+		{"add z0.h, z0.h, #32512 (derived shift)", "ADD", ExtFormImmediate,
+			[]ExtOperand{ExtImmediate(32512), ExtVector(0, ExtArrH)},
+			0x2560efe0, "2560efe0 is sqsub's GNU word for #32512; the classes share the encoding"},
+		{"sub z0.b, z0.b, #0", "SUB", ExtFormImmediate,
+			[]ExtOperand{ExtImmediate(0), ExtVector(0, ExtArrB)},
+			0x2521c000, "2521c000 sub z0.b, z0.b, #0"},
+		{"subr z0.b, z0.b, #0", "SUBR", ExtFormImmediate,
+			[]ExtOperand{ExtImmediate(0), ExtVector(0, ExtArrB)},
+			0x2523c000, "2523c000 subr z0.b, z0.b, #0"},
+		{"subr z0.b, z0.b, #255", "SUBR", ExtFormImmediate,
+			[]ExtOperand{ExtImmediate(255), ExtVector(0, ExtArrB)},
+			0x2523dfe0, "2523dfe0 subr z0.b, z0.b, #255"},
+		{"subr z0.h, z0.h, #0, lsl #8", "SUBR", ExtFormImmediate,
+			[]ExtOperand{ExtShiftedImmediate(0, 8), ExtVector(0, ExtArrH)},
+			0x2563e000, "2563e000 subr z0.h, z0.h, #0, lsl #8"},
+		{"sqadd z0.b, z0.b, #0", "SQADD", ExtFormImmediate,
+			[]ExtOperand{ExtImmediate(0), ExtVector(0, ExtArrB)},
+			0x2524c000, "2524c000 sqadd z0.b, z0.b, #0"},
+		{"uqadd z0.b, z0.b, #0", "UQADD", ExtFormImmediate,
+			[]ExtOperand{ExtImmediate(0), ExtVector(0, ExtArrB)},
+			0x2525c000, "2525c000 uqadd z0.b, z0.b, #0"},
+		{"sqsub z0.b, z0.b, #0", "SQSUB", ExtFormImmediate,
+			[]ExtOperand{ExtImmediate(0), ExtVector(0, ExtArrB)},
+			0x2526c000, "2526c000 sqsub z0.b, z0.b, #0"},
+		{"sqsub z0.b, z0.b, #255", "SQSUB", ExtFormImmediate,
+			[]ExtOperand{ExtImmediate(255), ExtVector(0, ExtArrB)},
+			0x2526dfe0, "2526dfe0 sqsub z0.b, z0.b, #255"},
+		{"sqsub z0.h, z0.h, #0, lsl #8", "SQSUB", ExtFormImmediate,
+			[]ExtOperand{ExtShiftedImmediate(0, 8), ExtVector(0, ExtArrH)},
+			0x2566e000, "2566e000 sqsub z0.h, z0.h, #0, lsl #8"},
+		{"uqsub z0.b, z0.b, #0", "UQSUB", ExtFormImmediate,
+			[]ExtOperand{ExtImmediate(0), ExtVector(0, ExtArrB)},
+			0x2527c000, "2527c000 uqsub z0.b, z0.b, #0 (class vector from the GNU table and LLVM: sve_int_arith_imm0 opc 0b111)"},
+		{"mul z0.b, z0.b, #0", "MUL", ExtFormSignedImmediate,
+			[]ExtOperand{ExtImmediate(0), ExtVector(0, ExtArrB)},
+			0x2530c000, "2530c000 mul z0.b, z0.b, #0"},
+		{"mul z0.b, z0.b, #127", "MUL", ExtFormSignedImmediate,
+			[]ExtOperand{ExtImmediate(127), ExtVector(0, ExtArrB)},
+			0x2530cfe0, "2530cfe0 mul z0.b, z0.b, #127"},
+		{"mul z0.b, z0.b, #-128", "MUL", ExtFormSignedImmediate,
+			[]ExtOperand{ExtImmediate(-128), ExtVector(0, ExtArrB)},
+			0x2530d000, "2530d000 mul z0.b, z0.b, #-128"},
+		{"mul z0.b, z0.b, #-1", "MUL", ExtFormSignedImmediate,
+			[]ExtOperand{ExtImmediate(-1), ExtVector(0, ExtArrB)},
+			0x2530dfe0, "2530dfe0 mul z0.b, z0.b, #-1"},
+		{"mul z0.h, z0.h, #0", "MUL", ExtFormSignedImmediate,
+			[]ExtOperand{ExtImmediate(0), ExtVector(0, ExtArrH)},
+			0x2570c000, "2570c000 mul z0.h, z0.h, #0"},
+	} {
+		in := extInstruction(t, tt.mnem, tt.form)
+		got, err := in.Encode(tt.ops)
+		if err != nil {
+			t.Errorf("%s: encode: %v", tt.name, err)
+			continue
+		}
+		if want := hex.EncodeToString(extWordLE(tt.want)); hex.EncodeToString(got) != want {
+			t.Errorf("%s:\n  got  %x\n  want %s", tt.name, got, want)
+		}
+	}
+}
+
+// TestArm64ExtGoldenSources pins the cross-check contract: every encoding
+// class in the table carries at least one GNU-assembler vector, so no class
+// rests on transcription alone.
+func TestArm64ExtGoldenSources(t *testing.T) {
+	classes := map[ExtForm]bool{}
+	for _, tt := range []struct {
+		mnem string
+		form ExtForm
+	}{
+		{"ADD", ExtFormVectors}, {"SUB", ExtFormVectors}, {"SQADD", ExtFormVectors},
+		{"UQADD", ExtFormVectors}, {"SQSUB", ExtFormVectors}, {"UQSUB", ExtFormVectors},
+		{"MUL", ExtFormVectors}, {"SMULH", ExtFormVectors}, {"UMULH", ExtFormVectors},
+		{"ADD", ExtFormPredicated}, {"SUB", ExtFormPredicated}, {"SUBR", ExtFormPredicated},
+		{"MUL", ExtFormPredicated}, {"SMULH", ExtFormPredicated}, {"UMULH", ExtFormPredicated},
+		{"ADD", ExtFormImmediate}, {"SUB", ExtFormImmediate}, {"SUBR", ExtFormImmediate},
+		{"SQADD", ExtFormImmediate}, {"UQADD", ExtFormImmediate},
+		{"SQSUB", ExtFormImmediate}, {"UQSUB", ExtFormImmediate},
+		{"MUL", ExtFormSignedImmediate},
+	} {
+		if _, ok := extInstructionQuiet(tt.mnem, tt.form); !ok {
+			t.Errorf("the table lacks %s with the %s form", tt.mnem, tt.form)
+		}
+		classes[tt.form] = true
+	}
+	for _, form := range []ExtForm{ExtFormVectors, ExtFormPredicated, ExtFormImmediate, ExtFormSignedImmediate} {
+		if !classes[form] {
+			t.Errorf("no golden vectors cover the %s form", form)
+		}
+	}
+}
+
+func extInstructionQuiet(mnem string, form ExtForm) (ExtInstr, bool) {
+	for _, in := range Extensions(ARM64) {
+		if in.Name == mnem && in.Form == form {
+			return in, true
+		}
+	}
+	return ExtInstr{}, false
+}
+
+// TestArm64ExtTableIntegrity checks the metadata contract: every entry names
+// its manual reference, summary and feature, and the element-size field sits
+// at bits 23..22 where the manual puts it for every class in the family.
+func TestArm64ExtTableIntegrity(t *testing.T) {
+	for _, in := range Extensions(ARM64) {
+		if in.Name == "" || in.Summary == "" || in.Ref == "" {
+			t.Errorf("%+v: name, summary and reference are mandatory", in)
+		}
+		if in.Feature != ExtFeatureSVE && in.Feature != ExtFeatureSVE2 {
+			t.Errorf("%s: feature %q is neither sve nor sve2", in.Name, in.Feature)
+		}
+		if in.Form.Arity() < 2 || in.Form.Arity() > 3 {
+			t.Errorf("%s: form %d carries an unusable arity %d", in.Name, in.Form, in.Form.Arity())
+		}
+		if in.Size.Off != 22 || in.Size.Width != 2 {
+			t.Errorf("%s: the size field sits at bits %d..%d, the classes here put it at 23..22",
+				in.Name, in.Size.Off, in.Size.Off+in.Size.Width-1)
+		}
+		// The destination register field and the element-size field are
+		// operands everywhere in this family, so Word carries both zero; the
+		// class opcodes live around them and stay where they are.
+		if in.Word&0x1f != 0 || in.Word&(0x3<<22) != 0 {
+			t.Errorf("%s: word %08x carries destination or size bits, want them zero", in.Name, in.Word)
+		}
+	}
+}
+
+func TestArm64ExtRejects(t *testing.T) {
+	rgb := func(rs ...int) []ExtOperand {
+		ops := make([]ExtOperand, len(rs))
+		for i, r := range rs {
+			ops[i] = ExtVector(r, ExtArrB)
+		}
+		return ops
+	}
+	for _, tt := range []struct {
+		name  string
+		mnem  string
+		form  ExtForm
+		ops   []ExtOperand
+		quote string // a fragment the error carries
+	}{
+		{"no arrangement", "ADD", ExtFormVectors,
+			[]ExtOperand{{Kind: ExtZReg, Reg: 0}, ExtVector(0, ExtArrB), ExtVector(0, ExtArrB)},
+			"no arrangement"},
+		{"mismatched arrangements", "ADD", ExtFormVectors,
+			[]ExtOperand{ExtVector(0, ExtArrB), ExtVector(0, ExtArrS), ExtVector(0, ExtArrB)},
+			"want .B"},
+		{"wrong arity", "ADD", ExtFormVectors, rgb(0, 0), "takes 3 operands"},
+		{"predicate in a vector position", "ADD", ExtFormVectors,
+			[]ExtOperand{ExtPredicate(0, ExtQualNone), ExtVector(0, ExtArrB), ExtVector(0, ExtArrB)},
+			"scalable vector register"},
+		{"quadword arrangement has no size encoding", "ADD", ExtFormVectors,
+			[]ExtOperand{ExtVector(0, ExtArrQ), ExtVector(0, ExtArrQ), ExtVector(0, ExtArrQ)},
+			"no size encoding"},
+		{"zeroing qualifier", "ADD", ExtFormPredicated,
+			[]ExtOperand{ExtVector(0, ExtArrB), ExtPredicate(0, ExtQualZeroing), ExtVector(0, ExtArrB)},
+			"/M"},
+		{"predicate beyond the 3-bit field", "ADD", ExtFormPredicated,
+			[]ExtOperand{ExtVector(0, ExtArrB), ExtPredicate(8, ExtQualMerging), ExtVector(0, ExtArrB)},
+			"P0-P7"},
+		{"predicate arrangement suffix", "ADD", ExtFormPredicated,
+			[]ExtOperand{ExtVector(0, ExtArrB), ExtOperand{Kind: ExtPReg, Reg: 0, Qual: ExtQualMerging, Arr: ExtArrB}, ExtVector(0, ExtArrB)},
+			"arrangement"},
+		{"predicate operands disagree on arrangement", "ADD", ExtFormPredicated,
+			[]ExtOperand{ExtVector(0, ExtArrB), ExtPredicate(0, ExtQualMerging), ExtVector(0, ExtArrS)},
+			"must match"},
+		{"bare 256 on .B", "ADD", ExtFormImmediate,
+			[]ExtOperand{ExtImmediate(256), ExtVector(0, ExtArrB)},
+			"immediate 256"},
+		{"negative unsigned immediate", "ADD", ExtFormImmediate,
+			[]ExtOperand{ExtImmediate(-1), ExtVector(0, ExtArrB)},
+			"immediate -1"},
+		{"multiple of 256 beyond the imm8 span", "ADD", ExtFormImmediate,
+			[]ExtOperand{ExtImmediate(65536), ExtVector(0, ExtArrH)},
+			"immediate 65536"},
+		{"shift amount other than 0 or 8", "ADD", ExtFormImmediate,
+			[]ExtOperand{ExtShiftedImmediate(1, 4), ExtVector(0, ExtArrS)},
+			"0 or 8"},
+		{"shifted constant on .B", "ADD", ExtFormImmediate,
+			[]ExtOperand{ExtShiftedImmediate(1, 8), ExtVector(0, ExtArrB)},
+			".B takes no shift"},
+		{"register where the immediate belongs", "ADD", ExtFormImmediate,
+			[]ExtOperand{ExtVector(0, ExtArrB), ExtVector(0, ExtArrB)},
+			"wants an immediate"},
+		{"signed immediate over the top", "MUL", ExtFormSignedImmediate,
+			[]ExtOperand{ExtImmediate(128), ExtVector(0, ExtArrB)},
+			"128"},
+		{"signed immediate under the floor", "MUL", ExtFormSignedImmediate,
+			[]ExtOperand{ExtImmediate(-129), ExtVector(0, ExtArrB)},
+			"-129"},
+		{"shift in the signed class", "MUL", ExtFormSignedImmediate,
+			[]ExtOperand{ExtShiftedImmediate(1, 8), ExtVector(0, ExtArrB)},
+			"no shift"},
+	} {
+		in := extInstruction(t, tt.mnem, tt.form)
+		_, err := in.Encode(tt.ops)
+		if err == nil {
+			t.Errorf("%s: encode succeeded, want an error", tt.name)
+			continue
+		}
+		if !strings.Contains(err.Error(), tt.quote) {
+			t.Errorf("%s: error %q lacks %q", tt.name, err, tt.quote)
+		}
+	}
+}
+
+// TestExtensionsArchBinding pins the registry's architecture binding: the
+// extended layer exists for arm64 alone until an amd64 table attaches, and no
+// other architecture sees a single SVE instruction.
+func TestExtensionsArchBinding(t *testing.T) {
+	for _, a := range []Arch{AMD64, RISCV, LOONG64, Unknown} {
+		if got := Extensions(a); len(got) != 0 {
+			t.Errorf("Extensions(%s) carries %d instructions, want none", a, len(got))
+		}
+	}
+	if got := Extensions(ARM64); len(got) == 0 {
+		t.Error("Extensions(ARM64) is empty")
+	}
+}

arch/arm64_gen.go

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

arch/common_gen.go

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

arch/loong64_gen.go

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

arch/riscv_gen.go

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

asm/aarch64_goobj_test.go

@@ -11,7 +11,7 @@ import (
 	"strings"
 	"testing"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
 )
 
 // TestGOObjectAARCH64Structure checks the basic structure of the emitted
@@ -178,26 +178,10 @@ func main() {
 	if err != nil {
 		t.Fatalf("baseline build: %v\n%s", err, buildLog)
 	}
-	var work, linkLine, asmObj string
-	for line := range strings.SplitSeq(string(buildLog), "\n") {
-		switch {
-		case strings.HasPrefix(line, "WORK="):
-			work = strings.TrimPrefix(line, "WORK=")
-		case strings.Contains(line, "/asm ") && strings.Contains(line, "main_arm64.s") && !strings.Contains(line, "-gensymabis"):
-			asmObj = fieldAfter(line, "-o")
-		case strings.Contains(line, "/link ") && strings.Contains(line, "-importcfg"):
-			linkLine = line
-		}
-	}
-	if work == "" || asmObj == "" {
-		t.Skipf("could not parse build log (work=%q asmObj=%q)", work, asmObj)
-	}
-	defer os.RemoveAll(work)
+	st := parseBuildLog(t, buildLog, "main_arm64.s")
+	defer os.RemoveAll(st.work)
 
-	// Expand $WORK in the object path.
-	asmObj = strings.ReplaceAll(asmObj, "$WORK", work)
-
-	// Read the toolchain-produced object and assemble the same source with gasm.
+	// Assemble the same source with gasm and substitute the object.
 	src, err := os.ReadFile(filepath.Join(dir, "main_arm64.s"))
 	if err != nil {
 		t.Fatal(err)
@@ -210,30 +194,16 @@ func main() {
 	if err != nil {
 		t.Fatalf("AssembleFileARM64: %v", err)
 	}
-	gasmObj, err := img.GOObjectAARCH64("a64link", "main_arm64.s")
+	// The package path is "main", the prefix the Go code's references carry.
+	gasmObj, err := img.GOObjectAARCH64("main", "main_arm64.s")
 	if err != nil {
 		t.Fatalf("GOObjectAARCH64: %v", err)
 	}
-
-	// Replace the toolchain-produced object with gasm's.
-	if err := os.WriteFile(asmObj, gasmObj, 0o644); err != nil {
-		t.Fatalf("write gasm object: %v", err)
-	}
-
-	// Re-link.
-	if linkLine == "" {
-		t.Skip("could not find link command in build log")
-	}
-	// Expand $WORK in the link command.
-	linkLine = strings.ReplaceAll(linkLine, "$WORK", work)
-	linkCmd := exec.Command("bash", "-c", "cd "+dir+" && "+linkLine)
-	linkCmd.Env = append(os.Environ(), "GOARCH=arm64")
-	if out, err := linkCmd.CombinedOutput(); err != nil {
-		t.Fatalf("re-link with gasm object: %v\n%s", err, out)
-	}
+	substituteAndRelink(t, goBin, dir, st, filepath.Join(dir, "prog2"),
+		gasmObj, "GOARCH=arm64")
 
 	// Verify the binary exists and contains the symbol.
-	binPath := filepath.Join(dir, "prog")
+	binPath := filepath.Join(dir, "prog2")
 	if _, err := os.Stat(binPath); err != nil {
 		t.Fatalf("binary not found: %v", err)
 	}
@@ -296,23 +266,8 @@ func main() {
 	if err != nil {
 		t.Fatalf("baseline build: %v\n%s", err, buildLog)
 	}
-	var work, linkLine, asmObj string
-	for line := range strings.SplitSeq(string(buildLog), "\n") {
-		switch {
-		case strings.HasPrefix(line, "WORK="):
-			work = strings.TrimPrefix(line, "WORK=")
-		case strings.Contains(line, "/asm ") && strings.Contains(line, "main_arm64.s") && !strings.Contains(line, "-gensymabis"):
-			asmObj = fieldAfter(line, "-o")
-		case strings.Contains(line, "/link ") && strings.Contains(line, "-importcfg"):
-			linkLine = line
-		}
-	}
-	if work == "" || asmObj == "" || linkLine == "" {
-		t.Skipf("could not parse build log (work=%q asmObj=%q link=%q)", work, asmObj, linkLine)
-	}
-	defer os.RemoveAll(work)
-	asmObj = strings.ReplaceAll(asmObj, "$WORK", work)
-	linkLine = strings.ReplaceAll(linkLine, "$WORK", work)
+	st := parseBuildLog(t, buildLog, "main_arm64.s")
+	defer os.RemoveAll(st.work)
 
 	src, err := os.ReadFile(filepath.Join(dir, "main_arm64.s"))
 	if err != nil {
@@ -326,19 +281,13 @@ func main() {
 	if err != nil {
 		t.Fatalf("AssembleFileARM64: %v", err)
 	}
-	gasmObj, err := img.GOObjectAARCH64("a64dlink", "main_arm64.s")
+	gasmObj, err := img.GOObjectAARCH64("main", "main_arm64.s")
 	if err != nil {
 		t.Fatalf("GOObjectAARCH64: %v", err)
 	}
-	if err := os.WriteFile(asmObj, gasmObj, 0o644); err != nil {
-		t.Fatalf("write gasm object: %v", err)
-	}
-	linkCmd := exec.Command("bash", "-c", "cd "+dir+" && "+linkLine)
-	linkCmd.Env = append(os.Environ(), "GOARCH=arm64")
-	if out, err := linkCmd.CombinedOutput(); err != nil {
-		t.Fatalf("re-link with gasm object: %v\n%s", err, out)
-	}
-	binData, err := os.ReadFile(filepath.Join(dir, "prog"))
+	substituteAndRelink(t, goBin, dir, st, filepath.Join(dir, "prog2"),
+		gasmObj, "GOARCH=arm64")
+	binData, err := os.ReadFile(filepath.Join(dir, "prog2"))
 	if err != nil {
 		t.Fatal(err)
 	}

asm/arm64_assemble.go

@@ -9,7 +9,7 @@ import (
 	"strconv"
 	"strings"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/ast"
+	"sourcedock.dev/petrbalvin/gasm-sdk/ast"
 )
 
 // assembleARM64 assembles an AArch64 (arm64) TEXT function body into machine
@@ -458,10 +458,13 @@ func encodeARM64Instr(instr *ast.Instr, pc int, offsets map[string]int, fi arm64
 		return encodeARM64Excl(mnem, enc.op, ops)
 	}
 
-	// LSE atomics (LDADD, CAS, SWP).
+	// LSE atomics (LDADD, CAS, SWP) and the compare-and-swap pair.
 	if enc, ok := a64InstrTable[mnem]; ok && enc.format == a64FLSE {
 		return encodeARM64LSEAtom(mnem, enc.op, ops)
 	}
+	if enc, ok := a64InstrTable[mnem]; ok && enc.format == a64FCASP {
+		return encodeARM64CASP(mnem, enc.op, ops)
+	}
 
 	// Bitfield/shift (ASR, LSL, LSR, ROR, BFI, BFXIL, SBFM, UBFM).
 	if enc, ok := a64InstrTable[mnem]; ok && enc.format == a64FBitfield {
@@ -555,16 +558,26 @@ func encodeARM64Instr(instr *ast.Instr, pc int, offsets map[string]int, fi arm64
 	// VPMULL, VRAX1 and friends).  VADD, VSUB and VMUL appear here too, so
 	// this check precedes the plain SIMD3 path below.  VCMLE and VCMLT exist
 	// only in the zero-immediate form (a64SimdVZero), so they route here with
-	// an empty register-form spec.
+	// an empty register-form spec.  VSQSHL/VUQSHL keep their shift-by-
+	// immediate route when the first operand is an immediate.
 	if spec, ok := a64SimdVTable[mnem]; ok || a64SimdVZero[mnem] != 0 {
+		if !arm64SimdShiftImmRoute(mnem, ops) {
 			return encodeARM64SimdV(mnem, spec, ops)
 		}
+	}
 
 	// Arrangement-aware SIMD two-register (VREV32, VREV64, VUADDLV, VMOV).
 	if spec, ok := a64SimdV2Table[mnem]; ok {
 		return encodeARM64SimdV2(mnem, spec, ops)
 	}
 
+	// Narrow/long/wide SIMD families whose size and Q bits read off one
+	// designated operand (VXTN, VSXTL, VUADDW, VUMULL, VSHRN, VSSHLL, VFCVTN
+	// and friends).
+	if spec, ok := a64SimdNLTable[mnem]; ok {
+		return encodeARM64SimdNL(mnem, spec, ops)
+	}
+
 	// SIMD four-register and immediate three-register (VEOR3, VBCAX, VXAR,
 	// VEXT).
 	if enc, ok := a64InstrTable[mnem]; ok && enc.format == a64FSIMDV4 {
@@ -587,6 +600,11 @@ func encodeARM64Instr(instr *ast.Instr, pc int, offsets map[string]int, fi arm64
 		return encodeARM64ShiftImm(mnem, enc.op, ops)
 	}
 
+	// SIMD move immediate: VMOVI $imm8, Vd.B8/B16.
+	if enc, ok := a64InstrTable[mnem]; ok && enc.format == a64FVMoviImm {
+		return encodeARM64MoviImm(ops)
+	}
+
 	// VMOVS/VMOVD/VMOVQ with a large constant: a literal pool load.
 	if enc, ok := a64InstrTable[mnem]; ok && enc.format == a64FMoviLit {
 		return encodeARM64MoviLit(mnem, enc.op, ops, relocs, lits)
@@ -637,6 +655,20 @@ func encodeARM64Branch(mnem string, ops []*ast.Operand, pc int, offsets map[stri
 		return a64wordLE(a64UncondBranch(opc, uint32(rn), 0)), nil
 	}
 
+	// The bare spelling BL R9 is the same indirect branch: the parser reads
+	// a bare identifier as a symbol, and one named for a register is an
+	// indirect branch through it, which the toolchain accepts alongside the
+	// parenthesised form (BL (R3) and BL R3 both encode BLR R3).
+	if op.Addr.Sym != nil && op.Addr.Sym.Pseudo == "" && op.Addr.Base == "" && op.Addr.Index == "" {
+		if rn := arm64RegNum(op.Addr.Sym.Name); rn >= 0 {
+			opc := uint32(0) // BR
+			if link {
+				opc = 1 // BLR
+			}
+			return a64wordLE(a64UncondBranch(opc, uint32(rn), 0)), nil
+		}
+	}
+
 	// Symbol reference: BL sym(SB), or B sym(SB) for a tail call, against a
 	// relocation (R_CALLARM64 either way).
 	if op.Addr.Sym != nil && op.Addr.Sym.Pseudo == "SB" {
@@ -1454,6 +1486,15 @@ func encodeARM64Mov(instr *ast.Instr, mnem string, wb string, fi arm64FrameInfo,
 			}
 			return encodeARM64SBAddr(src.Imm.Sym, rd, relocs), nil
 		}
+		// Immediate → memory: only storing zero is encodable (the ZR
+		// register); the toolchain rejects any other immediate-to-memory
+		// combination ("illegal combination").
+		if isMemOperand(dst) {
+			if arm64Imm64(src) != 0 {
+				return nil, fmt.Errorf("%s: illegal combination: an immediate store must be zero", mnem)
+			}
+			return encodeARM64MemOp(mnem, dst, 31, false, fi, "")
+		}
 		rd := arm64RegNum(operandRegName(dst))
 		if rd < 0 {
 			return nil, fmt.Errorf("%s $imm: invalid destination register", mnem)
@@ -2545,6 +2586,263 @@ func encodeARM64LSEAtom(mnem string, baseOp uint32, ops []*ast.Operand) ([]byte,
 	return a64wordLE(baseOp | uint32(rs)<<16 | uint32(rn)<<5 | uint32(rt)), nil
 }
 
+// encodeARM64CASP encodes the compare-and-swap pair: CASP (Rs, Rs+1), (Rn),
+// (Rt, Rt+1).  Both pairs must start on an even register and be contiguous;
+// the second register of each pair rides no encoding field.
+func encodeARM64CASP(mnem string, baseOp uint32, ops []*ast.Operand) ([]byte, error) {
+	if len(ops) != 3 {
+		return nil, fmt.Errorf("%s expects (Rs, Rs+1), (Rn), (Rt, Rt+1), got %d operands", mnem, len(ops))
+	}
+	rs, rs1, ok := arm64PairOf(ops[0])
+	if !ok {
+		return nil, fmt.Errorf("%s expects a source register pair (Rs, Rs+1)", mnem)
+	}
+	rn, err := arm64ExclMem(mnem, ops[1])
+	if err != nil {
+		return nil, err
+	}
+	rt, rt1, ok := arm64PairOf(ops[2])
+	if !ok {
+		return nil, fmt.Errorf("%s expects a destination register pair (Rt, Rt+1)", mnem)
+	}
+	if rs&1 != 0 {
+		return nil, fmt.Errorf("%s: source register pair must start from an even register", mnem)
+	}
+	if rt&1 != 0 {
+		return nil, fmt.Errorf("%s: destination register pair must start from an even register", mnem)
+	}
+	if rs != rs1-1 {
+		return nil, fmt.Errorf("%s: source register pair must be contiguous", mnem)
+	}
+	if rt != rt1-1 {
+		return nil, fmt.Errorf("%s: destination register pair must be contiguous", mnem)
+	}
+	if rt == 31 {
+		return nil, fmt.Errorf("%s: illegal destination register", mnem)
+	}
+	return a64wordLE(baseOp | uint32(rs)<<16 | uint32(rn)<<5 | uint32(rt)), nil
+}
+
+// encodeARM64MoviImm encodes VMOVI $imm8, Vd.B8/B16: the modified-immediate
+// form of the SIMD move (asm7.go case 86).  Only the byte arrangements exist
+// and the immediate is one unsigned byte.
+func encodeARM64MoviImm(ops []*ast.Operand) ([]byte, error) {
+	if len(ops) != 2 || !isImmOperand(ops[0]) {
+		return nil, fmt.Errorf("VMOVI expects $immediate, Vd.<T>")
+	}
+	vd, ok := arm64VecOf(ops[1])
+	if !ok || vd.hasIdx || (vd.arr != "B8" && vd.arr != "B16") {
+		return nil, fmt.Errorf("VMOVI: destination arrangement must be B8 or B16")
+	}
+	imm := arm64Imm64(ops[0])
+	if imm < 0 || imm > 255 {
+		return nil, fmt.Errorf("VMOVI: immediate constant %d out of range (0..255)", imm)
+	}
+	q := uint32(0)
+	if vd.arr == "B16" {
+		q = 1 << 30
+	}
+	w := 0x0f00e400 | q | uint32(imm>>5&7)<<16 | uint32(imm&0x1f)<<5 | uint32(vd.reg)
+	return a64wordLE(w), nil
+}
+
+// arm64SimdShiftImmRoute reports whether a mnemonic carries both a shift-by-
+// immediate and a register form and the operands spell the immediate one: the
+// dedicated shift route keeps them.
+func arm64SimdShiftImmRoute(mnem string, ops []*ast.Operand) bool {
+	if mnem != "VSQSHL" && mnem != "VUQSHL" {
+		return false
+	}
+	return len(ops) > 0 && isImmOperand(ops[0])
+}
+
+// arm64SimdNLArr describes one arrangement for the narrow/long/wide families:
+// the element width in bytes and whether the spelling names the 128-bit form.
+func arm64SimdNLArr(arr string) (esize int, wide bool, ok bool) {
+	switch arr {
+	case "B8", "B16":
+		return 1, arr == "B16", true
+	case "H4", "H8":
+		return 2, arr == "H8", true
+	case "S2", "S4":
+		return 4, arr == "S4", true
+	case "D1", "D2":
+		return 8, arr == "D2", true
+	}
+	return 0, false, false
+}
+
+// arm64SimdLongPair validates a long pairing (source narrow, destination
+// wide): the destination element is twice the source's, the destination is
+// always spelled the wide way (H8/S4/D2) and the source carries the 128-bit
+// flag exactly for the .2 spellings.
+func arm64SimdLongPair(mnem, src, dst string, two bool) error {
+	se, sw, ok1 := arm64SimdNLArr(src)
+	de, dw, ok2 := arm64SimdNLArr(dst)
+	if !ok1 || !ok2 || de != 2*se {
+		return fmt.Errorf("%s: incompatible arrangements %s and %s", mnem, src, dst)
+	}
+	if !dw || sw != two {
+		return fmt.Errorf("%s: operand mismatch for the %s spelling", mnem, mnem)
+	}
+	return nil
+}
+
+// arm64SimdNarrowPair validates a narrow pairing (source wide, destination
+// narrow): the mirror image of arm64SimdLongPair.
+func arm64SimdNarrowPair(mnem, src, dst string, two bool) error {
+	se, sw, ok1 := arm64SimdNLArr(src)
+	de, dw, ok2 := arm64SimdNLArr(dst)
+	if !ok1 || !ok2 || se != 2*de {
+		return fmt.Errorf("%s: incompatible arrangements %s and %s", mnem, src, dst)
+	}
+	if !sw || dw != two {
+		return fmt.Errorf("%s: operand mismatch for the %s spelling", mnem, mnem)
+	}
+	return nil
+}
+
+// arm64SimdNLArrBits returns the arrangement bits a narrow/long/wide
+// instruction contributes: the driving arrangement's size and Q bits, or for
+// the FCVT family only the Q bit, whose size field is fixed in the base.
+func arm64SimdNLArrBits(spec a64SimdNLSpec, drive string, two bool) uint32 {
+	if spec.qonly {
+		if two {
+			return 1 << 30
+		}
+		return 0
+	}
+	return a64ArrBits[a64ArrIndex(drive)]
+}
+
+// encodeARM64SimdNL encodes the narrow/long/wide SIMD families
+// (a64SimdNLTable): XTN and FCVTN narrow a wide source, SXTL and FCVTL
+// lengthen, the MULL/MLAL/MLSL group multiplies long, UADDW widens, and the
+// SSHLL/USHLL and SHRN shifts carry their immediate in the immh:immb field.
+// The size and Q bits read off the designated driving operand, and the .2
+// spellings force the 128-bit side through their own arrangement.
+func encodeARM64SimdNL(mnem string, spec a64SimdNLSpec, ops []*ast.Operand) ([]byte, error) {
+	two := strings.HasSuffix(mnem, "2")
+
+	// vecAt parses operand i as a vector register with an arrangement.
+	vecAt := func(i int) (a64Vec, bool) {
+		if i >= len(ops) {
+			return a64Vec{}, false
+		}
+		v, ok := arm64VecOf(ops[i])
+		if !ok || v.hasIdx {
+			return a64Vec{}, false
+		}
+		return v, true
+	}
+
+	switch spec.form {
+	case a64NLTwoNarrow, a64NLTwoLong:
+		if len(ops) != 2 {
+			return nil, fmt.Errorf("%s expects 2 operands, got %d", mnem, len(ops))
+		}
+		vn, ok1 := vecAt(0)
+		vd, ok2 := vecAt(1)
+		if !ok1 || !ok2 {
+			return nil, fmt.Errorf("invalid register operand in %s", mnem)
+		}
+		var drive string
+		var pairErr error
+		if spec.form == a64NLTwoNarrow {
+			// XTN/FCVTN: wide source into a narrow destination; the
+			// arrangement bits follow the destination.
+			drive, pairErr = vd.arr, arm64SimdNarrowPair(mnem, vn.arr, vd.arr, two)
+		} else {
+			// SXTL/UXTL/FCVTL: narrow source into a wide destination; the
+			// arrangement bits follow the source.
+			drive, pairErr = vn.arr, arm64SimdLongPair(mnem, vn.arr, vd.arr, two)
+		}
+		if pairErr != nil {
+			return nil, pairErr
+		}
+		arrBits := arm64SimdNLArrBits(spec, drive, two)
+		return a64wordLE(spec.base | arrBits | uint32(vn.reg)<<5 | uint32(vd.reg)), nil
+	case a64NLThreeLongMul, a64NLThreeWide:
+		if len(ops) != 3 {
+			return nil, fmt.Errorf("%s expects 3 operands, got %d", mnem, len(ops))
+		}
+		vm, ok1 := vecAt(0)
+		vn, ok2 := vecAt(1)
+		vd, ok3 := vecAt(2)
+		if !ok1 || !ok2 || !ok3 {
+			return nil, fmt.Errorf("invalid register operand in %s", mnem)
+		}
+		var drive string
+		var pairErr error
+		if spec.form == a64NLThreeWide {
+			// UADDW: Vn and Vd spell the wide arrangement, Vm the narrow one;
+			// the arrangement bits follow the wide side.
+			drive, pairErr = vn.arr, arm64SimdLongPair(mnem, vm.arr, vn.arr, two)
+		} else {
+			// MULL/MLAL/MLSL: Vm and Vn spell the narrow arrangement, Vd the
+			// wide one; the arrangement bits follow the narrow source.
+			drive, pairErr = vn.arr, arm64SimdLongPair(mnem, vn.arr, vd.arr, two)
+		}
+		if pairErr != nil {
+			return nil, pairErr
+		}
+		if spec.form == a64NLThreeWide && vd.arr != vn.arr {
+			return nil, fmt.Errorf("%s: operand mismatch: %s and %s", mnem, vn.arr, vd.arr)
+		}
+		if spec.form == a64NLThreeLongMul && vm.arr != vn.arr {
+			return nil, fmt.Errorf("%s: operand mismatch: %s and %s", mnem, vm.arr, vn.arr)
+		}
+		arrBits := arm64SimdNLArrBits(spec, drive, two)
+		return a64wordLE(spec.base | arrBits | uint32(vm.reg)<<16 | uint32(vn.reg)<<5 | uint32(vd.reg)), nil
+	case a64NLThreeLongShift, a64NLThreeNarrowShift:
+		if len(ops) != 3 || !isImmOperand(ops[0]) {
+			return nil, fmt.Errorf("%s expects ($shift, Vn.<T>, Vd.<T>)", mnem)
+		}
+		sh := arm64Imm64(ops[0])
+		vn, ok1 := vecAt(1)
+		vd, ok2 := vecAt(2)
+		if !ok1 || !ok2 {
+			return nil, fmt.Errorf("invalid register operand in %s", mnem)
+		}
+		if spec.form == a64NLThreeLongShift {
+			// SSHLL/USHLL: the narrow source drives the immediate's size
+			// (immh:immb = esize + shift), so the arrangement bits carry
+			// the Q bit alone: the size field belongs to immh, and ORing
+			// the source's size bits into it would collide with immb.
+			if err := arm64SimdLongPair(mnem, vn.arr, vd.arr, two); err != nil {
+				return nil, err
+			}
+			se, _, _ := arm64SimdNLArr(vn.arr)
+			esize := se * 8
+			if sh < 0 || sh >= int64(esize) {
+				return nil, fmt.Errorf("%s: shift %d out of range (0..%d)", mnem, sh, esize-1)
+			}
+			var qBit uint32
+			if two {
+				qBit = 1 << 30
+			}
+			return a64wordLE(spec.base | uint32(esize+int(sh))<<16 | qBit | uint32(vn.reg)<<5 | uint32(vd.reg)), nil
+		}
+		// SHRN: the narrow destination drives the immediate's size
+		// (immh:immb = esize - shift over the wide source element), so the
+		// arrangement bits carry the Q bit alone, exactly as above.
+		if err := arm64SimdNarrowPair(mnem, vn.arr, vd.arr, two); err != nil {
+			return nil, err
+		}
+		se, _, _ := arm64SimdNLArr(vn.arr)
+		esize := se * 8
+		if sh < 1 || sh >= int64(esize) {
+			return nil, fmt.Errorf("%s: shift %d out of range (1..%d)", mnem, sh, esize-1)
+		}
+		var qBit uint32
+		if two {
+			qBit = 1 << 30
+		}
+		return a64wordLE(spec.base | uint32(esize-int(sh))<<16 | qBit | uint32(vn.reg)<<5 | uint32(vd.reg)), nil
+	}
+	return nil, fmt.Errorf("unsupported arm64 instruction %q", mnem)
+}
+
 // encodeARM64DP1 encodes a data-processing (1 source) instruction:
 // RBIT, REV, CLZ and friends take (Rn, Rd), word = base | Rn<<5 | Rd.
 func encodeARM64DP1(mnem string, baseOp uint32, ops []*ast.Operand) ([]byte, error) {
@@ -2561,7 +2859,8 @@ func encodeARM64DP1(mnem string, baseOp uint32, ops []*ast.Operand) ([]byte, err
 
 // encodeARM64ADR encodes ADR/ADRP: (label, Rd), the byte distance to the
 // target split into immlo (bits 30:29) and immhi (bits 23:5), with bit 31
-// selecting the page form.
+// selecting the page form.  An n(PC) operand resolves to the instruction's
+// own address: the toolchain rewrites it away and encodes displacement 0.
 func encodeARM64ADR(mnem string, page uint32, ops []*ast.Operand, pc int, offsets map[string]int, resolve func(string) string) ([]byte, error) {
 	if len(ops) != 2 {
 		return nil, fmt.Errorf("%s expects 2 operands, got %d", mnem, len(ops))
@@ -2570,15 +2869,18 @@ func encodeARM64ADR(mnem string, page uint32, ops []*ast.Operand, pc int, offset
 	if rd < 0 {
 		return nil, fmt.Errorf("invalid register operand in %s", mnem)
 	}
+	var rel int64
+	if _, pcRel := arm64PCRelOffset(ops[0]); !pcRel {
 		target := resolve(arm64Label(ops[0]))
 		targetOff, ok := offsets[target]
 		if !ok {
 			return nil, fmt.Errorf("undefined label %q", target)
 		}
-	rel := int64(targetOff - pc)
+		rel = int64(targetOff - pc)
 		if rel < -(1<<20) || rel >= 1<<20 {
 			return nil, fmt.Errorf("%s to %q too far (21-bit range)", mnem, target)
 		}
+	}
 	return a64wordLE(a64ADR(page, int32(rel>>2), uint32(rel)&3, uint32(rd))), nil
 }
 
@@ -2856,11 +3158,16 @@ func encodeARM64AcqRel(mnem string, baseOp uint32, ops []*ast.Operand) ([]byte,
 //	BRK [$imm16]          SVC $imm16
 //	DMB|DSB|ISB $imm4     DC <op>, Rn
 //	MRS <sysreg>, Rd      MSR $imm4, <sysreg>
-//	PRFM (Rn), $imm|<op>
+//	PRFM (Rn), $imm|<op>  RPRFM (Rn), Rm, <op|$imm6>
+//	SYS $imm[, Rn]        SYSL $imm, Rd
+//	TLBI <op>[, Rn]       SB, PACIASP, PACIBSP
+//
+// The system registers, TLBI and DC aliases and the range-prefetch operations
+// come from the toolchain's own data tables in arm64_sysregs.go.
 func encodeARM64Sys(mnem string, ops []*ast.Operand) ([]byte, error) {
 	// Operand-less returns and pointer-authentication hints.
 	if w, ok := map[string]uint32{"DRPS": 0xd6bf03e0, "ERET": 0xd69f03e0,
-		"AUTIASP": 0xd50323bf, "AUTIBSP": 0xd50323ff,
+		"AUTIASP": 0xd50323bf, "AUTIBSP": 0xd50323ff, "PACIASP": 0xd503233f, "PACIBSP": 0xd503237f,
 		"AUTIA1716": 0xd503211f, "AUTIB1716": 0xd503213f,
 		"YIELD": 0xd503203d, "WFE": 0xd503205f, "WFI": 0xd503207f,
 		"SEVL": 0xd50320bf, "SEV": 0xd503209f}[mnem]; ok {
@@ -2896,6 +3203,12 @@ func encodeARM64Sys(mnem string, ops []*ast.Operand) ([]byte, error) {
 		}
 		base := map[string]uint32{"DMB": 0xd50330bf, "DSB": 0xd503309f, "ISB": 0xd50330df, "CLREX": 0xd503305f}[mnem]
 		return a64wordLE(base | uint32(v)<<8), nil
+	case "SB":
+		// Speculation barrier: DSB with a fixed barrier domain.
+		if len(ops) != 0 {
+			return nil, fmt.Errorf("%s expects no operand", mnem)
+		}
+		return a64wordLE(0xd50330ff), nil
 	case "HINT":
 		if len(ops) != 1 || !isImmOperand(ops[0]) {
 			return nil, fmt.Errorf("%s expects $immediate", mnem)
@@ -2934,7 +3247,7 @@ func encodeARM64Sys(mnem string, ops []*ast.Operand) ([]byte, error) {
 		if len(ops) != 2 {
 			return nil, fmt.Errorf("DC expects <op>, Rn")
 		}
-		base, ok := a64DCOps[operandRegName(ops[0])]
+		inst, ok := a64DCOps2[operandRegName(ops[0])]
 		if !ok {
 			return nil, fmt.Errorf("DC: unknown cache operation %q", operandRegName(ops[0]))
 		}
@@ -2942,51 +3255,115 @@ func encodeARM64Sys(mnem string, ops []*ast.Operand) ([]byte, error) {
 		if rn < 0 {
 			return nil, fmt.Errorf("DC: invalid register operand")
 		}
-		return a64wordLE(base | uint32(rn)&31), nil
+		w := 0xd5080000 | inst.op1<<16 | 7<<12 | inst.cm<<8 | inst.op2<<5
+		return a64wordLE(w | uint32(rn)&31), nil
+	case "TLBI":
+		// The register operand is optional: TLBI VMALLE1IS alone means ZR.
+		if len(ops) != 1 && len(ops) != 2 {
+			return nil, fmt.Errorf("TLBI expects <op>[, Rn]")
+		}
+		inst, ok := a64TLBIOps[operandRegName(ops[0])]
+		if !ok {
+			return nil, fmt.Errorf("TLBI: unknown operation %q", operandRegName(ops[0]))
+		}
+		rt := 31
+		if len(ops) == 2 {
+			if rt = arm64RegNum(operandRegName(ops[1])); rt < 0 {
+				return nil, fmt.Errorf("TLBI: invalid register operand")
+			}
+		}
+		w := 0xd5080000 | inst.op1<<16 | 8<<12 | inst.cm<<8 | inst.op2<<5
+		return a64wordLE(w | uint32(rt)&31), nil
+	case "SYS", "SYSL":
+		// SYS $imm[, Rn] / SYSL $imm, Rd: the immediate packs
+		// op1<<16 | CRn<<12 | CRm<<8 | op2<<5, the register defaults to ZR.
+		if len(ops) != 1 && len(ops) != 2 {
+			return nil, fmt.Errorf("%s expects $immediate[, Rn]", mnem)
+		}
+		if len(ops) == 1 && mnem == "SYSL" {
+			return nil, fmt.Errorf("SYSL expects $immediate, Rd")
+		}
+		if !isImmOperand(ops[0]) {
+			return nil, fmt.Errorf("%s expects $immediate[, Rn]", mnem)
+		}
+		imm := arm64Imm64(ops[0])
+		if imm < 0 || imm&^0x7FFE0 != 0 {
+			return nil, fmt.Errorf("%s: illegal SYS argument %d", mnem, imm)
+		}
+		rt := 31
+		if len(ops) == 2 {
+			if rt = arm64RegNum(operandRegName(ops[1])); rt < 0 {
+				return nil, fmt.Errorf("%s: invalid register operand", mnem)
+			}
+		}
+		base := uint32(0xd5080000)
+		if mnem == "SYSL" {
+			base = 0xd5280000
+		}
+		return a64wordLE(base | uint32(imm) | uint32(rt)&31), nil
 	case "MRS":
 		if len(ops) != 2 {
 			return nil, fmt.Errorf("MRS expects <sysreg>, Rd")
 		}
-		base, ok := a64MRSOps[operandRegName(ops[0])]
+		reg, ok := a64SysRegs[operandRegName(ops[0])]
 		if !ok {
 			return nil, fmt.Errorf("MRS: unknown system register %q", operandRegName(ops[0]))
 		}
+		if !reg.read {
+			return nil, fmt.Errorf("MRS: system register is not readable: %q", operandRegName(ops[0]))
+		}
 		rd := arm64RegNum(operandRegName(ops[1]))
 		if rd < 0 {
 			return nil, fmt.Errorf("MRS: invalid register operand")
 		}
-		return a64wordLE(base | uint32(rd)&31), nil
+		return a64wordLE(0xd5300000 | reg.v | uint32(rd)&31), nil
 	case "MSR":
 		if len(ops) != 2 {
 			return nil, fmt.Errorf("MSR expects $immediate, <sysreg> or Rn, <sysreg>")
 		}
-		if !isImmOperand(ops[0]) {
-			// Register form: MSR Rn, <sysreg> (the a64MSRRegOps words).
-			base, ok := a64MSRRegOps[operandRegName(ops[1])]
+		if isImmOperand(ops[0]) {
+			v := arm64Imm64(ops[0])
+			// The PSTATE fields keep their dedicated immediate form.
+			if base, ok := a64MSROps[operandRegName(ops[1])]; ok {
+				if v < 0 || v > 15 {
+					return nil, fmt.Errorf("MSR: immediate %d out of range (0..15)", v)
+				}
+				return a64wordLE(base | uint32(v)<<8 | 31), nil
+			}
+			// A $0 against a full system register writes it from ZR, exactly
+			// the way the toolchain preprocesses the constant away; any other
+			// immediate is the PSTATE-form error.
+			if v != 0 {
+				return nil, fmt.Errorf("MSR: illegal PSTATE field for immediate move: %q", operandRegName(ops[1]))
+			}
+			reg, ok := a64SysRegs[operandRegName(ops[1])]
 			if !ok {
 				return nil, fmt.Errorf("MSR: unknown system register %q", operandRegName(ops[1]))
 			}
+			if !reg.write {
+				return nil, fmt.Errorf("MSR: system register is not writable: %q", operandRegName(ops[1]))
+			}
+			return a64wordLE(0xd5100000 | reg.v | 31), nil
+		}
+		// Register form: MSR Rn, <sysreg>.
+		reg, ok := a64SysRegs[operandRegName(ops[1])]
+		if !ok {
+			return nil, fmt.Errorf("MSR: unknown system register %q", operandRegName(ops[1]))
+		}
+		if !reg.write {
+			return nil, fmt.Errorf("MSR: system register is not writable: %q", operandRegName(ops[1]))
+		}
 		rs := arm64RegNum(operandRegName(ops[0]))
 		if rs < 0 {
 			return nil, fmt.Errorf("MSR: invalid source register")
 		}
-			return a64wordLE(base | uint32(rs)&31), nil
-		}
-		base, ok := a64MSROps[operandRegName(ops[1])]
-		if !ok {
-			return nil, fmt.Errorf("MSR: unknown system register %q", operandRegName(ops[1]))
-		}
-		v := arm64Imm64(ops[0])
-		if v < 0 || v > 15 {
-			return nil, fmt.Errorf("MSR: immediate %d out of range (0..15)", v)
-		}
-		return a64wordLE(base | uint32(v)<<8 | 31), nil
+		return a64wordLE(0xd5100000 | reg.v | uint32(rs)&31), nil
 	case "PRFM":
 		if len(ops) != 2 {
 			return nil, fmt.Errorf("PRFM expects (Rn), $immediate|<op>")
 		}
 		rn, off := arm64MemWithFrame(ops[0], arm64FrameInfo{})
-		if rn < 0 || off != 0 {
+		if rn < 0 || off < 0 || off%8 != 0 || off/8 >= 4096 {
 			return nil, fmt.Errorf("PRFM: invalid memory operand")
 		}
 		var prfop int64
@@ -3002,7 +3379,36 @@ func encodeARM64Sys(mnem string, ops []*ast.Operand) ([]byte, error) {
 			}
 			prfop = int64(p)
 		}
-		return a64wordLE(0xf9800000 | uint32(rn)<<5 | uint32(prfop)), nil
+		return a64wordLE(0xf9800000 | uint32(off/8)<<10 | uint32(rn)<<5 | uint32(prfop)), nil
+	case "RPRFM":
+		// RPRFM (Rn), Rm, <op|$imm6>: the 6-bit operation scatters across
+		// bits 15, 13, 12 and 2:0 (asm7.go case 110).
+		if len(ops) != 3 {
+			return nil, fmt.Errorf("RPRFM expects (Rn), Rm, <op|$immediate>")
+		}
+		rn, off := arm64MemWithFrame(ops[0], arm64FrameInfo{})
+		if rn < 0 || off != 0 {
+			return nil, fmt.Errorf("RPRFM: invalid memory operand")
+		}
+		rm := arm64RegNum(operandRegName(ops[1]))
+		if rm < 0 {
+			return nil, fmt.Errorf("RPRFM: invalid register operand")
+		}
+		var op uint64
+		if isImmOperand(ops[2]) {
+			op = uint64(arm64Imm64(ops[2]))
+			if op > 63 {
+				return nil, fmt.Errorf("RPRFM: range prefetch immediate %d out of range (0..63)", op)
+			}
+		} else {
+			v, ok := a64RPRFOps[operandRegName(ops[2])]
+			if !ok {
+				return nil, fmt.Errorf("RPRFM: unknown range prefetch operation %q", operandRegName(ops[2]))
+			}
+			op = uint64(v)
+		}
+		scatter := (op&(1<<5))<<10 | (op&(1<<4))<<9 | (op&(1<<3))<<9 | op&7
+		return a64wordLE(0xf8a04818 | uint32(rm)<<16 | uint32(rn)<<5 | uint32(scatter)), nil
 	}
 	return nil, fmt.Errorf("unsupported arm64 instruction %q", mnem)
 }
@@ -3046,29 +3452,29 @@ func encodeARM64MoveWide(mnem string, baseOp uint32, ops []*ast.Operand) ([]byte
 	// base, so MOVZ and MOVN come along for free.
 	opc := baseOp >> 29 & 3
 	sf := baseOp >> 31 & 1
-	v := arm64Imm64(ops[0])
-	if v < 0 {
-		return nil, fmt.Errorf("%s: negative immediate %d", mnem, v)
+	// The toolchain's optab case 33, shared by the whole family in both
+	// widths: the immediate is one unsigned 64-bit pattern (a high-lane
+	// constant such as $(40000<<48) arrives negative through int64
+	// folding), it must occupy exactly one 16-bit lane, zero is rejected,
+	// and the W forms cannot reach the top half.
+	u := uint64(arm64Imm64(ops[0]))
+	if u == 0 {
+		return nil, fmt.Errorf("%s: zero immediate cannot be handled", mnem)
 	}
 	hw := -1
-	for i := range 4 {
-		if v>>(uint(i)*16)&0xFFFF != 0 {
-			hw = i
+	for lane := range 4 {
+		if u&^(uint64(0xFFFF)<<(lane*16)) == 0 {
+			hw = lane
 			break
 		}
 	}
 	if hw < 0 {
-		hw = 0 // zero: every chunk is zero, hw = 0 carries it
-	}
-	for i := hw + 1; i < 4; i++ {
-		if v>>(uint(i)*16)&0xFFFF != 0 {
-			return nil, fmt.Errorf("%s: immediate %d does not fit one 16-bit chunk", mnem, v)
-		}
+		return nil, fmt.Errorf("%s: immediate %#x does not fit one 16-bit chunk", mnem, u)
 	}
 	if sf == 0 && hw > 1 {
-		return nil, fmt.Errorf("%s: immediate %d out of range for the 32-bit form", mnem, v)
+		return nil, fmt.Errorf("%s: immediate %#x out of range for the 32-bit form", mnem, u)
 	}
-	return a64wordLE(a64MoveWide(sf, opc, uint32(hw), uint32(v>>uint(hw*16)&0xFFFF), uint32(rd))), nil
+	return a64wordLE(a64MoveWide(sf, opc, uint32(hw), uint32(u>>uint(hw*16)&0xFFFF), uint32(rd))), nil
 }
 
 // ---- Bitfield/EXTR encoding ----
@@ -3425,7 +3831,7 @@ func encodeARM64VTBL(mnem string, ops []*ast.Operand) ([]byte, error) {
 		return nil, fmt.Errorf("invalid destination register in VTBL")
 	}
 	for i, t := range ts {
-		if t.hasIdx || t.reg != ts[0].reg+i {
+		if t.hasIdx || (ts[0].reg+i)&31 != t.reg {
 			return nil, fmt.Errorf("VTBL table registers must be consecutive")
 		}
 	}
@@ -3582,15 +3988,17 @@ func encodeARM64Dup(mnem string, ops []*ast.Operand) ([]byte, error) {
 
 // encodeARM64VLDST encodes the SIMD structure loads and stores:
 //
-//	VLD1 (Rn), [Vt.arr, ...]      VST1 [Vt.arr, ...], (Rn)
-//	VLD1.P off(Rn), [Vt.arr, ...] VST1.P [Vt.arr, ...], off(Rn)
-//	VLD1.P off(Rn), Vt.T[i]       VST1.P Vt.T[i], off(Rn)   (one lane)
-//	VLD1R (Rn), [Vt.arr]          VLD4R (Rn), [Vt.arr, Vt+1, Vt+2, Vt+3]
+//	VLD1|2|3|4 (Rn), [Vt.arr, ...]        VST1|2|3|4 [Vt.arr, ...], (Rn)
+//	VLD1|2|3|4.P off(Rn), [Vt.arr, ...]   VST1|2|3|4.P [Vt.arr, ...], off(Rn)
+//	VLD1|2|3|4.P (Rn)(Rm), [Vt.arr, ...]  VST1|2|3|4.P [Vt.arr, ...], (Rn)(Rm)
+//	VLD1|2|3|4R (Rn), [Vt.arr, ...]       (replicating loads)
+//	VLD1 off(Rn), Vt.T[i]                 VST1 Vt.T[i], off(Rn)   (one lane)
 //
-// The post-index forms set the post bit and Rm = 11111.  A spelled offset
-// rides along (the encoding ignores it; the toolchain only checks that it
-// matches the access size), and a multi-register post-index list takes no
-// offset at all, the increment following from the list.
+// The post-index forms set the post bit and carry Rm: 11111 for an immediate
+// increment, the spelled register for (Rn)(Rm).  A register list may wrap
+// around V31: the toolchain checks only (first+i) mod 32.  A spelled offset
+// rides along on the one-lane forms (the toolchain only checks that it
+// matches the access size).
 func encodeARM64VLDST(mnem string, post uint32, ops []*ast.Operand) ([]byte, error) {
 	load := strings.HasPrefix(mnem, "VLD")
 
