feat(verify): add analyzeO1Range and fastStereoSums differential tests

Assisted-by: Qwen 3.8 Max Preview

arch/arch.go

@@ -156,6 +156,16 @@ func (t *Table) Lookup(mnemonic string) (Instr, bool) {
 			}
 		}
 	}
+	// amd64 EVEX instructions take a .Z zeroing suffix (masking is written as
+	// an explicit K operand rather than a suffix); strip it so the base
+	// instruction is still recognised.
+	if t.Arch == AMD64 {
+		if base, ok := strings.CutSuffix(key, ".Z"); ok {
+			if in, found := t.instrs[base]; found {
+				return in, true
+			}
+		}
+	}
 	return Instr{}, false
 }
 

asm/assemble.go

@@ -23,16 +23,19 @@ import (
 // operands require relocations and are not yet supported; the SIMD (VEX/AVX2)
 // integer and shuffle/extract/permute/move set is in.
 func Assemble(t *ast.Text) ([]byte, map[string]int, error) {
-	code, _, labels, err := assemble(t, nil)
+	code, _, labels, _, err := assemble(t, nil)
 	return code, labels, err
 }
 
 // linkInfo carries file-level symbol context into a single-function assembly:
 // the set of static symbols a GLOBL in the same file defines.  A nil link
 // rejects SB operands outright (single-function assembly cannot resolve
-// them).
+// 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.
 type linkInfo struct {
-	symbols map[string]bool
+	symbols       map[string]bool
+	allowExternal bool
 }
 
 // sbPatch is a function-relative static-symbol relocation: the disp32 field
@@ -45,9 +48,19 @@ type sbPatch struct {
 	addend int64
 }
 
+// spadjStep is one stack-adjustment boundary within a function: Value is the
+// SP delta from the entry state (just below the return address) in effect
+// from PC (function-relative) until the next step.  The steps feed the
+// pcsp table of the object-file emitters.
+type spadjStep struct {
+	pc    int
+	value int
+}
+
 // assemble encodes a TEXT body, returning the machine code, the static-symbol
-// patch sites (for the file-level layout to resolve) and the label table.
+// patch sites (for the file-level layout to resolve), the label table and the
-func assemble(t *ast.Text, link *linkInfo) ([]byte, []sbPatch, map[string]int, error) {
+// stack-adjustment boundaries.
+func assemble(t *ast.Text, link *linkInfo) ([]byte, []sbPatch, map[string]int, []spadjStep, error) {
 	fi := computeFrame(t)
 	chain := jumpChain(t)
 	resolve := func(name string) string {
@@ -71,7 +84,7 @@ func assemble(t *ast.Text, link *linkInfo) ([]byte, []sbPatch, map[string]int, e
 			case *ast.Instr:
 				sz, err := instrSize(s, fi, long[i], link)
 				if err != nil {
-					return nil, nil, nil, fmt.Errorf("%s: %w", s.Mnemonic.Text, err)
+					return nil, nil, nil, nil, fmt.Errorf("%s: %w", s.Mnemonic.Text, err)
 				}
 				sizes[i] = sz
 				pcs[i] = pos
@@ -111,24 +124,42 @@ func assemble(t *ast.Text, link *linkInfo) ([]byte, []sbPatch, map[string]int, e
 	// Pass 2: emit.
 	out := append([]byte(nil), fi.prologue...)
 	var patches []sbPatch
+	var steps []spadjStep
+	if fi.useFP {
+		// PUSHQ BP saves the return-address-relative base (+8); the MOVQ
+		// changes nothing; SUBQ $size, SP completes the frame.
+		steps = append(steps,
+			spadjStep{1, 8},
+			spadjStep{len(fi.prologue), 8 + fi.size},
+		)
+	}
 	pos := len(fi.prologue)
 	for i, stmt := range t.Body {
 		s, ok := stmt.(*ast.Instr)
 		if !ok {
 			continue
 		}
+		if strings.ToUpper(s.Mnemonic.Text) == "RET" && fi.useFP {
+			// The RET's epilogue prefix unwinds: ADDQ $size, SP restores
+			// the saved-BP-only stack, POPQ BP the entry state.
+			epi := len(fi.epilogue)
+			steps = append(steps,
+				spadjStep{pos + epi - 1, 8},
+				spadjStep{pos + epi, 0},
+			)
+		}
 		code, ps, err := encodeInstr(s, pos, offsets, fi, long[i], resolve, link)
 		if err != nil {
-			return nil, nil, nil, fmt.Errorf("%s: %w", s.Mnemonic.Text, err)
+			return nil, nil, nil, nil, fmt.Errorf("%s: %w", s.Mnemonic.Text, err)
 		}
 		if len(code) != sizes[i] {
-			return nil, nil, nil, fmt.Errorf("%s: size mismatch (%d vs %d)", s.Mnemonic.Text, len(code), sizes[i])
+			return nil, nil, nil, nil, fmt.Errorf("%s: size mismatch (%d vs %d)", s.Mnemonic.Text, len(code), sizes[i])
 		}
 		patches = append(patches, ps...)
 		out = append(out, code...)
 		pos += len(code)
 	}
-	return out, patches, offsets, nil
+	return out, patches, offsets, steps, nil
 }
 
 // jumpChain precomputes jump-to-jump folding: a label whose first instruction
@@ -425,7 +456,9 @@ func operandFromAST(op *ast.Operand, size int, fi frameInfo, link *linkInfo) (Op
 				if a.Sym.Static {
 					return nil, fmt.Errorf("undefined symbol %q", a.Sym.Name)
 				}
-				return nil, fmt.Errorf("external symbol %q needs object-file emission", a.Sym.Name)
+				if !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
 		}

asm/elf.go

@@ -0,0 +1,301 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package asm
+
+import (
+	"encoding/binary"
+	"fmt"
+)
+
+// This file emits ELF64 relocatable objects (ET_REL) from an assembled
+// Image: a .text section holding the function bodies, a .data section
+// holding the GLOBL initialisers, a symbol table with one symbol per TEXT
+// and GLOBL (file-local <> symbols are STB_LOCAL, the rest STB_GLOBAL), and
+// a .rela.text relocation table — one R_X86_64_PC32 entry per static-symbol
+// reference, internal references resolving against the local data symbols
+// and external ones against undefined globals.  The output links with the
+// system toolchain (cc/ld) the way a hand-assembled .o would.
+
+// ELF constants (ELF64, little-endian, System V).
+const (
+	elfClass64 = 2
+	elfDataLSB = 1
+	elfVersion = 1
+
+	etREL   = 1 // relocatable object
+	emX8664 = 62
+
+	shtNull     = 0
+	shtProgbits = 1
+	shtSymtab   = 2
+	shtStrtab   = 3
+	shtRela     = 4
+
+	shfWrite     = 1
+	shfAlloc     = 2
+	shfExecInstr = 4
+
+	stbLocal  = 0
+	stbGlobal = 1
+
+	sttNotype   = 0
+	sttObject   = 1
+	sttFunc     = 2
+	sttSection  = 3
+	stInfoShift = 4
+
+	shnUndef = 0
+
+	rX8664PC32 = 2
+)
+
+// elfSym is one symbol-table entry in construction.
+type elfSym struct {
+	name  string
+	info  byte
+	shndx uint16
+	value uint64
+	size  uint64
+}
+
+// ELFObject returns the image as an ELF64 relocatable object file, ready for
+// the system linker.  Symbol names are the TEXT and GLOBL identifiers as
+// written (the middle dot stripped); a package prefix, when present, is
+// joined with a dot.  Every static-symbol reference becomes an
+// R_X86_64_PC32 relocation, so the code is position-independent and links
+// at any address.
+func (img *Image) ELFObject() ([]byte, error) {
+	le := binary.LittleEndian
+
+	// Section indices: 0 NULL, 1 .text, 2 .data; the tables follow.
+	const (
+		secText = 1
+		secData = 2
+	)
+
+	// Build the symbol table: the null entry and the two section symbols
+	// come first, then the local symbols (static TEXT and GLOBL), then the
+	// globals (exported TEXT and GLOBL, and the undefined externals) — ELF
+	// requires every local to precede every global, and sh_info records the
+	// boundary.  symIdx maps a symbol name to its index for the relocations.
+	var locals, globals []elfSym
+	for _, fn := range img.Funcs {
+		s := elfSym{
+			name:  objectName(fn.Pkg, fn.Name),
+			info:  sttFunc,
+			shndx: secText,
+			value: uint64(fn.Offset),
+			size:  uint64(fn.Size),
+		}
+		if fn.Static {
+			locals = append(locals, s)
+		} else {
+			s.info |= stbGlobal << stInfoShift
+			globals = append(globals, s)
+		}
+	}
+	for _, d := range img.DataSyms {
+		s := elfSym{
+			name:  objectName(d.Pkg, d.Name),
+			info:  sttObject,
+			shndx: secData,
+			value: uint64(d.Offset),
+			size:  uint64(d.Size),
+		}
+		if d.Static {
+			locals = append(locals, s)
+		} else {
+			s.info |= stbGlobal << stInfoShift
+			globals = append(globals, s)
+		}
+	}
+	for _, name := range img.Externals {
+		globals = append(globals, elfSym{name: name, info: stbGlobal << stInfoShift})
+	}
+	syms := []elfSym{
+		{}, // the mandatory null entry
+		{name: ".text", info: sttSection, shndx: secText},
+		{name: ".data", info: sttSection, shndx: secData},
+	}
+	syms = append(syms, locals...)
+	shInfo := len(syms) // first global symbol
+	syms = append(syms, globals...)
+	symIdx := map[string]int{}
+	for i, s := range syms {
+		symIdx[s.name] = i
+	}
+
+	// Build the relocations.
+	type elfRela struct {
+		off    uint64
+		sym    int
+		addend int64
+	}
+	var relas []elfRela
+	for _, fn := range img.Funcs {
+		for _, r := range fn.Relocs {
+			idx, ok := symIdx[r.Name]
+			if !ok {
+				return nil, fmt.Errorf("relocation references unknown symbol %q", r.Name)
+			}
+			relas = append(relas, elfRela{
+				off: uint64(fn.Offset + r.Off),
+				sym: idx,
+				// R_X86_64_PC32 computes S + A − P with P the patch site; the
+				// assembler measures the symbol from the instruction end,
+				// After − Off bytes past the field, so the addend carries
+				// that distance with a negative sign.
+				addend: r.Addend - int64(r.After-r.Off),
+			})
+		}
+	}
+
+	// Serialise the string tables.
+	stNames := newElfStrtab()
+	for _, s := range syms {
+		stNames.add(s.name)
+	}
+	stSections := newElfStrtab()
+	for _, n := range []string{".text", ".data", ".symtab", ".strtab", ".rela.text", ".shstrtab"} {
+		stSections.add(n)
+	}
+
+	// Section presence: .rela.text only when there are relocations.
+	hasRela := len(relas) > 0
+	nSections := 6 // NULL, .text, .data, .symtab, .strtab, .shstrtab
+	if hasRela {
+		nSections = 7
+	}
+	secSymtab, secStrtab := 3, 4
+	secShstr := nSections - 1
+
+	// Lay the file out: header, section data, section headers.
+	var out []byte
+	out = append(out, make([]byte, 64)...) // ELF header, filled last
+
+	align := func(n int) {
+		for len(out)%n != 0 {
+			out = append(out, 0)
+		}
+	}
+
+	align(16)
+	textOff := len(out)
+	out = append(out, img.Code...)
+
+	align(16)
+	dataOff := len(out)
+	out = append(out, img.Data...)
+
+	align(8)
+	symtabOff := len(out)
+	for _, s := range syms {
+		var b [24]byte
+		le.PutUint32(b[0:], uint32(stNames.at(s.name)))
+		b[4] = s.info
+		b[5] = 0 // st_other
+		le.PutUint16(b[6:], s.shndx)
+		le.PutUint64(b[8:], s.value)
+		le.PutUint64(b[16:], s.size)
+		out = append(out, b[:]...)
+	}
+
+	strtabOff := len(out)
+	out = append(out, stNames.bytes()...)
+
+	var relaOff int
+	if hasRela {
+		align(8)
+		relaOff = len(out)
+		for _, r := range relas {
+			var b [24]byte
+			le.PutUint64(b[0:], r.off)
+			le.PutUint64(b[8:], uint64(r.sym)<<32|rX8664PC32)
+			le.PutUint64(b[16:], uint64(r.addend))
+			out = append(out, b[:]...)
+		}
+	}
+
+	shstrOff := len(out)
+	out = append(out, stSections.bytes()...)
+
+	align(8)
+	shoff := len(out)
+
+	// Section headers.
+	putSh := func(name string, typ int, flags uint64, off, size int, link, info int, alignV, entsize uint64) {
+		var b [64]byte
+		le.PutUint32(b[0:], uint32(stSections.at(name)))
+		le.PutUint32(b[4:], uint32(typ))
+		le.PutUint64(b[8:], flags)
+		le.PutUint64(b[16:], 0) // sh_addr
+		le.PutUint64(b[24:], uint64(off))
+		le.PutUint64(b[32:], uint64(size))
+		le.PutUint32(b[40:], uint32(link))
+		le.PutUint32(b[44:], uint32(info))
+		le.PutUint64(b[48:], alignV)
+		le.PutUint64(b[56:], entsize)
+		out = append(out, b[:]...)
+	}
+	putSh("", shtNull, 0, 0, 0, 0, 0, 0, 0)
+	putSh(".text", shtProgbits, shfAlloc|shfExecInstr, textOff, len(img.Code), 0, 0, 16, 0)
+	putSh(".data", shtProgbits, shfAlloc|shfWrite, dataOff, len(img.Data), 0, 0, 16, 0)
+	putSh(".symtab", shtSymtab, 0, symtabOff, 24*len(syms), secStrtab, shInfo, 8, 24)
+	putSh(".strtab", shtStrtab, 0, strtabOff, len(stNames.bytes()), 0, 0, 1, 0)
+	if hasRela {
+		putSh(".rela.text", shtRela, 0, relaOff, 24*len(relas), secSymtab, secText, 8, 24)
+	}
+	putSh(".shstrtab", shtStrtab, 0, shstrOff, len(stSections.bytes()), 0, 0, 1, 0)
+
+	// The ELF header.
+	hdr := out[:64]
+	copy(hdr[0:], []byte{0x7f, 'E', 'L', 'F', elfClass64, elfDataLSB, elfVersion, 0})
+	le.PutUint16(hdr[16:], etREL)
+	le.PutUint16(hdr[18:], emX8664)
+	le.PutUint32(hdr[20:], elfVersion)
+	le.PutUint64(hdr[24:], 0)             // e_entry
+	le.PutUint64(hdr[32:], 0)             // e_phoff
+	le.PutUint64(hdr[40:], uint64(shoff)) // e_shoff
+	le.PutUint32(hdr[48:], 0)             // e_flags
+	le.PutUint16(hdr[52:], 64)            // e_ehsize
+	le.PutUint16(hdr[54:], 0)             // e_phentsize
+	le.PutUint16(hdr[56:], 0)             // e_phnum
+	le.PutUint16(hdr[58:], 64)            // e_shentsize
+	le.PutUint16(hdr[60:], uint16(nSections))
+	le.PutUint16(hdr[62:], uint16(secShstr))
+
+	return out, nil
+}
+
+// objectName renders a symbol's object-file name: the identifier as written,
+// with an explicit package prefix joined by a dot.
+func objectName(pkg, name string) string {
+	if pkg == "" {
+		return name
+	}
+	return pkg + "." + name
+}
+
+// elfStrtab is an ELF string table under construction.
+type elfStrtab struct {
+	buf []byte
+	off map[string]int
+}
+
+func newElfStrtab() *elfStrtab {
+	return &elfStrtab{buf: []byte{0}, off: map[string]int{"": 0}}
+}
+
+func (s *elfStrtab) add(name string) {
+	if _, ok := s.off[name]; ok {
+		return
+	}
+	s.off[name] = len(s.buf)
+	s.buf = append(s.buf, name...)
+	s.buf = append(s.buf, 0)
+}
+
+func (s *elfStrtab) at(name string) int { return s.off[name] }
+
+func (s *elfStrtab) bytes() []byte { return s.buf }

asm/elf_test.go

@@ -0,0 +1,310 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package asm
+
+import (
+	"bytes"
+	"debug/elf"
+	"encoding/binary"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"testing"
+
+	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+)
+
+// The object-file tests share one source: two exported functions, one
+// file-local constant reached through a relocation, and one external symbol
+// the linker must resolve.  The functions take their arguments in the System
+// V registers (not the Go stack ABI) so a C driver can call them directly.
+const elfTestSrc = `
+#include "textflag.h"
+
+TEXT ·addq(SB), NOSPLIT, $0
+	LEAQ (DI)(SI*1), AX
+	RET
+
+TEXT ·getanswer(SB), NOSPLIT, $0
+	MOVQ answer<>(SB), AX
+	RET
+
+TEXT ·useextern(SB), NOSPLIT, $0
+	MOVQ extvar(SB), AX
+	RET
+
+GLOBL answer<>(SB), RODATA, $8
+DATA answer<>+0(SB)/8, $42
+`
+
+func elfTestImage(t *testing.T) *Image {
+	t.Helper()
+	f, errs := parser.Parse("t_amd64.s", elfTestSrc)
+	if len(errs) > 0 {
+		t.Fatalf("parse: %v", errs)
+	}
+	img, err := AssembleFile(f)
+	if err != nil {
+		t.Fatalf("AssembleFile: %v", err)
+	}
+	return img
+}
+
+// TestAssembleFileExternals checks that a reference to a symbol no GLOBL
+// defines is recorded as an external relocation instead of failing — the
+// raw image leaves the displacement zero, the object emitters carry it.
+func TestAssembleFileExternals(t *testing.T) {
+	img := elfTestImage(t)
+	if len(img.Externals) != 1 || img.Externals[0] != "extvar" {
+		t.Fatalf("Externals = %v, want [extvar]", img.Externals)
+	}
+	var ext, local int
+	for _, fn := range img.Funcs {
+		for _, r := range fn.Relocs {
+			if r.External {
+				ext++
+				if r.Name != "extvar" {
+					t.Errorf("external reloc names %q, want extvar", r.Name)
+				}
+			} else {
+				local++
+				if r.Name != "answer" {
+					t.Errorf("local reloc names %q, want answer", r.Name)
+				}
+			}
+		}
+	}
+	if ext != 1 || local != 1 {
+		t.Errorf("relocs = %d external, %d local; want 1 and 1", ext, local)
+	}
+}
+
+// TestELFObject checks the structure of the emitted ELF64 relocatable
+// object: sections, the symbol table (bindings, types, values, sizes) and
+// the .rela.text relocations, parsed back with debug/elf.
+func TestELFObject(t *testing.T) {
+	img := elfTestImage(t)
+	obj, err := img.ELFObject()
+	if err != nil {
+		t.Fatalf("ELFObject: %v", err)
+	}
+	f, err := elf.NewFile(bytes.NewReader(obj))
+	if err != nil {
+		t.Fatalf("parse emitted object: %v", err)
+	}
+	defer f.Close()
+
+	if f.Type != elf.ET_REL || f.Machine != elf.EM_X86_64 {
+		t.Errorf("type/machine = %v/%v, want ET_REL/EM_X86_64", f.Type, f.Machine)
+	}
+
+	text := f.Section(".text")
+	data := f.Section(".data")
+	if text == nil || data == nil {
+		t.Fatal("missing .text or .data section")
+	}
+	if text.Flags&elf.SHF_EXECINSTR == 0 || text.Flags&elf.SHF_ALLOC == 0 {
+		t.Errorf(".text flags = %v", text.Flags)
+	}
+	if data.Flags&elf.SHF_WRITE == 0 {
+		t.Errorf(".data flags = %v", data.Flags)
+	}
+	textData, err := text.Data()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !bytes.Equal(textData, img.Code) {
+		t.Errorf(".text contents differ from the image code")
+	}
+
+	syms, err := f.Symbols()
+	if err != nil {
+		t.Fatalf("symbols: %v", err)
+	}
+	byName := map[string]elf.Symbol{}
+	for _, s := range syms {
+		byName[s.Name] = s
+	}
+	wantSym := func(name string, bind elf.SymBind, typ elf.SymType, section elf.SectionIndex, size uint64) {
+		t.Helper()
+		s, ok := byName[name]
+		if !ok {
+			t.Errorf("symbol %q not found", name)
+			return
+		}
+		if elf.ST_BIND(s.Info) != bind || elf.ST_TYPE(s.Info) != typ {
+			t.Errorf("%s: bind/type = %v/%v, want %v/%v", name, elf.ST_BIND(s.Info), elf.ST_TYPE(s.Info), bind, typ)
+		}
+		if s.Section != section {
+			t.Errorf("%s: section = %v, want %v", name, s.Section, section)
+		}
+		if s.Size != size {
+			t.Errorf("%s: size = %d, want %d", name, s.Size, size)
+		}
+	}
+	// The emitted layout is fixed: 0 NULL, 1 .text, 2 .data.
+	if f.Sections[1].Name != ".text" || f.Sections[2].Name != ".data" {
+		t.Fatalf("section layout = %s, %s; want .text, .data", f.Sections[1].Name, f.Sections[2].Name)
+	}
+	textIdx := elf.SectionIndex(1)
+	dataIdx := elf.SectionIndex(2)
+	wantSym("addq", elf.STB_GLOBAL, elf.STT_FUNC, textIdx, 5)
+	wantSym("getanswer", elf.STB_GLOBAL, elf.STT_FUNC, textIdx, 8)
+	wantSym("useextern", elf.STB_GLOBAL, elf.STT_FUNC, textIdx, 8)
+	wantSym("answer", elf.STB_LOCAL, elf.STT_OBJECT, dataIdx, 8)
+	wantSym("extvar", elf.STB_GLOBAL, elf.STT_NOTYPE, elf.SHN_UNDEF, 0)
+
+	// Relocations: one for the file-local constant (resolving against the
+	// local data symbol) and one for the external (against the undefined
+	// global), both R_X86_64_PC32 with the −4 addend the PC-relative form
+	// needs.  debug/elf does not surface rela entries, so read the section
+	// directly.
+	relaSec := f.Section(".rela.text")
+	if relaSec == nil {
+		t.Fatal("missing .rela.text")
+	}
+	raw, err := relaSec.Data()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(raw)%24 != 0 || len(raw)/24 != 2 {
+		t.Fatalf(".rela.text has %d bytes, want two 24-byte entries", len(raw))
+	}
+	// Symbol names straight from the raw tables: r_info carries an index
+	// into .symtab including the null entry, which debug/elf's Symbols()
+	// slice may not mirror.
+	symtabRaw, err := f.Section(".symtab").Data()
+	if err != nil {
+		t.Fatal(err)
+	}
+	strtabRaw, err := f.Section(".strtab").Data()
+	if err != nil {
+		t.Fatal(err)
+	}
+	symName := func(idx int) string {
+		stName := binary.LittleEndian.Uint32(symtabRaw[idx*24:])
+		end := bytes.IndexByte(strtabRaw[stName:], 0)
+		return string(strtabRaw[stName : int(stName)+end])
+	}
+	for i := 0; i < 2; i++ {
+		e := raw[i*24 : (i+1)*24]
+		off := binary.LittleEndian.Uint64(e[0:])
+		info := binary.LittleEndian.Uint64(e[8:])
+		addend := int64(binary.LittleEndian.Uint64(e[16:]))
+		typ := info & 0xffffffff
+		sym := int(info >> 32)
+		if typ != uint64(elf.R_X86_64_PC32) {
+			t.Errorf("reloc %d: type %d, want R_X86_64_PC32", i, typ)
+		}
+		if addend != -4 {
+			t.Errorf("reloc %d: addend %d, want -4", i, addend)
+		}
+		if name := symName(sym); name != "answer" && name != "extvar" {
+			t.Errorf("reloc %d: symbol %q, want answer or extvar", i, name)
+		}
+		// The relocation offset lands on the disp32 field: the four bytes
+		// before a RET-terminated eight-byte MOVQ.
+		if off+4 > uint64(len(textData)) {
+			t.Errorf("reloc %d: offset %d outside .text", i, off)
+		}
+	}
+}
+
+// TestELFObjectNoRelocations checks a file with no static-symbol references
+// emits a valid object without a .rela.text section.
+func TestELFObjectNoRelocations(t *testing.T) {
+	f, errs := parser.Parse("n_amd64.s", `
+#include "textflag.h"
+TEXT ·nop(SB), NOSPLIT, $0
+	RET
+`)
+	if len(errs) > 0 {
+		t.Fatalf("parse: %v", errs)
+	}
+	img, err := AssembleFile(f)
+	if err != nil {
+		t.Fatalf("AssembleFile: %v", err)
+	}
+	obj, err := img.ELFObject()
+	if err != nil {
+		t.Fatalf("ELFObject: %v", err)
+	}
+	ef, err := elf.NewFile(bytes.NewReader(obj))
+	if err != nil {
+		t.Fatalf("parse emitted object: %v", err)
+	}
+	defer ef.Close()
+	if ef.Section(".rela.text") != nil {
+		t.Error("unexpected .rela.text section")
+	}
+	syms, err := ef.Symbols()
+	if err != nil {
+		t.Fatal(err)
+	}
+	found := false
+	for _, s := range syms {
+		if s.Name == "nop" && elf.ST_TYPE(s.Info) == elf.STT_FUNC {
+			found = true
+		}
+	}
+	if !found {
+		t.Error("function symbol nop not found")
+	}
+}
+
+// TestELFLinkAndRun is the end-to-end check: assemble the test functions,
+// link the emitted object with a C driver that defines the external symbol,
+// and run the result.  Skipped when no C compiler is available.
+func TestELFLinkAndRun(t *testing.T) {
+	cc, err := exec.LookPath("cc")
+	if err != nil {
+		t.Skip("no C compiler available")
+	}
+	dir := t.TempDir()
+
+	img := elfTestImage(t)
+	obj, err := img.ELFObject()
+	if err != nil {
+		t.Fatalf("ELFObject: %v", err)
+	}
+	objPath := filepath.Join(dir, "t.o")
+	if err := os.WriteFile(objPath, obj, 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	const driver = `
+#include <stdio.h>
+
+long addq(long a, long b);
+long getanswer(void);
+long useextern(void);
+
+long extvar = 7;
+
+int main(void) {
+	printf("%ld %ld %ld\n", addq(41, 1), getanswer(), useextern());
+	return 0;
+}
+`
+	driverPath := filepath.Join(dir, "driver.c")
+	if err := os.WriteFile(driverPath, []byte(driver), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	// -no-pie: the encoder emits R_X86_64_PC32 for external references,
+	// which a position-independent executable would reject (it wants
+	// PLT32/GOT relocations, a future increment).
+	appPath := filepath.Join(dir, "app")
+	out, err := exec.Command(cc, "-no-pie", "-o", appPath, driverPath, objPath).CombinedOutput()
+	if err != nil {
+		t.Fatalf("link failed: %v\n%s", err, out)
+	}
+	run, err := exec.Command(appPath).CombinedOutput()
+	if err != nil {
+		t.Fatalf("run failed: %v\n%s", err, run)
+	}
+	if got := string(run); got != "42 42 7\n" {
+		t.Errorf("output %q, want \"42 42 7\\n\"", got)
+	}
+}

asm/encode.go

@@ -51,9 +51,17 @@ func (e *enc) encode(mnem string, ops []Operand) error {
 
 	// VEX (AVX/AVX2) and EVEX (AVX-512) instructions: the trailing
 	// B/W/L/Q/D is part of the mnemonic, not a size suffix, so dispatch
-	// before splitSize.
+	// before splitSize.  EVEX suffixes (.Z, .SAE, rounding, .BCST) split
-	if isVex(upper) || isEvex(upper) || upper == "KMOVW" {
+	// off the mnemonic too.
-		return e.encodeVec(upper, ops)
+	base, sfx, err := parseEvexSuffix(upper)
+	if err != nil {
+		return err
+	}
+	if isVex(base) || isEvex(base) || isKOp(base) || isGather(base) || isScatter(base) || base == "KMOVW" || base == "KMOVQ" {
+		return e.encodeVec(base, ops, sfx)
+	}
+	if sfx.any() {
+		return fmt.Errorf("%s: the suffix requires an EVEX instruction", mnem)
 	}
 
 	// CMOVcc and SETcc carry the condition in the mnemonic (CMOVLGT, SETNE).
@@ -93,6 +101,8 @@ func (e *enc) encode(mnem string, ops []Operand) error {
 		return e.encodeMovExtend(base, ops)
 	case "CVTSL2SD", "CVTSQ2SD":
 		return e.encodeCvtsi2sd(base == "CVTSQ2SD", ops)
+	case "MOVOU", "MOVO", "MOVUPS", "MOVAPS", "MOVUPD", "MOVAPD", "MOVSD", "MOVSS":
+		return e.encodeSSEMove(sseMoveTable[base], ops)
 	}
 	return fmt.Errorf("unsupported instruction %q", mnem)
 }
@@ -119,14 +129,32 @@ func splitSize(upper string) (base string, size int) {
 // its own direction-dependent opcodes; KTESTW is always VEX; everything else
 // takes EVEX when an operand demands it (a ZMM or K register, or an
 // EVEX-only mnemonic) and VEX otherwise.
-func (e *enc) encodeVec(upper string, ops []Operand) error {
+func (e *enc) encodeVec(upper string, ops []Operand, sfx evexSuffix) error {
-	if upper == "KMOVW" {
+	if gs, ok := gatherTable[upper]; ok {
-		return e.encodeKmovw(ops)
+		return e.encodeGather(upper, gs, ops, sfx)
 	}
-	if upper == "KTESTW" || !evexRequired(upper, ops) {
+	if ss, ok := scatterTable[upper]; ok {
+		return e.encodeScatter(upper, ss, ops, sfx)
+	}
+	if upper == "KMOVW" || upper == "KMOVQ" {
+		if sfx.any() {
+			return fmt.Errorf("%s takes no EVEX suffixes", upper)
+		}
+		return e.encodeKmov(upper, ops)
+	}
+	if isKOp(upper) {
+		if sfx.any() {
+			return fmt.Errorf("%s takes no EVEX suffixes", upper)
+		}
+		return e.encodeKOp(upper, ops)
+	}
+	if upper == "KTESTW" || (!evexRequired(upper, ops) && !sfx.evexOnly()) {
+		if sfx.any() {
+			return fmt.Errorf("%s: the .Z suffix requires an EVEX instruction", upper)
+		}
 		return e.encodeVex(upper, ops)
 	}
-	return e.encodeEvex(upper, ops)
+	return e.encodeEvex(upper, ops, sfx)
 }
 
 // --- instruction components -------------------------------------------------

asm/encode_test.go

@@ -127,6 +127,52 @@ func TestControl(t *testing.T) {
 	checkOp(t, x86asm.JBE, "JLS", Imm(0))
 }
 
+// TestSSEMoveGroundTruth checks the legacy (non-VEX) SSE moves byte for byte
+// against the Go assembler.  wantOp is the decoder's name, which differs from
+// the Plan 9 spelling for the octa moves (MOVOU = MOVDQU, MOVO = MOVDQA).
+func TestSSEMoveGroundTruth(t *testing.T) {
+	cases := []struct {
+		name   string
+		mnem   string
+		ops    []Operand
+		want   string
+		wantOp string
+	}{
+		{"MOVOU (SI),X1", "MOVOU", []Operand{Ptr(SI, 0, 16), vreg(t, "X1")}, "f30f6f0e", "MOVDQU"},
+		{"MOVOU X3,(DI)", "MOVOU", []Operand{vreg(t, "X3"), Ptr(DI, 0, 16)}, "f30f7f1f", "MOVDQU"},
+		{"MOVOU X1,X2", "MOVOU", []Operand{vreg(t, "X1"), vreg(t, "X2")}, "f30f6fd1", "MOVDQU"},
+		{"MOVOU (SI)(BX*4),X9", "MOVOU", []Operand{Idx(SI, BX, 4, 0, 16), vreg(t, "X9")}, "f3440f6f0c9e", "MOVDQU"},
+		{"MOVO (SI),X1", "MOVO", []Operand{Ptr(SI, 0, 16), vreg(t, "X1")}, "660f6f0e", "MOVDQA"},
+		{"MOVO X3,(DI)", "MOVO", []Operand{vreg(t, "X3"), Ptr(DI, 0, 16)}, "660f7f1f", "MOVDQA"},
+		{"MOVUPS (SI),X1", "MOVUPS", []Operand{Ptr(SI, 0, 16), vreg(t, "X1")}, "0f100e", "MOVUPS"},
+		{"MOVAPS X3,(DI)", "MOVAPS", []Operand{vreg(t, "X3"), Ptr(DI, 0, 16)}, "0f291f", "MOVAPS"},
+		{"MOVUPD (SI),X1", "MOVUPD", []Operand{Ptr(SI, 0, 16), vreg(t, "X1")}, "660f100e", "MOVUPD"},
+		{"MOVAPD X3,(DI)", "MOVAPD", []Operand{vreg(t, "X3"), Ptr(DI, 0, 16)}, "660f291f", "MOVAPD"},
+		{"MOVSD (SI),X1", "MOVSD", []Operand{Ptr(SI, 0, 8), vreg(t, "X1")}, "f20f100e", "MOVSD_XMM"},
+		{"MOVSD X1,X2", "MOVSD", []Operand{vreg(t, "X1"), vreg(t, "X2")}, "f20f10d1", "MOVSD_XMM"},
+		{"MOVSS X3,(DI)", "MOVSS", []Operand{vreg(t, "X3"), Ptr(DI, 0, 4)}, "f30f111f", "MOVSS"},
+	}
+	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)
+			continue
+		}
+		inst, err := x86asm.Decode(code, 64)
+		if err != nil {
+			t.Errorf("%s: Decode(%x): %v", c.name, code, err)
+			continue
+		}
+		if inst.Op.String() != c.wantOp {
+			t.Errorf("%s: decoded as %s", c.name, inst.Op.String())
+		}
+	}
+}
+
 // TestGoFlacScalarTail encodes the scalar tail of an analyze kernel to confirm
 // the encoder handles a realistic instruction sequence.
 func TestGoFlacScalarTail(t *testing.T) {

asm/evex_test.go

@@ -61,6 +61,27 @@ func TestEvexGroundTruth(t *testing.T) {
 		{"VMOVDQU32 16(SI)(R15*4),Z4", "VMOVDQU32", []Operand{Idx(SI, vreg(t, "R15"), 4, 16, 64), vreg(t, "Z4")}, "62b17e486fa4be10000000"},
 		{"VMOVDQU32 Z0,4(SI)(AX*1)", "VMOVDQU32", []Operand{vreg(t, "Z0"), Idx(SI, AX, 1, 4, 64)}, "62f17e487f840604000000"},
 		{"VMOVDQU32 Z3,(DI)(R15*4)", "VMOVDQU32", []Operand{vreg(t, "Z3"), Idx(DI, vreg(t, "R15"), 4, 0, 64)}, "62b17e487f1cbf"},
+		// VMOVDQU64 — the W1 qword variant.
+		{"VMOVDQU64 (SI)(R15*4),Z3", "VMOVDQU64", []Operand{Idx(SI, vreg(t, "R15"), 4, 0, 64), vreg(t, "Z3")}, "62b1fe486f1cbe"},
+		{"VMOVDQU64 Z0,4(SI)(AX*1)", "VMOVDQU64", []Operand{vreg(t, "Z0"), Idx(SI, AX, 1, 4, 64)}, "62f1fe487f840604000000"},
+		{"VMOVDQU64 Z1,Z2", "VMOVDQU64", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f1fe487fca"},
+		// The wider AVX-512 F/BW integer set.
+		{"VPADDB Z1,Z2,Z3", "VPADDB", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f16d48fcd9"},
+		{"VPSUBW Z1,Z2,Z3", "VPSUBW", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f16d48f9d9"},
+		{"VPANDQ Z1,Z2,Z3", "VPANDQ", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f1ed48dbd9"},
+		{"VPANDND Z1,Z2,Z3", "VPANDND", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f16d48dfd9"},
+		{"VPMULLW Z1,Z2,Z3", "VPMULLW", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f16d48d5d9"},
+		{"VPMINUB Z1,Z2,Z3", "VPMINUB", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f16d48dad9"},
+		{"VPMAXUQ Z1,Z2,Z3", "VPMAXUQ", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f2ed483fd9"},
+		{"VPAVGW Z1,Z2,Z3", "VPAVGW", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f16d48e3d9"},
+		{"VPSLLVQ Z3,Z1,Z2", "VPSLLVQ", []Operand{vreg(t, "Z3"), vreg(t, "Z1"), vreg(t, "Z2")}, "62f2f54847d3"},
+		{"VPSRAVQ Z3,Z1,Z2", "VPSRAVQ", []Operand{vreg(t, "Z3"), vreg(t, "Z1"), vreg(t, "Z2")}, "62f2f54846d3"},
+		{"VPSHUFD $0x1B,Z1,Z2", "VPSHUFD", []Operand{Imm(0x1B), vreg(t, "Z1"), vreg(t, "Z2")}, "62f17d4870d11b"},
+		{"VPSHUFB Z1,Z2,Z3", "VPSHUFB", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f26d4800d9"},
+		{"VMOVDQU8 Z1,Z2", "VMOVDQU8", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f17f487fca"},
+		{"VMOVDQU16 Z1,Z2", "VMOVDQU16", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f1ff487fca"},
+		// Indices 16–31: rm[4] rides in X̄ for register operands.
+		{"VPSHUFD $1,X16,X17", "VPSHUFD", []Operand{Imm(1), vreg(t, "X16"), vreg(t, "X17")}, "62a17d0870c801"},
 		{"VMOVUPD (DI),Z14", "VMOVUPD", []Operand{Ptr(DI, 0, 64), vreg(t, "Z14")}, "6271fd481037"},
 		{"VMOVUPD 64(DI),Z14", "VMOVUPD", []Operand{Ptr(DI, 64, 64), vreg(t, "Z14")}, "6271fd48107701"},
 		// Conversions and narrowing stores (reg = wide source).
@@ -80,6 +101,32 @@ func TestEvexGroundTruth(t *testing.T) {
 		{"VPBROADCASTQ AX,Z9", "VPBROADCASTQ", []Operand{AX, vreg(t, "Z9")}, "6272fd487cc8"},
 		// Register indices 16–31 exist only in EVEX encodings.
 		{"VPBROADCASTD AX,Y30", "VPBROADCASTD", []Operand{AX, vreg(t, "Y30")}, "62627d287cf0"},
+		// Packed double arithmetic / unpack (EVEX forms carry W=1).
+		{"VSUBPD Z1,Z2,Z3", "VSUBPD", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f1ed485cd9"},
+		{"VDIVPD Z4,Z5,Z6", "VDIVPD", []Operand{vreg(t, "Z4"), vreg(t, "Z5"), vreg(t, "Z6")}, "62f1d5485ef4"},
+		{"VMINPD Z7,Z8,Z9", "VMINPD", []Operand{vreg(t, "Z7"), vreg(t, "Z8"), vreg(t, "Z9")}, "6271bd485dcf"},
+		{"VMAXPD Z10,Z11,Z12", "VMAXPD", []Operand{vreg(t, "Z10"), vreg(t, "Z11"), vreg(t, "Z12")}, "6251a5485fe2"},
+		{"VUNPCKLPD Z1,Z2,Z3", "VUNPCKLPD", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f1ed4814d9"},
+		{"VUNPCKHPD Z1,Z2,Z3", "VUNPCKHPD", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f1ed4815d9"},
+		{"VSUBPD 64(AX),Z1,Z2", "VSUBPD", []Operand{Ptr(AX, 64, 64), vreg(t, "Z1"), vreg(t, "Z2")}, "62f1f5485c5001"},
+		{"VSUBPD Z17,Z18,Z19", "VSUBPD", []Operand{vreg(t, "Z17"), vreg(t, "Z18"), vreg(t, "Z19")}, "62a1ed405cd9"},
+		// VMOVDDUP — duplicate the low double; disp8×N = 64 at 512 bits, and
+		// X16/X17 force EVEX (the mod=11 rm[4] extension rides in X̄).
+		{"VMOVDDUP Z1,Z2", "VMOVDDUP", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f1ff4812d1"},
+		{"VMOVDDUP 64(AX),Z1", "VMOVDDUP", []Operand{Ptr(AX, 64, 64), vreg(t, "Z1")}, "62f1ff48124801"},
+		{"VMOVDDUP X16,X17", "VMOVDDUP", []Operand{vreg(t, "X16"), vreg(t, "X17")}, "62a1ff0812c8"},
+		// Conversions: DQ→PS, PS→PD (pp = 00, the Go assembler's choice),
+		// DQ→PD (the destination sets the length).
+		{"VCVTDQ2PS Z1,Z2", "VCVTDQ2PS", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f17c485bd1"},
+		{"VCVTPS2PD Y1,Z2", "VCVTPS2PD", []Operand{vreg(t, "Y1"), vreg(t, "Z2")}, "62f17c485ad1"},
+		{"VCVTPS2PD 32(AX),Z2", "VCVTPS2PD", []Operand{Ptr(AX, 32, 32), vreg(t, "Z2")}, "62f17c485a5001"},
+		{"VCVTDQ2PD Y1,Z2", "VCVTDQ2PD", []Operand{vreg(t, "Y1"), vreg(t, "Z2")}, "62f17e48e6d1"},
+		// PD→DQ conversions: the source is the wide operand and fixes the
+		// length (ZMM source → L'L = 10 even with an XMM destination; a
+		// memory source takes the length the mnemonic's spelling implies).
+		{"VCVTPD2DQ Z1,Y2", "VCVTPD2DQ", []Operand{vreg(t, "Z1"), vreg(t, "Y2")}, "62f1ff48e6d1"},
+		{"VCVTPD2DQ 64(AX),Y2", "VCVTPD2DQ", []Operand{Ptr(AX, 64, 64), vreg(t, "Y2")}, "62f1ff48e65001"},
+		{"VCVTTPD2DQ Z3,Y4", "VCVTTPD2DQ", []Operand{vreg(t, "Z3"), vreg(t, "Y4")}, "62f1fd48e6e3"},
 	}
 	for _, c := range cases {
 		want := strings.ReplaceAll(c.want, " ", "")
@@ -106,6 +153,478 @@ func TestEvexGroundTruth(t *testing.T) {
 	}
 }
 
+// TestEvexMasking checks the AVX-512 mask operand (K1–K7, placed freely among
+// the operands) and the .Z zeroing suffix, byte for byte against the Go
+// assembler.
+func TestEvexMasking(t *testing.T) {
+	cases := []struct {
+		name string
+		mnem string
+		ops  []Operand
+		want string
+	}{
+		// Masked arithmetic: K anywhere among the operands; .Z sets the z bit.
+		{"VPADDD.Z merging+zeroing", "VPADDD.Z", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "K2"), vreg(t, "Z3")}, "62f16dcafed9"},
+		{"VPADDD merging", "VPADDD", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "K1"), vreg(t, "Z3")}, "62f16d49fed9"},
+		{"VADDPD.Z", "VADDPD.Z", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "K2"), vreg(t, "Z3")}, "62f1edca58d9"},
+		{"VPMINSD.Z", "VPMINSD.Z", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "K5"), vreg(t, "Z3")}, "62f26dcd39d9"},
+		{"VPMINSQ.Z", "VPMINSQ.Z", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "K5"), vreg(t, "Z3")}, "62f2edcd39d9"},
+		// Masked immediate shift (K before the destination).
+		{"VPSRAD.Z", "VPSRAD.Z", []Operand{Imm(1), vreg(t, "Z2"), vreg(t, "K1"), vreg(t, "Z3")}, "62f165c972e201"},
+		{"VPSLLD merge", "VPSLLD", []Operand{Imm(4), vreg(t, "Z1"), vreg(t, "K2"), vreg(t, "Z3")}, "62f1654a72f104"},
+		// Masked align.
+		{"VALIGND", "VALIGND", []Operand{Imm(12), vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "K3"), vreg(t, "Z4")}, "62f36d4b03e10c"},
+		// Masked conversion and extract.
+		{"VCVTQQ2PD.Z", "VCVTQQ2PD.Z", []Operand{vreg(t, "Z1"), vreg(t, "K2"), vreg(t, "Z3")}, "62f1fecae6d9"},
+		{"VEXTRACTI64X4", "VEXTRACTI64X4", []Operand{Imm(1), vreg(t, "Z1"), vreg(t, "K2"), vreg(t, "Y3")}, "62f3fd4a3bcb01"},
+		// Masked moves: K sits between the register and memory operands.
+		{"VMOVDQU8 store", "VMOVDQU8", []Operand{vreg(t, "Z1"), vreg(t, "K3"), Ptr(SI, 0, 64)}, "62f17f4b7f0e"},
+		{"VMOVDQU32 load", "VMOVDQU32", []Operand{Ptr(SI, 0, 64), vreg(t, "K4"), vreg(t, "Z1")}, "62f17e4c6f0e"},
+		{"VMOVDQU32 store", "VMOVDQU32", []Operand{vreg(t, "Z1"), vreg(t, "K4"), Ptr(DI, 0, 64)}, "62f17e4c7f0f"},
+		// Masked comparison with a K destination: dst K1, mask K2.
+		{"VPCMPEQD k-dst+mask", "VPCMPEQD", []Operand{vreg(t, "Z0"), vreg(t, "Z3"), vreg(t, "K2"), vreg(t, "K1")}, "62f1654a76c8"},
+		// Masked floating point: packed double, the scalar SD/SS forms (which
+		// exist under EVEX only for masked and zeroing use) and conversions.
+		{"VSUBPD.Z", "VSUBPD.Z", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "K3"), vreg(t, "Z4")}, "62f1edcb5ce1"},
+		{"VADDSD merge", "VADDSD", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "K3"), vreg(t, "X4")}, "62f1ef0b58e1"},
+		{"VSUBSD.Z", "VSUBSD.Z", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "K5"), vreg(t, "X3")}, "62f1ef8d5cd9"},
+		{"VADDSS merge", "VADDSS", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "K1"), vreg(t, "X3")}, "62f16e0958d9"},
+		{"VCVTPD2DQ merge", "VCVTPD2DQ", []Operand{vreg(t, "Z1"), vreg(t, "K2"), vreg(t, "Y3")}, "62f1ff4ae6d9"},
+		{"VCVTTPD2DQ.Z", "VCVTTPD2DQ.Z", []Operand{vreg(t, "Z1"), vreg(t, "K2"), vreg(t, "Y3")}, "62f1fdcae6d9"},
+		{"VCVTDQ2PS.Z", "VCVTDQ2PS.Z", []Operand{vreg(t, "Z1"), vreg(t, "K4"), vreg(t, "Z2")}, "62f17ccc5bd1"},
+		{"VCVTDQ2PD merge", "VCVTDQ2PD", []Operand{vreg(t, "X1"), vreg(t, "K2"), vreg(t, "X3")}, "62f17e0ae6d9"},
+		{"VCVTDQ2PD.Z", "VCVTDQ2PD.Z", []Operand{vreg(t, "Y1"), vreg(t, "K2"), vreg(t, "Z2")}, "62f17ecae6d1"},
+		{"VCVTPS2PD.Z", "VCVTPS2PD.Z", []Operand{vreg(t, "Y1"), vreg(t, "K3"), vreg(t, "Z2")}, "62f17ccb5ad1"},
+	}
+	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)
+			continue
+		}
+		inst, err := x86asm.Decode(code, 64)
+		if err != nil {
+			t.Errorf("%s: Decode(%x): %v", c.name, code, err)
+			continue
+		}
+		want := c.mnem
+		if i := len(want) - 2; i > 0 && want[i:] == ".Z" {
+			want = want[:i]
+		}
+		if inst.Op.String() != want {
+			t.Errorf("%s: decoded as %s", c.name, inst.Op.String())
+		}
+	}
+
+	// Error cases.
+	bad := []struct {
+		name string
+		mnem string
+		ops  []Operand
+	}{
+		{"zeroing without mask", "VPADDD.Z", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}},
+		{"K0 mask", "VPADDD", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "K0"), vreg(t, "Z3")}},
+		{"two masks", "VPADDD", []Operand{vreg(t, "Z1"), vreg(t, "K1"), vreg(t, "K2"), vreg(t, "Z3")}},
+		{".Z on VEX-only", "VPSHUFD.Z", []Operand{Imm(1), vreg(t, "X0"), vreg(t, "X1")}},
+		{"broadcast unsupported", "VPXORD.BCST", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}},
+		{"rounding unsupported", "VPXORD.RN_SAE", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}},
+		{"bcst with rounding", "VADDPD.BCST.RN_SAE", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}},
+		{"Z not last", "VADDPD.Z.RN_SAE", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}},
+		{"duplicate suffix", "VADDPD.Z.Z", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}},
+		{"KMOVW.Z", "KMOVW.Z", []Operand{vreg(t, "K1"), vreg(t, "K2")}},
+	}
+	for _, c := range bad {
+		if _, err := Encode(c.mnem, c.ops...); err == nil {
+			t.Errorf("%s: expected an error, got none", c.name)
+		}
+	}
+}
+
+// TestEvexExtendedGroundTruth covers the wider EVEX/AVX-512 set — ternary
+// logic, lane shuffles/inserts/extracts, compares with a K destination,
+// permutes, the wider integer families, expand/compress, broadcasts,
+// rotates and word shifts, the opmask instructions, the EVEX suffixes
+// (rounding/SAE/broadcast) and the aligned/scalar moves — byte for byte
+// against the Go assembler.
+func TestEvexExtendedGroundTruth(t *testing.T) {
+	mem64 := func(base Reg) Operand { return Ptr(base, 0, 64) }
+	cases := []struct {
+		name string
+		mnem string
+		ops  []Operand
+		want string
+	}{
+		// Ternary logic and lane shuffles (NDS + imm8).
+		{"VPTERNLOGD", "VPTERNLOGD", []Operand{Imm(0xE8), vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f36d4825d9e8"},
+		{"VPTERNLOGQ", "VPTERNLOGQ", []Operand{Imm(0x96), vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f3ed4825d996"},
+		{"VSHUFI32X4", "VSHUFI32X4", []Operand{Imm(0x4E), vreg(t, "Y1"), vreg(t, "Y2"), vreg(t, "Y3")}, "62f36d2843d94e"},
+		{"VSHUFF64X2", "VSHUFF64X2", []Operand{Imm(1), vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f3ed4823d901"},
+		{"VPALIGNR", "VPALIGNR", []Operand{Imm(7), vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f36d480fd907"},
+		// Permutes.
+		{"VPERMB", "VPERMB", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f26d488dd9"},
+		{"VPERMW", "VPERMW", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f2ed488dd9"},
+		{"VPERMI2D", "VPERMI2D", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f26d4876d9"},
+		{"VPERMT2PD", "VPERMT2PD", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f2ed487fd9"},
+		// Compare with a K destination (and an immediate predicate).
+		{"VCMPPD", "VCMPPD", []Operand{Imm(4), vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "K3")}, "62f1ed48c2d904"},
+		{"VCMPPS", "VCMPPS", []Operand{Imm(0), vreg(t, "Y1"), vreg(t, "Y2"), vreg(t, "K4")}, "62f16c28c2e100"},
+		{"VCMPSD", "VCMPSD", []Operand{Imm(17), vreg(t, "X1"), vreg(t, "X2"), vreg(t, "K5")}, "62f1ef08c2e911"},
+		// Rounding / SAE / broadcast suffixes.
+		{"VADDPD.RN_SAE", "VADDPD.RN_SAE", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f1ed1858d9"},
+		{"VMULPD.RZ_SAE.Z", "VMULPD.RZ_SAE.Z", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "K1"), vreg(t, "Z3")}, "62f1edf959d9"},
+		{"VMAXPD.SAE", "VMAXPD.SAE", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f1ed585fd9"},
+		{"VADDPD.BCST", "VADDPD.BCST", []Operand{mem64(AX), vreg(t, "Z1"), vreg(t, "Z2")}, "62f1f5585810"},
+		// Packed single arithmetic (same opcodes, no mandatory prefix) —
+		// ZMM, YMM and XMM widths, rounding and broadcast.
+		{"VADDPS", "VADDPS", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f16c4858d9"},
+		{"VMULPS", "VMULPS", []Operand{vreg(t, "Y1"), vreg(t, "Y2"), vreg(t, "Y3")}, "c5ec59d9"},
+		{"VMAXPS", "VMAXPS", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "c5e85fd9"},
+		{"VDIVPS.RD_SAE", "VDIVPS.RD_SAE", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f16c385ed9"},
+		{"VADDPS.BCST", "VADDPS.BCST", []Operand{mem64(AX), vreg(t, "Z1"), vreg(t, "Z2")}, "62f174585810"},
+		// Compress / expand.
+		{"VCOMPRESSPD", "VCOMPRESSPD", []Operand{vreg(t, "Z1"), mem64(DI)}, "62f2fd488a0f"},
+		{"VEXPANDPS", "VEXPANDPS", []Operand{mem64(SI), vreg(t, "Y2")}, "62f27d288816"},
+		{"VPCOMPRESSD.Z", "VPCOMPRESSD.Z", []Operand{vreg(t, "Z1"), vreg(t, "K2"), mem64(DI)}, "62f27dca8b0f"},
+		// Broadcasts.
+		{"VPBROADCASTB gpr", "VPBROADCASTB", []Operand{BX, vreg(t, "Z1")}, "62f27d487acb"},
+		{"VPBROADCASTW mem", "VPBROADCASTW", []Operand{mem64(AX), vreg(t, "Z2")}, "62f27d487910"},
+		{"VBROADCASTSS", "VBROADCASTSS", []Operand{mem64(AX), vreg(t, "Y3")}, "c4e27d1818"},
+		{"VBROADCASTSD", "VBROADCASTSD", []Operand{mem64(AX), vreg(t, "Z4")}, "62f2fd481920"},
+		// Wider integer families.
+		{"VPMADDWD", "VPMADDWD", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f16d48f5d9"},
+		{"VPMADDUBSW", "VPMADDUBSW", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f26d4804d9"},
+		{"VPMULHUW", "VPMULHUW", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f16d48e4d9"},
+		{"VPSLLVW", "VPSLLVW", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f2ed4812d9"},
+		{"VPACKSSWB", "VPACKSSWB", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f16d4863d9"},
+		{"VPACKUSDW", "VPACKUSDW", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f26d482bd9"},
+		// Absolute values and replicating moves.
+		{"VPABSD", "VPABSD", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f27d481ed1"},
+		{"VPABSQ mem", "VPABSQ", []Operand{mem64(AX), vreg(t, "Z2")}, "62f2fd481f10"},
+		{"VMOVSLDUP", "VMOVSLDUP", []Operand{vreg(t, "X1"), vreg(t, "X2")}, "c5fa12d1"},
+		{"VMOVSHDUP", "VMOVSHDUP", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f17e4816d1"},
+		// Rotates and word/qword shifts.
+		{"VPROLD", "VPROLD", []Operand{Imm(5), vreg(t, "Z1"), vreg(t, "Z2")}, "62f16d4872c905"},
+		{"VPRORQ", "VPRORQ", []Operand{Imm(63), vreg(t, "Z1"), vreg(t, "Z2")}, "62f1ed4872c13f"},
+		{"VPSLLW", "VPSLLW", []Operand{Imm(9), vreg(t, "X1"), vreg(t, "X2")}, "c5e971f109"},
+		{"VPSRLQ", "VPSRLQ", []Operand{Imm(3), vreg(t, "Z1"), vreg(t, "Z2")}, "62f1ed4873d103"},
+		// Opmask instructions (VEX-encoded, the width in the L/W/pp bits).
+		{"KANDW", "KANDW", []Operand{vreg(t, "K1"), vreg(t, "K2"), vreg(t, "K3")}, "c5ec41d9"},
+		{"KORD", "KORD", []Operand{vreg(t, "K4"), vreg(t, "K5"), vreg(t, "K6")}, "c4e1d545f4"},
+		{"KXNORQ", "KXNORQ", []Operand{vreg(t, "K1"), vreg(t, "K2"), vreg(t, "K3")}, "c4e1ec46d9"},
+		{"KNOTB", "KNOTB", []Operand{vreg(t, "K4"), vreg(t, "K5")}, "c5f944ec"},
+		{"KUNPCKBW", "KUNPCKBW", []Operand{vreg(t, "K1"), vreg(t, "K2"), vreg(t, "K3")}, "c5ed4bd9"},
+		{"KSHIFTLW", "KSHIFTLW", []Operand{Imm(2), vreg(t, "K1"), vreg(t, "K2")}, "c4e3f932d102"},
+		{"KADDQ", "KADDQ", []Operand{vreg(t, "K1"), vreg(t, "K2"), vreg(t, "K3")}, "c4e1ec4ad9"},
+		{"KORTESTD", "KORTESTD", []Operand{vreg(t, "K1"), vreg(t, "K2")}, "c4e1f998d1"},
+		{"KMOVQ k,k", "KMOVQ", []Operand{vreg(t, "K1"), vreg(t, "K2")}, "c4e1f890d1"},
+		{"KMOVQ gpr,k", "KMOVQ", []Operand{BX, vreg(t, "K1")}, "c4e1fb92cb"},
+		// Lane extract / insert.
+		{"VEXTRACTF32X4", "VEXTRACTF32X4", []Operand{Imm(1), vreg(t, "Y1"), vreg(t, "X2")}, "62f37d2819ca01"},
+		{"VEXTRACTI64X2", "VEXTRACTI64X2", []Operand{Imm(1), vreg(t, "Y1"), vreg(t, "X2")}, "62f3fd2839ca01"},
+		{"VINSERTF32X8", "VINSERTF32X8", []Operand{Imm(1), vreg(t, "Y1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f36d481ad901"},
+		{"VINSERTI64X4", "VINSERTI64X4", []Operand{Imm(1), vreg(t, "Y1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f3ed483ad901"},
+		// Aligned moves and the scalar single move.
+		{"VMOVAPS", "VMOVAPS", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f17c4829ca"},
+		{"VMOVDQA64 mem", "VMOVDQA64", []Operand{mem64(AX), vreg(t, "Z2")}, "62f1fd486f10"},
+		{"VMOVSS mem", "VMOVSS", []Operand{mem64(AX), vreg(t, "X2")}, "c5fa1010"},
+		// Conversions and extending/narrowing moves.
+		{"VCVTPS2DQ", "VCVTPS2DQ", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f17d485bd1"},
+		{"VCVTTPS2DQ", "VCVTTPS2DQ", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f17e485bd1"},
+		{"VPMOVZXBW", "VPMOVZXBW", []Operand{vreg(t, "X1"), vreg(t, "Y2")}, "c4e27d30d1"},
+		{"VPMOVSXBW mem", "VPMOVSXBW", []Operand{mem64(AX), vreg(t, "Z2")}, "62f27d482010"},
+		{"VPMOVWB", "VPMOVWB", []Operand{vreg(t, "Z1"), vreg(t, "Y2")}, "62f27e4830ca"},
+		{"VPMOVQB", "VPMOVQB", []Operand{vreg(t, "Z1"), vreg(t, "X2")}, "62f27e4832ca"},
+	}
+	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)
+			continue
+		}
+		inst, err := x86asm.Decode(code, 64)
+		if err != nil {
+			t.Errorf("%s: Decode(%x): %v", c.name, code, err)
+			continue
+		}
+		want := c.mnem
+		if i := strings.IndexByte(want, '.'); i > 0 {
+			want = want[:i]
+		}
+		if inst.Op.String() != want {
+			t.Errorf("%s: decoded as %s", c.name, inst.Op.String())
+		}
+	}
+}
+
+// TestEvexHelperGroundTruth covers the floating-point helper and conversion
+// tail of the EVEX set — reciprocals, rsqrt, getexp/getmant, scalef,
+// rndscale, reduce, fixupimm, range, fpclass, the remaining conversions —
+// plus gather/scatter with VSIB addressing, byte for byte against the Go
+// assembler.
+func TestEvexHelperGroundTruth(t *testing.T) {
+	vsib := func(base, idx string, scale int) Operand {
+		return Idx(vreg(t, base), vreg(t, idx), scale, 0, 0)
+	}
+	cases := []struct {
+		name string
+		mnem string
+		ops  []Operand
+		want string
+	}{
+		// Reciprocals and rsqrt (packed RM, scalar NDS).
+		{"VRCP14PD", "VRCP14PD", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f2fd484cd1"},
+		{"VRCP14PS", "VRCP14PS", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f27d484cd1"},
+		{"VRCP14SD", "VRCP14SD", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "62f2ed084dd9"},
+		{"VRCP14SS", "VRCP14SS", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "62f26d084dd9"},
+		{"VRSQRT14PD", "VRSQRT14PD", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f2fd484ed1"},
+		{"VRSQRT14PS", "VRSQRT14PS", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f27d484ed1"},
+		{"VRSQRT14SD", "VRSQRT14SD", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "62f2ed084fd9"},
+		{"VRSQRT14SS", "VRSQRT14SS", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "62f26d084fd9"},
+		// Getexp (packed RM, scalar NDS).
+		{"VGETEXPPD", "VGETEXPPD", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f2fd4842d1"},
+		{"VGETEXPPS", "VGETEXPPS", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f27d4842d1"},
+		{"VGETEXPSD", "VGETEXPSD", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "62f2ed0843d9"},
+		{"VGETEXPSS", "VGETEXPSS", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "62f26d0843d9"},
+		// Scalef (NDS).
+		{"VSCALEFPD", "VSCALEFPD", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f2ed482cd9"},
+		{"VSCALEFPS", "VSCALEFPS", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f26d482cd9"},
+		{"VSCALEFSD", "VSCALEFSD", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "62f2ed082dd9"},
+		{"VSCALEFSS", "VSCALEFSS", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "62f26d082dd9"},
+		// Rndscale / getmant / reduce (packed $imm,src,dst; scalar NDS+imm).
+		{"VRNDSCALEPD", "VRNDSCALEPD", []Operand{Imm(4), vreg(t, "Z1"), vreg(t, "Z2")}, "62f3fd4809d104"},
+		{"VRNDSCALEPS", "VRNDSCALEPS", []Operand{Imm(4), vreg(t, "Z1"), vreg(t, "Z2")}, "62f37d4808d104"},
+		{"VRNDSCALESD", "VRNDSCALESD", []Operand{Imm(4), vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "62f3ed080bd904"},
+		{"VRNDSCALESS", "VRNDSCALESS", []Operand{Imm(4), vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "62f36d080ad904"},
+		{"VGETMANTPD", "VGETMANTPD", []Operand{Imm(3), vreg(t, "Z1"), vreg(t, "Z2")}, "62f3fd4826d103"},
+		{"VGETMANTPS", "VGETMANTPS", []Operand{Imm(3), vreg(t, "Z1"), vreg(t, "Z2")}, "62f37d4826d103"},
+		{"VGETMANTSD", "VGETMANTSD", []Operand{Imm(3), vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "62f3ed0827d903"},
+		{"VGETMANTSS", "VGETMANTSS", []Operand{Imm(3), vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "62f36d0827d903"},
+		{"VREDUCEPD", "VREDUCEPD", []Operand{Imm(4), vreg(t, "Z1"), vreg(t, "Z2")}, "62f3fd4856d104"},
+		{"VREDUCEPS", "VREDUCEPS", []Operand{Imm(4), vreg(t, "Z1"), vreg(t, "Z2")}, "62f37d4856d104"},
+		{"VREDUCESD", "VREDUCESD", []Operand{Imm(4), vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "62f3ed0857d904"},
+		{"VREDUCESS", "VREDUCESS", []Operand{Imm(4), vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "62f36d0857d904"},
+		// Fixupimm / range (NDS + imm8).
+		{"VFIXUPIMMPD", "VFIXUPIMMPD", []Operand{Imm(2), vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f3ed4854d902"},
+		{"VFIXUPIMMPS", "VFIXUPIMMPS", []Operand{Imm(2), vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f36d4854d902"},
+		{"VFIXUPIMMSD", "VFIXUPIMMSD", []Operand{Imm(2), vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "62f3ed0855d902"},
+		{"VFIXUPIMMSS", "VFIXUPIMMSS", []Operand{Imm(2), vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "62f36d0855d902"},
+		{"VRANGEPD", "VRANGEPD", []Operand{Imm(1), vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f3ed4850d901"},
+		{"VRANGEPS", "VRANGEPS", []Operand{Imm(1), vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f36d4850d901"},
+		{"VRANGESD", "VRANGESD", []Operand{Imm(1), vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "62f3ed0851d901"},
+		{"VRANGESS", "VRANGESS", []Operand{Imm(1), vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "62f36d0851d901"},
+		// FP class test ($imm, src, kdst; packed forms carry the length in
+		// the X/Y/Z mnemonic suffix the decoder drops).
+		{"VFPCLASSPDZ", "VFPCLASSPDZ", []Operand{Imm(4), vreg(t, "Z1"), vreg(t, "K2")}, "62f3fd4866d104"},
+		{"VFPCLASSPSY", "VFPCLASSPSY", []Operand{Imm(4), vreg(t, "Y1"), vreg(t, "K2")}, "62f37d2866d104"},
+		{"VFPCLASSSD", "VFPCLASSSD", []Operand{Imm(4), vreg(t, "X1"), vreg(t, "K2")}, "62f3fd0867d104"},
+		{"VFPCLASSSS", "VFPCLASSSS", []Operand{Imm(4), vreg(t, "X1"), vreg(t, "K2")}, "62f37d0867d104"},
+		// Gather: VEX spelling (mask register, VSIB, destination) and EVEX
+		// spelling (VSIB, K mask, destination; L'L follows the VSIB index).
+		{"VGATHERDPS vex", "VGATHERDPS", []Operand{vreg(t, "X2"), vsib("SI", "X1", 4), vreg(t, "X3")}, "c4e269921c8e"},
+		{"VPGATHERDD vex", "VPGATHERDD", []Operand{vreg(t, "Y2"), vsib("SI", "Y1", 4), vreg(t, "Y3")}, "c4e26d901c8e"},
+		{"VGATHERDPS evex", "VGATHERDPS", []Operand{vsib("SI", "X1", 4), vreg(t, "K2"), vreg(t, "X3")}, "62f27d0a921c8e"},
+		{"VPGATHERQD evex", "VPGATHERQD", []Operand{vsib("SI", "Z1", 8), vreg(t, "K2"), vreg(t, "Y3")}, "62f27d4a911cce"},
+		// Scatter (EVEX only: source, K mask, VSIB).
+		{"VSCATTERDPS", "VSCATTERDPS", []Operand{vreg(t, "X3"), vreg(t, "K1"), vsib("SI", "X1", 4)}, "62f27d09a21c8e"},
+		{"VSCATTERQPD", "VSCATTERQPD", []Operand{vreg(t, "Z3"), vreg(t, "K1"), vsib("SI", "Z1", 8)}, "62f2fd49a31cce"},
+		// The remaining conversions.
+		{"VCVTDQ2PS", "VCVTDQ2PS", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f17c485bd1"},
+		{"VCVTQQ2PS", "VCVTQQ2PS", []Operand{vreg(t, "Z1"), vreg(t, "Y2")}, "62f1fc485bd1"},
+		{"VCVTPD2QQ", "VCVTPD2QQ", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f1fd487bd1"},
+		{"VCVTPS2QQ", "VCVTPS2QQ", []Operand{vreg(t, "Y1"), vreg(t, "Z2")}, "62f17d487bd1"},
+		{"VCVTUDQ2PD", "VCVTUDQ2PD", []Operand{vreg(t, "X1"), vreg(t, "Y2")}, "62f17e287ad1"},
+		{"VCVTPH2PS", "VCVTPH2PS", []Operand{vreg(t, "Y1"), vreg(t, "Z2")}, "62f27d4813d1"},
+		{"VCVTPS2PH", "VCVTPS2PH", []Operand{Imm(4), vreg(t, "Y1"), vreg(t, "X2")}, "c4e37d1dca04"},
+	}
+	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)
+			continue
+		}
+		inst, err := x86asm.Decode(code, 64)
+		if err != nil {
+			t.Errorf("%s: Decode(%x): %v", c.name, code, err)
+			continue
+		}
+		want := c.mnem
+		got := inst.Op.String()
+		if got != want && !(len(want) > len(got) && want[:len(got)] == got) {
+			t.Errorf("%s: decoded as %s", c.name, got)
+		}
+	}
+}
+
+// TestEvexGprGroundTruth covers the scalar conversions between vector and
+// general-purpose registers — the signed and truncated VCVT{,T}S{D,S}2SI
+// forms (VEX and EVEX), the unsigned EVEX-only forms, and the GPR-to-vector
+// VCVTSI2*/VCVTUSI2* forms with the preserved vector source in vvvv — byte
+// for byte against the Go assembler, including memory sources and extended
+// GPRs.
+func TestEvexGprGroundTruth(t *testing.T) {
+	mem := func(b Reg) Operand { return Ptr(b, 0, 8) }
+	cases := []struct {
+		name string
+		mnem string
+		ops  []Operand
+		want string
+	}{
+		{"VCVTSD2SI", "VCVTSD2SI", []Operand{vreg(t, "X1"), AX}, "c5fb2dc1"},
+		{"VCVTSD2SIQ", "VCVTSD2SIQ", []Operand{vreg(t, "X1"), AX}, "c4e1fb2dc1"},
+		{"VCVTSS2SI", "VCVTSS2SI", []Operand{vreg(t, "X1"), AX}, "c5fa2dc1"},
+		{"VCVTSS2SIQ", "VCVTSS2SIQ", []Operand{vreg(t, "X1"), AX}, "c4e1fa2dc1"},
+		{"VCVTTSD2SI", "VCVTTSD2SI", []Operand{vreg(t, "X1"), AX}, "c5fb2cc1"},
+		{"VCVTTSD2SIQ", "VCVTTSD2SIQ", []Operand{vreg(t, "X1"), AX}, "c4e1fb2cc1"},
+		{"VCVTTSS2SI", "VCVTTSS2SI", []Operand{vreg(t, "X1"), AX}, "c5fa2cc1"},
+		{"VCVTTSS2SIQ", "VCVTTSS2SIQ", []Operand{vreg(t, "X1"), AX}, "c4e1fa2cc1"},
+		{"VCVTSD2USIL", "VCVTSD2USIL", []Operand{vreg(t, "X1"), AX}, "62f17f0879c1"},
+		{"VCVTSD2USIQ", "VCVTSD2USIQ", []Operand{vreg(t, "X1"), AX}, "62f1ff0879c1"},
+		{"VCVTSS2USIL", "VCVTSS2USIL", []Operand{vreg(t, "X1"), AX}, "62f17e0879c1"},
+		{"VCVTSS2USIQ", "VCVTSS2USIQ", []Operand{vreg(t, "X1"), AX}, "62f1fe0879c1"},
+		{"VCVTTSD2USIL", "VCVTTSD2USIL", []Operand{vreg(t, "X1"), AX}, "62f17f0878c1"},
+		{"VCVTTSD2USIQ", "VCVTTSD2USIQ", []Operand{vreg(t, "X1"), AX}, "62f1ff0878c1"},
+		{"VCVTTSS2USIL", "VCVTTSS2USIL", []Operand{vreg(t, "X1"), AX}, "62f17e0878c1"},
+		{"VCVTTSS2USIQ", "VCVTTSS2USIQ", []Operand{vreg(t, "X1"), AX}, "62f1fe0878c1"},
+		{"VCVTSI2SDL", "VCVTSI2SDL", []Operand{AX, vreg(t, "X1"), vreg(t, "X2")}, "c5f32ad0"},
+		{"VCVTSI2SDQ", "VCVTSI2SDQ", []Operand{AX, vreg(t, "X1"), vreg(t, "X2")}, "c4e1f32ad0"},
+		{"VCVTSI2SSL", "VCVTSI2SSL", []Operand{AX, vreg(t, "X1"), vreg(t, "X2")}, "c5f22ad0"},
+		{"VCVTSI2SSQ", "VCVTSI2SSQ", []Operand{AX, vreg(t, "X1"), vreg(t, "X2")}, "c4e1f22ad0"},
+		{"VCVTUSI2SDL", "VCVTUSI2SDL", []Operand{AX, vreg(t, "X1"), vreg(t, "X2")}, "62f177087bd0"},
+		{"VCVTUSI2SDQ", "VCVTUSI2SDQ", []Operand{AX, vreg(t, "X1"), vreg(t, "X2")}, "62f1f7087bd0"},
+		{"VCVTUSI2SSL", "VCVTUSI2SSL", []Operand{AX, vreg(t, "X1"), vreg(t, "X2")}, "62f176087bd0"},
+		{"VCVTUSI2SSQ", "VCVTUSI2SSQ", []Operand{AX, vreg(t, "X1"), vreg(t, "X2")}, "62f1f6087bd0"},
+		{"VCVTSD2SI mem", "VCVTSD2SI", []Operand{mem(AX), BX}, "c5fb2d18"},
+		{"VCVTSI2SDQ mem", "VCVTSI2SDQ", []Operand{mem(BX), vreg(t, "X1"), vreg(t, "X2")}, "c4e1f32a13"},
+		{"VCVTSD2SIQ hi gpr", "VCVTSD2SIQ", []Operand{vreg(t, "X1"), vreg(t, "R9")}, "c461fb2dc9"},
+		{"VCVTSI2SDQ hi gpr", "VCVTSI2SDQ", []Operand{vreg(t, "R10"), vreg(t, "X1"), vreg(t, "X2")}, "c4c1f32ad2"},
+	}
+	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)
+			continue
+		}
+		inst, err := x86asm.Decode(code, 64)
+		if err != nil {
+			t.Errorf("%s: Decode(%x): %v", c.name, code, err)
+			continue
+		}
+		// The decoder does not distinguish the Plan 9 SIQ spelling (the
+		// 64-bit GPR destination) from the base name; the W bit carries it.
+		want := c.mnem
+		got := inst.Op.String()
+		if got != want && !(len(want) > len(got) && want[:len(got)] == got) {
+			t.Errorf("%s: decoded as %s", c.name, got)
+		}
+	}
+}
+
+// TestEvexConversionGroundTruth covers the unsigned and truncating VCVT*
+// conversions, the remaining sign/zero-extending moves, the signed/unsigned
+// narrowing stores and the mask/vector conversions, byte for byte against
+// the Go assembler.
+func TestEvexConversionGroundTruth(t *testing.T) {
+	cases := []struct {
+		name string
+		mnem string
+		ops  []Operand
+		want string
+	}{
+		// Unsigned and truncating conversions.
+		{"VCVTPD2PS", "VCVTPD2PS", []Operand{vreg(t, "Z1"), vreg(t, "Y2")}, "62f1fd485ad1"},
+		{"VCVTPD2PSX", "VCVTPD2PSX", []Operand{vreg(t, "X1"), vreg(t, "X2")}, "c5f95ad1"},
+		{"VCVTPD2PSY", "VCVTPD2PSY", []Operand{vreg(t, "Y1"), vreg(t, "X2")}, "c5fd5ad1"},
+		{"VCVTPD2UDQ", "VCVTPD2UDQ", []Operand{vreg(t, "Z1"), vreg(t, "Y2")}, "62f1fc4879d1"},
+		{"VCVTPD2UDQX", "VCVTPD2UDQX", []Operand{vreg(t, "X1"), vreg(t, "X2")}, "62f1fc0879d1"},
+		{"VCVTTPD2UDQ", "VCVTTPD2UDQ", []Operand{vreg(t, "Z1"), vreg(t, "Y2")}, "62f1fc4878d1"},
+		{"VCVTTPD2UDQY", "VCVTTPD2UDQY", []Operand{vreg(t, "Y1"), vreg(t, "X2")}, "62f1fc2878d1"},
+		{"VCVTTPD2UQQ", "VCVTTPD2UQQ", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f1fd4878d1"},
+		{"VCVTPS2UDQ", "VCVTPS2UDQ", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f17c4879d1"},
+		{"VCVTTPS2UDQ", "VCVTTPS2UDQ", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f17c4878d1"},
+		{"VCVTPS2UQQ", "VCVTPS2UQQ", []Operand{vreg(t, "Y1"), vreg(t, "Z2")}, "62f17d4879d1"},
+		{"VCVTTPS2UQQ", "VCVTTPS2UQQ", []Operand{vreg(t, "Y1"), vreg(t, "Z2")}, "62f17d4878d1"},
+		{"VCVTTPD2QQ", "VCVTTPD2QQ", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f1fd487ad1"},
+		{"VCVTTPS2QQ", "VCVTTPS2QQ", []Operand{vreg(t, "Y1"), vreg(t, "Z2")}, "62f17d487ad1"},
+		{"VCVTUQQ2PD", "VCVTUQQ2PD", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f1fe487ad1"},
+		{"VCVTUQQ2PS", "VCVTUQQ2PS", []Operand{vreg(t, "Z1"), vreg(t, "Y2")}, "62f1ff487ad1"},
+		{"VCVTUQQ2PSX", "VCVTUQQ2PSX", []Operand{vreg(t, "X1"), vreg(t, "X2")}, "62f1ff087ad1"},
+		{"VCVTQQ2PSX", "VCVTQQ2PSX", []Operand{vreg(t, "X1"), vreg(t, "X2")}, "62f1fc085bd1"},
+		{"VCVTQQ2PSY", "VCVTQQ2PSY", []Operand{vreg(t, "Y1"), vreg(t, "X2")}, "62f1fc285bd1"},
+		// The remaining sign/zero-extending moves.
+		{"VPMOVSXBD", "VPMOVSXBD", []Operand{vreg(t, "X1"), vreg(t, "Y2")}, "c4e27d21d1"},
+		{"VPMOVSXBQ evex", "VPMOVSXBQ", []Operand{vreg(t, "X1"), vreg(t, "Z2")}, "62f27d4822d1"},
+		{"VPMOVSXWQ", "VPMOVSXWQ", []Operand{vreg(t, "X1"), vreg(t, "Y2")}, "c4e27d24d1"},
+		{"VPMOVSXWD", "VPMOVSXWD", []Operand{vreg(t, "X1"), vreg(t, "Y2")}, "c4e27d23d1"},
+		{"VPMOVZXBD", "VPMOVZXBD", []Operand{vreg(t, "X1"), vreg(t, "Y2")}, "c4e27d31d1"},
+		{"VPMOVZXBQ evex", "VPMOVZXBQ", []Operand{vreg(t, "X1"), vreg(t, "Z2")}, "62f27d4832d1"},
+		{"VPMOVZXWD", "VPMOVZXWD", []Operand{vreg(t, "X1"), vreg(t, "Y2")}, "c4e27d33d1"},
+		{"VPMOVZXWQ", "VPMOVZXWQ", []Operand{vreg(t, "X1"), vreg(t, "Y2")}, "c4e27d34d1"},
+		// Signed narrowing stores.
+		{"VPMOVSDB", "VPMOVSDB", []Operand{vreg(t, "Z1"), vreg(t, "X2")}, "62f27e4821ca"},
+		{"VPMOVSDW", "VPMOVSDW", []Operand{vreg(t, "Z1"), vreg(t, "Y2")}, "62f27e4823ca"},
+		{"VPMOVSQB", "VPMOVSQB", []Operand{vreg(t, "Z1"), vreg(t, "X2")}, "62f27e4822ca"},
+		{"VPMOVSQD", "VPMOVSQD", []Operand{vreg(t, "Z1"), vreg(t, "Y2")}, "62f27e4825ca"},
+		{"VPMOVSQW", "VPMOVSQW", []Operand{vreg(t, "Z1"), vreg(t, "X2")}, "62f27e4824ca"},
+		{"VPMOVSWB", "VPMOVSWB", []Operand{vreg(t, "Z1"), vreg(t, "Y2")}, "62f27e4820ca"},
+		// Unsigned narrowing stores.
+		{"VPMOVUSDB", "VPMOVUSDB", []Operand{vreg(t, "Z1"), vreg(t, "X2")}, "62f27e4811ca"},
+		{"VPMOVUSDW", "VPMOVUSDW", []Operand{vreg(t, "Z1"), vreg(t, "Y2")}, "62f27e4813ca"},
+		{"VPMOVUSQB", "VPMOVUSQB", []Operand{vreg(t, "Z1"), vreg(t, "X2")}, "62f27e4812ca"},
+		{"VPMOVUSQD", "VPMOVUSQD", []Operand{vreg(t, "Z1"), vreg(t, "Y2")}, "62f27e4815ca"},
+		{"VPMOVUSQW", "VPMOVUSQW", []Operand{vreg(t, "Z1"), vreg(t, "X2")}, "62f27e4814ca"},
+		{"VPMOVUSWB", "VPMOVUSWB", []Operand{vreg(t, "Z1"), vreg(t, "Y2")}, "62f27e4810ca"},
+		{"VPMOVDB", "VPMOVDB", []Operand{vreg(t, "Z1"), vreg(t, "X2")}, "62f27e4831ca"},
+		{"VPMOVQW", "VPMOVQW", []Operand{vreg(t, "Z1"), vreg(t, "X2")}, "62f27e4834ca"},
+		// Mask/vector conversions (the K register is an operand, not a
+		// mask).
+		{"VPMOVM2B", "VPMOVM2B", []Operand{vreg(t, "K1"), vreg(t, "X2")}, "62f27e0828d1"},
+		{"VPMOVM2W", "VPMOVM2W", []Operand{vreg(t, "K1"), vreg(t, "X2")}, "62f2fe0828d1"},
+		{"VPMOVM2D", "VPMOVM2D", []Operand{vreg(t, "K1"), vreg(t, "X2")}, "62f27e0838d1"},
+		{"VPMOVM2Q", "VPMOVM2Q", []Operand{vreg(t, "K1"), vreg(t, "Z2")}, "62f2fe4838d1"},
+		{"VPMOVB2M", "VPMOVB2M", []Operand{vreg(t, "X1"), vreg(t, "K2")}, "62f27e0829d1"},
+		{"VPMOVW2M", "VPMOVW2M", []Operand{vreg(t, "X1"), vreg(t, "K2")}, "62f2fe0829d1"},
+		{"VPMOVD2M", "VPMOVD2M", []Operand{vreg(t, "Z1"), vreg(t, "K2")}, "62f27e4839d1"},
+		{"VPMOVQ2M", "VPMOVQ2M", []Operand{vreg(t, "Z1"), vreg(t, "K2")}, "62f2fe4839d1"},
+	}
+	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)
+			continue
+		}
+		inst, err := x86asm.Decode(code, 64)
+		if err != nil {
+			t.Errorf("%s: Decode(%x): %v", c.name, code, err)
+			continue
+		}
+		want := c.mnem
+		got := inst.Op.String()
+		if got != want && !(len(want) > len(got) && want[:len(got)] == got) {
+			t.Errorf("%s: decoded as %s", c.name, got)
+		}
+	}
+}
+
 // TestEvexErrors checks the EVEX-specific error paths.
 func TestEvexErrors(t *testing.T) {
 	cases := []struct {
@@ -121,7 +640,7 @@ func TestEvexErrors(t *testing.T) {
 		{"VPMOVDW src", "VPMOVDW", []Operand{AX, vreg(t, "Y0")}},
 		{"align arity", "VALIGND", []Operand{Imm(1), vreg(t, "Z0"), vreg(t, "Z1")}},
 		// VEX-only mnemonics reject registers only EVEX can encode.
-		{"VPSHUFD X16", "VPSHUFD", []Operand{Imm(1), vreg(t, "X16"), vreg(t, "X17")}},
+		{"VMOVMSKPS X16", "VMOVMSKPS", []Operand{vreg(t, "X16"), AX}},
 	}
 	for _, c := range cases {
 		if _, err := Encode(c.mnem, c.ops...); err == nil {