@@ -3628,24 +4036,31 @@ func encodeARM64VLDST(mnem string, post uint32, ops []*ast.Operand) ([]byte, err
 	if off != 0 && post == 0 {
 		return nil, fmt.Errorf("%s: offset %d not supported, plain and list accesses take a plain (Rn) operand", mnem, off)
 	}
-
-	// VLD1R loads one register and replicates; VLD4R loads four.
-	if strings.HasPrefix(mnem, "VLD1R") || strings.HasPrefix(mnem, "VLD4R") {
-		want := 1
-		base := uint32(0x0d40c000)
-		if strings.HasPrefix(mnem, "VLD4R") {
-			want, base = 4, 0x0d60e000
+	// The post-index increment: 11111 for an immediate offset, else the
+	// spelled (Rn)(Rm) register.
+	rm := 31
+	if post != 0 {
+		if idx := ops[memIdx].Addr.Index; idx != "" {
+			if rm = arm64RegNum(idx); rm < 0 {
+				return nil, fmt.Errorf("%s: invalid post-index register %q", mnem, idx)
 			}
-		if len(vs) != want {
-			return nil, fmt.Errorf("%s expects a list of %d registers", mnem, want)
+		}
+	}
+
+	// The replicating loads: VLD1R through VLD4R load one register and
+	// replicate it across the whole list.
+	if base := strings.TrimSuffix(mnem, ".P"); load && strings.HasSuffix(base, "R") && len(base) == 5 {
+		n := int(base[3] - '0')
+		if len(vs) != n {
+			return nil, fmt.Errorf("%s expects a list of %d registers", mnem, n)
 		}
 		size, q, ok := a64ArrSizeQ(vs[0].arr)
 		if !ok {
 			return nil, fmt.Errorf("%s: invalid arrangement %q", mnem, vs[0].arr)
 		}
-		w := base | q<<30 | size<<10 | uint32(rn)<<5 | uint32(vs[0].reg)
+		w := a64VLDNReplicate[n] | q<<30 | size<<10 | uint32(rn)<<5 | uint32(vs[0].reg)
 		if post != 0 {
-			w |= 1<<23 | 0x1f<<16
+			w |= 1<<23 | uint32(rm)<<16
 		}
 		return a64wordLE(w), nil
 	}
@@ -3654,7 +4069,7 @@ func encodeARM64VLDST(mnem string, post uint32, ops []*ast.Operand) ([]byte, err
 		return nil, fmt.Errorf("%s expects a list of one to four registers", mnem)
 	}
 	for i, v := range vs {
-		if v.hasIdx || v.reg != vs[0].reg+i {
+		if v.hasIdx || (vs[0].reg+i)&31 != v.reg {
 			return nil, fmt.Errorf("%s: register list must be consecutive", mnem)
 		}
 		_, _, okArr := a64ArrSizeQ(v.arr)
@@ -3666,21 +4081,36 @@ func encodeARM64VLDST(mnem string, post uint32, ops []*ast.Operand) ([]byte, err
 	if !ok {
 		return nil, fmt.Errorf("%s: invalid arrangement %q", mnem, vs[0].arr)
 	}
-	base := a64VLD1Base[len(vs)]
+	n := len(vs)
+	base := a64VLD1Base[n]
 	if !load {
-		base = a64VST1Base[len(vs)]
+		base = a64VST1Base[n]
+	}
+	// VLD2/VLD3/VLD4 and VST2/VST3/VST4 name the register count in the
+	// mnemonic and carry their own opcode fields.  The count digit sits at
+	// index 3 of the mnemonic (VLD2, VST3.P, ...), before any .P suffix.
+	if stem := strings.TrimSuffix(mnem, ".P"); len(stem) >= 4 && stem[3] >= '2' && stem[3] <= '4' {
+		n := int(stem[3] - '0')
+		if n != len(vs) {
+			return nil, fmt.Errorf("%s expects a list of %d registers", mnem, n)
+		}
+		if load {
+			base = a64VLDNBase[n]
+		} else {
+			base = a64VSTNBase[n]
+		}
 	}
 	postBits := uint32(0)
 	if post != 0 {
-		postBits = 0x9f0000
+		postBits = 1<<23 | uint32(rm)<<16
 	}
 	return a64wordLE(base | q<<30 | size<<10 | postBits | uint32(rn)<<5 | uint32(vs[0].reg)), nil
 }
 
 // encodeARM64VLDSTLane encodes the one-lane structure forms:
-// VLD1 off(Rn), Vt.T[i] (post-index adds the post bit and Rm=11111) and
-// VST1.P Vt.T[i], off(Rn); the plain VST1 lane form does not exist in the
-// toolchain's table and is rejected.
+// VLD1 off(Rn), Vt.T[i] and VST1 Vt.T[i], off(Rn); the post-index spellings
+// add the post bit and Rm: 11111 for an immediate increment, the spelled
+// register for (Rn)(Rm).
 func encodeARM64VLDSTLane(mnem string, post uint32, ops []*ast.Operand, load bool, laneIdx int, v a64Vec) ([]byte, error) {
 	if len(ops) != 2 {
 		return nil, fmt.Errorf("%s expects a memory operand and one Vt.T[i] lane operand", mnem)
@@ -3690,14 +4120,19 @@ func encodeARM64VLDSTLane(mnem string, post uint32, ops []*ast.Operand, load boo
 	if rn < 0 {
 		return nil, fmt.Errorf("%s: invalid memory operand", mnem)
 	}
-	if !load && post == 0 {
-		return nil, fmt.Errorf("%s: the toolchain only spells a post-index single-lane store", mnem)
+	rm := 31
+	if post != 0 {
+		if idx := ops[memIdx].Addr.Index; idx != "" {
+			if rm = arm64RegNum(idx); rm < 0 {
+				return nil, fmt.Errorf("%s: invalid post-index register %q", mnem, idx)
+			}
+		}
 	}
 	w := uint32(0x0d400000)
 	switch strings.ToUpper(v.arr) {
 	case "B":
-		// Index at bits 12:10 (the size field doubles as the low index bits).
-		w |= uint32(v.idx) << 10
+		// Index<3> rides bit 30, index<2:0> the size field at bits 12:10.
+		w |= uint32(v.idx&7)<<10 | uint32(v.idx>>3&1)<<30
 	case "H":
 		// Index<2> at bit 30, index<1> at bit 12, index<0> at bit 11.
 		w |= 1<<14 | uint32(v.idx&1)<<11 | uint32(v.idx>>1&1)<<12 | uint32(v.idx>>2&1)<<30
@@ -3711,12 +4146,12 @@ func encodeARM64VLDSTLane(mnem string, post uint32, ops []*ast.Operand, load boo
 		return nil, fmt.Errorf("%s: invalid lane arrangement %q", mnem, v.arr)
 	}
 	// The base carries bit 22 (L) set; a store clears it.  The post-index
-	// forms add bit 23 and Rm = 11111.
+	// forms add bit 23 and Rm.
 	if !load {
 		w &^= 1 << 22
 	}
 	if post != 0 {
-		w |= 1<<23 | 0x1f<<16
+		w |= 1<<23 | uint32(rm)<<16
 	}
 	return a64wordLE(w | uint32(rn)<<5 | uint32(v.reg)), nil
 }
@@ -3980,6 +4415,7 @@ func AssembleFileARM64(f *ast.File) (*Image, error) {
 			Size:   d.size,
 			Static: d.static,
 			Rodata: d.rodata,
+			Noptr:  d.noptr,
 			Dupok:  d.dupok,
 		})
 	}

asm/arm64_encode.go

@@ -33,7 +33,7 @@ import (
 	"strconv"
 	"strings"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/ast"
+	"sourcedock.dev/petrbalvin/gasm-sdk/ast"
 )
 
 // arm64RegNum returns the 5-bit register number for an AArch64 register name:
@@ -374,6 +374,7 @@ const (
 	a64FPair                           // load/store pair: LDP, STP, LDPW, STPW, FLDPD, FSTPD
 	a64FAcqRel                         // acquire/release: LDAR family, STLR family
 	a64FSys                            // system: BRK, SVC, DMB, DSB, ISB, DC, MRS, MSR, PRFM
+	a64FCASP                           // compare and swap pair: CASP
 	a64FCrypto2                        // crypto 2-register: AESD, AESE, AESIMC, AESMC, SHA1H, ...
 	a64FCrypto3                        // crypto 3-register: SHA1C, SHA256H, SHA512SU1, ...
 	a64FSIMDV                          // SIMD 3-register with arrangement: VADD, VAND, VCMEQ, VZIP1, ...
@@ -384,6 +385,7 @@ const (
 	a64FDUP                            // SIMD element moves: VDUP, VMOV with element indices
 	a64FVLDST                          // SIMD structure loads/stores: VLD1, VST1, VLD1R, VLD4R
 	a64FShiftImm                       // SIMD shift by immediate: VSHL, VUSHR, VSRI
+	a64FVMoviImm                       // SIMD move immediate: VMOVI $imm8, Vd.B8/B16
 	a64FMoviLit                        // VMOVS/VMOVD/VMOVQ with a large constant (literal pool)
 )
 
@@ -770,7 +772,7 @@ func init() {
 	a64InstrTable["CCMNW"] = a64Enc{format: a64FCondCmp, op: 0x3a400000}
 
 	// ---- system operations ----
-	for _, m := range []string{"BRK", "SVC", "DMB", "DSB", "ISB", "CLREX", "HINT", "BTI", "HLT", "SMC", "HVC", "DCPS1", "DCPS2", "DCPS3", "DRPS", "ERET", "AUTIASP", "AUTIBSP", "AUTIA1716", "AUTIB1716", "SEVL", "SEV", "WFE", "WFI", "YIELD", "DC", "MRS", "MSR", "PRFM"} {
+	for _, m := range []string{"BRK", "SVC", "DMB", "DSB", "ISB", "CLREX", "HINT", "BTI", "HLT", "SMC", "HVC", "DCPS1", "DCPS2", "DCPS3", "DRPS", "ERET", "AUTIASP", "AUTIBSP", "AUTIA1716", "AUTIB1716", "SEVL", "SEV", "WFE", "WFI", "YIELD", "DC", "MRS", "MSR", "PRFM", "RPRFM", "SYS", "SYSL", "TLBI", "SB", "PACIASP", "PACIBSP"} {
 		a64InstrTable[m] = a64Enc{format: a64FSys}
 	}
 
@@ -783,12 +785,20 @@ func init() {
 	a64InstrTable["TBNZ"] = a64Enc{format: a64FTestBranch, op: 0x37000000}
 
 	// ---- load/store pair (signed offset) ----
+	// The scale column of a64LoadTable does not reach the pair forms, so each
+	// entry states its own access width through the imm7 divisor the pair
+	// encoder derives from the opc field (8 for D, 4 for W and SW, 16 for Q).
 	a64InstrTable["LDP"] = a64Enc{format: a64FPair, op: 0xa9400000}
 	a64InstrTable["LDPW"] = a64Enc{format: a64FPair, op: 0x29400000}
+	a64InstrTable["LDPSW"] = a64Enc{format: a64FPair, op: 0x69400000}
 	a64InstrTable["STP"] = a64Enc{format: a64FPair, op: 0xa9000000}
 	a64InstrTable["STPW"] = a64Enc{format: a64FPair, op: 0x29000000}
 	a64InstrTable["FLDPD"] = a64Enc{format: a64FPair, op: 0x6d400000}
 	a64InstrTable["FSTPD"] = a64Enc{format: a64FPair, op: 0x6d000000}
+	a64InstrTable["FLDPS"] = a64Enc{format: a64FPair, op: 0x2d400000}
+	a64InstrTable["FSTPS"] = a64Enc{format: a64FPair, op: 0x2d000000}
+	a64InstrTable["FLDPQ"] = a64Enc{format: a64FPair, op: 0xad400000}
+	a64InstrTable["FSTPQ"] = a64Enc{format: a64FPair, op: 0xad000000}
 
 	// ---- acquire/release loads and stores ----
 	a64InstrTable["LDAR"] = a64Enc{format: a64FAcqRel, op: 0xc8dffc00}
@@ -806,11 +816,23 @@ func init() {
 	lse := map[string]uint32{
 		"CASALD":   0xc8e0fc00,
 		"CASALW":   0x88e0fc00,
+		"CASB":     0x08a07c00,
+		"CASAB":    0x08e07c00,
+		"CASH":     0x48a07c00,
+		"CASLD":    0xc8a0fc00,
+		"CASLH":    0x48a0fc00,
+		"CASAW":    0x88e07c00,
+		"CASAD":    0xc8e07c00,
+		"CASALH":   0x48e07c00,
 		"LDADDALD": 0xf8e00000,
 		"LDADDALW": 0xb8e00000,
+		"LDADDAD":  0xf8a00000,
+		"LDADDAW":  0xb8a00000,
 		"LDCLRALB": 0x38e01000,
 		"LDCLRALW": 0xb8e01000,
 		"LDCLRALD": 0xf8e01000,
+		"LDCLRAD":  0xf8a01000,
+		"LDCLRAW":  0xb8a01000,
 		"LDORALB":  0x38e03000,
 		"LDORALW":  0xb8e03000,
 		"LDORALD":  0xf8e03000,
@@ -889,6 +911,12 @@ func init() {
 		a64InstrTable[m] = a64Enc{format: a64FLSE, op: op}
 	}
 
+	// Compare and swap pair: the second register of each pair is implicit
+	// (Rs+1 and Rt+1), so the encoding carries Rs and Rt alone over a preset
+	// fixed field (asm7.go atomicCASP).
+	a64InstrTable["CASPD"] = a64Enc{format: a64FCASP, op: 1<<30 | 0x41<<21 | 0x1f<<10}
+	a64InstrTable["CASPW"] = a64Enc{format: a64FCASP, op: 0x41<<21 | 0x1f<<10}
+
 	// ---- carry-setting/carry-using arithmetic and widening multiply ----
 	// MUL and SMULH/UMULH are the MADD/MSUB layout with the accumulate
 	// register preset to ZR (bits 14:10 = 11111).
@@ -940,11 +968,13 @@ func init() {
 	a64InstrTable["VMOVS"] = a64Enc{format: a64FMoviLit, op: 0xbd400000}
 	a64InstrTable["VMOVD"] = a64Enc{format: a64FMoviLit, op: 0xfd400000}
 	a64InstrTable["VMOVQ"] = a64Enc{format: a64FMoviLit, op: 0x3dc00000}
+	a64InstrTable["VMOVI"] = a64Enc{format: a64FVMoviImm}
 	a64InstrTable["VSHL"] = a64Enc{format: a64FShiftImm, op: 0x0f000000 | 21<<10}
 	a64InstrTable["VUSHR"] = a64Enc{format: a64FShiftImm, op: 0x2f000000 | 1<<10}
 	a64InstrTable["VSRI"] = a64Enc{format: a64FShiftImm, op: 0x2f000000 | 17<<10}
 	a64InstrTable["VSSHR"] = a64Enc{format: a64FShiftImm, op: 0x0f000000 | 1<<10}
-	a64InstrTable["VSRA"] = a64Enc{format: a64FShiftImm, op: 0x0f000000 | 17<<10}
+	a64InstrTable["VSRA"] = a64Enc{format: a64FShiftImm, op: 0x0f000000 | 7<<10}
+	a64InstrTable["VUSRA"] = a64Enc{format: a64FShiftImm, op: 0x2f000000 | 5<<10}
 	a64InstrTable["VSRSHR"] = a64Enc{format: a64FShiftImm, op: 0x0f000000 | 9<<10}
 	a64InstrTable["VSLI"] = a64Enc{format: a64FShiftImm, op: 0x2f000000 | 21<<10}
 	a64InstrTable["VSQSHL"] = a64Enc{format: a64FShiftImm, op: 0x0f000000 | 29<<10}
@@ -957,6 +987,24 @@ func init() {
 	a64InstrTable["VLD1R.P"] = a64Enc{format: a64FVLDST, op: 1}
 	a64InstrTable["VLD4R"] = a64Enc{format: a64FVLDST}
 	a64InstrTable["VLD4R.P"] = a64Enc{format: a64FVLDST, op: 1}
+	// Multi-register structure accesses beyond VLD1/VST1: VLD2/VLD3/VLD4 and
+	// the replicate loads VLD2R/VLD3R, each with the post-index spelling.
+	a64InstrTable["VLD2"] = a64Enc{format: a64FVLDST}
+	a64InstrTable["VLD2.P"] = a64Enc{format: a64FVLDST, op: 1}
+	a64InstrTable["VLD3"] = a64Enc{format: a64FVLDST}
+	a64InstrTable["VLD3.P"] = a64Enc{format: a64FVLDST, op: 1}
+	a64InstrTable["VLD4"] = a64Enc{format: a64FVLDST}
+	a64InstrTable["VLD4.P"] = a64Enc{format: a64FVLDST, op: 1}
+	a64InstrTable["VLD2R"] = a64Enc{format: a64FVLDST}
+	a64InstrTable["VLD2R.P"] = a64Enc{format: a64FVLDST, op: 1}
+	a64InstrTable["VLD3R"] = a64Enc{format: a64FVLDST}
+	a64InstrTable["VLD3R.P"] = a64Enc{format: a64FVLDST, op: 1}
+	a64InstrTable["VST2"] = a64Enc{format: a64FVLDST}
+	a64InstrTable["VST2.P"] = a64Enc{format: a64FVLDST, op: 1}
+	a64InstrTable["VST3"] = a64Enc{format: a64FVLDST}
+	a64InstrTable["VST3.P"] = a64Enc{format: a64FVLDST, op: 1}
+	a64InstrTable["VST4"] = a64Enc{format: a64FVLDST}
+	a64InstrTable["VST4.P"] = a64Enc{format: a64FVLDST, op: 1}
 }
 
 // a64SimdVSpec is one arrangement-aware SIMD instruction: the 8B base word,
@@ -1120,6 +1168,78 @@ var a64SimdVTable = map[string]a64SimdVSpec{
 	"VRAX1":    {0xce608c00, 1 << a64Arr2D, true}, // SHA3 group, D2 only
 	"VPMULL":   {0x0e20e000, 1<<a64Arr8B | 1<<a64ArrD1, false},
 	"VPMULL2":  {0x0e20e000, 1<<a64Arr16B | 1<<a64Arr2D, false},
+	// Saturating shifts, register forms (the immediate spellings route to
+	// a64FShiftImm).
+	"VSQSHL": {0x0e204c00, 0x7f, false},
+	"VUQSHL": {0x2e204c00, 0x7f, false},
+}
+
+// a64SimdNLForm classifies the narrow/long/wide SIMD families whose
+// arrangement does not travel on every operand: the encoding's size and Q
+// bits read off one designated operand and the element widths pair up across
+// the operands.
+type a64SimdNLForm uint8
+
+const (
+	a64NLTwoNarrow        a64SimdNLForm = iota // (Vn.wide, Vd.narrow): size/Q from Vd
+	a64NLTwoLong                               // (Vn.narrow, Vd.long): size/Q from Vn
+	a64NLThreeLongMul                          // (Vm.narrow, Vn.narrow, Vd.long): size/Q from Vn
+	a64NLThreeWide                             // (Vm.narrow, Vn.wide, Vd.wide): size/Q from Vn
+	a64NLThreeLongShift                        // ($sh, Vn.narrow, Vd.long): size/Q from Vn, immh = esize+sh
+	a64NLThreeNarrowShift                      // ($sh, Vn.wide, Vd.narrow): size/Q from Vd, immh = esize-sh
+)
+
+// a64SimdNLSpec is one narrow/long/wide instruction: the base word (U, opcode
+// and fixed bits positioned) and the arrangement form.  qonly marks the FCVT
+// family, whose size field is fixed in the base and only the Q bit follows
+// the driving arrangement.
+type a64SimdNLSpec struct {
+	base  uint32
+	form  a64SimdNLForm
+	qonly bool
+}
+
+// a64SimdNLTable holds the families the arrangement-driven three-register and
+// two-register encoders cannot express.  The .2 spellings force the 128-bit
+// side of the pair through their operand arrangements, so the base carries no
+// arrangement bits of its own.
+var a64SimdNLTable = map[string]a64SimdNLSpec{
+	"VSHRN":    {0x0f008400, a64NLThreeNarrowShift, false},
+	"VSHRN2":   {0x0f008400, a64NLThreeNarrowShift, false},
+	"VSXTL":    {0x0f00a400, a64NLTwoLong, false},
+	"VSXTL2":   {0x0f00a400, a64NLTwoLong, false},
+	"VUXTL":    {0x2f00a400, a64NLTwoLong, false},
+	"VUXTL2":   {0x2f00a400, a64NLTwoLong, false},
+	"VXTN":     {0x0e202800, a64NLTwoNarrow, false},
+	"VXTN2":    {0x0e202800, a64NLTwoNarrow, false},
+	"VSQXTN":   {0x0e204800, a64NLTwoNarrow, false},
+	"VSQXTN2":  {0x0e204800, a64NLTwoNarrow, false},
+	"VSQXTUN":  {0x2e202800, a64NLTwoNarrow, false},
+	"VSQXTUN2": {0x2e202800, a64NLTwoNarrow, false},
+	"VUQXTN":   {0x2e204800, a64NLTwoNarrow, false},
+	"VUQXTN2":  {0x2e204800, a64NLTwoNarrow, false},
+	"VFCVTN":   {0x0e206800, a64NLTwoNarrow, true},
+	"VFCVTN2":  {0x0e206800, a64NLTwoNarrow, true},
+	"VFCVTL":   {0x0e217800, a64NLTwoLong, true},
+	"VFCVTL2":  {0x0e217800, a64NLTwoLong, true},
+	"VSSHLL":   {0x0f00a400, a64NLThreeLongShift, false},
+	"VSSHLL2":  {0x0f00a400, a64NLThreeLongShift, false},
+	"VUSHLL":   {0x2f00a400, a64NLThreeLongShift, false},
+	"VUSHLL2":  {0x2f00a400, a64NLThreeLongShift, false},
+	"VUADDW":   {0x2e201000, a64NLThreeWide, false},
+	"VUADDW2":  {0x2e201000, a64NLThreeWide, false},
+	"VUMULL":   {0x2e20c000, a64NLThreeLongMul, false},
+	"VUMULL2":  {0x2e20c000, a64NLThreeLongMul, false},
+	"VSMULL":   {0x0e20c000, a64NLThreeLongMul, false},
+	"VSMULL2":  {0x0e20c000, a64NLThreeLongMul, false},
+	"VUMLAL":   {0x2e208000, a64NLThreeLongMul, false},
+	"VUMLAL2":  {0x2e208000, a64NLThreeLongMul, false},
+	"VSMLAL":   {0x0e208000, a64NLThreeLongMul, false},
+	"VSMLAL2":  {0x0e208000, a64NLThreeLongMul, false},
+	"VUMLSL":   {0x2e20a000, a64NLThreeLongMul, false},
+	"VUMLSL2":  {0x2e20a000, a64NLThreeLongMul, false},
+	"VSMLSL":   {0x0e20a000, a64NLThreeLongMul, false},
+	"VSMLSL2":  {0x0e20a000, a64NLThreeLongMul, false},
 }
 
 // a64SimdVZero holds the compare-against-zero words of the SIMD compares
@@ -1243,6 +1363,16 @@ var a64PRFOps = map[string]int{
 var a64VLD1Base = [5]uint32{0, 0x0c407000, 0x0c40a000, 0x0c406000, 0x0c402000}
 var a64VST1Base = [5]uint32{0, 0x0c007000, 0x0c00a000, 0x0c006000, 0x0c002000}
 
+// a64VLDNBase and a64VSTNBase hold the VLD2/VLD3/VLD4 and VST2/VST3/VST4
+// fixed words (indexed by register count 2..4): the opcode field at bits
+// 15:12 carries the access kind.
+var a64VLDNBase = [5]uint32{0, 0, 0x0c408000, 0x0c404000, 0x0c400000}
+var a64VSTNBase = [5]uint32{0, 0, 0x0c008000, 0x0c004000, 0x0c000000}
+
+// a64VLDNReplicate holds the VLD2R/VLD3R fixed words beside the existing
+// VLD1R (0x0d40c000) and VLD4R (0x0d60e000) bases.
+var a64VLDNReplicate = [5]uint32{0, 0x0d40c000, 0x0d60c000, 0x0d40e000, 0x0d60e000}
+
 // a64Vec is a parsed vector operand: the register number, the arrangement
 // ("" when the operand spells none) and, for element forms, the lane index.
 type a64Vec struct {
@@ -1366,20 +1496,22 @@ type a64LSType struct {
 	size  int // 0=byte, 1=half, 2=word, 3=dword
 	V     int // 0=integer, 1=FP
 	opc   int // 00=store/unsigned load, 01=store FP, 10=signed load, 11=load FP
+	scale int // access width in bytes; the unsigned offset divides by it
 }
 
 // a64LoadTable maps MOV width mnemonics to their load/store encoding parameters.
 // For loads, opc selects signed vs unsigned; for stores, we flip the opc.
 var a64LoadTable = map[string]a64LSType{
-	"MOVD":  {3, 0, 1}, // LDR X (64-bit, unsigned offset)
-	"MOVWU": {2, 0, 1}, // LDR W (32-bit unsigned)
-	"MOVW":  {2, 0, 2}, // LDRSW (32-bit signed → 64-bit)
-	"MOVHU": {1, 0, 1}, // LDRH (16-bit unsigned)
-	"MOVH":  {1, 0, 2}, // LDRSH (16-bit signed)
-	"MOVBU": {0, 0, 1}, // LDRB (8-bit unsigned)
-	"MOVB":  {0, 0, 2}, // LDRSB (8-bit signed)
-	"FMOVS": {2, 1, 1}, // LDR S (32-bit FP)
-	"FMOVD": {3, 1, 1}, // LDR D (64-bit FP)
+	"MOVD":  {3, 0, 1, 8},  // LDR X (64-bit, unsigned offset)
+	"MOVWU": {2, 0, 1, 4},  // LDR W (32-bit unsigned)
+	"MOVW":  {2, 0, 2, 4},  // LDRSW (32-bit signed → 64-bit)
+	"MOVHU": {1, 0, 1, 2},  // LDRH (16-bit unsigned)
+	"MOVH":  {1, 0, 2, 2},  // LDRSH (16-bit signed)
+	"MOVBU": {0, 0, 1, 1},  // LDRB (8-bit unsigned)
+	"MOVB":  {0, 0, 2, 1},  // LDRSB (8-bit signed)
+	"FMOVS": {2, 1, 1, 4},  // LDR S (32-bit FP)
+	"FMOVD": {3, 1, 1, 8},  // LDR D (64-bit FP)
+	"FMOVQ": {0, 1, 3, 16}, // LDR/STR Q (128-bit FP): opc=11 selects it
 }
 
 // a64StoreOpc returns the store opc for a given load type: integer and FP

asm/arm64_encode_test.go

@@ -7,8 +7,8 @@ import (
 	"strings"
 	"testing"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/ast"
-	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+	"sourcedock.dev/petrbalvin/gasm-sdk/ast"
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
 )
 
 func TestArm64LDRSTREncoding(t *testing.T) {
@@ -1051,6 +1051,41 @@ func TestArm64MOVK(t *testing.T) {
 	}
 }
 
+// TestArm64MOVKHighLane pins the shifted high-lane immediate the arm64 test
+// kernels write: $(40000<<48) folds to a negative int64, and the toolchain
+// reads the value as an unsigned 64-bit pattern when it picks the lane.
+func TestArm64MOVKHighLane(t *testing.T) {
+	got := arm64Words(t, "\tMOVK $(40000<<48), R0\n\tMOVK $0x9c40000000000000, R1\n")
+	want := []uint32{
+		0xf2f38800, // MOVK $(40000<<48), R0 (go tool asm: f2f38800)
+		0xf2f38801, // MOVK hw=3
+		0xd65f03c0,
+	}
+	if len(got) != len(want) {
+		t.Fatalf("word count = %d, want %d", len(got), len(want))
+	}
+	for i := range want {
+		if got[i] != want[i] {
+			t.Errorf("word %d = %08x, want %08x", i, got[i], want[i])
+		}
+	}
+}
+
+// TestArm64MoveWideZeroImmediate pins the toolchain's rejection of a zero
+// immediate in the move-wide family (optab case 33: "zero shifts cannot be
+// handled"): every lane is zero, so no hw field can carry it.
+func TestArm64MoveWideZeroImmediate(t *testing.T) {
+	for _, mnem := range []string{"MOVK", "MOVZ", "MOVN"} {
+		f, errs := parser.Parse("test_arm64.s", "#include \"textflag.h\"\n\nTEXT ·f(SB), NOSPLIT, $0-0\n\t"+mnem+" $0, R0\n\tRET\n")
+		if len(errs) > 0 {
+			t.Fatalf("%s: parse: %v", mnem, errs)
+		}
+		if _, err := AssembleFileARM64(f); err == nil {
+			t.Errorf("%s $0: expected error, got nil", mnem)
+		}
+	}
+}
+
 // TestArm64LoadImm64 tests 64-bit immediate loading.
 func TestArm64LoadImm64(t *testing.T) {
 	src := `#include "textflag.h"

asm/arm64_frame.go

@@ -53,7 +53,7 @@ package asm
 import (
 	"strings"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/ast"
+	"sourcedock.dev/petrbalvin/gasm-sdk/ast"
 )
 
 // arm64FrameInfo holds the frame layout derived from a TEXT directive.

asm/arm64_reloc_test.go

@@ -7,7 +7,7 @@ import (
 	"encoding/binary"
 	"testing"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
 )
 
 // parseArm64File is a helper assembling one arm64 source file.

asm/arm64_sysregs.go

@@ -0,0 +1,588 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package asm
+
+// arm64 system registers and system-instruction aliases.
+//
+// The tables are transcribed from the data the Go toolchain itself carries
+// (cmd/internal/obj/arm64/sysRegEnc.go and the sysInstFields map of asm7.go),
+// which the ARM ARM defines: every system register is the packed field set
+// op0<<19 | op1<<16 | CRn<<12 | CRm<<8 | op2<<5, and the read/write flags are
+// the toolchain's own access classification.  The encoding tables live here so
+// the encoder stays testable against the GOROOT testdata word for word.
+
+// a64SysReg is one system register: the packed encoding fields and the
+// directions the register supports.
+type a64SysReg struct {
+	v     uint32
+	read  bool
+	write bool
+}
+
+// a64SysRegs maps the system register names the toolchain knows to their
+// encodings.  MRS reads 0xd5300000 | v | Rd and MSR writes
+// 0xd5100000 | v | Rt.
+var a64SysRegs = map[string]a64SysReg{
+	"ACTLR_EL1":         a64SysReg{0x181020, true, true},
+	"AFSR0_EL1":         a64SysReg{0x185100, true, true},
+	"AFSR1_EL1":         a64SysReg{0x185120, true, true},
+	"AIDR_EL1":          a64SysReg{0x1900e0, true, false},
+	"AMAIR_EL1":         a64SysReg{0x18a300, true, true},
+	"AMCFGR_EL0":        a64SysReg{0x1bd220, true, false},
+	"AMCGCR_EL0":        a64SysReg{0x1bd240, true, false},
+	"AMCNTENCLR0_EL0":   a64SysReg{0x1bd280, true, true},
+	"AMCNTENCLR1_EL0":   a64SysReg{0x1bd300, true, true},
+	"AMCNTENSET0_EL0":   a64SysReg{0x1bd2a0, true, true},
+	"AMCNTENSET1_EL0":   a64SysReg{0x1bd320, true, true},
+	"AMCR_EL0":          a64SysReg{0x1bd200, true, true},
+	"AMEVCNTR00_EL0":    a64SysReg{0x1bd400, true, true},
+	"AMEVCNTR01_EL0":    a64SysReg{0x1bd420, true, true},
+	"AMEVCNTR02_EL0":    a64SysReg{0x1bd440, true, true},
+	"AMEVCNTR03_EL0":    a64SysReg{0x1bd460, true, true},
+	"AMEVCNTR04_EL0":    a64SysReg{0x1bd480, true, true},
+	"AMEVCNTR05_EL0":    a64SysReg{0x1bd4a0, true, true},
+	"AMEVCNTR06_EL0":    a64SysReg{0x1bd4c0, true, true},
+	"AMEVCNTR07_EL0":    a64SysReg{0x1bd4e0, true, true},
+	"AMEVCNTR08_EL0":    a64SysReg{0x1bd500, true, true},
+	"AMEVCNTR09_EL0":    a64SysReg{0x1bd520, true, true},
+	"AMEVCNTR010_EL0":   a64SysReg{0x1bd540, true, true},
+	"AMEVCNTR011_EL0":   a64SysReg{0x1bd560, true, true},
+	"AMEVCNTR012_EL0":   a64SysReg{0x1bd580, true, true},
+	"AMEVCNTR013_EL0":   a64SysReg{0x1bd5a0, true, true},
+	"AMEVCNTR014_EL0":   a64SysReg{0x1bd5c0, true, true},
+	"AMEVCNTR015_EL0":   a64SysReg{0x1bd5e0, true, true},
+	"AMEVCNTR10_EL0":    a64SysReg{0x1bdc00, true, true},
+	"AMEVCNTR11_EL0":    a64SysReg{0x1bdc20, true, true},
+	"AMEVCNTR12_EL0":    a64SysReg{0x1bdc40, true, true},
+	"AMEVCNTR13_EL0":    a64SysReg{0x1bdc60, true, true},
+	"AMEVCNTR14_EL0":    a64SysReg{0x1bdc80, true, true},
+	"AMEVCNTR15_EL0":    a64SysReg{0x1bdca0, true, true},
+	"AMEVCNTR16_EL0":    a64SysReg{0x1bdcc0, true, true},
+	"AMEVCNTR17_EL0":    a64SysReg{0x1bdce0, true, true},
+	"AMEVCNTR18_EL0":    a64SysReg{0x1bdd00, true, true},
+	"AMEVCNTR19_EL0":    a64SysReg{0x1bdd20, true, true},
+	"AMEVCNTR110_EL0":   a64SysReg{0x1bdd40, true, true},
+	"AMEVCNTR111_EL0":   a64SysReg{0x1bdd60, true, true},
+	"AMEVCNTR112_EL0":   a64SysReg{0x1bdd80, true, true},
+	"AMEVCNTR113_EL0":   a64SysReg{0x1bdda0, true, true},
+	"AMEVCNTR114_EL0":   a64SysReg{0x1bddc0, true, true},
+	"AMEVCNTR115_EL0":   a64SysReg{0x1bdde0, true, true},
+	"AMEVTYPER00_EL0":   a64SysReg{0x1bd600, true, false},
+	"AMEVTYPER01_EL0":   a64SysReg{0x1bd620, true, false},
+	"AMEVTYPER02_EL0":   a64SysReg{0x1bd640, true, false},
+	"AMEVTYPER03_EL0":   a64SysReg{0x1bd660, true, false},
+	"AMEVTYPER04_EL0":   a64SysReg{0x1bd680, true, false},
+	"AMEVTYPER05_EL0":   a64SysReg{0x1bd6a0, true, false},
+	"AMEVTYPER06_EL0":   a64SysReg{0x1bd6c0, true, false},
+	"AMEVTYPER07_EL0":   a64SysReg{0x1bd6e0, true, false},
+	"AMEVTYPER08_EL0":   a64SysReg{0x1bd700, true, false},
+	"AMEVTYPER09_EL0":   a64SysReg{0x1bd720, true, false},
+	"AMEVTYPER010_EL0":  a64SysReg{0x1bd740, true, false},
+	"AMEVTYPER011_EL0":  a64SysReg{0x1bd760, true, false},
+	"AMEVTYPER012_EL0":  a64SysReg{0x1bd780, true, false},
+	"AMEVTYPER013_EL0":  a64SysReg{0x1bd7a0, true, false},
+	"AMEVTYPER014_EL0":  a64SysReg{0x1bd7c0, true, false},
+	"AMEVTYPER015_EL0":  a64SysReg{0x1bd7e0, true, false},
+	"AMEVTYPER10_EL0":   a64SysReg{0x1bde00, true, true},
+	"AMEVTYPER11_EL0":   a64SysReg{0x1bde20, true, true},
+	"AMEVTYPER12_EL0":   a64SysReg{0x1bde40, true, true},
+	"AMEVTYPER13_EL0":   a64SysReg{0x1bde60, true, true},
+	"AMEVTYPER14_EL0":   a64SysReg{0x1bde80, true, true},
+	"AMEVTYPER15_EL0":   a64SysReg{0x1bdea0, true, true},
+	"AMEVTYPER16_EL0":   a64SysReg{0x1bdec0, true, true},
+	"AMEVTYPER17_EL0":   a64SysReg{0x1bdee0, true, true},
+	"AMEVTYPER18_EL0":   a64SysReg{0x1bdf00, true, true},
+	"AMEVTYPER19_EL0":   a64SysReg{0x1bdf20, true, true},
+	"AMEVTYPER110_EL0":  a64SysReg{0x1bdf40, true, true},
+	"AMEVTYPER111_EL0":  a64SysReg{0x1bdf60, true, true},
+	"AMEVTYPER112_EL0":  a64SysReg{0x1bdf80, true, true},
+	"AMEVTYPER113_EL0":  a64SysReg{0x1bdfa0, true, true},
+	"AMEVTYPER114_EL0":  a64SysReg{0x1bdfc0, true, true},
+	"AMEVTYPER115_EL0":  a64SysReg{0x1bdfe0, true, true},
+	"AMUSERENR_EL0":     a64SysReg{0x1bd260, true, true},
+	"APDAKeyHi_EL1":     a64SysReg{0x182220, true, true},
+	"APDAKeyLo_EL1":     a64SysReg{0x182200, true, true},
+	"APDBKeyHi_EL1":     a64SysReg{0x182260, true, true},
+	"APDBKeyLo_EL1":     a64SysReg{0x182240, true, true},
+	"APGAKeyHi_EL1":     a64SysReg{0x182320, true, true},
+	"APGAKeyLo_EL1":     a64SysReg{0x182300, true, true},
+	"APIAKeyHi_EL1":     a64SysReg{0x182120, true, true},
+	"APIAKeyLo_EL1":     a64SysReg{0x182100, true, true},
+	"APIBKeyHi_EL1":     a64SysReg{0x182160, true, true},
+	"APIBKeyLo_EL1":     a64SysReg{0x182140, true, true},
+	"CCSIDR2_EL1":       a64SysReg{0x190040, true, false},
+	"CCSIDR_EL1":        a64SysReg{0x190000, true, false},
+	"CLIDR_EL1":         a64SysReg{0x190020, true, false},
+	"CNTFRQ_EL0":        a64SysReg{0x1be000, true, true},
+	"CNTKCTL_EL1":       a64SysReg{0x18e100, true, true},
+	"CNTP_CTL_EL0":      a64SysReg{0x1be220, true, true},
+	"CNTP_CVAL_EL0":     a64SysReg{0x1be240, true, true},
+	"CNTP_TVAL_EL0":     a64SysReg{0x1be200, true, true},
+	"CNTPCT_EL0":        a64SysReg{0x1be020, true, false},
+	"CNTPS_CTL_EL1":     a64SysReg{0x1fe220, true, true},
+	"CNTPS_CVAL_EL1":    a64SysReg{0x1fe240, true, true},
+	"CNTPS_TVAL_EL1":    a64SysReg{0x1fe200, true, true},
+	"CNTV_CTL_EL0":      a64SysReg{0x1be320, true, true},
+	"CNTV_CVAL_EL0":     a64SysReg{0x1be340, true, true},
+	"CNTV_TVAL_EL0":     a64SysReg{0x1be300, true, true},
+	"CNTVCT_EL0":        a64SysReg{0x1be040, true, false},
+	"CONTEXTIDR_EL1":    a64SysReg{0x18d020, true, true},
+	"CPACR_EL1":         a64SysReg{0x181040, true, true},
+	"CSSELR_EL1":        a64SysReg{0x1a0000, true, true},
+	"CTR_EL0":           a64SysReg{0x1b0020, true, false},
+	"CurrentEL":         a64SysReg{0x184240, true, false},
+	"DAIF":              a64SysReg{0x1b4220, true, true},
+	"DBGAUTHSTATUS_EL1": a64SysReg{0x107ec0, true, false},
+	"DBGBCR0_EL1":       a64SysReg{0x1000a0, true, true},
+	"DBGBCR1_EL1":       a64SysReg{0x1001a0, true, true},
+	"DBGBCR2_EL1":       a64SysReg{0x1002a0, true, true},
+	"DBGBCR3_EL1":       a64SysReg{0x1003a0, true, true},
+	"DBGBCR4_EL1":       a64SysReg{0x1004a0, true, true},
+	"DBGBCR5_EL1":       a64SysReg{0x1005a0, true, true},
+	"DBGBCR6_EL1":       a64SysReg{0x1006a0, true, true},
+	"DBGBCR7_EL1":       a64SysReg{0x1007a0, true, true},
+	"DBGBCR8_EL1":       a64SysReg{0x1008a0, true, true},
+	"DBGBCR9_EL1":       a64SysReg{0x1009a0, true, true},
+	"DBGBCR10_EL1":      a64SysReg{0x100aa0, true, true},
+	"DBGBCR11_EL1":      a64SysReg{0x100ba0, true, true},
+	"DBGBCR12_EL1":      a64SysReg{0x100ca0, true, true},
+	"DBGBCR13_EL1":      a64SysReg{0x100da0, true, true},
+	"DBGBCR14_EL1":      a64SysReg{0x100ea0, true, true},
+	"DBGBCR15_EL1":      a64SysReg{0x100fa0, true, true},
+	"DBGBVR0_EL1":       a64SysReg{0x100080, true, true},
+	"DBGBVR1_EL1":       a64SysReg{0x100180, true, true},
+	"DBGBVR2_EL1":       a64SysReg{0x100280, true, true},
+	"DBGBVR3_EL1":       a64SysReg{0x100380, true, true},
+	"DBGBVR4_EL1":       a64SysReg{0x100480, true, true},
+	"DBGBVR5_EL1":       a64SysReg{0x100580, true, true},
+	"DBGBVR6_EL1":       a64SysReg{0x100680, true, true},
+	"DBGBVR7_EL1":       a64SysReg{0x100780, true, true},
+	"DBGBVR8_EL1":       a64SysReg{0x100880, true, true},
+	"DBGBVR9_EL1":       a64SysReg{0x100980, true, true},
+	"DBGBVR10_EL1":      a64SysReg{0x100a80, true, true},
+	"DBGBVR11_EL1":      a64SysReg{0x100b80, true, true},
+	"DBGBVR12_EL1":      a64SysReg{0x100c80, true, true},
+	"DBGBVR13_EL1":      a64SysReg{0x100d80, true, true},
+	"DBGBVR14_EL1":      a64SysReg{0x100e80, true, true},
+	"DBGBVR15_EL1":      a64SysReg{0x100f80, true, true},
+	"DBGCLAIMCLR_EL1":   a64SysReg{0x1079c0, true, true},
+	"DBGCLAIMSET_EL1":   a64SysReg{0x1078c0, true, true},
+	"DBGDTR_EL0":        a64SysReg{0x130400, true, true},
+	"DBGDTRRX_EL0":      a64SysReg{0x130500, true, false},
+	"DBGDTRTX_EL0":      a64SysReg{0x130500, false, true},
+	"DBGPRCR_EL1":       a64SysReg{0x101480, true, true},
+	"DBGWCR0_EL1":       a64SysReg{0x1000e0, true, true},
+	"DBGWCR1_EL1":       a64SysReg{0x1001e0, true, true},
+	"DBGWCR2_EL1":       a64SysReg{0x1002e0, true, true},
+	"DBGWCR3_EL1":       a64SysReg{0x1003e0, true, true},
+	"DBGWCR4_EL1":       a64SysReg{0x1004e0, true, true},
+	"DBGWCR5_EL1":       a64SysReg{0x1005e0, true, true},
+	"DBGWCR6_EL1":       a64SysReg{0x1006e0, true, true},
+	"DBGWCR7_EL1":       a64SysReg{0x1007e0, true, true},
+	"DBGWCR8_EL1":       a64SysReg{0x1008e0, true, true},
+	"DBGWCR9_EL1":       a64SysReg{0x1009e0, true, true},
+	"DBGWCR10_EL1":      a64SysReg{0x100ae0, true, true},
+	"DBGWCR11_EL1":      a64SysReg{0x100be0, true, true},
+	"DBGWCR12_EL1":      a64SysReg{0x100ce0, true, true},
+	"DBGWCR13_EL1":      a64SysReg{0x100de0, true, true},
+	"DBGWCR14_EL1":      a64SysReg{0x100ee0, true, true},
+	"DBGWCR15_EL1":      a64SysReg{0x100fe0, true, true},
+	"DBGWVR0_EL1":       a64SysReg{0x1000c0, true, true},
+	"DBGWVR1_EL1":       a64SysReg{0x1001c0, true, true},
+	"DBGWVR2_EL1":       a64SysReg{0x1002c0, true, true},
+	"DBGWVR3_EL1":       a64SysReg{0x1003c0, true, true},
+	"DBGWVR4_EL1":       a64SysReg{0x1004c0, true, true},
+	"DBGWVR5_EL1":       a64SysReg{0x1005c0, true, true},
+	"DBGWVR6_EL1":       a64SysReg{0x1006c0, true, true},
+	"DBGWVR7_EL1":       a64SysReg{0x1007c0, true, true},
+	"DBGWVR8_EL1":       a64SysReg{0x1008c0, true, true},
+	"DBGWVR9_EL1":       a64SysReg{0x1009c0, true, true},
+	"DBGWVR10_EL1":      a64SysReg{0x100ac0, true, true},
+	"DBGWVR11_EL1":      a64SysReg{0x100bc0, true, true},
+	"DBGWVR12_EL1":      a64SysReg{0x100cc0, true, true},
+	"DBGWVR13_EL1":      a64SysReg{0x100dc0, true, true},
+	"DBGWVR14_EL1":      a64SysReg{0x100ec0, true, true},
+	"DBGWVR15_EL1":      a64SysReg{0x100fc0, true, true},
+	"DCZID_EL0":         a64SysReg{0x1b00e0, true, false},
+	"DISR_EL1":          a64SysReg{0x18c120, true, true},
+	"DIT":               a64SysReg{0x1b42a0, true, true},
+	"DLR_EL0":           a64SysReg{0x1b4520, true, true},
+	"DSPSR_EL0":         a64SysReg{0x1b4500, true, true},
+	"ELR_EL1":           a64SysReg{0x184020, true, true},
+	"ERRIDR_EL1":        a64SysReg{0x185300, true, false},
+	"ERRSELR_EL1":       a64SysReg{0x185320, true, true},
+	"ERXADDR_EL1":       a64SysReg{0x185460, true, true},
+	"ERXCTLR_EL1":       a64SysReg{0x185420, true, true},
+	"ERXFR_EL1":         a64SysReg{0x185400, true, false},
+	"ERXMISC0_EL1":      a64SysReg{0x185500, true, true},
+	"ERXMISC1_EL1":      a64SysReg{0x185520, true, true},
+	"ERXMISC2_EL1":      a64SysReg{0x185540, true, true},
+	"ERXMISC3_EL1":      a64SysReg{0x185560, true, true},
+	"ERXPFGCDN_EL1":     a64SysReg{0x1854c0, true, true},
+	"ERXPFGCTL_EL1":     a64SysReg{0x1854a0, true, true},
+	"ERXPFGF_EL1":       a64SysReg{0x185480, true, false},
+	"ERXSTATUS_EL1":     a64SysReg{0x185440, true, true},
+	"ESR_EL1":           a64SysReg{0x185200, true, true},
+	"FAR_EL1":           a64SysReg{0x186000, true, true},
+	"FPCR":              a64SysReg{0x1b4400, true, true},
+	"FPSR":              a64SysReg{0x1b4420, true, true},
+	"GCR_EL1":           a64SysReg{0x1810c0, true, true},
+	"GMID_EL1":          a64SysReg{0x31400, true, false},
+	"ICC_AP0R0_EL1":     a64SysReg{0x18c880, true, true},
+	"ICC_AP0R1_EL1":     a64SysReg{0x18c8a0, true, true},
+	"ICC_AP0R2_EL1":     a64SysReg{0x18c8c0, true, true},
+	"ICC_AP0R3_EL1":     a64SysReg{0x18c8e0, true, true},
+	"ICC_AP1R0_EL1":     a64SysReg{0x18c900, true, true},
+	"ICC_AP1R1_EL1":     a64SysReg{0x18c920, true, true},
+	"ICC_AP1R2_EL1":     a64SysReg{0x18c940, true, true},
+	"ICC_AP1R3_EL1":     a64SysReg{0x18c960, true, true},
+	"ICC_ASGI1R_EL1":    a64SysReg{0x18cbc0, false, true},
+	"ICC_BPR0_EL1":      a64SysReg{0x18c860, true, true},
+	"ICC_BPR1_EL1":      a64SysReg{0x18cc60, true, true},
+	"ICC_CTLR_EL1":      a64SysReg{0x18cc80, true, true},
+	"ICC_DIR_EL1":       a64SysReg{0x18cb20, false, true},
+	"ICC_EOIR0_EL1":     a64SysReg{0x18c820, false, true},
+	"ICC_EOIR1_EL1":     a64SysReg{0x18cc20, false, true},
+	"ICC_HPPIR0_EL1":    a64SysReg{0x18c840, true, false},
+	"ICC_HPPIR1_EL1":    a64SysReg{0x18cc40, true, false},
+	"ICC_IAR0_EL1":      a64SysReg{0x18c800, true, false},
+	"ICC_IAR1_EL1":      a64SysReg{0x18cc00, true, false},
+	"ICC_IGRPEN0_EL1":   a64SysReg{0x18ccc0, true, true},
+	"ICC_IGRPEN1_EL1":   a64SysReg{0x18cce0, true, true},
+	"ICC_PMR_EL1":       a64SysReg{0x184600, true, true},
+	"ICC_RPR_EL1":       a64SysReg{0x18cb60, true, false},
+	"ICC_SGI0R_EL1":     a64SysReg{0x18cbe0, false, true},
+	"ICC_SGI1R_EL1":     a64SysReg{0x18cba0, false, true},
+	"ICC_SRE_EL1":       a64SysReg{0x18cca0, true, true},
+	"ICV_AP0R0_EL1":     a64SysReg{0x18c880, true, true},
+	"ICV_AP0R1_EL1":     a64SysReg{0x18c8a0, true, true},
+	"ICV_AP0R2_EL1":     a64SysReg{0x18c8c0, true, true},
+	"ICV_AP0R3_EL1":     a64SysReg{0x18c8e0, true, true},
+	"ICV_AP1R0_EL1":     a64SysReg{0x18c900, true, true},
+	"ICV_AP1R1_EL1":     a64SysReg{0x18c920, true, true},
+	"ICV_AP1R2_EL1":     a64SysReg{0x18c940, true, true},
+	"ICV_AP1R3_EL1":     a64SysReg{0x18c960, true, true},
+	"ICV_BPR0_EL1":      a64SysReg{0x18c860, true, true},
+	"ICV_BPR1_EL1":      a64SysReg{0x18cc60, true, true},
+	"ICV_CTLR_EL1":      a64SysReg{0x18cc80, true, true},
+	"ICV_DIR_EL1":       a64SysReg{0x18cb20, false, true},
+	"ICV_EOIR0_EL1":     a64SysReg{0x18c820, false, true},
+	"ICV_EOIR1_EL1":     a64SysReg{0x18cc20, false, true},
+	"ICV_HPPIR0_EL1":    a64SysReg{0x18c840, true, false},
+	"ICV_HPPIR1_EL1":    a64SysReg{0x18cc40, true, false},
+	"ICV_IAR0_EL1":      a64SysReg{0x18c800, true, false},
+	"ICV_IAR1_EL1":      a64SysReg{0x18cc00, true, false},
+	"ICV_IGRPEN0_EL1":   a64SysReg{0x18ccc0, true, true},
+	"ICV_IGRPEN1_EL1":   a64SysReg{0x18cce0, true, true},
+	"ICV_PMR_EL1":       a64SysReg{0x184600, true, true},
+	"ICV_RPR_EL1":       a64SysReg{0x18cb60, true, false},
+	"ID_AA64AFR0_EL1":   a64SysReg{0x180580, true, false},
+	"ID_AA64AFR1_EL1":   a64SysReg{0x1805a0, true, false},
+	"ID_AA64DFR0_EL1":   a64SysReg{0x180500, true, false},
+	"ID_AA64DFR1_EL1":   a64SysReg{0x180520, true, false},
+	"ID_AA64ISAR0_EL1":  a64SysReg{0x180600, true, false},
+	"ID_AA64ISAR1_EL1":  a64SysReg{0x180620, true, false},
+	"ID_AA64MMFR0_EL1":  a64SysReg{0x180700, true, false},
+	"ID_AA64MMFR1_EL1":  a64SysReg{0x180720, true, false},
+	"ID_AA64MMFR2_EL1":  a64SysReg{0x180740, true, false},
+	"ID_AA64PFR0_EL1":   a64SysReg{0x180400, true, false},
+	"ID_AA64PFR1_EL1":   a64SysReg{0x180420, true, false},
+	"ID_AA64ZFR0_EL1":   a64SysReg{0x180480, true, false},
+	"ID_AFR0_EL1":       a64SysReg{0x180160, true, false},
+	"ID_DFR0_EL1":       a64SysReg{0x180140, true, false},
+	"ID_ISAR0_EL1":      a64SysReg{0x180200, true, false},
+	"ID_ISAR1_EL1":      a64SysReg{0x180220, true, false},
+	"ID_ISAR2_EL1":      a64SysReg{0x180240, true, false},
+	"ID_ISAR3_EL1":      a64SysReg{0x180260, true, false},
+	"ID_ISAR4_EL1":      a64SysReg{0x180280, true, false},
+	"ID_ISAR5_EL1":      a64SysReg{0x1802a0, true, false},
+	"ID_ISAR6_EL1":      a64SysReg{0x1802e0, true, false},
+	"ID_MMFR0_EL1":      a64SysReg{0x180180, true, false},
+	"ID_MMFR1_EL1":      a64SysReg{0x1801a0, true, false},
+	"ID_MMFR2_EL1":      a64SysReg{0x1801c0, true, false},
+	"ID_MMFR3_EL1":      a64SysReg{0x1801e0, true, false},
+	"ID_MMFR4_EL1":      a64SysReg{0x1802c0, true, false},
+	"ID_PFR0_EL1":       a64SysReg{0x180100, true, false},
+	"ID_PFR1_EL1":       a64SysReg{0x180120, true, false},
+	"ID_PFR2_EL1":       a64SysReg{0x180380, true, false},
+	"ISR_EL1":           a64SysReg{0x18c100, true, false},
+	"LORC_EL1":          a64SysReg{0x18a460, true, true},
+	"LOREA_EL1":         a64SysReg{0x18a420, true, true},
+	"LORID_EL1":         a64SysReg{0x18a4e0, true, false},
+	"LORN_EL1":          a64SysReg{0x18a440, true, true},
+	"LORSA_EL1":         a64SysReg{0x18a400, true, true},
+	"MAIR_EL1":          a64SysReg{0x18a200, true, true},
+	"MDCCINT_EL1":       a64SysReg{0x100200, true, true},
+	"MDCCSR_EL0":        a64SysReg{0x130100, true, false},
+	"MDRAR_EL1":         a64SysReg{0x101000, true, false},
+	"MDSCR_EL1":         a64SysReg{0x100240, true, true},
+	"MIDR_EL1":          a64SysReg{0x180000, true, false},
+	"MPAM0_EL1":         a64SysReg{0x18a520, true, true},
+	"MPAM1_EL1":         a64SysReg{0x18a500, true, true},
+	"MPAMIDR_EL1":       a64SysReg{0x18a480, true, false},
+	"MPIDR_EL1":         a64SysReg{0x1800a0, true, false},
+	"MVFR0_EL1":         a64SysReg{0x180300, true, false},
+	"MVFR1_EL1":         a64SysReg{0x180320, true, false},
+	"MVFR2_EL1":         a64SysReg{0x180340, true, false},
+	"NZCV":              a64SysReg{0x1b4200, true, true},
+	"OSDLR_EL1":         a64SysReg{0x101380, true, true},
+	"OSDTRRX_EL1":       a64SysReg{0x100040, true, true},
+	"OSDTRTX_EL1":       a64SysReg{0x100340, true, true},
+	"OSECCR_EL1":        a64SysReg{0x100640, true, true},
+	"OSLAR_EL1":         a64SysReg{0x101080, false, true},
+	"OSLSR_EL1":         a64SysReg{0x101180, true, false},
+	"PAN":               a64SysReg{0x184260, true, true},
+	"PAR_EL1":           a64SysReg{0x187400, true, true},
+	"PMBIDR_EL1":        a64SysReg{0x189ae0, true, false},
+	"PMBLIMITR_EL1":     a64SysReg{0x189a00, true, true},
+	"PMBPTR_EL1":        a64SysReg{0x189a20, true, true},
+	"PMBSR_EL1":         a64SysReg{0x189a60, true, true},
+	"PMCCFILTR_EL0":     a64SysReg{0x1befe0, true, true},
+	"PMCCNTR_EL0":       a64SysReg{0x1b9d00, true, true},
+	"PMCEID0_EL0":       a64SysReg{0x1b9cc0, true, false},
+	"PMCEID1_EL0":       a64SysReg{0x1b9ce0, true, false},
+	"PMCNTENCLR_EL0":    a64SysReg{0x1b9c40, true, true},
+	"PMCNTENSET_EL0":    a64SysReg{0x1b9c20, true, true},
+	"PMCR_EL0":          a64SysReg{0x1b9c00, true, true},
+	"PMEVCNTR0_EL0":     a64SysReg{0x1be800, true, true},
+	"PMEVCNTR1_EL0":     a64SysReg{0x1be820, true, true},
+	"PMEVCNTR2_EL0":     a64SysReg{0x1be840, true, true},
+	"PMEVCNTR3_EL0":     a64SysReg{0x1be860, true, true},
+	"PMEVCNTR4_EL0":     a64SysReg{0x1be880, true, true},
+	"PMEVCNTR5_EL0":     a64SysReg{0x1be8a0, true, true},
+	"PMEVCNTR6_EL0":     a64SysReg{0x1be8c0, true, true},
+	"PMEVCNTR7_EL0":     a64SysReg{0x1be8e0, true, true},
+	"PMEVCNTR8_EL0":     a64SysReg{0x1be900, true, true},
+	"PMEVCNTR9_EL0":     a64SysReg{0x1be920, true, true},
+	"PMEVCNTR10_EL0":    a64SysReg{0x1be940, true, true},
+	"PMEVCNTR11_EL0":    a64SysReg{0x1be960, true, true},
+	"PMEVCNTR12_EL0":    a64SysReg{0x1be980, true, true},
+	"PMEVCNTR13_EL0":    a64SysReg{0x1be9a0, true, true},
+	"PMEVCNTR14_EL0":    a64SysReg{0x1be9c0, true, true},
+	"PMEVCNTR15_EL0":    a64SysReg{0x1be9e0, true, true},
+	"PMEVCNTR16_EL0":    a64SysReg{0x1bea00, true, true},
+	"PMEVCNTR17_EL0":    a64SysReg{0x1bea20, true, true},
+	"PMEVCNTR18_EL0":    a64SysReg{0x1bea40, true, true},
+	"PMEVCNTR19_EL0":    a64SysReg{0x1bea60, true, true},
+	"PMEVCNTR20_EL0":    a64SysReg{0x1bea80, true, true},
+	"PMEVCNTR21_EL0":    a64SysReg{0x1beaa0, true, true},
+	"PMEVCNTR22_EL0":    a64SysReg{0x1beac0, true, true},
+	"PMEVCNTR23_EL0":    a64SysReg{0x1beae0, true, true},
+	"PMEVCNTR24_EL0":    a64SysReg{0x1beb00, true, true},
+	"PMEVCNTR25_EL0":    a64SysReg{0x1beb20, true, true},
+	"PMEVCNTR26_EL0":    a64SysReg{0x1beb40, true, true},
+	"PMEVCNTR27_EL0":    a64SysReg{0x1beb60, true, true},
+	"PMEVCNTR28_EL0":    a64SysReg{0x1beb80, true, true},
+	"PMEVCNTR29_EL0":    a64SysReg{0x1beba0, true, true},
+	"PMEVCNTR30_EL0":    a64SysReg{0x1bebc0, true, true},
+	"PMEVTYPER0_EL0":    a64SysReg{0x1bec00, true, true},
+	"PMEVTYPER1_EL0":    a64SysReg{0x1bec20, true, true},
+	"PMEVTYPER2_EL0":    a64SysReg{0x1bec40, true, true},
+	"PMEVTYPER3_EL0":    a64SysReg{0x1bec60, true, true},
+	"PMEVTYPER4_EL0":    a64SysReg{0x1bec80, true, true},
+	"PMEVTYPER5_EL0":    a64SysReg{0x1beca0, true, true},
+	"PMEVTYPER6_EL0":    a64SysReg{0x1becc0, true, true},
+	"PMEVTYPER7_EL0":    a64SysReg{0x1bece0, true, true},
+	"PMEVTYPER8_EL0":    a64SysReg{0x1bed00, true, true},
+	"PMEVTYPER9_EL0":    a64SysReg{0x1bed20, true, true},
+	"PMEVTYPER10_EL0":   a64SysReg{0x1bed40, true, true},
+	"PMEVTYPER11_EL0":   a64SysReg{0x1bed60, true, true},
+	"PMEVTYPER12_EL0":   a64SysReg{0x1bed80, true, true},
+	"PMEVTYPER13_EL0":   a64SysReg{0x1beda0, true, true},
+	"PMEVTYPER14_EL0":   a64SysReg{0x1bedc0, true, true},
+	"PMEVTYPER15_EL0":   a64SysReg{0x1bede0, true, true},
+	"PMEVTYPER16_EL0":   a64SysReg{0x1bee00, true, true},
+	"PMEVTYPER17_EL0":   a64SysReg{0x1bee20, true, true},
+	"PMEVTYPER18_EL0":   a64SysReg{0x1bee40, true, true},
+	"PMEVTYPER19_EL0":   a64SysReg{0x1bee60, true, true},
+	"PMEVTYPER20_EL0":   a64SysReg{0x1bee80, true, true},
+	"PMEVTYPER21_EL0":   a64SysReg{0x1beea0, true, true},
+	"PMEVTYPER22_EL0":   a64SysReg{0x1beec0, true, true},
+	"PMEVTYPER23_EL0":   a64SysReg{0x1beee0, true, true},
+	"PMEVTYPER24_EL0":   a64SysReg{0x1bef00, true, true},
+	"PMEVTYPER25_EL0":   a64SysReg{0x1bef20, true, true},
+	"PMEVTYPER26_EL0":   a64SysReg{0x1bef40, true, true},
+	"PMEVTYPER27_EL0":   a64SysReg{0x1bef60, true, true},
+	"PMEVTYPER28_EL0":   a64SysReg{0x1bef80, true, true},
+	"PMEVTYPER29_EL0":   a64SysReg{0x1befa0, true, true},
+	"PMEVTYPER30_EL0":   a64SysReg{0x1befc0, true, true},
+	"PMINTENCLR_EL1":    a64SysReg{0x189e40, true, true},
+	"PMINTENSET_EL1":    a64SysReg{0x189e20, true, true},
+	"PMMIR_EL1":         a64SysReg{0x189ec0, true, false},
+	"PMOVSCLR_EL0":      a64SysReg{0x1b9c60, true, true},
+	"PMOVSSET_EL0":      a64SysReg{0x1b9e60, true, true},
+	"PMSCR_EL1":         a64SysReg{0x189900, true, true},
+	"PMSELR_EL0":        a64SysReg{0x1b9ca0, true, true},
+	"PMSEVFR_EL1":       a64SysReg{0x1899a0, true, true},
+	"PMSFCR_EL1":        a64SysReg{0x189980, true, true},
+	"PMSICR_EL1":        a64SysReg{0x189940, true, true},
+	"PMSIDR_EL1":        a64SysReg{0x1899e0, true, false},
+	"PMSIRR_EL1":        a64SysReg{0x189960, true, true},
+	"PMSLATFR_EL1":      a64SysReg{0x1899c0, true, true},
+	"PMSWINC_EL0":       a64SysReg{0x1b9c80, false, true},
+	"PMUSERENR_EL0":     a64SysReg{0x1b9e00, true, true},
+	"PMXEVCNTR_EL0":     a64SysReg{0x1b9d40, true, true},
+	"PMXEVTYPER_EL0":    a64SysReg{0x1b9d20, true, true},
+	"REVIDR_EL1":        a64SysReg{0x1800c0, true, false},
+	"RGSR_EL1":          a64SysReg{0x1810a0, true, true},
+	"RMR_EL1":           a64SysReg{0x18c040, true, true},
+	"RNDR":              a64SysReg{0x1b2400, true, false},
+	"RNDRRS":            a64SysReg{0x1b2420, true, false},
+	"RVBAR_EL1":         a64SysReg{0x18c020, true, false},
+	"SCTLR_EL1":         a64SysReg{0x181000, true, true},
+	"SCXTNUM_EL0":       a64SysReg{0x1bd0e0, true, true},
+	"SCXTNUM_EL1":       a64SysReg{0x18d0e0, true, true},
+	"SP_EL0":            a64SysReg{0x184100, true, true},
+	"SP_EL1":            a64SysReg{0x1c4100, true, true},
+	"SPSel":             a64SysReg{0x184200, true, true},
+	"SPSR_abt":          a64SysReg{0x1c4320, true, true},
+	"SPSR_EL1":          a64SysReg{0x184000, true, true},
+	"SPSR_fiq":          a64SysReg{0x1c4360, true, true},
+	"SPSR_irq":          a64SysReg{0x1c4300, true, true},
+	"SPSR_und":          a64SysReg{0x1c4340, true, true},
+	"SSBS":              a64SysReg{0x1b42c0, true, true},
+	"TCO":               a64SysReg{0x1b42e0, true, true},
+	"TCR_EL1":           a64SysReg{0x182040, true, true},
+	"TFSR_EL1":          a64SysReg{0x185600, true, true},
+	"TFSRE0_EL1":        a64SysReg{0x185620, true, true},
+	"TPIDR_EL0":         a64SysReg{0x1bd040, true, true},
+	"TPIDR_EL1":         a64SysReg{0x18d080, true, true},
+	"TPIDRRO_EL0":       a64SysReg{0x1bd060, true, true},
+	"TRFCR_EL1":         a64SysReg{0x181220, true, true},
+	"TTBR0_EL1":         a64SysReg{0x182000, true, true},
+	"TTBR1_EL1":         a64SysReg{0x182020, true, true},
+	"UAO":               a64SysReg{0x184280, true, true},
+	"VBAR_EL1":          a64SysReg{0x18c000, true, true},
+	"ZCR_EL1":           a64SysReg{0x181200, true, true},
+}
+
+// a64SysInst is one TLBI alias: the fields the SYS encoding carries beside
+// the fixed op0 = 01 and CRn = 8.
+type a64SysInst struct {
+	op1, cm, op2 uint32
+}
+
+// a64TLBIOps maps the TLBI operation names to their fields; the register
+// operand is optional and defaults to ZR.
+var a64TLBIOps = map[string]a64SysInst{
+	"ALLE1":        {0x4, 0x7, 0x4},
+	"ALLE1IS":      {0x4, 0x3, 0x4},
+	"ALLE1OS":      {0x4, 0x1, 0x4},
+	"ALLE2":        {0x4, 0x7, 0x0},
+	"ALLE2IS":      {0x4, 0x3, 0x0},
+	"ALLE2OS":      {0x4, 0x1, 0x0},
+	"ALLE3":        {0x6, 0x7, 0x0},
+	"ALLE3IS":      {0x6, 0x3, 0x0},
+	"ALLE3OS":      {0x6, 0x1, 0x0},
+	"ASIDE1":       {0x0, 0x7, 0x2},
+	"ASIDE1IS":     {0x0, 0x3, 0x2},
+	"ASIDE1OS":     {0x0, 0x1, 0x2},
+	"IPAS2E1":      {0x4, 0x4, 0x1},
+	"IPAS2E1IS":    {0x4, 0x0, 0x1},
+	"IPAS2E1OS":    {0x4, 0x4, 0x0},
+	"IPAS2LE1":     {0x4, 0x4, 0x5},
+	"IPAS2LE1IS":   {0x4, 0x0, 0x5},
+	"IPAS2LE1OS":   {0x4, 0x4, 0x4},
+	"RIPAS2E1":     {0x4, 0x4, 0x2},
+	"RIPAS2E1IS":   {0x4, 0x0, 0x2},
+	"RIPAS2E1OS":   {0x4, 0x4, 0x3},
+	"RIPAS2LE1":    {0x4, 0x4, 0x6},
+	"RIPAS2LE1IS":  {0x4, 0x0, 0x6},
+	"RIPAS2LE1OS":  {0x4, 0x4, 0x7},
+	"RVAAE1":       {0x0, 0x6, 0x3},
+	"RVAAE1IS":     {0x0, 0x2, 0x3},
+	"RVAAE1OS":     {0x0, 0x5, 0x3},
+	"RVAALE1":      {0x0, 0x6, 0x7},
+	"RVAALE1IS":    {0x0, 0x2, 0x7},
+	"RVAALE1OS":    {0x0, 0x5, 0x7},
+	"RVAE1":        {0x0, 0x6, 0x1},
+	"RVAE1IS":      {0x0, 0x2, 0x1},
+	"RVAE1OS":      {0x0, 0x5, 0x1},
+	"RVAE2":        {0x4, 0x6, 0x1},
+	"RVAE2IS":      {0x4, 0x2, 0x1},
+	"RVAE2OS":      {0x4, 0x5, 0x1},
+	"RVAE3":        {0x6, 0x6, 0x1},
+	"RVAE3IS":      {0x6, 0x2, 0x1},
+	"RVAE3OS":      {0x6, 0x5, 0x1},
+	"RVALE1":       {0x0, 0x6, 0x5},
+	"RVALE1IS":     {0x0, 0x2, 0x5},
+	"RVALE1OS":     {0x0, 0x5, 0x5},
+	"RVALE2":       {0x4, 0x6, 0x5},
+	"RVALE2IS":     {0x4, 0x2, 0x5},
+	"RVALE2OS":     {0x4, 0x5, 0x5},
+	"RVALE3":       {0x6, 0x6, 0x5},
+	"RVALE3IS":     {0x6, 0x2, 0x5},
+	"RVALE3OS":     {0x6, 0x5, 0x5},
+	"VAAE1":        {0x0, 0x7, 0x3},
+	"VAAE1IS":      {0x0, 0x3, 0x3},
+	"VAAE1OS":      {0x0, 0x1, 0x3},
+	"VAALE1":       {0x0, 0x7, 0x7},
+	"VAALE1IS":     {0x0, 0x3, 0x7},
+	"VAALE1OS":     {0x0, 0x1, 0x7},
+	"VAE1":         {0x0, 0x7, 0x1},
+	"VAE1IS":       {0x0, 0x3, 0x1},
+	"VAE1OS":       {0x0, 0x1, 0x1},
+	"VAE2":         {0x4, 0x7, 0x1},
+	"VAE2IS":       {0x4, 0x3, 0x1},
+	"VAE2OS":       {0x4, 0x1, 0x1},
+	"VAE3":         {0x6, 0x7, 0x1},
+	"VAE3IS":       {0x6, 0x3, 0x1},
+	"VAE3OS":       {0x6, 0x1, 0x1},
+	"VALE1":        {0x0, 0x7, 0x5},
+	"VALE1IS":      {0x0, 0x3, 0x5},
+	"VALE1OS":      {0x0, 0x1, 0x5},
+	"VALE2":        {0x4, 0x7, 0x5},
+	"VALE2IS":      {0x4, 0x3, 0x5},
+	"VALE2OS":      {0x4, 0x1, 0x5},
+	"VALE3":        {0x6, 0x7, 0x5},
+	"VALE3IS":      {0x6, 0x3, 0x5},
+	"VALE3OS":      {0x6, 0x1, 0x5},
+	"VMALLE1":      {0x0, 0x7, 0x0},
+	"VMALLE1IS":    {0x0, 0x3, 0x0},
+	"VMALLE1OS":    {0x0, 0x1, 0x0},
+	"VMALLS12E1":   {0x4, 0x7, 0x6},
+	"VMALLS12E1IS": {0x4, 0x3, 0x6},
+	"VMALLS12E1OS": {0x4, 0x1, 0x6},
+}
+
+// a64DCOps2 maps the DC operation names to their fields; the register
+// operand is mandatory.
+var a64DCOps2 = map[string]a64SysInst{
+	"CGDSW":   {0x0, 0xa, 0x6},
+	"CGDVAC":  {0x3, 0xa, 0x5},
+	"CGDVADP": {0x3, 0xd, 0x5},
+	"CGDVAP":  {0x3, 0xc, 0x5},
+	"CGSW":    {0x0, 0xa, 0x4},
+	"CGVAC":   {0x3, 0xa, 0x3},
+	"CGVADP":  {0x3, 0xd, 0x3},
+	"CGVAP":   {0x3, 0xc, 0x3},
+	"CIGDSW":  {0x0, 0xe, 0x6},
+	"CIGDVAC": {0x3, 0xe, 0x5},
+	"CIGSW":   {0x0, 0xe, 0x4},
+	"CIGVAC":  {0x3, 0xe, 0x3},
+	"CISW":    {0x0, 0xe, 0x2},
+	"CIVAC":   {0x3, 0xe, 0x1},
+	"CSW":     {0x0, 0xa, 0x2},
+	"CVAC":    {0x3, 0xa, 0x1},
+	"CVADP":   {0x3, 0xd, 0x1},
+	"CVAP":    {0x3, 0xc, 0x1},
+	"CVAU":    {0x3, 0xb, 0x1},
+	"GVA":     {0x3, 0x4, 0x3},
+	"GZVA":    {0x3, 0x4, 0x4},
+	"IGDSW":   {0x0, 0x6, 0x6},
+	"IGDVAC":  {0x0, 0x6, 0x5},
+	"IGSW":    {0x0, 0x6, 0x4},
+	"IGVAC":   {0x0, 0x6, 0x3},
+	"ISW":     {0x0, 0x6, 0x2},
+	"IVAC":    {0x0, 0x6, 0x1},
+	"ZVA":     {0x3, 0x4, 0x1},
+}
+
+// a64RPRFOps maps the range-prefetch operation names to their 6-bit values.
+var a64RPRFOps = map[string]uint32{
+	"PLDKEEP": 0,
+	"PLDSTRM": 4,
+	"PSTKEEP": 1,
+	"PSTSTRM": 5,
+}