asm/goobj.go

@@ -0,0 +1,453 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package asm
+
+import (
+	"bytes"
+	"encoding/binary"
+	"fmt"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"sync"
+)
+
+// This file emits GOOBJ — the Go toolchain's object format, which cmd/link
+// consumes directly — so gasm-assembled functions drop into a go build
+// without the Go assembler.  The layout follows cmd/internal/goobj: a
+// toolchain preamble ("go object ...\n!\n"), the go120ld header with its
+// block offsets, a string table, symbol definitions, the relocation /
+// aux / data index arrays, and the three blocks themselves.
+//
+// The object carries what the linker requires of an assembly object: the
+// functions (non-package symbols, as cmd/asm emits them), the GLOBL data,
+// one FuncInfo per function, and the pc-value tables (pcsp, pcfile,
+// pcline, pcinline).  DWARF and the implicit funcdata symbols are omitted;
+// the linker fills their defaults.
+
+// GOOBJ block indices (cmd/internal/goobj).
+const (
+	blkAutolib = iota
+	blkPkgIdx
+	blkFile
+	blkSymdef
+	blkHashed64def
+	blkHasheddef
+	blkNonpkgdef
+	blkNonpkgref
+	blkRefFlags
+	blkHash64
+	blkHash
+	blkRelocIdx
+	blkAuxIdx
+	blkDataIdx
+	blkReloc
+	blkAux
+	blkData
+	blkRefName
+	blkEnd
+)
+
+// Symbol kinds used by assembly objects (cmd/internal/objabi).
+const (
+	kindSTEXT   = 1
+	kindSRODATA = 3
+	kindSDATA   = 7
+)
+
+// Symbol flags (cmd/internal/goobj).
+const (
+	symFlagDupok   = 0x01
+	symFlagNoSplit = 0x10
+	symFlag2Link   = 0x10 // asm objects flag every named symbol as linkname
+	symABIStatic   = 0xffff
+)
+
+// Aux entry types (cmd/internal/goobj).
+const (
+	auxFuncInfo = 1
+	auxPcsp     = 7
+	auxPcfile   = 8
+	auxPcline   = 9
+	auxPcinline = 10
+)
+
+// FuncInfo flags (internal/abi).
+const (
+	funcFlagSPWrite = 2
+	funcFlagAsm     = 4
+)
+
+// Relocation types (cmd/internal/objabi).
+const relocPCRel = 14
+
+// Special package indices for symbol references.
+const (
+	pkgIdxNone = 0x7fffffff
+	pkgIdxSelf = 0x7ffffffb
+)
+
+const goobjMagic = "\x00go120ld"
+
+// goSym is one symbol definition under construction.
+type goSym struct {
+	name  string
+	abi   uint16
+	typ   uint8
+	flag  uint8
+	flag2 uint8
+	size  uint32
+	align uint32
+}
+
+func (s goSym) append(b []byte, strOff map[string]uint32) []byte {
+	b = binary.LittleEndian.AppendUint32(b, uint32(len(s.name)))
+	b = binary.LittleEndian.AppendUint32(b, strOff[s.name])
+	b = binary.LittleEndian.AppendUint16(b, s.abi)
+	b = append(b, s.typ, s.flag, s.flag2)
+	b = binary.LittleEndian.AppendUint32(b, s.size)
+	return binary.LittleEndian.AppendUint32(b, s.align)
+}
+
+// GOObject returns the image as a GOOBJ object file for the given package
+// path (the linker qualifies the exported symbols with it, the way cmd/asm
+// does with its -p flag).  srcPath names the source file recorded in the
+// object's file table and line tables.  The toolchain's object preamble is
+// captured from the installed go tool asm, so the output links with the
+// toolchain it was produced on — exactly like a real assembly object.
+func (img *Image) GOObject(pkgPath, srcPath string) ([]byte, error) {
+	if pkgPath == "" {
+		return nil, fmt.Errorf("GOOBJ emission requires a package path (-p)")
+	}
+	pre, err := toolchainObjectPreamble()
+	if err != nil {
+		return nil, err
+	}
+
+	// The symbol tables.  Package definitions: the GLOBL symbols, then one
+	// anonymous FuncInfo symbol per function.  Non-package definitions: the
+	// pc-value tables and the functions themselves, as cmd/asm lays them
+	// out.  defIdx maps a GLOBL's bare name to its definition index for the
+	// relocations; fnNpIdx maps a function to its non-package index.
+	var defs []goSym
+	var defData [][]byte
+	defIdx := map[string]int{}
+	for _, d := range img.DataSyms {
+		name := d.Name
+		if !d.Static {
+			name = pkgPath + "." + name
+		}
+		typ := uint8(kindSDATA)
+		if d.Rodata {
+			typ = kindSRODATA
+		}
+		flag := uint8(0)
+		if d.Dupok {
+			flag = symFlagDupok
+		}
+		abi := uint16(0)
+		if d.Static {
+			abi = symABIStatic
+		}
+		defIdx[d.Name] = len(defs)
+		defs = append(defs, goSym{name: name, abi: abi, typ: typ, flag: flag, flag2: symFlag2Link, size: uint32(d.Size)})
+		defData = append(defData, img.Data[d.Offset:d.Offset+d.Size])
+	}
+	fnFiIdx := make([]int, len(img.Funcs))
+	for i := range img.Funcs {
+		data := marshalFuncInfo(img.Funcs[i])
+		fnFiIdx[i] = len(defs)
+		defs = append(defs, goSym{typ: kindSDATA, size: uint32(len(data))})
+		defData = append(defData, data)
+	}
+
+	type npSym struct {
+		sym  goSym
+		data []byte
+	}
+	var nps []npSym
+	type pcRefs struct{ sp, file, line, inl int }
+	pcIdx := make([]pcRefs, len(img.Funcs))
+	fnNpIdx := make([]int, len(img.Funcs))
+	for i, fn := range img.Funcs {
+		tables := []struct {
+			data []byte
+			dst  *int
+		}{
+			{pcspTable(fn), &pcIdx[i].sp},
+			{pcValueFlat(0, fn.Size), &pcIdx[i].file},
+			{pcValueFlat(int32(fn.Line), fn.Size), &pcIdx[i].line},
+			{pcValueFlat(-1, fn.Size), &pcIdx[i].inl},
+		}
+		for _, t := range tables {
+			*t.dst = len(nps)
+			nps = append(nps, npSym{
+				sym:  goSym{typ: kindSRODATA, size: uint32(len(t.data)), align: 1},
+				data: t.data,
+			})
+		}
+		name := fn.Name
+		abi := uint16(0)
+		if fn.Static {
+			abi = symABIStatic
+		} else {
+			name = pkgPath + "." + name
+		}
+		flag := uint8(0)
+		if fn.NoSplit {
+			flag |= symFlagNoSplit
+		}
+		fnNpIdx[i] = len(nps)
+		code := append([]byte(nil), img.Code[fn.Offset:fn.Offset+fn.Size]...)
+		for _, r := range fn.Relocs {
+			// The linker writes the resolved displacement into the field;
+			// leave it zero, as cmd/asm's object does.
+			if r.Off >= 0 && r.Off+4 <= len(code) {
+				code[r.Off], code[r.Off+1], code[r.Off+2], code[r.Off+3] = 0, 0, 0, 0
+			}
+		}
+		nps = append(nps, npSym{
+			sym:  goSym{name: name, abi: abi, typ: kindSTEXT, flag: flag, flag2: symFlag2Link, size: uint32(fn.Size)},
+			data: code,
+		})
+	}
+
+	// Relocations, per defined symbol in definition order (package defs,
+	// then non-package defs).  Only file-local GLOBL references resolve;
+	// external symbols need the import machinery of a later increment.
+	nsyms := len(defs) + len(nps)
+	symRelocs := make([][]byte, nsyms) // flat 23-byte records
+	for i, fn := range img.Funcs {
+		si := len(defs) + fnNpIdx[i]
+		for _, r := range fn.Relocs {
+			if r.External {
+				return nil, fmt.Errorf("GOOBJ emission: external symbol %q is not supported yet", r.Name)
+			}
+			di, ok := defIdx[r.Name]
+			if !ok {
+				return nil, fmt.Errorf("GOOBJ emission: reference to unknown symbol %q", r.Name)
+			}
+			var rec [23]byte
+			binary.LittleEndian.PutUint32(rec[0:], uint32(int32(r.Off)))
+			rec[4] = 4 // field width
+			binary.LittleEndian.PutUint16(rec[5:], relocPCRel)
+			binary.LittleEndian.PutUint64(rec[7:], uint64(r.Addend))
+			binary.LittleEndian.PutUint32(rec[15:], pkgIdxSelf)
+			binary.LittleEndian.PutUint32(rec[19:], uint32(di))
+			symRelocs[si] = append(symRelocs[si], rec[:]...)
+		}
+	}
+
+	// Aux entries per function: FuncInfo, then the four pc tables.
+	// References into the non-package table use pkgIdxNone.
+	symAux := make([][]byte, nsyms)
+	for i := range img.Funcs {
+		si := len(defs) + fnNpIdx[i]
+		aux := func(typ uint8, pkg, idx uint32) {
+			var rec [9]byte
+			rec[0] = typ
+			binary.LittleEndian.PutUint32(rec[1:], pkg)
+			binary.LittleEndian.PutUint32(rec[5:], idx)
+			symAux[si] = append(symAux[si], rec[:]...)
+		}
+		aux(auxFuncInfo, pkgIdxSelf, uint32(fnFiIdx[i]))
+		aux(auxPcsp, pkgIdxNone, uint32(len(defs)+pcIdx[i].sp))
+		aux(auxPcfile, pkgIdxNone, uint32(len(defs)+pcIdx[i].file))
+		aux(auxPcline, pkgIdxNone, uint32(len(defs)+pcIdx[i].line))
+		aux(auxPcinline, pkgIdxNone, uint32(len(defs)+pcIdx[i].inl))
+	}
+
+	// The string table.  Absolute offsets: it starts right after the
+	// 96-byte header (magic, fingerprint, flags, the 19 block offsets).
+	const headerSize = 8 + 8 + 4 + 4*(blkEnd+1)
+	strTab := []byte{}
+	strOff := map[string]uint32{}
+	addStr := func(s string) {
+		if _, ok := strOff[s]; ok {
+			return
+		}
+		strOff[s] = uint32(headerSize + len(strTab))
+		strTab = append(strTab, s...)
+	}
+	addStr("")
+	addStr(srcPath)
+	for _, s := range defs {
+		addStr(s.name)
+	}
+	for _, s := range nps {
+		addStr(s.sym.name)
+	}
+	stringRef := func(b []byte, s string) []byte {
+		b = binary.LittleEndian.AppendUint32(b, uint32(len(s)))
+		return binary.LittleEndian.AppendUint32(b, strOff[s])
+	}
+
+	// Serialise the block bodies.
+	var symdefBlk, npdefBlk []byte
+	for _, s := range defs {
+		symdefBlk = s.append(symdefBlk, strOff)
+	}
+	for _, s := range nps {
+		npdefBlk = s.sym.append(npdefBlk, strOff)
+	}
+	pkgIdxBlk := stringRef(nil, "") // index 0: the dummy invalid package
+	fileBlk := stringRef(nil, srcPath)
+
+	var relocBlk, auxBlk, dataBlk []byte
+	relocIdxBlk := make([]byte, 0, 4*(nsyms+1))
+	auxIdxBlk := make([]byte, 0, 4*(nsyms+1))
+	dataIdxBlk := make([]byte, 0, 4*(nsyms+1))
+	var nr, na, nd uint32
+	for si := 0; si < nsyms; si++ {
+		relocIdxBlk = binary.LittleEndian.AppendUint32(relocIdxBlk, nr)
+		auxIdxBlk = binary.LittleEndian.AppendUint32(auxIdxBlk, na)
+		dataIdxBlk = binary.LittleEndian.AppendUint32(dataIdxBlk, nd)
+		relocBlk = append(relocBlk, symRelocs[si]...)
+		auxBlk = append(auxBlk, symAux[si]...)
+		var d []byte
+		if si < len(defData) {
+			d = defData[si]
+		} else {
+			d = nps[si-len(defData)].data
+		}
+		dataBlk = append(dataBlk, d...)
+		nr += uint32(len(symRelocs[si])) / 23
+		na += uint32(len(symAux[si])) / 9
+		nd += uint32(len(d))
+	}
+	relocIdxBlk = binary.LittleEndian.AppendUint32(relocIdxBlk, nr)
+	auxIdxBlk = binary.LittleEndian.AppendUint32(auxIdxBlk, na)
+	dataIdxBlk = binary.LittleEndian.AppendUint32(dataIdxBlk, nd)
+
+	blocks := [blkEnd][]byte{
+		blkPkgIdx:    pkgIdxBlk,
+		blkFile:      fileBlk,
+		blkSymdef:    symdefBlk,
+		blkNonpkgdef: npdefBlk,
+		blkRelocIdx:  relocIdxBlk,
+		blkAuxIdx:    auxIdxBlk,
+		blkDataIdx:   dataIdxBlk,
+		blkReloc:     relocBlk,
+		blkAux:       auxBlk,
+		blkData:      dataBlk,
+	}
+
+	// Assemble the payload: header (offsets filled once known), string
+	// table, blocks in order.
+	payload := make([]byte, headerSize)
+	copy(payload, goobjMagic)
+	// The fingerprint stays zero, as cmd/asm leaves it.
+	binary.LittleEndian.PutUint32(payload[16:], 4) // ObjFlagFromAssembly
+	off := uint32(headerSize + len(strTab))
+	for i := 0; i < blkEnd; i++ {
+		binary.LittleEndian.PutUint32(payload[20+4*i:], off)
+		off += uint32(len(blocks[i]))
+	}
+	binary.LittleEndian.PutUint32(payload[20+4*blkEnd:], off)
+	payload = append(payload, strTab...)
+	for _, blk := range blocks {
+		payload = append(payload, blk...)
+	}
+
+	out := make([]byte, 0, len(pre)+len(payload))
+	out = append(out, pre...)
+	return append(out, payload...), nil
+}
+
+// marshalFuncInfo serialises a function's goobj.FuncInfo: sizes, flags,
+// start line, the one-element file table and an empty inline tree.
+func marshalFuncInfo(fn FuncLayout) []byte {
+	flag := uint8(funcFlagAsm)
+	if fn.SPWrite {
+		flag |= funcFlagSPWrite
+	}
+	b := make([]byte, 0, 28)
+	b = binary.LittleEndian.AppendUint32(b, uint32(fn.Args))
+	b = binary.LittleEndian.AppendUint32(b, uint32(fn.Frame))
+	b = append(b, 0, flag, 0, 0) // FuncID normal, flags, padding
+	b = binary.LittleEndian.AppendUint32(b, uint32(int32(fn.Line)))
+	b = binary.LittleEndian.AppendUint32(b, 1) // one file
+	b = binary.LittleEndian.AppendUint32(b, 0) // file index 0
+	b = binary.LittleEndian.AppendUint32(b, 0) // no inline tree
+	return b
+}
+
+// pcValueFlat encodes a pc-value table holding v over the whole function.
+func pcValueFlat(v int32, size int) []byte {
+	// The table is delta-encoded from an implicit value of -1: a varint
+	// value delta, an unsigned pc delta to the end, and a zero terminator.
+	out := binary.AppendVarint(nil, int64(v)+1)
+	out = binary.AppendUvarint(out, uint64(size))
+	return append(out, 0)
+}
+
+// pcspTable encodes the stack-adjustment table: the SP delta in effect at
+// every pc, from the function's prologue and epilogue boundaries.
+func pcspTable(fn FuncLayout) []byte {
+	if len(fn.Spadj) == 0 {
+		return pcValueFlat(0, fn.Size)
+	}
+	pts := make([]SpadjStep, 0, len(fn.Spadj)+1)
+	pts = append(pts, SpadjStep{PC: 0, Value: 0})
+	pts = append(pts, fn.Spadj...)
+	out := binary.AppendVarint(nil, int64(pts[0].Value)+1)
+	cur, old := pts[0].PC, pts[0].Value
+	for _, p := range pts[1:] {
+		out = binary.AppendUvarint(out, uint64(p.PC-cur))
+		out = binary.AppendVarint(out, int64(p.Value-old))
+		cur, old = p.PC, p.Value
+	}
+	out = binary.AppendUvarint(out, uint64(fn.Size-cur))
+	return append(out, 0)
+}
+
+// toolchainObjectPreamble returns the "go object ...\n!\n" header the
+// installed go tool asm writes, captured by assembling a one-instruction
+// probe.  The linker compares this string verbatim against its own, so it
+// must come from the toolchain itself, not be reconstructed.
+var (
+	preambleOnce sync.Once
+	preamble     []byte
+	preambleErr  error
+)
+
+func toolchainObjectPreamble() ([]byte, error) {
+	preambleOnce.Do(func() {
+		goBin, err := exec.LookPath("go")
+		if err != nil {
+			preambleErr = fmt.Errorf("GOOBJ emission needs the Go toolchain: %w", err)
+			return
+		}
+		dir, err := os.MkdirTemp("", "gasm-preamble")
+		if err != nil {
+			preambleErr = err
+			return
+		}
+		defer os.RemoveAll(dir)
+		src := filepath.Join(dir, "probe_amd64.s")
+		if err := os.WriteFile(src, []byte("TEXT \u00b7x(SB), $0-0\n\tRET\n"), 0o644); err != nil {
+			preambleErr = err
+			return
+		}
+		obj := filepath.Join(dir, "probe.o")
+		cmd := exec.Command(goBin, "tool", "asm", "-p", "probe", "-o", obj, src)
+		cmd.Env = append(os.Environ(), "GOARCH=amd64")
+		if out, err := cmd.CombinedOutput(); err != nil {
+			preambleErr = fmt.Errorf("probing the assembler for the object header: %v\n%s", err, out)
+			return
+		}
+		data, err := os.ReadFile(obj)
+		if err != nil {
+			preambleErr = err
+			return
+		}
+		i := bytes.Index(data, []byte("\n!\n"))
+		if i < 0 || !bytes.HasPrefix(data[i+3:], []byte(goobjMagic)) {
+			preambleErr = fmt.Errorf("unrecognised assembler object layout")
+			return
+		}
+		preamble = data[:i+3]
+	})
+	return preamble, preambleErr
+}