asm/arm64_sysregs_test.go

@@ -0,0 +1,164 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package asm
+
+import (
+	"encoding/binary"
+	"fmt"
+	"os"
+	"path/filepath"
+	"slices"
+	"strings"
+	"testing"
+
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
+)
+
+// TestARM64SysRegsDifferential proves the whole system-register table against
+// the toolchain at once: one TEXT whose body reads every register the table
+// carries (and writes every writable one), assembled by gasm and by
+// go tool asm, must agree byte for byte.  A single wrong op0/op1/CRn/CRm/op2
+// packing names its register through the first differing word.
+func TestARM64SysRegsDifferential(t *testing.T) {
+	names := make([]string, 0, len(a64SysRegs))
+	for name := range a64SysRegs {
+		names = append(names, name)
+	}
+	slices.Sort(names)
+
+	var body strings.Builder
+	for i, name := range names {
+		// R18 is the arm64 platform register and R29-R31 carry dedicated
+		// meanings; a plain read/write destination keeps to R0-R17.
+		reg := fmt.Sprintf("R%d", i%18)
+		if a64SysRegs[name].read {
+			body.WriteString(fmt.Sprintf("\tMRS %s, %s\n", name, reg))
+		}
+		if a64SysRegs[name].write {
+			body.WriteString(fmt.Sprintf("\tMSR %s, %s\n", reg, name))
+		}
+	}
+	src := "#include \"textflag.h\"\n\nTEXT ·sysregs(SB), NOSPLIT, $0\n" + body.String() + "\tRET\n"
+
+	dir := t.TempDir()
+	path := filepath.Join(dir, "sysregs_arm64.s")
+	if err := os.WriteFile(path, []byte(src), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	assertARM64Differential(t, path, src, "sysregs")
+}
+
+// TestARM64FamiliesDifferential pins the non-sysreg families the arm64
+// campaign added: the LSE compare-and-swap pairs, the VMOVI immediate, the
+// SIMD narrow/long shift pairs, the VLD2/VLD3/VLD4 and VST2/VST3/VST4
+// structure accesses with their post-index and replicate forms, LDPSW, the
+// pointer-authentication hint and the DC maintenance operation.  Every
+// spelling is the toolchain's own, taken from its arm64 testdata, and the
+// bytes must agree word for word.
+func TestARM64FamiliesDifferential(t *testing.T) {
+	src := `#include "textflag.h"
+
+TEXT ·families(SB), NOSPLIT, $0
+	CASPD (R2, R3), (R2), (R8, R9)
+	CASPW (R6, R7), (R8), (R4, R5)
+	VMOVI $82, V0.B16
+	VMOVI $146, V22.B16
+	VSSHLL $0, V1.B8, V2.H8
+	VSSHLL $7, V1.B8, V2.H8
+	VSSHLL2 $0, V1.B16, V2.H8
+	VSHRN $7, V1.H8, V0.B8
+	VSHRN2 $31, V1.D2, V0.S4
+	VLD2 (R29), [V23.H8, V24.H8]
+	VLD2.P 16(R0), [V18.B8, V19.B8]
+	VLD2.P (R1)(R2), [V15.S2, V16.S2]
+	VLD3 (R27), [V11.S4, V12.S4, V13.S4]
+	VLD3.P 48(RSP), [V11.S4, V12.S4, V13.S4]
+	VLD4 (R15), [V10.H4, V11.H4, V12.H4, V13.H4]
+	VLD4.P 32(R24), [V31.B8, V0.B8, V1.B8, V2.B8]
+	VLD1R (R1), [V9.B8]
+	VLD1R.P (R0), [V0.B16]
+	VLD1R.P 2(R1), [V2.H4]
+	VLD2R (R15), [V15.H4, V16.H4]
+	VLD2R.P 16(R0), [V0.D2, V1.D2]
+	VLD4R (R0), [V0.B8, V1.B8, V2.B8, V3.B8]
+	VLD4R.P 16(RSP), [V31.S4, V0.S4, V1.S4, V2.S4]
+	VST2 [V22.H8, V23.H8], (R23)
+	VST2.P [V14.H4, V15.H4], 16(R17)
+	VST2.P [V14.H4, V15.H4], (R3)(R17)
+	VST3 [V1.D2, V2.D2, V3.D2], (R11)
+	VST3.P [V18.S4, V19.S4, V20.S4], 48(R25)
+	VST4 [V22.D2, V23.D2, V24.D2, V25.D2], (R3)
+	VST4.P [V14.D2, V15.D2, V16.D2, V17.D2], 64(R15)
+	LDPSW (R0), (R1, R2)
+	LDPSW 4(R0), (R1, R2)
+	LDPSW -4(R0), (R1, R2)
+	PACIASP
+	DC IVAC, R1
+	RET
+`
+	dir := t.TempDir()
+	path := filepath.Join(dir, "families_arm64.s")
+	if err := os.WriteFile(path, []byte(src), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	assertARM64Differential(t, path, src, "families")
+}
+
+// assertARM64Differential assembles the same source with gasm and with the
+// toolchain for arm64 and requires the named function's code bytes to agree.
+// The live oracle is a deliberate-run comparison, so -short skips it (the
+// push pipeline's mode); the golden bytes of the individual encoders are
+// pinned separately in every mode.
+func assertARM64Differential(t *testing.T, path, src, fn string) {
+	t.Helper()
+	oracle := oracleFuncCode(t, toolAsmObject(t, path, "arm64"))
+
+	// The oracle keys its functions by the qualified object name
+	// (pkg.name); match on the local part.
+	want := map[string][]byte{}
+	for name, code := range oracle {
+		if _, after, ok := strings.Cut(name, "."); ok {
+			want[after] = code
+		} else {
+			want[name] = code
+		}
+	}
+	if want[fn] == nil {
+		t.Fatalf("the oracle object carries no function %q (has %v)", fn, keysOf(want))
+	}
+
+	f, perrs := parser.Parse(path, src)
+	if len(perrs) > 0 {
+		t.Fatalf("parse: %v", perrs[0])
+	}
+	img, err := AssembleFileARM64(f)
+	if err != nil {
+		t.Fatalf("AssembleFileARM64: %v", err)
+	}
+	got := trimTrailingZeroWords(img.Code)
+	wantB := trimTrailingZeroWords(want[fn])
+	if len(got) != len(wantB) {
+		t.Fatalf("gasm %d bytes, oracle %d bytes", len(got), len(wantB))
+	}
+	for i := range wantB {
+		if got[i] != wantB[i] {
+			t.Fatalf("word %d differs: gasm %08x, oracle %08x", i/4,
+				binary.LittleEndian.Uint32(got[i:i+4]), binary.LittleEndian.Uint32(wantB[i:i+4]))
+		}
+	}
+}
+
+// trimTrailingZeroWords drops whole zero words off the end of a code span:
+// an object pads a function to its alignment, and the raw image does not.
+// A difference in the middle survives the trim untouched.
+func trimTrailingZeroWords(b []byte) []byte {
+	for len(b) >= 4 {
+		last := b[len(b)-4:]
+		if last[0]|last[1]|last[2]|last[3] != 0 {
+			break
+		}
+		b = b[:len(b)-4]
+	}
+	return b
+}

asm/assemble.go

@@ -8,7 +8,7 @@ import (
 	"strconv"
 	"strings"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/ast"
+	"sourcedock.dev/petrbalvin/gasm-sdk/ast"
 )
 
 // Assemble encodes the body of a TEXT function into x86-64 machine code,
@@ -38,10 +38,25 @@ func Assemble(t *ast.Text) ([]byte, map[string]int, error) {
 // rejects SB operands outright (single-function assembly cannot resolve
 // them).  When allowExternal is set, a reference to a symbol no GLOBL in the
 // file defines is recorded as an external relocation instead of failing
-// the object-file emitters resolve it at link time.
+// the object-file emitters resolve it at link time.  goos selects the TLS
+// access form: the empty default behaves as linux.
 type linkInfo struct {
 	symbols       map[string]bool
 	allowExternal bool
+	goos          string
+}
+
+// tlsOneInsn reports the one-instruction TLS form, obj6.go's
+// CanUse1InsnTLS for the GOOS gasm supports: the bare TLS load nops out and
+// the (TLS*1) index folds to a segment-absolute access.  Windows and plan9
+// keep the two-instruction form; shared linux does too, which gasm's raw
+// path does not model and therefore does not select.
+func (l *linkInfo) tlsOneInsn() bool {
+	switch l.goos {
+	case "", "linux", "freebsd":
+		return true
+	}
+	return false
 }
 
 // sbPatch is a function-relative static-symbol relocation: the disp32 field
@@ -86,6 +101,10 @@ func assemble(t *ast.Text, link *linkInfo) ([]byte, []sbPatch, map[string]int, [
 	// outgrows the short form.
 	long := make([]bool, len(t.Body))
 	sizes := make([]int, len(t.Body))
+	numTargets := make([]int, len(t.Body))
+	for i := range numTargets {
+		numTargets[i] = -1
+	}
 	offsets := map[string]int{}
 	pcs := make([]int, len(t.Body))
 	var guardJBlong, guardJBElong, moreJMPlong bool
@@ -94,23 +113,106 @@ func assemble(t *ast.Text, link *linkInfo) ([]byte, []sbPatch, map[string]int, [
 	for {
 		guard := fi.guardLen(guardJBlong, guardJBElong)
 		pos := guard + len(fi.prologue)
+		for i := range numTargets {
+			numTargets[i] = -1
+		}
+		idxAtPc := map[int]int{}
 		for i, stmt := range t.Body {
 			switch s := stmt.(type) {
 			case *ast.Label:
 				offsets[s.Name.Text] = pos
 			case *ast.Instr:
+				if strings.ToUpper(s.Mnemonic.Text) == "PCALIGN" {
+					// The alignment pseudo-statement: its size is the
+					// padding to the next boundary at this very position,
+					// filled with NOPs at emission.
+					pad, err := pcAlignPad(pcAlignValue(s), pos)
+					if err != nil {
+						return nil, nil, nil, nil, nil, nil, fmt.Errorf("PCALIGN: %w", err)
+					}
+					sizes[i] = pad
+					pcs[i] = pos
+					pos += pad
+					continue
+				}
 				sz, err := instrSize(s, fi, long[i], link)
 				if err != nil {
 					return nil, nil, nil, nil, nil, nil, fmt.Errorf("%s: %w", s.Mnemonic.Text, err)
 				}
 				sizes[i] = sz
 				pcs[i] = pos
+				idxAtPc[pos] = i
 				pos += sz
 			}
 		}
 		bodyLen := pos - (guard + len(fi.prologue))
 		// Expand any short jump whose displacement no longer fits rel8.
 		changed := false
+		// Numeric ±N(PC) jumps resolve against this iteration's layout; the
+		// emission pass reads the same table after the loop converges.  A
+		// target that is itself an unconditional local JMP is chased to the
+		// ultimate target: the toolchain's brloop pass collapses branch-to-
+		// branch chains before it encodes, so matching its bytes requires
+		// the same redirection.
+		for i := range numTargets {
+			numTargets[i] = -1
+		}
+		for i, stmt := range t.Body {
+			s, ok := stmt.(*ast.Instr)
+			if !ok {
+				continue
+			}
+			if len(s.Operands) == 1 {
+				if n, isNum := pcJumpOffset(s.Operands[0]); isNum {
+					if target, okT := pcJumpTarget(t, i, n, pcs); okT {
+						numTargets[i] = target
+					}
+				}
+			}
+		}
+		for i := range numTargets {
+			if numTargets[i] < 0 {
+				continue
+			}
+			tgt := numTargets[i]
+			for hop := 0; hop < len(t.Body); hop++ {
+				idx, ok := idxAtPc[tgt]
+				if !ok {
+					break
+				}
+				in, ok := t.Body[idx].(*ast.Instr)
+				if !ok || strings.ToUpper(in.Mnemonic.Text) != "JMP" || len(in.Operands) != 1 {
+					break
+				}
+				if name, isLabel := labelName(in.Operands[0]); isLabel {
+					tgt = offsets[resolve(name)]
+					continue
+				}
+				if n, isNum := pcJumpOffset(in.Operands[0]); isNum {
+					next, okT := pcJumpTarget(t, idx, n, pcs)
+					if !okT {
+						break
+					}
+					tgt = next
+					continue
+				}
+				break // JMP through a register or memory: the chain ends
+			}
+			numTargets[i] = tgt
+		}
+		for i, stmt := range t.Body {
+			s, ok := stmt.(*ast.Instr)
+			if !ok {
+				continue
+			}
+			if numTargets[i] >= 0 && !long[i] {
+				rel := int64(numTargets[i] - (pcs[i] + jumpSize(strings.ToUpper(s.Mnemonic.Text), false)))
+				if !fits8(rel) {
+					long[i] = true
+					changed = true
+				}
+			}
+		}
 		for i, stmt := range t.Body {
 			s, ok := stmt.(*ast.Instr)
 			if !ok {
@@ -232,7 +334,7 @@ func assemble(t *ast.Text, link *linkInfo) ([]byte, []sbPatch, map[string]int, [
 				spadjStep{pos + epi, 0},
 			)
 		}
-		code, ps, pool, err := encodeInstr(s, pos, offsets, fi, long[i], resolve, link)
+		code, ps, pool, err := encodeInstr(s, pos, offsets, fi, long[i], resolve, link, numTargets[i])
 		if err != nil {
 			return nil, nil, nil, nil, nil, nil, fmt.Errorf("%s: %w", s.Mnemonic.Text, err)
 		}
@@ -428,6 +530,99 @@ func computeFrame(t *ast.Text) frameInfo {
 	return fi
 }
 
+// pcJumpOffset recognises the numeric relative jump operand ±N(PC) and
+// returns N: the toolchain counts instructions, not bytes, so +2(PC) targets
+// the second instruction boundary after the branch.
+func pcJumpOffset(op *ast.Operand) (int, bool) {
+	if op.Kind != ast.OpAddr || op.Addr.Base != "PC" {
+		return 0, false
+	}
+	return int(op.Addr.Offset), true
+}
+
+// pcJumpTarget resolves a numeric jump at statement index j: N counts the
+// instruction statements after the jump itself (N = 0 is the jump's own
+// address, the classic park loop), and the target is the start of the Nth
+// one.  A negative N counts the same way backwards, before the jump: the
+// exit loops write JMP -3(PC) to land three instructions earlier.  Labels
+// count not, in either direction.  It reports false when the count runs
+// past the end of the function, or before its first instruction.
+func pcJumpTarget(t *ast.Text, j, n int, pcs []int) (int, bool) {
+	if n == 0 {
+		return pcs[j], true
+	}
+	if n < 0 {
+		seen := 0
+		for k := j - 1; k >= 0; k-- {
+			if _, ok := t.Body[k].(*ast.Instr); !ok {
+				continue
+			}
+			seen--
+			if seen == n {
+				return pcs[k], true
+			}
+		}
+		return 0, false
+	}
+	seen := 0
+	for k := j + 1; k < len(t.Body); k++ {
+		if _, ok := t.Body[k].(*ast.Instr); !ok {
+			continue
+		}
+		seen++
+		if seen == n {
+			return pcs[k], true
+		}
+	}
+	return 0, false
+}
+
+// x86 NOP encodings, single-instruction no-ops of lengths 1 to 9 (the
+// toolchain's asm6.go nop table); longer padding repeats the largest that
+// fits, greedy from the end.
+var x86Nops = [][]byte{
+	{0x90},
+	{0x66, 0x90},
+	{0x0F, 0x1F, 0x00},
+	{0x0F, 0x1F, 0x40, 0x00},
+	{0x0F, 0x1F, 0x44, 0x00, 0x00},
+	{0x66, 0x0F, 0x1F, 0x44, 0x00, 0x00},
+	{0x0F, 0x1F, 0x80, 0x00, 0x00, 0x00, 0x00},
+	{0x0F, 0x1F, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00},
+	{0x66, 0x0F, 0x1F, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00},
+}
+
+// fillNOPs fills p with the greedy largest single-instruction NOPs, exactly
+// the toolchain's fillnop.
+func fillNOPs(p []byte) {
+	for len(p) > 0 {
+		m := min(len(p), len(x86Nops))
+		copy(p[:m], x86Nops[m-1])
+		p = p[m:]
+	}
+}
+
+// pcAlignPad computes the padding PCALIGN $align inserts at pos: the
+// alignment must be a power of two in [8, 2048] and the padding runs to the
+// next boundary (zero when the position is already aligned).
+func pcAlignPad(align, pos int) (int, error) {
+	if align <= 0 || align&(align-1) != 0 || align < 8 || align > 2048 {
+		return 0, fmt.Errorf("alignment value of an instruction must be a power of two and in the range [8, 2048], got %d", align)
+	}
+	if lob := pos & (align - 1); lob != 0 {
+		return align - lob, nil
+	}
+	return 0, nil
+}
+
+// pcAlignValue reads a PCALIGN statement's alignment operand.
+func pcAlignValue(s *ast.Instr) int {
+	if len(s.Operands) == 1 && s.Operands[0].Kind == ast.OpImmediate && s.Operands[0].Imm.HasVal {
+		return int(s.Operands[0].Imm.Val)
+	}
+	return 0 // rejected by pcAlignPad's range check
+}
+
 // hasCall reports whether the function body contains a CALL instruction.
 func hasCall(t *ast.Text) bool {
 	for _, stmt := range t.Body {
@@ -602,7 +797,7 @@ func instrSize(s *ast.Instr, fi frameInfo, long bool, link *linkInfo) (int, erro
 		}
 		return jumpSize(mnem, long), nil
 	}
-	code, _, _, err := encodeInstr(s, 0, nil, fi, false, nil, link)
+	code, _, _, err := encodeInstr(s, 0, nil, fi, false, nil, link, -1)
 	if err != nil {
 		return 0, err
 	}
@@ -613,17 +808,43 @@ func isJumpMnemonic(mnem string) bool {
 	if mnem == "JMP" || mnem == "CALL" {
 		return true
 	}
+	if isLoopMnemonic(mnem) {
+		return true
+	}
 	_, ok := condCode(mnem)
 	return ok
 }
 
+// isLoopMnemonic reports the LOOP family, rel8 alone (E0-E2).
+func isLoopMnemonic(mnem string) bool {
+	switch mnem {
+	case "LOOP", "LOOPE", "LOOPNE":
+		return true
+	}
+	return false
+}
+
+// loopOpcode maps the LOOP family to its E0-E2 opcode.
+func loopOpcode(mnem string) byte {
+	switch mnem {
+	case "LOOPE":
+		return 0xE1
+	case "LOOPNE":
+		return 0xE0
+	}
+	return 0xE2
+}
+
 // jumpSize returns the length of a jump instruction in the requested form:
 // short (rel8) where available, otherwise the rel32 form.  CALL is always
-// rel32.
+// rel32; the LOOP family is rel8 alone.
 func jumpSize(mnem string, long bool) int {
 	if mnem == "CALL" {
 		return 5 // opcode + rel32
 	}
+	if isLoopMnemonic(mnem) {
+		return 2 // opcode + rel8, the only form
+	}
 	if !long {
 		return 2 // opcode + rel8
 	}
@@ -637,9 +858,21 @@ func jumpSize(mnem string, long bool) int {
 // (relative to pc, the instruction's own offset).  A RET in a frame-pointer
 // function is prefixed with the epilogue.  resolve, when non-nil, redirects a
 // jump label through the jump-to-jump chain before the offset lookup.
-func encodeInstr(s *ast.Instr, pc int, offsets map[string]int, fi frameInfo, long bool, resolve func(string) string, link *linkInfo) ([]byte, []sbPatch, []floatPoolEntry, error) {
+func encodeInstr(s *ast.Instr, pc int, offsets map[string]int, fi frameInfo, long bool, resolve func(string) string, link *linkInfo, numTarget int) ([]byte, []sbPatch, []floatPoolEntry, error) {
 	mnem := strings.ToUpper(s.Mnemonic.Text)
 
+	if mnem == "PCALIGN" {
+		// The layout pass already accounted the padding; emit the same
+		// amount of NOP bytes for the statement's own position.
+		pad, err := pcAlignPad(pcAlignValue(s), pc)
+		if err != nil {
+			return nil, nil, nil, err
+		}
+		out := make([]byte, pad)
+		fillNOPs(out)
+		return out, nil, nil, nil
+	}
+
 	var prefix []byte
 	if mnem == "RET" && fi.useFP {
 		prefix = fi.epilogue
@@ -677,7 +910,7 @@ func encodeInstr(s *ast.Instr, pc int, offsets map[string]int, fi frameInfo, lon
 			}
 			return append(prefix, code...), nil, nil, nil
 		}
-		code, err = encodeJump(s, mnem, pc+len(prefix), offsets, long, resolve)
+		code, err = encodeJump(s, mnem, pc+len(prefix), offsets, long, resolve, numTarget)
 	} else {
 		code, ps, pool, err = encodeNormal(s, fi, link)
 	}
@@ -703,6 +936,51 @@ func encodeNormal(s *ast.Instr, fi frameInfo, link *linkInfo) ([]byte, []sbPatch
 		}
 		return code, nil, nil, nil
 	}
+	// MOVQ $sym±off(SB), r64: the toolchain assembles a symbol immediate as
+	// LEAQ disp32(RIP), r64 with an R_PCREL relocation at the disp32 field,
+	// never as a 64-bit absolute immediate (verified against go tool asm).
+	// MOVD is the MOVQ alias; the narrower widths reject the form outright.
+	if (mnemUpper == "MOVQ" || mnemUpper == "MOVD") && len(s.Operands) == 2 &&
+		s.Operands[0].Kind == ast.OpImmediate && s.Operands[0].Imm.Sym != nil &&
+		s.Operands[0].Imm.Sym.Pseudo == "SB" {
+		mem := &ast.Operand{Kind: ast.OpAddr, Addr: ast.Address{Sym: s.Operands[0].Imm.Sym}}
+		src, err := operandFromAST(mnemUpper, mem, 8, fi, link)
+		if err != nil {
+			return nil, nil, nil, err
+		}
+		dst, err := operandFromAST(mnemUpper, s.Operands[1], 8, fi, link)
+		if err != nil {
+			return nil, nil, nil, err
+		}
+		e := &enc{}
+		if err := e.encodeLea([]Operand{src, dst}, 8); err != nil {
+			return nil, nil, nil, err
+		}
+		ps := make([]sbPatch, len(e.patches))
+		for i, p := range e.patches {
+			ps[i] = sbPatch{off: p.off, name: p.name, addend: p.addend}
+		}
+		return e.out, ps, nil, nil
+	}
+	// MOVQ/MOVL TLS, r: the bare TLS load.  The toolchain's progedit nops
+	// it out on the one-instruction TLS systems (linux and freebsd, not
+	// shared) and encodes the segment-prefixed load elsewhere; get_tls(r),
+	// the macro GOROOT's go_tls.h defines, expands to exactly this
+	// statement, and the toolchain's pairing pass removes it whenever the
+	// following instruction's (TLS*1) index folds.
+	if (mnemUpper == "MOVQ" || mnemUpper == "MOVL") && len(s.Operands) == 2 && isBareTLS(s.Operands[0]) {
+		return encodeTLSBaseLoad(s, fi, link)
+	}
+	// The old paired-register shift spelling, SHLL CX, R11:AX (a colon
+	// between the two registers), is the toolchain's SHLD family: SHLDL CL,
+	// AX, R11 with the count register first, the paired source in the reg
+	// field and the pair's head in r/m.
+	if code, ps, err := encodeColonShift(s, mnemUpper, fi, link); code != nil || err != nil {
+		if err != nil {
+			return nil, nil, nil, err
+		}
+		return code, ps, nil, nil
+	}
 	_, size := splitSize(mnemUpper)
 	if size == 0 {
 		size = 8
@@ -722,10 +1000,67 @@ func encodeNormal(s *ast.Instr, fi frameInfo, link *linkInfo) ([]byte, []sbPatch
 	ps := make([]sbPatch, len(e.patches))
 	for i, p := range e.patches {
 		ps[i] = sbPatch{off: p.off, name: p.name, addend: p.addend}
+		if p.tls {
+			ps[i].kind = RelTLSLE
+		}
 	}
 	return e.out, ps, e.floatPoolList(), nil
 }
 
+// isBareTLS reports whether the operand is the bare TLS pseudo-register
+// load source, the expansion of go_tls.h's get_tls(r) macro.
+func isBareTLS(op *ast.Operand) bool {
+	return op.Kind == ast.OpAddr && op.Addr.Sym != nil &&
+		op.Addr.Sym.Pseudo == "" && op.Addr.Sym.Name == "TLS" &&
+		op.Addr.Base == "" && op.Addr.Index == ""
+}
+
+// encodeTLSBaseLoad assembles MOVQ/MOVL TLS, r.  On the one-instruction TLS
+// systems (linux and freebsd outside -shared, obj6.go's CanUse1InsnTLS) the
+// statement nops out: the following (TLS*1) access folds to a direct
+// segment-absolute load.  The two-instruction systems keep the segment load,
+// nine bytes with the R_TLSLE patch site at the disp32.
+func encodeTLSBaseLoad(s *ast.Instr, fi frameInfo, link *linkInfo) ([]byte, []sbPatch, []floatPoolEntry, error) {
+	_, size := splitSize(strings.ToUpper(s.Mnemonic.Text))
+	if size == 0 {
+		size = 8
+	}
+	dst, err := operandFromAST("MOVQ", s.Operands[1], 8, fi, link)
+	if err != nil {
+		return nil, nil, nil, err
+	}
+	reg, ok := dst.(Reg)
+	if !ok || reg.isVec() {
+		return nil, nil, nil, fmt.Errorf("TLS: destination must be a general register")
+	}
+	if link == nil || link.tlsOneInsn() {
+		return nil, nil, nil, nil // noped out
+	}
+	seg := byte(0x64) // FS
+	if link.goos == "windows" {
+		seg = 0x65 // GS
+	}
+	e := &enc{}
+	i := &instr{
+		prefix: seg,
+		rexW:   size == 8,
+		rexR:   reg.idx >= 8,
+		opcode: []byte{0x8B},
+		modrm:  0x04 | (reg.idx&7)<<3,
+		sib:    0x25,
+		disp:   le32(0),
+		tls:    true,
+	}
+	if err := e.emit(i); err != nil {
+		return nil, nil, nil, err
+	}
+	ps := make([]sbPatch, len(e.patches))
+	for i, p := range e.patches {
+		ps[i] = sbPatch{off: p.off, name: p.name, addend: p.addend, kind: RelTLSLE}
+	}
+	return e.out, ps, nil, nil
+}
+
 // encodeBookkeeping accepts-and-ignores FUNCDATA and PCDATA at the statement
 // level, before operand conversion: the toolchain's shapes are FUNCDATA
 // $n, sym(SB) and PCDATA $n, $m, and neither contributes a byte to the
@@ -752,23 +1087,91 @@ func encodeBookkeeping(upper string, s *ast.Instr) ([]byte, error) {
 	return nil, nil
 }
 
+// encodeColonShift encodes the paired-register shift spellings, SHLx CX,
+// dst:src: the toolchain reads them as the SHLD family (double-precision
+// shift by CL), reg = the paired source, r/m = the pair's head.  The second
+// operand's raw text carries the colon; ok reports the spelling was found.
+func encodeColonShift(s *ast.Instr, mnemUpper string, fi frameInfo, link *linkInfo) ([]byte, []sbPatch, error) {
+	base, _ := strings.CutPrefix(mnemUpper, "SHL")
+	if base == mnemUpper || len(s.Operands) != 2 {
+		return nil, nil, nil
+	}
+	_, size := splitSize(mnemUpper)
+	raw := strings.ReplaceAll(s.Operands[1].Raw, " ", "")
+	head, tail, ok := strings.Cut(raw, ":")
+	if !ok || head == "" || tail == "" {
+		return nil, nil, nil
+	}
+	headReg, ok1 := ParseReg(head)
+	srcReg, ok2 := ParseReg(tail)
+	if !ok1 || !ok2 {
+		return nil, nil, fmt.Errorf("%s: invalid paired register %q", mnemUpper, s.Operands[1].Raw)
+	}
+	cnt, err := operandFromAST(mnemUpper, s.Operands[0], size, fi, link)
+	if err != nil {
+		return nil, nil, err
+	}
+	cntReg, ok := cnt.(Reg)
+	if !ok || cntReg.idx != 1 {
+		return nil, nil, fmt.Errorf("%s: the paired-register form counts in CL", mnemUpper)
+	}
+	// SHLD r/m, reg, CL: 0F A5 (REX.W for the 64-bit width).
+	e := &enc{}
+	i := &instr{rexW: size == 8, opcode: []byte{0x0F, 0xA5}, modrm: -1, sib: -1}
+	if err := setRM(i, srcReg, headReg, size); err != nil {
+		return nil, nil, err
+	}
+	if err := e.emit(i); err != nil {
+		return nil, nil, err
+	}
+	ps := make([]sbPatch, len(e.patches))
+	for j, p := range e.patches {
+		ps[j] = sbPatch{off: p.off, name: p.name, addend: p.addend}
+	}
+	return e.out, ps, nil
+}
+
+// trailingIndexGroup recovers a trailing "(index*scale)" or "(index)" group
+// from an operand's raw text: the symbol-pseudo parse returns before the
+// index group, so foo(SP)(AX*1) keeps its index only in the spelling.
+func trailingIndexGroup(raw string) (string, int, bool) {
+	compact := strings.ReplaceAll(raw, " ", "")
+	if !strings.HasSuffix(compact, ")") {
+		return "", 0, false
+	}
+	open := strings.LastIndex(compact, "(")
+	if open < 2 || !strings.Contains(compact[:open], ")") {
+		return "", 0, false // one group alone: no trailing index
+	}
+	name, scale, _, ok := cutParenGroup(compact[open:])
+	return name, scale, ok
+}
+
 // encodeJump encodes a JMP/CALL/Jcc with a relative offset resolved from the
-// target label, in the short (rel8) or long (rel32) form.
-func encodeJump(s *ast.Instr, mnem string, pc int, offsets map[string]int, long bool, resolve func(string) string) ([]byte, error) {
+// target label or from a numeric ±N(PC) instruction count, in the short
+// (rel8) or long (rel32) form.  numTarget is the resolved byte offset of a
+// numeric operand, negative when the operand is not one.
+func encodeJump(s *ast.Instr, mnem string, pc int, offsets map[string]int, long bool, resolve func(string) string, numTarget int) ([]byte, error) {
 	if len(s.Operands) != 1 {
 		return nil, fmt.Errorf("jump expects 1 operand, got %d", len(s.Operands))
 	}
-	name, ok := labelName(s.Operands[0])
-	if !ok {
+	name, isLabel := labelName(s.Operands[0])
+	if !isLabel && numTarget < 0 {
 		return nil, fmt.Errorf("jump target must be a local label")
 	}
+	var target int
+	if isLabel {
 		if resolve != nil && mnem != "CALL" {
 			name = resolve(name)
 		}
-	target, ok := offsets[name]
+		t, ok := offsets[name]
 		if !ok {
 			return nil, fmt.Errorf("undefined label %q", name)
 		}
+		target = t
+	} else {
+		target = numTarget
+	}
 	rel := int64(target - (pc + jumpSize(mnem, long)))
 
 	if !long {
@@ -778,9 +1181,15 @@ func encodeJump(s *ast.Instr, mnem string, pc int, offsets map[string]int, long
 		if mnem == "JMP" {
 			return []byte{0xEB, byte(int8(rel))}, nil
 		}
+		if isLoopMnemonic(mnem) {
+			return []byte{loopOpcode(mnem), byte(int8(rel))}, nil
+		}
 		cc, _ := condCode(mnem)
 		return []byte{0x70 + byte(cc), byte(int8(rel))}, nil
 	}
+	if isLoopMnemonic(mnem) {
+		return nil, fmt.Errorf("%s has no long form", mnem)
+	}
 	switch mnem {
 	case "JMP":
 		return append([]byte{0xE9}, le32(rel)...), nil
@@ -832,6 +1241,72 @@ func labelName(op *ast.Operand) (string, bool) {
 	return "", false
 }
 
+// jumpOperand returns the branch-target operand of a JMP/CALL, rewriting the
+// `*`-prefixed indirect spellings (JMP *(R12), JMP *4(SP)) into their plain
+// memory form.  The star marks an indirect target and changes no bytes; the
+// address parser leaves the operand's address empty because of the leading
+// star, so the fields are rebuilt from the raw text onto a copy of the
+// operand, never on the shared syntax tree.
+func jumpOperand(s *ast.Instr) *ast.Operand {
+	if len(s.Operands) != 1 {
+		return nil
+	}
+	op := s.Operands[0]
+	compact := strings.ReplaceAll(op.Raw, " ", "")
+	inner, ok := strings.CutPrefix(compact, "*")
+	if !ok {
+		return op
+	}
+	var addr ast.Address
+	if i := strings.IndexByte(inner, '('); i > 0 {
+		v, err := strconv.ParseInt(inner[:i], 0, 64)
+		if err != nil {
+			return op
+		}
+		addr.Offset, addr.HasOff = v, true
+		inner = inner[i:]
+	}
+	base, _, rest, ok := cutParenGroup(inner)
+	if !ok {
+		return op
+	}
+	if base != "" {
+		addr.Base = base
+	}
+	if rest != "" {
+		idx, scale, _, ok := cutParenGroup(rest)
+		if ok && idx != "" {
+			addr.Index = idx
+			addr.Scale = scale
+		}
+	}
+	c := *op
+	c.Addr = addr
+	return &c
+}
+
+// cutParenGroup splits a leading "(name)" or "(name*n)" off s, returning the
+// inner text, the scale it names (1 when the group spells no multiplier) and
+// the remainder.
+func cutParenGroup(s string) (name string, scale int, rest string, ok bool) {
+	if !strings.HasPrefix(s, "(") {
+		return "", 0, "", false
+	}
+	i := strings.IndexByte(s, ')')
+	if i < 0 {
+		return "", 0, "", false
+	}
+	inner, rest := s[1:i], s[i+1:]
+	if before, after, ok := strings.Cut(inner, "*"); ok {
+		n, err := strconv.Atoi(after)
+		if err != nil {
+			return "", 0, "", false
+		}
+		return before, n, rest, true
+	}
+	return inner, 1, rest, true
+}
+
 // 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
@@ -840,7 +1315,19 @@ func indirectJumpTarget(s *ast.Instr) bool {
 	if len(s.Operands) != 1 || s.Operands[0].Kind != ast.OpAddr {
 		return false
 	}
-	a := s.Operands[0].Addr
+	op := jumpOperand(s)
+	if op == nil {
+		return false
+	}
+	if op != s.Operands[0] {
+		return true // the star marker spells an indirect target
+	}
+	a := op.Addr
+	// ±N(PC) is the numeric relative form, the PC counts instructions from
+	// the branch: relative, not indirect.
+	if a.Base == "PC" || a.Index == "PC" {
+		return false
+	}
 	if a.Base != "" || a.Index != "" {
 		return true
 	}
@@ -853,18 +1340,20 @@ func indirectJumpTarget(s *ast.Instr) bool {
 }
 