asm/goobj_test.go

@@ -0,0 +1,477 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package asm
+
+import (
+	"bytes"
+	"encoding/binary"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+)
+
+// goobjView is a minimal parsed view of a GOOBJ payload, enough to check
+// the emitter's output block by block.
+type goobjView struct {
+	t      *testing.T
+	b      []byte
+	offs   [blkEnd + 1]uint32
+	strOff uint32
+}
+
+func openGoobj(t *testing.T, data []byte) *goobjView {
+	t.Helper()
+	i := bytes.Index(data, []byte(goobjMagic))
+	if i < 0 {
+		t.Fatal("no GOOBJ magic in output")
+	}
+	v := &goobjView{t: t, b: data[i:], strOff: uint32(i + 96)}
+	for j := 0; j <= blkEnd; j++ {
+		v.offs[j] = binary.LittleEndian.Uint32(v.b[20+4*j:])
+	}
+	return v
+}
+
+func (v *goobjView) blk(i int) []byte { return v.b[v.offs[i]:v.offs[i+1]] }
+
+func (v *goobjView) str(off, ln uint32) string {
+	return string(v.b[off : off+ln])
+}
+
+type goobjSymView struct {
+	name  string
+	abi   uint16
+	typ   uint8
+	flag  uint8
+	flag2 uint8
+	size  uint32
+	align uint32
+}
+
+func (v *goobjView) syms(i int) []goobjSymView {
+	var out []goobjSymView
+	for x := v.blk(i); len(x) >= 21; x = x[21:] {
+		le := binary.LittleEndian
+		out = append(out, goobjSymView{
+			name:  v.str(le.Uint32(x[4:]), le.Uint32(x[0:])),
+			abi:   le.Uint16(x[8:]),
+			typ:   x[10],
+			flag:  x[11],
+			flag2: x[12],
+			size:  le.Uint32(x[13:]),
+			align: le.Uint32(x[17:]),
+		})
+	}
+	return out
+}
+
+// TestGOObjectStructure checks the emitted object's blocks against the
+// ground truth captured from go tool asm: the symbol tables, the FuncInfo
+// contents, the pc-value tables, the relocation and the aux wiring.
+func TestGOObjectStructure(t *testing.T) {
+	f, errs := parser.Parse("t_amd64.s", `
+#include "textflag.h"
+
+TEXT ·addq(SB), NOSPLIT, $0-24
+	MOVQ a+0(FP), AX
+	MOVQ b+8(FP), CX
+	ADDQ CX, AX
+	MOVQ AX, ret+16(FP)
+	RET
+
+TEXT ·loadmask(SB), NOSPLIT, $0-8
+	VMOVDQU mask<>(SB), X0
+	VPMOVMSKB X0, AX
+	MOVQ AX, ret+0(FP)
+	RET
+
+GLOBL mask<>(SB), RODATA, $16
+DATA mask<>+0(SB)/8, $0x0807060504030201
+DATA mask<>+8(SB)/8, $0x800f0e0d0c0b0a09
+`)
+	if len(errs) > 0 {
+		t.Fatalf("parse: %v", errs)
+	}
+	img, err := AssembleFile(f)
+	if err != nil {
+		t.Fatalf("AssembleFile: %v", err)
+	}
+	obj, err := img.GOObject("testpkg", "t_amd64.s")
+	if err != nil {
+		t.Fatalf("GOObject: %v", err)
+	}
+	v := openGoobj(t, obj)
+
+	if flags := binary.LittleEndian.Uint32(v.b[16:]); flags != 4 {
+		t.Errorf("flags = %#x, want ObjFlagFromAssembly (4)", flags)
+	}
+
+	// Package defs: the static GLOBL, then one anonymous FuncInfo per
+	// function.
+	defs := v.syms(blkSymdef)
+	if len(defs) != 3 {
+		t.Fatalf("symdefs = %d, want 3", len(defs))
+	}
+	if defs[0].name != "mask" || defs[0].abi != 0xffff || defs[0].typ != kindSRODATA || defs[0].size != 16 || defs[0].flag2 != symFlag2Link {
+		t.Errorf("mask symbol = %+v", defs[0])
+	}
+	if defs[1].name != "" || defs[1].typ != kindSDATA || defs[1].size != 28 {
+		t.Errorf("funcinfo symbol = %+v", defs[1])
+	}
+
+	// Non-package defs: four pc tables and the function, per function.
+	nps := v.syms(blkNonpkgdef)
+	if len(nps) != 10 {
+		t.Fatalf("nonpkgdefs = %d, want 10", len(nps))
+	}
+	fn := nps[4]
+	if fn.name != "testpkg.addq" || fn.typ != kindSTEXT || fn.flag != symFlagNoSplit || fn.size != 19 {
+		t.Errorf("addq symbol = %+v", fn)
+	}
+	for i, s := range []int{0, 1, 2, 3, 5, 6, 7, 8} {
+		if nps[s].typ != kindSRODATA || nps[s].align != 1 || nps[s].name != "" {
+			t.Errorf("pc table %d = %+v", i, nps[s])
+		}
+	}
+
+	// FuncInfo: args 24, FuncFlag Asm, one file, no inline tree.
+	le := binary.LittleEndian
+	data := v.blk(blkData)
+	fi := data[16:44]
+	if le.Uint32(fi[0:]) != 24 || le.Uint32(fi[4:]) != 0 || fi[8] != 0 || fi[9] != funcFlagAsm ||
+		le.Uint32(fi[16:]) != 1 || le.Uint32(fi[20:]) != 0 || le.Uint32(fi[24:]) != 0 {
+		t.Errorf("funcinfo bytes %x", fi)
+	}
+
+	// pcsp: a flat zero over the whole function (zero-frame NOSPLIT).
+	if got := data[72:75]; !bytes.Equal(got, []byte{0x02, 19, 0x00}) {
+		t.Errorf("pcsp = %x, want 021300", got)
+	}
+	// pcinline: a flat -1.
+	if got := data[81:84]; !bytes.Equal(got, []byte{0x00, 19, 0x00}) {
+		t.Errorf("pcinline = %x, want 001300", got)
+	}
+
+	// The one relocation: R_PCREL, four bytes wide, against the GLOBL,
+	// with the field in the function code left zero.  The loadmask code's
+	// offset comes from the data index (symbol 3 defs + 9 non-package).
+	relocs := v.blk(blkReloc)
+	if len(relocs) != 23 {
+		t.Fatalf("relocs = %d bytes, want one 23-byte entry", len(relocs))
+	}
+	off := int32(le.Uint32(relocs[0:]))
+	if off != 4 || relocs[4] != 4 || le.Uint16(relocs[5:]) != relocPCRel ||
+		le.Uint64(relocs[7:]) != 0 || le.Uint32(relocs[15:]) != pkgIdxSelf || le.Uint32(relocs[19:]) != 0 {
+		t.Errorf("reloc = %x", relocs)
+	}
+	didx := v.blk(blkDataIdx)
+	lm := le.Uint32(didx[4*(3+9):])
+	code := data[lm : lm+18]
+	if !bytes.Equal(code[4:8], []byte{0, 0, 0, 0}) {
+		t.Errorf("relocated field = %x, want zeroed", code[4:8])
+	}
+
+	// Aux wiring: FuncInfo (package symbol), then the four pc tables
+	// (non-package symbols).
+	auxs := v.blk(blkAux)
+	if len(auxs) != 2*5*9 {
+		t.Fatalf("aux = %d bytes, want 10 entries", len(auxs))
+	}
+	wantAux := []struct {
+		typ uint8
+		pkg uint32
+		idx uint32
+	}{
+		{auxFuncInfo, pkgIdxSelf, 1},
+		{auxPcsp, pkgIdxNone, uint32(len(defs) + 0)},
+		{auxPcfile, pkgIdxNone, uint32(len(defs) + 1)},
+		{auxPcline, pkgIdxNone, uint32(len(defs) + 2)},
+		{auxPcinline, pkgIdxNone, uint32(len(defs) + 3)},
+		{auxFuncInfo, pkgIdxSelf, 2},
+		{auxPcsp, pkgIdxNone, uint32(len(defs) + 5)},
+		{auxPcfile, pkgIdxNone, uint32(len(defs) + 6)},
+		{auxPcline, pkgIdxNone, uint32(len(defs) + 7)},
+		{auxPcinline, pkgIdxNone, uint32(len(defs) + 8)},
+	}
+	for i, w := range wantAux {
+		e := auxs[i*9:]
+		if e[0] != w.typ || le.Uint32(e[1:]) != w.pkg || le.Uint32(e[5:]) != w.idx {
+			t.Errorf("aux[%d] = {%d,%d,%d}, want {%d,%d,%d}", i, e[0], le.Uint32(e[1:]), le.Uint32(e[5:]), w.typ, w.pkg, w.idx)
+		}
+	}
+}
+
+// decodePCValues decodes a pc-value table into (pc, value) steps.  The
+// table ends with a final unsigned pc delta covering the rest of the
+// function, followed by a zero byte that carries no value delta.
+func decodePCValues(b []byte) (pcs, vals []int64) {
+	val, n := binary.Varint(b)
+	b = b[n:]
+	val-- // the first delta is against the implicit -1
+	var pc int64
+	pcs = append(pcs, pc)
+	vals = append(vals, val)
+	for {
+		pcd, n := binary.Uvarint(b)
+		b = b[n:]
+		if pcd == 0 { // zero pc delta terminates the table
+			break
+		}
+		pc += int64(pcd)
+		if len(b) == 1 && b[0] == 0 { // final coverage, no value change
+			break
+		}
+		vd, n := binary.Varint(b)
+		b = b[n:]
+		val += vd
+		pcs = append(pcs, pc)
+		vals = append(vals, val)
+	}
+	return pcs, vals
+}
+
+// TestGOObjectPcspFrame checks the pcsp table of a frame-pointer function:
+// the prologue raises the stack delta to 8+frame, the RET's epilogue
+// restores it to zero.
+func TestGOObjectPcspFrame(t *testing.T) {
+	f, errs := parser.Parse("frame_amd64.s", `
+#include "textflag.h"
+TEXT ·framed(SB), NOSPLIT, $8-0
+	MOVQ BP, AX
+	RET
+`)
+	if len(errs) > 0 {
+		t.Fatalf("parse: %v", errs)
+	}
+	img, err := AssembleFile(f)
+	if err != nil {
+		t.Fatalf("AssembleFile: %v", err)
+	}
+	fn := img.Funcs[0]
+	pcs, vals := decodePCValues(pcspTable(fn))
+	// Prologue: PUSHQ BP (1 byte, +8), MOVQ SP, BP (3 bytes, no change),
+	// SUBQ $8, SP (4 bytes, +16 in total); the RET's epilogue unwinds
+	// ADDQ $8, SP (+8) then POPQ BP (0).
+	wantPCs := []int64{0, 1, 8}
+	wantVals := []int64{0, 8, 16}
+	if len(pcs) < len(wantPCs) {
+		t.Fatalf("pcsp pcs = %v vals = %v", pcs, vals)
+	}
+	for i := range wantPCs {
+		if pcs[i] != wantPCs[i] || vals[i] != wantVals[i] {
+			t.Errorf("pcsp[%d] = (%d,%d), want (%d,%d) — all: %v %v", i, pcs[i], vals[i], wantPCs[i], wantVals[i], pcs, vals)
+		}
+	}
+	// The last two steps unwind the epilogue to zero.
+	n := len(pcs)
+	if vals[n-1] != 0 || vals[n-2] != 8 {
+		t.Errorf("epilogue steps = %v %v, want …8, 0", pcs, vals)
+	}
+	// The table covers the whole function.
+	if last := pcs[n-1]; last >= int64(fn.Size) {
+		t.Errorf("last pc %d beyond function size %d", last, fn.Size)
+	}
+}
+
+// TestGOObjectExternalRejected checks that a reference to a symbol no GLOBL
+// defines is reported: GOOBJ emission resolves only file-local symbols so
+// far.
+func TestGOObjectExternalRejected(t *testing.T) {
+	f, errs := parser.Parse("ext_amd64.s", `
+#include "textflag.h"
+TEXT ·useext(SB), NOSPLIT, $0-8
+	MOVQ elsewhere(SB), AX
+	MOVQ AX, ret+0(FP)
+	RET
+`)
+	if len(errs) > 0 {
+		t.Fatalf("parse: %v", errs)
+	}
+	img, err := AssembleFile(f)
+	if err != nil {
+		t.Fatalf("AssembleFile: %v", err)
+	}
+	if _, err := img.GOObject("p", "ext_amd64.s"); err == nil || !strings.Contains(err.Error(), "external") {
+		t.Errorf("error = %v, want an external-symbol error", err)
+	}
+}
+
+// TestGOObjectLinkAndRun is the end-to-end check: assemble the test
+// functions to a GOOBJ, swap it into a go build in place of the toolchain's
+// assembly object, link, and run — the output must match the baseline
+// binary the Go assembler produced.  Skipped when no Go toolchain is
+// available.
+func TestGOObjectLinkAndRun(t *testing.T) {
+	goBin, err := exec.LookPath("go")
+	if err != nil {
+		t.Skip("no Go toolchain available")
+	}
+	dir := t.TempDir()
+
+	const asmSrc = `
+#include "textflag.h"
+
+TEXT ·addq(SB), NOSPLIT, $0-24
+	MOVQ a+0(FP), AX
+	MOVQ b+8(FP), CX
+	ADDQ CX, AX
+	MOVQ AX, ret+16(FP)
+	RET
+
+TEXT ·loadmask(SB), NOSPLIT, $0-8
+	VMOVDQU mask<>(SB), X0
+	VPMOVMSKB X0, AX
+	MOVQ AX, ret+0(FP)
+	RET
+
+GLOBL mask<>(SB), RODATA, $16
+DATA mask<>+0(SB)/8, $0x0807060504030201
+DATA mask<>+8(SB)/8, $0x800f0e0d0c0b0a09
+`
+	const mainSrc = `package main
+
+func addq(a, b int64) int64
+func loadmask() int64
+
+func main() {
+	println(addq(41, 1))
+	println(loadmask())
+}
+`
+	if err := os.WriteFile(filepath.Join(dir, "main_amd64.s"), []byte(asmSrc), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(dir, "main.go"), []byte(mainSrc), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(dir, "go.mod"), []byte("module goobjtest\n\ngo 1.26\n"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	// Baseline build with the toolchain's assembler; keep the work
+	// directory and the commands the build used.
+	cmd := exec.Command(goBin, "build", "-x", "-work", "-o", "app", ".")
+	cmd.Dir = dir
+	buildLog, err := cmd.CombinedOutput()
+	if err != nil {
+		t.Fatalf("baseline build: %v\n%s", err, buildLog)
+	}
+	var work string
+	var asmObj, pkgArch, linkLine string
+	for _, line := range strings.Split(string(buildLog), "\n") {
+		switch {
+		case strings.HasPrefix(line, "WORK="):
+			work = strings.TrimPrefix(line, "WORK=")
+		case strings.Contains(line, "/asm ") && strings.Contains(line, "-o ") && strings.Contains(line, "main_amd64.s") && !strings.Contains(line, "-gensymabis"):
+			asmObj = fieldAfter(line, "-o")
+		case strings.Contains(line, "pack r") && strings.Contains(line, "_pkg_.a"):
+			pkgArch = strings.TrimSpace(strings.SplitN(line, "pack r", 2)[1])
+			pkgArch = strings.Fields(strings.SplitN(pkgArch, "#", 2)[0])[0]
+		case strings.Contains(line, "/link ") && strings.Contains(line, "-importcfg"):
+			linkLine = line
+		}
+	}
+	if work == "" || asmObj == "" || pkgArch == "" || linkLine == "" {
+		t.Fatalf("could not locate the build steps:\n%s", buildLog)
+	}
+	asmObj = strings.ReplaceAll(asmObj, "$WORK", work)
+	pkgArch = strings.ReplaceAll(pkgArch, "$WORK", work)
+
+	// The baseline's answer.
+	baseOut, err := exec.Command(filepath.Join(dir, "app")).CombinedOutput()
+	if err != nil {
+		t.Fatalf("run baseline: %v\n%s", err, baseOut)
+	}
+
+	// Assemble the same source with gasm and swap the object in.
+	pf, perrs := parser.Parse(filepath.Join(dir, "main_amd64.s"), asmSrc)
+	if len(perrs) > 0 {
+		t.Fatalf("parse: %v", perrs)
+	}
+	img, err := AssembleFile(pf)
+	if err != nil {
+		t.Fatalf("AssembleFile: %v", err)
+	}
+	obj, err := img.GOObject("main", filepath.Join(dir, "main_amd64.s"))
+	if err != nil {
+		t.Fatalf("GOObject: %v", err)
+	}
+	if err := os.WriteFile(asmObj, obj, 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	// Rebuild the package archive with our object in place of the
+	// toolchain's (go tool pack has no replace-in-place that dedupes, so
+	// extract, substitute and repack).
+	extract := exec.Command(goBin, "tool", "pack", "x", pkgArch)
+	membersDir := filepath.Join(dir, "members")
+	if err := os.MkdirAll(membersDir, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	extract.Dir = membersDir
+	if out, err := extract.CombinedOutput(); err != nil {
+		t.Fatalf("pack x: %v\n%s", err, out)
+	}
+	listCmd := exec.Command(goBin, "tool", "pack", "t", pkgArch)
+	listOut, err := listCmd.CombinedOutput()
+	if err != nil {
+		t.Fatalf("pack t: %v\n%s", err, listOut)
+	}
+	newArch := filepath.Join(dir, "pkg.a")
+	args := []string{"tool", "pack", "c", newArch}
+	seen := map[string]bool{}
+	for _, m := range strings.Fields(string(listOut)) {
+		if seen[m] {
+			continue
+		}
+		seen[m] = true
+		if err := os.Chmod(filepath.Join(membersDir, m), 0o644); err != nil {
+			t.Fatal(err)
+		}
+		args = append(args, filepath.Join(membersDir, m))
+	}
+	pack := exec.Command(goBin, args...)
+	pack.Dir = membersDir
+	if out, err := pack.CombinedOutput(); err != nil {
+		t.Fatalf("pack c: %v\n%s", err, out)
+	}
+
+	// Link with our archive.  The link line carries a GOROOT assignment
+	// and $WORK placeholders; run it through the shell with the
+	// GOEXPERIMENT the toolchain expects (the linker compares the object
+	// header against its own, experiments included).
+	goExp, _ := exec.Command(goBin, "env", "GOEXPERIMENT").Output()
+	linkLine = strings.ReplaceAll(linkLine, "$WORK", work)
+	linkLine = strings.ReplaceAll(linkLine, filepath.Join(work, "b001", "_pkg_.a"), newArch)
+	linkLine = strings.ReplaceAll(linkLine, filepath.Join(work, "b001", "exe", "a.out"), filepath.Join(dir, "app2"))
+	link := exec.Command("sh", "-c", linkLine)
+	link.Dir = dir
+	link.Env = append(os.Environ(), "GOEXPERIMENT="+strings.TrimSpace(string(goExp)))
+	if out, err := link.CombinedOutput(); err != nil {
+		t.Fatalf("link with gasm object: %v\n%s", err, out)
+	}
+	got, err := exec.Command(filepath.Join(dir, "app2")).CombinedOutput()
+	if err != nil {
+		t.Fatalf("run gasm-linked binary: %v\n%s", err, got)
+	}
+	if !bytes.Equal(got, baseOut) {
+		t.Errorf("gasm-linked output %q, want baseline %q", got, baseOut)
+	}
+}
+
+// fieldAfter returns the whitespace-delimited field following the first
+// occurrence of flag in line.
+func fieldAfter(line, flag string) string {
+	fields := strings.Fields(line)
+	for i, f := range fields {
+		if f == flag && i+1 < len(fields) {
+			return fields[i+1]
+		}
+	}
+	return ""
+}

asm/instrs.go

@@ -661,6 +661,66 @@ func (e *enc) encodeMovExtend(base string, ops []Operand) error {
 	return e.emit(i)
 }
 
+// --- legacy SSE moves --------------------------------------------------------
+
+// sseMove describes a legacy (non-VEX) SSE move: a mandatory prefix plus a
+// load opcode (reg = destination, rm = source) and a store opcode (the
+// reverse).  The Plan 9 names MOVOU/MOVO are the integer unaligned/aligned
+// octa moves (MOVDQU/MOVDQA), not the packed-single ones.
+type sseMove struct {
+	prefix byte // 0, 0x66, 0xF2 or 0xF3
+	load   byte
+	store  byte
+}
+
+var sseMoveTable = map[string]sseMove{
+	"MOVOU":  {0xF3, 0x6F, 0x7F}, // MOVDQU — unaligned octa
+	"MOVO":   {0x66, 0x6F, 0x7F}, // MOVDQA — aligned octa
+	"MOVUPS": {0x00, 0x10, 0x11}, // unaligned packed single
+	"MOVAPS": {0x00, 0x28, 0x29}, // aligned packed single
+	"MOVUPD": {0x66, 0x10, 0x11}, // unaligned packed double
+	"MOVAPD": {0x66, 0x28, 0x29}, // aligned packed double
+	"MOVSD":  {0xF2, 0x10, 0x11}, // scalar double
+	"MOVSS":  {0xF3, 0x10, 0x11}, // scalar single
+}
+
+// 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 {
+	if len(ops) != 2 {
+		return fmt.Errorf("SSE move expects 2 operands, got %d", len(ops))
+	}
+	src, dst := ops[0], ops[1]
+	srcReg, srcVec := vecReg(src)
+	dstReg, dstVec := vecReg(dst)
+	op := m.store
+	var reg Reg
+	var rm Operand
+	switch {
+	case srcVec && dstVec:
+		op = m.load
+		reg, rm = dstReg, src
+	case srcVec:
+		if _, ok := dst.(Mem); !ok {
+			return fmt.Errorf("SSE move: invalid destination operand")
+		}
+		reg, rm = srcReg, dst
+	case dstVec:
+		if _, ok := src.(Mem); !ok {
+			return fmt.Errorf("SSE move: invalid source operand")
+		}
+		op = m.load
+		reg, rm = dstReg, src
+	default:
+		return fmt.Errorf("SSE move needs a vector register operand")
+	}
+	i := &instr{prefix: m.prefix, opcode: []byte{0x0F, op}, modrm: -1, sib: -1}
+	if err := setRM(i, reg, rm, 8); err != nil {
+		return err
+	}
+	return e.emit(i)
+}
+
 // --- CVTSL2SD / CVTSQ2SD -----------------------------------------------------
 
 // encodeCvtsi2sd encodes a signed integer to scalar double conversion

asm/link.go

@@ -5,27 +5,73 @@ package asm
 
 import (
 	"fmt"
+	"sort"
 
 	"sourcedock.dev/petrbalvin/gasm-devkit/ast"
 )
 
 // Image is an assembled file: the function bodies laid out in source order,
-// followed by the file's static data section (GLOBL/DATA).  Static-symbol
+// followed by the file's static data section (GLOBL/DATA).  References to
-// references are encoded RIP-relative and resolved within the image, so the
+// file-local static symbols are encoded RIP-relative and resolved within the
-// bytes are self-consistent and executable at any base address.
+// image, so the raw bytes are self-consistent and executable at any base
+// address; references to external symbols are recorded as relocations
+// (Funcs[i].Relocs, Externals) and left unresolved — the object-file
+// emitters turn them into linker relocations.
 type Image struct {
-	Code    []byte         // concatenated function bodies
+	Code      []byte         // concatenated function bodies
-	Data    []byte         // static data section
+	Data      []byte         // static data section
-	Funcs   []FuncLayout   // function positions, in source order
+	Funcs     []FuncLayout   // function positions, in source order
-	Symbols map[string]int // static symbol → byte offset within the image
+	Symbols   map[string]int // static symbol → byte offset within the image
+	DataSyms  []DataSymbol   // GLOBL symbols, in layout order
+	Externals []string       // referenced but undefined symbols, sorted
 }
 
 // FuncLayout describes one assembled function within an Image.
 type FuncLayout struct {
+	Name    string
+	Pkg     string // explicit package prefix ("" = the current package)
+	Static  bool   // the <> marker: file-local, not exported
+	Offset  int    // start offset within the image (== offset within Code)
+	Size    int
+	Args    int            // declared argument/result area (the TEXT size suffix)
+	Frame   int            // local frame size (the TEXT $framesize)
+	NoSplit bool           // the NOSPLIT flag
+	SPWrite bool           // the SPWRITE flag: writes an arbitrary value to SP
+	Line    int            // source line of the TEXT directive
+	Labels  map[string]int // local labels, function-relative
+	Relocs  []Reloc        // static-symbol references, in emission order
+	Spadj   []SpadjStep    // stack-adjustment boundaries, ascending by PC
+}
+
+// SpadjStep is one stack-adjustment boundary: Value is the SP delta from the
+// entry state in effect from PC (function-relative) until the next step.
+type SpadjStep struct {
+	PC    int
+	Value int
+}
+
+// Reloc is one static-symbol reference within a function body: the disp32
+// field at Off (function-relative) must reach the symbol plus Addend,
+// measured from After, the address just past the instruction.  An External
+// relocation names a symbol no GLOBL in the file defines; the object-file
+// emitters carry it into the output's relocation table.
+type Reloc struct {
+	Off      int
+	After    int
+	Name     string
+	Addend   int64
+	External bool
+}
+
+// DataSymbol describes one GLOBL symbol laid out in the data section.
+type DataSymbol struct {
 	Name   string
-	Offset int // start offset within the image (== offset within Code)
+	Pkg    string // explicit package prefix ("" = the current package)
+	Offset int    // byte offset within Data
 	Size   int
-	Labels map[string]int // local labels, function-relative
+	Static bool // the <> marker: file-local, not exported
+	Rodata bool // the RODATA flag: read-only data
+	Dupok  bool // the DUPOK flag: duplicate-OK
 }
 
 // Bytes returns the whole image: code, then data.
@@ -37,19 +83,21 @@ func (img *Image) Bytes() []byte {
 
 // 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
-// static-symbol reference becomes a RIP-relative load whose displacement is
+// reference to a file-local static symbol becomes a RIP-relative load whose
-// resolved against that layout.  External (non-file-local) symbol references
+// displacement is resolved against that layout; a reference to a symbol no
-// are rejected: they need object-file emission.
+// 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) {
-	syms, order, err := collectData(f)
+	dataSyms, err := collectData(f)
 	if err != nil {
 		return nil, err
 	}
-	known := make(map[string]bool, len(syms))
+	known := make(map[string]bool, len(dataSyms))
-	for name := range syms {
+	for _, d := range dataSyms {
-		known[name] = true
+		known[d.name] = true
 	}
-	link := &linkInfo{symbols: known}
+	link := &linkInfo{symbols: known, allowExternal: true}
 
 	img := &Image{Symbols: map[string]int{}}
 	type asmFunc struct {
@@ -62,50 +110,100 @@ func AssembleFile(f *ast.File) (*Image, error) {
 		if !ok {
 			continue
 		}
-		code, patches, labels, err := assemble(t, link)
+		code, patches, labels, steps, err := assemble(t, link)
 		if err != nil {
 			return nil, fmt.Errorf("%s: %w", t.Name.Name, err)
 		}
-		img.Funcs = append(img.Funcs, FuncLayout{
+		fl := FuncLayout{
 			Name:   t.Name.Name,
+			Pkg:    t.Name.Pkg,
+			Static: t.Name.Static,
 			Offset: len(img.Code),
 			Size:   len(code),
+			Frame:  frameSize(t),
+			Args:   argsSize(t),
+			Line:   t.Pos().Line,
 			Labels: labels,
-		})
+		}
+		for _, f := range t.Flags {
+			switch f {
+			case "NOSPLIT":
+				fl.NoSplit = true
+			case "SPWRITE":
+				fl.SPWrite = true
+			}
+		}
+		for _, s := range steps {
+			fl.Spadj = append(fl.Spadj, SpadjStep{PC: s.pc, Value: s.value})
+		}
+		img.Funcs = append(img.Funcs, fl)
 		img.Code = append(img.Code, code...)
 		funcs = append(funcs, asmFunc{name: t.Name.Name, patches: patches})
 	}
 
 	// Lay out the data section behind the code, each symbol 16-aligned.
 	dataStart := len(img.Code)
-	for _, name := range order {
+	for _, d := range dataSyms {
 		if pos := dataStart + len(img.Data); pos != align16(pos) {
 			img.Data = append(img.Data, make([]byte, align16(pos)-pos)...)
 		}
-		img.Symbols[name] = dataStart + len(img.Data)
+		img.Symbols[d.name] = dataStart + len(img.Data)
-		img.Data = append(img.Data, syms[name]...)
+		img.DataSyms = append(img.DataSyms, DataSymbol{
+			Name:   d.name,
+			Pkg:    d.pkg,
+			Offset: len(img.Data),
+			Size:   len(d.buf),
+			Static: d.static,
+			Rodata: d.rodata,
+			Dupok:  d.dupok,
+		})
+		img.Data = append(img.Data, d.buf...)
 	}
 
-	// Resolve the RIP-relative displacements now that every address is known.
+	// Resolve the RIP-relative displacements of file-local references now
+	// that every address is known, and record every reference (resolved or
+	// external) for the object-file emitters.
+	externals := map[string]bool{}
 	for i, fn := range funcs {
 		base := img.Funcs[i].Offset
 		code := img.Code[base : base+img.Funcs[i].Size]
 		for _, p := range fn.patches {
-			rel := int64(img.Symbols[p.name]) + p.addend - int64(base+p.after)
+			reloc := Reloc{Off: p.off, After: p.after, Name: p.name, Addend: p.addend}
-			if rel < -1<<31 || rel >= 1<<31 {
+			if imgOff, ok := img.Symbols[p.name]; ok {
-				return nil, fmt.Errorf("%s: displacement to %q out of rel32 range", fn.name, p.name)
+				rel := int64(imgOff) + p.addend - int64(base+p.after)
+				if rel < -1<<31 || rel >= 1<<31 {
+					return nil, fmt.Errorf("%s: displacement to %q out of rel32 range", fn.name, p.name)
+				}
+				copy(code[p.off:p.off+4], le32(rel))
+			} else {
+				reloc.External = true
+				externals[p.name] = true
 			}
-			copy(code[p.off:p.off+4], le32(rel))
+			img.Funcs[i].Relocs = append(img.Funcs[i].Relocs, reloc)
 		}
 	}
+	for name := range externals {
+		img.Externals = append(img.Externals, name)
+	}
+	sort.Strings(img.Externals)
 	return img, nil
 }
 
+// dataSym is one GLOBL symbol and its DATA initialiser.
+type dataSym struct {
+	name   string
+	pkg    string
+	buf    []byte
+	static bool
+	rodata bool
+	dupok  bool
+}
+
 // collectData gathers the file's static symbols (GLOBL) and their initial
 // contents (DATA) into byte buffers, in declaration order.
-func collectData(f *ast.File) (map[string][]byte, []string, error) {
+func collectData(f *ast.File) ([]dataSym, error) {
-	syms := map[string][]byte{}
+	index := map[string]int{}
-	var order []string
+	var syms []dataSym
 	for _, d := range f.Decls {
 		switch dd := d.(type) {
 		case *ast.Globl:
@@ -113,50 +211,88 @@ func collectData(f *ast.File) (map[string][]byte, []string, error) {
 				continue
 			}
 			name := dd.Name.Name
-			if _, dup := syms[name]; dup {
+			if _, dup := index[name]; dup {
-				return nil, nil, fmt.Errorf("duplicate GLOBL %q", name)
+				return nil, fmt.Errorf("duplicate GLOBL %q", name)
 			}
 			size := 0
 			if dd.Size != nil && dd.Size.Imm.HasVal {
 				size = int(dd.Size.Imm.Val)
 			}
-			syms[name] = make([]byte, size)
+			index[name] = len(syms)
-			order = append(order, name)
+			ds := dataSym{
+				name:   name,
+				pkg:    dd.Name.Pkg,
+				buf:    make([]byte, size),
+				static: dd.Name.Static,
+			}
+			for _, f := range dd.Flags {
+				switch f {
+				case "RODATA":
+					ds.rodata = true
+				case "DUPOK":
+					ds.dupok = true
+				case "1":
+					ds.dupok = true
+				case "8":
+					ds.rodata = true
+				case "9":
+					ds.dupok = true
+					ds.rodata = true
+				}
+			}
+			syms = append(syms, ds)
 
 		case *ast.Data:
 			if dd.Name == nil || dd.Name.Pseudo != "SB" {
 				continue
 			}
-			buf, ok := syms[dd.Name.Name]
+			i, ok := index[dd.Name.Name]
 			if !ok {
-				return nil, nil, fmt.Errorf("DATA %q: no matching GLOBL", dd.Name.Name)
+				return nil, fmt.Errorf("DATA %q: no matching GLOBL", dd.Name.Name)
 			}
 			if dd.Value == nil || !dd.Value.Imm.HasVal {
-				return nil, nil, fmt.Errorf("DATA %q: value must be an integer immediate", dd.Name.Name)
+				return nil, fmt.Errorf("DATA %q: value must be an integer immediate", dd.Name.Name)
 			}
 			w := dd.Width
 			switch w {
 			case 1, 2, 4, 8:
 			default:
-				return nil, nil, fmt.Errorf("DATA %q: invalid width %d (want 1, 2, 4 or 8)", dd.Name.Name, w)
+				return nil, fmt.Errorf("DATA %q: invalid width %d (want 1, 2, 4 or 8)", dd.Name.Name, w)
 			}
 			off := dd.Name.Offset
+			buf := syms[i].buf
 			if off < 0 || off+int64(w) > int64(len(buf)) {
-				return nil, nil, fmt.Errorf("DATA %q+%d/%d exceeds GLOBL size %d", dd.Name.Name, off, w, len(buf))
+				return nil, fmt.Errorf("DATA %q+%d/%d exceeds GLOBL size %d", dd.Name.Name, off, w, len(buf))
 			}
 			v := dd.Value.Imm.Val
 			if dd.Value.Imm.Neg {
 				v = -v
 			}
-			for i := 0; i < w; i++ {
+			for j := 0; j < w; j++ {
-				buf[off+int64(i)] = byte(v >> (8 * i))
+				buf[off+int64(j)] = byte(v >> (8 * j))
 			}
 		}
 	}
-	return syms, order, nil
+	return syms, nil
 }
 
 // align16 rounds n up to the next multiple of 16.
 func align16(n int) int {
 	return (n + 15) &^ 15
 }
+
+// frameSize returns the local frame size declared on the TEXT directive.
+func frameSize(t *ast.Text) int {
+	if t.Frame != nil && t.Frame.Imm.HasVal {
+		return int(t.Frame.Imm.Val)
+	}
+	return 0
+}
+
+// argsSize returns the argument/result area declared on the TEXT directive.
+func argsSize(t *ast.Text) int {
+	if t.Args != nil && t.Args.Imm.HasVal {
+		return int(t.Args.Imm.Val)
+	}
+	return 0
+}

asm/macho.go

@@ -0,0 +1,258 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package asm
+
+import (
+	"encoding/binary"
+	"fmt"
+)
+
+// This file emits Mach-O x86-64 objects (MH_OBJECT) from an assembled
+// Image, in the shape the Darwin assembler produces: one unnamed segment
+// carrying a __TEXT,__text and a __DATA,__data section laid out back to
+// back at addresses zero and len(code), a symbol table (locals first, then
+// exported definitions, then undefined externals) and one relocation entry
+// per static-symbol reference, of type X86_64_RELOC_SIGNED.
+//
+// The image's own address space carries straight over — the data section
+// starts immediately after the code, and the layout padding already lives
+// inside Image.Data — so every symbol keeps its image address as its
+// n_value, and a local (non-external) relocation leaves the displacement
+// the assembler resolved in place: the linker only adjusts it by the
+// section's final movement.
+
+// Mach-O constants.
+const (
+	machoMagic64   = 0xfeedfacf
+	machoCPUamd64  = 0x01000007 // CPU_TYPE_X86_64
+	machoCPUSubAll = 3          // CPU_SUBTYPE_X86_64_ALL
+	machoObj       = 1          // MH_OBJECT
+
+	machoSegment64 = 0x19 // LC_SEGMENT_64
+	machoSymtab    = 0x2  // LC_SYMTAB
+
+	machoSectTextFlags = 0x80000400 // S_ATTR_PURE_INSTRUCTIONS | S_ATTR_SOME_INSTRUCTIONS
+
+	nUndf = 0x00 // undefined symbol
+	nSect = 0x0e // defined in section number n_sect
+	nExt  = 0x01 // external (exported or undefined-global) bit
+
+	x8664RelocSigned = 1
+)
+
+// MachOObject returns the image as a Mach-O x86-64 relocatable object
+// (MH_OBJECT), the shape the Darwin toolchain links.  Symbol names follow
+// the same rules as the ELF output.  Every static-symbol reference becomes
+// an X86_64_RELOC_SIGNED relocation: external references against their
+// undefined symbol, file-local ones against the __DATA section with the
+// resolved displacement carried in the instruction bytes.
+func (img *Image) MachOObject() ([]byte, error) {
+	le := binary.LittleEndian
+
+	// Section ordinals (1-based, as Mach-O numbers them).
+	const (
+		sectText = 1
+		sectData = 2
+	)
+
+	// Object address space: code at 0, data immediately after (the layout
+	// padding is already part of img.Data, so image addresses are object
+	// addresses).
+	textAddr := uint64(0)
+	dataAddr := uint64(len(img.Code))
+	vmsize := dataAddr + uint64(len(img.Data))
+
+	// The code, with external displacements primed to addend − 4: the
+	// linker adds the symbol's address to the field as it stands.  Local
+	// displacements stay as the assembler resolved them.
+	code := append([]byte(nil), img.Code...)
+	for _, fn := range img.Funcs {
+		for _, r := range fn.Relocs {
+			if r.External {
+				// Prime the field to the addend measured from the patch
+				// site: the assembler records it from the instruction end,
+				// After − Off bytes past the field.
+				copy(code[fn.Offset+r.Off:], le32(r.Addend-int64(r.After-r.Off)))
+			}
+		}
+	}
+
+	// Symbols: locals first, then exported definitions, then undefined
+	// externals — the order the classic link editor expects.
+	type machoSym struct {
+		name  string
+		typ   byte
+		sect  byte
+		value uint64
+	}
+	var locals, globals, undefs []machoSym
+	for _, fn := range img.Funcs {
+		s := machoSym{name: objectName(fn.Pkg, fn.Name), typ: nSect, sect: sectText, value: textAddr + uint64(fn.Offset)}
+		if fn.Static {
+			locals = append(locals, s)
+		} else {
+			s.typ |= nExt
+			globals = append(globals, s)
+		}
+	}
+	for _, d := range img.DataSyms {
+		s := machoSym{name: objectName(d.Pkg, d.Name), typ: nSect, sect: sectData, value: dataAddr + uint64(d.Offset)}
+		if d.Static {
+			locals = append(locals, s)
+		} else {
+			s.typ |= nExt
+			globals = append(globals, s)
+		}
+	}
+	for _, name := range img.Externals {
+		undefs = append(undefs, machoSym{name: name, typ: nUndf | nExt})
+	}
+	syms := append(append(locals, globals...), undefs...)
+	symIdx := map[string]int{}
+	for i, s := range syms {
+		symIdx[s.name] = i
+	}
+
+	// Relocations, attached to the __text section.
+	type machoReloc struct {
+		addr   uint32
+		symnum uint32
+		extern bool
+	}
+	var relocs []machoReloc
+	for _, fn := range img.Funcs {
+		for _, r := range fn.Relocs {
+			rel := machoReloc{addr: uint32(fn.Offset + r.Off)}
+			if r.External {
+				idx, ok := symIdx[r.Name]
+				if !ok {
+					return nil, fmt.Errorf("relocation references unknown symbol %q", r.Name)
+				}
+				rel.symnum = uint32(idx)
+				rel.extern = true
+			} else {
+				// Section-relative: r_symbolnum carries the section number
+				// and the resolved displacement stays in the bytes.
+				rel.symnum = sectData
+			}
+			relocs = append(relocs, rel)
+		}
+	}
+
+	// The string table opens with the conventional " \0".
+	strtab := []byte{' ', 0}
+	strOff := map[string]int{}
+	for _, s := range syms {
+		if _, ok := strOff[s.name]; ok {
+			continue
+		}
+		strOff[s.name] = len(strtab)
+		strtab = append(strtab, s.name...)
+		strtab = append(strtab, 0)
+	}
+
+	// File layout: header, the two load commands, section data (code,
+	// data), the relocation table, the symbol table, the string table.
+	const (
+		hdrSize    = 32
+		segCmdSize = 72 + 2*80 // segment command with two sections
+		symCmdSize = 24
+	)
+	sizeofcmds := segCmdSize + symCmdSize
+	dataOff := hdrSize + sizeofcmds
+	reloff := dataOff + len(code) + len(img.Data)
+	symoff := reloff + 8*len(relocs)
+	stroff := symoff + 16*len(syms)
+
+	out := make([]byte, stroff+len(strtab))
+
+	// mach_header_64.
+	le.PutUint32(out[0:], machoMagic64)
+	le.PutUint32(out[4:], machoCPUamd64)
+	le.PutUint32(out[8:], machoCPUSubAll)
+	le.PutUint32(out[12:], machoObj)
+	le.PutUint32(out[16:], 2) // ncmds
+	le.PutUint32(out[20:], uint32(sizeofcmds))
+	le.PutUint32(out[24:], 0) // flags
+	le.PutUint32(out[28:], 0) // reserved
+
+	// LC_SEGMENT_64 with the two sections.
+	p := hdrSize
+	le.PutUint32(out[p:], machoSegment64)
+	le.PutUint32(out[p+4:], segCmdSize)
+	// segname: the empty string, zero-padded to 16 bytes.
+	le.PutUint64(out[p+8:], 0)
+	le.PutUint64(out[p+16:], 0)
+	le.PutUint64(out[p+24:], 0) // vmaddr
+	le.PutUint64(out[p+32:], vmsize)
+	le.PutUint64(out[p+40:], uint64(dataOff))
+	le.PutUint64(out[p+48:], vmsize)
+	le.PutUint32(out[p+56:], 7) // maxprot rwx
+	le.PutUint32(out[p+60:], 7) // initprot rwx
+	le.PutUint32(out[p+64:], 2) // nsects
+	le.PutUint32(out[p+68:], 0) // flags
+
+	// __TEXT,__text
+	s := p + 72
+	copy(out[s:], "__text")
+	copy(out[s+16:], "__TEXT")
+	le.PutUint64(out[s+32:], textAddr)
+	le.PutUint64(out[s+40:], uint64(len(code)))
+	le.PutUint32(out[s+48:], uint32(dataOff))
+	le.PutUint32(out[s+52:], 4) // align 2^4
+	le.PutUint32(out[s+56:], uint32(reloff))
+	le.PutUint32(out[s+60:], uint32(len(relocs)))
+	le.PutUint32(out[s+64:], machoSectTextFlags)
+
+	// __DATA,__data
+	s += 80
+	copy(out[s:], "__data")
+	copy(out[s+16:], "__DATA")
+	le.PutUint64(out[s+32:], dataAddr)
+	le.PutUint64(out[s+40:], uint64(len(img.Data)))
+	le.PutUint32(out[s+48:], uint32(dataOff+len(code)))
+	le.PutUint32(out[s+52:], 4) // align 2^4
+
+	// LC_SYMTAB.
+	p = hdrSize + segCmdSize
+	le.PutUint32(out[p:], machoSymtab)
+	le.PutUint32(out[p+4:], symCmdSize)
+	le.PutUint32(out[p+8:], uint32(symoff))
+	le.PutUint32(out[p+12:], uint32(len(syms)))
+	le.PutUint32(out[p+16:], uint32(stroff))
+	le.PutUint32(out[p+20:], uint32(len(strtab)))
+
+	// Section data.
+	copy(out[dataOff:], code)
+	copy(out[dataOff+len(code):], img.Data)
+
+	// Relocation entries.
+	for i, r := range relocs {
+		e := out[reloff+i*8:]
+		le.PutUint32(e[0:], r.addr)
+		bits := r.symnum & 0x00ffffff
+		bits |= 1 << 24 // r_pcrel
+		bits |= 2 << 25 // r_length = 4 bytes
+		if r.extern {
+			bits |= 1 << 27 // r_extern
+		}
+		bits |= x8664RelocSigned << 28
+		le.PutUint32(e[4:], bits)
+	}
+
+	// nlist_64 entries.
+	for i, s := range syms {
+		e := out[symoff+i*16:]
+		le.PutUint32(e[0:], uint32(strOff[s.name]))
+		e[4] = s.typ
+		e[5] = s.sect
+		le.PutUint16(e[6:], 0) // n_desc
+		le.PutUint64(e[8:], s.value)
+	}
+
+	// String table.
+	copy(out[stroff:], strtab)
+
+	return out, nil
+}

asm/macho_test.go

@@ -0,0 +1,127 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package asm
+
+import (
+	"bytes"
+	"debug/macho"
+	"encoding/binary"
+	"testing"
+)
+
+// TestMachOObject checks the structure of the emitted MH_OBJECT: the two
+// sections and their addresses, the symbol table (types, sections, values)
+// and the __text relocation entries, parsed back with debug/macho.  No
+// Darwin toolchain is available on the test hosts, so the check is
+// structural — the ELF output carries the end-to-end link-and-run proof of
+// the shared symbol and relocation model.
+func TestMachOObject(t *testing.T) {
+	img := elfTestImage(t)
+	obj, err := img.MachOObject()
+	if err != nil {
+		t.Fatalf("MachOObject: %v", err)
+	}
+	f, err := macho.NewFile(bytes.NewReader(obj))
+	if err != nil {
+		t.Fatalf("parse emitted object: %v", err)
+	}
+	defer f.Close()
+
+	if f.Type != macho.TypeObj {
+		t.Errorf("file type = %v, want MH_OBJECT", f.Type)
+	}
+	if f.Cpu != macho.CpuAmd64 {
+		t.Errorf("cpu = %v, want CpuAmd64", f.Cpu)
+	}
+
+	text := f.Section("__text")
+	data := f.Section("__data")
+	if text == nil || data == nil {
+		t.Fatal("missing __text or __data section")
+	}
+	if text.Addr != 0 || text.Size != uint64(len(img.Code)) {
+		t.Errorf("__text addr/size = %#x/%d, want 0/%d", text.Addr, text.Size, len(img.Code))
+	}
+	if data.Addr != uint64(len(img.Code)) {
+		t.Errorf("__data addr = %#x, want %#x", data.Addr, len(img.Code))
+	}
+
+	// Symbol table: locals, exported definitions, undefined externals.
+	syms := f.Symtab.Syms
+	byName := map[string]macho.Symbol{}
+	for _, s := range syms {
+		byName[s.Name] = s
+	}
+	wantSym := func(name string, typ, sect uint8, value uint64) {
+		t.Helper()
+		s, ok := byName[name]
+		if !ok {
+			t.Errorf("symbol %q not found", name)
+			return
+		}
+		if s.Type != typ || s.Sect != sect || s.Value != value {
+			t.Errorf("%s: type/sect/value = %#x/%d/%#x, want %#x/%d/%#x",
+				name, s.Type, s.Sect, s.Value, typ, sect, value)
+		}
+	}
+	const (
+		defined   = nSect | nExt
+		local     = nSect
+		undefined = nUndf | nExt
+	)
+	wantSym("addq", defined, 1, 0)
+	wantSym("getanswer", defined, 1, 5)
+	wantSym("useextern", defined, 1, 13)
+	answer := byName["answer"]
+	if answer.Type != local || answer.Sect != 2 {
+		t.Errorf("answer: type/sect = %#x/%d, want %#x/2", answer.Type, answer.Sect, local)
+	}
+	wantSym("extvar", undefined, 0, 0)
+
+	// Relocations: both X86_64_RELOC_SIGNED, PC-relative, 4 bytes wide.
+	// The local one carries its section number in Value, the external one
+	// its symbol number.
+	if len(text.Relocs) != 2 {
+		t.Fatalf("__text relocs = %d, want 2", len(text.Relocs))
+	}
+	var sawLocal, sawExternal bool
+	for _, r := range text.Relocs {
+		if !r.Pcrel || r.Len != 2 || r.Type != x8664RelocSigned {
+			t.Errorf("reloc at %#x: pcrel/len/type = %v/%d/%d", r.Addr, r.Pcrel, r.Len, r.Type)
+		}
+		switch {
+		case r.Extern:
+			if name := syms[r.Value].Name; name != "extvar" {
+				t.Errorf("external reloc at %#x names %q, want extvar", r.Addr, name)
+			}
+			sawExternal = true
+		default:
+			if r.Value != 2 { // __data, the second section
+				t.Errorf("local reloc at %#x: section %d, want 2 (__data)", r.Addr, r.Value)
+			}
+			sawLocal = true
+		}
+	}
+	if !sawLocal || !sawExternal {
+		t.Errorf("relocs seen: local=%v external=%v, want both", sawLocal, sawExternal)
+	}
+
+	// The __text bytes are the image code, with the external displacement
+	// primed to addend − 4 and the local one left resolved.
+	textData, err := text.Data()
+	if err != nil {
+		t.Fatal(err)
+	}
+	want := append([]byte(nil), img.Code...)
+	for _, fn := range img.Funcs {
+		for _, r := range fn.Relocs {
+			if r.Name == "extvar" {
+				binary.LittleEndian.PutUint32(want[fn.Offset+r.Off:], 0xfffffffc) // −4
+			}
+		}
+	}
+	if !bytes.Equal(textData, want) {
+		t.Errorf("__text bytes %x, want %x", textData, want)
+	}
+}

asm/vex.go

@@ -43,6 +43,13 @@ const (
 	// in ModRM.reg and the destination in r/m — the layout of the EVEX
 	// narrowing stores (VPMOVDW, VPMOVQD).
 	vexRMRev
+	// vexRMSrcLen is the two-operand conversion form `OP src, dst` whose
+	// vector length follows the source: the packed-double → dword
+	// conversions (VCVTPD2DQ/VCVTTPD2DQ and their X/Y spellings) narrow into
+	// an XMM destination, so the L bit rides with the wider source.  The
+	// mnemonic's spelling fixes the length (X = 128, Y = 256), which also
+	// covers a memory source.  ModRM.reg = dst, ModRM.rm = src, no vvvv.
+	vexRMSrcLen
 	// vexZero is the no-operand form (VZEROUPPER).
 	vexZero
 )
@@ -84,14 +91,38 @@ var vexTable = map[string]vexSpec{
 	"VPCMPGTQ": {2, 0x37, 0, 1, -1, vexNDS3},
 
 	// VEX.128/256.66.0F.WIG — packed double-precision arithmetic / logic.
-	"VADDPD":    {1, 0x58, 0, 1, -1, vexNDS3},
+	"VADDPD": {1, 0x58, 0, 1, -1, vexNDS3},
-	"VMULPD":    {1, 0x59, 0, 1, -1, vexNDS3},
+	"VMULPD": {1, 0x59, 0, 1, -1, vexNDS3},
+	"VSUBPD": {1, 0x5C, 0, 1, -1, vexNDS3},
+	"VDIVPD": {1, 0x5E, 0, 1, -1, vexNDS3},
+	"VMINPD": {1, 0x5D, 0, 1, -1, vexNDS3},
+	"VMAXPD": {1, 0x5F, 0, 1, -1, vexNDS3},
+	// VEX.128/256.0F.WIG — packed single-precision arithmetic.
+	"VADDPS":    {1, 0x58, 0, 0, -1, vexNDS3},
+	"VMULPS":    {1, 0x59, 0, 0, -1, vexNDS3},
+	"VSUBPS":    {1, 0x5C, 0, 0, -1, vexNDS3},
+	"VDIVPS":    {1, 0x5E, 0, 0, -1, vexNDS3},
+	"VMINPS":    {1, 0x5D, 0, 0, -1, vexNDS3},
+	"VMAXPS":    {1, 0x5F, 0, 0, -1, vexNDS3},
 	"VXORPD":    {1, 0x57, 0, 1, -1, vexNDS3},
 	"VUNPCKHPD": {1, 0x15, 0, 1, -1, vexNDS3},
+	"VUNPCKLPD": {1, 0x14, 0, 1, -1, vexNDS3},
 	// VEX.128.F2.0F.WIG — scalar double-precision arithmetic (the packed
 	// opcodes with an F2 pp).
 	"VADDSD": {1, 0x58, 0, 3, -1, vexNDS3},
+	"VSUBSD": {1, 0x5C, 0, 3, -1, vexNDS3},
 	"VMULSD": {1, 0x59, 0, 3, -1, vexNDS3},
+	"VDIVSD": {1, 0x5E, 0, 3, -1, vexNDS3},
+	"VMINSD": {1, 0x5D, 0, 3, -1, vexNDS3},
+	"VMAXSD": {1, 0x5F, 0, 3, -1, vexNDS3},
+	// VEX.128.F3.0F.WIG — scalar single-precision arithmetic (the packed
+	// opcodes with an F3 pp).
+	"VADDSS": {1, 0x58, 0, 2, -1, vexNDS3},
+	"VSUBSS": {1, 0x5C, 0, 2, -1, vexNDS3},
+	"VMULSS": {1, 0x59, 0, 2, -1, vexNDS3},
+	"VDIVSS": {1, 0x5E, 0, 2, -1, vexNDS3},
+	"VMINSS": {1, 0x5D, 0, 2, -1, vexNDS3},
+	"VMAXSS": {1, 0x5F, 0, 2, -1, vexNDS3},
 	// VEX.128/256.66.0F38.W1 — fused multiply-add (NDS form).
 	"VFMADD231PD": {2, 0xB8, 1, 1, -1, vexNDS3},
 
@@ -99,12 +130,33 @@ var vexTable = map[string]vexSpec{
 	// no vvvv).
 	"VPMOVSXWD":    {2, 0x23, 0, 1, -1, vexRM},
 	"VPMOVSXDQ":    {2, 0x25, 0, 1, -1, vexRM},
+	"VPMOVSXBD":    {2, 0x21, 0, 1, -1, vexRM},
+	"VPMOVSXBQ":    {2, 0x22, 0, 1, -1, vexRM},
+	"VPMOVSXWQ":    {2, 0x24, 0, 1, -1, vexRM},
 	"VPMOVZXDQ":    {2, 0x35, 0, 1, -1, vexRM},
+	"VPMOVZXBW":    {2, 0x30, 0, 1, -1, vexRM},
+	"VPMOVZXBD":    {2, 0x31, 0, 1, -1, vexRM},
+	"VPMOVZXBQ":    {2, 0x32, 0, 1, -1, vexRM},
+	"VPMOVZXWD":    {2, 0x33, 0, 1, -1, vexRM},
+	"VPMOVZXWQ":    {2, 0x34, 0, 1, -1, vexRM},
 	"VPBROADCASTD": {2, 0x58, 0, 1, -1, vexRM},
 	"VPBROADCASTQ": {2, 0x59, 0, 1, -1, vexRM},
 	// VEX.128/256.F3.0F.WIG — signed dword to packed double conversion
 	// (reg=dst, rm=src, no vvvv; the length follows the destination).
 	"VCVTDQ2PD": {1, 0xE6, 0, 2, -1, vexRM},
+	// VEX.128/256.0F.WIG — signed dword to packed single conversion
+	// (reg=dst, rm=src, no vvvv, no mandatory prefix).
+	"VCVTDQ2PS": {1, 0x5B, 0, 0, -1, vexRM},
+	// VEX.128/256.0F.WIG — packed single to packed double conversion
+	// (reg=dst, rm=src; the destination is the wide operand and sets the
+	// length).  Intel's maps prescribe the F3 prefix here (VEX.pp = 10), but
+	// the Go assembler emits the instruction with pp = 00, and gasm follows
+	// the Go assembler's bytes — its machine code is the oracle, not the
+	// manual.
+	"VCVTPS2PD": {1, 0x5A, 0, 0, -1, vexRM},
+	// VEX.128.F2.0F.WIG — duplicate the low double of each 128-bit lane
+	// (reg=dst, rm=src, no vvvv; the length follows the destination).
+	"VMOVDDUP": {1, 0x12, 0, 3, -1, vexRM},
 	// VEX.128/256.66.0F.WIG — move mask to a GPR (reg=gpr dst, rm=vec src).
 	"VPMOVMSKB": {1, 0xD7, 0, 1, -1, vexRM},
 	"VMOVMSKPS": {1, 0x50, 0, 0, -1, vexRM}, // no 66 prefix (that would be VMOVMSKPD)
@@ -132,12 +184,75 @@ var vexTable = map[string]vexSpec{
 	// VEX.256.66.0F3A.W0 — lane extract (reg=YMM src, rm=XMM/memory dst, imm8).
 	"VEXTRACTI128": {3, 0x39, 0, 1, -1, vexExtract},
 	"VEXTRACTF128": {3, 0x19, 0, 1, -1, vexExtract},
+	// VEX.128/256.66.0F3A.W0 — half-precision convert back ($imm, src, dst:
+	// reg=src, rm=XMM/memory dst, imm8 — the extract layout).
+	"VCVTPS2PH": {3, 0x1D, 0, 1, -1, vexExtract},
 
 	// VEX.128.0F.W0 — no operands.
 	"VZEROUPPER": {1, 0x77, 0, 0, -1, vexZero},
 
 	// VEX.128.0F.W0 — mask-register test (KTESTW k1, k2: reg = dst, rm = src).
 	"KTESTW": {1, 0x99, 0, 0, -1, vexRM},
+
+	// VEX.66.0F38.W0 — broadcast a single/double to all lanes (reg=dst,
+	// rm=scalar memory; SD is 256-bit only).
+	"VBROADCASTSS": {2, 0x18, 0, 1, -1, vexRM},
+	"VBROADCASTSD": {2, 0x19, 0, 1, -1, vexRM},
+	// VEX.66.0F38.W0 — half-precision convert (reg=dst, rm=half-width
+	// source).
+	"VCVTPH2PS": {2, 0x13, 0, 1, -1, vexRM},
+	// VEX.F3.0F.WIG — replicate even/odd singles (reg=dst, rm=src).
+	"VMOVSLDUP": {1, 0x12, 0, 2, -1, vexRM},
+	"VMOVSHDUP": {1, 0x16, 0, 2, -1, vexRM},
+	// VEX.66.0F.WIG — packed double to packed single conversion, the X/Y
+	// spellings: the destination is always XMM and the spelling fixes the
+	// source length (X = 128, Y = 256).
+	"VCVTPD2PSX": {1, 0x5A, 0, 1, -1, vexRMSrcLen},
+	"VCVTPD2PSY": {1, 0x5A, 0, 1, -1, vexRMSrcLen},
+
+	// VEX scalar conversions between vector and general-purpose registers.
+	// Vector to GPR (two operands: vec/mem source, GPR destination, vvvv
+	// unused; the length follows the source).
+	"VCVTSD2SI":   {1, 0x2D, 0, 3, -1, vexRM},
+	"VCVTSD2SIQ":  {1, 0x2D, 1, 3, -1, vexRM},
+	"VCVTSS2SI":   {1, 0x2D, 0, 2, -1, vexRM},
+	"VCVTSS2SIQ":  {1, 0x2D, 1, 2, -1, vexRM},
+	"VCVTTSD2SI":  {1, 0x2C, 0, 3, -1, vexRM},
+	"VCVTTSD2SIQ": {1, 0x2C, 1, 3, -1, vexRM},
+	"VCVTTSS2SI":  {1, 0x2C, 0, 2, -1, vexRM},
+	"VCVTTSS2SIQ": {1, 0x2C, 1, 2, -1, vexRM},
+	// GPR to vector (three operands: GPR/mem source in r/m, the preserved
+	// vector source in vvvv, vector destination in reg).
+	"VCVTSI2SDL": {1, 0x2A, 0, 3, -1, vexNDS3},
+	"VCVTSI2SDQ": {1, 0x2A, 1, 3, -1, vexNDS3},
+	"VCVTSI2SSL": {1, 0x2A, 0, 2, -1, vexNDS3},
+	"VCVTSI2SSQ": {1, 0x2A, 1, 2, -1, vexNDS3},
+
+	// VEX.128/256.66.0F.WIG — word shifts (opdigit selects the shift).
+	"VPSRLW": {1, 0x71, 0, 1, 2, vexShiftImm},
+	"VPSRAW": {1, 0x71, 0, 1, 4, vexShiftImm},
+	"VPSLLW": {1, 0x71, 0, 1, 6, vexShiftImm},
+
+	// VEX.F2.0F — packed double to packed dword conversions, truncating and
+	// non-truncating.  The destination is always XMM; the X/Y spellings fix
+	// the source length (XMM/YMM), and VEX.L follows it — see vexSrcLen.
+	"VCVTPD2DQX":  {1, 0xE6, 0, 3, -1, vexRMSrcLen},
+	"VCVTPD2DQY":  {1, 0xE6, 0, 3, -1, vexRMSrcLen},
+	"VCVTTPD2DQX": {1, 0xE6, 0, 1, -1, vexRMSrcLen},
+	"VCVTTPD2DQY": {1, 0xE6, 0, 1, -1, vexRMSrcLen},
+}
+
+// vexSrcLen maps a source-length conversion mnemonic (the X/Y spellings of
+// the packed-double → dword conversions) to its fixed vector length:
+// X = 128 (L = 0), Y = 256 (L = 1).  The spelling fixes the length even for
+// a memory source, matching the Go assembler's ytab.
+var vexSrcLen = map[string]int{
+	"VCVTPD2DQX":  0,
+	"VCVTPD2DQY":  1,
+	"VCVTTPD2DQX": 0,
+	"VCVTTPD2DQY": 1,
+	"VCVTPD2PSX":  0,
+	"VCVTPD2PSY":  1,
 }
 
 // vexVarShift maps the shift mnemonics to their variable-count opcode — the
@@ -182,6 +297,11 @@ var vexMoveTable = map[string]vexMoveSpec{
 	// VEX.128.F2.0F.WIG — scalar double move, memory operands only (the
 	// register form takes three operands and is not supported yet).
 	"VMOVSD": {1, 3, 0x10, 0x11, 0, 0, 0, 0, false, false, true},
+	// VEX.128.F3.0F.WIG — scalar single move, memory operands only.
+	"VMOVSS": {1, 2, 0x10, 0x11, 0, 0, 0, 0, false, false, true},
+	// VEX.128/256 — aligned packed moves.
+	"VMOVAPS": {1, 0, 0x28, 0x29, 0, 0, 0, 0, true, false, false},
+	"VMOVAPD": {1, 1, 0x28, 0x29, 0, 0, 0, 0, true, false, false},
 }
 
 // isVex reports whether the mnemonic is a VEX-encoded instruction we handle.
@@ -230,6 +350,8 @@ func (e *enc) encodeVex(mnemUpper string, ops []Operand) error {
 		return e.encodeVexNDS3Imm(spec, ops)
 	case vexExtract:
 		return e.encodeVexExtract(spec, ops)
+	case vexRMSrcLen:
+		return e.encodeVexRMSrcLen(mnemUpper, spec, ops)
 	case vexZero:
 		return e.encodeVexZero(mnemUpper, spec, ops)
 	}
@@ -295,6 +417,32 @@ func (e *enc) encodeVexRM(spec vexSpec, ops []Operand) error {
 	return e.emitVexFields(spec, l, regField, rBit, 15, src)
 }
 
+// encodeVexRMSrcLen encodes a length-narrowing conversion: OP src, dst with
+// the destination always XMM and the VEX.L bit following the source — fixed
+// by the mnemonic's spelling (VCVTPD2DQX = 128, VCVTPD2DQY = 256) even when
+// the source is memory.
+func (e *enc) encodeVexRMSrcLen(mnem string, spec vexSpec, ops []Operand) error {
+	if len(ops) != 2 {
+		return fmt.Errorf("conversion expects 2 operands, got %d", len(ops))
+	}
+	src, dst := ops[0], ops[1]
+	dstReg, ok := dst.(Reg)
+	if !ok || !dstReg.isVec() {
+		return fmt.Errorf("VEX destination must be a vector register")
+	}
+	ll, ok := vexSrcLen[mnem]
+	if !ok {
+		return fmt.Errorf("no fixed vector length for %s", mnem)
+	}
+	regField := dstReg.idx & 7
+	rBit := 0
+	if dstReg.idx >= 8 {
+		rBit = 1
+	}
+	// An unused vvvv field must be stored as all ones (v̄vvv = 1111).
+	return e.emitVexFields(spec, ll, regField, rBit, 15, src)
+}
+
 // encodeVexShiftImm encodes an immediate-shift instruction: OP $imm, src, dst.
 // The destination is carried in VEX.vvvv, the source in ModRM.rm, and the
 // shift kind in the ModRM.reg /digit.

asm/vex_test.go

@@ -39,11 +39,14 @@ func TestVexNDS3(t *testing.T) {
 		}
 		inst, err := x86asm.Decode(code, 64)
 		if err != nil {
-			t.Errorf("%s: Decode(% x): %v", mnem, code, err)
+			t.Errorf("%s: Decode(% x): %v", mnem, err, code)
 			continue
 		}
-		if inst.Op.String() != mnem {
+		// The decoder folds the Plan 9 L/Q GPR-width spellings (VCVTSI2SDL/
-			t.Errorf("%s: decoded as %s (% x)", mnem, inst.Op.String(), code)
+		// SDQ, SSL/SSQ) onto the base name; the W bit carries the width.
+		got := inst.Op.String()
+		if got != mnem && !(len(mnem) > len(got) && mnem[:len(got)] == got) {
+			t.Errorf("%s: decoded as %s (% x)", mnem, got, code)
 		}
 	}
 }
@@ -156,82 +159,121 @@ func TestVexShiftImm(t *testing.T) {
 // as well as every new operand form.
 func TestVexGroundTruth(t *testing.T) {
 	cases := []struct {
-		name string
+		name   string
-		mnem string
+		mnem   string
-		ops  []Operand
+		ops    []Operand
-		want string
+		want   string
+		wantOp string // decoded mnemonic, when it differs from mnem (the X/Y spellings)
 	}{
 		// Three-operand NDS form.
-		{"VPADDQ Y8,Y9,Y8", "VPADDQ", []Operand{vreg(t, "Y8"), vreg(t, "Y9"), vreg(t, "Y8")}, "c44135d4c0"},
+		{"VPADDQ Y8,Y9,Y8", "VPADDQ", []Operand{vreg(t, "Y8"), vreg(t, "Y9"), vreg(t, "Y8")}, "c44135d4c0", ""},
-		{"VPADDQ X9,X8,X8", "VPADDQ", []Operand{vreg(t, "X9"), vreg(t, "X8"), vreg(t, "X8")}, "c44139d4c1"},
+		{"VPADDQ X9,X8,X8", "VPADDQ", []Operand{vreg(t, "X9"), vreg(t, "X8"), vreg(t, "X8")}, "c44139d4c1", ""},
-		{"VPXOR X7,X7,X7", "VPXOR", []Operand{vreg(t, "X7"), vreg(t, "X7"), vreg(t, "X7")}, "c5c1efff"},
+		{"VPXOR X7,X7,X7", "VPXOR", []Operand{vreg(t, "X7"), vreg(t, "X7"), vreg(t, "X7")}, "c5c1efff", ""},
-		{"VPSHUFB Y1,Y2,Y3", "VPSHUFB", []Operand{vreg(t, "Y1"), vreg(t, "Y2"), vreg(t, "Y3")}, "c4e26d00d9"},
+		{"VPSHUFB Y1,Y2,Y3", "VPSHUFB", []Operand{vreg(t, "Y1"), vreg(t, "Y2"), vreg(t, "Y3")}, "c4e26d00d9", ""},
-		{"VPMULLD Y1,Y2,Y3", "VPMULLD", []Operand{vreg(t, "Y1"), vreg(t, "Y2"), vreg(t, "Y3")}, "c4e26d40d9"},
+		{"VPMULLD Y1,Y2,Y3", "VPMULLD", []Operand{vreg(t, "Y1"), vreg(t, "Y2"), vreg(t, "Y3")}, "c4e26d40d9", ""},
-		{"VPUNPCKLDQ Y4,Y3,Y5", "VPUNPCKLDQ", []Operand{vreg(t, "Y4"), vreg(t, "Y3"), vreg(t, "Y5")}, "c5e562ec"},
+		{"VPUNPCKLDQ Y4,Y3,Y5", "VPUNPCKLDQ", []Operand{vreg(t, "Y4"), vreg(t, "Y3"), vreg(t, "Y5")}, "c5e562ec", ""},
-		{"VPERMD Y1,Y2,Y3", "VPERMD", []Operand{vreg(t, "Y1"), vreg(t, "Y2"), vreg(t, "Y3")}, "c4e26d36d9"},
+		{"VPERMD Y1,Y2,Y3", "VPERMD", []Operand{vreg(t, "Y1"), vreg(t, "Y2"), vreg(t, "Y3")}, "c4e26d36d9", ""},
 		// Floating point (packed and scalar) and FMA — same NDS form, the pp
 		// bits and map select the operation.
-		{"VADDPD Y9,Y8,Y8", "VADDPD", []Operand{vreg(t, "Y9"), vreg(t, "Y8"), vreg(t, "Y8")}, "c4413d58c1"},
+		{"VADDPD Y9,Y8,Y8", "VADDPD", []Operand{vreg(t, "Y9"), vreg(t, "Y8"), vreg(t, "Y8")}, "c4413d58c1", ""},
-		{"VADDPD X1,X2,X3", "VADDPD", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "c5e958d9"},
+		{"VADDPD X1,X2,X3", "VADDPD", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "c5e958d9", ""},
-		{"VMULPD Y12,Y12,Y12", "VMULPD", []Operand{vreg(t, "Y12"), vreg(t, "Y12"), vreg(t, "Y12")}, "c4411d59e4"},
+		{"VMULPD Y12,Y12,Y12", "VMULPD", []Operand{vreg(t, "Y12"), vreg(t, "Y12"), vreg(t, "Y12")}, "c4411d59e4", ""},
-		{"VXORPD Y8,Y8,Y8", "VXORPD", []Operand{vreg(t, "Y8"), vreg(t, "Y8"), vreg(t, "Y8")}, "c4413d57c0"},
+		{"VXORPD Y8,Y8,Y8", "VXORPD", []Operand{vreg(t, "Y8"), vreg(t, "Y8"), vreg(t, "Y8")}, "c4413d57c0", ""},
-		{"VUNPCKHPD X8,X8,X9", "VUNPCKHPD", []Operand{vreg(t, "X8"), vreg(t, "X8"), vreg(t, "X9")}, "c4413915c8"},
+		{"VUNPCKHPD X8,X8,X9", "VUNPCKHPD", []Operand{vreg(t, "X8"), vreg(t, "X8"), vreg(t, "X9")}, "c4413915c8", ""},
-		{"VADDSD X9,X8,X8", "VADDSD", []Operand{vreg(t, "X9"), vreg(t, "X8"), vreg(t, "X8")}, "c4413b58c1"},
+		{"VADDSD X9,X8,X8", "VADDSD", []Operand{vreg(t, "X9"), vreg(t, "X8"), vreg(t, "X8")}, "c4413b58c1", ""},
-		{"VMULSD X0,X1,X1", "VMULSD", []Operand{vreg(t, "X0"), vreg(t, "X1"), vreg(t, "X1")}, "c5f359c8"},
+		{"VMULSD X0,X1,X1", "VMULSD", []Operand{vreg(t, "X0"), vreg(t, "X1"), vreg(t, "X1")}, "c5f359c8", ""},
-		{"VFMADD231PD Y14,Y12,Y8", "VFMADD231PD", []Operand{vreg(t, "Y14"), vreg(t, "Y12"), vreg(t, "Y8")}, "c4429db8c6"},
+		{"VFMADD231PD Y14,Y12,Y8", "VFMADD231PD", []Operand{vreg(t, "Y14"), vreg(t, "Y12"), vreg(t, "Y8")}, "c4429db8c6", ""},
-		{"VFMADD231PD (DI),Y12,Y8", "VFMADD231PD", []Operand{Ptr(DI, 0, 32), vreg(t, "Y12"), vreg(t, "Y8")}, "c4629db807"},
+		{"VFMADD231PD (DI),Y12,Y8", "VFMADD231PD", []Operand{Ptr(DI, 0, 32), vreg(t, "Y12"), vreg(t, "Y8")}, "c4629db807", ""},
 		// Two-operand reg/rm form (v̄vvv must be 1111).
-		{"VPMOVSXDQ X0,Y4", "VPMOVSXDQ", []Operand{vreg(t, "X0"), vreg(t, "Y4")}, "c4e27d25e0"},
+		{"VPMOVSXDQ X0,Y4", "VPMOVSXDQ", []Operand{vreg(t, "X0"), vreg(t, "Y4")}, "c4e27d25e0", ""},
-		{"VPMOVSXWD (SI),Y0", "VPMOVSXWD", []Operand{Ptr(SI, 0, 8), vreg(t, "Y0")}, "c4e27d2306"},
+		{"VPMOVSXWD (SI),Y0", "VPMOVSXWD", []Operand{Ptr(SI, 0, 8), vreg(t, "Y0")}, "c4e27d2306", ""},
-		{"VPBROADCASTD X0,Y15", "VPBROADCASTD", []Operand{vreg(t, "X0"), vreg(t, "Y15")}, "c4627d58f8"},
+		{"VPBROADCASTD X0,Y15", "VPBROADCASTD", []Operand{vreg(t, "X0"), vreg(t, "Y15")}, "c4627d58f8", ""},
-		{"VCVTDQ2PD X12,Y12", "VCVTDQ2PD", []Operand{vreg(t, "X12"), vreg(t, "Y12")}, "c4417ee6e4"},
+		{"VCVTDQ2PD X12,Y12", "VCVTDQ2PD", []Operand{vreg(t, "X12"), vreg(t, "Y12")}, "c4417ee6e4", ""},
-		{"VCVTDQ2PD (SI),Y4", "VCVTDQ2PD", []Operand{Ptr(SI, 0, 16), vreg(t, "Y4")}, "c5fee626"},
+		{"VCVTDQ2PD (SI),Y4", "VCVTDQ2PD", []Operand{Ptr(SI, 0, 16), vreg(t, "Y4")}, "c5fee626", ""},
-		{"VPMOVMSKB X11,AX", "VPMOVMSKB", []Operand{vreg(t, "X11"), AX}, "c4c179d7c3"},
+		{"VPMOVMSKB X11,AX", "VPMOVMSKB", []Operand{vreg(t, "X11"), AX}, "c4c179d7c3", ""},
-		{"VMOVMSKPS Y7,AX", "VMOVMSKPS", []Operand{vreg(t, "Y7"), AX}, "c5fc50c7"},
+		{"VMOVMSKPS Y7,AX", "VMOVMSKPS", []Operand{vreg(t, "Y7"), AX}, "c5fc50c7", ""},
 		// Immediate shifts.
-		{"VPSLLD $1,Y3,Y4", "VPSLLD", []Operand{Imm(1), vreg(t, "Y3"), vreg(t, "Y4")}, "c5dd72f301"},
+		{"VPSLLD $1,Y3,Y4", "VPSLLD", []Operand{Imm(1), vreg(t, "Y3"), vreg(t, "Y4")}, "c5dd72f301", ""},
-		{"VPSRLQ $2,Y5,Y6", "VPSRLQ", []Operand{Imm(2), vreg(t, "Y5"), vreg(t, "Y6")}, "c5cd73d502"},
+		{"VPSRLQ $2,Y5,Y6", "VPSRLQ", []Operand{Imm(2), vreg(t, "Y5"), vreg(t, "Y6")}, "c5cd73d502", ""},
 		// Variable-count shifts: the count lives in an XMM register or memory
 		// and the instruction takes the NDS form.
-		{"VPSRLQ X0,Y8,Y8", "VPSRLQ", []Operand{vreg(t, "X0"), vreg(t, "Y8"), vreg(t, "Y8")}, "c53dd3c0"},
+		{"VPSRLQ X0,Y8,Y8", "VPSRLQ", []Operand{vreg(t, "X0"), vreg(t, "Y8"), vreg(t, "Y8")}, "c53dd3c0", ""},
-		{"VPSRLQ (AX),Y8,Y8", "VPSRLQ", []Operand{Ptr(AX, 0, 16), vreg(t, "Y8"), vreg(t, "Y8")}, "c53dd300"},
+		{"VPSRLQ (AX),Y8,Y8", "VPSRLQ", []Operand{Ptr(AX, 0, 16), vreg(t, "Y8"), vreg(t, "Y8")}, "c53dd300", ""},
-		{"VPSLLD X0,Y1,Y2", "VPSLLD", []Operand{vreg(t, "X0"), vreg(t, "Y1"), vreg(t, "Y2")}, "c5f5f2d0"},
+		{"VPSLLD X0,Y1,Y2", "VPSLLD", []Operand{vreg(t, "X0"), vreg(t, "Y1"), vreg(t, "Y2")}, "c5f5f2d0", ""},
-		{"VPSRLD X0,Y1,Y2", "VPSRLD", []Operand{vreg(t, "X0"), vreg(t, "Y1"), vreg(t, "Y2")}, "c5f5d2d0"},
+		{"VPSRLD X0,Y1,Y2", "VPSRLD", []Operand{vreg(t, "X0"), vreg(t, "Y1"), vreg(t, "Y2")}, "c5f5d2d0", ""},
-		{"VPSRAD X0,Y1,Y2", "VPSRAD", []Operand{vreg(t, "X0"), vreg(t, "Y1"), vreg(t, "Y2")}, "c5f5e2d0"},
+		{"VPSRAD X0,Y1,Y2", "VPSRAD", []Operand{vreg(t, "X0"), vreg(t, "Y1"), vreg(t, "Y2")}, "c5f5e2d0", ""},
-		{"VPSLLQ X0,Y1,Y2", "VPSLLQ", []Operand{vreg(t, "X0"), vreg(t, "Y1"), vreg(t, "Y2")}, "c5f5f3d0"},
+		{"VPSLLQ X0,Y1,Y2", "VPSLLQ", []Operand{vreg(t, "X0"), vreg(t, "Y1"), vreg(t, "Y2")}, "c5f5f3d0", ""},
 		// Immediate shuffle (reg=dst, rm=src, imm8).
-		{"VPSHUFD $0xEE,X8,X9", "VPSHUFD", []Operand{Imm(0xEE), vreg(t, "X8"), vreg(t, "X9")}, "c4417970c8ee"},
+		{"VPSHUFD $0xEE,X8,X9", "VPSHUFD", []Operand{Imm(0xEE), vreg(t, "X8"), vreg(t, "X9")}, "c4417970c8ee", ""},
-		{"VPSHUFD $0xEE,Y1,Y2", "VPSHUFD", []Operand{Imm(0xEE), vreg(t, "Y1"), vreg(t, "Y2")}, "c5fd70d1ee"},
+		{"VPSHUFD $0xEE,Y1,Y2", "VPSHUFD", []Operand{Imm(0xEE), vreg(t, "Y1"), vreg(t, "Y2")}, "c5fd70d1ee", ""},
-		{"VPERMQ $0x1B,Y1,Y2", "VPERMQ", []Operand{Imm(0x1B), vreg(t, "Y1"), vreg(t, "Y2")}, "c4e3fd00d11b"},
+		{"VPERMQ $0x1B,Y1,Y2", "VPERMQ", []Operand{Imm(0x1B), vreg(t, "Y1"), vreg(t, "Y2")}, "c4e3fd00d11b", ""},
-		{"VPERMQ $0x1B,Y11,Y12", "VPERMQ", []Operand{Imm(0x1B), vreg(t, "Y11"), vreg(t, "Y12")}, "c443fd00e31b"},
+		{"VPERMQ $0x1B,Y11,Y12", "VPERMQ", []Operand{Imm(0x1B), vreg(t, "Y11"), vreg(t, "Y12")}, "c443fd00e31b", ""},
 		// Three-operand + immediate (reg=dst, vvvv=src1, rm=src2, imm8).
-		{"VSHUFPD $1,X1,X2,X3", "VSHUFPD", []Operand{Imm(1), vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "c5e9c6d901"},
+		{"VSHUFPD $1,X1,X2,X3", "VSHUFPD", []Operand{Imm(1), vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "c5e9c6d901", ""},
-		{"VSHUFPD $1,Y1,Y2,Y3", "VSHUFPD", []Operand{Imm(1), vreg(t, "Y1"), vreg(t, "Y2"), vreg(t, "Y3")}, "c5edc6d901"},
+		{"VSHUFPD $1,Y1,Y2,Y3", "VSHUFPD", []Operand{Imm(1), vreg(t, "Y1"), vreg(t, "Y2"), vreg(t, "Y3")}, "c5edc6d901", ""},
-		{"VPERM2I128 $0x31,Y1,Y2,Y3", "VPERM2I128", []Operand{Imm(0x31), vreg(t, "Y1"), vreg(t, "Y2"), vreg(t, "Y3")}, "c4e36d46d931"},
+		{"VPERM2I128 $0x31,Y1,Y2,Y3", "VPERM2I128", []Operand{Imm(0x31), vreg(t, "Y1"), vreg(t, "Y2"), vreg(t, "Y3")}, "c4e36d46d931", ""},
-		{"VINSERTI128 $1,X5,Y1,Y2", "VINSERTI128", []Operand{Imm(1), vreg(t, "X5"), vreg(t, "Y1"), vreg(t, "Y2")}, "c4e37538d501"},
+		{"VINSERTI128 $1,X5,Y1,Y2", "VINSERTI128", []Operand{Imm(1), vreg(t, "X5"), vreg(t, "Y1"), vreg(t, "Y2")}, "c4e37538d501", ""},
 		// Lane extract (reg=YMM source, rm=XMM/memory destination, imm8).
-		{"VEXTRACTI128 $1,Y8,X9", "VEXTRACTI128", []Operand{Imm(1), vreg(t, "Y8"), vreg(t, "X9")}, "c4437d39c101"},
+		{"VEXTRACTI128 $1,Y8,X9", "VEXTRACTI128", []Operand{Imm(1), vreg(t, "Y8"), vreg(t, "X9")}, "c4437d39c101", ""},
-		{"VEXTRACTI128 $1,Y8,(DI)", "VEXTRACTI128", []Operand{Imm(1), vreg(t, "Y8"), Ptr(DI, 0, 16)}, "c4637d390701"},
+		{"VEXTRACTI128 $1,Y8,(DI)", "VEXTRACTI128", []Operand{Imm(1), vreg(t, "Y8"), Ptr(DI, 0, 16)}, "c4637d390701", ""},
-		{"VEXTRACTF128 $1,Y8,X9", "VEXTRACTF128", []Operand{Imm(1), vreg(t, "Y8"), vreg(t, "X9")}, "c4437d19c101"},
+		{"VEXTRACTF128 $1,Y8,X9", "VEXTRACTF128", []Operand{Imm(1), vreg(t, "Y8"), vreg(t, "X9")}, "c4437d19c101", ""},
 		// Moves — each direction picks its own opcode and VEX.W.
-		{"VMOVDQU (SI),Y1", "VMOVDQU", []Operand{Ptr(SI, 0, 32), vreg(t, "Y1")}, "c5fe6f0e"},
+		{"VMOVDQU (SI),Y1", "VMOVDQU", []Operand{Ptr(SI, 0, 32), vreg(t, "Y1")}, "c5fe6f0e", ""},
-		{"VMOVDQU Y3,(DI)", "VMOVDQU", []Operand{vreg(t, "Y3"), Ptr(DI, 0, 32)}, "c5fe7f1f"},
+		{"VMOVDQU Y3,(DI)", "VMOVDQU", []Operand{vreg(t, "Y3"), Ptr(DI, 0, 32)}, "c5fe7f1f", ""},
-		{"VMOVDQU X1,X2", "VMOVDQU", []Operand{vreg(t, "X1"), vreg(t, "X2")}, "c5fa7fca"},
+		{"VMOVDQU X1,X2", "VMOVDQU", []Operand{vreg(t, "X1"), vreg(t, "X2")}, "c5fa7fca", ""},
-		{"VMOVUPD (DI),Y14", "VMOVUPD", []Operand{Ptr(DI, 0, 32), vreg(t, "Y14")}, "c57d1037"},
+		{"VMOVUPD (DI),Y14", "VMOVUPD", []Operand{Ptr(DI, 0, 32), vreg(t, "Y14")}, "c57d1037", ""},
-		{"VMOVUPD Y14,(DI)", "VMOVUPD", []Operand{vreg(t, "Y14"), Ptr(DI, 0, 32)}, "c57d1137"},
+		{"VMOVUPD Y14,(DI)", "VMOVUPD", []Operand{vreg(t, "Y14"), Ptr(DI, 0, 32)}, "c57d1137", ""},
-		{"VMOVUPD X1,X2", "VMOVUPD", []Operand{vreg(t, "X1"), vreg(t, "X2")}, "c5f911ca"},
+		{"VMOVUPD X1,X2", "VMOVUPD", []Operand{vreg(t, "X1"), vreg(t, "X2")}, "c5f911ca", ""},
-		{"VMOVQ X8,AX", "VMOVQ", []Operand{vreg(t, "X8"), AX}, "c461f97ec0"},
+		{"VMOVQ X8,AX", "VMOVQ", []Operand{vreg(t, "X8"), AX}, "c461f97ec0", ""},
-		{"VMOVQ AX,X9", "VMOVQ", []Operand{AX, vreg(t, "X9")}, "c461f96ec8"},
+		{"VMOVQ AX,X9", "VMOVQ", []Operand{AX, vreg(t, "X9")}, "c461f96ec8", ""},
-		{"VMOVQ X8,(DI)", "VMOVQ", []Operand{vreg(t, "X8"), Ptr(DI, 0, 8)}, "c461f97e07"},
+		{"VMOVQ X8,(DI)", "VMOVQ", []Operand{vreg(t, "X8"), Ptr(DI, 0, 8)}, "c461f97e07", ""},
-		{"VMOVQ (SI),X9", "VMOVQ", []Operand{Ptr(SI, 0, 8), vreg(t, "X9")}, "c461f96e0e"},
+		{"VMOVQ (SI),X9", "VMOVQ", []Operand{Ptr(SI, 0, 8), vreg(t, "X9")}, "c461f96e0e", ""},
-		{"VMOVQ X8,X2", "VMOVQ", []Operand{vreg(t, "X8"), vreg(t, "X2")}, "c579d6c2"},
+		{"VMOVQ X8,X2", "VMOVQ", []Operand{vreg(t, "X8"), vreg(t, "X2")}, "c579d6c2", ""},
-		{"VMOVQ X2,X8", "VMOVQ", []Operand{vreg(t, "X2"), vreg(t, "X8")}, "c4c179d6d0"},
+		{"VMOVQ X2,X8", "VMOVQ", []Operand{vreg(t, "X2"), vreg(t, "X8")}, "c4c179d6d0", ""},
-		{"VMOVD X0,(SI)", "VMOVD", []Operand{vreg(t, "X0"), Ptr(SI, 0, 4)}, "c5f97e06"},
+		{"VMOVD X0,(SI)", "VMOVD", []Operand{vreg(t, "X0"), Ptr(SI, 0, 4)}, "c5f97e06", ""},
-		{"VMOVD AX,X0", "VMOVD", []Operand{AX, vreg(t, "X0")}, "c5f96ec0"},
+		{"VMOVD AX,X0", "VMOVD", []Operand{AX, vreg(t, "X0")}, "c5f96ec0", ""},
-		{"VMOVSD (SI),X8", "VMOVSD", []Operand{Ptr(SI, 0, 8), vreg(t, "X8")}, "c57b1006"},
+		{"VMOVSD (SI),X8", "VMOVSD", []Operand{Ptr(SI, 0, 8), vreg(t, "X8")}, "c57b1006", ""},
-		{"VMOVSD X8,(SI)", "VMOVSD", []Operand{vreg(t, "X8"), Ptr(SI, 0, 8)}, "c57b1106"},
+		{"VMOVSD X8,(SI)", "VMOVSD", []Operand{vreg(t, "X8"), Ptr(SI, 0, 8)}, "c57b1106", ""},
+		// Packed double arithmetic and unpack — the NDS form, the opcode
+		// selects the operation.
+		{"VSUBPD Y1,Y2,Y3", "VSUBPD", []Operand{vreg(t, "Y1"), vreg(t, "Y2"), vreg(t, "Y3")}, "c5ed5cd9", ""},
+		{"VDIVPD X1,X2,X3", "VDIVPD", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "c5e95ed9", ""},
+		{"VMINPD Y1,Y2,Y3", "VMINPD", []Operand{vreg(t, "Y1"), vreg(t, "Y2"), vreg(t, "Y3")}, "c5ed5dd9", ""},
+		{"VMAXPD X4,X5,X6", "VMAXPD", []Operand{vreg(t, "X4"), vreg(t, "X5"), vreg(t, "X6")}, "c5d15ff4", ""},
+		{"VUNPCKLPD X1,X2,X3", "VUNPCKLPD", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "c5e914d9", ""},
+		{"VUNPCKLPD Y1,Y2,Y3", "VUNPCKLPD", []Operand{vreg(t, "Y1"), vreg(t, "Y2"), vreg(t, "Y3")}, "c5ed14d9", ""},
+		{"VSUBPD (AX),X1,X2", "VSUBPD", []Operand{Ptr(AX, 0, 16), vreg(t, "X1"), vreg(t, "X2")}, "c5f15c10", ""},
+		// Scalar double and single arithmetic (F2 / F3 pp, 128-bit only).
+		{"VSUBSD X1,X2,X3", "VSUBSD", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "c5eb5cd9", ""},
+		{"VDIVSD X7,X1,X2", "VDIVSD", []Operand{vreg(t, "X7"), vreg(t, "X1"), vreg(t, "X2")}, "c5f35ed7", ""},
+		{"VMINSD X1,X2,X3", "VMINSD", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "c5eb5dd9", ""},
+		{"VMAXSD X3,X4,X5", "VMAXSD", []Operand{vreg(t, "X3"), vreg(t, "X4"), vreg(t, "X5")}, "c5db5feb", ""},
+		{"VADDSS X1,X2,X3", "VADDSS", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "c5ea58d9", ""},
+		{"VSUBSS X1,X2,X3", "VSUBSS", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "c5ea5cd9", ""},
+		{"VMULSS X9,X10,X11", "VMULSS", []Operand{vreg(t, "X9"), vreg(t, "X10"), vreg(t, "X11")}, "c4412a59d9", ""},
+		{"VDIVSS X1,X2,X3", "VDIVSS", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "c5ea5ed9", ""},
+		{"VMINSS X6,X7,X8", "VMINSS", []Operand{vreg(t, "X6"), vreg(t, "X7"), vreg(t, "X8")}, "c5425dc6", ""},
+		{"VMAXSS X1,X2,X3", "VMAXSS", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "c5ea5fd9", ""},
+		{"VADDSD 8(AX),X1,X2", "VADDSD", []Operand{Ptr(AX, 8, 8), vreg(t, "X1"), vreg(t, "X2")}, "c5f3585008", ""},
+		// VMOVDDUP — duplicate the low double (reg=dst, rm=src, F2 pp).
+		{"VMOVDDUP X1,X2", "VMOVDDUP", []Operand{vreg(t, "X1"), vreg(t, "X2")}, "c5fb12d1", ""},
+		{"VMOVDDUP Y1,Y2", "VMOVDDUP", []Operand{vreg(t, "Y1"), vreg(t, "Y2")}, "c5ff12d1", ""},
+		{"VMOVDDUP 8(AX),X1", "VMOVDDUP", []Operand{Ptr(AX, 8, 8), vreg(t, "X1")}, "c5fb124808", ""},
+		// Conversions: DQ→PS (no prefix), PS→PD (Go emits it without the F3
+		// prefix — see the table comment), DQ→PD.
+		{"VCVTDQ2PS X1,X2", "VCVTDQ2PS", []Operand{vreg(t, "X1"), vreg(t, "X2")}, "c5f85bd1", ""},
+		{"VCVTDQ2PS Y3,Y4", "VCVTDQ2PS", []Operand{vreg(t, "Y3"), vreg(t, "Y4")}, "c5fc5be3", ""},
+		{"VCVTPS2PD X1,X2", "VCVTPS2PD", []Operand{vreg(t, "X1"), vreg(t, "X2")}, "c5f85ad1", ""},
+		{"VCVTPS2PD X1,Y2", "VCVTPS2PD", []Operand{vreg(t, "X1"), vreg(t, "Y2")}, "c5fc5ad1", ""},
+		// PD→DQ conversions: the X/Y spellings fix the source length and the
+		// destination is always XMM; the decoder reports the base mnemonic.
+		{"VCVTPD2DQX X1,X2", "VCVTPD2DQX", []Operand{vreg(t, "X1"), vreg(t, "X2")}, "c5fbe6d1", "VCVTPD2DQ"},
+		{"VCVTPD2DQY Y1,X2", "VCVTPD2DQY", []Operand{vreg(t, "Y1"), vreg(t, "X2")}, "c5ffe6d1", "VCVTPD2DQ"},
+		{"VCVTTPD2DQX X3,X4", "VCVTTPD2DQX", []Operand{vreg(t, "X3"), vreg(t, "X4")}, "c5f9e6e3", "VCVTTPD2DQ"},
+		{"VCVTTPD2DQY Y5,X6", "VCVTTPD2DQY", []Operand{vreg(t, "Y5"), vreg(t, "X6")}, "c5fde6f5", "VCVTTPD2DQ"},
+		{"VCVTPD2DQY (AX),X1", "VCVTPD2DQY", []Operand{Ptr(AX, 0, 32), vreg(t, "X1")}, "c5ffe608", "VCVTPD2DQ"},
 		// No-operand.
-		{"VZEROUPPER", "VZEROUPPER", nil, "c5f877"},
+		{"VZEROUPPER", "VZEROUPPER", nil, "c5f877", ""},
 	}
 	for _, c := range cases {
 		code, err := Encode(c.mnem, c.ops...)
@@ -251,7 +293,11 @@ func TestVexGroundTruth(t *testing.T) {
 		if inst.Len != len(code) {
 			t.Errorf("%s: Decode consumed %d of %d bytes", c.name, inst.Len, len(code))
 		}
-		if inst.Op.String() != c.mnem {
+		wantOp := c.wantOp
+		if wantOp == "" {
+			wantOp = c.mnem
+		}
+		if inst.Op.String() != wantOp {
 			t.Errorf("%s: decoded as %s", c.name, inst.Op.String())
 		}
 	}