 // encodeIndirectJump assembles a JMP/CALL through a register or memory
-// operand, which carries no relocation and no label to resolve.
+// operand, which carries no relocation and no label to resolve.  The
+// `*`-prefixed spellings go through jumpOperand first, their star rebuilt
+// into a plain memory operand.
 func encodeIndirectJump(s *ast.Instr, mnem string) ([]byte, error) {
-	ops := make([]Operand, len(s.Operands))
-	for i, op := range s.Operands {
+	op := s.Operands[0]
+	if cleaned := jumpOperand(s); cleaned != nil {
+		op = cleaned
+	}
 	o, err := operandFromAST(mnem, op, 8, frameInfo{}, nil)
 	if err != nil {
 		return nil, err
 	}
-		ops[i] = o
-	}
 	e := &enc{}
-	if err := e.encodeIndirectBranch(mnem, ops); err != nil {
+	if err := e.encodeIndirectBranch(mnem, []Operand{o}); err != nil {
 		return nil, err
 	}
 	return e.out, nil
@@ -933,37 +1422,89 @@ func operandFromAST(mnemUpper string, op *ast.Operand, size int, fi frameInfo, l
 			off := a.Sym.Offset + fi.fpAdjust
 			return Mem{Base: spReg, Disp: off, HasBase: true, Size: size}, nil
 		}
-		// SP-relative local: x-N(SP) → (spAdjust + offset)(SP).
+		// SP-relative local: x-N(SP) → (spAdjust + offset)(SP), keeping a scaled
+		// index beside the virtual stack pointer (foo(SP)(AX*1)).  The
+		// symbol-pseudo parse returns before the index group, so the index
+		// is recovered from the raw text when the address lacks it.
 		if a.Sym != nil && a.Sym.Pseudo == "SP" && a.Base == "" {
 			off := fi.spAdjust + a.Sym.Offset
-			return Mem{Base: spReg, Disp: off, HasBase: true, Size: size}, nil
+			m := Mem{Base: spReg, Disp: off, HasBase: true, Size: size}
+			if name, scale, ok := trailingIndexGroup(op.Raw); ok {
+				idx, ok := ParseReg(name)
+				if !ok {
+					return nil, fmt.Errorf("unknown index register %q", name)
+				}
+				m.Index = idx
+				m.Scale = scale
+				m.HasIndex = true
+			}
+			return m, nil
 		}
 		// SB (global symbol): a symbol defined in the same file (GLOBL) is
 		// encoded RIP-relative and resolved by the file-level layout;
-		// anything not defined here needs object-file emission.
+		// anything not defined here needs object-file emission.  A static
+		// (file-local) spelling of an undefined symbol defers the same way
+		// the toolchain does: the relocation names it and the linker decides.
 		if a.Sym != nil && a.Sym.Pseudo == "SB" {
 			if link == nil || link.symbols == nil {
 				return nil, fmt.Errorf("symbol %q needs file-level assembly (AssembleFile)", a.Sym.Name)
 			}
-			if !link.symbols[a.Sym.Name] {
-				if a.Sym.Static {
-					return nil, fmt.Errorf("undefined symbol %q", a.Sym.Name)
-				}
-				if !link.allowExternal {
+			if !link.symbols[a.Sym.Name] && !link.allowExternal {
 				return nil, fmt.Errorf("external symbol %q needs object-file emission", a.Sym.Name)
 			}
-			}
 			return sbMem{size: size, name: a.Sym.Name, addend: a.Sym.Offset}, nil
 		}
 
 		// Memory with a real base register: (base), off(base), (base)(index*scale).
 		if a.Base != "" {
+			// The TLS pseudo-base, off(TLS): the segment-prefixed absolute
+			// the thread-local access lowers to, 64 8B 04 25 with its
+			// R_TLS_LE patch site on the disp32.
+			if a.Base == "TLS" {
+				seg := byte(0x64) // FS on linux, freebsd, plan9
+				if link != nil && link.goos == "windows" {
+					seg = 0x65 // GS
+				}
+				return TLSMem{Disp: a.Offset, Size: size, Seg: seg}, nil
+			}
+			// Segment-absolute: 0x30(GS) and 0x28(FS), the windows TLS
+			// spellings.  The segment override prefixes a disp32 absolute
+			// reference with no relocation.
+			if a.Base == "GS" || a.Base == "FS" {
+				seg := byte(0x64)
+				if a.Base == "GS" {
+					seg = 0x65
+				}
+				return SegAbs{Disp: a.Offset, Size: size, Seg: seg}, nil
+			}
 			base, ok := ParseReg(a.Base)
 			if !ok {
 				return nil, fmt.Errorf("unknown base register %q", a.Base)
 			}
 			m := Mem{Base: base, Disp: a.Offset, HasBase: true, Size: size}
 			if a.Index != "" {
+				if a.Index == "TLS" {
+					// off(base)(TLS*1): the thread-local annotation.  The
+					// one-instruction TLS form folds it to off(TLS), the
+					// segment-prefixed absolute whose disp32 carries an
+					// R_TLS_LE patch site; the base register disappears
+					// from the encoding, exactly as the toolchain's
+					// progedit rewrites the address.
+					seg := byte(0x64) // FS on linux, freebsd, plan9
+					if link != nil && link.goos == "windows" {
+						seg = 0x65 // GS
+					}
+					return TLSMem{Disp: a.Offset, Size: size, Seg: seg}, nil
+				}
+				if a.Index == "GS" || a.Index == "FS" {
+					// 0(CX)(GS): the segment annotation rides the base
+					// access as the override prefix.
+					m.Seg = 0x64
+					if a.Index == "GS" {
+						m.Seg = 0x65
+					}
+					return m, nil
+				}
 				idx, ok := ParseReg(a.Index)
 				if !ok {
 					return nil, fmt.Errorf("unknown index register %q", a.Index)
@@ -984,6 +1525,11 @@ func operandFromAST(mnemUpper string, op *ast.Operand, size int, fi frameInfo, l
 			}
 			return Mem{Index: idx, Scale: a.Scale, Disp: a.Offset, HasIndex: true, Size: size}, nil
 		}
+		// A bare displacement with no base: the absolute address form,
+		// MOVL $0xf1, 0xf1.  No segment and no relocation.
+		if a.Sym == nil && a.Base == "" && a.Index == "" && a.HasOff {
+			return SegAbs{Disp: a.Offset, Size: size}, nil
+		}
 		// Bare register.
 		if a.Sym != nil && a.Sym.Pseudo == "" && a.Sym.Name != "" {
 			if r, ok := ParseReg(a.Sym.Name); ok {

asm/assemble_test.go

@@ -10,8 +10,8 @@ import (
 
 	"golang.org/x/arch/x86/x86asm"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/ast"
-	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+	"sourcedock.dev/petrbalvin/gasm-sdk/ast"
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
 )
 
 // firstText parses src and returns its first TEXT function.
@@ -370,6 +370,58 @@ end:
 	}
 }
 
+// TestAssembleNumericPCJumps pins the numeric ±N(PC) branch operands: N
+// counts instruction statements, skipping labels, in both directions (the
+// runtime's exit loops write JMP -3(PC)), N = 0 parks on the jump itself.
+func TestAssembleNumericPCJumps(t *testing.T) {
+	fn := firstText(t, `
+#include "textflag.h"
+TEXT ·exit(SB), NOSPLIT, $0
+	MOVB $1, AL
+lab:
+	MOVB $2, AL
+	MOVB $3, AL
+	JMP -3(PC)
+	MOVB $4, AL
+park:
+	JMP 0(PC)
+	MOVB $5, AL
+	JMP 2(PC)
+	MOVB $6, AL
+	RET
+`)
+	code, _, err := Assemble(fn)
+	if err != nil {
+		t.Fatalf("Assemble: %v", err)
+	}
+	// From the Go-assembled function:
+	//   MOVB $1, AL       b001
+	//   MOVB $2, AL       b002
+	//   MOVB $3, AL       b003
+	//   JMP -3(PC)        ebf8      (three instructions back, past lab:)
+	//   MOVB $4, AL       b004
+	//   JMP 0(PC)         ebfe      (the park loop)
+	//   MOVB $5, AL       b005
+	//   JMP 2(PC)         eb02      (over MOVB $6 to the RET)
+	//   MOVB $6, AL       b006
+	//   RET               c3
+	want := []byte{
+		0xb0, 0x01,
+		0xb0, 0x02,
+		0xb0, 0x03,
+		0xeb, 0xf8,
+		0xb0, 0x04,
+		0xeb, 0xfe,
+		0xb0, 0x05,
+		0xeb, 0x02,
+		0xb0, 0x06,
+		0xc3,
+	}
+	if hexBytes(code) != hexBytes(want) {
+		t.Errorf("numeric-PC mismatch:\n got:  %s\n want: %s", hexBytes(code), hexBytes(want))
+	}
+}
+
 func TestAssemblePrefetch(t *testing.T) {
 	fn := firstText(t, `
 #include "textflag.h"

asm/elf_dwarf_test.go

@@ -8,7 +8,7 @@ import (
 	"encoding/binary"
 	"testing"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
 )
 
 // ulebIter reads ULEB128 values, the .debug_abbrev and line-header

asm/elf_test.go

@@ -12,8 +12,8 @@ import (
 	"path/filepath"
 	"testing"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/ast"
-	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+	"sourcedock.dev/petrbalvin/gasm-sdk/ast"
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
 )
 
 // The object-file tests share one source: two exported functions, one

asm/elfarm64_test.go

@@ -9,7 +9,7 @@ import (
 	"encoding/binary"
 	"testing"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
 )
 
 // TestELFAARCH64Object checks the structure of the emitted AArch64 ELF64

asm/elfloong64_test.go

@@ -9,7 +9,7 @@ import (
 	"encoding/binary"
 	"testing"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
 )
 
 // TestELFLOONG64Object checks the structure of the emitted LoongArch ELF64

asm/elfriscv_test.go

@@ -9,7 +9,7 @@ import (
 	"encoding/binary"
 	"testing"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
 )
 
 // TestELFRISCVObjectDataRelocation checks that a symbol-valued DATA field

asm/encodable.go

@@ -20,11 +20,24 @@ func Encodable(mnemonic string) bool {
 	switch upper {
 	case "RET", "NOP", "CALL", "JMP",
 		"POPFQ", "PUSHFQ", "INT", "LDMXCSR", "STMXCSR", "CMPSD", "SHA256RNDS2",
+		// The SSE compare family sharing CMPSD's predicate-last shape, the
+		// far return with its stack pop, the loop family, the bank-crossing
+		// MMX moves and the one-operand system controls.
+		"CMPSS", "CMPPS", "CMPPD", "RETFL",
+		"LOOP", "LOOPE", "LOOPNE",
+		"MOVDQ2Q", "MOVQ2DQ",
+		"ENDBR64", "CLWB", "TPAUSE", "UMONITOR", "UMWAIT", "RDPID", "CLDEMOTE",
 		// The literal-data pseudo-ops, the accepted-and-ignored END and
 		// bookkeeping statements, and the SP adjust.
 		"BYTE", "WORD", "LONG", "QUAD", "END", "ADJSP", "FUNCDATA", "PCDATA":
 		return true
 	}
+	if _, ok := sysUnaryTable[upper]; ok {
+		return true
+	}
+	if _, ok := sseStoreOnly[upper]; ok {
+		return true
+	}
 	if _, ok := noOperandTable[upper]; ok {
 		return true
 	}

asm/encode.go

@@ -64,6 +64,7 @@ type encPatch struct {
 	off    int
 	name   string
 	addend int64
+	tls    bool // a TLS slot offset: the patch is R_TLSLE with no symbol
 }
 
 func (e *enc) encode(mnem string, ops []Operand) error {
@@ -72,9 +73,23 @@ func (e *enc) encode(mnem string, ops []Operand) error {
 	// Fixed-name instructions (no size suffix).
 	switch {
 	case upper == "RET":
+		// RET sym(SB), the absolute return: the toolchain encodes it as a
+		// tail jump, E9 rel32 with a call relocation against the symbol.
+		if len(ops) == 1 {
+			if m, ok := ops[0].(sbMem); ok {
+				return e.emit(&instr{opcode: []byte{0xE9}, modrm: -1, sib: -1, disp: le32(0), sb: &sbRef{name: m.name, addend: m.addend}})
+			}
+			return fmt.Errorf("RET: unsupported operand")
+		}
+		if len(ops) != 0 {
+			return fmt.Errorf("RET expects no operands, got %d", len(ops))
+		}
 		return e.encodeRet()
 	case upper == "NOP":
-		return e.emit(&instr{opcode: []byte{0x90}, modrm: -1, sib: -1})
+		// The toolchain consumes every NOP statement as a pseudo and emits
+		// nothing for it, operands included (a bare NOP, NOP AX and
+		// NOP sym(SB) all vanish from the object).
+		return nil
 	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.
@@ -101,6 +116,15 @@ func (e *enc) encode(mnem string, ops []Operand) error {
 		}
 		return e.emit(&instr{opcode: op, modrm: -1, sib: -1})
 	}
+	// One-operand system instructions whose reg field is a fixed digit:
+	// the cache and wait controls under 0F AE/0F 1C and the RDPID read.
+	if m, ok := sysUnaryTable[upper]; ok {
+		return e.encodeSysUnary(upper, m, ops)
+	}
+	// The store-only SSE moves (the non-temporal store).
+	if m, ok := sseStoreOnly[upper]; ok {
+		return e.encodeSSEStoreOnly(upper, m, ops)
+	}
 	// 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 {
@@ -116,14 +140,63 @@ func (e *enc) encode(mnem string, ops []Operand) error {
 		return e.emit(&instr{opcode: []byte{0x9C}, modrm: -1, sib: -1})
 	case "INT":
 		return e.encodeInt(ops)
+	// The LOOP family outside the assembler's label settlement: the operand
+	// is the already-computed rel8 (E0-E2).
+	case "LOOP", "LOOPE", "LOOPNE":
+		if len(ops) != 1 {
+			return fmt.Errorf("%s expects 1 operand, got %d", upper, len(ops))
+		}
+		imm, ok := ops[0].(Imm)
+		if !ok || !fits8(int64(imm)) {
+			return fmt.Errorf("%s: relative offset must be a signed byte", upper)
+		}
+		return e.emit(&instr{opcode: []byte{loopOpcode(upper)}, modrm: -1, sib: -1, imm: []byte{byte(int8(imm))}})
 	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).
+	// LAST in Plan 9 order (src, dst, $imm); the family shares the shape.
 	case "CMPSD":
-		return e.encodeCmpsd(ops)
+		return e.encodeSSECmp("CMPSD", 0xF2, ops)
+	case "CMPSS":
+		return e.encodeSSECmp("CMPSS", 0xF3, ops)
+	case "CMPPS":
+		return e.encodeSSECmp("CMPPS", 0x00, ops)
+	case "CMPPD":
+		return e.encodeSSECmp("CMPPD", 0x66, ops)
+	// RETFL pops the immediate's worth of bytes after the far return
+	// (LRET iw: CA imm16), the toolchain's RETF spelling with a stack
+	// adjustment.
+	case "RETFL":
+		if len(ops) != 1 {
+			return fmt.Errorf("RETFL expects 1 operand, got %d", len(ops))
+		}
+		imm, ok := ops[0].(Imm)
+		if !ok {
+			return fmt.Errorf("RETFL expects an immediate")
+		}
+		return e.emit(&instr{opcode: []byte{0xCA}, modrm: -1, sib: -1, imm: le16(int64(imm))})
+	// MOVDQ2Q/MOVQ2DQ cross the MMX and XMM banks (F2 0F D6), the register
+	// in the reg field, the other bank's in r/m.
+	case "MOVDQ2Q", "MOVQ2DQ":
+		if len(ops) != 2 {
+			return fmt.Errorf("%s expects 2 operands, got %d", upper, len(ops))
+		}
+		srcReg, ok1 := ops[0].(Reg)
+		dstReg, ok2 := ops[1].(Reg)
+		if !ok1 || !ok2 {
+			return fmt.Errorf("%s takes register operands alone", upper)
+		}
+		if upper == "MOVDQ2Q" && (!srcReg.isVec() || !dstReg.mmx) ||
+			upper == "MOVQ2DQ" && (!srcReg.mmx || !dstReg.isVec()) {
+			return fmt.Errorf("%s crosses the XMM and MMX banks in that order", upper)
+		}
+		i := &instr{prefix: 0xF2, opcode: []byte{0x0F, 0xD6}, modrm: -1, sib: -1}
+		if err := setRM(i, dstReg, srcReg, 8); err != nil {
+			return err
+		}
+		return e.emit(i)
 	// SHA256RNDS2 carries the round constant in a literal X0 first operand.
 	case "SHA256RNDS2":
 		return e.encodeSha256rnds2(ops)
@@ -578,6 +651,7 @@ type instr struct {
 	disp      []byte
 	imm       []byte
 	sb        *sbRef // static-symbol displacement in disp, awaiting resolution
+	tls       bool   // the displacement is a TLS slot offset, patched R_TLSLE
 }
 
 // sbRef records that an instruction's displacement refers to a static symbol
@@ -620,6 +694,9 @@ func (e *enc) emit(i *instr) error {
 	if i.sb != nil {
 		e.patches = append(e.patches, encPatch{off: len(e.out), name: i.sb.name, addend: i.sb.addend})
 	}
+	if i.tls {
+		e.patches = append(e.patches, encPatch{off: len(e.out), tls: true})
+	}
 	e.out = append(e.out, i.disp...)
 	e.out = append(e.out, i.imm...)
 	return nil
@@ -674,12 +751,30 @@ func setRMReg(i *instr, regField int, rexR, regForced bool, rm Operand, opSize i
 		i.disp = le32(0)
 		i.sb = &sbRef{name: r.name, addend: r.addend}
 		return nil
+	case TLSMem:
+		// off(TLS): the segment-prefixed absolute access, mod=00 with the
+		// SIB escape's disp32 absolute form.  The displacement is the TLS
+		// slot offset, patched by the linker's TLS relocation.
+		i.prefix = r.Seg
+		i.modrm = 0x04 | regField<<3
+		i.sib = 0x25
+		i.disp = le32(r.Disp)
+		i.tls = true
+		return nil
+	case SegAbs:
+		// 0x30(GS): the segment override with the SIB escape's disp32
+		// absolute form, no relocation.
+		setSegAbs(i, regField, r)
+		return nil
 	default:
 		return fmt.Errorf("invalid r/m operand %T", rm)
 	}
 }
 
 func setMem(i *instr, regField int, m Mem) error {
+	if m.Seg != 0 {
+		i.prefix = m.Seg
+	}
 	modrm, sib, disp, xBit, bBit, err := memComponents(regField, m)
 	if err != nil {
 		return err
@@ -692,6 +787,16 @@ func setMem(i *instr, regField int, m Mem) error {
 	return nil
 }
 
+// setSegAbs assembles a segment-absolute operand, 0x30(GS): the segment
+// override with the mod=00 SIB escape's disp32 absolute form and no
+// relocation.
+func setSegAbs(i *instr, regField int, m SegAbs) {
+	i.prefix = m.Seg
+	i.modrm = 0x04 | regField<<3
+	i.sib = 0x25
+	i.disp = le32(m.Disp)
+}
+
 // memComponents computes the ModR/M byte (with the given reg field), the SIB
 // byte (-1 if none), the displacement bytes, and the high index/base bits, for
 // a memory operand.  It is shared by the REX (scalar) and VEX (vector) paths.

asm/encode_test.go

@@ -10,8 +10,8 @@ import (
 
 	"golang.org/x/arch/x86/x86asm"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/ast"
-	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+	"sourcedock.dev/petrbalvin/gasm-sdk/ast"
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
 )
 
 // decode encodes an instruction and decodes it back, returning the decoded
@@ -265,7 +265,11 @@ func TestImul(t *testing.T) {
 
 func TestControl(t *testing.T) {
 	checkSyntax(t, "ret", "RET")
-	checkSyntax(t, "nop", "NOP")
+	// NOP contributes nothing on amd64, consumed whole by the toolchain as
+	// a pseudo; only the bytes pin it (no decodable instruction remains).
+	if code, err := Encode("NOP"); err != nil || len(code) != 0 {
+		t.Errorf("NOP: bytes %x (err %v), want empty", code, err)
+	}
 	checkOp(t, x86asm.JMP, "JMP", Imm(0))
 	checkOp(t, x86asm.CALL, "CALL", Imm(0))
 	checkOp(t, x86asm.JGE, "JGE", Imm(0))
@@ -1193,7 +1197,7 @@ func TestBookkeepingGroundTruth(t *testing.T) {
 	if err != nil {
 		t.Fatalf("assemble: %v", err)
 	}
-	want := "90c3"
+	want := "c3"
 	if got := hexCompact(img.Code); got != want {
 		t.Errorf("body %s, want %s (the bookkeeping lines contribute nothing)", got, want)
 	}
@@ -1221,3 +1225,153 @@ func mustParse(t *testing.T, src string) *ast.File {
 	}
 	return f
 }
+
+// TestCorpusTailSystem pins the system and control forms the toolchain's own
+// amd64 testdata carries, byte for byte: the one-operand IMUL, the compare
+// family, the far return, the loop, the MMX moves, the CR/DR and segment
+// register moves, the TLS pseudo-base and the 0F AE/1C/C7 controls.
+func TestCorpusTailSystem(t *testing.T) {
+	regBPT := Reg{idx: 5, size: 8}
+	X0, X1, X2 := vreg(t, "X0"), vreg(t, "X1"), vreg(t, "X2")
+	Y1, Y2, Y7 := vreg(t, "Y1"), vreg(t, "Y2"), vreg(t, "Y7")
+	X5, X20 := vreg(t, "X5"), vreg(t, "X20")
+	cases := []struct {
+		name string
+		mnem string
+		ops  []Operand
+		want string
+	}{
+		{"IMUL one-op byte", "IMULB", []Operand{DX}, "f6ea"},
+		{"IMUL one-op long", "IMULL", []Operand{AX}, "f7e8"},
+		{"CMPPD", "CMPPD", []Operand{X1, X2, Imm(4)}, "660fc2d104"},
+		{"CMPSS", "CMPSS", []Operand{X1, X2, Imm(4)}, "f30fc2d104"},
+		{"CMPPS", "CMPPS", []Operand{X1, X2, Imm(4)}, "0fc2d104"},
+		{"RETFL", "RETFL", []Operand{Imm(4)}, "ca0400"},
+		{"LOOP", "LOOP", []Operand{Imm(-2)}, "e2fe"},
+		{"LOOPE", "LOOPE", []Operand{Imm(-2)}, "e1fe"},
+		{"LOOPNE", "LOOPNE", []Operand{Imm(-2)}, "e0fe"},
+		{"PADDD MMX", "PADDD", []Operand{Reg{idx: 2, size: 8, mmx: true}, Reg{idx: 1, size: 8, mmx: true}}, "0ffeca"},
+		{"MOVDQ2Q", "MOVDQ2Q", []Operand{X1, Reg{idx: 1, size: 8, mmx: true}}, "f20fd6c9"},
+		{"MOVNTDQ", "MOVNTDQ", []Operand{X1, Ptr(AX, 0, 16)}, "660fe708"},
+		{"MOVQ mmx load", "MOVQ", []Operand{Ptr(AX, 0, 8), Reg{idx: 0, size: 8, mmx: true}}, "0f6f00"},
+		{"MOVQ mmx store", "MOVQ", []Operand{Reg{idx: 0, size: 8, mmx: true}, Ptr(SI, 0, 8)}, "0f7f06"},
+		{"MOVQ CR0 load", "MOVQ", []Operand{Reg{idx: 0, size: 8, ctl: 1}, AX}, "0f20c0"},
+		{"MOVQ CR4 load", "MOVQ", []Operand{Reg{idx: 4, size: 8, ctl: 1}, DI}, "0f20e7"},
+		{"MOVQ CR0 store", "MOVQ", []Operand{AX, Reg{idx: 0, size: 8, ctl: 1}}, "0f22c0"},
+		{"MOVQ DR0 load", "MOVQ", []Operand{Reg{idx: 0, size: 8, ctl: 2}, AX}, "0f21c0"},
+		{"MOVQ DR7 load", "MOVQ", []Operand{Reg{idx: 7, size: 8, ctl: 2}, SI}, "0f21fe"},
+		{"PUSHQ FS", "PUSHQ", []Operand{Reg{idx: 4, size: 2, seg: 5}}, "0fa0"},
+		{"PUSHQ GS", "PUSHQ", []Operand{Reg{idx: 5, size: 2, seg: 6}}, "0fa8"},
+		{"POPQ FS", "POPQ", []Operand{Reg{idx: 4, size: 2, seg: 5}}, "0fa1"},
+		{"POPQ GS", "POPQ", []Operand{Reg{idx: 5, size: 2, seg: 6}}, "0fa9"},
+		{"ENDBR64", "ENDBR64", nil, "f30f1efa"},
+		{"CLWB", "CLWB", []Operand{Ptr(BX, 0, 8)}, "660fae33"},
+		{"CLDEMOTE", "CLDEMOTE", []Operand{Ptr(BX, 0, 8)}, "0f1c03"},
+		{"TPAUSE", "TPAUSE", []Operand{BX}, "660faef3"},
+		{"UMONITOR", "UMONITOR", []Operand{BX}, "f30faef3"},
+		{"UMWAIT", "UMWAIT", []Operand{BX}, "f20faef3"},
+		{"RDPID", "RDPID", []Operand{DX}, "f30fc7fa"},
+		{"RDPID r11", "RDPID", []Operand{Reg{idx: 11, size: 8}}, "f3410fc7fb"},
+		{"LEAL wide disp", "LEAL", []Operand{Idx(regBPT, Reg{idx: 10, size: 8}, 1, 0x8f1bbcdc, 8), regBPT}, "428dac15dcbc1b8f"},
+		{"VPERMPD", "VPERMPD", []Operand{Imm(0xd8), Y7, Y7}, "c4e3fd01ffd8"},
+		{"VPERMILPD", "VPERMILPD", []Operand{Imm(0xff), X1, X2}, "c4e37905d1ff"},
+		{"VPERMILPS", "VPERMILPS", []Operand{Imm(0xff), X1, X2}, "c4e37904d1ff"},
+		{"VROUNDPD", "VROUNDPD", []Operand{Imm(-1), X1, X2}, "c4e37909d1ff"},
+		{"VROUNDPS", "VROUNDPS", []Operand{Imm(-1), Y1, Y2}, "c4e37d08d1ff"},
+		{"VAESKEYGENASSIST", "VAESKEYGENASSIST", []Operand{Imm(-1), X1, X2}, "c4e379dfd1ff"},
+		{"VPCMPESTRI", "VPCMPESTRI", []Operand{Imm(-1), X1, X2}, "c4e37961d1ff"},
+		{"VPCMPESTRM", "VPCMPESTRM", []Operand{Imm(-1), X1, X2}, "c4e37960d1ff"},
+		{"VPCMPISTRI", "VPCMPISTRI", []Operand{Imm(-1), X1, X2}, "c4e37963d1ff"},
+		{"VPCMPISTRM", "VPCMPISTRM", []Operand{Imm(-1), X1, X2}, "c4e37962d1ff"},
+		{"VEXTRACTPS", "VEXTRACTPS", []Operand{Imm(-1), X1, AX}, "c4e37917c8ff"},
+		{"VPEXTRW", "VPEXTRW", []Operand{Imm(0xff), X1, AX}, "c4e37915c8ff"},
+		{"VPBLENDVB", "VPBLENDVB", []Operand{X0, Ptr(BX, 0, 16), X1, X2}, "c4e3714c1300"},
+		{"VMOVHPD load", "VMOVHPD", []Operand{Ptr(AX, 0, 8), X5, X5}, "c5d11628"},
+		{"VMOVHPD load disp", "VMOVHPD", []Operand{Ptr(DX, 7, 8), X5, X5}, "c5d1166a07"},
+		{"VMOVHPD store", "VMOVHPD", []Operand{X5, Ptr(AX, 0, 8)}, "c5f91728"},
+		{"VMOVLPD load", "VMOVLPD", []Operand{Ptr(AX, 0, 8), X5, X5}, "c5d11228"},
+		{"VMOVLPD store", "VMOVLPD", []Operand{X5, Ptr(AX, 0, 8)}, "c5f91328"},
+		{"VMOVQ EVEX gpr load", "VMOVQ", []Operand{Reg{idx: 4, size: 8}, X20}, "62e1fd086ee4"},
+		{"VMOVQ EVEX mem store", "VMOVQ", []Operand{X20, Ptr(AX, 0, 8)}, "62e1fd087e20"},
+		{"VMOVQ EVEX mem load", "VMOVQ", []Operand{Ptr(AX, 0, 8), X20}, "62e1fd086e20"},
+	}
+	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)
+		}
+	}
+}
+
+// TestCorpusTailFileForms pins the file-level forms the toolchain's amd64
+// testdata carries: the star-marked indirect jumps, the TLS pseudo-base, the
+// paired-register shift spelling, the absolute RET and the jump to an
+// undefined static symbol (its displacement and the TLS slot offsets are
+// relocation sites, zeroed here as the kernel parity suites do).
+func TestCorpusTailFileForms(t *testing.T) {
+	mask32 := func(b []byte, at int) { b[at], b[at+1], b[at+2], b[at+3] = 0, 0, 0, 0 }
+	cases := []struct {
+		name string
+		src  string
+		want string // hex, with X marking a masked 32-bit relocation site
+	}{
+		{"star reg jump", "\tJMP *(R12)\n\tRET\n", "41ff2424c3"},
+		{"star sp jump", "\tJMP *4(SP)\n\tRET\n", "ff642404c3"},
+		{"star indexed jump", "\tJMP *(R12)(R13*4)\n\tRET\n", "43ff24acc3"},
+		{"TLS load", "\tMOVQ (TLS), AX\n\tRET\n", "64488b0425XXXXXXXXc3"},
+		{"TLS load offset", "\tMOVQ 8(TLS), DX\n\tRET\n", "64488b1425XXXXXXXXc3"},
+		{"colon shift", "\tSHLL CX, R11:AX\n\tRET\n", "410fa5c3c3"},
+		{"SP indexed local", "\tMOVQ foo(SP)(AX*1), BX\n\tRET\n", "488b1c04c3"},
+	}
+	for _, c := range cases {
+		f, errs := parser.Parse("t_amd64.s", "#include \"textflag.h\"\nTEXT ·f(SB), NOSPLIT, $0\n"+c.src)
+		if len(errs) > 0 {
+			t.Errorf("%s: parse: %v", c.name, errs)
+			continue
+		}
+		img, err := AssembleFile(f)
+		if err != nil {
+			t.Errorf("%s: assemble: %v", c.name, err)
+			continue
+		}
+		fn := img.Funcs[0]
+		code := append([]byte(nil), img.Code[fn.Offset:fn.Offset+fn.Size]...)
+		if at := strings.Index(c.want, "XXXXXXXX"); at >= 0 {
+			mask32(code, at/2) // the masked relocation site
+		}
+		if got := hexCompact(code); got != strings.ReplaceAll(c.want, "X", "0") {
+			t.Errorf("%s: bytes %s, want %s", c.name, got, c.want)
+		}
+	}
+	// JMP to an undefined static symbol and the absolute RET: their rel32
+	// carries a call relocation against the symbol, masked to zero here.
+	for _, c := range []struct{ name, src string }{
+		{"static external jump", "\tJMP bar<>+4(SB)\n\tRET\n"},
+		{"static external indexed jump", "\tJMP bar<>+4(SB)(R11*4)\n\tRET\n"},
+		{"absolute ret", "\tRET\n\tRET foo(SB)\n"},
+	} {
+		f, errs := parser.Parse("t_amd64.s", "#include \"textflag.h\"\nTEXT ·f(SB), NOSPLIT, $0\n"+c.src)
+		if len(errs) > 0 {
+			t.Errorf("%s: parse: %v", c.name, errs)
+			continue
+		}
+		img, err := AssembleFile(f)
+		if err != nil {
+			t.Errorf("%s: assemble: %v", c.name, err)
+			continue
+		}
+		fn := img.Funcs[0]
+		code := maskCode(append([]byte(nil), img.Code[fn.Offset:fn.Offset+fn.Size]...), fn.Relocs)
+		want := "e900000000c3"
+		if c.name == "absolute ret" {
+			want = "c3e900000000"
+		}
+		if got := hexCompact(code); got != want {
+			t.Errorf("%s: bytes %s, want %s", c.name, got, want)
+		}
+	}
+}

asm/evex.go

@@ -856,37 +856,41 @@ type evexMoveSpec struct {
 	vecOK   bool // the non-memory operand may be a vector register
 	xmmOnly bool // wider than XMM registers are rejected
 	nds3    bool // a three-operand register form exists (VMOVSD/VMOVSS)
+	gprOK   bool // the r/m side may be a general-purpose register (VMOVQ)
 }
 
 // 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}, true, false, false},
+	"VMOVDQU32": {1, 2, 0x6F, 0x7F, 0, [3]int{16, 32, 64}, true, false, false, false},
 	// EVEX.128/256/512.F3.0F.W1, unaligned qword move.
-	"VMOVDQU64": {1, 2, 0x6F, 0x7F, 1, [3]int{16, 32, 64}, true, false, false},
+	"VMOVDQU64": {1, 2, 0x6F, 0x7F, 1, [3]int{16, 32, 64}, true, false, false, 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}, true, false, false},
+	"VMOVDQU8": {1, 3, 0x6F, 0x7F, 0, [3]int{16, 32, 64}, true, false, false, 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}, true, false, false},
+	"VMOVDQU16": {1, 3, 0x6F, 0x7F, 1, [3]int{16, 32, 64}, true, false, false, false},
 	// EVEX.128/256/512.66.0F.W1, unaligned packed double move.
-	"VMOVUPD": {1, 1, 0x10, 0x11, 1, [3]int{16, 32, 64}, true, false, false},
+	"VMOVUPD": {1, 1, 0x10, 0x11, 1, [3]int{16, 32, 64}, true, false, false, false},
 	// EVEX.128/256/512, aligned packed moves.
-	"VMOVAPS": {1, 0, 0x28, 0x29, 0, [3]int{16, 32, 64}, true, false, false},
-	"VMOVAPD": {1, 1, 0x28, 0x29, 1, [3]int{16, 32, 64}, true, false, false},
+	"VMOVAPS": {1, 0, 0x28, 0x29, 0, [3]int{16, 32, 64}, true, false, false, false},
+	"VMOVAPD": {1, 1, 0x28, 0x29, 1, [3]int{16, 32, 64}, true, false, false, false},
 	// EVEX.128/256/512.66.0F, aligned integer moves.
-	"VMOVDQA32": {1, 1, 0x6F, 0x7F, 0, [3]int{16, 32, 64}, true, false, false},
-	"VMOVDQA64": {1, 1, 0x6F, 0x7F, 1, [3]int{16, 32, 64}, true, false, false},
+	"VMOVDQA32": {1, 1, 0x6F, 0x7F, 0, [3]int{16, 32, 64}, true, false, false, false},
+	"VMOVDQA64": {1, 1, 0x6F, 0x7F, 1, [3]int{16, 32, 64}, true, false, false, 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}, false, true, true},
+	"VMOVSS": {1, 2, 0x10, 0x11, 0, [3]int{4, 4, 4}, false, true, true, false},
 	// EVEX.128.F2.0F.W1, scalar double move: memory operands and the
 	// three-operand register form (VMOVSD dst, src1, src2).
-	"VMOVSD": {1, 3, 0x10, 0x11, 1, [3]int{8, 8, 8}, false, true, true},
+	"VMOVSD": {1, 3, 0x10, 0x11, 1, [3]int{8, 8, 8}, false, true, true, false},
 	// EVEX.128/256/512.0F.W0, unaligned packed single move.
-	"VMOVUPS": {1, 0, 0x10, 0x11, 0, [3]int{16, 32, 64}, true, false, false},
+	"VMOVUPS": {1, 0, 0x10, 0x11, 0, [3]int{16, 32, 64}, true, false, false, false},
+	// EVEX.128.66.0F.W1, the 64-bit GPR/memory ↔ XMM move (VMOVQ RSP, X20
+	// and friends, the EVEX spelling the high registers demand).
+	"VMOVQ": {1, 1, 0x6E, 0x7E, 1, [3]int{8, 8, 8}, true, true, false, true},
 }
 
 // isEvex reports whether the mnemonic has an EVEX encoding we handle.
@@ -900,6 +904,9 @@ func isEvex(mnemUpper string) bool {
 	if _, ok := evexMoveTable[mnemUpper]; ok {
 		return true
 	}
+	if _, ok := evexHptrTable[mnemUpper]; ok {
+		return true
+	}
 	return isEvexQuad(mnemUpper)
 }
 
@@ -911,8 +918,14 @@ func evexRequired(upper string, ops []Operand) bool {
 	_, inVex := vexTable[upper]
 	_, inVexMove := vexMoveTable[upper]
 	if !inVex && !inVexMove {
+		// The dual-shape moves pick their VEX form by operand count, so
+		// they are not EVEX-only either.
+		switch upper {
+		case "VMOVHPD", "VMOVLPD":
+		default:
 			return true // EVEX-only mnemonic
 		}
+	}
 	// The byte-quad shifts have VEX register forms but EVEX-only memory
 	// forms: a memory count source forces the EVEX encoding.
 	if upper == "VPSLLDQ" || upper == "VPSRLDQ" {
@@ -1018,6 +1031,8 @@ var evexRound = map[string]bool{
 	"VCVTTSD2USIL": true, "VCVTTSD2USIQ": true, "VCVTTSS2USIL": true, "VCVTTSS2USIQ": true,
 	"VCVTSI2SDQ": true, "VCVTSI2SSL": true, "VCVTSI2SSQ": true,
 	"VCVTUSI2SDQ": true, "VCVTUSI2SSL": true, "VCVTUSI2SSQ": true,
+	// The scalar compares suppress exceptions on their LIG encoding.
+	"VCMPSD": true, "VCMPSS": true,
 }
 
 // evexBcstN maps an instruction accepting .BCST to the broadcast element
@@ -1082,6 +1097,14 @@ func (e *enc) encodeEvex(mnemUpper string, ops []Operand, sfx evexSuffix) error
 		return e.encodeEvexRM(spec, ops, 0, sfx)
 	}
 	spec, inTable := evexTable[mnemUpper]
+	// A high/low half move that lives in the hptr table alone (the packed
+	// double twins) reaches the same inTable block below, which completes
+	// its spec from the hptr entry.
+	if !inTable {
+		if _, ok := evexHptrTable[mnemUpper]; ok {
+			inTable = true
+		}
+	}
 	if q, ok := evexQuadTable[mnemUpper]; ok {
 		// The quad-register family carries no rounding, SAE or broadcast;
 		// only masking and zeroing apply.
@@ -1446,6 +1469,19 @@ func (e *enc) encodeEvexExtractGPR(spec evexSpec, ops []Operand, mask int, sfx e
 // assembler.  The scalar moves also carry a three-operand register form
 // (VMOVSD dst, src1, src2: the load opcode with vvvv = src1), which ms.nds3
 // opens.
+// validEvexMoveOther reports whether the non-vector side of an EVEX move may
+// take the operand: memory always, a general-purpose register when gprOK.
+func validEvexMoveOther(ms evexMoveSpec, op Operand) bool {
+	if memOperand(op) {
+		return true
+	}
+	if !ms.gprOK {
+		return false
+	}
+	r, ok := op.(Reg)
+	return ok && !r.isVec() && !r.mask && r.ctl == 0 && !r.mmx && !r.fp
+}
+
 func (e *enc) encodeEvexMove(mnem string, ms evexMoveSpec, ops []Operand, mask int, sfx evexSuffix) error {
 	if len(ops) == 3 {
 		if !ms.nds3 {
@@ -1453,7 +1489,7 @@ func (e *enc) encodeEvexMove(mnem string, ms evexMoveSpec, ops []Operand, mask i
 		}
 		// The masked scalar register form keeps the Go assembler's own
 		// layout: the store opcode with reg = op0, vvvv = op1 and the
-		// destination in r/m (op2) — the bytes go tool asm emits, not
+		// destination in r/m (op2), the bytes go tool asm emits, not
 		// the manual's NDS reading.
 		src, src1, dst := ops[0], ops[1], ops[2]
 		reg, ok := src.(Reg)
@@ -1494,12 +1530,12 @@ func (e *enc) encodeEvexMove(mnem string, ms evexMoveSpec, ops []Operand, mask i
 		}
 		reg, rm = srcReg, dst
 	case srcIsVec:
-		if !memOperand(dst) {
+		if !validEvexMoveOther(ms, dst) {
 			return fmt.Errorf("%s: invalid destination operand", mnem)
 		}
 		reg, rm = srcReg, dst
 	case dstIsVec:
-		if !memOperand(src) {
+		if !validEvexMoveOther(ms, src) {
 			return fmt.Errorf("%s: invalid source operand", mnem)
 		}
 		op = ms.load
@@ -1890,6 +1926,15 @@ var evexHptrTable = map[string]evexHptrSpec{
 	"VMOVLHPS": {
 		insert: evexSpec{mapSel: 1, opcode: 0x16, w: 0, pp: 0, opdigit: -1, form: vexNDS3, n: [3]int{8, 0, 0}},
 	},
+	// The packed-double twins, 66-prefixed.
+	"VMOVHPD": {
+		insert: evexSpec{mapSel: 1, opcode: 0x16, w: 1, pp: 1, opdigit: -1, form: vexNDS3, n: [3]int{8, 0, 0}},
+		store:  evexSpec{mapSel: 1, opcode: 0x17, w: 1, pp: 1, opdigit: -1, form: vexRMRev, n: [3]int{8, 0, 0}},
+	},
+	"VMOVLPD": {
+		insert: evexSpec{mapSel: 1, opcode: 0x12, w: 1, pp: 1, opdigit: -1, form: vexNDS3, n: [3]int{8, 0, 0}},
+		store:  evexSpec{mapSel: 1, opcode: 0x13, w: 1, pp: 1, opdigit: -1, form: vexRMRev, n: [3]int{8, 0, 0}},
+	},
 }
 
 // encodeEvexPrefGather encodes a gather/scatter prefetch hint: OP K, vsib.
@@ -1954,7 +1999,7 @@ func (e *enc) encodeGather(upper string, gs gatherSpec, ops []Operand, sfx evexS
 	if !ok || !maskReg.isVec() {
 		return fmt.Errorf("%s: mask must be a vector register", upper)
 	}
-	vsib, _, err := vsibLen(rest[1], upper)
+	vsib, idxLen, err := vsibLen(rest[1], upper)
 	if err != nil {
 		return err
 	}
@@ -1962,12 +2007,16 @@ func (e *enc) encodeGather(upper string, gs gatherSpec, ops []Operand, sfx evexS
 	if !ok || !dst.isVec() {
 		return fmt.Errorf("%s: destination must be a vector register", upper)
 	}
+	// The L bit is the wider of the data register and the VSIB index
+	// lengths (a YMM index under an XMM destination selects 256-bit, the
+	// bytes go tool asm emits).
+	ll := max(idxLen, dst.vecLenBit())
 	spec := vexSpec{mapSel: 2, opcode: gs.opcode, w: gs.w, pp: 1, opdigit: -1}
 	rBit := 0
 	if dst.idx >= 8 {
 		rBit = 1
 	}
-	return e.emitVexFields(spec, dst.vecLenBit(), dst.idx&7, rBit, 15-maskReg.idx, vsib)
+	return e.emitVexFields(spec, ll, dst.idx&7, rBit, 15-maskReg.idx, vsib)
 }
 
 // encodeScatter encodes a scatter (EVEX only): OP src, K, vsib, reg = src,

asm/extension.go

@@ -0,0 +1,159 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+// The extended-instruction registry: the lookup over and above the generated
+// architecture tables.  The generated tables (arch/*_gen.go) list the
+// mnemonics the Go toolchain knows; the extension layer carries the
+// instructions it does not, and this file indexes them per architecture so
+// the assembler and the linter can consult the layer without touching the
+// generated lists or the main encoders.  A later hook wires
+// ExtensionEncodable into the Encodable mirror and EncodeExtension into the
+// per-architecture assembly paths; nothing existing changes until then.
+
+package asm
+
+import (
+	"fmt"
+	"slices"
+	"strings"
+
+	"sourcedock.dev/petrbalvin/gasm-sdk/arch"
+)
+
+// extensionIndex is the per-architecture index of the extension layer, keyed
+// by upper-case mnemonic.  One mnemonic registers several forms (the SVE ADD
+// carries unpredicated, predicated and immediate shapes), so the value is the
+// full candidate list in table order.
+type extensionIndex struct {
+	byName map[string][]arch.ExtInstr
+}
+
+// extensionIndexes builds one index per known architecture.  Architectures
+// whose extension layer is not built yet get an empty index, which keeps the
+// queries answering false rather than failing on a missing entry.
+var extensionIndexes = buildExtensionIndexes()
+
+func buildExtensionIndexes() map[arch.Arch]*extensionIndex {
+	m := make(map[arch.Arch]*extensionIndex)
+	for _, a := range []arch.Arch{arch.AMD64, arch.ARM64, arch.RISCV, arch.LOONG64} {
+		idx := &extensionIndex{byName: make(map[string][]arch.ExtInstr)}
+		for _, in := range arch.Extensions(a) {
+			key := strings.ToUpper(in.Name)
+			idx.byName[key] = append(idx.byName[key], in)
+		}
+		m[a] = idx
+	}
+	return m
+}
+
+// LookupExtension returns the extended instructions registered for the
+// mnemonic on a, outside the generated architecture table.  It reports false
+// when a carries no extended layer or the mnemonic is not in it; a mnemonic
+// the base table knows is not thereby covered, the layers stay independent.
+func LookupExtension(a arch.Arch, mnemonic string) ([]arch.ExtInstr, bool) {
+	idx, ok := extensionIndexes[a]
+	if !ok || idx == nil {
+		return nil, false
+	}
+	cands, ok := idx.byName[strings.ToUpper(mnemonic)]
+	return cands, ok && len(cands) > 0
+}
+
+// ExtensionNames returns the mnemonics the extension layer of a registers,
+// in table order, without duplicates.
+func ExtensionNames(a arch.Arch) []string {
+	var names []string
+	seen := make(map[string]bool)
+	for _, in := range arch.Extensions(a) {
+		key := strings.ToUpper(in.Name)
+		if !seen[key] {
+			seen[key] = true
+			names = append(names, in.Name)
+		}
+	}
+	return names
+}
+
+// EncodeExtension encodes one extended instruction on a: it resolves the
+// mnemonic through the extension registry, picks the registered form whose
+// arity matches the operands and encodes against it.  The first form that
+// encodes wins.  When every matching form rejects the operands, the error
+// comes from the form whose operand kinds the list points at (the one with
+// the most matching positions), so a mis-spelled predicate qualifier is
+// diagnosed as one, not as the unpredicated form's register complaint.
+func EncodeExtension(a arch.Arch, mnemonic string, ops ...arch.ExtOperand) ([]byte, error) {
+	cands, ok := LookupExtension(a, mnemonic)
+	if !ok {
+		return nil, fmt.Errorf("%s registers no extended instruction %q", a, mnemonic)
+	}
+	var bestErr error
+	var bestScore int
+	var tried int
+	for _, in := range cands {
+		if in.Form.Arity() != len(ops) {
+			continue
+		}
+		tried++
+		b, err := in.Encode(ops)
+		if err == nil {
+			return b, nil
+		}
+		if score := kindScore(in.Form, ops); bestErr == nil || score > bestScore {
+			bestErr, bestScore = err, score
+		}
+	}
+	if tried == 0 {
+		return nil, fmt.Errorf("%s: extended %q takes %s, got %d operands",
+			a, mnemonic, extensionAritySummary(cands), len(ops))
+	}
+	return nil, bestErr
+}
+
+// kindScore counts the positions whose operand kind matches what the form
+// wants, the tie-break that picks the most specific rejection.
+func kindScore(form arch.ExtForm, ops []arch.ExtOperand) int {
+	kinds := form.Kinds()
+	score := 0
+	for i, op := range ops {
+		if i < len(kinds) && op.Kind == kinds[i] {
+			score++
+		}
+	}
+	return score
+}
+
+// ExtensionEncodable reports whether the extension layer of a encodes the
+// mnemonic with these operands.  It mirrors asm.Encodable for the extension
+// layer: the predicate the linter consults once the hook wires it in.
+func ExtensionEncodable(a arch.Arch, mnemonic string, ops ...arch.ExtOperand) bool {
+	_, err := EncodeExtension(a, mnemonic, ops...)
+	return err == nil
+}
+
+// extensionAritySummary describes the operand counts the candidate forms
+// take, "2 or 3" style, for the arity error.
+func extensionAritySummary(cands []arch.ExtInstr) string {
+	counts := make([]int, 0, len(cands))
+	seen := make(map[int]bool)
+	for _, in := range cands {
+		n := in.Form.Arity()
+		if !seen[n] {
+			seen[n] = true
+			counts = append(counts, n)
+		}
+	}
+	slices.Sort(counts)
+	var b strings.Builder
+	for i, n := range counts {
+		if i > 0 {
+			if i == len(counts)-1 {
+				b.WriteString(" or ")
+			} else {
+				b.WriteString(", ")
+			}
+		}
+		fmt.Fprintf(&b, "%d", n)
+	}
+	b.WriteString(" operands")
+	return b.String()
+}

asm/extension_test.go

@@ -0,0 +1,197 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package asm
+
+import (
+	"encoding/hex"
+	"strings"
+	"testing"
+
+	"sourcedock.dev/petrbalvin/gasm-sdk/arch"
+)
+
+// TestExtensionRegistryARM64 checks the mnemonic lookup over and above the
+// generated arm64 table: one mnemonic, several forms, case-insensitive, and
+// nothing offered for a spelling the layer does not carry.
+func TestExtensionRegistryARM64(t *testing.T) {
+	add, ok := LookupExtension(arch.ARM64, "ADD")
+	if !ok {
+		t.Fatal("LookupExtension(ARM64, ADD) found nothing")
+	}
+	var forms []arch.ExtForm
+	for _, in := range add {
+		if in.Name != "ADD" {
+			t.Errorf("candidate %q leaked into the ADD lookup", in.Name)
+		}
+		forms = append(forms, in.Form)
+	}
+	if len(forms) != 3 ||
+		forms[0] != arch.ExtFormVectors ||
+		forms[1] != arch.ExtFormPredicated ||
+		forms[2] != arch.ExtFormImmediate {
+		t.Errorf("ADD registers forms %v, want unpredicated, predicated and immediate", forms)
+	}
+	if _, ok := LookupExtension(arch.ARM64, "add"); !ok {
+		t.Error("the lookup is case-sensitive")
+	}
+	if _, ok := LookupExtension(arch.ARM64, "NOSUCHINSTR"); ok {
+		t.Error("a non-extended mnemonic resolved")
+	}
+	sqadd, ok := LookupExtension(arch.ARM64, "SQADD")
+	if !ok || len(sqadd) != 2 {
+		t.Errorf("SQADD registers %d forms, want the unpredicated and immediate pair", len(sqadd))
+	}
+}
+
+// TestExtensionAboveGeneratedTable pins the layering: SQADD is nowhere in the
+// generated arm64 table (the toolchain knows only the NEON spelling VSQADD)
+// yet the extension layer carries it, while ADD sits in both layers
+// independently.
+func TestExtensionAboveGeneratedTable(t *testing.T) {
+	if _, found := arch.ForArch(arch.ARM64).Lookup("SQADD"); found {
+		t.Error("SQADD is in the generated table, the layering assumption broke")
+	}
+	if _, ok := LookupExtension(arch.ARM64, "SQADD"); !ok {
+		t.Error("SQADD is missing from the extension layer")
+	}
+	if _, found := arch.ForArch(arch.ARM64).Lookup("ADD"); !found {
+		t.Error("ADD vanished from the generated table")
+	}
+	if add, ok := LookupExtension(arch.ARM64, "ADD"); !ok || len(add) != 3 {
+		t.Errorf("ADD carries %d extension forms, want 3", len(add))
+	}
+}
+
+// TestEncodeExtensionGolden encodes through the registry and pins the same
+// golden words the arch table tests pin, proving the registry resolves to the
+// right encoding.
+func TestEncodeExtensionGolden(t *testing.T) {
+	for _, tt := range []struct {
+		name string
+		mnem string
+		ops  []arch.ExtOperand
+		want uint32
+	}{
+		{"unpredicated add", "ADD",
+			[]arch.ExtOperand{
+				arch.ExtVector(2, arch.ExtArrB), arch.ExtVector(0, arch.ExtArrB), arch.ExtVector(0, arch.ExtArrB),
+			},
+			0x04200040},
+		{"predicated mul", "MUL",
+			[]arch.ExtOperand{
+				arch.ExtVector(0, arch.ExtArrB), arch.ExtPredicate(2, arch.ExtQualMerging), arch.ExtVector(0, arch.ExtArrB),
+			},
+			0x04100800},
+		{"immediate add with derived shift", "ADD",
+			[]arch.ExtOperand{arch.ExtImmediate(32512), arch.ExtVector(0, arch.ExtArrH)},
+			0x2560efe0},
+		{"signed immediate mul", "MUL",
+			[]arch.ExtOperand{arch.ExtImmediate(-1), arch.ExtVector(0, arch.ExtArrB)},
+			0x2530dfe0},
+	} {
+		got, err := EncodeExtension(arch.ARM64, tt.mnem, tt.ops...)
+		if err != nil {
+			t.Errorf("%s: encode: %v", tt.name, err)
+			continue
+		}
+		if want := hex.EncodeToString([]byte{
+			byte(tt.want), byte(tt.want >> 8), byte(tt.want >> 16), byte(tt.want >> 24),
+		}); hex.EncodeToString(got) != want {
+			t.Errorf("%s:\n  got  %x\n  want %s", tt.name, got, want)
+		}
+	}
+}
+
+// TestEncodeExtensionErrors checks the registry's diagnostics: a wrong arity
+// names every form's count, an operand the first candidate rejects surfaces
+// its own message once a later form takes over.
+func TestEncodeExtensionErrors(t *testing.T) {
+	if _, err := EncodeExtension(arch.ARM64, "ADD", arch.ExtVector(0, arch.ExtArrB)); err == nil {
+		t.Error("one operand encoded, want an arity error")
+	} else if !strings.Contains(err.Error(), "2 or 3 operands") {
+		t.Errorf("arity error %q does not name the counts", err)
+	}
+	// The predicated candidate must answer for its own operands: the /Z
+	// qualifier is rejected with the merging message, not the unpredicated
+	// form's register-kind complaint.
+	_, err := EncodeExtension(arch.ARM64, "ADD",
+		arch.ExtVector(0, arch.ExtArrB), arch.ExtPredicate(0, arch.ExtQualZeroing), arch.ExtVector(0, arch.ExtArrB))
+	if err == nil {
+		t.Fatal("/Z encoded, want an error")
+	}
+	if !strings.Contains(err.Error(), "/M") {
+		t.Errorf("error %q does not name the merging qualifier", err)
+	}
+	if _, err := EncodeExtension(arch.ARM64, "NOSUCHINSTR", arch.ExtVector(0, arch.ExtArrB)); err == nil ||
+		!strings.Contains(err.Error(), "registers no extended instruction") {
+		t.Errorf("unknown mnemonic error = %v", err)
+	}
+}
+
+// TestExtensionEncodable checks the predicate the later Encodable hook will
+// call: true exactly when the registry encodes the operand list.
+func TestExtensionEncodable(t *testing.T) {
+	if !ExtensionEncodable(arch.ARM64, "ADD",
+		arch.ExtVector(0, arch.ExtArrS), arch.ExtVector(1, arch.ExtArrS), arch.ExtVector(2, arch.ExtArrS)) {
+		t.Error("an encodable unpredicated add reported false")
+	}
+	if !ExtensionEncodable(arch.ARM64, "ADD",
+		arch.ExtVector(1, arch.ExtArrS), arch.ExtPredicate(0, arch.ExtQualMerging), arch.ExtVector(0, arch.ExtArrS)) {
+		t.Error("an encodable predicated add reported false")
+	}
+	if ExtensionEncodable(arch.ARM64, "ADD",
+		arch.ExtVector(0, arch.ExtArrB), arch.ExtVector(0, arch.ExtArrS), arch.ExtVector(0, arch.ExtArrB)) {
+		t.Error("mismatched arrangements reported encodable")
+	}
+	if ExtensionEncodable(arch.ARM64, "ADD", arch.ExtVector(0, arch.ExtArrB)) {
+		t.Error("a one-operand add reported encodable")
+	}
+	if ExtensionEncodable(arch.ARM64, "NOSUCHINSTR") {
+		t.Error("an unregistered mnemonic reported encodable")
+	}
+}
+
+// TestExtensionArchIsolation is the architecture-binding negative case: the
+// extension layer is registered for arm64 alone, and no other architecture
+// answers its queries, not even for a mnemonic the amd64 base table carries.
+func TestExtensionArchIsolation(t *testing.T) {
+	ops := []arch.ExtOperand{
+		arch.ExtVector(0, arch.ExtArrB), arch.ExtVector(0, arch.ExtArrB), arch.ExtVector(0, arch.ExtArrB),
+	}
+	for _, a := range []arch.Arch{arch.AMD64, arch.RISCV, arch.LOONG64, arch.Unknown} {
+		if cands, ok := LookupExtension(a, "ADD"); ok || cands != nil {
+			t.Errorf("LookupExtension(%s, ADD) offered %d candidates", a, len(cands))
+		}
+		if cands, ok := LookupExtension(a, "MUL"); ok || cands != nil {
+			t.Errorf("LookupExtension(%s, MUL) offered %d candidates", a, len(cands))
+		}
+		if got, err := EncodeExtension(a, "ADD", ops...); err == nil {
+			t.Errorf("EncodeExtension(%s, ADD) encoded %x, want a refusal", a, got)
+		} else if !strings.Contains(err.Error(), string(a)) {
+			t.Errorf("EncodeExtension(%s) error %q does not name the architecture", a, err)
+		}
+		if ExtensionEncodable(a, "ADD", ops...) {
+			t.Errorf("ExtensionEncodable(%s, ADD) reported true", a)
+		}
+		if names := ExtensionNames(a); len(names) != 0 {
+			t.Errorf("ExtensionNames(%s) = %v, want none", a, names)
+		}
+		if got := arch.Extensions(a); len(got) != 0 {
+			t.Errorf("arch.Extensions(%s) carries %d instructions", a, len(got))
+		}
+	}
+}
+
+// TestExtensionNamesARM64 checks the completion-facing name list: every
+// distinct mnemonic of the family, first-occurrence order, no duplicates.
+func TestExtensionNamesARM64(t *testing.T) {
+	want := []string{"ADD", "SUB", "SQADD", "UQADD", "SQSUB", "UQSUB", "MUL", "SMULH", "UMULH", "SUBR"}
+	got := ExtensionNames(arch.ARM64)
+	if strings.Join(got, ",") != strings.Join(want, ",") {
+		t.Errorf("ExtensionNames(ARM64) = %v, want %v", got, want)
+	}
+	if n := len(arch.Extensions(arch.ARM64)); n != 23 {
+		t.Errorf("the family registers %d instructions, want 23", n)
+	}
+}

asm/goobj.go

@@ -63,6 +63,7 @@ const (
 const (
 	kindSTEXT       = 1
 	kindSRODATA     = 3
+	kindSNOPTRDATA  = 5
 	kindSDATA       = 7
 	kindSDWARFFCN   = 14
 	kindSDWARFLINES = 20
@@ -305,9 +306,16 @@ func (img *Image) emitGOObject(pkgPath, srcPath string, pre []byte, minLC int, r
 		if !d.Static {
 			name = pkgPath + "." + name
 		}
+		// RODATA implies no pointers, so it wins over NOPTR: the kind is
+		// SRODATA either way, exactly as the toolchain chooses it.  Plain
+		// NOPTR data is SNOPTRDATA, which the linker keeps out of the GC's
+		// type scan; a plain SDATA symbol would demand Go type information
+		// no assembly file can supply, and the link would fail.
 		typ := uint8(kindSDATA)
 		if d.Rodata {
 			typ = kindSRODATA
+		} else if d.Noptr {
+			typ = kindSNOPTRDATA
 		}
 		flag := uint8(0)
 		if d.Dupok {

asm/goobj_resolve_test.go

@@ -45,6 +45,9 @@ func TestReadRuntimeSymbols(t *testing.T) {
 // TestResolveExternalSymbols verifies end-to-end resolution of external
 // symbol references.
 func TestResolveExternalSymbols(t *testing.T) {
+	if testing.Short() {
+		t.Skip("resolves through a live go list -export: skipped in -short mode")
+	}
 	if _, err := exec.LookPath("go"); err != nil {
 		t.Skip("go toolchain not available")
 	}

asm/goobj_test.go

@@ -12,7 +12,7 @@ import (
 	"strings"
 	"testing"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
 )
 