cmd/gasm/main.go

@@ -11,7 +11,9 @@ import (
 	"flag"
 	"fmt"
 	"io"
+	"io/fs"
 	"os"
+	"path/filepath"
 	"strings"
 
 	"sourcedock.dev/petrbalvin/gasm-devkit/arch"
@@ -22,11 +24,12 @@ import (
 	"sourcedock.dev/petrbalvin/gasm-devkit/lint"
 	"sourcedock.dev/petrbalvin/gasm-devkit/lsp"
 	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+	"sourcedock.dev/petrbalvin/gasm-devkit/verify"
 )
 
 // version is the release version, stamped at build time via
 // -ldflags "-X main.version=…" (defaulting to the current release).
-var version = "0.5.0"
+var version = "0.22.0"
 
 func main() {
 	if len(os.Args) < 2 {
@@ -44,33 +47,79 @@ func main() {
 		os.Exit(cmdLint(os.Args[2:]))
 	case "asm":
 		os.Exit(cmdAsm(os.Args[2:]))
+	case "verify":
+		os.Exit(cmdVerify(os.Args[2:]))
 	case "lsp":
 		os.Exit(cmdLSP(os.Args[2:]))
 	case "version", "--version", "-V":
-		fmt.Printf("gasm %s\n", version)
+		os.Exit(cmdVersion())
-	case "help", "-h", "--help":
+	case "help", "--help", "-h":
 		usage(os.Stdout)
 	default:
-		fmt.Fprintf(os.Stderr, "gasm: unknown command %q\n\n", os.Args[1])
+		fmt.Fprintf(os.Stderr, "gasm: unknown command %q — run \"gasm --help\" for usage\n", os.Args[1])
-		usage(os.Stderr)
 		os.Exit(2)
 	}
 }
 
+// cmdVersion prints the release version.
+func cmdVersion() int {
+	fmt.Printf("gasm %s\n", version)
+	return 0
+}
+
 func usage(w io.Writer) {
-	fmt.Fprintf(w, `gasm %s — developer tooling for Go's Plan 9 assembler
+	fmt.Fprintf(w, `gasm %s — developer tooling for Go's Plan 9 assembler (GAsm)
+
+gasm bundles a lexer, parser, formatter, linter, standalone assembler and
+language server for Plan 9 assembly into one self-contained binary.
 
 Usage:
-  gasm tokens <file>          print the lexical token stream
+  gasm <command> [arguments]
-  gasm parse  <file>          parse and report syntax errors
+  gasm [flags]
-  gasm fmt    [-w] <file...>  canonicalise formatting (-w writes in place)
+
-  gasm lint   <file...>       run static checks
+Commands:
-  gasm asm    [-o out.bin] <file>  assemble to machine code (amd64, Phase 2)
+  tokens    print the lexical token stream
-  gasm lsp                    run the language server over stdio
+  parse     parse and report syntax errors
-  gasm version                print the version
+  fmt       canonicalise formatting (gofmt for assembly)
+  lint      run static checks
+  asm       assemble .s files to machine code (amd64)
+  verify    JIT-assemble and run dynamic checks (amd64)
+  lsp       run the language server over stdio
+  version   print the version (same as --version)
+
+Flags:
+  -h, --help      show this help
+  -V, --version   print the version
+
+Run "gasm <command> -h" for a command's usage and flags.
+
+Examples:
+  gasm fmt                  reformat every .s below the current directory
+  gasm lint go-flac/*.s     run static checks over the kernels
+  gasm asm -o k.bin kern_amd64.s
+  gasm asm --format elf -o k.o kern_amd64.s
+  gasm asm --format goobj -p pkg/path -o k.o kern_amd64.s
 `, version)
 }
 
+// newCommand returns the FlagSet of a subcommand whose -h/--help prints a
+// proper usage block: the one-line usage, the long description and the flag
+// defaults.  The flag package routes -h/--help to fs.Usage and exits 0.
+func newCommand(name, usageLine, long string) *flag.FlagSet {
+	fs := flag.NewFlagSet(name, flag.ExitOnError)
+	fs.Usage = func() {
+		w := fs.Output()
+		fmt.Fprintf(w, "Usage: %s\n\n%s\n", usageLine, strings.TrimSpace(long))
+		hasFlags := false
+		fs.VisitAll(func(*flag.Flag) { hasFlags = true })
+		if hasFlags {
+			fmt.Fprintln(w, "\nFlags:")
+			fs.PrintDefaults()
+		}
+	}
+	return fs
+}
+
 // readSource returns the contents of path, or stdin when path is "-".
 func readSource(path string) (string, error) {
 	if path == "-" {
@@ -82,7 +131,10 @@ func readSource(path string) (string, error) {
 }
 
 func cmdTokens(args []string) int {
-	fs := flag.NewFlagSet("tokens", flag.ExitOnError)
+	fs := newCommand("tokens", "gasm tokens <file>", `
+Print the lexical token stream of FILE: position, token kind and text, one
+token per line.  FILE may be "-" to read standard input.
+`)
 	fs.Parse(args)
 	if fs.NArg() != 1 {
 		fmt.Fprintln(os.Stderr, "usage: gasm tokens <file>")
@@ -100,7 +152,11 @@ func cmdTokens(args []string) int {
 }
 
 func cmdParse(args []string) int {
-	fs := flag.NewFlagSet("parse", flag.ExitOnError)
+	fs := newCommand("parse", "gasm parse <file>", `
+Parse FILE and report syntax errors on stderr.  On success, print how many
+declarations and TEXT functions the file contains.  FILE may be "-" to read
+standard input.
+`)
 	fs.Parse(args)
 	if fs.NArg() != 1 {
 		fmt.Fprintln(os.Stderr, "usage: gasm parse <file>")
@@ -130,15 +186,49 @@ func cmdParse(args []string) int {
 }
 
 func cmdFmt(args []string) int {
-	fs := flag.NewFlagSet("fmt", flag.ExitOnError)
+	fs := newCommand("fmt", "gasm fmt [-w] [path...]", `
+Canonicalise the formatting of Plan 9 assembly sources: indentation, operand
+spacing, per-function mnemonic alignment and blank-line layout (exactly one
+blank line before each label, TEXT and GLOBL block).  Formatting is
+idempotent and preserves every line, comments included.
+
+With no paths — or a directory path — every .s file below it is reformatted
+in place and the changed files are listed, the way go fmt does; "." and "_"
+directories are skipped.  Explicit file paths print to stdout unless -w is
+given.
+`)
 	write := fs.Bool("w", false, "write result to the source file")
 	fs.Parse(args)
-	if fs.NArg() == 0 {
+	// Like go fmt: with no arguments, or with a directory argument, every .s
-		fmt.Fprintln(os.Stderr, "usage: gasm fmt [-w] <file...>")
+	// file below the directory is formatted in place and the names of the
-		return 2
+	// changed files are listed; explicit file arguments keep the -w / stdout
+	// behaviour.
+	paths := fs.Args()
+	dirMode := len(paths) == 0
+	if dirMode {
+		paths = []string{"."}
+	}
+	var files []string
+	for _, p := range paths {
+		info, err := os.Stat(p)
+		if err != nil {
+			fmt.Fprintln(os.Stderr, "gasm:", err)
+			return 1
+		}
+		if info.IsDir() {
+			dirMode = true
+			found, err := asmFiles(p)
+			if err != nil {
+				fmt.Fprintln(os.Stderr, "gasm:", err)
+				return 1
+			}
+			files = append(files, found...)
+			continue
+		}
+		files = append(files, p)
 	}
 	rc := 0
-	for _, path := range fs.Args() {
+	for _, path := range files {
 		src, err := readSource(path)
 		if err != nil {
 			fmt.Fprintln(os.Stderr, "gasm:", err)
@@ -146,11 +236,15 @@ func cmdFmt(args []string) int {
 			continue
 		}
 		out := format.Source(path, src)
-		if *write {
+		if dirMode || *write {
 			if out != src {
 				if err := os.WriteFile(path, []byte(out), 0o644); err != nil {
 					fmt.Fprintln(os.Stderr, "gasm:", err)
 					rc = 1
+					continue
+				}
+				if dirMode {
+					fmt.Println(path)
 				}
 			}
 			continue
@@ -160,8 +254,40 @@ func cmdFmt(args []string) int {
 	return rc
 }
 
+// asmFiles collects the .s files below dir, skipping directories whose name
+// starts with "." or "_" — as the go tooling does, which keeps .git and
+// scratch or reference trees (e.g. _refs) untouched.
+func asmFiles(dir string) ([]string, error) {
+	var out []string
+	err := filepath.WalkDir(dir, func(path string, d fs.DirEntry, err error) error {
+		if err != nil {
+			return err
+		}
+		if d.IsDir() {
+			if path != dir && (strings.HasPrefix(d.Name(), ".") || strings.HasPrefix(d.Name(), "_")) {
+				return filepath.SkipDir
+			}
+			return nil
+		}
+		if strings.HasSuffix(d.Name(), ".s") {
+			out = append(out, path)
+		}
+		return nil
+	})
+	return out, err
+}
+
 func cmdLint(args []string) int {
-	fs := flag.NewFlagSet("lint", flag.ExitOnError)
+	fs := newCommand("lint", "gasm lint <file...>", `
+Run the static checks over the given files and print diagnostics as
+"file:line:col: severity: message [code]".  The exit status is non-zero when
+an error-severity diagnostic is found; warnings (e.g. the register-clobber
+audit) do not affect it.
+
+Rules include unknown-instruction, operand-count, undefined-label,
+duplicate-label, missing-ret, missing-textflag-include, abi-argsize,
+unreachable-code, register-clobber and funcdata-pcdata.
+`)
 	disable := fs.String("disable", "", "comma-separated rule codes to disable")
 	fs.Parse(args)
 	if fs.NArg() == 0 {
@@ -202,7 +328,13 @@ func cmdLint(args []string) int {
 }
 
 func cmdLSP(args []string) int {
-	fs := flag.NewFlagSet("lsp", flag.ExitOnError)
+	fs := newCommand("lsp", "gasm lsp", `
+Run the language server over standard input/output: JSON-RPC 2.0 with
+Content-Length framing.  Point an LSP-capable editor at the binary and
+associate it with .s files; the target architecture is inferred from the file
+suffix (_amd64.s, _arm64.s, _riscv64.s, _loong64.s).  Provides completion,
+hover, document symbols, diagnostics and semantic-token highlighting.
+`)
 	fs.Parse(args)
 	srv := lsp.New(os.Stdin, os.Stdout)
 	if err := srv.Run(); err != nil {
@@ -213,11 +345,26 @@ func cmdLSP(args []string) int {
 }
 
 func cmdAsm(args []string) int {
-	fs := flag.NewFlagSet("asm", flag.ExitOnError)
+	fs := newCommand("asm", "gasm asm [--format raw|elf|macho|goobj] [-p pkg] [-o out] <file>", `
-	out := fs.String("o", "", "write the concatenated machine code to this file")
+Assemble FILE (amd64) without the Go toolchain: every TEXT function is
+encoded to machine code — scalar, VEX/AVX2 and EVEX/AVX-512 instructions,
+FP/SP frame mapping, local labels and file-local static symbols (GLOBL/DATA)
+resolved RIP-relative — and printed as a hex dump.
+
+With -o the output is written to a file instead.  The --format flag selects
+what is written: raw (the default) concatenates the functions and the data
+section into one self-consistent image; elf and macho emit a relocatable
+object (.text/.data sections, a symbol table and one PC32 relocation per
+static-symbol reference) that links with the system toolchain; goobj emits
+the Go toolchain's own object format, which cmd/link consumes directly (it
+requires -p, the package path, and the installed Go toolchain).
+`)
+	out := fs.String("o", "", "write the output to this file")
+	format := fs.String("format", "raw", "output format: raw (concatenated image), elf, macho or goobj (Go object)")
+	pkg := fs.String("p", "", "package path for --format goobj (qualifies the exported symbols)")
 	fs.Parse(args)
 	if fs.NArg() != 1 {
-		fmt.Fprintln(os.Stderr, "usage: gasm asm [-o out.bin] <file>")
+		fmt.Fprintln(os.Stderr, "usage: gasm asm [--format raw|elf|macho|goobj] [-p pkg] [-o out] <file>")
 		return 2
 	}
 	path := fs.Arg(0)
@@ -288,12 +435,123 @@ func cmdAsm(args []string) int {
 		}
 	}
 	if *out != "" {
-		all := img.Bytes()
+		var obj []byte
-		if err := os.WriteFile(*out, all, 0o644); err != nil {
+		var err error
+		var kind string
+		switch *format {
+		case "raw":
+			if len(img.Externals) > 0 {
+				fmt.Fprintf(os.Stderr, "gasm asm: external symbol %q needs an object file (use --format elf or --format macho)\n", img.Externals[0])
+				return 1
+			}
+			obj, kind = img.Bytes(), "raw image"
+		case "elf":
+			obj, err = img.ELFObject()
+			kind = "ELF object"
+		case "macho":
+			obj, err = img.MachOObject()
+			kind = "Mach-O object"
+		case "goobj":
+			obj, err = img.GOObject(*pkg, path)
+			kind = "Go object"
+		default:
+			fmt.Fprintf(os.Stderr, "gasm asm: unknown format %q (want raw, elf, macho or goobj)\n", *format)
+			return 2
+		}
+		if err != nil {
 			fmt.Fprintln(os.Stderr, "gasm asm:", err)
 			return 1
 		}
-		fmt.Printf("wrote %d bytes to %s\n", len(all), *out)
+		if err := os.WriteFile(*out, obj, 0o644); err != nil {
+			fmt.Fprintln(os.Stderr, "gasm asm:", err)
+			return 1
+		}
+		fmt.Printf("wrote %d bytes to %s (%s)\n", len(obj), *out, kind)
 	}
 	return 0
 }
+
+func cmdVerify(args []string) int {
+	fs := newCommand("verify", "gasm verify [-smoke] [-abi] [-profile] <file.s>", `
+Assemble FILE (amd64), map it into executable memory and report the available
+functions.  This confirms the assembled image is self-consistent (no
+unresolved external symbols) and executable — the prerequisite for dynamic
+testing.
+
+With -smoke, each NOSPLIT function is called with a zeroed argument block to
+confirm the JIT trampoline works end-to-end.  This is safe only for functions
+that tolerate nil pointers and zero lengths in their arguments.
+
+With -abi, each function is called with sentinel values in the callee-saved
+registers (BP, R14) and a red-zone canary below SP; violations are reported.
+
+With -profile, the static basic-block structure is listed for each function.
+`)
+	smoke := fs.Bool("smoke", false, "call each NOSPLIT function with zeroed args")
+	abi := fs.Bool("abi", false, "run ABI-checking calls (sentinel registers + red zone)")
+	profile := fs.Bool("profile", false, "list basic-block structure per function")
+	fs.Parse(args)
+	if fs.NArg() != 1 {
+		fmt.Fprintln(os.Stderr, "usage: gasm verify [-smoke] [-abi] [-profile] <file.s>")
+		return 2
+	}
+	path := fs.Arg(0)
+	if arch.FromFilename(path) != arch.AMD64 {
+		fmt.Fprintln(os.Stderr, "gasm verify: only amd64 is supported")
+		return 1
+	}
+
+	k, err := verify.Load(path)
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "gasm verify: %v\n", err)
+		return 1
+	}
+	defer k.Close()
+
+	names := k.FuncNames()
+	fmt.Printf("%s: %d functions JIT-loaded\n", path, len(names))
+	rc := 0
+	for _, name := range names {
+		fl, _ := k.Func(name)
+		flags := ""
+		if fl.NoSplit {
+			flags = " NOSPLIT"
+		}
+		fmt.Printf("  %s: %d bytes, args=%d, frame=%d%s\n", name, fl.Size, fl.Args, fl.Frame, flags)
+
+		if *profile {
+			blocks, err := k.Blocks(name)
+			if err != nil {
+				fmt.Printf("    profile: %v\n", err)
+			} else {
+				fmt.Printf("    blocks: %d\n", len(blocks))
+			}
+		}
+
+		if *smoke && fl.NoSplit {
+			args := make([]byte, fl.Args)
+			_, err := k.CallFunc(name, args)
+			if err != nil {
+				fmt.Printf("    smoke: FAIL — %v\n", err)
+				rc = 1
+			} else {
+				fmt.Printf("    smoke: OK\n")
+			}
+		}
+
+		if *abi && fl.NoSplit {
+			args := make([]byte, fl.Args)
+			_, report, err := k.CallFuncChecked(name, args)
+			if err != nil {
+				fmt.Printf("    abi: FAIL — %v\n", err)
+				rc = 1
+			} else if !report.OK() {
+				fmt.Printf("    abi: %s\n", report)
+				rc = 1
+			} else {
+				fmt.Printf("    abi: clean\n")
+			}
+		}
+	}
+	return rc
+}

cmd/gasm/main_test.go

@@ -52,6 +52,54 @@ func capture(fn func() int) (stdout, stderr string, code int) {
 	return string(ob), string(eb), code
 }
 
+// TestCmdFmtRecursive checks the go-fmt-style directory mode: with no
+// arguments every .s file below the working directory is formatted in place
+// ("." and "_" directories skipped), changed files are listed, and a second
+// run is a no-op.
+func TestCmdFmtRecursive(t *testing.T) {
+	tmp := t.TempDir()
+	t.Chdir(tmp)
+	unformatted := []byte("TEXT ·f(SB),NOSPLIT,$0\nRET\n")
+	write := func(path string) {
+		if err := os.MkdirAll(filepath.Dir(path), 0o755); err != nil {
+			t.Fatal(err)
+		}
+		if err := os.WriteFile(path, unformatted, 0o644); err != nil {
+			t.Fatal(err)
+		}
+	}
+	write("a_amd64.s")
+	write(filepath.Join("sub", "b_amd64.s"))
+	write(filepath.Join("_refs", "c_amd64.s"))
+	write(filepath.Join(".git", "d_amd64.s"))
+
+	out, errOut, code := capture(func() int { return cmdFmt(nil) })
+	if code != 0 {
+		t.Fatalf("code = %d (%s)", code, errOut)
+	}
+	if out != "a_amd64.s\n"+filepath.Join("sub", "b_amd64.s")+"\n" {
+		t.Errorf("listed files unexpected:\n%s", out)
+	}
+	for _, p := range []string{"a_amd64.s", filepath.Join("sub", "b_amd64.s")} {
+		b, _ := os.ReadFile(p)
+		if !strings.Contains(string(b), "\tRET") {
+			t.Errorf("%s not formatted in place:\n%s", p, b)
+		}
+	}
+	for _, p := range []string{filepath.Join("_refs", "c_amd64.s"), filepath.Join(".git", "d_amd64.s")} {
+		b, _ := os.ReadFile(p)
+		if string(b) != string(unformatted) {
+			t.Errorf("%s must not be touched:\n%s", p, b)
+		}
+	}
+
+	// Second pass: everything is canonical, nothing is listed.
+	out, _, code = capture(func() int { return cmdFmt(nil) })
+	if code != 0 || out != "" {
+		t.Errorf("second pass: code=%d out=%q, want a no-op", code, out)
+	}
+}
+
 func TestCmdTokens(t *testing.T) {
 	path := writeTemp(t, "f_amd64.s", clean)
 	out, _, code := capture(func() int { return cmdTokens([]string{path}) })
@@ -152,15 +200,32 @@ func TestCmdFmtWrite(t *testing.T) {
 func TestUsage(t *testing.T) {
 	var b bytes.Buffer
 	usage(&b)
-	if !strings.Contains(b.String(), "gasm") {
+	out := b.String()
-		t.Errorf("usage text unexpected:\n%s", b.String())
+	for _, want := range []string{
+		"gasm", "Commands:", "Flags:", "--help", "--version",
+		"tokens", "parse", "fmt", "lint", "asm", "lsp", "version",
+	} {
+		if !strings.Contains(out, want) {
+			t.Errorf("usage text missing %q:\n%s", want, out)
+		}
+	}
+}
+
+func TestCmdVersion(t *testing.T) {
+	out, _, code := capture(func() int { return cmdVersion() })
+	if code != 0 {
+		t.Fatalf("code = %d", code)
+	}
+	if !strings.Contains(out, version) {
+		t.Errorf("version output %q does not mention %q", out, version)
 	}
 }
 
 func TestCmdArgErrors(t *testing.T) {
-	// Missing file arguments produce a usage error (code 2).
+	// A missing path is an error (code 1); cmdFmt with no arguments is the
-	if _, _, code := capture(func() int { return cmdFmt(nil) }); code != 2 {
+	// recursive mode now, covered by TestCmdFmtRecursive.
-		t.Errorf("cmdFmt() code = %d, want 2", code)
+	if _, _, code := capture(func() int { return cmdFmt([]string{"no/such/path"}) }); code != 1 {
+		t.Errorf("cmdFmt(missing path) code = %d, want 1", code)
 	}
 	if _, _, code := capture(func() int { return cmdLint(nil) }); code != 2 {
 		t.Errorf("cmdLint() code = %d, want 2", code)

docs/ARCHITECTURE.md

@@ -136,13 +136,18 @@ Two deeper analyses sit on top of the AST:
   control-flow graph (basic blocks split at labels and after branches, with
   fall-through and jump-target edges), computes a conservative per-instruction
   register def/use, and runs the standard backward liveness iteration to a fixed
-  point.  On top of that it flags a **callee-saved register that is written but
+  point.  On top of that it flags writes to the registers the **Go ABI** fixes
-  never saved and restored** — the per-architecture callee-saved set is amd64
+  across calls that are never saved and restored — calibrated from
-  `BX/BP/R12–R15`, arm64 `R19–R30`, riscv64 `X1/X8/X9/X18–X27`, loong64
+  `cmd/compile/abi-internal.md`, *not* the platform ABI: Go's stack-based ABI0
-  `R1/R22–R31`.  This is an *audit*: the runtime's own assembly clobbers these
+  has no System V style callee-saved registers (amd64 `BX`, `R12`–`R15` and
-  registers freely (it controls both sides of the call), so the rule is
+  the like are caller-saved or permanent scratch, and hand-written kernels may
-  advisory there, but in hand-written kernels called from ordinary Go code a
+  clobber them freely).  The audited set is the frame pointer and the
-  clobber is a genuine ABI violation.  It runs only on macro-free files, where
+  the frame pointer, the goroutine pointer per architecture (amd64 `BP`/`R14`, arm64 `R18`/`R28`/
+  `R29`, riscv64 `X27`, loong64 `R22`); the goroutine pointer is reported only
+  when the function can reach the runtime — it is not `NOSPLIT` or makes a
+  call — since the ABI0 transition machinery restores it on those paths, and
+  NOSPLIT call-free leaves may use it (the runtime's own assembly does).  It
+  runs only on macro-free files, where
   no opaque macro can perform the save/restore.
 - **`funcdata-pcdata`.**  `FUNCDATA $idx, sym(SB)` and `PCDATA $idx, $val` are
   checked for well-formed operands (arity, immediate index and value, symbol
@@ -152,9 +157,15 @@ Two deeper analyses sit on top of the AST:
 ### `format`
 
 The formatter works on the **token stream, not the AST**, so it preserves
-every line — comments and blanks included.  It only normalises indentation,
+every line — comments and blanks included.  It normalises indentation, operand
-operand spacing and per-function mnemonic alignment.  It is idempotent and its
+spacing, per-function mnemonic alignment and blank-line layout: a new block
-output always round-trips through the parser.
+(a label, `TEXT` or `GLOBL`) is preceded by exactly one blank line (comments
+leading a block stay with it), runs of blanks collapse to one, and a `RET`
+terminates the body so the next function's doc comment stays at column 0.  It
+is idempotent and its output always round-trips through the parser.  With a
+directory argument — or none — it reformats every `.s` file below it in
+place and lists the files changed, the way `go fmt` does (`.` and `_`
+directories are skipped).
 