 // goobjView is a minimal parsed view of a GOOBJ payload, enough to check
@@ -514,6 +514,145 @@ func fieldAfter(line, flag string) string {
 	return ""
 }
 
+// buildLogSteps is what a substitution test needs from a `go build -x -work`
+// log: the work directory, the assembler's object, the package archive and
+// the link command line.
+type buildLogSteps struct {
+	work     string
+	asmObj   string // $WORK expanded
+	pkgArch  string // $WORK expanded
+	linkLine string // still carries $WORK placeholders
+}
+
+// parseBuildLog extracts the build steps from a `go build -x -work` log.
+// asmFile names the assembly file whose object the test substitutes.  A
+// missing step is a failure, not a skip: the toolchain changed shape and the
+// substitution would silently test nothing.
+func parseBuildLog(t *testing.T, log []byte, asmFile string) buildLogSteps {
+	t.Helper()
+	var st buildLogSteps
+	for line := range strings.SplitSeq(string(log), "\n") {
+		switch {
+		case strings.HasPrefix(line, "WORK="):
+			st.work = strings.TrimPrefix(line, "WORK=")
+		case strings.Contains(line, "/asm ") && strings.Contains(line, asmFile) && !strings.Contains(line, "-gensymabis"):
+			st.asmObj = fieldAfter(line, "-o")
+		case strings.Contains(line, "pack r") && strings.Contains(line, "_pkg_.a"):
+			rest := strings.TrimSpace(strings.SplitN(line, "pack r", 2)[1])
+			st.pkgArch = strings.Fields(strings.SplitN(rest, "#", 2)[0])[0]
+		case strings.Contains(line, "/link ") && strings.Contains(line, "-importcfg"):
+			st.linkLine = line
+		}
+	}
+	if st.work == "" || st.asmObj == "" || st.pkgArch == "" || st.linkLine == "" {
+		t.Fatalf("could not locate the build steps (work=%q asmObj=%q pkgArch=%q link=%q):\n%s",
+			st.work, st.asmObj, st.pkgArch, st.linkLine, log)
+	}
+	st.asmObj = strings.ReplaceAll(st.asmObj, "$WORK", st.work)
+	st.pkgArch = strings.ReplaceAll(st.pkgArch, "$WORK", st.work)
+	return st
+}
+
+// substituteAndRelink swaps the gasm object into the package archive the
+// baseline build produced and re-runs the captured link line against the
+// rebuilt archive, writing the binary to outBin (the -x log's link step
+// always targets the action graph's internal a.out, which the helper
+// redirects; the copy to the -o target is a separate build action the helper
+// does not need).  The archive handed to the linker is proven to carry the
+// gasm object byte for byte, so a build-layout change that skipped the
+// substitution fails here instead of passing vacuously.
+func substituteAndRelink(t *testing.T, goBin, dir string, st buildLogSteps, outBin string, gasmObj []byte, extraEnv ...string) {
+	t.Helper()
+
+	// The deliberate-run boundary: this path drives a real `go build` and
+	// cmd/link per invocation, minutes-scale work on the small single-core
+	// CI runner.  Under -short (the push pipeline's mode) it skips; the
+	// local test gate and the dispatched workflows run it in full.
+	if testing.Short() {
+		t.Skip("end-to-end go build and link: skipped in -short mode")
+	}
+
+	// Extract the archive, overwrite the assembler's member with the gasm
+	// object and repack (go tool pack has no replace-in-place).
+	membersDir := filepath.Join(dir, "members")
+	if err := os.MkdirAll(membersDir, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	extract := exec.Command(goBin, "tool", "pack", "x", st.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(st.asmObj))
+	if _, err := os.Stat(member); err != nil {
+		t.Fatalf("the assembler's archive member was not extracted: %v", err)
+	}
+	if err := os.Chmod(member, 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(member, gasmObj, 0o644); err != nil {
+		t.Fatal(err)
+	}
+	listCmd := exec.Command(goBin, "tool", "pack", "t", st.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, 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)
+	}
+
+	// Prove the substitution: the archive the linker is about to consume
+	// holds the gasm object, byte for byte.
+	checkDir := filepath.Join(dir, "check")
+	if err := os.MkdirAll(checkDir, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	check := exec.Command(goBin, "tool", "pack", "x", newArch)
+	check.Dir = checkDir
+	if out, err := check.CombinedOutput(); err != nil {
+		t.Fatalf("pack x (verification): %v\n%s", err, out)
+	}
+	got, err := os.ReadFile(filepath.Join(checkDir, filepath.Base(st.asmObj)))
+	if err != nil {
+		t.Fatalf("read the substituted member back: %v", err)
+	}
+	if !bytes.Equal(got, gasmObj) {
+		t.Fatal("the repacked archive does not carry the gasm object")
+	}
+
+	// Re-link.  The line carries a GOROOT assignment and $WORK placeholders;
+	// GOEXPERIMENT must match the toolchain's own, because the linker
+	// compares the object header against its configuration.
+	goExp, _ := exec.Command(goBin, "env", "GOEXPERIMENT").Output()
+	linkLine := strings.ReplaceAll(st.linkLine, "$WORK", st.work)
+	linkLine = strings.ReplaceAll(linkLine, filepath.Join(st.work, "b001", "_pkg_.a"), newArch)
+	linkLine = strings.ReplaceAll(linkLine, filepath.Join(st.work, "b001", "exe", "a.out"), outBin)
+	env := append(os.Environ(), "GOEXPERIMENT="+strings.TrimSpace(string(goExp)))
+	env = append(env, extraEnv...)
+	link := exec.Command("sh", "-c", linkLine)
+	link.Dir = dir
+	link.Env = env
+	if out, err := link.CombinedOutput(); err != nil {
+		t.Fatalf("link with the gasm object: %v\n%s", err, out)
+	}
+}
+
 // 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
@@ -524,6 +663,9 @@ func fieldAfter(line, flag string) string {
 // 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) {
+	if testing.Short() {
+		t.Skip("end-to-end go build and link: skipped in -short mode")
+	}
 	goBin, err := exec.LookPath("go")
 	if err != nil {
 		t.Skip("no Go toolchain available")

asm/guard_test.go

@@ -9,8 +9,8 @@ import (
 	"strings"
 	"testing"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/ast"
-	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+	"sourcedock.dev/petrbalvin/gasm-sdk/ast"
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
 )
 
 // The expected bytes are pinned from `go tool asm` output (Go 1.27, amd64,
@@ -307,12 +307,12 @@ func TestStackGuardBytesLOONG64(t *testing.T) {
 func TestStackGuardGOObjInternalCall(t *testing.T) {
 	for _, tt := range []struct {
 		src      string
-		assemble func(*ast.File) (*Image, error)
+		assemble func(*ast.File, ...AssembleOption) (*Image, error)
 	}{
 		{"g_amd64.s", AssembleFile},
-		{"g_arm64.s", AssembleFileARM64},
-		{"g_riscv64.s", AssembleFileRISCV},
-		{"g_loong64.s", AssembleFileLOONG64},
+		{"g_arm64.s", func(f *ast.File, _ ...AssembleOption) (*Image, error) { return AssembleFileARM64(f) }},
+		{"g_riscv64.s", func(f *ast.File, _ ...AssembleOption) (*Image, error) { return AssembleFileRISCV(f) }},
+		{"g_loong64.s", func(f *ast.File, _ ...AssembleOption) (*Image, error) { return AssembleFileLOONG64(f) }},
 	} {
 		f, errs := parser.Parse(tt.src, "TEXT \u00b7callsmall(SB), $16-0\n\tCALL \u00b7other(SB)\n\tRET\nTEXT \u00b7other(SB), NOSPLIT, $0\n\tRET\n")
 		if len(errs) > 0 {

asm/instrs.go

@@ -90,6 +90,40 @@ var noOperandTable = map[string][]byte{
 	"LOCK":    {0xF0},
 	"REP":     {0xF3},
 	"REPN":    {0xF2},
+	"ENDBR64": {0xF3, 0x0F, 0x1E, 0xFA},
+}
+
+// sysUnaryTable maps the one-operand system instructions to their bytes:
+// the prefix, the opcode and the /digit the reg field carries.  The operand
+// is a register or memory in r/m.
+var sysUnaryTable = map[string]struct {
+	prefix byte
+	opcode []byte
+	digit  int
+}{
+	"CLWB":     {0x66, []byte{0x0F, 0xAE}, 6},
+	"TPAUSE":   {0x66, []byte{0x0F, 0xAE}, 6},
+	"UMONITOR": {0xF3, []byte{0x0F, 0xAE}, 6},
+	"UMWAIT":   {0xF2, []byte{0x0F, 0xAE}, 6},
+	"RDPID":    {0xF3, []byte{0x0F, 0xC7}, 7},
+	"CLDEMOTE": {0x00, []byte{0x0F, 0x1C}, 0},
+}
+
+// encodeSysUnary emits a one-operand system instruction: the operand in r/m
+// under the fixed /digit, no REX.W.
+func (e *enc) encodeSysUnary(mnem string, m struct {
+	prefix byte
+	opcode []byte
+	digit  int
+}, ops []Operand) error {
+	if len(ops) != 1 {
+		return fmt.Errorf("%s expects 1 operand, got %d", mnem, len(ops))
+	}
+	i := &instr{prefix: m.prefix, opcode: m.opcode, modrm: -1, sib: -1}
+	if err := setRMDigit(i, m.digit, ops[0], 8); err != nil {
+		return err
+	}
+	return e.emit(i)
 }
 
 // --- MOV --------------------------------------------------------------------
@@ -111,6 +145,76 @@ func (e *enc) encodeMov(ops []Operand, size int) error {
 	// silently emit REX.W 8B with the wrong operand meaning.
 	_, srcVec := vecReg(src)
 	dstReg, dstVec := vecReg(dst)
+
+	// Control and debug register moves: 0F 20 (CRn→r64), 0F 22 (r64→CRn),
+	// 0F 21 (DRn→r64) and 0F 23 (r64→DRn).  The CR/DR number rides the reg
+	// field, the general register r/m; CR8+/DR8+ take REX.R.
+	if c, ok := src.(Reg); ok && c.ctl != 0 {
+		g, ok := dst.(Reg)
+		if !ok || g.isVec() || g.ctl != 0 {
+			return fmt.Errorf("MOV: control/debug register load needs a general register destination")
+		}
+		opc := byte(0x20)
+		if c.ctl == 2 {
+			opc = 0x21
+		}
+		return e.emit(&instr{
+			opcode: []byte{0x0F, opc},
+			modrm:  0xC0 | (c.idx&7)<<3 | (g.idx & 7),
+			sib:    -1, rexR: c.idx >= 8, rexB: g.idx >= 8,
+		})
+	}
+	if c, ok := dst.(Reg); ok && c.ctl != 0 {
+		g, ok := src.(Reg)
+		if !ok || g.isVec() || g.ctl != 0 {
+			return fmt.Errorf("MOV: control/debug register store needs a general register source")
+		}
+		opc := byte(0x22)
+		if c.ctl == 2 {
+			opc = 0x23
+		}
+		return e.emit(&instr{
+			opcode: []byte{0x0F, opc},
+			modrm:  0xC0 | (c.idx&7)<<3 | (g.idx & 7),
+			sib:    -1, rexR: c.idx >= 8, rexB: g.idx >= 8,
+		})
+	}
+
+	// MMX register moves: MOVQ M0, mem and MOVQ mem, M0 are the MMX
+	// load/store pair 0F 6F/0F 7F (no prefix); a register pair takes the
+	// load opcode.  The XMM MOVQ forms follow below.
+	if m, ok := src.(Reg); ok && m.mmx {
+		switch d := dst.(type) {
+		case Reg:
+			if !d.mmx {
+				return fmt.Errorf("MOV: MMX register moves stay inside the M bank")
+			}
+			i := &instr{opcode: []byte{0x0F, 0x6F}, modrm: -1, sib: -1}
+			if err := setRM(i, d, src, 8); err != nil {
+				return err
+			}
+			return e.emit(i)
+		case Mem:
+			i := &instr{opcode: []byte{0x0F, 0x7F}, modrm: -1, sib: -1}
+			if err := setRM(i, m, d, 8); err != nil {
+				return err
+			}
+			return e.emit(i)
+		}
+		return fmt.Errorf("MOV: invalid MMX destination")
+	}
+	if m, ok := dst.(Reg); ok && m.mmx {
+		srcM, ok := src.(Mem)
+		if !ok {
+			return fmt.Errorf("MOV: MMX load takes a memory source")
+		}
+		i := &instr{opcode: []byte{0x0F, 0x6F}, modrm: -1, sib: -1}
+		if err := setRM(i, m, srcM, 8); err != nil {
+			return err
+		}
+		return e.emit(i)
+	}
+
 	if srcVec || dstVec {
 		if dstVec {
 			if g, ok := src.(Reg); ok && !g.isVec() {
@@ -192,6 +296,30 @@ func (e *enc) encodeMov(ops []Operand, size int) error {
 		}
 		return e.emit(i)
 
+	case TLSMem:
+		if !dstIsReg {
+			return fmt.Errorf("MOV: two memory operands")
+		}
+		// MOV r, off(TLS): the segment-prefixed absolute load, reg=dst,
+		// rm=src(tlsMem) through the SIB escape; the disp32 is the TLS slot
+		// offset with its R_TLSLE patch site.
+		i := newInstr(size, []byte{movRR(size)})
+		if err := setRM(i, dstReg, src, size); err != nil {
+			return err
+		}
+		return e.emit(i)
+
+	case SegAbs:
+		if !dstIsReg {
+			return fmt.Errorf("MOV: two memory operands")
+		}
+		// MOV r, 0x30(GS): the segment-absolute load.
+		i := newInstr(size, []byte{movRR(size)})
+		if err := setRM(i, dstReg, src, size); err != nil {
+			return err
+		}
+		return e.emit(i)
+
 	case Imm:
 		if dstIsReg {
 			v := int64(src)
@@ -232,11 +360,24 @@ func (e *enc) encodeMov(ops []Operand, size int) error {
 			i.imm = imm
 			return e.emit(i)
 		}
-		// MOV r/m, imm: 0xC6 (8-bit) / 0xC7 /0.
+		// MOV r/m, imm: 0xC6 (8-bit) / 0xC7 /0.  An immediate in the
+		// destination slot is the absolute-address crash-store spelling,
+		// MOVL $0xf1, 0xf1: the parser reads the trailing bare constant
+		// as an immediate, and the store's disp32 carries the address.
 		op := byte(0xC7)
 		if size == 1 {
 			op = 0xC6
 		}
+		if d, ok := dst.(Imm); ok {
+			i := newInstr(size, []byte{op})
+			setSegAbs(i, 0, SegAbs{Disp: int64(d)})
+			immBytes, err := immediate(int64(src), size, false)
+			if err != nil {
+				return err
+			}
+			i.imm = immBytes
+			return e.emit(i)
+		}
 		i := newInstr(size, []byte{op})
 		if err := setRMDigit(i, 0, dst, size); err != nil {
 			return err
@@ -481,6 +622,14 @@ func (e *enc) encodeLea(ops []Operand, size int) error {
 	default:
 		return fmt.Errorf("LEA: source must be a memory operand")
 	}
+	// LEA accepts the full unsigned 32-bit displacement span where the
+	// loads and stores reject it beyond the signed one; the wide values
+	// ride the same disp32 bytes as their two's-complement bit pattern.
+	if m, ok := src.(Mem); ok && m.Disp >= 1<<31 && m.Disp <= (1<<32)-1 {
+		c := m
+		c.Disp = int64(int32(uint32(m.Disp)))
+		src = c
+	}
 	i := newInstr(size, []byte{0x8D})
 	if err := setRM(i, dstReg, src, size); err != nil {
 		return err
@@ -626,8 +775,29 @@ func (e *enc) encodeDoubleShift(base string, ops []Operand, size int) error {
 
 func (e *enc) encodeImul(ops []Operand, size int) error {
 	switch len(ops) {
+	case 1:
+		// The one-operand form, IMUL r/m: F6/F7 /5 with AL/AX/EAX/RAX as the
+		// implied destination (the toolchain's one-register shape).
+		opc := byte(0xF7)
+		if size == 1 {
+			opc = 0xF6
+		}
+		i := newInstr(size, []byte{opc})
+		if err := setRMDigit(i, 5, ops[0], size); err != nil {
+			return err
+		}
+		return e.emit(i)
 	case 2:
-		// IMUL r, r/m: 0x0F 0xAF.
+		// Two shapes. The leading-immediate spelling IMUL $imm, r multiplies
+		// r in place (dst = rm = r): the shape GOROOT's clock code writes.
+		// Otherwise IMUL r, r/m: 0x0F 0xAF.
+		if imm, ok := ops[0].(Imm); ok {
+			dstReg, isReg := ops[1].(Reg)
+			if !isReg {
+				return fmt.Errorf("IMUL: destination must be a register")
+			}
+			return e.encodeImulImm(imm, dstReg, dstReg, size)
+		}
 		dstReg, ok := ops[1].(Reg)
 		if !ok {
 			return fmt.Errorf("IMUL: destination must be a register")
@@ -647,17 +817,26 @@ func (e *enc) encodeImul(ops []Operand, size int) error {
 		if !ok {
 			return fmt.Errorf("IMUL: immediate operand expected first")
 		}
-		// Plan 9 order: IMUL $imm, src, dst.
+		// Plan 9 order: IMUL $imm, src, dst; the source stays a general
+		// r/m operand (setRM takes registers and memory alike).
+		return e.encodeImulImm(imm, ops[1], dstReg, size)
+	}
+	return fmt.Errorf("IMUL expects 1, 2 or 3 operands, got %d", len(ops))
+}
+
+// encodeImulImm emits the immediate multiply: 0x6B with a sign-extended imm8
+// when the value fits, 0x69 with a 32-bit immediate otherwise.
+func (e *enc) encodeImulImm(imm Imm, rm Operand, dst Reg, size int) error {
 	if fits8(int64(imm)) {
 		i := newInstr(size, []byte{0x6B})
-			if err := setRM(i, dstReg, ops[1], size); err != nil {
+		if err := setRM(i, dst, rm, size); err != nil {
 			return err
 		}
 		i.imm = []byte{byte(int8(imm))}
 		return e.emit(i)
 	}
 	i := newInstr(size, []byte{0x69})
-		if err := setRM(i, dstReg, ops[1], size); err != nil {
+	if err := setRM(i, dst, rm, size); err != nil {
 		return err
 	}
 	immBytes, err := immediate(int64(imm), size, false)
@@ -666,8 +845,6 @@ func (e *enc) encodeImul(ops []Operand, size int) error {
 	}
 	i.imm = immBytes
 	return e.emit(i)
-	}
-	return fmt.Errorf("IMUL expects 2 or 3 operands, got %d", len(ops))
 }
 
 // --- PUSH / POP -------------------------------------------------------------
@@ -689,6 +866,27 @@ func (e *enc) encodePushPop(ops []Operand, size int, push bool) error {
 	w16 := size == 2
 	switch op := ops[0].(type) {
 	case Reg:
+		// Segment registers: FS and GS carry their own one-byte opcodes
+		// under 0F (A0/A8 push, A1/A9 pop); the other four spellings are
+		// not pushable in 64-bit mode.
+		if n, isSeg := op.segNumber(); isSeg {
+			switch n {
+			case 4: // FS
+				if push {
+					return e.emit(&instr{opcode: []byte{0x0F, 0xA0}, modrm: -1, sib: -1})
+				}
+				return e.emit(&instr{opcode: []byte{0x0F, 0xA1}, modrm: -1, sib: -1})
+			case 5: // GS
+				if push {
+					return e.emit(&instr{opcode: []byte{0x0F, 0xA8}, modrm: -1, sib: -1})
+				}
+				return e.emit(&instr{opcode: []byte{0x0F, 0xA9}, modrm: -1, sib: -1})
+			}
+			return fmt.Errorf("PUSH/POP: only FS and GS are encodable in 64-bit mode")
+		}
+		if op.mmx || op.isVec() || op.fp || op.ctl != 0 {
+			return fmt.Errorf("PUSH/POP: invalid register operand")
+		}
 		base := byte(0x50) // PUSH r; POP is 0x58
 		if !push {
 			base = 0x58
@@ -1040,6 +1238,38 @@ var sseMoveTable = map[string]sseMove{
 	"MOVSS":  {0xF3, 0x10, 0x11}, // scalar single
 }
 
+// sseStoreOnly holds the store-only SSE forms, OP xmm, mem: the XMM register
+// rides the reg field and memory r/m (the non-temporal store).
+var sseStoreOnly = map[string]struct {
+	prefix byte
+	op     byte
+}{
+	"MOVNTDQ": {0x66, 0xE7},
+}
+
+// encodeSSEStoreOnly encodes OP xmm, mem (reg = the XMM source, r/m = the
+// destination memory).
+func (e *enc) encodeSSEStoreOnly(mnem string, m struct {
+	prefix byte
+	op     byte
+}, ops []Operand) error {
+	if len(ops) != 2 {
+		return fmt.Errorf("%s expects 2 operands, got %d", mnem, len(ops))
+	}
+	srcReg, ok := ops[0].(Reg)
+	if !ok || !srcReg.isVec() {
+		return fmt.Errorf("%s source must be a vector register", mnem)
+	}
+	if !isX86Mem(ops[1]) {
+		return fmt.Errorf("%s destination must be a memory operand", mnem)
+	}
+	i := &instr{prefix: m.prefix, opcode: []byte{0x0F, m.op}, modrm: -1, sib: -1}
+	if err := setRM(i, srcReg, ops[1], 8); err != nil {
+		return err
+	}
+	return e.emit(i)
+}
+
 // encodeSSEMove encodes a legacy SSE move: a vector-to-vector move uses the
 // load form (reg = destination), matching the Go assembler.
 func (e *enc) encodeSSEMove(m sseMove, ops []Operand) error {
@@ -1255,14 +1485,23 @@ func (e *enc) encodeSSEBin(m sseBin, ops []Operand) error {
 	}
 	src, dst := ops[0], ops[1]
 	dstReg, ok := dst.(Reg)
-	if !ok || !dstReg.isVec() {
+	if !ok || (!dstReg.isVec() && !dstReg.mmx) {
 		return fmt.Errorf("SSE binary destination must be a vector register")
 	}
+	// The MMX twins of the packed-integer SSE2 ops drop the 0x66 prefix:
+	// PADDD M2, M1 is 0F FE where the XMM form is 66 0F FE.
+	prefix := m.prefix
+	if dstReg.mmx {
+		if prefix != 0x66 {
+			return fmt.Errorf("SSE binary: this form takes no MMX register operand")
+		}
+		prefix = 0
+	}
 	opcode := []byte{0x0F, m.op}
 	if m.map38 {
 		opcode = []byte{0x0F, 0x38, m.op}
 	}
-	i := &instr{prefix: m.prefix, opcode: opcode, modrm: -1, sib: -1}
+	i := &instr{prefix: prefix, opcode: opcode, modrm: -1, sib: -1}
 	if err := setRM(i, dstReg, src, 8); err != nil {
 		return err
 	}
@@ -1738,12 +1977,19 @@ func (e *enc) encodeSSEShift(name string, ops []Operand) error {
 // immediate LAST in Plan 9 order (src, dst, $imm), unlike the shuffle family:
 // F2 0F C2 with reg = dst, rm = src.
 func (e *enc) encodeCmpsd(ops []Operand) error {
+	return e.encodeSSECmp("CMPSD", 0xF2, ops)
+}
+
+// encodeSSECmp encodes the SSE compare family (CMPSD/CMPSS/CMPPS/CMPPD):
+// 0F C2 /r ib with the predicate immediate last in Plan 9 order
+// (src, dst, $imm) and the packed forms' prefixes.
+func (e *enc) encodeSSECmp(mnem string, prefix byte, ops []Operand) error {
 	if len(ops) != 3 {
-		return fmt.Errorf("CMPSD expects 3 operands (src, dst, $imm), got %d", len(ops))
+		return fmt.Errorf("%s expects 3 operands (src, dst, $imm), got %d", mnem, len(ops))
 	}
 	imm, ok := ops[2].(Imm)
 	if !ok {
-		return fmt.Errorf("CMPSD predicate must be an immediate")
+		return fmt.Errorf("%s predicate must be an immediate", mnem)
 	}
 	immByte, err := imm8(int64(imm))
 	if err != nil {
@@ -1751,9 +1997,9 @@ func (e *enc) encodeCmpsd(ops []Operand) error {
 	}
 	dstReg, ok2 := ops[1].(Reg)
 	if !ok2 || !dstReg.isVec() {
-		return fmt.Errorf("CMPSD destination must be a vector register")
+		return fmt.Errorf("%s destination must be a vector register", mnem)
 	}
-	i := &instr{prefix: 0xF2, opcode: []byte{0x0F, 0xC2}, modrm: -1, sib: -1}
+	i := &instr{prefix: prefix, opcode: []byte{0x0F, 0xC2}, modrm: -1, sib: -1}
 	if err := setRM(i, dstReg, ops[0], 8); err != nil {
 		return err
 	}

asm/integration_test.go

@@ -16,7 +16,7 @@ import (
 
 	"golang.org/x/arch/x86/x86asm"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
 )
 
 // TestAssembleGoFlacAVX2Kernel assembles the whole production AVX2 kernel;
@@ -25,7 +25,7 @@ import (
 func TestAssembleGoFlacAVX2Kernel(t *testing.T) {
 	path := "../../go-libraries/go-flac/avx2_amd64.s"
 	if _, err := os.Stat(path); err != nil {
-		t.Skip("go-libraries repository not present next to gasm-devkit")
+		t.Skip("go-libraries repository not present next to gasm-sdk")
 	}
 	src, err := os.ReadFile(path)
 	if err != nil {
@@ -86,7 +86,7 @@ func TestAssembleGoFlacAVX2Kernel(t *testing.T) {
 func TestAssembleGoFlacAVX512Kernel(t *testing.T) {
 	path := "../../go-libraries/go-flac/avx512_amd64.s"
 	if _, err := os.Stat(path); err != nil {
-		t.Skip("go-libraries repository not present next to gasm-devkit")
+		t.Skip("go-libraries repository not present next to gasm-sdk")
 	}
 	src, err := os.ReadFile(path)
 	if err != nil {

asm/kernels_differential_test.go

@@ -13,7 +13,7 @@ import (
 	"strings"
 	"testing"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
 )
 
 // The differential kernels for the DATA-path and front-end gaps are kept in
@@ -23,9 +23,18 @@ import (
 // bytes must agree with the relocation sites masked on both sides.
 
 // toolAsmObject assembles path with the installed toolchain's assembler for
-// goarch ("" = the host) and returns the object bytes.
+// goarch ("" = the host) and returns the object bytes.  Every live-oracle
+// comparison funnels through here, so this is also where the deliberate-run
+// boundary sits: under -short (the push pipeline's mode) the comparisons
+// skip, because each spawns a go tool asm subprocess and the small single-
+// core runner pays seconds per spawn.  The encodings stay pinned by the
+// golden-byte tests in every mode; the live oracle runs in the local test
+// gate and the dispatched workflows.
 func toolAsmObject(t *testing.T, path, goarch string) []byte {
 	t.Helper()
+	if testing.Short() {
+		t.Skip("live go tool asm oracle: skipped in -short mode")
+	}
 	goBin, err := exec.LookPath("go")
 	if err != nil {
 		t.Skip("no Go toolchain available")
@@ -130,6 +139,7 @@ func TestDifferentialKernels(t *testing.T) {
 		{filepath.Join("..", "testdata", "verify", "quadreg_amd64.s"), "", false},
 		{filepath.Join("..", "testdata", "verify", "floatimm_amd64.s"), "", false},
 		{filepath.Join("..", "testdata", "verify", "bookkeep_amd64.s"), "", false},
+		{filepath.Join("..", "testdata", "verify", "forms_amd64.s"), "", false},
 		{filepath.Join("..", "testdata", "verify", "datarel_arm64.s"), "arm64", true},
 		{filepath.Join("..", "testdata", "verify", "divslash_arm64.s"), "arm64", true},
 	} {

asm/l64_goobj_test.go

@@ -12,7 +12,7 @@ import (
 	"strings"
 	"testing"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
 )
 
 // TestGOObjectLOONG64Structure checks the emitted loong64 object's blocks:

asm/link.go

@@ -9,7 +9,7 @@ import (
 	"sort"
 	"strconv"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/ast"
+	"sourcedock.dev/petrbalvin/gasm-sdk/ast"
 )
 
 // Image is an assembled file: the function bodies laid out in source order,
@@ -134,7 +134,8 @@ type DataSymbol struct {
 	Offset int    // byte offset within Data
 	Size   int
 	Static bool // the <> marker: file-local, not exported
-	Rodata bool // the RODATA flag: read-only data
+	Rodata bool // the RODATA flag: read-only data (implies no pointers)
+	Noptr  bool // the NOPTR flag: data with no pointers, kept out of GC scanning
 	Dupok  bool // the DUPOK flag: duplicate-OK
 	// Relocs carries the symbol-valued DATA initialisers ("DATA s+0(SB)/8,
 	// $other(SB)"): fields of this symbol's data that hold another symbol's
@@ -150,6 +151,18 @@ func (img *Image) Bytes() []byte {
 	return append(out, img.Data...)
 }
 
+// AssembleOption adjusts the file-level assembly context.
+type AssembleOption func(*linkInfo)
+
+// WithGOOS selects the target operating system for the forms that depend on
+// it, the TLS access shape above all: linux and freebsd take the
+// one-instruction form, windows and plan9 keep the two-instruction load.
+func WithGOOS(goos string) AssembleOption {
+	return func(l *linkInfo) {
+		l.goos = goos
+	}
+}
+
 // AssembleFile assembles every TEXT function of a parsed file and lays out
 // its static symbols (GLOBL/DATA) in a data section behind the code.  Each
 // reference to a file-local static symbol becomes a RIP-relative load whose
@@ -157,7 +170,7 @@ func (img *Image) Bytes() []byte {
 // GLOBL defines is recorded as an external relocation (Externals) with its
 // displacement left zero, the object-file emitters resolve it at link
 // time, while the raw image (Bytes) cannot represent it.
-func AssembleFile(f *ast.File) (*Image, error) {
+func AssembleFile(f *ast.File, opts ...AssembleOption) (*Image, error) {
 	dataSyms, err := collectData(f)
 	if err != nil {
 		return nil, err
@@ -166,7 +179,17 @@ func AssembleFile(f *ast.File) (*Image, error) {
 	for _, d := range dataSyms {
 		known[d.name] = true
 	}
+	// TEXT symbols are file-level definitions too: a symbol immediate
+	// ($fn(SB)) may name one, exactly as a data reference names a GLOBL.
+	for _, d := range f.Decls {
+		if t, ok := d.(*ast.Text); ok {
+			known[t.Name.Name] = true
+		}
+	}
 	link := &linkInfo{symbols: known, allowExternal: true}
+	for _, o := range opts {
+		o(link)
+	}
 	poolSeen := map[string]bool{}
 
 	img := &Image{Symbols: map[string]int{}, SourcePath: f.Path}
@@ -247,6 +270,7 @@ func AssembleFile(f *ast.File) (*Image, error) {
 			Size:   len(d.buf),
 			Static: d.static,
 			Rodata: d.rodata,
+			Noptr:  d.noptr,
 			Dupok:  d.dupok,
 		})
 		img.Data = append(img.Data, d.buf...)
@@ -391,6 +415,7 @@ func AssembleFileRISCV(f *ast.File) (*Image, error) {
 			Size:   d.size,
 			Static: d.static,
 			Rodata: d.rodata,
+			Noptr:  d.noptr,
 			Dupok:  d.dupok,
 		})
 	}
@@ -463,6 +488,7 @@ func AssembleFileLOONG64(f *ast.File) (*Image, error) {
 			Size:   d.size,
 			Static: d.static,
 			Rodata: d.rodata,
+			Noptr:  d.noptr,
 			Dupok:  d.dupok,
 		})
 	}
@@ -524,6 +550,7 @@ type dataSym struct {
 	size   int
 	static bool
 	rodata bool
+	noptr  bool
 	dupok  bool
 	// relocs are the symbol-valued DATA fields, in declaration order; Off
 	// is relative to the symbol's data start.
@@ -565,12 +592,14 @@ func collectData(f *ast.File) ([]dataSym, error) {
 			switch f {
 			case "RODATA":
 				ds.rodata = true
+			case "NOPTR":
+				ds.noptr = true
 			case "DUPOK":
 				ds.dupok = true
 			default:
 				// Legacy numeric flag constants (runtime/textflag.h):
-				// DUPOK is 2, RODATA is 8; combinations arrive as one
-				// number (e.g. 10 = RODATA|DUPOK).
+				// DUPOK is 2, RODATA is 8, NOPTR is 16; combinations arrive
+				// as one number (e.g. 10 = RODATA|DUPOK).
 				if n, err := strconv.Atoi(f); err == nil {
 					if n&2 != 0 {
 						ds.dupok = true
@@ -578,6 +607,9 @@ func collectData(f *ast.File) ([]dataSym, error) {
 					if n&8 != 0 {
 						ds.rodata = true
 					}
+					if n&16 != 0 {
+						ds.noptr = true
+					}
 				}
 			}
 		}

asm/link_test.go

@@ -12,7 +12,7 @@ import (
 	"strings"
 	"testing"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
 )
 
 // TestAssembleFileStaticData checks the whole-image layout; code, padding
@@ -60,23 +60,15 @@ DATA small<>+0(SB)/4, $0x1234
 	}
 }
 
-// TestAssembleFileErrors checks the static-symbol error paths.
+// TestAssembleFileErrors checks the static-symbol error paths.  A reference
+// to a static symbol no GLOBL defines defers to the linker exactly as the
+// toolchain does (an external relocation), so it is not an error here.
 func TestAssembleFileErrors(t *testing.T) {
 	cases := []struct {
 		name string
 		src  string
 		want string // substring of the error
 	}{
-		{
-			"undefined symbol",
-			`
-#include "textflag.h"
-TEXT ·f(SB), NOSPLIT, $0
-	VMOVDQU nope<>(SB), X0
-	RET
-`,
-			"undefined symbol",
-		},
 		{
 			"DATA without GLOBL",
 			`
@@ -369,23 +361,8 @@ func main() {
 	if err != nil {
 		t.Fatalf("baseline build: %v\n%s", err, buildLog)
 	}
-	var work, linkLine, asmObj string
-	for line := range strings.SplitSeq(string(buildLog), "\n") {
-		switch {
-		case strings.HasPrefix(line, "WORK="):
-			work = strings.TrimPrefix(line, "WORK=")
-		case strings.Contains(line, "/asm ") && strings.Contains(line, "main_amd64.s") && !strings.Contains(line, "-gensymabis"):
-			asmObj = fieldAfter(line, "-o")
-		case strings.Contains(line, "/link ") && strings.Contains(line, "-importcfg"):
-			linkLine = line
-		}
-	}
-	if work == "" || asmObj == "" || linkLine == "" {
-		t.Skipf("could not parse build log (work=%q asmObj=%q link=%q)", work, asmObj, linkLine)
-	}
-	defer os.RemoveAll(work)
-	asmObj = strings.ReplaceAll(asmObj, "$WORK", work)
-	linkLine = strings.ReplaceAll(linkLine, "$WORK", work)
+	st := parseBuildLog(t, buildLog, "main_amd64.s")
+	defer os.RemoveAll(st.work)
 
 	// Assemble the same source with gasm and substitute the object.
 	src, err := os.ReadFile(filepath.Join(dir, "main_amd64.s"))
@@ -400,21 +377,18 @@ func main() {
 	if err != nil {
 		t.Fatalf("AssembleFile: %v", err)
 	}
-	gasmObj, err := img.GOObject("dlink", "main_amd64.s")
+	// The package path is "main": the linker resolves the Go code's
+	// references against main.<name>, so the object must define the symbols
+	// under that prefix whatever the module is called.
+	gasmObj, err := img.GOObject("main", "main_amd64.s")
 	if err != nil {
 		t.Fatalf("GOObject: %v", err)
 	}
-	if err := os.WriteFile(asmObj, gasmObj, 0o644); err != nil {
-		t.Fatalf("write gasm object: %v", err)
-	}
-	linkCmd := exec.Command("bash", "-c", "cd "+dir+" && "+linkLine)
-	if out, err := linkCmd.CombinedOutput(); err != nil {
-		t.Fatalf("re-link with gasm object: %v\n%s", err, out)
-	}
+	substituteAndRelink(t, goBin, dir, st, filepath.Join(dir, "prog2"), gasmObj)
 
 	// The linked program must run and find the right function behind the
 	// data word.
-	out, err := exec.Command(filepath.Join(dir, "prog")).CombinedOutput()
+	out, err := exec.Command(filepath.Join(dir, "prog2")).CombinedOutput()
 	if err != nil {
 		t.Fatalf("linked program failed: %v\n%s", err, out)
 	}

asm/loong64_assemble.go

@@ -9,7 +9,7 @@ import (
 	"strconv"
 	"strings"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/ast"
+	"sourcedock.dev/petrbalvin/gasm-sdk/ast"
 )
 
 // assembleLOONG64 assembles a LoongArch (loong64) TEXT function body into
@@ -371,6 +371,13 @@ func loong64InstrSize(instr *ast.Instr, fi loong64FrameInfo) int {
 	if mnem == "RET" {
 		return len(loong64Return(fi))
 	}
+	// BYTE lays down one raw byte per operand, a front-end pseudo-op the
+	// toolchain spells only on x86 but accepts here the same way the arm64
+	// and riscv64 encoders do (a superset spelling, shippable via the goobj
+	// path).
+	if mnem == "BYTE" {
+		return len(ops)
+	}
 	switch mnem {
 	case "END", "FUNCDATA", "PCDATA":
 		return 0 // bookkeeping statements contribute no bytes
@@ -484,6 +491,18 @@ func encodeLOONG64Instr(instr *ast.Instr, pc int, offsets map[string]int, fi loo
 			return nil, fmt.Errorf("WORD expects 1 operand, got %d", len(ops))
 		}
 		return l64wordLE(uint32(immFromOperand(ops[0]))), nil
+	case "BYTE":
+		// BYTE $b lays down one raw byte per operand, the same front-end
+		// pseudo-op the arm64 and riscv64 encoders accept.
+		var out []byte
+		for _, op := range ops {
+			b := l64Imm64(op)
+			if b < 0 || b > 0xFF {
+				return nil, fmt.Errorf("BYTE: immediate %d does not fit a byte", b)
+			}
+			out = append(out, byte(b))
+		}
+		return out, nil
 	case "END", "FUNCDATA", "PCDATA", "GETCALLERPC":
 		// The assembler's bookkeeping statements.  END, FUNCDATA and PCDATA
 		// contribute no bytes, the same shapes GOARCH=loong64 go tool asm
@@ -701,6 +720,35 @@ func encodeLOONG64Instr(instr *ast.Instr, pc int, offsets map[string]int, fi loo
 		}
 		return l64wordLE(l64rr(enc.op, rj, rd)), nil
 
+	case l64Fllsc:
+		// LLACQ{W,V} (Rj), Rd loads and SCREL{W,V} Rd, (Rj) stores, both
+		// 2R encodings op | rj<<5 | rd against a zero-offset memory operand
+		// (the toolchain's C_ZOREG, which rejects any displacement).
+		rd, rj, off, _, err := l64MemOperands(ops, fi)
+		if err != nil {
+			return nil, fmt.Errorf("%s: %w", mnem, err)
+		}
+		if off != 0 {
+			return nil, fmt.Errorf("%s: only a zero-offset memory operand is allowed", mnem)
+		}
+		return l64wordLE(l64rr(enc.op, rj, rd)), nil
+
+	case l64Fscq:
+		// SCQ first, middle, (base): op | middle<<10 | base<<5 | first,
+		// against a zero-offset memory operand as with the LL/SC pair.
+		if len(ops) != 3 || !isMemOperand(ops[2]) || isMemOperand(ops[0]) || isMemOperand(ops[1]) {
+			return nil, fmt.Errorf("%s expects reg, reg, (reg)", mnem)
+		}
+		first, middle := l64Reg(ops[0]), l64Reg(ops[1])
+		rj, off := l64MemWithFrame(ops[2], fi)
+		if first < 0 || middle < 0 || rj < 0 {
+			return nil, fmt.Errorf("%s: invalid register operand", mnem)
+		}
+		if off != 0 {
+			return nil, fmt.Errorf("%s: only a zero-offset memory operand is allowed", mnem)
+		}
+		return l64wordLE(l64rrr(enc.op, middle, rj, first)), nil
+
 	case l64Firr:
 		// LU52ID: INSTR $imm, rd  or  INSTR $imm, rj, rd.
 		if len(ops) < 2 || !isImmOperand(ops[0]) {
@@ -1229,6 +1277,31 @@ func l64MemOperands(ops []*ast.Operand, fi loong64FrameInfo) (rd, rj int, off in
 	return rd, rj, off, load, nil
 }
 
+// l64ImmMem reads the `$off(rj)` immediate form off an operand's raw text:
+// the shared immediate parse reduces it to the bare number and keeps only
+// the text as a witness of the base register.  ok reports the form was
+// found, with the base's register number (or -1 when the name is not a
+// general register).
+func l64ImmMem(op *ast.Operand) (off int32, base int, ok bool) {
+	if op.Kind != ast.OpImmediate || !op.Imm.HasVal {
+		return 0, 0, false
+	}
+	raw := strings.ReplaceAll(op.Raw, " ", "")
+	if !strings.HasPrefix(raw, "$") || !strings.HasSuffix(raw, ")") {
+		return 0, 0, false
+	}
+	open := strings.LastIndexByte(raw, '(')
+	if open < 2 {
+		return 0, 0, false
+	}
+	base = loong64RegNum(raw[open+1 : len(raw)-1])
+	v := op.Imm.Val
+	if op.Imm.Neg {
+		v = -v
+	}
+	return int32(v), base, base >= 0
+}
+
 // ---- the MOV pseudo-instruction ----
 
 // encodeLOONG64Mov encodes the MOV family, the load/store/immediate
@@ -1262,6 +1335,29 @@ func encodeLOONG64Mov(instr *ast.Instr, mnem string, fi loong64FrameInfo, relocs
 			}
 			return encodeLOONG64SBAddr(src.Imm.Sym, rd, relocs), nil
 		}
+		// MOVx $off(rj), rd computes an address: the toolchain's `mov
+		// $soreg, r` case, a plain addi.d whatever the move's width (both
+		// MOVW and MOVV $4(R4), R5 encode the same addi.d in its testdata).
+		// A wider offset materialises in R30 first (lu12i.w + ori + add.d,
+		// its case 10).  The immediate's Raw carries the base register,
+		// which the shared immediate parse reduces to the bare number.
+		if off, base, ok := l64ImmMem(src); ok {
+			rd := l64Reg(dst)
+			if rd < 0 {
+				return nil, fmt.Errorf("%s $imm(rj): invalid destination register", mnem)
+			}
+			if loong64RegClass(operandRegName(dst)) == l64ClsFP {
+				return nil, fmt.Errorf("%s $imm(rj): illegal combination with an F register destination", mnem)
+			}
+			if off >= -2048 && off <= 2047 {
+				return l64wordLE(l64irr(l64DualTable["ADDV"].imm, int(off), base, rd)), nil
+			}
+			return l64WordsLE(
+				l64ir(l64InstrTable["LU12IW"].op, int(off)>>12, 30),
+				l64irr(l64DualTable["OR"].imm, int(off)&0xFFF, 30, 30),
+				l64rrr(l64DualTable["ADDV"].rrr, 30, base, rd),
+			), nil
+		}
 		rd := l64Reg(dst)
 		if rd < 0 {
 			return nil, fmt.Errorf("%s $imm: invalid destination register", mnem)
@@ -1356,6 +1452,14 @@ func loong64MovSize(mnem string, ops []*ast.Operand, fi loong64FrameInfo) int {
 		if src.Imm.Sym != nil && src.Imm.Sym.Pseudo == "SB" {
 			return 8 // pcalau12i + addi.d
 		}
+		// The $off(rj) address immediate: addi.d in the 12-bit window,
+		// lu12i.w + ori + add.d beyond it (the toolchain's case 10).
+		if off, _, ok := l64ImmMem(src); ok {
+			if off >= -2048 && off <= 2047 {
+				return 4
+			}
+			return 12
+		}
 		if loong64RegClass(operandRegName(dst)) == l64ClsFP {
 			return 8 // ori/addi.w r30 + movgr2fr.w (an encode-time diagnostic when invalid)
 		}
@@ -1868,6 +1972,19 @@ func l64MemWithFrame(op *ast.Operand, fi loong64FrameInfo) (rj int, off int32) {
 	return l64Mem(op)
 }
 
+// l64VmovqMem resolves a VMOVQ/XVMOVQ memory operand.  The toolchain's
+// vector table falls back to the zero register as the FP-relative base
+// (`VMOVQ V2, y+16(FP)` stores through R0 while MOVW reads the same operand
+// through R3), so the vector moves keep the resolved offset but the zero
+// base, exactly as `go tool asm` emits them.
+func l64VmovqMem(op *ast.Operand, fi loong64FrameInfo) (rj int, off int32) {
+	rj, off = l64MemWithFrame(op, fi)
+	if op.Addr.Sym != nil && op.Addr.Sym.Pseudo == "FP" {
+		rj = 0
+	}
+	return rj, off
+}
+
 // l64MemOffset returns the resolved byte offset of a memory operand.
 func l64MemOffset(op *ast.Operand, fi loong64FrameInfo) int32 {
 	_, off := l64MemWithFrame(op, fi)
@@ -1892,7 +2009,7 @@ func l64Label(op *ast.Operand) string {
 type l64VecOperand struct {
 	num    int  // 5-bit register number
 	lasx   bool // X bank (LASX) rather than V (LSX)
-	width  byte // suffix width letter (B/H/W/V), 0 on a bare register
+	width  byte // suffix width letter (B/H/W/V/Q), 0 on a bare register
 	lanes  int  // lane count of a width suffix (B16 → 16)
 	elem   int  // element index of a .T[i] suffix
 	hasEl  bool // the suffix names an element (.T[i])
@@ -1932,7 +2049,7 @@ func l64ParseVecOperand(op *ast.Operand) (v l64VecOperand, ok bool) {
 	}
 	i++
 	w := name[i]
-	if w != 'B' && w != 'H' && w != 'W' && w != 'V' {
+	if w != 'B' && w != 'H' && w != 'W' && w != 'V' && w != 'Q' {
 		return v, false
 	}
 	v.width, v.hasSuf = w, true
@@ -2174,6 +2291,10 @@ func encodeLOONG64Vector(instr *ast.Instr, mnem string, fi loong64FrameInfo) ([]
 //	VMOVQ rj, vd.T        vreplgr2vr (duplicate a general register)
 //	VMOVQ vj.T[i], rd     vpickve2gr (extract one element)
 //	VMOVQ rj, vd.T[i]     vinsgr2vr  (insert one element)
+//	VMOVQ vj.T[i], vd.T   vreplvei   (broadcast one element, LSX)
+//	XVMOVQ xj, xd.T       xvreplve0  (broadcast element zero, LASX)
+//	XVMOVQ xj, xd.T[i]    xvinsve0   (insert element zero, LASX)
+//	XVMOVQ xj.T[i], xd    xvpickve   (extract one element, LASX)
 func encodeLOONG64Vmovq(lasx bool, ops []*ast.Operand, fi loong64FrameInfo) ([]byte, error) {
 	enc := l64VmovqTable[lasx]
 	bank := "V"
@@ -2200,6 +2321,118 @@ func encodeLOONG64Vmovq(lasx bool, ops []*ast.Operand, fi loong64FrameInfo) ([]b
 		return loong64RegNum(name), nil
 	}
 
+	// Element broadcast: VMOVQ vj.T[i], vd.T (vreplvei.{b,h,w,d}), the
+	// source element width matching the destination arrangement.  An LSX-only
+	// form: the toolchain's table gives vreplvei no LASX counterpart.
+	if srcVec && dstVec && src.hasEl && dst.hasSuf && !dst.hasEl {
+		if lasx || src.lasx || dst.lasx {
+			return nil, fmt.Errorf("VMOVQ: vreplvei has no %s-bank form", bank)
+		}
+		if src.unsig {
+			return nil, fmt.Errorf("VMOVQ: vreplvei takes no unsigned element suffix")
+		}
+		if src.width != dst.width {
+			return nil, fmt.Errorf("VMOVQ: element width does not match arrangement %q", ops[1].Raw)
+		}
+		if _, ok := l64VecSuffixWidth(false, dst); !ok {
+			return nil, fmt.Errorf("VMOVQ: invalid arrangement %q", ops[1].Raw)
+		}
+		var op uint32
+		limit := 0
+		switch src.width {
+		case 'B':
+			op, limit = enc.rveiB, 15
+		case 'H':
+			op, limit = enc.rveiH, 7
+		case 'W':
+			op, limit = enc.rveiW, 3
+		default:
+			op, limit = enc.rveiD, 1
+		}
+		if src.elem > limit {
+			return nil, fmt.Errorf("VMOVQ: element index %d out of range [0, %d]", src.elem, limit)
+		}
+		return l64wordLE(op | uint32(src.elem)<<10 | uint32(src.num)<<5 | uint32(dst.num)), nil
+	}
+
+	// Broadcast of element zero: XVMOVQ xj, xd.T (xvreplve0.{b,h,w,d,q}),
+	// a bare X source into an arranged X destination.  LASX only.
+	if srcVec && dstVec && !src.hasSuf && dst.hasSuf && !dst.hasEl {
+		if !lasx || src.lasx != lasx || dst.lasx != lasx {
+			return nil, fmt.Errorf("XVMOVQ: xvreplve0 is the %s-bank form alone", bank)
+		}
+		var op uint32
+		switch dst.width {
+		case 'B':
+			if dst.lanes != 32 {
+				return nil, fmt.Errorf("XVMOVQ: invalid arrangement %q", ops[1].Raw)
+			}
+			op = enc.rve0B
+		case 'H':
+			if dst.lanes != 16 {
+				return nil, fmt.Errorf("XVMOVQ: invalid arrangement %q", ops[1].Raw)
+			}
+			op = enc.rve0H
+		case 'W':
+			if dst.lanes != 8 {
+				return nil, fmt.Errorf("XVMOVQ: invalid arrangement %q", ops[1].Raw)
+			}
+			op = enc.rve0W
+		case 'V':
+			if dst.lanes != 4 {
+				return nil, fmt.Errorf("XVMOVQ: invalid arrangement %q", ops[1].Raw)
+			}
+			op = enc.rve0D
+		case 'Q':
+			if dst.lanes != 2 {
+				return nil, fmt.Errorf("XVMOVQ: invalid arrangement %q", ops[1].Raw)
+			}
+			op = enc.rve0Q
+		default:
+			return nil, fmt.Errorf("XVMOVQ: invalid arrangement %q", ops[1].Raw)
+		}
+		return l64wordLE(op | uint32(src.num)<<5 | uint32(dst.num)), nil
+	}
+
+	// Insert of element zero: XVMOVQ xj, xd.T[i] (xvinsve0.{w,d}), a bare X
+	// source into one word or double-word lane.  LASX only.
+	if srcVec && dstVec && !src.hasSuf && dst.hasEl {
+		if !lasx || src.lasx != lasx || dst.lasx != lasx {
+			return nil, fmt.Errorf("XVMOVQ: xvinsve0 is the %s-bank form alone", bank)
+		}
+		op, limit := enc.xinsW, 7
+		if dst.width != 'W' {
+			op, limit = enc.xinsD, 3
+			if dst.width != 'V' {
+				return nil, fmt.Errorf("XVMOVQ: xvinsve0 takes word or double-word lanes, got %q", ops[1].Raw)
+			}
+		}
+		if dst.elem > limit {
+			return nil, fmt.Errorf("XVMOVQ: element index %d out of range [0, %d]", dst.elem, limit)
+		}
+		return l64wordLE(op | uint32(dst.elem)<<10 | uint32(src.num)<<5 | uint32(dst.num)), nil
+	}
+
+	// Element extract into a vector register: XVMOVQ xj.T[i], xd
+	// (xvpickve.{w,d}), one word or double-word lane out to a bare X
+	// register.  LASX only.
+	if srcVec && src.hasEl && dstVec && !dst.hasSuf {
+		if !lasx || src.lasx != lasx || dst.lasx != lasx {
+			return nil, fmt.Errorf("XVMOVQ: xvpickve is the %s-bank form alone", bank)
+		}
+		op, limit := enc.xpickW, 7
+		if src.width != 'W' {
+			op, limit = enc.xpickD, 3
+			if src.width != 'V' {
+				return nil, fmt.Errorf("XVMOVQ: xvpickve takes word or double-word lanes, got %q", ops[0].Raw)
+			}
+		}
+		if src.elem > limit {
+			return nil, fmt.Errorf("XVMOVQ: element index %d out of range [0, %d]", src.elem, limit)
+		}
+		return l64wordLE(op | uint32(src.elem)<<10 | uint32(src.num)<<5 | uint32(dst.num)), nil
+	}
+
 	// Register move: VMOVQ vj, vd (vori.b/xvori.b with the zero constant),
 	// both operands bare registers of the same bank.
 	if srcVec && dstVec {
@@ -2224,7 +2457,7 @@ func encodeLOONG64Vmovq(lasx bool, ops []*ast.Operand, fi loong64FrameInfo) ([]b
 			}
 			return l64wordLE(l64rrr(enc.stx, rk, rj, src.num)), nil
 		}
-		rj, off := l64MemWithFrame(ops[1], fi)
+		rj, off := l64VmovqMem(ops[1], fi)
 		if rj < 0 || off < -2048 || off > 2047 {
 			return nil, fmt.Errorf("VMOVQ: store offset out of range [-2048, 2047]")
 		}
@@ -2247,9 +2480,9 @@ func encodeLOONG64Vmovq(lasx bool, ops []*ast.Operand, fi loong64FrameInfo) ([]b
 			}
 			return l64wordLE(l64rrr(enc.ldx, rk, rj, dst.num)), nil
 		}
-		rj, off := l64MemWithFrame(ops[0], fi)
-		if rj < 0 || off < -2048 || off > 2047 {
-			return nil, fmt.Errorf("VMOVQ: load offset out of range [-2048, 2047]")
+		rj, off := l64VmovqMem(ops[0], fi)
+		if rj < 0 {
+			return nil, fmt.Errorf("VMOVQ: invalid load operand")
 		}
 		op := enc.ld
 		if dst.hasSuf {
@@ -2257,16 +2490,34 @@ func encodeLOONG64Vmovq(lasx bool, ops []*ast.Operand, fi loong64FrameInfo) ([]b
 			if !ok {
 				return nil, fmt.Errorf("VMOVQ: invalid replicate width suffix %q", ops[1].Raw)
 			}
+			// vldrepl keeps the byte offset raw for bytes and scales it by
+			// the element width for the wider forms, the immediate field
+			// shrinking a bit per scale exactly as the toolchain encodes it
+			// (the field mask keeps the two's complement inside its width).
+			scale, mask, lo, hi := 1, int32(0xFFF), -2048, 2047
 			switch w {
 			case 0:
 				op = enc.replB
 			case 1:
 				op = enc.replH
+				scale, mask, lo, hi = 2, 0x7FF, -1024, 1023
 			case 2:
 				op = enc.replW
+				scale, mask, lo, hi = 4, 0x3FF, -512, 511
 			default:
 				op = enc.replD
+				scale, mask, lo, hi = 8, 0x1FF, -256, 255
 			}
+			if off%int32(scale) != 0 {
+				return nil, fmt.Errorf("VMOVQ: offset %d must be a multiple of %d", off, scale)
+			}
+			off /= int32(scale)
+			if off < int32(lo) || off > int32(hi) {
+				return nil, fmt.Errorf("VMOVQ: offset out of range [%d, %d]", lo*scale, hi*scale)
+			}
+			off &= mask
+		} else if off < -2048 || off > 2047 {
+			return nil, fmt.Errorf("VMOVQ: load offset out of range [-2048, 2047]")
 		}
 		return l64wordLE(l64irr(op, int(off), rj, dst.num)), nil
 	}

asm/loong64_encode.go

@@ -279,6 +279,8 @@ const (
 	l64Fvvv                     // 3R vector (LSX/LASX): op | vk<<10 | vj<<5 | vd
 	l64Fvcf                     // vector-to-condition: op | subop<<10 | vj<<5 | fcc
 	l64Fvvvv                    // 4R vector shuffle: op | va<<15 | vk<<10 | vj<<5 | vd
+	l64Fllsc                    // acquire/release LL/SC (2R against a zero-offset memory operand)
+	l64Fscq                     // sc.q: op | middle<<10 | base<<5 | first against a zero-offset memory operand
 )
 
 // l64Enc is one instruction's encoding: its bit layout (format) and the
@@ -341,8 +343,9 @@ var l64Vec2R = map[string]bool{}
 // such as vshuf.b).
 var l64Vec4R = map[string]bool{}
 
-// l64VmovqOps holds the VMOVQ/XVMOVQ opcode constants (pre-shifted to bit
-// 15), read off `go tool objdump` of GOARCH=loong64 `go tool asm` kernels.
+// l64VmovqOps holds the VMOVQ/XVMOVQ opcode constants, each pre-shifted to
+// its exact bit range, read off `go tool objdump` of GOARCH=loong64
+// `go tool asm` kernels and the toolchain's specialLsxMovInst table.
 type l64VmovqEnc struct {
 	ld, st, ldx, stx           uint32 // plain and indexed load/store
 	replB, replH, replW, replD uint32 // vldrepl: load and replicate element
@@ -350,6 +353,11 @@ type l64VmovqEnc struct {
 	ins                        uint32 // vinsgr2vr element insert
 	dup                        uint32 // vreplgr2vr duplicate (width in [11:10])
 	move                       uint32 // vori.b/xvori.b $0 register move
+	rveiB, rveiH, rveiW, rveiD uint32 // vreplvei: broadcast one element (LSX)
+	rve0B, rve0H, rve0W        uint32 // xvreplve0 broadcast of element zero (LASX)
+	rve0D, rve0Q               uint32 // xvreplve0.{d,q}, ditto
+	xinsW, xinsD               uint32 // xvinsve0: insert element zero (LASX)
+	xpickW, xpickD             uint32 // xvpickve: extract element (LASX)
 }
 
 var l64VmovqTable = map[bool]l64VmovqEnc{
@@ -358,12 +366,17 @@ var l64VmovqTable = map[bool]l64VmovqEnc{
 		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,
+		rveiB: 0x01CBDE << 14, rveiH: 0x0397BE << 13, rveiW: 0x072F7E << 12, rveiD: 0x0E5EFE << 11,
 	},
 	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,
+		rve0B: 0x1DC1C0 << 10, rve0H: 0x1DC1E0 << 10, rve0W: 0x1DC1F0 << 10,
+		rve0D: 0x1DC1F8 << 10, rve0Q: 0x1DC1FC << 10,
+		xinsW: 0x03B7FE << 13, xinsD: 0x076FFE << 12,
+		xpickW: 0x03B81E << 13, xpickD: 0x07703E << 12,
 	},
 }
 
@@ -373,7 +386,7 @@ func init() {
 		"ADD": 0x20 << 15, "ADDW": 0x20 << 15, "ADDV": 0x21 << 15, "ADDVU": 0x21 << 15,
 		"SUB": 0x22 << 15, "SUBW": 0x22 << 15, "SUBV": 0x23 << 15, "SUBVU": 0x23 << 15,
 		"SGT": 0x24 << 15, "SGTU": 0x25 << 15,
-		"MASKEQZ": 0x26 << 15, "MASKNEZ": 0x27 << 15, "SCQ": 0x070AE << 15,
+		"MASKEQZ": 0x26 << 15, "MASKNEZ": 0x27 << 15,
 		"NOR": 0x28 << 15, "AND": 0x29 << 15, "OR": 0x2a << 15, "XOR": 0x2b << 15,
 		"ORN": 0x2c << 15, "ANDN": 0x2d << 15,
 		"SLL": 0x2e << 15, "SRL": 0x2f << 15, "SRA": 0x30 << 15,
@@ -467,6 +480,20 @@ func init() {
 	l64InstrTable["RDTIMEHW"] = l64Enc{format: l64Frdtime, op: 0x19 << 10}
 	l64InstrTable["RDTIMED"] = l64Enc{format: l64Frdtime, op: 0x1a << 10}
 
+	// Acquire/release LL/SC (2R against a zero-offset memory operand):
+	// LLACQV (Rj), Rd loads, SCRELV Rd, (Rj) stores, both encoding
+	// op | rj<<5 | rd.  Opcodes from cmd/internal/obj/loong64/instOp.go
+	// (ll.acq.{w,d}, sc.rel.{w,d}).
+	l64InstrTable["LLACQW"] = l64Enc{format: l64Fllsc, op: 0x0E15E0 << 10}
+	l64InstrTable["SCRELW"] = l64Enc{format: l64Fllsc, op: 0x0E15E1 << 10}
+	l64InstrTable["LLACQV"] = l64Enc{format: l64Fllsc, op: 0x0E15E2 << 10}
+	l64InstrTable["SCRELV"] = l64Enc{format: l64Fllsc, op: 0x0E15E3 << 10}
+
+	// SCQ (sc.q first, middle, (base)) keeps its own operand order: the
+	// encoding is op | middle<<10 | base<<5 | first, the memory operand's
+	// base in the rj field, not the toolchain's generic 3R layout.
+	l64InstrTable["SCQ"] = l64Enc{format: l64Fscq, op: 0x070AE << 15}
+
 	// The dual-form arithmetic mnemonics (register 3R + immediate 2RI12),
 	// selected by the operand kind; the shift mnemonics pair the 3R form
 	// with a 5/6-bit shift immediate.
@@ -868,7 +895,7 @@ func init() {
 		"VNORB":  {0xE7B8 << 15, false, 0, 255, 0, 0xFF},
 		"XVNORB": {0xEFB8 << 15, true, 0, 255, 0, 0xFF},
 		"VSEQB":  {0xE500 << 15, false, -16, 15, 0, 0x1F},
-		"XVSEQB": {0xE900 << 15, true, -16, 15, 0, 0x1F},
+		"XVSEQB": {0xED00 << 15, true, -16, 15, 0, 0x1F},
 		// vseqi.h/w accept the same si5 window as vseqi.b; vseqi.d carries a
 		// 7-bit field, but the toolchain range-checks it down to si5 as well
 		// (GOARCH=loong64 go tool asm rejects VSEQV $32 and VSEQV $-64).
@@ -877,7 +904,7 @@ func init() {
 		"VSEQW":  {0xE502 << 15, false, -16, 15, 0, 0x1F},
 		"XVSEQW": {0xED02 << 15, true, -16, 15, 0, 0x1F},
 		"VSEQV":  {0xE503 << 15, false, -16, 15, 0, 0x7F},
-		"XVSEQV": {0xE903 << 15, true, -16, 15, 0, 0x7F},
+		"XVSEQV": {0xED03 << 15, true, -16, 15, 0, 0x7F},
 		// vslti compares against a signed (or, in the U spellings, unsigned)
 		// si5/ui5 constant.
 		"VSLTB":   {0xE50C << 15, false, -16, 15, 0, 0x1F},

asm/loong64_encode_test.go

@@ -8,8 +8,8 @@ import (
 	"encoding/binary"
 	"testing"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/ast"
-	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+	"sourcedock.dev/petrbalvin/gasm-sdk/ast"
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
 )
 
 // firstTextLOONG64 parses assembly source and returns the first TEXT body.
@@ -831,3 +831,230 @@ TEXT ·atoms(SB), NOSPLIT, $0
 		0x4C000020,
 	)
 }
+
+// TestLOONG64_llacqScrel pins the acquire/release LL/SC pair.  The oracle
+// words come from GOARCH=loong64 go tool objdump and the toolchain's own
+// loong64enc1.s golden bytes.
+func TestLOONG64_llacqScrel(t *testing.T) {
+	fn := firstTextLOONG64(t, `#include "textflag.h"
+TEXT ·llsc(SB), NOSPLIT, $0
+	LLACQW (R5), R4
+	LLACQV (R5), R4
+	SCRELW R4, (R6)
+	SCRELV R4, (R6)
+	RET
+`)
+	code := assembleLOONG64Helper(t, fn)
+	wantWords(t, code,
+		0x385780A4, // ll.acq.w r4, r5
+		0x385788A4, // ll.acq.d r4, r5
+		0x385784C4, // sc.rel.w r4, r6
+		0x38578CC4, // sc.rel.d r4, r6
+		0x4C000020,
+	)
+	// The toolchain accepts the zero-offset memory form alone.
+	for i, src := range []string{
+		`TEXT ·e(SB), NOSPLIT, $0
+	LLACQW 4(R5), R4
+	RET
+`,
+		`TEXT ·e(SB), NOSPLIT, $0
+	SCRELV R4, 8(R6)
+	RET
+`,
+	} {
+		fn := firstTextLOONG64(t, src)
+		if _, _, _, _, _, err := assembleLOONG64(fn); err == nil {
+			t.Errorf("case %d: expected an error, got none", i)
+		}
+	}
+}
+
+// TestLOONG64_vmovqSuffixed pins the element-broadcast and element-move
+// VMOVQ/XVMOVQ forms, with the oracle words lifted verbatim from the
+// toolchain's loong64enc1.s.
+func TestLOONG64_vmovqSuffixed(t *testing.T) {
+	fn := firstTextLOONG64(t, `#include "textflag.h"
+TEXT ·vmovq(SB), NOSPLIT, $0
+	VMOVQ V1.B[3], V9.B16
+	VMOVQ V2.H[2], V8.H8
+	VMOVQ V3.W[1], V7.W4
+	VMOVQ V4.V[0], V6.V2
+	XVMOVQ X0, X31.B32
+	XVMOVQ X1, X30.H16
+	XVMOVQ X2, X29.W8
+	XVMOVQ X3, X28.V4
+	XVMOVQ X3, X27.Q2
+	XVMOVQ X0, X31.W[7]
+	XVMOVQ X1, X29.W[0]
+	XVMOVQ X3, X28.V[3]
+	XVMOVQ X4, X27.V[0]
+	XVMOVQ X31.W[7], X0
+	XVMOVQ X29.W[0], X1
+	XVMOVQ X28.V[3], X8
+	XVMOVQ X27.V[0], X9
+	RET
+`)
+	code := assembleLOONG64Helper(t, fn)
+	wantWords(t, code,
+		0x72F78C29, // vreplvei.b v9, v1, 3
+		0x72F7C848, // vreplvei.h v8, v2, 2
+		0x72F7E467, // vreplvei.w v7, v3, 1
+		0x72F7F086, // vreplvei.d v6, v4, 0
+		0x7707001F, // xvreplve0.b x31, x0
+		0x7707803E, // xvreplve0.h x30, x1
+		0x7707C05D, // xvreplve0.w x29, x2
+		0x7707E07C, // xvreplve0.d x28, x3
+		0x7707F07B, // xvreplve0.q x27, x3
+		0x76FFDC1F, // xvinsve0.w x31, x0, 7
+		0x76FFC03D, // xvinsve0.w x29, x1, 0
+		0x76FFEC7C, // xvinsve0.d x28, x3, 3
+		0x76FFE09B, // xvinsve0.d x27, x4, 0
+		0x7703DFE0, // xvpickve.w x0, x31, 7
+		0x7703C3A1, // xvpickve.w x1, x29, 0
+		0x7703EF88, // xvpickve.d x8, x28, 3
+		0x7703E369, // xvpickve.d x9, x27, 0
+		0x4C000020,
+	)
+	// The rejected shapes: a width mismatch between the element and the
+	// arrangement, an element index past the lane count, a wrong-bank
+	// vreplvei and an arrangement the LASX bank does not spell.
+	for i, src := range []string{
+		`TEXT ·e(SB), NOSPLIT, $0
+	VMOVQ V1.H[3], V9.B16
+	RET
+`,
+		`TEXT ·e(SB), NOSPLIT, $0
+	VMOVQ V1.B[16], V9.B16
+	RET
+`,
+		`TEXT ·e(SB), NOSPLIT, $0
+	XVMOVQ X1.B[3], X9.B32
+	RET
+`,
+		`TEXT ·e(SB), NOSPLIT, $0
+	XVMOVQ X0, X31.B16
+	RET
+`,
+		`TEXT ·e(SB), NOSPLIT, $0
+	XVMOVQ X0, X31.W[8]
+	RET
+`,
+	} {
+		fn := firstTextLOONG64(t, src)
+		if _, _, _, _, _, err := assembleLOONG64(fn); err == nil {
+			t.Errorf("case %d: expected an error, got none", i)
+		}
+	}
+}
+
+// TestLOONG64_parityFixes pins the operand forms whose encodings were found
+// diverging from the toolchain by the loong64enc1.s differential: the
+// $off(reg) address immediate (addi.d), the SCQ operand order, the scaled
+// vldrepl offsets (with their field masks), the XVSEQB/XVSEQV immediate
+// opcodes and the zero-register base the toolchain gives FP-relative
+// VMOVQ/XVMOVQ memory operands.  Golden words from loong64enc1.s.
+func TestLOONG64_parityFixes(t *testing.T) {
+	fn := firstTextLOONG64(t, `#include "textflag.h"
+TEXT ·parity(SB), NOSPLIT, $0-32
+	MOVW $4(R4), R5
+	MOVV $4(R4), R5
+	MOVW $65536(R4), R5
+	MOVW $-4096(R4), R5
+	SCQ R4, R5, (R6)
+	VMOVQ 2(R4), V1.H8
+	VMOVQ -6(R4), V1.H8
+	VMOVQ -12(R4), V2.W4
+	VMOVQ -16(R4), V3.V2
+	XVMOVQ -10(R4), X1.H16
+	XVSEQB $0, X2, X4
+	XVSEQH $3, X2, X4
+	XVSEQW $12, X2, X4
+	XVSEQV $15, X2, X4
+	XVSEQV $-15, X2, X4
+	VMOVQ V2, y+16(FP)
+	VMOVQ y+16(FP), V2
+	VMOVQ V2, x+2030(FP)
+	XVMOVQ X6, y+16(FP)
+	RET
+`)
+	code := assembleLOONG64Helper(t, fn)
+	wantWords(t, code,
+		0x02C01085, // addi.d $4, r4, r5
+		0x02C01085, // addi.d $4, r4, r5 (MOVW keeps the 64-bit addi.d)
+		0x1400021E, // lu12i.w $16, r30
+		0x038003DE, // ori $0, r30, r30
+		0x0010F885, // add.d r5, r4, r30
+		0x15FFFFFE, // lu12i.w $-1, r30
+		0x038003DE, // ori $0, r30, r30
+		0x0010F885, // add.d r5, r4, r30
+		0x385714C4, // sc.q r4, r5, (r6): middle<<10 | base<<5 | first
+		0x30400481, // vldrepl.h v1, 2(r4)
+		0x305FF481, // vldrepl.h v1, -6(r4)
+		0x302FF482, // vldrepl.w v2, -12(r4)
+		0x3017F883, // vldrepl.d v3, -16(r4)
+		0x325FEC81, // xvldrepl.h x1, -10(r4)
+		0x76800044, // xvseqi.b x4, x2, 0
+		0x76808C44, // xvseqi.h x4, x2, 3
+		0x76813044, // xvseqi.w x4, x2, 12
+		0x7681BC44, // xvseqi.d x4, x2, 15
+		0x7681C444, // xvseqi.d x4, x2, -15
+		0x2C406002, // vst v2, 24(r0): FP-relative keeps the zero base
+		0x2C006002, // vld v2, 24(r0)
+		0x2C5FD802, // vst v2, 2038(r0)
+		0x2CC06006, // xvst x6, 24(r0)
+		0x4C000020,
+	)
+	// Misaligned vldrepl offsets are rejected, as the toolchain does.
+	for i, src := range []string{
+		`TEXT ·e(SB), NOSPLIT, $0
+	VMOVQ 3(R4), V1.H8
+	RET
+`,
+		`TEXT ·e(SB), NOSPLIT, $0
+	MOVW $4(R4), F1
+	RET
+`,
+	} {
+		fn := firstTextLOONG64(t, src)
+		if _, _, _, _, _, err := assembleLOONG64(fn); err == nil {
+			t.Errorf("case %d: expected an error, got none", i)
+		}
+	}
+}
+
+// TestLOONG64_bytePseudo pins the BYTE literal-data pseudo-op, which the
+// loong64 toolchain does not spell but the arm64 and riscv64 encoders of
+// this package already accept for byte-exact data layout (a superset
+// spelling, shippable via the goobj path).
+func TestLOONG64_bytePseudo(t *testing.T) {
+	fn := firstTextLOONG64(t, `#include "textflag.h"
+TEXT ·bytes(SB), NOSPLIT, $0
+	BYTE $2
+	BYTE $1; BYTE $0
+	BYTE $255
+	RET
+`)
+	code := assembleLOONG64Helper(t, fn)
+	// Four literal bytes, then RET (jirl r0, r1, 0); the trailing bytes pad
+	// the final word the way any sub-word tail does.
+	want := []byte{2, 1, 0, 0xFF, 0x20, 0x00, 0x00, 0x4C}
+	if !bytes.Equal(code[:len(want)], want) {
+		t.Errorf("bytes = % x, want % x", code, want)
+	}
+	for _, src := range []string{
+		`TEXT ·e(SB), NOSPLIT, $0
+	BYTE $256
+	RET
+`,
+		`TEXT ·e(SB), NOSPLIT, $0
+	BYTE $-1
+	RET
+`,
+	} {
+		fn := firstTextLOONG64(t, src)
+		if _, _, _, _, _, err := assembleLOONG64(fn); err == nil {
+			t.Errorf("%q: expected an error, got none", src)
+		}
+	}
+}

asm/loong64_frame.go

@@ -6,7 +6,7 @@ package asm
 import (
 	"strings"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/ast"
+	"sourcedock.dev/petrbalvin/gasm-sdk/ast"
 )
 
 // Loong64 frame mapping, matching the Go toolchain's loong64 backend.

asm/loong64_more_test.go

@@ -7,7 +7,7 @@ import (
 	"bytes"
 	"testing"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
 )
 
 // TestLOONG64_sys exercises the no-operand system instructions and the

asm/loong64_reloc_test.go

@@ -6,7 +6,7 @@ package asm
 import (
 	"testing"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
 )
 
 // TestLOONG64RelocOffsetsIncludePrologue pins the function-relative

asm/operand.go

@@ -36,6 +36,29 @@ type FloatImm struct {
 
 func (FloatImm) isOperand() {}
 
+// TLSMem is a thread-local access, the source form off(base)(TLS*1) with the
+// base dropped: the toolchain's one-instruction TLS rewrite assembles it as
+// the segment-prefixed absolute whose disp32 carries an R_TLS_LE patch site
+// (the linker fills the TLS slot offset).
+type TLSMem struct {
+	Disp int64
+	Size int
+	Seg  byte // the segment override: FS (0x64) or GS (0x65) on windows
+}
+
+func (TLSMem) isOperand() {}
+
+// SegAbs is a segment-absolute access, 0x30(GS): the segment override
+// prefixes a disp32 absolute reference with no relocation.  The base
+// register spellings GS and FS produce it.
+type SegAbs struct {
+	Disp int64
+	Size int
+	Seg  byte // 0x64 FS, 0x65 GS
+}
+
+func (SegAbs) isOperand() {}
+
 // Mem is a memory operand of the form disp(base)(index*scale).
 type Mem struct {
 	Base     Reg
@@ -45,6 +68,7 @@ type Mem struct {
 	Size     int // operand width in bytes
 	HasBase  bool
 	HasIndex bool
+	Seg      byte // segment override prefix (0x64 FS, 0x65 GS); 0 = none
 }
 
 func (Mem) isOperand() {}

asm/reg.go

@@ -17,13 +17,28 @@ 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, the fp flag the x87 stack registers F0-F7.
+// registers K0-K7, the fp flag the x87 stack registers F0-F7, the mmx flag the
+// MMX registers M0-M7, the seg field a bare segment register (FS, GS) and the
+// ctl field the control and debug registers, whose number rides an
+// instruction's reg field rather than r/m.
 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
+	mmx  bool // M0-M7 MMX register
+	seg  int  // segment register number plus one (ES=1..GS=6); 0 = not one
+	ctl  byte // 0 none, 1 CRn control register, 2 DRn debug register
+}
+
+// segNumber returns the segment register number (ES=0..GS=5) when r names a
+// bare segment register.
+func (r Reg) segNumber() (int, bool) {
+	if r.seg == 0 {
+		return 0, false
+	}
+	return r.seg - 1, true
 }
 
 // Index returns the register number (0-15 for GPRs, 0-31 for vectors).
@@ -149,6 +164,20 @@ func buildRegByName() map[string]Reg {
 	for i := 0; i <= 7; i++ {
 		m["F"+itoa(i)] = Reg{idx: i, size: 8, fp: true}
 	}
+	// MMX: M0..M7.
+	for i := 0; i <= 7; i++ {
+		m["M"+itoa(i)] = Reg{idx: i, size: 8, mmx: true}
+	}
+	// Bare segment registers: ES, CS, SS, DS, FS, GS (the memory-base and
+	// index spellings of FS and GS are handled before register lookup).
+	for i, n := range []string{"ES", "CS", "SS", "DS", "FS", "GS"} {
+		m[n] = Reg{idx: i, size: 2, seg: i + 1}
+	}
+	// Control and debug registers: CR0..CR15, DR0..DR15.
+	for i := 0; i <= 15; i++ {
+		m["CR"+itoa(i)] = Reg{idx: i, size: 8, ctl: 1}
+		m["DR"+itoa(i)] = Reg{idx: i, size: 8, ctl: 2}
+	}
 	return m
 }
 

asm/riscv_assemble.go

@@ -10,7 +10,7 @@ import (
 	"slices"
 	"strings"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/ast"
+	"sourcedock.dev/petrbalvin/gasm-sdk/ast"
 )
 
 // assembleRISCV assembles a RISC-V TEXT function body into machine code.

asm/riscv_encode_test.go

@@ -10,8 +10,8 @@ import (
 	"strings"
 	"testing"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/ast"
-	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+	"sourcedock.dev/petrbalvin/gasm-sdk/ast"
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
 )
 
 // firstTextRISCV parses assembly source and returns the first TEXT function body.

asm/riscv_frame.go

@@ -7,7 +7,7 @@ import (
 	"fmt"
 	"strings"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/ast"
+	"sourcedock.dev/petrbalvin/gasm-sdk/ast"
 )
 
 // RISC-V frame mapping, matching the Go toolchain's riscv64 backend.

asm/riscv_frame_test.go

@@ -6,7 +6,7 @@ package asm
 import (
 	"testing"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
 )
 
 // TestRISCVFrameSpadjAndLines checks that a framed function records its

asm/riscv_goobj_test.go

@@ -13,7 +13,7 @@ import (
 	"strings"
 	"testing"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
 )
 
 // TestGOObjectRISCVCallReloc checks that CALL sym(SB) emits a single JAL

asm/vex.go

@@ -76,6 +76,10 @@ const (
 	// carries a vector length, so the register the L'L field follows is the
 	// XMM source.
 	vexExtractGPR
+	// vexBlend4 is the four-operand variable blend `OP mask, src2, src1,
+	// dst` (VPBLENDVB): ModRM.reg = dst (op3), VEX.vvvv = src1 (op2),
+	// ModRM.rm = src2 (op1) and the mask register in the /is4 byte (op0).
+	vexBlend4
 )
 
 // vexSpec describes one VEX instruction's encoding parameters.
@@ -202,8 +206,28 @@ var vexTable = map[string]vexSpec{
 
 	// VEX.128/256.66.0F.WIG, immediate shuffle (reg=dst, rm=src, imm8).
 	"VPSHUFD": {1, 0x70, 0, 1, -1, vexImmRM},
-	// VEX.256.66.0F3A.W1, qword permute (reg=dst, rm=src, imm8).
+	// VEX.256.66.0F3A.W1, qword permute (reg=dst, rm=src, imm8), and its
+	// double twin under op 01; the in-lane permutes under 04/05.
 	"VPERMQ":    {3, 0x00, 1, 1, -1, vexImmRM},
+	"VPERMPD":   {3, 0x01, 1, 1, -1, vexImmRM},
+	"VPERMILPS": {3, 0x04, 0, 1, -1, vexImmRM},
+	"VPERMILPD": {3, 0x05, 0, 1, -1, vexImmRM},
+	// VEX.66.0F3A.W0, the immediate-controlled AVX tail: the rounding
+	// pair, the AES key assistant and the string compares.
+	"VROUNDPD":         {3, 0x09, 0, 1, -1, vexImmRM},
+	"VROUNDPS":         {3, 0x08, 0, 1, -1, vexImmRM},
+	"VAESKEYGENASSIST": {3, 0xDF, 0, 1, -1, vexImmRM},
+	"VPCMPESTRI":       {3, 0x61, 0, 1, -1, vexImmRM},
+	"VPCMPESTRM":       {3, 0x60, 0, 1, -1, vexImmRM},
+	"VPCMPISTRI":       {3, 0x63, 0, 1, -1, vexImmRM},
+	"VPCMPISTRM":       {3, 0x62, 0, 1, -1, vexImmRM},
+	// VEX.128.66.0F3A.W0, the scalar lane extract to a GPR or memory
+	// (reg = the XMM source, r/m = the destination).
+	"VEXTRACTPS": {3, 0x17, 0, 1, -1, vexExtractGPR},
+	"VPEXTRW":    {3, 0x15, 0, 1, -1, vexExtractGPR},
+	// VEX.128.66.0F3A.W0, the four-operand variable blend with its mask
+	// register in the /is4 byte.
+	"VPBLENDVB": {3, 0x4C, 0, 1, -1, vexBlend4},
 
 	// VEX.128/256.66.0F.WIG, two-source shuffle (reg=dst, vvvv=src1, rm=src2,
 	// imm8).
@@ -524,8 +548,16 @@ func isVex(mnemUpper string) bool {
 	if _, ok := vexTable[mnemUpper]; ok {
 		return true
 	}
-	_, ok := vexMoveTable[mnemUpper]
-	return ok
+	if _, ok := vexMoveTable[mnemUpper]; ok {
+		return true
+	}
+	// The dual-shape moves (VMOVHPD/VMOVLPD) pick their VEX form by operand
+	// count in encodeVex.
+	switch mnemUpper {
+	case "VMOVHPD", "VMOVLPD":
+		return true
+	}
+	return false
 }
 
 // encodeVex encodes a VEX instruction with operands in Plan 9 order.
@@ -559,6 +591,22 @@ func (e *enc) encodeVex(mnemUpper string, ops []Operand) error {
 			return e.encodeVexNDS3(vexSpec{mapSel: 1, opcode: op, pp: 1, opdigit: -1, form: vexNDS3}, ops)
 		}
 	}
+	// The high/low double moves split by operand count: three operands
+	// load-and-insert (mem, src, dst, an NDS form), two store (xmm, m64,
+	// the reversed store layout).
+	if mnemUpper == "VMOVHPD" || mnemUpper == "VMOVLPD" {
+		loadOp, storeOp := byte(0x16), byte(0x17)
+		if mnemUpper == "VMOVLPD" {
+			loadOp, storeOp = 0x12, 0x13
+		}
+		switch len(ops) {
+		case 3:
+			return e.encodeVexNDS3(vexSpec{mapSel: 1, opcode: loadOp, w: 0, pp: 1, opdigit: -1, form: vexNDS3}, ops)
+		case 2:
+			return e.encodeVexRMRev(vexSpec{mapSel: 1, opcode: storeOp, w: 0, pp: 1, opdigit: -1, form: vexRMRev}, ops)
+		}
+		return fmt.Errorf("%s expects 2 or 3 operands, got %d", mnemUpper, len(ops))
+	}
 	spec := vexTable[mnemUpper]
 	switch spec.form {
 	case vexNDS3:
@@ -573,6 +621,10 @@ func (e *enc) encodeVex(mnemUpper string, ops []Operand) error {
 		return e.encodeVexNDS3Imm(spec, ops)
 	case vexExtract:
 		return e.encodeVexExtract(spec, ops)
+	case vexExtractGPR:
+		return e.encodeVexExtractGPR(spec, ops)
+	case vexBlend4:
+		return e.encodeVexBlend4(spec, ops)
 	case vexRMSrcLen:
 		return e.encodeVexRMSrcLen(mnemUpper, spec, ops)
 	case vexZero:
@@ -964,6 +1016,73 @@ func (e *enc) encodeVexRMRev(spec vexSpec, ops []Operand) error {
 	return e.emitVexFields(spec, srcReg.vecLenBit(), srcReg.idx&7, rBit, 15, ops[1])
 }
 
+// encodeVexExtractGPR encodes the lane extract to a general-purpose register
+// or memory (VEXTRACTPS): OP $imm, xsrc, gpr/mem with the XMM source in
+// ModRM.reg and the destination in r/m, L = 0.
+func (e *enc) encodeVexExtractGPR(spec vexSpec, ops []Operand) error {
+	if len(ops) != 3 {
+		return fmt.Errorf("extract expects 3 operands ($imm, xsrc, dst), got %d", len(ops))
+	}
+	imm, src, dst := ops[0], ops[1], ops[2]
+	immVal, ok := imm.(Imm)
+	if !ok {
+		return fmt.Errorf("extract lane must be an immediate")
+	}
+	srcReg, ok := src.(Reg)
+	if !ok || !srcReg.isVec() || srcReg.size != 16 {
+		return fmt.Errorf("extract source must be an XMM register")
+	}
+	if _, isReg := dst.(Reg); !isReg && !memOperand(dst) {
+		return fmt.Errorf("extract destination must be a register or memory")
+	}
+	rBit := 0
+	if srcReg.idx >= 8 {
+		rBit = 1
+	}
+	if err := e.emitVexFields(spec, 0, srcReg.idx&7, rBit, 15, dst); err != nil {
+		return err
+	}
+	immByte, err := imm8(int64(immVal))
+	if err != nil {
+		return err
+	}
+	e.out = append(e.out, immByte)
+	return nil
+}
+
+// encodeVexBlend4 encodes the four-operand variable blend (VPBLENDVB):
+// OP mask, src2, src1, dst with ModRM.reg = dst, VEX.vvvv = src1, r/m =
+// src2 and the mask XMM register in the trailing /is4 byte.
+func (e *enc) encodeVexBlend4(spec vexSpec, ops []Operand) error {
+	if len(ops) != 4 {
+		return fmt.Errorf("blend expects 4 operands (mask, src2, src1, dst), got %d", len(ops))
+	}
+	mask, src2, src1, dst := ops[0], ops[1], ops[2], ops[3]
+	maskReg, ok := mask.(Reg)
+	if !ok || !maskReg.isVec() || maskReg.size != 16 {
+		return fmt.Errorf("blend mask must be an XMM register")
+	}
+	vvvvReg, ok := src1.(Reg)
+	if !ok || !vvvvReg.isVec() {
+		return fmt.Errorf("blend second source must be a vector register")
+	}
+	dstReg, ok := dst.(Reg)
+	if !ok || !dstReg.isVec() {
+		return fmt.Errorf("blend destination must be a vector register")
+	}
+	rBit := 0
+	if dstReg.idx >= 8 {
+		rBit = 1
+	}
+	if err := e.emitVexFields(spec, dstReg.vecLenBit(), dstReg.idx&7, rBit, 15-(vvvvReg.idx&15), src2); err != nil {
+		return err
+	}
+	// The /is4 byte names the mask register: bits [3:0] its low nibble,
+	// bit 7 the fourth register bit (X8-X15).
+	e.out = append(e.out, byte(maskReg.idx&7)|byte((maskReg.idx&8)<<4))
+	return nil
+}
+
 // 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

ast/ast.go

@@ -8,7 +8,7 @@
 // to the arch package; the AST records syntax only.
 package ast
 
-import "sourcedock.dev/petrbalvin/gasm-devkit/token"
+import "sourcedock.dev/petrbalvin/gasm-sdk/token"
 
 // File is the parsed representation of one .s source file.
 type File struct {

ast/ast_test.go

@@ -6,7 +6,7 @@ package ast
 import (
 	"testing"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/token"
+	"sourcedock.dev/petrbalvin/gasm-sdk/token"
 )
 
 func pos(line, col int) token.Position { return token.Position{Line: line, Column: col} }

cmd/gasm/asmhdr.go

@@ -17,7 +17,7 @@ import (
 	"regexp"
 	"strings"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/arch"
+	"sourcedock.dev/petrbalvin/gasm-sdk/arch"
 )
 
 // go_asm.h is the header the Go compiler writes for every package that

cmd/gasm/asmhdr_test.go

@@ -5,6 +5,7 @@ package main
 
 import (
 	"os"
+	"os/exec"
 	"path/filepath"
 	"strings"
 	"testing"
@@ -392,14 +393,67 @@ func TestRunCorpusAuditGOOS(t *testing.T) {
 	}
 }
 
+// TestRunCorpusAuditBuildConstraint covers the //go:build classification end
+// to end: a generic-named file whose constraint admits one target is
+// attempted there alone (cpu_x86.s on amd64), and a file whose constraint
+// admits none of the four targets is never attempted (the msan and
+// goexperiment trees).
+func TestRunCorpusAuditBuildConstraint(t *testing.T) {
+	dir := t.TempDir()
+	write := func(name, src string) {
+		t.Helper()
+		if err := os.WriteFile(filepath.Join(dir, name), []byte(src), 0o644); err != nil {
+			t.Fatal(err)
+		}
+	}
+	write("x86.s", "//go:build 386 || amd64\n\nTEXT \xc2\xb7f(SB), NOSPLIT, $0\n\tRET\n")
+	write("racey.s", "//go:build race\n\nTEXT \xc2\xb7r(SB), NOSPLIT, $0\n\tRET\n")
+	write("plain.s", "TEXT \xc2\xb7p(SB), NOSPLIT, $0\n\tRET\n")
+
+	stats, err := runCorpusAudit(dir, nil)
+	if err != nil {
+		t.Fatalf("runCorpusAudit: %v", err)
+	}
+	tally := func(name string) *corpusTally {
+		for i, tg := range stats.targets {
+			if tg.name == name {
+				return stats.tallies[i]
+			}
+		}
+		t.Fatalf("no tally for %s", name)
+		return nil
+	}
+	if stats.narrowed != 1 || stats.excluded != 1 || stats.generic != 1 {
+		t.Errorf("buckets = narrowed %d, excluded %d, generic %d; want 1, 1, 1", stats.narrowed, stats.excluded, stats.generic)
+	}
+	if a := tally("amd64"); a.attempted != 2 || a.assembled != 2 {
+		t.Errorf("amd64 = %d/%d, want 2/2 (x86.s and plain.s)", a.assembled, a.attempted)
+	}
+	for _, name := range []string{"arm64", "riscv64", "loong64"} {
+		if a := tally(name); a.attempted != 1 || a.assembled != 1 {
+			t.Errorf("%s = %d/%d, want 1/1 (plain.s only)", name, a.assembled, a.attempted)
+		}
+	}
+	if stats.full != 2 {
+		t.Errorf("full = %d, want 2 (x86.s over its one target, plain.s over all four)", stats.full)
+	}
+}
+
 // TestGenerateGoAsmHeaderRuntime pins the generator against the real thing:
-// the runtime package, whose header the toolchain's own -asmhdr output was
-// sampled from.  Skipped in short mode: it type-checks the whole package.
+// the runtime package of the ambient toolchain, whose header the toolchain's
+// own -asmhdr output was sampled from.  Skipped in short mode: it type-checks
+// the whole package.  The GOROOT comes from the go command itself, so the
+// test follows whatever toolchain the host provides.
 func TestGenerateGoAsmHeaderRuntime(t *testing.T) {
 	if testing.Short() {
 		t.Skip("type-checks the whole runtime package")
 	}
-	dir, err := generateGoAsmHeader("/usr/local/go/src/runtime", "", "amd64", t.TempDir())
+	out, err := exec.Command("go", "env", "GOROOT").Output()
+	if err != nil {
+		t.Skipf("no Go toolchain: %v", err)
+	}
+	runtimeDir := filepath.Join(strings.TrimSpace(string(out)), "src", "runtime")
+	dir, err := generateGoAsmHeader(runtimeDir, "", "amd64", t.TempDir())
 	if err != nil {
 		t.Fatalf("generateGoAsmHeader(runtime): %v", err)
 	}

cmd/gasm/audit.go

@@ -5,6 +5,7 @@ package main
 
 import (
 	"fmt"
+	"go/build/constraint"
 	"maps"
 	"os"
 	"os/exec"
@@ -15,9 +16,9 @@ import (
 	"strconv"
 	"strings"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/arch"
-	"sourcedock.dev/petrbalvin/gasm-devkit/asm"
-	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+	"sourcedock.dev/petrbalvin/gasm-sdk/arch"
+	"sourcedock.dev/petrbalvin/gasm-sdk/asm"
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
 )
 
 // cmdAuditInstructions cross-checks a gasm encoder against the Go toolchain's
@@ -38,7 +39,7 @@ 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 [--corpus [dir]] [-I dir] [amd64|arm64|riscv64|loong64]", `
+	fs := newCommand("audit-instructions", "gasm audit-instructions [--corpus [dir]] [--list] [-I dir] [amd64|arm64|riscv64|loong64]", `
 Compare the gasm encoder for the given architecture (default amd64) against
 go tool asm and print the diff: superset encodings (gasm-only, shippable via
 gasm asm --format goobj) and known-but-unencodable names (the backlog).  The
@@ -55,16 +56,19 @@ 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.
+the encodability backlog by frequency rather than by table order.  With
+-list the report also prints every failing file with its reason, per
+architecture.
 `)
 	corpus := fs.Bool("corpus", false, "assemble a corpus of .s files and report pass rates and failure reasons")
+	list := fs.Bool("list", false, "with --corpus, list every failing file with its reason, per architecture")
 	var dirs includeDirs
 	fs.Var(&dirs, "I", "directory to search for #include files (may be repeated)")
 	if err := fs.Parse(args); err != nil {
 		return err
 	}
 	if *corpus {
-		return cmdAuditCorpus(fs.Args(), dirs)
+		return cmdAuditCorpus(fs.Args(), dirs, *list)
 	}
 	archName := "amd64"
 	switch n := len(fs.Args()); {
@@ -403,20 +407,30 @@ type corpusTally struct {
 	assembled int
 	reasons   map[string]int    // failure reason → count
 	example   map[string]string // failure reason → one representative file
+	fails     []corpusFailure   // every failure, in file order, for --list
 }
 
-func (t *corpusTally) fail(path, reason string) {
+// corpusFailure is one failed attempt, recorded for the --list report.
+type corpusFailure struct {
+	path   string
+	reason string
+	detail string
+}
+
+func (t *corpusTally) fail(path string, err error) {
+	reason := corpusReason(err)
 	t.reasons[reason]++
 	if t.example[reason] == "" {
 		t.example[reason] = path
 	}
+	t.fails = append(t.fails, corpusFailure{path: path, reason: reason, detail: firstLine(err.Error())})
 }
 
 // cmdAuditCorpus implements audit-instructions --corpus.  The include
 // directories carry #include resolution over a corpus whose files refer to
 // headers such as GOROOT/pkg/include, the same -I a toolchain comparison
 // needs.
-func cmdAuditCorpus(args []string, dirs includeDirs) error {
+func cmdAuditCorpus(args []string, dirs includeDirs, list bool) error {
 	if len(args) > 1 {
 		return &usageError{fmt.Errorf("audit-instructions --corpus takes at most one directory argument")}
 	}
@@ -454,7 +468,7 @@ func cmdAuditCorpus(args []string, dirs includeDirs) error {
 	if err != nil {
 		return err
 	}
-	printCorpusStats(stats)
+	printCorpusStats(stats, list)
 	return nil
 }
 
@@ -463,8 +477,10 @@ type corpusStats struct {
 	root      string
 	files     int
 	generic   int // files attempted for all four architectures
+	narrowed  int // files whose //go:build admits a proper subset of the four
+	excluded  int // files whose //go:build admits none of the four: never compiled
 	otherPort int // files named for another Go port: never attempted
-	full      int // files that assembled for every target architecture
+	full      int // files that assembled for every applicable target architecture
 	targets   []corpusTarget
 	tallies   []*corpusTally
 }
@@ -560,6 +576,57 @@ func otherGOOSFile(path string) bool {
 	return false
 }
 
+// buildConstraint returns the file's leading //go:build expression, or nil
+// when the file carries none.  The constraint governs the same header block
+// go/build reads: blank lines and comments may precede it, and the first
+// line that is neither ends the block.  A constraint that does not parse
+// narrows nothing, so the file stays in the attempted set: the audit must
+// never exclude a file the toolchain would compile.
+func buildConstraint(src string) constraint.Expr {
+	for line := range strings.SplitSeq(src, "\n") {
+		t := strings.TrimSpace(line)
+		switch {
+		case t == "":
+			continue
+		case strings.HasPrefix(t, "//"):
+			if constraint.IsGoBuild(t) {
+				e, err := constraint.Parse(t)
+				if err != nil {
+					return nil
+				}
+				return e
+			}
+			continue
+		default:
+			return nil
+		}
+	}
+	return nil
+}
+
+// unixOS is go/build's unixOS set: the GOOSes the unix build tag admits.
+var unixOS = map[string]bool{
+	"aix": true, "android": true, "darwin": true, "dragonfly": true,
+	"freebsd": true, "hurd": true, "illumos": true, "ios": true,
+	"linux": true, "netbsd": true, "openbsd": true, "solaris": true,
+}
+
+// constraintTags answers the build tags a plain `go build` sets for a
+// target: the GOOS and GOARCH, gc, and unix on the unix-like GOOSes.  No
+// experiment, sanitiser or cgo tag is ever true: the audit models the
+// default build, and no GOROOT assembly file's constraint hinges on cgo.
+func constraintTags(goarch, goos string) func(string) bool {
+	return func(tag string) bool {
+		switch tag {
+		case goarch, goos, "gc":
+			return true
+		case "unix":
+			return unixOS[goos]
+		}
+		return false
+	}
+}
+
 func runCorpusAudit(root string, dirs includeDirs) (*corpusStats, error) {
 	files, err := asmFiles(root)
 	if err != nil {
@@ -577,8 +644,8 @@ func runCorpusAudit(root string, dirs includeDirs) (*corpusStats, error) {
 		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
+	// its build admits assembles it.
+	full, generic, otherPort, narrowedCount, excluded := 0, 0, 0, 0, 0
 
 	// Header generation is created on first use, so a corpus with no
 	// go_asm.h includes never pays for a temp directory.
@@ -601,28 +668,78 @@ func runCorpusAudit(root string, dirs includeDirs) (*corpusStats, error) {
 		// invisible to a file-name rule).
 		goos := goosFromFilename(path)
 
+		// The GOOS the header generation type-checks under follows the
+		// file's name when the name carries one; the ambient GOOS is the
+		// honest guess otherwise.
+		namedArch := arch.FromFilename(path)
 		var wanted []int // indexes into targets
-		if a := arch.FromFilename(path); a != arch.Unknown {
+		other := false
+		switch {
+		case namedArch != arch.Unknown:
 			for i, tg := range targets {
-				if tg.a == a {
+				if tg.a == namedArch {
 					wanted = append(wanted, i)
 				}
 			}
-		} else if otherPortFile(path) {
+		case otherPortFile(path):
 			// A file named for a Go port gasm does not support (arm,
 			// 386, s390x, ...) or for another GOOS is compiled by no
 			// supported-arch build, so it is neither generic nor a
 			// per-arch attempt: counting it as generic would make the
 			// headline unreachably low for reasons no supported target
 			// can fix.
+			other = true
 			otherPort++
-		} else {
-			generic++
+		default:
 			for i := range targets {
 				wanted = append(wanted, i)
 			}
 		}
 
+		// A //go:build constraint narrows the set of targets the file is
+		// assembled for, the way the go command compiles the file only for
+		// the targets the expression admits: cpu_x86.s belongs to the x86
+		// build alone, and a file whose constraint admits none of the four
+		// targets (the goexperiment.runtimesecret and msan trees) is
+		// compiled by no supported build.  The tags mirror what a plain
+		// `go build` sets: the GOOS and GOARCH, gc, and unix on the
+		// unix-like GOOSes; no experiment, sanitiser or cgo tag is ever
+		// true.  The GOOS is the file's own when the name carries one,
+		// else the ambient one.
+		goosForEval := goos
+		if goosForEval == "" {
+			goosForEval = runtime.GOOS
+		}
+		narrowed := false
+		if len(wanted) > 0 {
+			if ce := buildConstraint(src); ce != nil {
+				kept := make([]int, 0, len(wanted))
+				for _, i := range wanted {
+					tg := targets[i]
+					if ce.Eval(constraintTags(goarchName(tg.a), goosForEval)) {
+						kept = append(kept, i)
+					}
+				}
+				if len(kept) < len(wanted) {
+					narrowed = true
+				}
+				wanted = kept
+			}
+		}
+
+		switch {
+		case other:
+			// already tallied above
+		case len(wanted) == 0:
+			excluded++
+		case namedArch != arch.Unknown:
+			// a per-arch attempt over the constraint's subset
+		case narrowed:
+			narrowedCount++
+		default:
+			generic++
+		}
+
 		// A file that includes go_asm.h parses against a per-target header:
 		// the defines differ per architecture (internal/cpu's layout, for
 		// one) and per GOOS (sys_darwin_arm64.s's trampoline constants,
@@ -645,21 +762,22 @@ func runCorpusAudit(root string, dirs includeDirs) (*corpusStats, error) {
 				hdrDir, err := hdr.dirFor(pkgDir, goos, goarchName(tg.a))
 				if err != nil {
 					ok = false
-					t.fail(path, corpusReason(err))
+					t.fail(path, err)
 					continue
 				}
 				f, errs := parser.ParseWithOptions(path, src, parser.Options{
 					Expand:      true,
 					IncludeDirs: append(slices.Clone(dirs), hdrDir),
+					Predefines:  platformPredefinesFor(goarchName(tg.a), goos),
 				})
 				if len(errs) > 0 {
 					ok = false
-					t.fail(path, corpusReason(errs[0]))
+					t.fail(path, errs[0])
 					continue
 				}
-				if _, err := assembleFile(tg.a, f); err != nil {
+				if _, err := assembleFile(tg.a, f, goos); err != nil {
 					ok = false
-					t.fail(path, corpusReason(err))
+					t.fail(path, err)
 					continue
 				}
 				t.assembled++
@@ -670,21 +788,28 @@ func runCorpusAudit(root string, dirs includeDirs) (*corpusStats, error) {
 			continue
 		}
 
-		f, errs := parser.ParseWithOptions(path, src, parser.Options{Expand: true, IncludeDirs: dirs})
-
 		ok := true
 		for _, i := range wanted {
 			tg, t := targets[i], tallies[i]
 			t.attempted++
+			// The parse carries the target's platform predefines, so it
+			// cannot be shared across targets the way a header-free file's
+			// could: a #ifdef GOARCH_arm block must be live on arm64 and
+			// dead everywhere else.
+			f, errs := parser.ParseWithOptions(path, src, parser.Options{
+				Expand:      true,
+				IncludeDirs: dirs,
+				Predefines:  platformPredefinesFor(goarchName(tg.a), goos),
+			})
 			var err error
 			if len(errs) > 0 {
 				err = errs[0] // a parse failure is a failure for every target
 			} else {
-				_, err = assembleFile(tg.a, f)
+				_, err = assembleFile(tg.a, f, goos)
 			}
 			if err != nil {
 				ok = false
-				t.fail(path, corpusReason(err))
+				t.fail(path, err)
 				continue
 			}
 			t.assembled++
@@ -698,6 +823,8 @@ func runCorpusAudit(root string, dirs includeDirs) (*corpusStats, error) {
 		root:      root,
 		files:     len(files),
 		generic:   generic,
+		narrowed:  narrowedCount,
+		excluded:  excluded,
 		otherPort: otherPort,
 		full:      full,
 		targets:   targets,
@@ -706,14 +833,16 @@ func runCorpusAudit(root string, dirs includeDirs) (*corpusStats, error) {
 }
 
 // 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)
+func printCorpusStats(s *corpusStats, list bool) {
+	fmt.Printf("corpus %s: %d files (%d generic, attempted for all architectures; %d narrowed by //go:build; %d excluded by //go:build; %d named for other Go ports, never attempted)\n",
+		s.root, s.files, s.generic, s.narrowed, s.excluded, s.otherPort)
 	// The rate is over the files a supported build would attempt: the
-	// other ports' files 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))
+	// other ports' files and the ones no supported target compiles sit in
+	// the count for completeness but can never assemble, so counting them
+	// in the denominator would report the gap of platforms gasm
+	// deliberately does not target.
+	attemptable := max(s.files-s.otherPort-s.excluded, 1)
+	fmt.Printf("  assemble for every applicable target: %d of %d attemptable (%.1f%%)\n", s.full, attemptable, 100*float64(s.full)/float64(attemptable))
 	for i, tg := range s.targets {
 		t := s.tallies[i]
 		fmt.Printf("  %s: %d/%d attempted\n", tg.name, t.assembled, t.attempted)
@@ -721,6 +850,13 @@ func printCorpusStats(s *corpusStats) {
 			fmt.Printf("    %4d  %s\n", t.reasons[r], r)
 			fmt.Printf("         e.g. %s\n", t.example[r])
 		}
+		if !list {
+			continue
+		}
+		for _, f := range t.fails {
+			fmt.Printf("    FAIL %s\n", f.path)
+			fmt.Printf("         %s: %s\n", f.reason, f.detail)
+		}
 	}
 }
 

cmd/gasm/audit_test.go

@@ -4,9 +4,10 @@
 package main
 
 import (
+	"runtime"
 	"testing"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/arch"
+	"sourcedock.dev/petrbalvin/gasm-sdk/arch"
 )
 
 func TestDerivedFamily(t *testing.T) {
@@ -64,3 +65,62 @@ func TestGasmEncodable(t *testing.T) {
 		}
 	}
 }
+
+// TestBuildConstraint pins the //go:build reader: the constraint governs the
+// leading comment block, the first non-comment line ends it (a tag below a
+// #include governs nothing, exactly as go/build drops it), and a file
+// without one admits every target.
+func TestBuildConstraint(t *testing.T) {
+	admits := func(src, goarch, goos string) bool {
+		t.Helper()
+		e := buildConstraint(src)
+		if e == nil {
+			return true
+		}
+		return e.Eval(constraintTags(goarch, goos))
+	}
+	const ret = "TEXT \xc2\xb7f(SB), NOSPLIT, $0\n\tRET\n"
+	cases := []struct {
+		name         string
+		src          string
+		amd64, arm64 bool
+	}{
+		{"no constraint", ret, true, true},
+		{"x86 only", "//go:build 386 || amd64\n\n" + ret, true, false},
+		{"arm64 and linux", "//go:build arm64 && linux\n\n" + ret, false, true},
+		{"msan never", "//go:build msan\n\n" + ret, false, false},
+		{"experiment never", "//go:build goexperiment.runtimesecret\n\n" + ret, false, false},
+		{"below an include governs nothing", "#include \"textflag.h\"\n//go:build amd64\n" + ret, true, true},
+		{"unparsable narrows nothing", "//go:build (amd64\n" + ret, true, true},
+	}
+	for _, c := range cases {
+		t.Run(c.name, func(t *testing.T) {
+			if got := admits(c.src, "amd64", runtime.GOOS); got != c.amd64 {
+				t.Errorf("amd64 admission = %v, want %v", got, c.amd64)
+			}
+			if got := admits(c.src, "arm64", runtime.GOOS); got != c.arm64 {
+				t.Errorf("arm64 admission = %v, want %v", got, c.arm64)
+			}
+		})
+	}
+}
+
+// TestConstraintTags pins the tag set a plain `go build` sets: the GOOS and
+// GOARCH, gc, unix on the unix-like GOOSes; nothing else is ever true.
+func TestConstraintTags(t *testing.T) {
+	ok := constraintTags("amd64", "linux")
+	for _, tag := range []string{"amd64", "linux", "gc", "unix"} {
+		if !ok(tag) {
+			t.Errorf("tag %q = false, want true", tag)
+		}
+	}
+	for _, tag := range []string{"arm64", "freebsd", "darwin", "cgo", "race", "msan", "goexperiment.runtimesecret"} {
+		if ok(tag) {
+			t.Errorf("tag %q = true, want false", tag)
+		}
+	}
+	fb := constraintTags("arm64", "freebsd")
+	if !fb("unix") {
+		t.Error("unix on freebsd = false, want true")
+	}
+}

cmd/gasm/debug_other.go

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

cmd/gasm/debug_ptrace.go

@@ -1,7 +1,7 @@
 // Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
 // SPDX-License-Identifier: BSD-3-Clause
 