 ### `lsp`
 
@@ -202,15 +213,46 @@ three-operand-plus-immediate form (`VSHUFPD`,
 `VPERM2I128`, `VINSERTI128`), the lane-extract form (`VEXTRACTI128`,
 `VEXTRACTF128`, where the YMM source occupies the reg field and the XMM or
 memory destination r/m), the direction-sensitive moves (`VMOVDQU`, `VMOVUPD`,
-`VMOVD`, `VMOVQ`, `VMOVSD`), the floating-point and FMA arithmetic (`VADDPD`,
+`VMOVD`, `VMOVQ`, `VMOVSD`), the floating-point and FMA arithmetic — the
-`VMULPD`, `VXORPD`, `VUNPCKHPD`, the scalar `VADDSD`/`VMULSD`, `VCVTDQ2PD`,
+packed double operations (`VADDPD`/`VSUBPD`/`VMULPD`/`VDIVPD`/`VMINPD`/
-`VFMADD231PD`) and the no-operand `VZEROUPPER` — together with `VPERMD` and
+`VMAXPD`), the unpacks (`VUNPCKHPD`/`VUNPCKLPD`), the scalar SD and SS
+operations, `VMOVDDUP`, `VXORPD`, the width-changing conversions
+(`VCVTDQ2PS`, `VCVTPS2PD`, `VCVTDQ2PD`, and the `VCVTPD2DQX`/`Y` and
+`VCVTTPD2DQX`/`Y` spellings, whose length follows the wider source) and
+`VFMADD231PD` — and the no-operand `VZEROUPPER`, together with `VPERMD` and
 the scalar families (`CMOVcc`, `SETcc`, `LZCNT`/`TZCNT`, the extending moves,
 `CVTSx2SD`, `IMUL3`) and the EVEX (AVX-512) prefix — the four-byte prefix with
-5-bit register fields (Z0–Z31, X/Y 16–31), opmask registers as operands and
+5-bit register fields (Z0–Z31, X/Y 16–31, with the mod=11 quirk that carries
-mask destinations, and the compressed disp8×N displacement, whose multiplier
+rm[4] in X̄), opmask registers (K0–K7 as operands, mask destinations and
-follows the memory operand's size — covering every instruction the go-flac
+explicit merging/zeroing masks — written the way Go writes them, as a K
-AVX2 and AVX-512 kernels use.  Every encoding is validated two ways: by
+operand among the operands plus a `.Z` mnemonic suffix), and the compressed
+disp8×N displacement, whose multiplier follows the memory operand's size —
+covering every instruction the go-flac and go-lz4 AVX2/AVX-512 kernels use,
+plus the common AVX-512 F/BW integer set, the floating-point and conversion
+set (the packed double and single arithmetic, the scalar SD/SS forms —
+whose EVEX encodings serve masked and zeroing use — `VMOVDDUP`, the
+replicating moves, and the width-changing conversions, including the
+`VCVTPD2DQ`/`VCVTTPD2DQ` family whose length follows the wider source
+operand), and the wider AVX-512 set: ternary logic, lane shuffles, inserts
+and extracts, compares with an opmask destination, the permutes, the
+expand/compress family, the broadcasts, the opmask-register instructions
+(KAND/KOR/KXNOR/KADD/KUNPCK/KNOT/KSHIFTL/KORTEST and KMOVQ), the aligned
+moves and the remaining extending/narrowing moves, the floating-point
+helper and conversion tail (VRCP14*, VRSQRT14*, VGETEXP*, VGETMANT*,
+VSCALEF*, VRNDSCALE*, VREDUCE*, VFIXUPIMM*, VRANGE*, VFPCLASS* with an
+opmask destination, and the VCVT* conversions — signed, unsigned and
+truncating, including the length-suffixed X/Y spellings and the
+mask/vector conversions VPMOVM2*/VPMOV*2M, and the scalar conversions
+between vector and general-purpose registers (VCVT{,T}S{D,S}2SI{,Q} and
+the unsigned forms, VCVTSI2*/VCVTUSI2*), and gather/scatter with VSIB addressing — both the
+VEX spelling with a vector mask register and the EVEX spelling with an
+explicit K mask, where the EVEX length follows the VSIB index register,
+not the data register.  The EVEX mnemonic
+suffixes — rounding modes (.RN_SAE/.RD_SAE/.RU_SAE/.RZ_SAE),
+suppress-all-exceptions (.SAE) and memory broadcast (.BCST) — set the EVEX
+b bit and the L'L rounding-control field (broadcast keeps the vector length
+and scales disp8 by the element size), and combine with the .Z zeroing
+suffix.  Every encoding is validated two ways: by
 round-trip decoding through `golang.org/x/arch`, and byte-for-byte against
 the machine code the real Go assembler emits — a comparison that holds for
 whole functions: all 27 functions of both kernels assemble to exactly the Go
@@ -221,9 +263,51 @@ File-level assembly (`AssembleFile`) goes beyond single functions: it
 materialises the file's static symbols (`GLOBL`/`DATA`) in a data section
 behind the code and resolves references to them (`mask<>(SB)`) to
 RIP-relative loads whose displacements point inside the resulting image, so
-the bytes are self-consistent at any base address.  External (non-file-local)
+the bytes are self-consistent at any base address.  References to symbols no
-symbols are rejected: they need object-file emission, which — together with
+`GLOBL` defines are kept as relocations on the function layout, and the
-EVEX masking/zeroing and the other architectures — is the rest of Phase 2.
+object-file emitters turn the whole image into a linkable object: the ELF
+and Mach-O writers (`gasm asm --format elf|macho`) lay the code and data out
+as `.text`/`.data` (or `__text`/`__data`) sections, export a symbol per
+`TEXT` and `GLOBL` (the `<>` ones local, the rest global) and emit one
+PC-relative relocation per static-symbol reference — undefined external
+symbols included, so the output links with the system toolchain.  The GOOBJ
+emitter (`gasm asm --format goobj`) writes the format the Go linker consumes
+directly: the functions as non-package symbols (the way `cmd/asm` records
+assembly symbols), the `GLOBL` data, one `FuncInfo` per function and the
+pc-value tables — `pcsp` built from the prologue and epilogue stack
+boundaries, plus flat `pcfile`, `pcline` and `pcinline` tables — so a
+gasm-assembled object drops into a `go build` in place of the toolchain's.
+The object preamble (the version-and-experiment header the linker compares
+verbatim) is captured from the installed `go tool asm`, so the output is
+always consistent with the toolchain that links it.  External cross-package
+references and the implicit funcdata/DWARF symbols remain future work (the
+linker fills the latter's defaults); the rest of Phase 2 is those, the
+remaining EVEX forms and the other architectures.
+
+### `verify`
+
+The dynamic-analysis substrate (Phase 3).  It JIT-loads assembled images into
+executable memory and invokes them directly, enabling differential testing,
+runtime ABI checks and coverage profiling.
+
+The execution model is pure Go (stdlib only).  `Map` copies machine code into
+an anonymous `syscall.Mmap` mapping and enforces W^X (write the bytes, then
+`mprotect` to read-execute).  `Call` prepares a stack whose first word is the
+address of an assembly trampoline (`leaveJIT`), lays the ABI0 argument
+block after it, switches to that stack via `enterJIT` (which saves the Go
+stack pointer in a package global and jumps to the target), and recovers
+control when the function RETs into `leaveJIT` (which restores the Go stack
+and returns).  A 64-byte pad below the return address accommodates the
+ABIInternal wrapper that the Go runtime interposes on assembly functions.
+
+`Load` / `LoadSource` / `LoadAST` parse, assemble and map a `.s` file in one
+step, returning a `Kernel` whose `CallFunc` method marshals the argument block
+by name.  The image must be self-contained (no external relocations); the
+assembler’s `Image.Bytes()` provides the code-and-data concatenation.
+
+The `gasm verify` CLI subcommand exposes this: it loads a file, reports the
+available functions and (with `-smoke`) calls each NOSPLIT function with zeroed
+arguments to confirm the trampoline round-trips.
 
 ## Extension points
 

docs/DEFERRED.md

@@ -0,0 +1,56 @@
+# Deferred decisions
+
+Design decisions deliberately postponed, with enough context to pick them up
+again without re-deriving the analysis.  Each entry records what is deferred,
+why, the options on the table, and the trigger that should reopen it.
+
+---
+
+## GOOBJ external (cross-package) symbol references
+
+**Status:** deferred (v0.15.0, 2026-08-02).  The GOOBJ emitter resolves only
+symbols defined in the file being assembled; a reference to any other symbol
+is rejected.
+
+**Why it is deferred.**  GOOBJ symbol references are *positional*: a
+reference is a `{PkgIdx, SymIdx}` pair, where `SymIdx` is the index of the
+symbol in the *referenced package's* symbol-definition table.  That ordering
+is not derivable from the reference site — it lives in the referenced
+package's gc export data (the iexport binary format, which evolves with the
+toolchain).  `cmd/asm` reads it with `cmd/internal` readers gasm cannot
+import, so emitting external references means either parsing export data
+ourselves or taking a dependency that does.
+
+**What works today.**  Single-package objects: every symbol the file defines
+(as `TEXT` or `GLOBL`, static or exported) and every reference to them.
+This covers the production use case — the go-flac / go-lz4 kernels carry no
+`FUNCDATA`/`PCDATA`, hence no references into `runtime`, and the Go side
+references the assembly symbols, never the reverse.  Such a package builds
+with its assembly object replaced by a gasm-emitted one.
+
+**The options, when we return.**
+
+1. **`golang.org/x/tools/go/gcexportdata` as a production dependency.**
+   The straightforward path: read each imported package's export file
+   (paths from `-importcfg` or `go list -export`), assign symbol indices in
+   its symbol order, write `PkgIndex`/`Autolib` entries (fingerprints from
+   the export files' build IDs) and positional references.  Robust across
+   toolchain versions — `x/tools` tracks the format.  **Cost:** the first
+   production dependency beyond the standard library, an explicit deviation
+   from the "production code depends only on the standard library"
+   principle in the README.  Requires the user's explicit agreement.
+2. **A minimal iexport parser of our own.**  Preserves self-containment.
+   Substantial effort and inherently fragile: the format is an internal
+   contract that changes with Go releases, so the parser needs a
+   version-gated fallback and regression tests against several toolchains.
+3. **Shell out to the toolchain for symbol metadata.**  Consistent with the
+   existing GOOBJ preamble probe (which already runs `go tool asm`), but no
+   toolchain command exposes a package's symbols *in definition-index
+   order* — `go tool nm` sorts differently — so this does not solve the
+   core problem on its own; it would only feed option 1 or 2.
+
+**Trigger to reopen.**  An assembly file that needs a cross-package
+reference — in practice `FUNCDATA $…, runtime·…(SB)` (stack maps / GC
+metadata written in assembly), or any kernel that calls into another
+package directly.  Until then, option 3's limitation is moot and the
+single-package emitter suffices.

format/format.go

@@ -27,14 +27,6 @@ func Source(path, src string) string {
 		mnemLen int
 		funcID  int
 	}
-	const (
-		kBlank = iota
-		kComment
-		kPreproc
-		kDirective
-		kLabel
-		kInstr
-	)
 
 	infos := make([]info, len(lines))
 	funcID := -1
@@ -70,8 +62,8 @@ func Source(path, src string) string {
 		infos[i] = inf
 	}
 
-	// Second pass: render.
+	// Second pass: render each line.
-	var b strings.Builder
+	outs := make([]outLine, 0, len(lines))
 	inBody := false
 	for i, line := range lines {
 		inf := infos[i]
@@ -99,11 +91,94 @@ func Source(path, src string) string {
 			}
 		case kInstr:
 			out = renderInstr(line, maxWidth[inf.funcID])
+			// A RET ends the body for indentation purposes: comments that
+			// follow it — typically the next function's doc comment — belong
+			// at column 0, not inside the finished function.
+			if strings.EqualFold(line[0].Text, "RET") {
+				inBody = false
+			}
 		}
-		b.WriteString(strings.TrimRight(out, " \t"))
+		outs = append(outs, outLine{kind: inf.kind, text: strings.TrimRight(out, " \t")})
-		b.WriteByte('\n')
 	}
-	return b.String()
+	return normalizeSpacing(outs)
+}
+
+// Line classification, shared by the formatting passes.
+const (
+	kBlank = iota
+	kComment
+	kPreproc
+	kDirective
+	kLabel
+	kInstr
+)
+
+// outLine is one rendered line together with its classification.
+type outLine struct {
+	kind int
+	text string
+}
+
+// normalizeSpacing enforces the canonical blank-line layout: runs of blank
+// lines collapse to one, and a new block — a label, or a TEXT or GLOBL
+// directive — is preceded by exactly one blank line.  Comments immediately
+// above a block belong to it, so the blank line is inserted before them.  No
+// blank line is forced at the top of the file, right after a TEXT (the
+// function's first label), or between stacked labels that share an address.
+func normalizeSpacing(outs []outLine) string {
+	blockStart := func(ol outLine) bool {
+		switch ol.kind {
+		case kLabel:
+			return true
+		case kDirective:
+			// TEXT and GLOBL open a block; DATA continues a GLOBL block.
+			return strings.HasPrefix(ol.text, "TEXT") || strings.HasPrefix(ol.text, "GLOBL")
+		}
+		return false
+	}
+	insert := make([]bool, len(outs))
+	for i, ol := range outs {
+		if !blockStart(ol) {
+			continue
+		}
+		j := i
+		for j > 0 && outs[j-1].kind == kComment {
+			j--
+		}
+		if j == 0 {
+			continue // top of file
+		}
+		switch prev := outs[j-1]; {
+		case prev.kind == kBlank, prev.kind == kLabel:
+			continue // already separated, or stacked labels
+		case prev.kind == kDirective && strings.HasPrefix(prev.text, "TEXT"):
+			continue // the function's first label
+		}
+		insert[j] = true
+	}
+
+	var b strings.Builder
+	prevBlank := true // also suppresses leading blanks
+	for i, ol := range outs {
+		if insert[i] && !prevBlank {
+			b.WriteByte('\n')
+		}
+		if ol.kind == kBlank {
+			if !prevBlank {
+				b.WriteByte('\n')
+			}
+			prevBlank = true
+			continue
+		}
+		b.WriteString(ol.text)
+		b.WriteByte('\n')
+		prevBlank = false
+	}
+	out := strings.TrimRight(b.String(), "\n")
+	if out == "" {
+		return ""
+	}
+	return out + "\n"
 }
 
 // renderInstr renders an instruction line: a tab, the mnemonic padded to the

format/format_test.go

@@ -39,6 +39,102 @@ func TestGolden(t *testing.T) {
 	}
 }
 
+// TestDocCommentIndent checks that a doc comment preceding a TEXT directive
+// sits at column 0 even when another function (ending in RET) precedes it —
+// the RET must terminate the previous body for indentation purposes.
+func TestDocCommentIndent(t *testing.T) {
+	in := "#include \"textflag.h\"\n" +
+		"\n" +
+		"// func first()\n" +
+		"TEXT ·first(SB), NOSPLIT, $0\n" +
+		"XORQ AX, AX\n" +
+		"RET\n" +
+		"\n" +
+		"// func second()\n" +
+		"TEXT ·second(SB), NOSPLIT, $0\n" +
+		"RET\n"
+
+	want := "#include \"textflag.h\"\n" +
+		"\n" +
+		"// func first()\n" +
+		"TEXT ·first(SB), NOSPLIT, $0\n" +
+		"\tXORQ AX, AX\n" +
+		"\tRET\n" +
+		"\n" +
+		"// func second()\n" +
+		"TEXT ·second(SB), NOSPLIT, $0\n" +
+		"\tRET\n"
+
+	got := Source("d_amd64.s", in)
+	if got != want {
+		t.Fatalf("formatting mismatch:\n--- got ---\n%q\n--- want ---\n%q", got, want)
+	}
+	// Body comments stay indented.
+	body := "#include \"textflag.h\"\nTEXT ·f(SB), NOSPLIT, $0\n// inside the body\nXORQ AX, AX\nRET\n"
+	gotBody := Source("b_amd64.s", body)
+	if !strings.Contains(gotBody, "\t// inside the body\n") {
+		t.Fatalf("body comment must stay indented:\n%q", gotBody)
+	}
+}
+
+// TestBlankLines checks the blank-line canonicalisation: exactly one blank
+// line before a new block (a label, or TEXT/GLOBL), runs of blanks collapsed
+// to one, and no blank forced after TEXT, between stacked labels, or at the
+// top of the file.  Leading comments belong to the block they precede.
+func TestBlankLines(t *testing.T) {
+	in := "#include \"textflag.h\"\n" +
+		"TEXT ·f(SB), NOSPLIT, $0\n" +
+		"first:\n" + // first label: no blank after TEXT
+		"XORQ AX, AX\n" +
+		"JMP next\n" + // unlabeled glue: fmt inserts a blank before next:
+		"next:\n" +
+		"stacked:\n" + // stacked labels share an address: no blank between
+		"INCQ AX\n" +
+		"\n" +
+		"\n" + // two blanks collapse to one
+		"// separated block\n" + // comment belongs to the label below
+		"later:\n" +
+		"RET\n" +
+		"// func g()\n" + // doc comment: blank goes before it
+		"TEXT ·g(SB), NOSPLIT, $0\n" +
+		"RET\n" +
+		"GLOBL ·mask(SB), RODATA, $8\n" + // blank before GLOBL…
+		"DATA ·mask+0(SB)/4, $1\n" + // …but not before DATA
+		"\n" +
+		"\n" +
+		"\n" // trailing blanks dropped
+
+	want := "#include \"textflag.h\"\n" +
+		"\n" +
+		"TEXT ·f(SB), NOSPLIT, $0\n" +
+		"first:\n" +
+		"\tXORQ AX, AX\n" +
+		"\tJMP  next\n" +
+		"\n" +
+		"next:\n" +
+		"stacked:\n" +
+		"\tINCQ AX\n" +
+		"\n" +
+		"\t// separated block\n" + // body comment before a label stays indented
+		"later:\n" +
+		"\tRET\n" +
+		"\n" +
+		"// func g()\n" +
+		"TEXT ·g(SB), NOSPLIT, $0\n" +
+		"\tRET\n" +
+		"\n" +
+		"GLOBL ·mask(SB), RODATA, $8\n" +
+		"DATA ·mask+0(SB)/4, $1\n"
+
+	got := Source("b_amd64.s", in)
+	if got != want {
+		t.Fatalf("formatting mismatch:\n--- got ---\n%q\n--- want ---\n%q", got, want)
+	}
+	if again := Source("b_amd64.s", got); again != got {
+		t.Fatalf("not idempotent:\n%q", again)
+	}
+}
+
 func TestOperandSpacing(t *testing.T) {
 	cases := map[string]string{
 		"4(SI)":           "4(SI)",

justfile

@@ -3,7 +3,7 @@
 
 # gasm-devkit — developer tooling for Go's Plan 9 assembler (GAsm).
 
-version := "0.5.0"
+version := "0.22.0"
 
 default:
     @just --list

lint/lint.go

@@ -242,7 +242,7 @@ func lintText(t *ast.Text, tab *arch.Table, archKnown bool, cfg Config, macros m
 				}
 			}
 
-			if archKnown && !cfg.Disable[CodeOperandCount] && !isMacroInvocation(mnem, macros) {
+			if archKnown && !cfg.Disable[CodeOperandCount] && !isMacroInvocation(mnem, macros) && !maskedEvex(mnem, st.Operands) {
 				if in, ok := tab.Lookup(mnem); ok && in.MinOps >= 0 {
 					n := len(st.Operands)
 					if n < in.MinOps || n > in.MaxOps {
@@ -319,18 +319,27 @@ func lintText(t *ast.Text, tab *arch.Table, archKnown bool, cfg Config, macros m
 		}
 	}
 
-	// Register liveness: a callee-saved register that is written but never
+	// Register liveness: a register the Go ABI fixes across calls that is
-	// saved and restored is clobbered across the call.  The check runs over the
+	// written but never saved and restored is clobbered.  The check runs over
-	// control-flow graph and is skipped for macro-using files, where an opaque
+	// the control-flow graph and is skipped for macro-using files, where an
-	// macro may perform the save/restore.
+	// opaque macro may perform the save/restore.
 	if doLabelChecks && archKnown && !cfg.Disable[CodeRegisterClobber] {
 		live := analyzeLiveness(t, cfg.Arch)
-		if clobbered := clobberedCalleeSaved(live, cfg.Arch); len(clobbered) > 0 {
+		always, rt := clobberedGoFixed(live, cfg.Arch, reachesRuntime(t))
+		if len(always) > 0 {
 			out = append(out, Diagnostic{
 				Pos:      t.Keyword.Pos,
 				Severity: Warning,
 				Code:     CodeRegisterClobber,
-				Message:  fmt.Sprintf("callee-saved register(s) %s written but never saved/restored", strings.Join(clobbered, ", ")),
+				Message:  fmt.Sprintf("register(s) %s written but never saved/restored: fixed by the Go ABI (frame/goroutine pointer)", strings.Join(always, ", ")),
+			})
+		}
+		if len(rt) > 0 {
+			out = append(out, Diagnostic{
+				Pos:      t.Keyword.Pos,
+				Severity: Warning,
+				Code:     CodeRegisterClobber,
+				Message:  fmt.Sprintf("goroutine-pointer register(s) %s written but never saved/restored in a function that can reach the Go runtime", strings.Join(rt, ", ")),
 			})
 		}
 	}
@@ -341,6 +350,29 @@ func lintText(t *ast.Text, tab *arch.Table, archKnown bool, cfg Config, macros m
 	return out
 }
 
+// reachesRuntime reports whether a function can reach the Go runtime: it is
+// not NOSPLIT (so the stack-split and traceback machinery runs) or it makes a
+// CALL.  Goroutine-pointer registers must survive such functions; a NOSPLIT
+// leaf may clobber them, since the ABI0 transition restores them (the
+// runtime's own assembly relies on this, e.g. R14 on amd64).
+func reachesRuntime(t *ast.Text) bool {
+	nosplit := false
+	for _, f := range t.Flags {
+		if strings.EqualFold(f, "NOSPLIT") {
+			nosplit = true
+		}
+	}
+	for _, s := range t.Body {
+		if in, ok := s.(*ast.Instr); ok {
+			switch strings.ToUpper(in.Mnemonic.Text) {
+			case "CALL", "BL", "JAL": // amd64, arm64/loong64, riscv64 calls
+				return true
+			}
+		}
+	}
+	return !nosplit
+}
+
 // usesFPArgs reports whether a function references its arguments through the FP
 // pseudo-register — i.e. it uses the stack-based ABI0 layout, where the
 // declared argument size must match the signature.
@@ -406,6 +438,28 @@ func isMacroInvocation(mnem string, macros map[string]bool) bool {
 	return strings.Contains(mnem, "_") || macros[mnem]
 }
 
+// maskedEvex reports whether the instruction is a masked EVEX form: the
+// mnemonic carries a .Z suffix, or the operand list contains an opmask
+// register (K1–K7).  Either way the operand count differs from the unmasked
+// form, so count checks are skipped.
+func maskedEvex(mnem string, ops []*ast.Operand) bool {
+	if strings.Contains(mnem, ".") {
+		return true
+	}
+	for _, op := range ops {
+		if op.Kind == ast.OpAddr && op.Addr.Sym != nil && op.Addr.Base == "" &&
+			op.Addr.Index == "" && op.Addr.Sym.Pseudo == "" && isMaskReg(op.Addr.Sym.Name) {
+			return true
+		}
+	}
+	return false
+}
+
+// isMaskReg reports whether name is an opmask register K0–K7.
+func isMaskReg(name string) bool {
+	return len(name) == 2 && name[0] == 'K' && name[1] >= '0' && name[1] <= '7'
+}
+
 // isConditionalDirective reports whether a preprocessor directive (the text
 // after '#') is a conditional-compilation directive whose branches the parser
 // cannot resolve.

lint/lint_test.go

@@ -48,11 +48,11 @@ func TestFixtureIsClean(t *testing.T) {
 	if len(errs) > 0 {
 		t.Fatalf("parse: %v", errs)
 	}
-	// The fixture mirrors the go-flac kernels, which use callee-saved registers
+	// The fixture mirrors the go-flac kernels, which write the Go ABI0
-	// (BX, R13) without saving them; the register-clobber audit flags that by
+	// scratch registers (BX, R13) without saving them — legal under Go's
-	// design.  This test targets the other rules, so the audit is disabled here
+	// stack-based ABI, so the register-clobber audit stays silent and the
-	// (it is covered by TestRegisterClobber).
+	// fixture must lint entirely clean.
-	diags := File(f, Config{Arch: arch.AMD64, Disable: map[string]bool{CodeRegisterClobber: true}})
+	diags := File(f, Config{Arch: arch.AMD64})
 	if len(diags) != 0 {
 		t.Fatalf("expected no diagnostics on the fixture, got %+v", diags)
 	}
@@ -187,6 +187,26 @@ done:
 	}
 }
 
+// TestEvexMaskingRecognised checks that masked EVEX forms — the .Z suffix and
+// an explicit K operand — are recognised and exempt from operand-count
+// checks.
+func TestEvexMaskingRecognised(t *testing.T) {
+	diags := lintSrc(t, `
+#include "textflag.h"
+TEXT ·f(SB), NOSPLIT, $0
+	VPADDD.Z Z1, Z2, K2, Z3
+	VPMINSD Z1, Z2, K5, Z3
+	VMOVDQU8 Z1, K3, (SI)
+	RET
+`)
+	if codes(diags)[CodeUnknownInstr] != 0 {
+		t.Fatalf("masked EVEX must be recognised: %+v", diags)
+	}
+	if codes(diags)[CodeOperandCount] != 0 {
+		t.Fatalf("masked operand counts must not be flagged: %+v", diags)
+	}
+}
+
 func TestArm64AddressingSuffix(t *testing.T) {
 	// .W (pre-index) and .P (post-index) suffixes must resolve to the base
 	// instruction.

lint/liveness.go

@@ -4,7 +4,6 @@
 package lint
 
 import (
-	"fmt"
 	"sort"
 	"strings"
 
@@ -248,7 +247,7 @@ func instrEffect(in *ast.Instr, a arch.Arch) regEffect {
 	}
 
 	compare := isCompare(mnem)
-	dstIdx := dstIndex(in, a)
+	dstIdx := dstIndex(in)
 
 	for i, op := range in.Operands {
 		r := gprName(op, a)
@@ -281,13 +280,11 @@ func instrEffect(in *ast.Instr, a arch.Arch) regEffect {
 	return eff
 }
 
-// dstIndex returns the operand index of the destination register: last for the
+// dstIndex returns the operand index of the destination register: in Plan 9
-// Plan 9 (amd64) spelling, first for arm64/riscv64/loong64.
+// notation the destination is the last operand on every architecture Go
-func dstIndex(in *ast.Instr, a arch.Arch) int {
+// supports (amd64, arm64, riscv64 and loong64 alike).
-	if a == arch.AMD64 {
+func dstIndex(in *ast.Instr) int {
-		return len(in.Operands) - 1
+	return len(in.Operands) - 1
-	}
-	return 0
 }
 
 // isCompare reports whether the mnemonic only reads its operands (setting flags).
@@ -372,41 +369,37 @@ func sameSet(a, b map[string]bool) bool {
 	return true
 }
 
-// calleeSavedGPRs returns the general-purpose registers an assembly function
+// goFixedGPRs returns the general-purpose registers the Go ABI designates as
-// must preserve for its caller, using the register names the assembler accepts
+// fixed across calls — the ones hand-written assembly must not permanently
-// for each architecture.
+// clobber.  This follows cmd/compile/abi-internal.md, not the platform ABI:
-func calleeSavedGPRs(a arch.Arch) map[string]bool {
+// Go's stack-based ABI0 (which hand-written assembly uses) has no System V
+// style callee-saved registers, so clobbering the argument and scratch
+// registers (amd64 BX, R12, R13, R15, …) is legal.
+//
+// Two groups are returned.  always holds registers whose loss is never safe.
+// runtime holds registers that survive an ABI0 leaf only because the
+// transition machinery restores them (on amd64 the g pointer is reloaded
+// from TLS): clobbering them is safe exactly in NOSPLIT functions that make
+// no calls, which is how the runtime's own assembly uses them.
+func goFixedGPRs(a arch.Arch) (always, runtime map[string]bool) {
 	switch a {
 	case arch.AMD64:
-		return gprSet("BX", "BP", "R12", "R13", "R14", "R15")
+		// BP maintains the frame chain; R14 holds the current goroutine.
+		// R15 is scratch except in dynamically linked binaries, so it is not
+		// flagged.
+		return gprSet("BP"), gprSet("R14")
 	case arch.ARM64:
-		names := []string{"R29", "R30"} // FP, LR
+		// R18 is reserved for the OS on some platforms, R28 holds the current
-		for i := 19; i <= 28; i++ {
+		// goroutine, R29 is the frame pointer.
-			names = append(names, fmt.Sprintf("R%d", i))
+		return gprSet("R18", "R28", "R29"), nil
-		}
-		return gprSet(names...)
 	case arch.RISCV:
-		// RA (X1) and the S registers (X8, X9, X18–X27) are callee-saved.
+		// X27 holds the current goroutine.
-		names := []string{"X1", "RA", "X8", "X9", "S0", "S1", "FP"}
+		return gprSet("X27"), nil
-		for i := 18; i <= 27; i++ {
-			names = append(names, fmt.Sprintf("X%d", i))
-		}
-		for i := 2; i <= 11; i++ {
-			names = append(names, fmt.Sprintf("S%d", i))
-		}
-		return gprSet(names...)
 	case arch.LOONG64:
-		// RA (R1), FP (R22) and S0–S8 (R23–R31) are callee-saved.
+		// R22 holds the current goroutine.
-		names := []string{"R1", "RA", "R22", "FP"}
+		return gprSet("R22"), nil
-		for i := 23; i <= 31; i++ {
-			names = append(names, fmt.Sprintf("R%d", i))
-		}
-		for i := 0; i <= 8; i++ {
-			names = append(names, fmt.Sprintf("S%d", i))
-		}
-		return gprSet(names...)
 	}
-	return nil
+	return nil, nil
 }
 
 func gprSet(names ...string) map[string]bool {
@@ -417,15 +410,16 @@ func gprSet(names ...string) map[string]bool {
 	return m
 }
 
-// clobberedCalleeSaved returns the callee-saved registers a function writes
+// clobberedGoFixed returns the Go-ABI-fixed registers a function writes
-// without also saving and restoring them — i.e. registers whose caller-owned
+// without also saving and restoring them.  The first result lists registers
-// value is lost across the call.  It walks the blocks of the liveness analysis
+// whose loss is never safe; the second lists the goroutine-pointer class,
-// (so the control-flow graph is what supplies the instruction set) and
+// whose loss is reported only when reachesRuntime is true (a non-NOSPLIT
-// aggregates each instruction's register effects.
+// function, or one that makes calls — the ABI0 transition machinery restores
-func clobberedCalleeSaved(l *liveness, a arch.Arch) []string {
+// the g pointer only on such paths).
-	callee := calleeSavedGPRs(a)
+func clobberedGoFixed(l *liveness, a arch.Arch, reachesRuntime bool) (always, runtime []string) {
-	if len(callee) == 0 {
+	alwaysSet, runtimeSet := goFixedGPRs(a)
-		return nil
+	if len(alwaysSet) == 0 && len(runtimeSet) == 0 {
+		return nil, nil
 	}
 	def := map[string]bool{}
 	saved := map[string]bool{}
@@ -444,12 +438,19 @@ func clobberedCalleeSaved(l *liveness, a arch.Arch) []string {
 			}
 		}
 	}
-	var out []string
+	clobbered := func(set map[string]bool) []string {
-	for r := range callee {
+		var out []string
-		if def[r] && !(saved[r] && restored[r]) {
+		for r := range set {
-			out = append(out, r)
+			if def[r] && !(saved[r] && restored[r]) {
+				out = append(out, r)
+			}
 		}
+		sort.Strings(out)
+		return out
 	}
-	sort.Strings(out)
+	always = clobbered(alwaysSet)
-	return out
+	if reachesRuntime {
+		runtime = clobbered(runtimeSet)
+	}
+	return always, runtime
 }

lint/liveness_test.go

@@ -5,36 +5,132 @@ package lint
 
 import "testing"
 
-// TestRegisterClobber detects writes to callee-saved registers that are not
+// TestRegisterClobber checks the register-clobber audit is calibrated to the
-// saved and restored.
+// Go ABI (cmd/compile/abi-internal.md), not the platform ABI: Go's
+// stack-based ABI0 — which hand-written assembly uses — has no System V
+// style callee-saved registers, so argument and scratch registers may be
+// clobbered freely.  Only the registers the ABI fixes across calls (the
+// frame pointer, the goroutine pointer, OS-reserved registers) are audited.
 func TestRegisterClobber(t *testing.T) {
-	// BX (callee-saved on amd64) is written but never saved → clobbered.
+	// amd64: BX, R12, R13 and R15 are argument/permanent-scratch registers in
-	clob := lintSrc(t, "#include \"textflag.h\"\n"+
+	// Go ABI0 — writing them unsaved is legal (a System V calibration would
+	// report all of these).
+	scratch := lintSrc(t, "#include \"textflag.h\"\n"+
 		"TEXT ·f(SB), NOSPLIT, $0\n"+
 		"\tMOVQ CX, BX\n"+
+		"\tXORL R12, R12\n"+
+		"\tXORL R13, R13\n"+
+		"\tXORL R15, R15\n"+
 		"\tRET\n")
-	if codes(clob)[CodeRegisterClobber] != 1 {
+	if codes(scratch)[CodeRegisterClobber] != 0 {
-		t.Fatalf("unsaved callee-saved write should be flagged: %+v", clob)
+		t.Fatalf("Go ABI0 scratch registers must not be flagged: %+v", scratch)
 	}
 
-	// Saved and restored → preserved.
+	// amd64: R14 (the goroutine pointer) in a NOSPLIT function without calls
+	// is the runtime's own pattern — the ABI0 transition restores it — so it
+	// is not flagged.
+	leaf := lintSrc(t, "#include \"textflag.h\"\n"+
+		"TEXT ·f(SB), NOSPLIT, $0\n"+
+		"\tXORL R14, R14\n"+
+		"\tRET\n")
+	if codes(leaf)[CodeRegisterClobber] != 0 {
+		t.Fatalf("R14 in a NOSPLIT leaf must not be flagged: %+v", leaf)
+	}
+
+	// amd64: R14 in a function that makes a call is a genuine hazard.
+	withCall := lintSrc(t, "#include \"textflag.h\"\n"+
+		"TEXT ·f(SB), NOSPLIT, $0\n"+
+		"\tXORL R14, R14\n"+
+		"\tCALL ·g(SB)\n"+
+		"\tRET\n")
+	if codes(withCall)[CodeRegisterClobber] != 1 {
+		t.Fatalf("unsaved R14 with a call should be flagged: %+v", withCall)
+	}
+
+	// amd64: R14 in a non-NOSPLIT function is a hazard regardless of calls.
+	split := lintSrc(t, "#include \"textflag.h\"\n"+
+		"TEXT ·f(SB), $0\n"+
+		"\tMOVQ CX, R14\n"+
+		"\tRET\n")
+	if codes(split)[CodeRegisterClobber] != 1 {
+		t.Fatalf("unsaved R14 in a non-NOSPLIT function should be flagged: %+v", split)
+	}
+
+	// amd64: R14 saved and restored around the call is preserved.
 	saved := lintSrc(t, "#include \"textflag.h\"\n"+
 		"TEXT ·f(SB), NOSPLIT, $8\n"+
-		"\tPUSHQ BX\n"+
+		"\tPUSHQ R14\n"+
-		"\tMOVQ CX, BX\n"+
+		"\tXORL R14, R14\n"+
-		"\tPOPQ BX\n"+
+		"\tCALL ·g(SB)\n"+
+		"\tPOPQ R14\n"+
 		"\tRET\n")
 	if codes(saved)[CodeRegisterClobber] != 0 {
-		t.Fatalf("saved/restored register must not be flagged: %+v", saved)
+		t.Fatalf("saved/restored R14 must not be flagged: %+v", saved)
 	}
 