-//go:build linux
+//go:build linux || (freebsd && (amd64 || arm64 || riscv64))
 
 package main
 
@@ -13,8 +13,8 @@ import (
 	"strings"
 	"time"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/debug"
-	"sourcedock.dev/petrbalvin/gasm-devkit/verify"
+	"sourcedock.dev/petrbalvin/gasm-sdk/debug"
+	"sourcedock.dev/petrbalvin/gasm-sdk/verify"
 )
 
 func cmdDebug(args []string) int {
@@ -239,10 +239,12 @@ REPL commands:
 				fmt.Printf("gasm debug: cover: stopped on signal %v\n", sig)
 				break
 			}
+			reason, _ := sess.StopInfo()
 			regs, rerr := sess.GetRegs()
 			if rerr != nil {
 				break
 			}
+			trapPC := regs.GetPC()
 			// HandleTrap restores the original byte, rewinds PC and counts
 			// the hit on the breakpoint itself.  Single-step over the
 			// restored instruction so the reinsertion at the top of the
@@ -251,6 +253,12 @@ REPL commands:
 				if err := sess.Step(); err != nil {
 					break
 				}
+			} else if debug.TrapStray(sess, reason, trapPC) {
+				// A breakpoint-class trap that matches none of ours and left
+				// the PC in place: resuming would re-execute the trapping
+				// instruction forever, so the coverage run stops here.
+				fmt.Printf("gasm debug: cover: SIGTRAP at %#x matches no breakpoint; the PC did not advance\n", trapPC)
+				break
 			}
 		}
 		hits := map[uint64]int{}

cmd/gasm/dis.go

@@ -9,9 +9,9 @@ import (
 	"sort"
 	"strings"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/arch"
-	"sourcedock.dev/petrbalvin/gasm-devkit/disasm"
-	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+	"sourcedock.dev/petrbalvin/gasm-sdk/arch"
+	"sourcedock.dev/petrbalvin/gasm-sdk/disasm"
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
 )
 
 // cmdDis disassembles machine code: either a raw binary (standard input with
@@ -84,7 +84,7 @@ func disSource(path string, target arch.Arch) int {
 	if len(errs) > 0 {
 		return 1
 	}
-	img, err := assembleFile(target, f)
+	img, err := assembleFile(target, f, "")
 	if err != nil {
 		fmt.Fprintf(os.Stderr, "gasm dis: %v\n", err)
 		return 1

cmd/gasm/main.go

@@ -27,15 +27,15 @@ import (
 	"sync"
 	"syscall"
 
-	"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/verify"
+	"sourcedock.dev/petrbalvin/gasm-sdk/arch"
+	"sourcedock.dev/petrbalvin/gasm-sdk/asm"
+	"sourcedock.dev/petrbalvin/gasm-sdk/ast"
+	"sourcedock.dev/petrbalvin/gasm-sdk/format"
+	"sourcedock.dev/petrbalvin/gasm-sdk/lexer"
+	"sourcedock.dev/petrbalvin/gasm-sdk/lint"
+	"sourcedock.dev/petrbalvin/gasm-sdk/lsp"
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
+	"sourcedock.dev/petrbalvin/gasm-sdk/verify"
 )
 
 // version reports the release the toolchain recorded for this build: the
@@ -574,7 +574,7 @@ naming the package.
 		defer cleanup()
 		dirs = append(dirs, hdrDir)
 	}
-	f, errs := parser.ParseWithOptions(path, src, parser.Options{Expand: true, IncludeDirs: dirs})
+	f, errs := parser.ParseWithOptions(path, src, parser.Options{Expand: true, IncludeDirs: dirs, Predefines: platformPredefinesFor(string(targetArch), goos)})
 	for _, e := range errs {
 		fmt.Fprintf(os.Stderr, "%s: %v\n", path, e)
 	}
@@ -582,7 +582,7 @@ naming the package.
 		return 1
 	}
 
-	img, err := assembleFile(targetArch, f)
+	img, err := assembleFile(targetArch, f, goos)
 	if err != nil {
 		fmt.Fprintf(os.Stderr, "%s: %v\n", path, err)
 		return 1
@@ -789,11 +789,34 @@ e.g. --map wideCopyAVX2=wideCopyAVX512 pairs the two regardless of suffix.
 	return 1
 }
 
+// platformPredefines mirrors the go command's assembler invocation, which
+// defines GOOS_<goos> and GOARCH_<arch> as -D macros: GOROOT headers
+// (go_tls.h, asm_riscv64.h) select their platform blocks with #ifdef on
+// exactly those names, so an assembler without them cannot see the platform
+// definitions at all.
+func platformPredefines(goarch, goos string) map[string]string {
+	return map[string]string{
+		"GOARCH_" + goarch: "1",
+		"GOOS_" + goos:     "1",
+	}
+}
+
+// platformPredefinesFor resolves the ambient GOOS the way a build would: a
+// file whose name carries one (sys_darwin_arm64.s) is compiled for that GOOS
+// and nothing else.
+func platformPredefinesFor(goarch string, fileGoos string) map[string]string {
+	goos := fileGoos
+	if goos == "" {
+		goos = runtime.GOOS
+	}
+	return platformPredefines(goarch, goos)
+}
+
 // assembleFile assembles a parsed file for the given architecture and returns the image.
-func assembleFile(targetArch arch.Arch, f *ast.File) (*asm.Image, error) {
+func assembleFile(targetArch arch.Arch, f *ast.File, goos string) (*asm.Image, error) {
 	switch targetArch {
 	case arch.AMD64:
-		return asm.AssembleFile(f)
+		return asm.AssembleFile(f, asm.WithGOOS(goos))
 	case arch.RISCV:
 		return asm.AssembleFileRISCV(f)
 	case arch.ARM64:
@@ -813,18 +836,18 @@ func assemblePath(path string, forced arch.Arch, dirs includeDirs) (*asm.Image,
 	if err != nil {
 		return nil, err
 	}
-	f, errs := parser.ParseWithOptions(path, src, parser.Options{Expand: true, IncludeDirs: dirs})
+	target := forced
+	if target == arch.Unknown {
+		target = arch.FromFilename(path)
+	}
+	f, errs := parser.ParseWithOptions(path, src, parser.Options{Expand: true, IncludeDirs: dirs, Predefines: platformPredefinesFor(string(target), "")})
 	for _, e := range errs {
 		fmt.Fprintf(os.Stderr, "%s: %v\n", path, e)
 	}
 	if len(errs) > 0 {
 		return nil, fmt.Errorf("parse errors")
 	}
-	target := forced
-	if target == arch.Unknown {
-		target = arch.FromFilename(path)
-	}
-	return assembleFile(target, f)
+	return assembleFile(target, f, "")
 }
 
 // printByteDiff shows the first few byte differences between two code blocks.
@@ -920,7 +943,7 @@ func cmdVerifyNonJIT(path string, targetArch arch.Arch, groundTruth, profile boo
 	if len(errs) > 0 {
 		return 1
 	}
-	img, err := assembleFile(targetArch, f)
+	img, err := assembleFile(targetArch, f, "")
 	if err != nil {
 		fmt.Fprintf(os.Stderr, "gasm verify: %v\n", err)
 		return 1

cmd/gasm/main_test.go

@@ -14,8 +14,8 @@ import (
 	"syscall"
 	"testing"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/arch"
-	"sourcedock.dev/petrbalvin/gasm-devkit/asm"
+	"sourcedock.dev/petrbalvin/gasm-sdk/arch"
+	"sourcedock.dev/petrbalvin/gasm-sdk/asm"
 )
 
 const clean = "#include \"textflag.h\"\n" +

cmd/gasm/scaffold.go

@@ -11,8 +11,8 @@ import (
 	"os"
 	"strings"
 
-	gasmast "sourcedock.dev/petrbalvin/gasm-devkit/ast"
-	gasmparser "sourcedock.dev/petrbalvin/gasm-devkit/parser"
+	gasmast "sourcedock.dev/petrbalvin/gasm-sdk/ast"
+	gasmparser "sourcedock.dev/petrbalvin/gasm-sdk/parser"
 )
 
 // cmdScaffold generates a differential test skeleton for every kernel in a

debug/breakpoint.go

@@ -1,13 +1,17 @@
 // Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
 // SPDX-License-Identifier: BSD-3-Clause
 
-//go:build linux
+//go:build linux || (freebsd && (amd64 || arm64 || riscv64))
 
 package debug
 
 import "strings"
 
-import "fmt"
+import (
+	"cmp"
+	"fmt"
+	"slices"
+)
 
 // Breakpoint is one software breakpoint in the debuggee.
 type Breakpoint struct {
@@ -149,15 +153,16 @@ func (bm *Breakpoints) SetWithCond(addr uint64, label string, cond *Condition) (
 	return bp, nil
 }
 
-// Info returns a formatted list of all breakpoints.
+// Info returns a formatted list of all breakpoints, ordered by address so
+// the numbering is stable across calls (map iteration order is not).
 func (bm *Breakpoints) Info() string {
 	if len(bm.bps) == 0 {
 		return "no breakpoints set\n"
 	}
 	var result strings.Builder
-	i := 0
-	for _, bp := range bm.bps {
-		i++
+	bps := bm.All()
+	slices.SortFunc(bps, func(a, b *Breakpoint) int { return cmp.Compare(a.Addr, b.Addr) })
+	for i, bp := range bps {
 		status := "enabled"
 		if !bp.Enabled {
 			status = "disabled"
@@ -170,7 +175,7 @@ func (bm *Breakpoints) Info() string {
 		if bp.Cond != nil {
 			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))
+		result.WriteString(fmt.Sprintf("  %d: %s at %#x [%s, %d hits]%s\n", i+1, label, bp.Addr, status, bp.hits, cond))
 	}
 	return result.String()
 }
@@ -238,6 +243,25 @@ func (bm *Breakpoints) All() []*Breakpoint {
 // Hits returns how many times the breakpoint has been hit.
 func (bp *Breakpoint) Hits() int { return bp.hits }
 
+// TrapStray reports whether a stop is a breakpoint-class trap that matches
+// no breakpoint of ours and cannot be resumed: the PC still stands on the
+// trapping instruction (the kernel's own BRK, EBREAK or break, on an
+// architecture that reports the trap in place), so the next resume would
+// re-execute it and trap forever.  trapPC is the PC the stop reported,
+// before any HandleTrap rewinding; reason is the stop's StopInfo class.
+// The debuggee's SIGSTOP barriers also stop without PC movement, and they
+// never carry the breakpoint class, so they are unaffected.
+func TrapStray(s *Session, reason StopReason, trapPC uint64) bool {
+	if reason != StopBreakpoint {
+		return false
+	}
+	after, err := s.GetRegs()
+	if err != nil {
+		return false
+	}
+	return after.GetPC() <= trapPC-uint64(breakpointPCAdjust)
+}
+
 func (bm *Breakpoints) HandleTrap(regs *Regs) *Breakpoint {
 	// On amd64 the kernel reports the trap with RIP past the INT3; on the
 	// other supported architectures the PC still stands on the trap

debug/debug_arch_test.go

@@ -10,6 +10,8 @@ package debug
 // build on every supported linux architecture.
 
 import (
+	"fmt"
+	"slices"
 	"strings"
 	"testing"
 )
@@ -263,3 +265,45 @@ func TestConditionString(t *testing.T) {
 		}
 	}
 }
+
+// TestBreakpointsInfoOrdered proves the listing is ordered by address: the
+// numbers it prints are map keys rendered in iteration order otherwise, so
+// the same set of breakpoints would renumber itself between calls.
+func TestBreakpointsInfoOrdered(t *testing.T) {
+	tr := newMockTracer()
+	bm := NewBreakpoints(tr)
+	addrs := []uint64{0x9000, 0x1000, 0x7000, 0x3000, 0x8000, 0x2000,
+		0x6000, 0x4000, 0x5000, 0xa000}
+	for i, a := range addrs {
+		if _, err := bm.Set(a, fmt.Sprintf("bp%d", i)); err != nil {
+			t.Fatalf("Set(%#x): %v", a, err)
+		}
+	}
+	sorted := append([]uint64(nil), addrs...)
+	slices.Sort(sorted)
+	info := bm.Info()
+	for i, a := range sorted {
+		want := fmt.Sprintf("  %d: bp%d at %#x", i+1, indexOf(addrs, a), a)
+		if !strings.Contains(info, want) {
+			t.Errorf("Info() missing %q; listing:\n%s", want, info)
+		}
+	}
+	// The numbers themselves must ascend: "1:" before "2" ... "10".
+	pos := 0
+	for i := range len(addrs) {
+		next := strings.Index(info[pos:], fmt.Sprintf("  %d: ", i+1))
+		if next < 0 {
+			t.Fatalf("Info() has no entry %d; listing:\n%s", i+1, info)
+		}
+		pos += next
+	}
+}
+
+func indexOf(addrs []uint64, a uint64) int {
+	for i, v := range addrs {
+		if v == a {
+			return i
+		}
+	}
+	return -1
+}

debug/debug_audit_test.go

@@ -0,0 +1,325 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+//go:build linux && amd64
+
+package debug
+
+import (
+	"fmt"
+	"os"
+	"runtime"
+	"strings"
+	"testing"
+	"time"
+)
+
+// Regression tests for the debugger audit: memory access at mapping
+// boundaries, watchpoint slot attribution, launch failure latency, stray
+// trap instructions and the REPL's argument validation.  All drive a real
+// ptrace session, so they run on amd64 hosts only.
+
+// memMap is one line of /proc/pid/maps.
+type memMap struct {
+	lo, hi uint64
+	perms  string
+	name   string
+}
+
+// readMaps parses the debuggee's memory map.
+func readMaps(t *testing.T, pid int) []memMap {
+	t.Helper()
+	data, err := os.ReadFile(fmt.Sprintf("/proc/%d/maps", pid))
+	if err != nil {
+		t.Fatalf("read maps: %v", err)
+	}
+	var out []memMap
+	for line := range strings.SplitSeq(string(data), "\n") {
+		fields := strings.Fields(line)
+		if len(fields) < 2 {
+			continue
+		}
+		var lo, hi uint64
+		if _, err := fmt.Sscanf(fields[0], "%x-%x", &lo, &hi); err != nil {
+			continue
+		}
+		m := memMap{lo: lo, hi: hi, perms: fields[1]}
+		if len(fields) >= 6 {
+			m.name = fields[5]
+		}
+		out = append(out, m)
+	}
+	return out
+}
+
+// boundaryByte returns the last byte of a writable, ordinary mapping that is
+// followed by an unmapped gap: an access there is inside the mapping, while
+// the 8-byte word starting at it crosses into unmapped memory.
+func boundaryByte(t *testing.T, pid int) uint64 {
+	t.Helper()
+	maps := readMaps(t, pid)
+	for i, m := range maps {
+		if !strings.Contains(m.perms, "rw") ||
+			strings.Contains(m.name, "vvar") || strings.Contains(m.name, "vdso") ||
+			strings.Contains(m.name, "vsyscall") {
+			continue
+		}
+		gap := uint64(1) << 62
+		if i+1 < len(maps) {
+			gap = maps[i+1].lo - m.hi
+		}
+		if gap >= 4096 {
+			return m.hi - 1
+		}
+	}
+	t.Skip("no writable mapping followed by a hole; cannot construct the boundary")
+	return 0
+}
+
+// TestReadMemoryPageBoundary proves ReadMemory never reads past the requested
+// range: one byte at the end of a mapping followed by a hole must be
+// readable, which the old word-at-a-time tail read failed because its final
+// 8-byte Peek crossed into the unmapped page.
+func TestReadMemoryPageBoundary(t *testing.T) {
+	sess, _, _ := launchKernel(t, buildGasm(t), boundaryKernel(t), "boundary", nil)
+	addr := boundaryByte(t, sess.Pid())
+	mem, err := sess.ReadMemory(addr, 1)
+	if err != nil {
+		t.Fatalf("ReadMemory(%#x, 1): %v (the read must not cross into the unmapped page)", addr, err)
+	}
+	if len(mem) != 1 {
+		t.Fatalf("ReadMemory returned %d bytes, want 1", len(mem))
+	}
+	// A request whose own range crosses into the hole must still fail.
+	if _, err := sess.ReadMemory(addr, 8); err == nil {
+		t.Fatal("ReadMemory past the mapping end should fail")
+	}
+}
+
+// TestDisassemblePageBoundary proves the disassembler shrinks its read
+// window at a mapping end instead of failing: the instruction stream cannot
+// be decoded at all when the fixed 15-byte read crosses into the hole.
+func TestDisassemblePageBoundary(t *testing.T) {
+	sess, _, _ := launchKernel(t, buildGasm(t), boundaryKernel(t), "boundary", nil)
+	addr := boundaryByte(t, sess.Pid())
+	if _, _, err := sess.Disassemble(addr); err != nil {
+		t.Fatalf("Disassemble(%#x): %v (the read window must shrink at the mapping end)", addr, err)
+	}
+}
+
+// TestWriteMemoryPageBoundary proves WriteMemory writes exactly the bytes it
+// is given: one byte at the end of a mapping followed by a hole must be
+// writable, which the old read-modify-write of the final partial word failed
+// because its Peek crossed into the unmapped page.
+func TestWriteMemoryPageBoundary(t *testing.T) {
+	sess, _, _ := launchKernel(t, buildGasm(t), boundaryKernel(t), "boundary", nil)
+	addr := boundaryByte(t, sess.Pid())
+	orig, err := sess.ReadMemory(addr, 1)
+	if err != nil {
+		t.Fatalf("ReadMemory(%#x, 1): %v", addr, err)
+	}
+	if err := sess.WriteMemory(addr, []byte{orig[0]}); err != nil {
+		t.Fatalf("WriteMemory(%#x, 1): %v (the write must not read past the range)", addr, err)
+	}
+}
+
+// TestWatchpointSlotAttribution proves a hit is attributed to the slot that
+// fired, not to an earlier one whose DR6 status bit is still set: the B0-B3
+// bits are sticky, so they must be acknowledged when read.
+func TestWatchpointSlotAttribution(t *testing.T) {
+	bin := buildGasm(t)
+	const kernel = `#include "textflag.h"
+
+// func wptwo(x, y int64) (a, b int64)
+TEXT ·wptwo(SB), NOSPLIT, $0-32
+	MOVQ $0x1111, AX
+	MOVQ AX, a+16(FP)
+	MOVQ $0x2222, BX
+	MOVQ BX, b+24(FP)
+	RET
+`
+	path := writeKernel(t, kernel)
+	sess, bm, fl := launchKernel(t, bin, path, "wptwo", 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)
+	}
+	// FP sits one word above the entry stack pointer (the return address
+	// occupies [RSP]), so a+16(FP) = RSP+24 and b+24(FP) = RSP+32.
+	watchA := regs.RSP + 24
+	watchB := regs.RSP + 32
+
+	if err := sess.SetWatchpoint(0, watchA, WatchWrite, 8); err != nil {
+		t.Fatalf("SetWatchpoint(0): %v", err)
+	}
+	if err := sess.SetWatchpoint(1, watchB, WatchWrite, 8); err != nil {
+		t.Fatalf("SetWatchpoint(1): %v", err)
+	}
+
+	for i, want := range []uint64{watchA, watchB} {
+		if err := sess.Continue(); err != nil {
+			t.Fatalf("Continue (hit %d): %v", i+1, err)
+		}
+		reason, addr := sess.StopInfo()
+		if reason != StopWatchpoint {
+			t.Fatalf("hit %d: stop reason = %v, want StopWatchpoint", i+1, reason)
+		}
+		if addr != want {
+			t.Fatalf("hit %d reported %#x, want %#x (the sticky DR6 bit misattributes the slot)", i+1, addr, want)
+		}
+	}
+
+	// Clearing a watchpoint must zero its address register: a stale
+	// address in a disabled slot turns any sticky status bit into a
+	// misattributed report later.
+	if err := sess.ClearWatchpoint(0); err != nil {
+		t.Fatalf("ClearWatchpoint(0): %v", err)
+	}
+	dr0, err := ptracePeekUser(sess.Pid(), drOffset)
+	if err != nil {
+		t.Fatalf("read DR0: %v", err)
+	}
+	if dr0 != 0 {
+		t.Fatalf("DR0 = %#x after ClearWatchpoint, want 0 (the address register must be cleared)", dr0)
+	}
+}
+
+// TestLaunchFailsFastOnDeadDebuggee proves a debuggee that dies before
+// signalling readiness surfaces promptly: the ready poll used to run its
+// full 2.5 seconds before the wait discovered the exit.
+func TestLaunchFailsFastOnDeadDebuggee(t *testing.T) {
+	runtime.LockOSThread()
+	defer runtime.UnlockOSThread()
+	bin := buildGasm(t)
+	path := boundaryKernel(t)
+	start := time.Now()
+	sess, err := Launch(bin, path, "nosuchfunction", nil)
+	elapsed := time.Since(start)
+	if err == nil {
+		sess.Kill()
+		t.Fatal("Launch with an unknown function should fail")
+	}
+	if !strings.Contains(err.Error(), "before signalling readiness") &&
+		!strings.Contains(err.Error(), "debuggee exited") {
+		t.Errorf("error does not name the dead debuggee: %v", err)
+	}
+	if elapsed >= 1500*time.Millisecond {
+		t.Fatalf("Launch took %v to report the dead debuggee; the readiness poll must detect the exit, not time out", elapsed)
+	}
+}
+
+// TestStrayTrapRunsThrough proves the continue loop survives a trap
+// instruction planted in the kernel itself (BYTE $0xCC, the same byte the
+// debugger patches in): on architectures that report the trap in place the
+// loop must surface the stop, and on amd64 it runs through to the exit.  A
+// regression here hangs, so a watchdog fails the run.
+func TestStrayTrapRunsThrough(t *testing.T) {
+	bin := buildGasm(t)
+	const kernel = `#include "textflag.h"
+
+// func stray() int64
+TEXT ·stray(SB), NOSPLIT, $0-8
+	MOVQ $7, AX
+	BYTE $0xCC
+	MOVQ AX, ret+0(FP)
+	RET
+`
+	path := writeKernel(t, kernel)
+	sess, bm, _ := launchKernel(t, bin, path, "stray", nil)
+
+	timer := time.AfterFunc(time.Minute, func() {
+		panic("watchdog: the continue loop hung on the stray trap instruction")
+	})
+	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, "debuggee exited") {
+		t.Errorf("the stray trap wedged the continue loop; output:\n%s", out)
+	}
+}
+
+// TestStepIntoFaultReportsSignal proves the step command reports a genuine
+// signal-delivery-stop instead of silently printing the faulting
+// instruction as if the step had succeeded.
+func TestStepIntoFaultReportsSignal(t *testing.T) {
+	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, fl := launchKernel(t, bin, path, "crash", 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)
+
+	out := captureStdout(t, func() {
+		REPL(sess, bm, sess.CodeBase(), fl.Offset, fl.Size, fl.Args, nil, nil,
+			strings.NewReader("step 2\nquit\n"))
+	})
+	if !strings.Contains(out, "stopped on signal") {
+		t.Errorf("stepping into the fault did not report the signal; output:\n%s", out)
+	}
+}
+
+// TestREPLRejectsBadArguments proves the command loop reports malformed
+// input instead of silently defaulting: an unknown label for x would read
+// address 0, and a malformed count would silently step one instruction.
+func TestREPLRejectsBadArguments(t *testing.T) {
+	bin := buildGasm(t)
+	path := boundaryKernel(t)
+	sess, bm, fl := launchKernel(t, bin, path, "boundary", 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)
+
+	out := captureStdout(t, func() {
+		REPL(sess, bm, sess.CodeBase(), fl.Offset, fl.Size, fl.Args, nil, nil,
+			strings.NewReader("x nosuchlabel\nstep abc\ndisas abc\nwatch 0x1000 q 8\nquit\n"))
+	})
+	for _, want := range []string{
+		"unknown address: nosuchlabel",
+		"invalid count: abc",
+		"unknown watchpoint type: q",
+	} {
+		if !strings.Contains(out, want) {
+			t.Errorf("output missing %q:\n%s", want, out)
+		}
+	}
+	if got := strings.Count(out, "invalid count: abc"); got != 2 {
+		t.Errorf("invalid count reported %d times, want 2 (step and disas):\n%s", got, out)
+	}
+}
+
+// boundaryKernel is a minimal kernel for the boundary tests, which only need
+// a live, stopped debuggee.
+func boundaryKernel(t *testing.T) string {
+	t.Helper()
+	const kernel = `#include "textflag.h"
+
+// func boundary() int64
+TEXT ·boundary(SB), NOSPLIT, $0-8
+	MOVQ $1, AX
+	MOVQ AX, ret+0(FP)
+	RET
+`
+	return writeKernel(t, kernel)
+}

debug/disasm_freebsd_amd64.go

@@ -0,0 +1,65 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+//go:build freebsd && amd64
+
+package debug
+
+import (
+	"fmt"
+	"strings"
+
+	"sourcedock.dev/petrbalvin/gasm-sdk/arch"
+	"sourcedock.dev/petrbalvin/gasm-sdk/disasm"
+)
+
+// Disassemble decodes the instruction at the given address in the debuggee's
+// memory and returns its text representation and length in bytes.  An amd64
+// instruction is up to 15 bytes long, but the read must not reach past the
+// end of the mapping: when the full 15-byte window crosses into unmapped
+// memory the window shrinks, because an instruction at the mapping's end is
+// by construction no longer than the readable bytes that hold it.
+func (s *Session) Disassemble(addr uint64) (string, int, error) {
+	var lastErr error
+	for _, n := range []int{15, 8, 4, 2, 1} {
+		mem, err := s.ReadMemory(addr, n)
+		if err != nil {
+			lastErr = err
+			continue
+		}
+		ins, derr := disasm.Decode(arch.AMD64, mem, addr)
+		if derr != nil {
+			return "", 0, derr
+		}
+		return ins.Text, ins.Len, nil
+	}
+	return "", 0, lastErr
+}
+
+// DisassembleN decodes up to n instructions starting at addr and returns
+// them as a formatted string with addresses and byte offsets.
+func (s *Session) DisassembleN(addr uint64, n int) string {
+	var result strings.Builder
+	pc := addr
+	for range n {
+		text, length, err := s.Disassemble(pc)
+		if err != nil {
+			result.WriteString(fmt.Sprintf("  %#08x: <error: %v>\n", pc, err))
+			break
+		}
+		result.WriteString(fmt.Sprintf("  %#08x: %s\n", pc, text))
+		if length == 0 {
+			length = 1
+		}
+		pc += uint64(length)
+	}
+	return result.String()
+}
+
+// isCallInsn reports whether disassembled text (x86asm.IntelSyntax) is a
+// call.  The first token must match exactly: a prefix test would also catch
+// unrelated mnemonics.
+func isCallInsn(text string) bool {
+	m, _, _ := strings.Cut(text, " ")
+	return strings.ToLower(m) == "call"
+}

debug/disasm_freebsd_arm64.go

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

debug/disasm_freebsd_riscv64.go

@@ -0,0 +1,70 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+//go:build freebsd && riscv64
+
+package debug
+
+import (
+	"fmt"
+	"strings"
+
+	"sourcedock.dev/petrbalvin/gasm-sdk/arch"
+	"sourcedock.dev/petrbalvin/gasm-sdk/disasm"
+)
+
+// Disassemble decodes the instruction at the given address in the debuggee's
+// memory and returns its text representation and length in bytes.  The read
+// shrinks from 4 to 2 bytes when the full word crosses into unmapped memory:
+// a compressed instruction at the mapping's end still fits the shorter
+// window, and an instruction can never extend past the mapping that holds it.
+func (s *Session) Disassemble(addr uint64) (string, int, error) {
+	var lastErr error
+	for _, n := range []int{4, 2} {
+		mem, err := s.ReadMemory(addr, n)
+		if err != nil {
+			lastErr = err
+			continue
+		}
+		ins, derr := disasm.Decode(arch.RISCV, mem, addr)
+		if derr != nil {
+			return "", 0, derr
+		}
+		return ins.Text, ins.Len, nil
+	}
+	return "", 0, lastErr
+}
+
+// DisassembleN decodes up to n instructions starting at addr and returns
+// them as a formatted string with addresses and byte offsets.
+func (s *Session) DisassembleN(addr uint64, n int) string {
+	var result strings.Builder
+	pc := addr
+	for range n {
+		text, length, err := s.Disassemble(pc)
+		if err != nil {
+			result.WriteString(fmt.Sprintf("  %#08x: <error: %v>\n", pc, err))
+			break
+		}
+		result.WriteString(fmt.Sprintf("  %#08x: %s\n", pc, text))
+		if length == 0 {
+			length = 1
+		}
+		pc += uint64(length)
+	}
+	return result.String()
+}
+
+// isCallInsn reports whether disassembled text (riscv64asm.GoSyntax) is a
+// call.  GoSyntax renders jal and jalr calls as CALL; the native mnemonics
+// are accepted too.  The first token must match exactly: a prefix test on
+// "bl" would catch branches on other architectures, and jalr as ret prints
+// RET, which must not be stepped over.
+func isCallInsn(text string) bool {
+	m, _, _ := strings.Cut(text, " ")
+	switch strings.ToLower(m) {
+	case "call", "jal", "jalr":
+		return true
+	}
+	return false
+}

debug/disasm_linux_amd64.go

@@ -9,22 +9,31 @@ import (
 	"fmt"
 	"strings"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/arch"
-	"sourcedock.dev/petrbalvin/gasm-devkit/disasm"
+	"sourcedock.dev/petrbalvin/gasm-sdk/arch"
+	"sourcedock.dev/petrbalvin/gasm-sdk/disasm"
 )
 
 // Disassemble decodes the instruction at the given address in the debuggee's
-// memory and returns its text representation and length in bytes.
+// memory and returns its text representation and length in bytes.  An amd64
+// instruction is up to 15 bytes long, but the read must not reach past the
+// end of the mapping: when the full 15-byte window crosses into unmapped
+// memory the window shrinks, because an instruction at the mapping's end is
+// by construction no longer than the readable bytes that hold it.
 func (s *Session) Disassemble(addr uint64) (string, int, error) {
-	mem, err := s.ReadMemory(addr, 15)
+	var lastErr error
+	for _, n := range []int{15, 8, 4, 2, 1} {
+		mem, err := s.ReadMemory(addr, n)
 		if err != nil {
-		return "", 0, err
+			lastErr = err
+			continue
 		}
-	ins, err := disasm.Decode(arch.AMD64, mem, addr)
-	if err != nil {
-		return "", 0, err
+		ins, derr := disasm.Decode(arch.AMD64, mem, addr)
+		if derr != nil {
+			return "", 0, derr
 		}
 		return ins.Text, ins.Len, nil
+	}
+	return "", 0, lastErr
 }
 
 // DisassembleN decodes up to n instructions starting at addr and returns

debug/disasm_linux_arm64.go

@@ -9,8 +9,8 @@ import (
 	"fmt"
 	"strings"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/arch"
-	"sourcedock.dev/petrbalvin/gasm-devkit/disasm"
+	"sourcedock.dev/petrbalvin/gasm-sdk/arch"
+	"sourcedock.dev/petrbalvin/gasm-sdk/disasm"
 )
 
 // Disassemble decodes the instruction at the given address in the debuggee's

debug/disasm_linux_loong64.go

@@ -9,8 +9,8 @@ import (
 	"fmt"
 	"strings"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/arch"
-	"sourcedock.dev/petrbalvin/gasm-devkit/disasm"
+	"sourcedock.dev/petrbalvin/gasm-sdk/arch"
+	"sourcedock.dev/petrbalvin/gasm-sdk/disasm"
 )
 
 // Disassemble decodes the instruction at the given address in the debuggee's

debug/disasm_linux_riscv64.go

@@ -9,22 +9,30 @@ import (
 	"fmt"
 	"strings"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/arch"
-	"sourcedock.dev/petrbalvin/gasm-devkit/disasm"
+	"sourcedock.dev/petrbalvin/gasm-sdk/arch"
+	"sourcedock.dev/petrbalvin/gasm-sdk/disasm"
 )
 
 // Disassemble decodes the instruction at the given address in the debuggee's
-// memory and returns its text representation and length in bytes.
+// memory and returns its text representation and length in bytes.  The read
+// shrinks from 4 to 2 bytes when the full word crosses into unmapped memory:
+// a compressed instruction at the mapping's end still fits the shorter
+// window, and an instruction can never extend past the mapping that holds it.
 func (s *Session) Disassemble(addr uint64) (string, int, error) {
-	mem, err := s.ReadMemory(addr, 4)
+	var lastErr error
+	for _, n := range []int{4, 2} {
+		mem, err := s.ReadMemory(addr, n)
 		if err != nil {
-		return "", 0, err
+			lastErr = err
+			continue
 		}
-	ins, err := disasm.Decode(arch.RISCV, mem, addr)
-	if err != nil {
-		return "", 0, err
+		ins, derr := disasm.Decode(arch.RISCV, mem, addr)
+		if derr != nil {
+			return "", 0, derr
 		}
 		return ins.Text, ins.Len, nil
+	}
+	return "", 0, lastErr
 }
 
 // DisassembleN decodes up to n instructions starting at addr.

debug/display_freebsd_amd64.go

@@ -0,0 +1,136 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+//go:build freebsd && amd64
+
+package debug
+
+import "fmt"
+
+func printRegs(regs *Regs, codeBase, funcOff uint64) {
+	fmt.Printf("  RIP = %#016x  (func+%#x)\n", regs.RIP, regs.RIP-codeBase-funcOff)
+	fmt.Printf("  RSP = %#016x  RBP = %#016x\n", regs.RSP, regs.RBP)
+	fmt.Printf("  RAX = %#016x  RBX = %#016x\n", regs.RAX, regs.RBX)
+	fmt.Printf("  RCX = %#016x  RDX = %#016x\n", regs.RCX, regs.RDX)
+	fmt.Printf("  RSI = %#016x  RDI = %#016x\n", regs.RSI, regs.RDI)
+	fmt.Printf("  R8  = %#016x  R9  = %#016x\n", regs.R8, regs.R9)
+	fmt.Printf("  R10 = %#016x  R11 = %#016x\n", regs.R10, regs.R11)
+	fmt.Printf("  R12 = %#016x  R13 = %#016x\n", regs.R12, regs.R13)
+	fmt.Printf("  R14 = %#016x  R15 = %#016x\n", regs.R14, regs.R15)
+	fmt.Printf("  RFLAGS = %#x  [%s]\n", regs.RFLAGS, decodeRflags(regs.RFLAGS))
+}
+
+func printVectorRegs(v *VectorRegs) {
+	fmt.Println("\n  Vector registers (YMM):")
+	for i := 0; i < 16; i += 2 {
+		fmt.Printf("  YMM%-2d = ", i)
+		printYMM(v.YMM[i][:])
+		fmt.Printf("  YMM%-2d = ", i+1)
+		printYMM(v.YMM[i+1][:])
+		fmt.Println()
+	}
+}
+
+func printYMM(b []byte) {
+	for j := 0; j < 32; j += 4 {
+		v := uint32(b[j]) | uint32(b[j+1])<<8 | uint32(b[j+2])<<16 | uint32(b[j+3])<<24
+		fmt.Printf("%08x ", v)
+	}
+}
+
+func decodeRflags(f uint64) string {
+	var flags string
+	if f&1 != 0 {
+		flags += "CF "
+	}
+	if f&(1<<2) != 0 {
+		flags += "PF "
+	}
+	if f&(1<<4) != 0 {
+		flags += "AF "
+	}
+	if f&(1<<6) != 0 {
+		flags += "ZF "
+	}
+	if f&(1<<7) != 0 {
+		flags += "SF "
+	}
+	if f&(1<<8) != 0 {
+		flags += "TF "
+	}
+	if f&(1<<9) != 0 {
+		flags += "IF "
+	}
+	if f&(1<<10) != 0 {
+		flags += "DF "
+	}
+	if f&(1<<11) != 0 {
+		flags += "OF "
+	}
+	if flags == "" {
+		return "none"
+	}
+	return flags[:len(flags)-1]
+}
+
+// SetReg modifies a register value in the debuggee.
+func (s *Session) SetReg(name string, value uint64) error {
+	regs, err := s.GetRegs()
+	if err != nil {
+		return err
+	}
+	switch name {
+	case "rax", "eax", "ax", "al":
+		regs.RAX = value
+	case "rbx", "ebx", "bx", "bl":
+		regs.RBX = value
+	case "rcx", "ecx", "cx", "cl":
+		regs.RCX = value
+	case "rdx", "edx", "dx", "dl":
+		regs.RDX = value
+	case "rsi", "esi", "si":
+		regs.RSI = value
+	case "rdi", "edi", "di":
+		regs.RDI = value
+	case "rbp", "ebp", "bp":
+		regs.RBP = value
+	case "rsp", "esp", "sp":
+		regs.RSP = value
+	case "r8":
+		regs.R8 = value
+	case "r9":
+		regs.R9 = value
+	case "r10":
+		regs.R10 = value
+	case "r11":
+		regs.R11 = value
+	case "r12":
+		regs.R12 = value
+	case "r13":
+		regs.R13 = value
+	case "r14":
+		regs.R14 = value
+	case "r15":
+		regs.R15 = value
+	case "rip", "eip":
+		regs.RIP = value
+	default:
+		return fmt.Errorf("debug: unknown register %q", name)
+	}
+	return s.SetRegs(&regs)
+}
+
+// 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.  FreeBSD exposes no mapping list, so the
+// walk degenerates to the raw entry convention, [SP] before any push.
+func archReturnAddr(s *Session, regs *Regs) (uint64, error) {
+	return s.Peek(regs.RSP)
+}
+
+// archSPLabel returns the SP register name for display.
+func archSPLabel() string { return "RSP" }

debug/display_freebsd_arm64.go

@@ -0,0 +1,127 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+//go:build freebsd && arm64
+
+package debug
+
+import (
+	"encoding/binary"
+	"fmt"
+)
+
+func printRegs(regs *Regs, codeBase, funcOff uint64) {
+	fmt.Printf("  PC  = %#016x  (func+%#x)\n", regs.PC, regs.PC-codeBase-funcOff)
+	fmt.Printf("  SP  = %#016x  FP  = %#016x\n", regs.SP, regs.X29)
+	fmt.Printf("  LR  = %#016x\n", regs.X30)
+	fmt.Printf("  X0  = %#016x  X1  = %#016x\n", regs.X0, regs.X1)
+	fmt.Printf("  X2  = %#016x  X3  = %#016x\n", regs.X2, regs.X3)
+	fmt.Printf("  X4  = %#016x  X5  = %#016x\n", regs.X4, regs.X5)
+	fmt.Printf("  X6  = %#016x  X7  = %#016x\n", regs.X6, regs.X7)
+	fmt.Printf("  X8  = %#016x  X9  = %#016x\n", regs.X8, regs.X9)
+	fmt.Printf("  X10 = %#016x  X11 = %#016x\n", regs.X10, regs.X11)
+	fmt.Printf("  X12 = %#016x  X13 = %#016x\n", regs.X12, regs.X13)
+	fmt.Printf("  X14 = %#016x  X15 = %#016x\n", regs.X14, regs.X15)
+	fmt.Printf("  X16 = %#016x  X17 = %#016x\n", regs.X16, regs.X17)
+	fmt.Printf("  X18 = %#016x  X19 = %#016x\n", regs.X18, regs.X19)
+	fmt.Printf("  X20 = %#016x  X21 = %#016x\n", regs.X20, regs.X21)
+	fmt.Printf("  X22 = %#016x  X23 = %#016x\n", regs.X22, regs.X23)
+	fmt.Printf("  X24 = %#016x  X25 = %#016x\n", regs.X24, regs.X25)
+	fmt.Printf("  X26 = %#016x  X27 = %#016x\n", regs.X26, regs.X27)
+	fmt.Printf("  X28 = %#016x  PSTATE = %#x\n", regs.X28, regs.PSTATE)
+}
+
+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, 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]))
+	}
+}
+
+// SetReg modifies a register value in the debuggee.
+func (s *Session) SetReg(name string, value uint64) error {
+	regs, err := s.GetRegs()
+	if err != nil {
+		return err
+	}
+	switch name {
+	case "x0":
+		regs.X0 = value
+	case "x1":
+		regs.X1 = value
+	case "x2":
+		regs.X2 = value
+	case "x3":
+		regs.X3 = value
+	case "x4":
+		regs.X4 = value
+	case "x5":
+		regs.X5 = value
+	case "x6":
+		regs.X6 = value
+	case "x7":
+		regs.X7 = value
+	case "x8":
+		regs.X8 = value
+	case "x9":
+		regs.X9 = value
+	case "x10":
+		regs.X10 = value
+	case "x11":
+		regs.X11 = value
+	case "x12":
+		regs.X12 = value
+	case "x13":
+		regs.X13 = value
+	case "x14":
+		regs.X14 = value
+	case "x15":
+		regs.X15 = value
+	case "x16":
+		regs.X16 = value
+	case "x17":
+		regs.X17 = value
+	case "x18":
+		regs.X18 = value
+	case "x19":
+		regs.X19 = value
+	case "x20":
+		regs.X20 = value
+	case "x21":
+		regs.X21 = value
+	case "x22":
+		regs.X22 = value
+	case "x23":
+		regs.X23 = value
+	case "x24":
+		regs.X24 = value
+	case "x25":
+		regs.X25 = value
+	case "x26":
+		regs.X26 = value
+	case "x27":
+		regs.X27 = value
+	case "x28":
+		regs.X28 = value
+	case "x29", "fp":
+		regs.X29 = value
+	case "x30", "lr":
+		regs.X30 = value
+	case "sp":
+		regs.SP = value
+	case "pc":
+		regs.PC = value
+	default:
+		return fmt.Errorf("debug: unknown register %q", name)
+	}
+	return s.SetRegs(&regs)
+}
+
+// archReturnAddr reads the return address from LR (arm64 convention).
+func archReturnAddr(s *Session, regs *Regs) (uint64, error) {
+	return regs.X30, nil
+}
+
+// archSPLabel returns the SP register name for display.
+func archSPLabel() string { return "SP" }

debug/display_freebsd_riscv64.go

@@ -0,0 +1,121 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+//go:build freebsd && riscv64
+
+package debug
+
+import "fmt"
+
+func printRegs(regs *Regs, codeBase, funcOff uint64) {
+	fmt.Printf("  PC  = %#016x  (func+%#x)\n", regs.PC, regs.PC-codeBase-funcOff)
+	fmt.Printf("  SP  = %#016x  FP  = %#016x\n", regs.Sp, regs.S0)
+	fmt.Printf("  RA  = %#016x\n", regs.Ra)
+	fmt.Printf("  A0  = %#016x  A1  = %#016x\n", regs.A0, regs.A1)
+	fmt.Printf("  A2  = %#016x  A3  = %#016x\n", regs.A2, regs.A3)
+	fmt.Printf("  A4  = %#016x  A5  = %#016x\n", regs.A4, regs.A5)
+	fmt.Printf("  A6  = %#016x  A7  = %#016x\n", regs.A6, regs.A7)
+	fmt.Printf("  T0  = %#016x  T1  = %#016x\n", regs.T0, regs.T1)
+	fmt.Printf("  T2  = %#016x  T3  = %#016x\n", regs.T2, regs.T3)
+	fmt.Printf("  T4  = %#016x  T5  = %#016x\n", regs.T4, regs.T5)
+	fmt.Printf("  T6  = %#016x\n", regs.T6)
+	fmt.Printf("  S1  = %#016x  S2  = %#016x\n", regs.S1, regs.S2)
+	fmt.Printf("  S3  = %#016x  S4  = %#016x\n", regs.S3, regs.S4)
+	fmt.Printf("  S5  = %#016x  S6  = %#016x\n", regs.S5, regs.S6)
+	fmt.Printf("  S7  = %#016x  S8  = %#016x\n", regs.S7, regs.S8)
+	fmt.Printf("  S9  = %#016x  S10 = %#016x\n", regs.S9, regs.S10)
+	fmt.Printf("  S11 = %#016x\n", regs.S11)
+}
+
+func printVectorRegs(v *VectorRegs) {
+	fmt.Println("\n  FP registers (F0-F31):")
+	for i := 0; i < 32; i += 2 {
+		fmt.Printf("  F%-2d  = %#018x  F%-2d  = %#018x\n", i, v.F[i], i+1, v.F[i+1])
+	}
+	fmt.Printf("  FCSR = %#x\n", v.FCSR)
+}
+
+// SetReg modifies a register value in the debuggee.
+func (s *Session) SetReg(name string, value uint64) error {
+	regs, err := s.GetRegs()
+	if err != nil {
+		return err
+	}
+	switch name {
+	case "pc":
+		regs.PC = value
+	case "ra", "x1":
+		regs.Ra = value
+	case "sp", "x2":
+		regs.Sp = value
+	case "gp", "x3":
+		regs.Gp = value
+	case "tp", "x4":
+		regs.Tp = value
+	case "t0", "x5":
+		regs.T0 = value
+	case "t1", "x6":
+		regs.T1 = value
+	case "t2", "x7":
+		regs.T2 = value
+	case "s0", "fp", "x8":
+		regs.S0 = value
+	case "s1", "x9":
+		regs.S1 = value
+	case "a0", "x10":
+		regs.A0 = value
+	case "a1", "x11":
+		regs.A1 = value
+	case "a2", "x12":
+		regs.A2 = value
+	case "a3", "x13":
+		regs.A3 = value
+	case "a4", "x14":
+		regs.A4 = value
+	case "a5", "x15":
+		regs.A5 = value
+	case "a6", "x16":
+		regs.A6 = value
+	case "a7", "x17":
+		regs.A7 = value
+	case "s2", "x18":
+		regs.S2 = value
+	case "s3", "x19":
+		regs.S3 = value
+	case "s4", "x20":
+		regs.S4 = value
+	case "s5", "x21":
+		regs.S5 = value
+	case "s6", "x22":
+		regs.S6 = value
+	case "s7", "x23":
+		regs.S7 = value
+	case "s8", "x24":
+		regs.S8 = value
+	case "s9", "x25":
+		regs.S9 = value
+	case "s10", "x26":
+		regs.S10 = value
+	case "s11", "x27":
+		regs.S11 = value
+	case "t3", "x28":
+		regs.T3 = value
+	case "t4", "x29":
+		regs.T4 = value
+	case "t5", "x30":
+		regs.T5 = value
+	case "t6", "x31":
+		regs.T6 = value
+	default:
+		return fmt.Errorf("debug: unknown register %q", name)
+	}
+	return s.SetRegs(&regs)
+}
+
+// archReturnAddr reads the return address from RA (riscv64 convention).
+func archReturnAddr(s *Session, regs *Regs) (uint64, error) {
+	return regs.Ra, nil
+}
+
+// archSPLabel returns the SP register name for display.
+func archSPLabel() string { return "SP" }

debug/ptrace_extra_test.go

@@ -17,8 +17,8 @@ import (
 	"time"
 	"unsafe"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/asm"
-	"sourcedock.dev/petrbalvin/gasm-devkit/verify"
+	"sourcedock.dev/petrbalvin/gasm-sdk/asm"
+	"sourcedock.dev/petrbalvin/gasm-sdk/verify"
 )
 
 // Integration tests beyond the basic entry breakpoint: hardware watchpoints,

debug/ptrace_freebsd.go

@@ -0,0 +1,310 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+//go:build freebsd && (amd64 || arm64 || riscv64)
+
+package debug
+
+import (
+	"fmt"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"runtime"
+	"strings"
+	"syscall"
+	"time"
+
+	"golang.org/x/sys/unix"
+)
+
+// Session is a ptrace debugging session controlling one debuggee process.
+// The FreeBSD implementation sits behind the same surface as the Linux one:
+// PT_TRACE_ME from the debuggee, PT_CONTINUE/PT_STEP from the tracer, and
+// tracee memory through PT_IO (FreeBSD has no /proc/pid/mem to fall back
+// on, so PT_IO is the only supported route).
+type Session struct {
+	pid      int
+	cmd      *exec.Cmd
+	stopped  bool
+	exited   bool
+	codeBase uint64   // base address of the JIT code in the debuggee
+	tmpDir   string   // scratch directory of the session, removed on Kill
+	wpSlots  [16]bool // hardware watchpoint slots in use (DR0-DR3, arm64 dbw 0-15)
+	// 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
+// attaches to it via ptrace.
+func Launch(gasmBin, asmPath, funcName string, args []byte) (*Session, error) {
+	sess, _, err := LaunchWithBuffers(gasmBin, asmPath, funcName, args, "")
+	return sess, err
+}
+
+// LaunchWithBuffers is like Launch but also allocates buffers in the debuggee.
+//
+// It pins the calling goroutine to its OS thread and leaves it pinned: the
+// debuggee's PT_TRACE_ME binds the tracer relation to the forking thread,
+// and every ptrace request on the session must come from that same thread.
+// All Session methods must therefore be called from the goroutine that
+// launched the session (the REPL and coverage loops do exactly that).
+func LaunchWithBuffers(gasmBin, asmPath, funcName string, args []byte, bufSpec string) (*Session, []uint64, error) {
+	runtime.LockOSThread() // ptrace requests must stay on the forking thread
+	self, err := os.Executable()
+	if err != nil {
+		return nil, nil, fmt.Errorf("debug: cannot find gasm binary: %w", err)
+	}
+	if gasmBin != "" {
+		self = gasmBin
+	}
+
+	tmpDir, err := os.MkdirTemp("", "gasm-debug-*")
+	if err != nil {
+		return nil, nil, fmt.Errorf("debug: tempdir: %w", err)
+	}
+	argsFile := filepath.Join(tmpDir, "args.bin")
+	if err := os.WriteFile(argsFile, args, 0o644); err != nil {
+		os.RemoveAll(tmpDir)
+		return nil, nil, fmt.Errorf("debug: write args: %w", err)
+	}
+
+	if bufSpec != "" {
+		if err := os.WriteFile(filepath.Join(tmpDir, "bufspec"), []byte(bufSpec), 0o644); err != nil {
+			os.RemoveAll(tmpDir)
+			return nil, nil, fmt.Errorf("debug: write bufspec: %w", err)
+		}
+	}
+
+	cmd := exec.Command(self, "debug", "--func", funcName, "--args", argsFile, asmPath)
+	cmd.Env = append(os.Environ(), "GASM_DEBUG_TARGET=1", "GASM_DEBUG_TMP="+tmpDir)
+	cmd.Stdout = nil
+	cmd.Stderr = os.Stderr
+	cmd.SysProcAttr = &syscall.SysProcAttr{}
+
+	if err := cmd.Start(); err != nil {
+		os.RemoveAll(tmpDir)
+		return nil, nil, fmt.Errorf("debug: start debuggee: %w", err)
+	}
+
+	s := &Session{pid: cmd.Process.Pid, cmd: cmd, tmpDir: tmpDir}
+
+	readyFile := filepath.Join(tmpDir, "ready")
+	for range 500 {
+		if _, err := os.Stat(readyFile); err == nil {
+			break
+		}
+		// A debuggee that died before signalling readiness (unknown
+		// function, unparseable source) writes its failure notice to the
+		// handshake directory; read it and fail fast.  The poll never
+		// waits on the child: a wait here could consume the SIGSTOP park
+		// that waitStopped below must receive, hanging the launch.
+		if err := s.deadReason(); err != nil {
+			cmd.Wait()
+			os.RemoveAll(tmpDir)
+			return nil, nil, err
+		}
+		time.Sleep(5 * time.Millisecond)
+	}
+
+	// The debuggee parks itself with SIGSTOP once the JIT code is mapped.
+	// A Go tracee also reports SIGURG preemption as signal-delivery-stops,
+	// so the wait loops until a stop the debugger cares about instead of
+	// assuming the first event is the SIGSTOP.
+	if _, err := s.waitStopped(); err != nil {
+		cmd.Process.Kill()
+		os.RemoveAll(tmpDir)
+		return nil, nil, fmt.Errorf("debug: wait for debuggee: %w", err)
+	}
+	s.stopped = true
+
+	// The debuggee reports its JIT mapping in the codebase file; that is
+	// the supported path on FreeBSD, where no /proc/pid/maps exists to
+	// scan for the RWX region as a fallback.
+	if data, err := os.ReadFile(filepath.Join(tmpDir, "codebase")); err == nil {
+		fmt.Sscanf(string(data), "%d", &s.codeBase)
+	}
+
+	var bufAddrs []uint64
+	if bufSpec != "" {
+		addrFile := filepath.Join(tmpDir, "bufaddrs")
+		if data, err := os.ReadFile(addrFile); err == nil {
+			for line := range strings.SplitSeq(strings.TrimSpace(string(data)), "\n") {
+				var addr uint64
+				if _, err := fmt.Sscanf(line, "%d", &addr); err == nil {
+					bufAddrs = append(bufAddrs, addr)
+				}
+			}
+		}
+	}
+
+	return s, bufAddrs, nil
+}
+
+// deadReason reports the debuggee's own failure notice, the file its
+// failure paths write before exiting.  A debuggee killed without a notice
+// (a crash, SIGKILL) surfaces through waitStopped after the poll instead,
+// which is why the poll's budget stays finite.
+func (s *Session) deadReason() error {
+	data, err := os.ReadFile(filepath.Join(s.tmpDir, "dead"))
+	if err != nil {
+		return nil
+	}
+	return fmt.Errorf("debug: debuggee failed before signalling readiness: %s", strings.TrimSpace(string(data)))
+}
+
+// waitStopped consumes ptrace-stop events until one the debugger cares
+// 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.
+func (s *Session) waitStopped() (syscall.Signal, error) {
+	for {
+		var ws syscall.WaitStatus
+		if _, err := syscall.Wait4(s.pid, &ws, syscall.WUNTRACED, nil); err != nil {
+			return 0, err
+		}
+		if ws.Exited() {
+			s.exited = true
+			return 0, fmt.Errorf("debuggee exited with status %d", ws.ExitStatus())
+		}
+		if ws.Signaled() {
+			s.exited = true
+			return 0, fmt.Errorf("debuggee killed by signal %v", ws.Signal())
+		}
+		switch sig := ws.StopSignal(); sig {
+		case syscall.SIGTRAP, syscall.SIGSTOP:
+			s.stopped = true
+			s.lastSignal = 0
+			return sig, nil
+		case syscall.SIGURG:
+			// Go runtime asynchronous preemption: resume the tracee
+			// without delivering the signal.
+			s.lastSignal = 0
+			if err := unix.PtraceCont(s.pid, 0); err != nil {
+				return 0, fmt.Errorf("debug: PT_CONTINUE: %w", err)
+			}
+		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, through
+// PT_IO with PIOD_READ_D.
+func (s *Session) Peek(addr uint64) (uint64, error) {
+	var buf [8]byte
+	if _, err := unix.PtraceIO(unix.PIOD_READ_D, s.pid, uintptr(addr), buf[:], len(buf)); err != nil {
+		return 0, fmt.Errorf("debug: read mem %#x: %w", addr, err)
+	}
+	return uint64(buf[0]) | uint64(buf[1])<<8 | uint64(buf[2])<<16 | uint64(buf[3])<<24 |
+		uint64(buf[4])<<32 | uint64(buf[5])<<40 | uint64(buf[6])<<48 | uint64(buf[7])<<56, nil
+}
+
+// Poke writes a word (8 bytes) to the debuggee's memory at addr, through
+// PT_IO with PIOD_WRITE_D.
+func (s *Session) Poke(addr, val uint64) error {
+	buf := []byte{byte(val), byte(val >> 8), byte(val >> 16), byte(val >> 24),
+		byte(val >> 32), byte(val >> 40), byte(val >> 48), byte(val >> 56)}
+	if _, err := unix.PtraceIO(unix.PIOD_WRITE_D, s.pid, uintptr(addr), buf, len(buf)); err != nil {
+		return fmt.Errorf("debug: write mem %#x: %w", addr, err)
+	}
+	return nil
+}
+
+// ReadMemory reads len bytes from the debuggee's memory at addr in one
+// PT_IO request, the shape the request is built for.
+func (s *Session) ReadMemory(addr uint64, length int) ([]byte, error) {
+	out := make([]byte, length)
+	n, err := unix.PtraceIO(unix.PIOD_READ_D, s.pid, uintptr(addr), out, length)
+	return out[:n], err
+}
+
+// WriteMemory writes bytes to the debuggee's memory at addr in one PT_IO
+// request.
+func (s *Session) WriteMemory(addr uint64, data []byte) error {
+	_, err := unix.PtraceIO(unix.PIOD_WRITE_D, s.pid, uintptr(addr), data, len(data))
+	return err
+}
+
+// Step executes a single instruction in the debuggee.
+func (s *Session) Step() error {
+	if s.exited {
+		return fmt.Errorf("debug: debuggee has exited")
+	}
+	if err := unix.PtraceSingleStep(s.pid); err != nil {
+		return fmt.Errorf("debug: PT_STEP: %w", err)
+	}
+	_, err := s.waitStopped()
+	return err
+}
+
+// Continue resumes execution until the next breakpoint or exit.
+func (s *Session) Continue() error {
+	if s.exited {
+		return fmt.Errorf("debug: debuggee has exited")
+	}
+	if err := unix.PtraceCont(s.pid, 0); err != nil {
+		return fmt.Errorf("debug: PT_CONTINUE: %w", err)
+	}
+	_, err := s.waitStopped()
+	return err
+}
+
+// Exited returns true if the debuggee has terminated.
+func (s *Session) Exited() bool { return s.exited }
+
+// Pid returns the debuggee's process ID.
+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 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)
+		syscall.Wait4(s.pid, nil, 0, nil)
+		s.exited = true
+	}
+	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 is a stub on FreeBSD: there is no /proc/pid/maps to parse,
+// and procfs(5) is not guaranteed to be mounted.  The callers degrade
+// gracefully: archReturnAddr falls back to the raw stack convention and
+// the mapping scan is skipped.
+func execRanges(pid int) []execRange { return nil }
+
+// findRWXMapping is a stub on FreeBSD for the same reason: the codebase
+// handshake file is the supported way the JIT region is located.
+func findRWXMapping(pid int) uint64 { return 0 }

debug/ptrace_freebsd_amd64.go

@@ -0,0 +1,160 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+//go:build freebsd && amd64
+
+package debug
+
+import (
+	"encoding/binary"
+	"fmt"
+	"unsafe"
+
+	"golang.org/x/sys/unix"
+)
+
+// GetRegs reads the general-purpose registers of the stopped debuggee and
+// converts the FreeBSD struct reg into the portable layout.
+func (s *Session) GetRegs() (Regs, error) {
+	var ur unix.Reg
+	if err := unix.PtraceGetRegs(s.pid, &ur); err != nil {
+		return Regs{}, fmt.Errorf("debug: PT_GETREGS: %w", err)
+	}
+	return Regs{
+		R15:    uint64(ur.R15),
+		R14:    uint64(ur.R14),
+		R13:    uint64(ur.R13),
+		R12:    uint64(ur.R12),
+		R11:    uint64(ur.R11),
+		R10:    uint64(ur.R10),
+		R9:     uint64(ur.R9),
+		R8:     uint64(ur.R8),
+		RDI:    uint64(ur.Rdi),
+		RSI:    uint64(ur.Rsi),
+		RBP:    uint64(ur.Rbp),
+		RBX:    uint64(ur.Rbx),
+		RDX:    uint64(ur.Rdx),
+		RCX:    uint64(ur.Rcx),
+		RAX:    uint64(ur.Rax),
+		RIP:    uint64(ur.Rip),
+		CS:     uint64(ur.Cs),
+		RFLAGS: uint64(ur.Rflags),
+		RSP:    uint64(ur.Rsp),
+		SS:     uint64(ur.Ss),
+		FS:     uint64(ur.Fs),
+		GS:     uint64(ur.Gs),
+		DS:     uint64(ur.Ds),
+		ES:     uint64(ur.Es),
+	}, nil
+}
+
+// SetRegs writes the general-purpose registers of the stopped debuggee.
+func (s *Session) SetRegs(regs *Regs) error {
+	// Read-modify-write keeps the fields FreeBSD owns (trapno, err) intact.
+	var ur unix.Reg
+	if err := unix.PtraceGetRegs(s.pid, &ur); err != nil {
+		return fmt.Errorf("debug: PT_GETREGS: %w", err)
+	}
+	ur.R15 = int64(regs.R15)
+	ur.R14 = int64(regs.R14)
+	ur.R13 = int64(regs.R13)
+	ur.R12 = int64(regs.R12)
+	ur.R11 = int64(regs.R11)
+	ur.R10 = int64(regs.R10)
+	ur.R9 = int64(regs.R9)
+	ur.R8 = int64(regs.R8)
+	ur.Rdi = int64(regs.RDI)
+	ur.Rsi = int64(regs.RSI)
+	ur.Rbp = int64(regs.RBP)
+	ur.Rbx = int64(regs.RBX)
+	ur.Rdx = int64(regs.RDX)
+	ur.Rcx = int64(regs.RCX)
+	ur.Rax = int64(regs.RAX)
+	ur.Rip = int64(regs.RIP)
+	ur.Cs = int64(regs.CS)
+	ur.Rflags = int64(regs.RFLAGS)
+	ur.Rsp = int64(regs.RSP)
+	ur.Ss = int64(regs.SS)
+	return unix.PtraceSetRegs(s.pid, &ur)
+}
+
+// FPRegs holds the x87 FPU and SSE (XMM) register state, the FXSAVE image
+// the FreeBSD struct fpreg mirrors: XMM0-15 at the same offsets.
+type FPRegs struct {
+	XMM [16][16]byte // XMM0-15
+}
+
+// GetFPRegs retrieves the FPU/SSE register state via PT_GETFPREGS.  The
+// FreeBSD struct fpreg mirrors the FXSAVE image: the x87 environment and
+// stack in Env/Acc, XMM0-15 in Xacc.
+func (s *Session) GetFPRegs() (FPRegs, error) {
+	var fp FPRegs
+	var fr unix.FpReg
+	if err := unix.PtraceGetFpRegs(s.pid, &fr); err != nil {
+		return fp, fmt.Errorf("debug: PT_GETFPREGS: %w", err)
+	}
+	for i := range 16 {
+		copy(fp.XMM[i][:], fr.Xacc[i][:])
+	}
+	return fp, nil
+}
+
+// VectorRegs holds the YMM register state.
+type VectorRegs struct {
+	YMM [16][32]byte // YMM0-15 (full 256-bit values)
+}
+
+// The XSAVE area the PT_GETXSTATE request returns follows the architectural
+// layout (Intel SDM vol 1, "XSAVE"): 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; XFEATURE_STATE_BIT_AVX is bit 2 of xstate_bv.
+const (
+	xsaveXMMOffset    = 160
+	xsaveHeaderOffset = 512
+	xsaveBVOffset     = xsaveHeaderOffset
+	ymmOffset         = xsaveHeaderOffset + 64 // 576
+	ymmSize           = 256                    // 16 registers, 16 bytes each
+	xfeatureMaskYMM   = 1 << 2
+	xstateMaxBuffer   = 4096 // PT_GETXSTATE_INFO bounds the size far below this
+)
+
+// GetVectorRegs retrieves the YMM registers via PT_GETXSTATE.  The low
+// (XMM) halves always come from the legacy image; the high halves are
+// copied only when xstate_bv reports the AVX state, and read as zero
+// otherwise.  When the 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)
+	n, _, errno := unix.Syscall6(
+		unix.SYS_PTRACE,
+		uintptr(unix.PT_GETXSTATE),
+		uintptr(s.pid),
+		0,
+		uintptr(unsafe.Pointer(&buf[0])),
+		0, 0,
+	)
+	if errno != 0 {
+		fp, err := s.GetFPRegs()
+		if err != nil {
+			return v, err
+		}
+		for i := range 16 {
+			copy(v.YMM[i][:16], fp.XMM[i][:])
+		}
+		return v, nil
+	}
+	for i := range 16 {
+		copy(v.YMM[i][:16], buf[xsaveXMMOffset+16*i:xsaveXMMOffset+16*i+16])
+	}
+	if int(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
+}

debug/ptrace_freebsd_arm64.go

@@ -0,0 +1,114 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+//go:build freebsd && arm64
+
+package debug
+
+import (
+	"fmt"
+
+	"golang.org/x/sys/unix"
+)
+
+// GetRegs reads the general-purpose registers of the stopped debuggee and
+// converts the FreeBSD struct reg (x[30], lr, sp, elr, spsr) into the
+// portable layout.
+func (s *Session) GetRegs() (Regs, error) {
+	var ur unix.Reg
+	if err := unix.PtraceGetRegs(s.pid, &ur); err != nil {
+		return Regs{}, fmt.Errorf("debug: PT_GETREGS: %w", err)
+	}
+	return Regs{
+		X0:     ur.X[0],
+		X1:     ur.X[1],
+		X2:     ur.X[2],
+		X3:     ur.X[3],
+		X4:     ur.X[4],
+		X5:     ur.X[5],
+		X6:     ur.X[6],
+		X7:     ur.X[7],
+		X8:     ur.X[8],
+		X9:     ur.X[9],
+		X10:    ur.X[10],
+		X11:    ur.X[11],
+		X12:    ur.X[12],
+		X13:    ur.X[13],
+		X14:    ur.X[14],
+		X15:    ur.X[15],
+		X16:    ur.X[16],
+		X17:    ur.X[17],
+		X18:    ur.X[18],
+		X19:    ur.X[19],
+		X20:    ur.X[20],
+		X21:    ur.X[21],
+		X22:    ur.X[22],
+		X23:    ur.X[23],
+		X24:    ur.X[24],
+		X25:    ur.X[25],
+		X26:    ur.X[26],
+		X27:    ur.X[27],
+		X28:    ur.X[28],
+		X29:    ur.X[29],
+		X30:    ur.Lr,
+		SP:     ur.Sp,
+		PC:     ur.Elr,
+		PSTATE: uint64(ur.Spsr),
+	}, nil
+}
+
+// SetRegs writes the general-purpose registers of the stopped debuggee.
+func (s *Session) SetRegs(regs *Regs) error {
+	var ur unix.Reg
+	ur.X = [30]uint64{
+		regs.X0, regs.X1, regs.X2, regs.X3, regs.X4, regs.X5, regs.X6,
+		regs.X7, regs.X8, regs.X9, regs.X10, regs.X11, regs.X12, regs.X13,
+		regs.X14, regs.X15, regs.X16, regs.X17, regs.X18, regs.X19, regs.X20,
+		regs.X21, regs.X22, regs.X23, regs.X24, regs.X25, regs.X26, regs.X27,
+		regs.X28, regs.X29,
+	}
+	ur.Lr = regs.X30
+	ur.Sp = regs.SP
+	ur.Elr = regs.PC
+	ur.Spsr = uint32(regs.PSTATE)
+	return unix.PtraceSetRegs(s.pid, &ur)
+}
+
+// FPRegs holds the arm64 FP/NEON register state: the 32 128-bit V
+// registers, then FPSR and FPCR (the user_fpsimd shape).
+type FPRegs struct {
+	V    [32][16]byte // V0-V31 (128-bit NEON/FP registers)
+	FPSR uint32
+	FPCR uint32
+}
+
+// GetFPRegs retrieves the FP/NEON register state via PT_GETFPREGS.  The
+// FreeBSD struct fpreg holds the 32 128-bit V registers followed by FPSR
+// and FPCR, the user_fpsimd shape.
+func (s *Session) GetFPRegs() (FPRegs, error) {
+	var fp FPRegs
+	var fr unix.FpReg
+	if err := unix.PtraceGetFpRegs(s.pid, &fr); err != nil {
+		return fp, fmt.Errorf("debug: PT_GETFPREGS: %w", err)
+	}
+	for i := range 32 {
+		copy(fp.V[i][:], fr.Q[i][:])
+	}
+	return fp, nil
+}
+
+// VectorRegs holds the full SIMD register state.
+type VectorRegs struct {
+	V [32][16]byte // V0-V31 (128-bit)
+}
+
+// GetVectorRegs retrieves the SIMD registers.
+func (s *Session) GetVectorRegs() (VectorRegs, error) {
+	var v VectorRegs
+	fp, err := s.GetFPRegs()
+	if err != nil {
+		return v, err
+	}
+	copy(v.V[:][:], fp.V[:][:])
+	return v, nil
+}

debug/ptrace_freebsd_riscv64.go

@@ -0,0 +1,115 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+//go:build freebsd && riscv64
+
+package debug
+
+import (
+	"fmt"
+
+	"golang.org/x/sys/unix"
+)
+
+// GetRegs reads the general-purpose registers of the stopped debuggee and
+// converts the FreeBSD struct reg into the portable layout.  Sstatus rides
+// the kernel's struct but the portable surface carries the GPRs and PC.
+func (s *Session) GetRegs() (Regs, error) {
+	var ur unix.Reg
+	if err := unix.PtraceGetRegs(s.pid, &ur); err != nil {
+		return Regs{}, fmt.Errorf("debug: PT_GETREGS: %w", err)
+	}
+	return Regs{
+		PC:  ur.Sepc,
+		Ra:  ur.Ra,
+		Sp:  ur.Sp,
+		Gp:  ur.Gp,
+		Tp:  ur.Tp,
+		T0:  ur.T[0],
+		T1:  ur.T[1],
+		T2:  ur.T[2],
+		S0:  ur.S[0],
+		S1:  ur.S[1],
+		A0:  ur.A[0],
+		A1:  ur.A[1],
+		A2:  ur.A[2],
+		A3:  ur.A[3],
+		A4:  ur.A[4],
+		A5:  ur.A[5],
+		A6:  ur.A[6],
+		A7:  ur.A[7],
+		S2:  ur.S[2],
+		S3:  ur.S[3],
+		S4:  ur.S[4],
+		S5:  ur.S[5],
+		S6:  ur.S[6],
+		S7:  ur.S[7],
+		S8:  ur.S[8],
+		S9:  ur.S[9],
+		S10: ur.S[10],
+		S11: ur.S[11],
+		T3:  ur.T[3],
+		T4:  ur.T[4],
+		T5:  ur.T[5],
+		T6:  ur.T[6],
+	}, nil
+}
+
+// SetRegs writes the general-purpose registers of the stopped debuggee.
+// Read-modify-write keeps sstatus, which the kernel owns, intact.
+func (s *Session) SetRegs(regs *Regs) error {
+	var ur unix.Reg
+	if err := unix.PtraceGetRegs(s.pid, &ur); err != nil {
+		return fmt.Errorf("debug: PT_GETREGS: %w", err)
+	}
+	ur.Sepc = regs.PC
+	ur.Ra = regs.Ra
+	ur.Sp = regs.Sp
+	ur.Gp = regs.Gp
+	ur.Tp = regs.Tp
+	ur.T = [7]uint64{regs.T0, regs.T1, regs.T2, regs.T3, regs.T4, regs.T5, regs.T6}
+	ur.S = [12]uint64{regs.S0, regs.S1, regs.S2, regs.S3, regs.S4, regs.S5,
+		regs.S6, regs.S7, regs.S8, regs.S9, regs.S10, regs.S11}
+	ur.A = [8]uint64{regs.A0, regs.A1, regs.A2, regs.A3, regs.A4, regs.A5, regs.A6, regs.A7}
+	return unix.PtraceSetRegs(s.pid, &ur)
+}
+
+// FPRegs holds the RISC-V FP register state (32 64-bit FP registers plus
+// fcsr).
+type FPRegs struct {
+	F    [32]uint64 // F0-F31 (64-bit FP registers)
+	FCSR uint32
+}
+
+// GetFPRegs retrieves the FP register state via PT_GETFPREGS.  The FreeBSD
+// struct fpreg carries each 64-bit FP register in a 128-bit slot (fp_x is
+// the flat [64]-word area the x/sys type renders as [32][2]); the low word
+// holds the register, and FCSR rides the tail.
+func (s *Session) GetFPRegs() (FPRegs, error) {
+	var fp FPRegs
+	var fr unix.FpReg
+	if err := unix.PtraceGetFpRegs(s.pid, &fr); err != nil {
+		return fp, fmt.Errorf("debug: PT_GETFPREGS: %w", err)
+	}
+	for i := range 32 {
+		fp.F[i] = fr.X[i][0]
+	}
+	fp.FCSR = uint32(fr.Fcsr)
+	return fp, nil
+}
+
+// VectorRegs holds the FP register state shown by the regs command
+// (riscv64 has 32 64-bit FP registers and fcsr).
+type VectorRegs struct {
+	F    [32]uint64
+	FCSR uint32
+}
+
+// GetVectorRegs retrieves the FP registers.
+func (s *Session) GetVectorRegs() (VectorRegs, error) {
+	fp, err := s.GetFPRegs()
+	if err != nil {
+		return VectorRegs{}, err
+	}
+	return VectorRegs{F: fp.F, FCSR: fp.FCSR}, nil
+}

debug/ptrace_integration_freebsd_test.go

@@ -0,0 +1,91 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+//go:build freebsd && amd64
+
+package debug
+
+import (
+	"os/exec"
+	"path/filepath"
+	"runtime"
+	"testing"
+
+	"sourcedock.dev/petrbalvin/gasm-sdk/verify"
+)
+
+// TestLaunchAndBreakpoint is the FreeBSD twin of the Linux integration
+// test: it drives the whole launch, breakpoint, trap and register-rewind
+// flow end to end.  It needs a real FreeBSD kernel (ptrace does not work
+// under emulation), so it only runs where it can.
+func TestLaunchAndBreakpoint(t *testing.T) {
+	if runtime.GOARCH != "amd64" {
+		t.Skip("runs only on amd64 hosts")
+	}
+	// The tracer is the OS thread that forked the debuggee (PT_TRACE_ME
+	// binds the relation to that thread); every ptrace request must come
+	// from the same thread, so pin the test goroutine to one thread.
+	runtime.LockOSThread()
+	defer runtime.UnlockOSThread()
+
+	bin := filepath.Join(t.TempDir(), "gasm")
+	out, err := exec.Command("go", "build", "-o", bin, "sourcedock.dev/petrbalvin/gasm-sdk/cmd/gasm").CombinedOutput()
+	if err != nil {
+		t.Fatalf("build gasm: %v: %s", err, out)
+	}
+
+	const kernelPath = "../testdata/verify/basic_amd64.s"
+	k, err := verify.Load(kernelPath)
+	if err != nil {
+		t.Fatalf("Load: %v", err)
+	}
+	t.Cleanup(k.Close)
+	fl, err := k.Func("wideCopy")
+	if err != nil {
+		t.Fatalf("Func: %v", err)
+	}
+
+	sess, err := Launch(bin, kernelPath, "wideCopy", make([]byte, fl.Args))
+	if err != nil {
+		t.Fatalf("Launch: %v", err)
+	}
+	t.Cleanup(sess.Kill)
+
+	bm := NewBreakpoints(sess)
+	entry := sess.CodeBase() + uint64(fl.Offset)
+	if _, err := bm.Set(entry, "entry"); err != nil {
+		t.Fatalf("Set: %v", err)
+	}
+
+	// The INT3 must be visible in the debuggee's memory.
+	word, err := sess.Peek(entry)
+	if err != nil {
+		t.Fatalf("Peek: %v", err)
+	}
+	if b := word & 0xFF; b != 0xCC {
+		t.Fatalf("int3 not patched: first byte %#02x at %#x", b, entry)
+	}
+
+	// The debuggee raises a second SIGSTOP after the launch barrier (the
+	// child's RunTarget marks its entry), so like the REPL and the cover
+	// mode the test keeps resuming until the breakpoint trap arrives.
+	for range 10 {
+		if err := sess.Continue(); err != nil {
+			t.Fatalf("Continue: %v", err)
+		}
+		if sess.Exited() {
+			t.Fatal("debuggee exited instead of trapping on the breakpoint")
+		}
+		regs, err := sess.GetRegs()
+		if err != nil {
+			t.Fatalf("GetRegs: %v", err)
+		}
+		if bp := bm.HandleTrap(&regs); bp != nil {
+			if bp.Addr != entry {
+				t.Fatalf("trap at %#x, want %#x", bp.Addr, entry)
+			}
+			return // trap on the entry breakpoint: the whole flow works
+		}
+	}
+	t.Fatal("no breakpoint trap after 10 resumes")
+}

debug/ptrace_integration_test.go

@@ -14,17 +14,26 @@ import (
 	"strings"
 	"testing"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/verify"
+	"sourcedock.dev/petrbalvin/gasm-sdk/verify"
 )
 
 // buildGasm produces the gasm binary the debugger spawns as its debuggee.
+// Every live ptrace test funnels through here, so this is also where the
+// deliberate-run boundary sits: under -short (the push pipeline's mode) the
+// live sessions skip, because a real debuggee's launch handshake needs the
+// machine to itself and a starved single-core runner turns each one into a
+// timeout that burns the step's whole budget. The local test gate and the
+// dispatched workflows run them in full.
 func buildGasm(t *testing.T) string {
 	t.Helper()
+	if testing.Short() {
+		t.Skip("live ptrace session: skipped in -short mode")
+	}
 	if p := os.Getenv("GASM_TEST_BIN"); p != "" {
 		return p
 	}
 	bin := filepath.Join(t.TempDir(), "gasm")
-	cmd := exec.Command("go", "build", "-o", bin, "sourcedock.dev/petrbalvin/gasm-devkit/cmd/gasm")
+	cmd := exec.Command("go", "build", "-o", bin, "sourcedock.dev/petrbalvin/gasm-sdk/cmd/gasm")
 	out, err := cmd.CombinedOutput()
 	if err != nil {
 		t.Fatalf("build gasm: %v: %s", err, out)

debug/ptrace_linux.go

@@ -91,6 +91,16 @@ func LaunchWithBuffers(gasmBin, asmPath, funcName string, args []byte, bufSpec s
 		if _, err := os.Stat(readyFile); err == nil {
 			break
 		}
+		// A debuggee that died before signalling readiness (unknown
+		// function, unparseable source) writes its failure notice to the
+		// handshake directory; read it and fail fast.  The poll never
+		// waits on the child: a wait here could consume the SIGSTOP park
+		// that waitStopped below must receive, hanging the launch.
+		if err := s.deadReason(); err != nil {
+			cmd.Wait()
+			os.RemoveAll(tmpDir)
+			return nil, nil, err
+		}
 		time.Sleep(5 * time.Millisecond)
 	}
 
@@ -131,6 +141,18 @@ func LaunchWithBuffers(gasmBin, asmPath, funcName string, args []byte, bufSpec s
 	return s, bufAddrs, nil
 }
 
+// deadReason reports the debuggee's own failure notice, the file its
+// failure paths write before exiting.  A debuggee killed without a notice
+// (a crash, SIGKILL) surfaces through waitStopped after the poll instead,
+// which is why the poll's budget stays finite.
+func (s *Session) deadReason() error {
+	data, err := os.ReadFile(filepath.Join(s.tmpDir, "dead"))
+	if err != nil {
+		return nil
+	}
+	return fmt.Errorf("debug: debuggee failed before signalling readiness: %s", strings.TrimSpace(string(data)))
+}
+
 // waitStopped consumes ptrace-stop events until one the debugger cares
 // 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
@@ -219,40 +241,41 @@ func (s *Session) Poke(addr, val uint64) error {
 	return nil
 }
 
-// ReadMemory reads len bytes from the debuggee's memory at addr.
+// ReadMemory reads len bytes from the debuggee's memory at addr.  The read
+// covers exactly the requested range: the old word-at-a-time loop read a
+// whole 8-byte word for the final partial word, so a request that ended
+// inside the last mapped page failed whenever the following page was
+// unmapped, even though every requested byte was readable.
 func (s *Session) ReadMemory(addr uint64, length int) ([]byte, error) {
 	out := make([]byte, length)
-	for i := 0; i < length; i += 8 {
-		word, err := s.Peek(addr + uint64(i))
+	mem, err := os.OpenFile(fmt.Sprintf("/proc/%d/mem", s.pid), os.O_RDONLY, 0)
 	if err != nil {
-			return out[:i], err
-		}
-		for j := 0; j < 8 && i+j < length; j++ {
-			out[i+j] = byte(word >> (8 * j))
+		return out, fmt.Errorf("debug: open /proc/%d/mem: %w", s.pid, err)
 	}
+	defer mem.Close()
+	n, err := mem.ReadAt(out, int64(addr))
+	if err != nil {
+		return out[:n], fmt.Errorf("debug: read mem %#x: %w", addr, err)
 	}
 	return out, nil
 }
 
-// WriteMemory writes bytes to the debuggee's memory at addr.
+// WriteMemory writes bytes to the debuggee's memory at addr.  The write
+// covers exactly the given bytes: /proc/pid/mem accepts writes of any
+// length at any offset, so the word loop's read-modify-write of the final
+// partial word (which read past the requested range and failed on an
+// unmapped following page) is unnecessary.
 func (s *Session) WriteMemory(addr uint64, data []byte) error {
-	for i := 0; i < len(data); i += 8 {
-		end := min(i+8, len(data))
-		var word uint64
-		for j := 0; j < end-i; j++ {
-			word |= uint64(data[i+j]) << (8 * j)
+	if len(data) == 0 {
+		return nil
 	}
-		if end-i < 8 {
-			existing, err := s.Peek(addr + uint64(i))
+	mem, err := os.OpenFile(fmt.Sprintf("/proc/%d/mem", s.pid), os.O_WRONLY, 0)
 	if err != nil {
-				return err
-			}
-			mask := ^((uint64(1) << (8 * (end - i))) - 1)
-			word = (existing & mask) | word
-		}
-		if err := s.Poke(addr+uint64(i), word); err != nil {
-			return err
+		return fmt.Errorf("debug: open /proc/%d/mem: %w", s.pid, err)
 	}
+	defer mem.Close()
+	if _, err := mem.WriteAt(data, int64(addr)); err != nil {
+		return fmt.Errorf("debug: write mem %#x: %w", addr, err)
 	}
 	return nil
 }

debug/regs_freebsd_amd64.go

@@ -0,0 +1,98 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+//go:build freebsd && amd64
+
+package debug
+
+// Regs holds the full general-purpose register set of a traced process
+// (the FreeBSD amd64 struct reg layout, sys/x86/include/reg.h).  FreeBSD
+// reports segment selectors (FS/GS/ES/DS), not the bases the Linux ptrace
+// surface carries, and has no ORIG_RAX slot.
+type Regs struct {
+	R15    uint64
+	R14    uint64
+	R13    uint64
+	R12    uint64
+	RBP    uint64
+	RBX    uint64
+	R11    uint64
+	R10    uint64
+	R9     uint64
+	R8     uint64
+	RAX    uint64
+	RCX    uint64
+	RDX    uint64
+	RSI    uint64
+	RDI    uint64
+	RIP    uint64
+	CS     uint64
+	RFLAGS uint64
+	RSP    uint64
+	SS     uint64
+	FS     uint64
+	GS     uint64
+	DS     uint64
+	ES     uint64
+}
+
+// GetPC returns the program counter.
+func (r *Regs) GetPC() uint64 { return r.RIP }
+
+// SetPC sets the program counter.
+func (r *Regs) SetPC(pc uint64) { r.RIP = pc }
+
+// GetSP returns the stack pointer.
+func (r *Regs) GetSP() uint64 { return r.RSP }
+
+// RegValue returns the value of the named register, or false if unknown.
+func (r *Regs) RegValue(name string) (uint64, bool) {
+	switch name {
+	case "rax", "eax", "ax", "al":
+		return r.RAX, true
+	case "rbx", "ebx", "bx", "bl":
+		return r.RBX, true
+	case "rcx", "ecx", "cx", "cl":
+		return r.RCX, true
+	case "rdx", "edx", "dx", "dl":
+		return r.RDX, true
+	case "rsi", "esi", "si":
+		return r.RSI, true
+	case "rdi", "edi", "di":
+		return r.RDI, true
+	case "rbp", "ebp", "bp":
+		return r.RBP, true
+	case "rsp", "esp", "sp":
+		return r.RSP, true
+	case "r8":
+		return r.R8, true
+	case "r9":
+		return r.R9, true
+	case "r10":
+		return r.R10, true
+	case "r11":
+		return r.R11, true
+	case "r12":
+		return r.R12, true
+	case "r13":
+		return r.R13, true
+	case "r14":
+		return r.R14, true
+	case "r15":
+		return r.R15, true
+	case "rip", "eip":
+		return r.RIP, true
+	default:
+		return 0, false
+	}
+}
+
+// breakpointInsn is the software breakpoint instruction.
+var breakpointInsn = []byte{0xCC} // INT3
+
+// breakpointPCAdjust is how far PC is past the breakpoint instruction after
+// a trap.  INT3 leaves the hardware PC on the following instruction (Intel
+// SDM vol 3, "Debug Exceptions") and the FreeBSD T_BPTFLT path delivers
+// that frame unmodified (sys/amd64/amd64/trap.c), so the trap address is
+// PC-1, the same correction the Linux side applies.
+const breakpointPCAdjust = 1

debug/regs_freebsd_arm64.go

@@ -0,0 +1,139 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+//go:build freebsd && arm64
+
+package debug
+
+// Regs holds the full general-purpose register set of a traced process
+// (the FreeBSD arm64 struct reg layout, sys/arm64/include/reg.h: x[30], lr,
+// sp, elr, spsr).
+type Regs struct {
+	X0     uint64
+	X1     uint64
+	X2     uint64
+	X3     uint64
+	X4     uint64
+	X5     uint64
+	X6     uint64
+	X7     uint64
+	X8     uint64
+	X9     uint64
+	X10    uint64
+	X11    uint64
+	X12    uint64
+	X13    uint64
+	X14    uint64
+	X15    uint64
+	X16    uint64
+	X17    uint64
+	X18    uint64
+	X19    uint64
+	X20    uint64
+	X21    uint64
+	X22    uint64
+	X23    uint64
+	X24    uint64
+	X25    uint64
+	X26    uint64
+	X27    uint64
+	X28    uint64
+	X29    uint64 // FP (frame pointer)
+	X30    uint64 // LR (link register)
+	SP     uint64
+	PC     uint64
+	PSTATE uint64
+}
+
+// GetPC returns the program counter.
+func (r *Regs) GetPC() uint64 { return r.PC }
+
+// SetPC sets the program counter.
+func (r *Regs) SetPC(pc uint64) { r.PC = pc }
+
+// GetSP returns the stack pointer.
+func (r *Regs) GetSP() uint64 { return r.SP }
+
+// RegValue returns the value of the named register, or false if unknown.
+func (r *Regs) RegValue(name string) (uint64, bool) {
+	switch name {
+	case "x0":
+		return r.X0, true
+	case "x1":
+		return r.X1, true
+	case "x2":
+		return r.X2, true
+	case "x3":
+		return r.X3, true
+	case "x4":
+		return r.X4, true
+	case "x5":
+		return r.X5, true
+	case "x6":
+		return r.X6, true
+	case "x7":
+		return r.X7, true
+	case "x8":
+		return r.X8, true
+	case "x9":
+		return r.X9, true
+	case "x10":
+		return r.X10, true
+	case "x11":
+		return r.X11, true
+	case "x12":
+		return r.X12, true
+	case "x13":
+		return r.X13, true
+	case "x14":
+		return r.X14, true
+	case "x15":
+		return r.X15, true
+	case "x16":
+		return r.X16, true
+	case "x17":
+		return r.X17, true
+	case "x18":
+		return r.X18, true
+	case "x19":
+		return r.X19, true
+	case "x20":
+		return r.X20, true
+	case "x21":
+		return r.X21, true
+	case "x22":
+		return r.X22, true
+	case "x23":
+		return r.X23, true
+	case "x24":
+		return r.X24, true
+	case "x25":
+		return r.X25, true
+	case "x26":
+		return r.X26, true
+	case "x27":
+		return r.X27, true
+	case "x28":
+		return r.X28, true
+	case "x29", "fp":
+		return r.X29, true
+	case "x30", "lr":
+		return r.X30, true
+	case "sp":
+		return r.SP, true
+	case "pc":
+		return r.PC, true
+	default:
+		return 0, false
+	}
+}
+
+// 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: 0.  The BRK synchronous exception leaves ELR_EL0 on the BRK
+// itself (ARM DDI 0487), and the FreeBSD EXCP_BRKPT_EL0 handler delivers
+// the frame's elr unmodified (sys/arm64/arm64/trap.c), so the trap address
+// is the PC as reported.
+const breakpointPCAdjust = 0

debug/regs_freebsd_riscv64.go

@@ -0,0 +1,134 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+//go:build freebsd && riscv64
+
+package debug
+
+// Regs holds the full general-purpose register set of a traced process
+// (the FreeBSD riscv64 struct reg layout: ra, sp, gp, tp, t0-t6, s0-s11,
+// a0-a7, sepc, sstatus).
+type Regs struct {
+	PC  uint64 // sepc
+	Ra  uint64 // x1 (return address)
+	Sp  uint64 // x2
+	Gp  uint64 // x3
+	Tp  uint64 // x4
+	T0  uint64 // x5
+	T1  uint64 // x6
+	T2  uint64 // x7
+	S0  uint64 // x8 (frame pointer)
+	S1  uint64 // x9
+	A0  uint64 // x10
+	A1  uint64 // x11
+	A2  uint64 // x12
+	A3  uint64 // x13
+	A4  uint64 // x14
+	A5  uint64 // x15
+	A6  uint64 // x16
+	A7  uint64 // x17
+	S2  uint64 // x18
+	S3  uint64 // x19
+	S4  uint64 // x20
+	S5  uint64 // x21
+	S6  uint64 // x22
+	S7  uint64 // x23
+	S8  uint64 // x24
+	S9  uint64 // x25
+	S10 uint64 // x26
+	S11 uint64 // x27
+	T3  uint64 // x28
+	T4  uint64 // x29
+	T5  uint64 // x30
+	T6  uint64 // x31
+}
+
+// GetPC returns the program counter.
+func (r *Regs) GetPC() uint64 { return r.PC }
+
+// SetPC sets the program counter.
+func (r *Regs) SetPC(pc uint64) { r.PC = pc }
+
+// GetSP returns the stack pointer.
+func (r *Regs) GetSP() uint64 { return r.Sp }
+
+// RegValue returns the value of the named register, or false if unknown.
+func (r *Regs) RegValue(name string) (uint64, bool) {
+	switch name {
+	case "pc":
+		return r.PC, true
+	case "ra", "x1":
+		return r.Ra, true
+	case "sp", "x2":
+		return r.Sp, true
+	case "gp", "x3":
+		return r.Gp, true
+	case "tp", "x4":
+		return r.Tp, true
+	case "t0", "x5":
+		return r.T0, true
+	case "t1", "x6":
+		return r.T1, true
+	case "t2", "x7":
+		return r.T2, true
+	case "s0", "fp", "x8":
+		return r.S0, true
+	case "s1", "x9":
+		return r.S1, true
+	case "a0", "x10":
+		return r.A0, true
+	case "a1", "x11":
+		return r.A1, true
+	case "a2", "x12":
+		return r.A2, true
+	case "a3", "x13":
+		return r.A3, true
+	case "a4", "x14":
+		return r.A4, true
+	case "a5", "x15":
+		return r.A5, true
+	case "a6", "x16":
+		return r.A6, true
+	case "a7", "x17":
+		return r.A7, true
+	case "s2", "x18":
+		return r.S2, true
+	case "s3", "x19":
+		return r.S3, true
+	case "s4", "x20":
+		return r.S4, true
+	case "s5", "x21":
+		return r.S5, true
+	case "s6", "x22":
+		return r.S6, true
+	case "s7", "x23":
+		return r.S7, true
+	case "s8", "x24":
+		return r.S8, true
+	case "s9", "x25":
+		return r.S9, true
+	case "s10", "x26":
+		return r.S10, true
+	case "s11", "x27":
+		return r.S11, true
+	case "t3", "x28":
+		return r.T3, true
+	case "t4", "x29":
+		return r.T4, true
+	case "t5", "x30":
+		return r.T5, true
+	case "t6", "x31":
+		return r.T6, true
+	default:
+		return 0, false
+	}
+}
+
+// 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: 0.  The EBREAK synchronous exception leaves sepc on the ebreak
+// itself (RISC-V privileged architecture), so the trap address is the PC as
+// reported.
+const breakpointPCAdjust = 0

debug/repl.go

@@ -1,7 +1,7 @@
 // Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
 // SPDX-License-Identifier: BSD-3-Clause
 
-//go:build linux
+//go:build linux || (freebsd && (amd64 || arm64 || riscv64))
 
 package debug
 
@@ -71,7 +71,12 @@ func REPL(s *Session, bm *Breakpoints, codeBase uint64, funcOffset, funcSize, ar
 		case "step", "s":
 			n := 1
 			if len(parts) > 1 {
-				n, _ = strconv.Atoi(parts[1])
+				v, err := strconv.Atoi(parts[1])
+				if err != nil || v < 0 {
+					fmt.Printf("invalid count: %s\n", parts[1])
+					continue
+				}
+				n = v
 			}
 			for range n {
 				if s.Exited() {
@@ -82,8 +87,13 @@ func REPL(s *Session, bm *Breakpoints, codeBase uint64, funcOffset, funcSize, ar
 					fmt.Println(err)
 					break
 				}
+				if sig := s.LastSignal(); sig != 0 {
+					regs, _ := s.GetRegs()
+					fmt.Printf("stopped on signal %v at %#x\n", sig, regs.GetPC())
+					break
 				}
-			if !s.Exited() {
+			}
+			if !s.Exited() && s.LastSignal() == 0 {
 				regs, _ := s.GetRegs()
 				pc := regs.GetPC()
 				text, _, _ := s.Disassemble(pc)
@@ -106,16 +116,16 @@ func REPL(s *Session, bm *Breakpoints, codeBase uint64, funcOffset, funcSize, ar
 				}
 				if err := s.Continue(); err != nil {
 					fmt.Println(err)
-					bm.Clear(afterAddr)
+					clearNextBp(bm, afterAddr)
 					continue
 				}
 				if s.Exited() {
-					bm.Clear(afterAddr)
+					clearNextBp(bm, afterAddr)
 					fmt.Println("debuggee exited")
 					continue
 				}
 				if sig := s.LastSignal(); sig != 0 {
-					bm.Clear(afterAddr)
+					clearNextBp(bm, afterAddr)
 					regs, _ := s.GetRegs()
 					fmt.Printf("stopped on signal %v at %#x\n", sig, regs.GetPC())
 					continue
@@ -125,12 +135,17 @@ func REPL(s *Session, bm *Breakpoints, codeBase uint64, funcOffset, funcSize, ar
 				// snapshot, and a stale SetRegs would clobber live state.
 				regs, _ = s.GetRegs()
 				bm.HandleTrap(&regs)
-				bm.Clear(afterAddr)
+				clearNextBp(bm, afterAddr)
 			} else {
 				if err := s.Step(); err != nil {
 					fmt.Println(err)
 					continue
 				}
+				if sig := s.LastSignal(); sig != 0 {
+					regs, _ := s.GetRegs()
+					fmt.Printf("stopped on signal %v at %#x\n", sig, regs.GetPC())
+					continue
+				}
 			}
 			if !s.Exited() {
 				regs, _ := s.GetRegs()
@@ -156,16 +171,16 @@ func REPL(s *Session, bm *Breakpoints, codeBase uint64, funcOffset, funcSize, ar
 			}
 			if err := s.Continue(); err != nil {
 				fmt.Println(err)
-				bm.Clear(retAddr)
+				clearNextBp(bm, retAddr)
 				continue
 			}
 			if s.Exited() {
-				bm.Clear(retAddr)
+				clearNextBp(bm, retAddr)
 				fmt.Println("debuggee exited")
 				continue
 			}
 			if sig := s.LastSignal(); sig != 0 {
-				bm.Clear(retAddr)
+				clearNextBp(bm, retAddr)
 				regs, _ := s.GetRegs()
 				fmt.Printf("stopped on signal %v at %#x\n", sig, regs.GetPC())
 				continue
@@ -174,7 +189,7 @@ func REPL(s *Session, bm *Breakpoints, codeBase uint64, funcOffset, funcSize, ar
 			// does: HandleTrap must see the PC the trap left behind.
 			regs, _ = s.GetRegs()
 			bm.HandleTrap(&regs)
-			bm.Clear(retAddr)
+			clearNextBp(bm, retAddr)
 			if s.Exited() {
 				fmt.Println("debuggee exited")
 			} else {
@@ -214,6 +229,7 @@ func REPL(s *Session, bm *Breakpoints, codeBase uint64, funcOffset, funcSize, ar
 					break
 				}
 				regs, _ := s.GetRegs()
+				trapPC := regs.GetPC()
 				if bp := bm.HandleTrap(&regs); bp != nil {
 					// Execute the instruction under the restored breakpoint
 					// so the next continue cannot re-trap on the same
@@ -229,6 +245,13 @@ func REPL(s *Session, bm *Breakpoints, codeBase uint64, funcOffset, funcSize, ar
 					fmt.Printf("breakpoint hit: %s (func+%#x)\n", name, bp.Addr-codeBase-uint64(funcOffset))
 					break
 				}
+				// The trap matched no breakpoint of ours.  When the PC still
+				// stands on the trapping instruction, resuming would re-execute
+				// it and trap forever, so surface the stop instead of spinning.
+				if TrapStray(s, reason, trapPC) {
+					fmt.Printf("SIGTRAP at %#x matches no breakpoint; the PC did not advance\n", trapPC-uint64(breakpointPCAdjust))
+					break
+				}
 			}
 
 		case "break", "b":
@@ -326,6 +349,10 @@ func REPL(s *Session, bm *Breakpoints, codeBase uint64, funcOffset, funcSize, ar
 			length := 64
 			if len(parts) > 1 {
 				addr, _ = resolveAddr(parts[1], codeBase, uint64(funcOffset), labels)
+				if addr == 0 {
+					fmt.Printf("unknown address: %s\n", parts[1])
+					continue
+				}
 			}
 			if len(parts) > 2 {
 				// A malformed or non-positive length would panic
@@ -399,9 +426,13 @@ func REPL(s *Session, bm *Breakpoints, codeBase uint64, funcOffset, funcSize, ar
 		case "disas", "u":
 			n := 5
 			if len(parts) > 1 {
-				n, _ = strconv.Atoi(parts[1])
-				if n <= 0 {
-					n = 5
+				v, err := strconv.Atoi(parts[1])
+				if err != nil {
+					fmt.Printf("invalid count: %s\n", parts[1])
+					continue
+				}
+				if v > 0 {
+					n = v
 				}
 			}
 			regs, _ := s.GetRegs()
@@ -496,10 +527,18 @@ func REPL(s *Session, bm *Breakpoints, codeBase uint64, funcOffset, funcSize, ar
 					typ = WatchRead
 				case "w":
 					typ = WatchWrite
+				default:
+					fmt.Printf("unknown watchpoint type: %s (want r or w)\n", parts[2])
+					continue
 				}
 			}
 			if len(parts) > 3 {
-				size, _ = strconv.Atoi(parts[3])
+				v, err := strconv.Atoi(parts[3])
+				if err != nil || v <= 0 {
+					fmt.Printf("invalid size: %s\n", parts[3])
+					continue
+				}
+				size = v
 			}
 			slot := s.FindFreeWatchpointSlot()
 			if slot < 0 {
@@ -543,6 +582,15 @@ func REPL(s *Session, bm *Breakpoints, codeBase uint64, funcOffset, funcSize, ar
 	s.Kill()
 }
 
+// clearNextBp removes one of the temporary breakpoints the next and finish
+// commands plant, reporting a failure instead of silently leaving the trap
+// instruction behind in the debuggee.
+func clearNextBp(bm *Breakpoints, addr uint64) {
+	if err := bm.Clear(addr); err != nil {
+		fmt.Printf("cannot remove temporary breakpoint at %#x: %v\n", addr, err)
+	}
+}
+
 func hexDump(addr uint64, data []byte) {
 	for i := 0; i < len(data); i += 16 {
 		end := min(i+16, len(data))

debug/stopinfo_freebsd.go

@@ -0,0 +1,73 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+//go:build freebsd && (amd64 || arm64 || riscv64)
+
+package debug
+
+import (
+	"encoding/binary"
+	"syscall"
+	"unsafe"
+
+	"golang.org/x/sys/unix"
+)
+
+// FreeBSD TRAP_* si_code values (sys/signal.h).  A breakpoint (INT3, BRK,
+// EBREAK) arrives as TRAP_BRKPT on every supported architecture; TRAP_TRACE
+// is shared by the completed single-step and the hardware watchpoint hit,
+// so the watchpoint layer disambiguates from the debug registers.
+const (
+	trapBRKPT = 1 // TRAP_BRKPT
+	trapTRACE = 2 // TRAP_TRACE
+)
+
+// StopReason describes why the debuggee stopped.
+type StopReason int
+
+const (
+	StopNone       StopReason = iota
+	StopBreakpoint            // software breakpoint hit
+	StopWatchpoint            // hardware watchpoint triggered
+	StopSingleStep            // single-step completed
+	StopSignal                // stopped by a signal
+	StopExited                // process exited
+)
+
+// StopInfo returns the reason the debuggee stopped and the faulting address
+// (for watchpoints, the watched address that was accessed).  FreeBSD has no
+// PTRACE_GETSIGINFO; the stop's signal information comes from PT_LWPINFO,
+// whose pl_siginfo carries the siginfo the kernel delivered.  A ptrace stop
+// with no signal behind it (a completed single-step, the initial attach)
+// fills no siginfo at all.
+func (s *Session) StopInfo() (StopReason, uint64) {
+	if s.exited {
+		return StopExited, 0
+	}
+	var info unix.PtraceLwpInfoStruct
+	if err := unix.PtraceLwpInfo(s.pid, &info); err != nil {
+		return StopNone, 0
+	}
+	// The siginfo layout is the FreeBSD siginfo_t: three leading ints
+	// (signo, errno, code), then the union, 8-byte aligned, whose _fault
+	// member puts the address at byte offset 16.  The read is byte-wise
+	// because the blob's alignment is not guaranteed.
+	si := (*[64]byte)(unsafe.Pointer(&info.Siginfo))
+	signo := int32(binary.LittleEndian.Uint32(si[0:4]))
+	code := int32(binary.LittleEndian.Uint32(si[8:12]))
+	switch {
+	case signo == 0:
+		// A pure ptrace stop: single-step completion, attach, or the
+		// events the kernel resolves internally.
+		return StopSingleStep, 0
+	case signo != int32(syscall.SIGTRAP):
+		return StopSignal, uint64(code)
+	case code == trapBRKPT:
+		return StopBreakpoint, 0
+	case code == trapTRACE:
+		addr := binary.LittleEndian.Uint64(si[16:24])
+		return archStopTrace(s, addr)
+	default:
+		return StopSingleStep, 0
+	}
+}

debug/target_freebsd.go

@@ -0,0 +1,206 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+//go:build freebsd && (amd64 || arm64 || riscv64)
+
+package debug
+
+import (
+	"encoding/hex"
+	"fmt"
+	"os"
+	"runtime"
+	"strconv"
+	"strings"
+	"syscall"
+	"unsafe"
+
+	"golang.org/x/sys/unix"
+
+	"sourcedock.dev/petrbalvin/gasm-sdk/asm"
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
+	"sourcedock.dev/petrbalvin/gasm-sdk/verify"
+)
+
+// mapRWX maps code into a read-write-execute region.
+func mapRWX(code []byte) ([]byte, error) {
+	const pageSize = 4096
+	size := (len(code) + pageSize - 1) &^ (pageSize - 1)
+	mem, err := syscall.Mmap(-1, 0, size,
+		syscall.PROT_READ|syscall.PROT_WRITE|syscall.PROT_EXEC,
+		syscall.MAP_PRIVATE|syscall.MAP_ANON)
+	if err != nil {
+		return nil, err
+	}
+	copy(mem, code)
+	return mem, nil
+}
+
+// setupBuffers allocates buffers in the debuggee's memory.
+func setupBuffers(spec string, args []byte, tmpDir string) ([]byte, error) {
+	type bufSpec struct {
+		name    string
+		size    int
+		pattern string
+	}
+	var specs []bufSpec
+	for part := range strings.SplitSeq(spec, ",") {
+		fields := strings.SplitN(part, ":", 3)
+		if len(fields) != 3 {
+			continue
+		}
+		size, err := strconv.Atoi(fields[1])
+		if err != nil || size <= 0 {
+			continue
+		}
+		specs = append(specs, bufSpec{name: fields[0], size: size, pattern: fields[2]})
+	}
+
+	if len(specs) == 0 {
+		return args, nil
+	}
+
+	var bufAddrs []uint64
+	for _, s := range specs {
+		buf, err := syscall.Mmap(-1, 0, s.size,
+			syscall.PROT_READ|syscall.PROT_WRITE,
+			syscall.MAP_PRIVATE|syscall.MAP_ANON)
+		if err != nil {
+			return nil, fmt.Errorf("mmap buffer %s: %w", s.name, err)
+		}
+		fillBuffer(buf, s.pattern)
+		bufAddrs = append(bufAddrs, uint64(uintptr(unsafe.Pointer(&buf[0]))))
+	}
+
+	addrFile, err := os.Create(tmpDir + "/bufaddrs")
+	if err != nil {
+		return nil, err
+	}
+	for _, addr := range bufAddrs {
+		fmt.Fprintf(addrFile, "%d\n", addr)
+	}
+	addrFile.Close()
+
+	return args, nil
+}
+
+// fillBuffer fills a buffer with the specified pattern.
+func fillBuffer(buf []byte, pattern string) {
+	switch pattern {
+	case "zero":
+	case "ones":
+		for i := range buf {
+			buf[i] = 0xFF
+		}
+	case "seq":
+		for i := range buf {
+			buf[i] = byte(i)
+		}
+	default:
+		if data, err := hex.DecodeString(pattern); err == nil && len(data) > 0 {
+			for i := range buf {
+				buf[i] = data[i%len(data)]
+			}
+		}
+	}
+}
+
+// RunTarget is the debuggee entry point (gasm debug --target).  A failure
+// is marked in the handshake directory before the process exits, so the
+// debugger's readiness poll fails fast on a dead debuggee instead of
+// waiting out its whole budget.
+func RunTarget(asmPath, funcName, argsFile, tmpDir string) error {
+	err := runTarget(asmPath, funcName, argsFile, tmpDir)
+	if err != nil {
+		markDead(tmpDir, err.Error())
+	}
+	return err
+}
+
+func runTarget(asmPath, funcName, argsFile, tmpDir string) error {
+	src, err := os.ReadFile(asmPath)
+	if err != nil {
+		return fmt.Errorf("debug target: %w", err)
+	}
+	file, errs := parser.Parse(asmPath, string(src))
+	if len(errs) > 0 {
+		return fmt.Errorf("debug target: parse: %v", errs[0])
+	}
+	img, err := asm.AssembleFile(file)
+	if err != nil {
+		return fmt.Errorf("debug target: assemble: %w", err)
+	}
+
+	var fl *asm.FuncLayout
+	for i := range img.Funcs {
+		if img.Funcs[i].Name == funcName {
+			fl = &img.Funcs[i]
+			break
+		}
+	}
+	if fl == nil {
+		return fmt.Errorf("debug target: function %q not found", funcName)
+	}
+
+	code := img.Bytes()
+	exec, err := mapRWX(code)
+	if err != nil {
+		return fmt.Errorf("debug target: mmap: %w", err)
+	}
+
+	codeBase := uintptr(unsafe.Pointer(&exec[0]))
+	if err := os.WriteFile(tmpDir+"/codebase", []byte(fmt.Sprintf("%d", codeBase)), 0o644); err != nil {
+		return fmt.Errorf("debug target: write codebase: %w", err)
+	}
+
+	meta := fmt.Sprintf("%d %d %d", fl.Offset, fl.Size, fl.Args)
+	os.WriteFile(tmpDir+"/funcmeta", []byte(meta), 0o644)
+
+	labelsFile, _ := os.Create(tmpDir + "/labels")
+	if labelsFile != nil {
+		for label, off := range fl.Labels {
+			fmt.Fprintf(labelsFile, "%s %d\n", label, off)
+		}
+		labelsFile.Close()
+	}
+
+	args, err := os.ReadFile(argsFile)
+	if err != nil {
+		return fmt.Errorf("debug target: read args: %w", err)
+	}
+	if len(args) < fl.Args {
+		padded := make([]byte, fl.Args)
+		copy(padded, args)
+		args = padded
+	}
+
+	bufSpecFile := tmpDir + "/bufspec"
+	if bufSpec, err := os.ReadFile(bufSpecFile); err == nil && len(bufSpec) > 0 {
+		args, err = setupBuffers(string(bufSpec), args, tmpDir)
+		if err != nil {
+			return fmt.Errorf("debug target: setup buffers: %w", err)
+		}
+	}
+
+	runtime.LockOSThread()
+
+	if _, _, errno := unix.RawSyscall(unix.SYS_PTRACE, uintptr(unix.PT_TRACE_ME), 0, 0); errno != 0 {
+		return fmt.Errorf("debug target: PT_TRACE_ME: %v", errno)
+	}
+	os.WriteFile(tmpDir+"/ready", []byte("ok"), 0o644)
+	syscall.Kill(syscall.Getpid(), syscall.SIGSTOP)
+
+	os.WriteFile(tmpDir+"/entry", []byte("ok"), 0o644)
+	syscall.Kill(syscall.Getpid(), syscall.SIGSTOP)
+
+	fnAddr := codeBase + uintptr(fl.Offset)
+	stackArgs := make([]byte, fl.Args)
+	copy(stackArgs, args)
+
+	if _, callErr := verify.Call(fnAddr, stackArgs); callErr != nil {
+		os.Exit(1)
+	}
+	// Success returns to the caller, which exits with status 0; the JIT
+	// code has already run to its own trampoline by the time Call returns.
+	return nil
+}

debug/target_linux_amd64.go

@@ -12,13 +12,24 @@ import (
 	"syscall"
 	"unsafe"
 
-	"sourcedock.dev/petrbalvin/gasm-devkit/asm"
-	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
-	"sourcedock.dev/petrbalvin/gasm-devkit/verify"
+	"sourcedock.dev/petrbalvin/gasm-sdk/asm"
+	"sourcedock.dev/petrbalvin/gasm-sdk/parser"
+	"sourcedock.dev/petrbalvin/gasm-sdk/verify"
 )
 
-// RunTarget is the debuggee entry point (gasm debug --target).
+// RunTarget is the debuggee entry point (gasm debug --target).  A failure
+// is marked in the handshake directory before the process exits, so the
+// debugger's readiness poll fails fast on a dead debuggee instead of
+// waiting out its whole budget.
 func RunTarget(asmPath, funcName, argsFile, tmpDir string) error {
+	err := runTarget(asmPath, funcName, argsFile, tmpDir)
+	if err != nil {
+		markDead(tmpDir, err.Error())
+	}
+	return err
+}
+
+func runTarget(asmPath, funcName, argsFile, tmpDir string) error {
 	src, err := os.ReadFile(asmPath)
 	if err != nil {
 		return fmt.Errorf("debug target: %w", err)