-	// A caller-saved register (CX) is fine to write.
+	// amd64: BP maintains the frame chain and is always audited.
-	caller := lintSrc(t, "#include \"textflag.h\"\n"+
+	bp := lintSrc(t, "#include \"textflag.h\"\n"+
 		"TEXT ·f(SB), NOSPLIT, $0\n"+
-		"\tMOVQ $1, CX\n"+
+		"\tMOVQ CX, BP\n"+
 		"\tRET\n")
-	if codes(caller)[CodeRegisterClobber] != 0 {
+	if codes(bp)[CodeRegisterClobber] != 1 {
-		t.Fatalf("caller-saved register must not be flagged: %+v", caller)
+		t.Fatalf("unsaved BP write should be flagged: %+v", bp)
+	}
+
+	// arm64: R20 is scratch; R28 (goroutine pointer) and R18 (OS-reserved)
+	// are fixed by the Go ABI.
+	armScratch := lintSrcArch(t, "t_arm64.s", "#include \"textflag.h\"\n"+
+		"TEXT ·f(SB), NOSPLIT, $0\n"+
+		"\tMOVD R0, R20\n"+
+		"\tRET\n")
+	if codes(armScratch)[CodeRegisterClobber] != 0 {
+		t.Fatalf("arm64 scratch register must not be flagged: %+v", armScratch)
+	}
+	armG := lintSrcArch(t, "t_arm64.s", "#include \"textflag.h\"\n"+
+		"TEXT ·f(SB), NOSPLIT, $0\n"+
+		"\tMOVD R0, R28\n"+
+		"\tRET\n")
+	if codes(armG)[CodeRegisterClobber] != 1 {
+		t.Fatalf("unsaved arm64 R28 write should be flagged: %+v", armG)
+	}
+	armReserved := lintSrcArch(t, "t_arm64.s", "#include \"textflag.h\"\n"+
+		"TEXT ·f(SB), NOSPLIT, $0\n"+
+		"\tMOVD R0, R18\n"+
+		"\tRET\n")
+	if codes(armReserved)[CodeRegisterClobber] != 1 {
+		t.Fatalf("arm64 R18 write should be flagged: %+v", armReserved)
+	}
+
+	// riscv64: X27 holds the goroutine; X5–X7 are scratch.
+	riscScratch := lintSrcArch(t, "t_riscv64.s", "#include \"textflag.h\"\n"+
+		"TEXT ·f(SB), NOSPLIT, $0\n"+
+		"\tMOV X5, X6\n"+
+		"\tRET\n")
+	if codes(riscScratch)[CodeRegisterClobber] != 0 {
+		t.Fatalf("riscv64 scratch register must not be flagged: %+v", riscScratch)
+	}
+	riscG := lintSrcArch(t, "t_riscv64.s", "#include \"textflag.h\"\n"+
+		"TEXT ·f(SB), NOSPLIT, $0\n"+
+		"\tMOV X5, X27\n"+
+		"\tRET\n")
+	if codes(riscG)[CodeRegisterClobber] != 1 {
+		t.Fatalf("unsaved riscv64 X27 write should be flagged: %+v", riscG)
+	}
+
+	// loong64: R22 holds the goroutine; R5–R19 are argument/scratch.
+	loongScratch := lintSrcArch(t, "t_loong64.s", "#include \"textflag.h\"\n"+
+		"TEXT ·f(SB), NOSPLIT, $0\n"+
+		"\tMOVV R5, R6\n"+
+		"\tRET\n")
+	if codes(loongScratch)[CodeRegisterClobber] != 0 {
+		t.Fatalf("loong64 scratch register must not be flagged: %+v", loongScratch)
+	}
+	loongG := lintSrcArch(t, "t_loong64.s", "#include \"textflag.h\"\n"+
+		"TEXT ·f(SB), NOSPLIT, $0\n"+
+		"\tMOVV R5, R22\n"+
+		"\tRET\n")
+	if codes(loongG)[CodeRegisterClobber] != 1 {
+		t.Fatalf("unsaved loong64 R22 write should be flagged: %+v", loongG)
 	}
 }
 

parser/parser.go

@@ -375,6 +375,11 @@ func parseImmediate(g []token.Token) ast.Immediate {
 		if v, ok := tryInt(text); ok {
 			imm.Val = v
 			imm.HasVal = true
+		} else if u, err := strconv.ParseUint(text, 0, 64); err == nil && !imm.Neg {
+			// Unsigned 64-bit literals (DATA mask<>+8(SB)/8, $0x8000…)
+			// overflow int64; keep the bit pattern.
+			imm.Val = int64(u)
+			imm.HasVal = true
 		} else {
 			imm.Float = text
 		}

testdata/verify/abi_amd64.s

@@ -0,0 +1,28 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+#include "textflag.h"
+
+// func cleanAdd(a, b int64) int64
+// A well-behaved function that preserves all callee-saved registers.
+TEXT ·cleanAdd(SB), NOSPLIT, $0-24
+	MOVQ a+0(FP), AX
+	ADDQ b+8(FP), AX
+	MOVQ AX, ret+16(FP)
+	RET
+
+// func dirtyBP(a int64) int64
+// Deliberately clobbers BP (an ABI violation for a NOSPLIT frame=0 function).
+TEXT ·dirtyBP(SB), NOSPLIT, $0-16
+	MOVQ $0x1234, BP
+	MOVQ a+0(FP), AX
+	MOVQ AX, ret+8(FP)
+	RET
+
+// func dirtyR14(a int64) int64
+// Deliberately clobbers R14 (the goroutine pointer — a serious ABI violation).
+TEXT ·dirtyR14(SB), NOSPLIT, $0-16
+	MOVQ $0x5678, R14
+	MOVQ a+0(FP), AX
+	MOVQ AX, ret+8(FP)
+	RET

testdata/verify/basic_amd64.s

@@ -0,0 +1,67 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+#include "textflag.h"
+
+// func add(a, b int64) int64
+TEXT ·add(SB), NOSPLIT, $0-24
+	MOVQ a+0(FP), AX
+	ADDQ b+8(FP), AX
+	MOVQ AX, ret+16(FP)
+	RET
+
+// func sum(data []int64) int64
+// Sums all elements of the slice.
+TEXT ·sum(SB), NOSPLIT, $0-32
+	MOVQ data_base+0(FP), SI
+	MOVQ data_len+8(FP), CX
+	XORQ AX, AX
+	TESTQ CX, CX
+	JZ   sum_done
+
+sum_loop:
+	ADDQ (SI), AX
+	ADDQ $8, SI
+	DECQ CX
+	JNZ  sum_loop
+
+sum_done:
+	MOVQ AX, ret+24(FP)
+	RET
+
+// func wideCopy(dst, src []byte)
+// Non-overlapping copy of min(len(dst), len(src)) bytes using 32-byte moves.
+TEXT ·wideCopy(SB), NOSPLIT, $0-48
+	MOVQ dst_base+0(FP), DI
+	MOVQ dst_len+8(FP), BX
+	MOVQ src_base+24(FP), SI
+	MOVQ src_len+32(FP), R8
+	CMPQ BX, R8
+	JLE  wc_have_n
+	MOVQ R8, BX
+
+wc_have_n:
+	CMPQ BX, $32
+	JB   wc_small
+
+	VMOVDQU (SI), Y0
+	VMOVDQU Y0, (DI)
+	VMOVDQU -32(SI)(BX*1), Y0
+	VMOVDQU Y0, -32(DI)(BX*1)
+	VZEROUPPER
+	RET
+
+wc_small:
+	TESTQ BX, BX
+	JZ    wc_done
+
+wc_byte:
+	MOVB (SI), R8B
+	MOVB R8B, (DI)
+	INCQ SI
+	INCQ DI
+	DECQ BX
+	JNZ  wc_byte
+
+wc_done:
+	RET

verify/abi_amd64.go

@@ -0,0 +1,130 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+//go:build amd64
+
+package verify
+
+import (
+	"encoding/binary"
+	"fmt"
+	"syscall"
+	"unsafe"
+)
+
+// abiResult records register-clobber violations detected by the ABI-checking
+// trampoline.  Bit 0: BP clobbered.  Bit 1: R14 clobbered.
+var abiResult uint64
+
+// savedBP holds the caller's frame pointer across the ABI-checked JIT call.
+// Referenced by enterJITChecked to satisfy go vet's save-before-clobber rule.
+var savedBP uintptr
+
+// leaveCheckedPtr is initialised by the linker from the GLOBL/DATA in
+// abi_amd64.s: it holds the raw address of leaveJITCheckedRaw (which has
+// no ABIInternal wrapper, so the JIT function RETs directly into it).
+var leaveCheckedPtr uintptr
+
+// enterJITChecked sets sentinels in BP and R14, switches to the prepared
+// stack and jumps to fn.
+//
+//go:nosplit
+func enterJITChecked(fn uintptr, stack uintptr)
+
+// leaveJITCheckedRaw is the raw return trampoline for ABI checks.  Its
+// address is obtained from the GLOBL in abi_amd64.s (leaveCheckedPtr),
+// which points to the .abi0 code — NOT the ABIInternal wrapper that this
+// declaration would generate.  The declaration exists solely to satisfy
+// go vet's "missing Go declaration" check.
+//
+//go:nosplit
+func leaveJITCheckedRaw()
+
+// ABIReport describes the result of an ABI-checking call.
+type ABIReport struct {
+	BPClobbered  bool // BP was modified by the function
+	R14Clobbered bool // R14 (goroutine pointer) was modified
+	RedZoneHit   bool // the 128-byte red zone below SP was written
+}
+
+// OK returns true when no violations were detected.
+func (r ABIReport) OK() bool {
+	return !r.BPClobbered && !r.R14Clobbered && !r.RedZoneHit
+}
+
+// String returns a human-readable summary.
+func (r ABIReport) String() string {
+	if r.OK() {
+		return "ABI clean"
+	}
+	s := "ABI violation:"
+	if r.BPClobbered {
+		s += " BP clobbered"
+	}
+	if r.R14Clobbered {
+		s += " R14 clobbered"
+	}
+	if r.RedZoneHit {
+		s += " red-zone written"
+	}
+	return s
+}
+
+// redZoneSize is the System V AMD64 red zone: 128 bytes below SP that a
+// leaf function may use without adjusting SP.  Go does not use the red zone,
+// so any write there is a bug.
+const redZoneSize = 128
+
+// redZoneFill is the byte pattern used to detect red-zone writes.
+const redZoneFill = 0xA5
+
+// CallChecked invokes the function with ABI sentinels and a red-zone
+// canary, returning both the argument block (with results) and an ABIReport.
+func CallChecked(fnAddr uintptr, args []byte) ([]byte, ABIReport, error) {
+	report := ABIReport{}
+
+	// Reset the global result.
+	abiResult = 0
+
+	// Prepare the stack: [red-zone canary][padding][leaveJITCheckedRaw][args...]
+	// The red zone sits below the initial SP, so the function would have to
+	// write below SP to corrupt it.
+	totalSize := redZoneSize + stackPad + 8 + len(args) + 64
+	stackMem, err := syscall.Mmap(-1, 0, totalSize,
+		syscall.PROT_READ|syscall.PROT_WRITE, syscall.MAP_PRIVATE|syscall.MAP_ANON)
+	if err != nil {
+		return nil, report, fmt.Errorf("verify: stack mmap: %w", err)
+	}
+	defer syscall.Munmap(stackMem)
+
+	// Fill the red zone with the canary pattern.
+	for i := 0; i < redZoneSize; i++ {
+		stackMem[i] = redZoneFill
+	}
+
+	// Return address and args after the red zone and padding.
+	retOff := redZoneSize + stackPad
+	binary.LittleEndian.PutUint64(stackMem[retOff:retOff+8], uint64(leaveCheckedPtr))
+	copy(stackMem[retOff+8:], args)
+
+	stackBase := uintptr(unsafe.Pointer(&stackMem[retOff]))
+	enterJITChecked(fnAddr, stackBase)
+
+	// Read the register-clobber result.
+	res := abiResult
+	report.BPClobbered = res&1 != 0
+	report.R14Clobbered = res&2 != 0
+
+	// Check the red zone.
+	for i := 0; i < redZoneSize; i++ {
+		if stackMem[i] != redZoneFill {
+			report.RedZoneHit = true
+			break
+		}
+	}
+
+	// Copy out the argument area.
+	out := make([]byte, len(args))
+	copy(out, stackMem[retOff+8:retOff+8+len(args)])
+	return out, report, nil
+}

verify/abi_amd64.s

@@ -0,0 +1,61 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+#include "textflag.h"
+
+// ABI-checking trampoline.  Sets sentinel values in the callee-saved
+// registers (BP, R14) before entering the JIT function and checks whether
+// they survived on return.
+//
+// The return trampoline (leaveJITCheckedRaw) is a raw TEXT symbol with no
+// Go function declaration, so the toolchain does NOT interpose an
+// ABIInternal wrapper — the JIT function RETs directly into the check code,
+// which sees the registers exactly as the function left them.
+//
+// Go ABI0 on amd64 guarantees:
+//   - BP is callee-saved (NOSPLIT frame=0 functions must not touch it).
+//   - R14 holds the goroutine pointer and must survive across any call.
+
+// Sentinel values chosen to be unlikely in normal execution.
+#define SENTINEL_BP  0xDEADBEEFCAFEF00D
+#define SENTINEL_R14 0x0BADF00DDEADBEEF
+
+// GLOBL holding the raw address of the leave trampoline, read by Go.
+GLOBL ·leaveCheckedPtr(SB), NOPTR, $8
+DATA  ·leaveCheckedPtr(SB)/8, $·leaveJITCheckedRaw(SB)
+
+// func enterJITChecked(fn uintptr, stack uintptr)
+// Sets sentinels in BP and R14, switches to the prepared stack and jumps
+// to fn.  The prepared stack's return address must be leaveJITCheckedRaw
+// (read from leaveCheckedPtr).
+TEXT ·enterJITChecked(SB), NOSPLIT, $0-16
+	MOVQ fn+0(FP), AX         // target (before SP switch)
+	MOVQ SP, ·savedSP(SB)     // preserve Go stack
+	MOVQ BP, ·savedBP(SB)     // preserve frame pointer (vet requires save before clobber)
+	MOVQ $SENTINEL_BP, BP     // sentinel in BP
+	MOVQ $SENTINEL_R14, R14   // sentinel in R14
+	MOVQ stack+8(FP), SP      // switch to prepared stack
+	JMP  AX
+
+// leaveJITCheckedRaw is the raw return trampoline.  It has NO Go function
+// declaration, so no ABIInternal wrapper is generated — the JIT function's
+// RET lands here directly, seeing BP and R14 exactly as the function left
+// them.  It checks the sentinels, records violations in abiResult, then
+// restores the Go stack and returns.
+TEXT ·leaveJITCheckedRaw(SB), NOSPLIT, $0-0
+	// Check BP against the sentinel.
+	MOVQ $SENTINEL_BP, CX
+	CMPQ BP, CX
+	JEQ  bp_ok
+	ORQ  $1, ·abiResult(SB)
+
+bp_ok:
+	// Check R14 against the sentinel.
+	MOVQ $SENTINEL_R14, CX
+	CMPQ R14, CX
+	JEQ  r14_ok
+	ORQ  $2, ·abiResult(SB)
+
+r14_ok:
+	MOVQ ·savedSP(SB), SP
+	RET

verify/abi_other.go

@@ -0,0 +1,26 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+//go:build !amd64
+
+package verify
+
+import "fmt"
+
+// ABIReport describes the result of an ABI-checking call.
+type ABIReport struct {
+	BPClobbered  bool
+	R14Clobbered bool
+	RedZoneHit   bool
+}
+
+// OK returns true when no violations were detected.
+func (r ABIReport) OK() bool { return false }
+
+// String returns a human-readable summary.
+func (r ABIReport) String() string { return "verify: ABI checks require amd64" }
+
+// CallChecked is unavailable on non-amd64 architectures.
+func CallChecked(fnAddr uintptr, args []byte) ([]byte, ABIReport, error) {
+	return nil, ABIReport{}, fmt.Errorf("verify: ABI checks require amd64")
+}

verify/abi_test.go

@@ -0,0 +1,145 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package verify
+
+import (
+	"testing"
+	"unsafe"
+)
+
+func loadABIKernel(t *testing.T) *Kernel {
+	t.Helper()
+	k, err := Load("../testdata/verify/abi_amd64.s")
+	if err != nil {
+		t.Fatalf("Load: %v", err)
+	}
+	t.Cleanup(k.Close)
+	return k
+}
+
+func TestABIClean(t *testing.T) {
+	k := loadABIKernel(t)
+
+	args := make([]byte, 24)
+	PutUint64(args, 0, 3)
+	PutUint64(args, 8, 4)
+
+	out, report, err := k.CallFuncChecked("cleanAdd", args)
+	if err != nil {
+		t.Fatalf("CallFuncChecked: %v", err)
+	}
+	if got := int64(GetUint64(out, 16)); got != 7 {
+		t.Errorf("cleanAdd(3, 4) = %d, want 7", got)
+	}
+	if !report.OK() {
+		t.Errorf("cleanAdd: %s", report)
+	}
+}
+
+func TestABIBPClobbered(t *testing.T) {
+	k := loadABIKernel(t)
+
+	args := make([]byte, 16)
+	PutUint64(args, 0, 42)
+
+	out, report, err := k.CallFuncChecked("dirtyBP", args)
+	if err != nil {
+		t.Fatalf("CallFuncChecked: %v", err)
+	}
+	if got := int64(GetUint64(out, 8)); got != 42 {
+		t.Errorf("dirtyBP(42) = %d, want 42", got)
+	}
+	if !report.BPClobbered {
+		t.Error("dirtyBP: expected BP clobbered, but report says clean")
+	}
+	if report.R14Clobbered {
+		t.Error("dirtyBP: R14 should not be clobbered")
+	}
+}
+
+func TestABIR14Clobbered(t *testing.T) {
+	k := loadABIKernel(t)
+
+	args := make([]byte, 16)
+	PutUint64(args, 0, 99)
+
+	out, report, err := k.CallFuncChecked("dirtyR14", args)
+	if err != nil {
+		t.Fatalf("CallFuncChecked: %v", err)
+	}
+	if got := int64(GetUint64(out, 8)); got != 99 {
+		t.Errorf("dirtyR14(99) = %d, want 99", got)
+	}
+	if !report.R14Clobbered {
+		t.Error("dirtyR14: expected R14 clobbered, but report says clean")
+	}
+	if report.BPClobbered {
+		t.Error("dirtyR14: BP should not be clobbered")
+	}
+}
+
+// TestABILZ4Kernels verifies that the production go-lz4 kernels are ABI-clean:
+// they preserve BP and R14 and do not write into the red zone.
+func TestABILZ4Kernels(t *testing.T) {
+	k := loadLZ4Kernel(t)
+
+	// wideCopyAVX2 with a real copy.
+	src := make([]byte, 128)
+	for i := range src {
+		src[i] = byte(i)
+	}
+	dst := make([]byte, 128)
+
+	args := make([]byte, 48)
+	PutPtr(args, 0, unsafe.Pointer(&dst[0]))
+	PutUint64(args, 8, 128)
+	PutUint64(args, 16, 128)
+	PutPtr(args, 24, unsafe.Pointer(&src[0]))
+	PutUint64(args, 32, 128)
+	PutUint64(args, 40, 128)
+
+	_, report, err := k.CallFuncChecked("wideCopyAVX2", args)
+	if err != nil {
+		t.Fatalf("CallFuncChecked(wideCopyAVX2): %v", err)
+	}
+	if !report.OK() {
+		t.Errorf("wideCopyAVX2: %s", report)
+	}
+
+	// decodeBlockAVX2 with a simple block.
+	decSrc := []byte{0x50, 'H', 'e', 'l', 'l', 'o'}
+	decDst := make([]byte, 64)
+
+	decArgs := make([]byte, 64)
+	PutPtr(decArgs, 0, unsafe.Pointer(&decSrc[0]))
+	PutUint64(decArgs, 8, uint64(len(decSrc)))
+	PutUint64(decArgs, 16, uint64(cap(decSrc)))
+	PutPtr(decArgs, 24, unsafe.Pointer(&decDst[0]))
+	PutUint64(decArgs, 32, uint64(len(decDst)))
+	PutUint64(decArgs, 40, uint64(cap(decDst)))
+
+	_, report, err = k.CallFuncChecked("decodeBlockAVX2", decArgs)
+	if err != nil {
+		t.Fatalf("CallFuncChecked(decodeBlockAVX2): %v", err)
+	}
+	if !report.OK() {
+		t.Errorf("decodeBlockAVX2: %s", report)
+	}
+}
+
+func TestCallFuncCheckedErrors(t *testing.T) {
+	k := loadABIKernel(t)
+
+	// Nonexistent function.
+	_, _, err := k.CallFuncChecked("nope", make([]byte, 8))
+	if err == nil {
+		t.Fatal("expected error for nonexistent function")
+	}
+
+	// Arg block too small.
+	_, _, err = k.CallFuncChecked("cleanAdd", make([]byte, 8))
+	if err == nil {
+		t.Fatal("expected error for too-small arg block")
+	}
+}

verify/avx512_test.go

@@ -0,0 +1,144 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package verify
+
+import (
+	"bytes"
+	"math/rand"
+	"os"
+	"testing"
+	"unsafe"
+)
+
+const lz4AVX512Path = "../../go-libraries/go-lz4/avx512_amd64.s"
+
+func loadLZ4AVX512Kernel(t *testing.T) *Kernel {
+	t.Helper()
+	if _, err := os.Stat(lz4AVX512Path); err != nil {
+		t.Skipf("sibling kernel not available: %v", err)
+	}
+	k, err := Load(lz4AVX512Path)
+	if err != nil {
+		t.Fatalf("Load(%s): %v", lz4AVX512Path, err)
+	}
+	t.Cleanup(k.Close)
+	return k
+}
+
+func TestAVX512DecodeKnownAnswers(t *testing.T) {
+	k := loadLZ4AVX512Kernel(t)
+
+	tests := []struct {
+		name     string
+		src      []byte
+		wantN    int
+		wantCode int
+	}{
+		{"literals_only", []byte{0x50, 'H', 'e', 'l', 'l', 'o'}, 5, 0},
+		{"literals_and_match", []byte{0x54, 'A', 'A', 'A', 'A', 'A', 0x05, 0x00, 0x30, 'B', 'B', 'B'}, 16, 0},
+		{"overlapping", []byte{0x14, 'X', 0x01, 0x00, 0x10, 'Y'}, 10, 0},
+		{"malformed", []byte{0x50, 'H', 'e'}, 0, 1},
+		{"zero_offset", []byte{0x14, 'X', 0x00, 0x00}, 0, 2},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			dst := make([]byte, 64)
+			args := make([]byte, 64)
+			PutPtr(args, 0, unsafe.Pointer(&tt.src[0]))
+			PutUint64(args, 8, uint64(len(tt.src)))
+			PutUint64(args, 16, uint64(cap(tt.src)))
+			PutPtr(args, 24, unsafe.Pointer(&dst[0]))
+			PutUint64(args, 32, uint64(len(dst)))
+			PutUint64(args, 40, uint64(cap(dst)))
+
+			out, err := k.CallFunc("decodeBlockAVX512", args)
+			if err != nil {
+				t.Fatalf("CallFunc: %v", err)
+			}
+			n := int(GetUint64(out, 48))
+			code := int(GetUint64(out, 56))
+			if n != tt.wantN || code != tt.wantCode {
+				t.Errorf("got (n=%d, code=%d), want (n=%d, code=%d)", n, code, tt.wantN, tt.wantCode)
+			}
+		})
+	}
+}
+
+func TestAVX512DifferentialFuzz(t *testing.T) {
+	k := loadLZ4AVX512Kernel(t)
+	rng := rand.New(rand.NewSource(77))
+
+	for i := 0; i < 3000; i++ {
+		wantSize := 1 + rng.Intn(4096)
+		src := genLZ4Block(rng, wantSize)
+		dstSize := wantSize + 64
+
+		goDst := make([]byte, dstSize)
+		goN, goCode := decodeBlockGo(src, goDst)
+
+		jitDst := make([]byte, dstSize)
+		args := make([]byte, 64)
+		if len(src) > 0 {
+			PutPtr(args, 0, unsafe.Pointer(&src[0]))
+		}
+		PutUint64(args, 8, uint64(len(src)))
+		PutUint64(args, 16, uint64(cap(src)))
+		if dstSize > 0 {
+			PutPtr(args, 24, unsafe.Pointer(&jitDst[0]))
+		}
+		PutUint64(args, 32, uint64(dstSize))
+		PutUint64(args, 40, uint64(cap(jitDst)))
+
+		out, err := k.CallFunc("decodeBlockAVX512", args)
+		if err != nil {
+			t.Fatalf("iter %d: %v", i, err)
+		}
+		jitN := int(GetUint64(out, 48))
+		jitCode := int(GetUint64(out, 56))
+
+		if jitCode != goCode {
+			t.Fatalf("iter %d: code mismatch: JIT=%d Go=%d", i, jitCode, goCode)
+		}
+		if jitCode != 0 {
+			continue
+		}
+		if jitN != goN {
+			t.Fatalf("iter %d: n mismatch: JIT=%d Go=%d", i, jitN, goN)
+		}
+		if !bytes.Equal(jitDst[:jitN], goDst[:goN]) {
+			t.Fatalf("iter %d: output mismatch (n=%d)", i, jitN)
+		}
+	}
+}
+
+func TestAVX512WideCopy(t *testing.T) {
+	k := loadLZ4AVX512Kernel(t)
+
+	sizes := []int{0, 1, 31, 32, 63, 64, 65, 127, 128, 256, 1024}
+	for _, n := range sizes {
+		src := make([]byte, n)
+		for i := range src {
+			src[i] = byte(i*11 + 3)
+		}
+		dst := make([]byte, n)
+
+		args := make([]byte, 48)
+		if n > 0 {
+			PutPtr(args, 0, unsafe.Pointer(&dst[0]))
+			PutPtr(args, 24, unsafe.Pointer(&src[0]))
+		}
+		PutUint64(args, 8, uint64(n))
+		PutUint64(args, 16, uint64(n))
+		PutUint64(args, 32, uint64(n))
+		PutUint64(args, 40, uint64(n))
+
+		_, err := k.CallFunc("wideCopyAVX512", args)
+		if err != nil {
+			t.Fatalf("wideCopyAVX512(n=%d): %v", n, err)
+		}
+		if !bytes.Equal(dst, src) {
+			t.Errorf("wideCopyAVX512(n=%d): mismatch", n)
+		}
+	}
+}

verify/call_amd64.go

@@ -0,0 +1,77 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+//go:build amd64
+
+package verify
+
+import (
+	"encoding/binary"
+	"fmt"
+	"reflect"
+	"syscall"
+	"unsafe"
+)
+
+// savedSP holds the Go stack pointer while a JIT call is in flight.
+// Referenced by the assembly trampoline (trampoline_amd64.s).
+var savedSP uintptr
+
+// enterJIT switches to the prepared stack and jumps to fn.
+// It does not return normally; the JIT function's RET transfers control
+// to leaveJIT, which restores the Go stack.
+//
+//go:nosplit
+func enterJIT(fn uintptr, stack uintptr)
+
+// leaveJIT restores the Go stack after a JIT function returns.
+// Its address is placed as the return address on the prepared stack.
+//
+//go:nosplit
+func leaveJIT()
+
+// leaveJITAddr is the machine address of leaveJIT, resolved once at init.
+var leaveJITAddr uintptr
+
+func init() {
+	leaveJITAddr = reflect.ValueOf(leaveJIT).Pointer()
+}
+
+// stackPad is padding below the return address on the prepared stack.
+// The ABIInternal wrapper that leaveJIT's address resolves to executes
+// PUSHQ BP and CALL before reaching the raw assembly, writing up to 16
+// bytes below the return-address slot.  64 bytes of headroom is ample.
+const stackPad = 64
+
+// Call invokes the assembled function at fnAddr with the given ABI0 argument
+// block (the raw bytes that would appear at FP+0).  It returns the argument
+// block after the call, which contains any results the function wrote back
+// (the ABI0 convention shares the argument area for inputs and outputs).
+//
+// The function must be NOSPLIT (no stack growth) and must not reference
+// external symbols — the image is self-contained.
+func Call(fnAddr uintptr, args []byte) ([]byte, error) {
+	// Prepare the stack: [padding][leaveJIT addr][args...]
+	stackSize := stackPad + 8 + len(args) + 64 // padding + ret + args + safety
+	stackMem, err := syscall.Mmap(-1, 0, stackSize,
+		syscall.PROT_READ|syscall.PROT_WRITE, syscall.MAP_PRIVATE|syscall.MAP_ANON)
+	if err != nil {
+		return nil, fmt.Errorf("verify: stack mmap: %w", err)
+	}
+	defer syscall.Munmap(stackMem)
+
+	// The return address sits after the padding; the function's SP will
+	// point here, leaving stackPad bytes below for the wrapper's pushes.
+	retOff := stackPad
+	binary.LittleEndian.PutUint64(stackMem[retOff:retOff+8], uint64(leaveJITAddr))
+	// The ABI0 argument area follows the return address.
+	copy(stackMem[retOff+8:], args)
+
+	stackBase := uintptr(unsafe.Pointer(&stackMem[retOff]))
+	enterJIT(fnAddr, stackBase)
+
+	// Copy out the (possibly modified) argument area.
+	out := make([]byte, len(args))
+	copy(out, stackMem[retOff+8:retOff+8+len(args)])
+	return out, nil
+}

verify/call_other.go

@@ -0,0 +1,13 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+//go:build !amd64
+
+package verify
+
+import "fmt"
+
+// Call is unavailable on non-amd64 architectures.
+func Call(fnAddr uintptr, args []byte) ([]byte, error) {
+	return nil, fmt.Errorf("verify: JIT execution requires amd64")
+}

verify/coverage.go

@@ -0,0 +1,103 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package verify
+
+import (
+	"fmt"
+	"sort"
+)
+
+// Block describes one basic block within a function: a maximal sequence of
+// instructions with a single entry point (a label or the function start) and
+// a single exit (a jump, conditional jump or RET).
+type Block struct {
+	Offset int    // byte offset within the function
+	Label  string // label name ("" for the entry block)
+}
+
+// Blocks identifies the basic blocks of a function from its local labels.
+// Each label is a potential jump target and therefore a block boundary; the
+// function entry (offset 0) is always a block.  The blocks are returned in
+// ascending offset order.
+func (k *Kernel) Blocks(name string) ([]Block, error) {
+	idx, ok := k.funcs[name]
+	if !ok {
+		return nil, fmt.Errorf("verify: function %q not found", name)
+	}
+	fl := k.img.Funcs[idx]
+
+	blocks := []Block{{Offset: 0, Label: "(entry)"}}
+	// Build a reverse map: offset → label name.
+	offToLabel := make(map[int]string, len(fl.Labels))
+	for label, off := range fl.Labels {
+		if off > 0 && off < fl.Size {
+			offToLabel[off] = label
+		}
+	}
+	// Collect and sort offsets.
+	offsets := make([]int, 0, len(offToLabel))
+	for off := range offToLabel {
+		offsets = append(offsets, off)
+	}
+	sort.Ints(offsets)
+	for _, off := range offsets {
+		blocks = append(blocks, Block{Offset: off, Label: offToLabel[off]})
+	}
+	return blocks, nil
+}
+
+// BlockCount returns the number of identified basic blocks for the function.
+func (k *Kernel) BlockCount(name string) (int, error) {
+	blocks, err := k.Blocks(name)
+	if err != nil {
+		return 0, err
+	}
+	return len(blocks), nil
+}
+
+// PathFingerprint is the observable output of one function execution: the
+// values written back into the result slots of the argument block.  Two
+// executions that produce the same fingerprint took observationally
+// equivalent paths (though they may differ internally).
+type PathFingerprint struct {
+	Results []uint64 // the result words from the arg block
+}
+
+// ProfilePaths runs the function with each of the given argument blocks and
+// collects the distinct output fingerprints.  This measures path diversity:
+// how many observationally different execution paths the input corpus
+// exercises.  Combined with Blocks (the static block count), it gives a
+// lower bound on code coverage.
+func (k *Kernel) ProfilePaths(name string, argSets [][]byte, resultOffsets []int) ([]PathFingerprint, error) {
+	idx, ok := k.funcs[name]
+	if !ok {
+		return nil, fmt.Errorf("verify: function %q not found", name)
+	}
+	fl := k.img.Funcs[idx]
+
+	seen := map[string]bool{}
+	var paths []PathFingerprint
+
+	for _, args := range argSets {
+		if len(args) < fl.Args {
+			return nil, fmt.Errorf("verify: %s: arg block too small", name)
+		}
+		out, err := k.CallFunc(name, args)
+		if err != nil {
+			return nil, err
+		}
+		fp := PathFingerprint{}
+		key := ""
+		for _, off := range resultOffsets {
+			v := GetUint64(out, off)
+			fp.Results = append(fp.Results, v)
+			key += fmt.Sprintf("%016x", v)
+		}
+		if !seen[key] {
+			seen[key] = true
+			paths = append(paths, fp)
+		}
+	}
+	return paths, nil
+}

verify/coverage_test.go

@@ -0,0 +1,85 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package verify
+
+import (
+	"testing"
+	"unsafe"
+)
+
+func TestBlocks(t *testing.T) {
+	k := loadBasic(t)
+
+	// The "sum" function has labels: sum_done, sum_loop.
+	blocks, err := k.Blocks("sum")
+	if err != nil {
+		t.Fatalf("Blocks(sum): %v", err)
+	}
+	if len(blocks) < 3 {
+		t.Errorf("sum: expected at least 3 blocks (entry + 2 labels), got %d", len(blocks))
+	}
+	if blocks[0].Offset != 0 {
+		t.Errorf("first block offset = %d, want 0", blocks[0].Offset)
+	}
+	t.Logf("sum blocks: %v", blocks)
+}
+
+func TestBlockCount(t *testing.T) {
+	k := loadLZ4Kernel(t)
+
+	n, err := k.BlockCount("decodeBlockAVX2")
+	if err != nil {
+		t.Fatalf("BlockCount: %v", err)
+	}
+	// The decoder has many labels (dec_loop, dec_malformed, etc.).
+	if n < 10 {
+		t.Errorf("decodeBlockAVX2: expected at least 10 blocks, got %d", n)
+	}
+	t.Logf("decodeBlockAVX2: %d basic blocks", n)
+}
+
+func TestProfilePaths(t *testing.T) {
+	k := loadLZ4Kernel(t)
+
+	// Build a corpus of varied LZ4 blocks.
+	var argSets [][]byte
+	blocks := []struct {
+		src     []byte
+		dstSize int
+	}{
+		{[]byte{0x00}, 16},                          // empty
+		{[]byte{0x50, 'H', 'e', 'l', 'l', 'o'}, 16}, // literals only
+		{[]byte{0x54, 'A', 'A', 'A', 'A', 'A', 5, 0, 0x30, 'B', 'B', 'B'}, 32}, // match
+		{[]byte{0x14, 'X', 1, 0, 0x10, 'Y'}, 16},                               // overlapping
+		{[]byte{0x50, 'H'}, 16},                                                // malformed
+		{[]byte{0x14, 'X', 0, 0}, 16},                                          // zero offset
+	}
+	for _, b := range blocks {
+		args := make([]byte, 64)
+		if len(b.src) > 0 {
+			PutPtr(args, 0, unsafe.Pointer(&b.src[0]))
+		}
+		PutUint64(args, 8, uint64(len(b.src)))
+		PutUint64(args, 16, uint64(cap(b.src)))
+		dst := make([]byte, b.dstSize)
+		if len(dst) > 0 {
+			PutPtr(args, 24, unsafe.Pointer(&dst[0]))
+		}
+		PutUint64(args, 32, uint64(len(dst)))
+		PutUint64(args, 40, uint64(cap(dst)))
+		argSets = append(argSets, args)
+	}
+
+	// Result offsets: n+48 and code+56.
+	paths, err := k.ProfilePaths("decodeBlockAVX2", argSets, []int{48, 56})
+	if err != nil {
+		t.Fatalf("ProfilePaths: %v", err)
+	}
+
+	// We expect at least 3 distinct paths: success (various n), malformed, zero offset.
+	if len(paths) < 3 {
+		t.Errorf("expected at least 3 distinct paths, got %d", len(paths))
+	}
+	t.Logf("decodeBlockAVX2: %d distinct output paths from %d inputs", len(paths), len(argSets))
+}

verify/differential_test.go

@@ -0,0 +1,295 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package verify
+
+import (
+	"bytes"
+	"math/rand"
+	"testing"
+	"unsafe"
+)
+
+// decodeBlockGo is a minimal portable LZ4 block decoder used as the
+// differential-testing oracle.  It mirrors the contract of
+// go-lz4's decodeBlockGo: (bytesWritten, code) where code is
+// 0 = ok, 1 = malformed, 2 = zero offset.
+func decodeBlockGo(src, dst []byte) (int, int) {
+	if len(src) == 0 {
+		return 0, 1
+	}
+	si, di := 0, 0
+	for {
+		if si >= len(src) {
+			return 0, 1 // truncated: no token
+		}
+		token := int(src[si])
+		si++
+
+		// Literals.
+		lLen := token >> 4
+		if lLen == 15 {
+			for {
+				if si >= len(src) {
+					return 0, 1
+				}
+				b := int(src[si])
+				si++
+				lLen += b
+				if b != 255 {
+					break
+				}
+			}
+		}
+		if si+lLen > len(src) {
+			return 0, 1 // truncated literals
+		}
+		if di+lLen > len(dst) {
+			return 0, 1 // destination overflow
+		}
+		copy(dst[di:di+lLen], src[si:si+lLen])
+		di += lLen
+		si += lLen
+
+		// End of block.
+		if si >= len(src) {
+			return di, 0
+		}
+
+		// Match offset.
+		if si+2 > len(src) {
+			return 0, 1
+		}
+		offset := int(src[si]) | int(src[si+1])<<8
+		si += 2
+		if offset == 0 {
+			return 0, 2
+		}
+
+		// Match length.
+		mLen := token & 15
+		if mLen == 15 {
+			for {
+				if si >= len(src) {
+					return 0, 1
+				}
+				b := int(src[si])
+				si++
+				mLen += b
+				if b != 255 {
+					break
+				}
+			}
+		}
+		mLen += 4
+
+		// Copy match (overlapping-safe).
+		if di-offset < 0 {
+			return 0, 1 // offset reaches before dst start
+		}
+		if di+mLen > len(dst) {
+			return 0, 1 // destination overflow
+		}
+		for i := 0; i < mLen; i++ {
+			dst[di+i] = dst[di-offset+i]
+		}
+		di += mLen
+	}
+}
+
+// genLZ4Block generates a random valid LZ4 block that decompresses into
+// approximately wantSize bytes.  The block is always well-formed (ends with
+// a literals-only sequence).
+func genLZ4Block(rng *rand.Rand, wantSize int) []byte {
+	var block []byte
+	produced := 0
+	for produced < wantSize {
+		remaining := wantSize - produced
+
+		// Decide: emit a literals+match sequence or the final literals.
+		if remaining <= 8 || rng.Intn(4) == 0 {
+			// Final literals-only sequence.
+			lLen := remaining
+			if lLen > 60 {
+				lLen = 1 + rng.Intn(60)
+			}
+			block = appendToken(block, lLen, 0)
+			for i := 0; i < lLen; i++ {
+				block = append(block, byte(rng.Intn(256)))
+			}
+			produced += lLen
+			break
+		}
+
+		// Literals + match.
+		lLen := rng.Intn(min(16, remaining))
+		if produced+lLen == 0 {
+			lLen = 1 // must have at least 1 literal before the first match
+		}
+		mLenRaw := rng.Intn(12) // match length = mLenRaw + 4
+		mLen := mLenRaw + 4
+		if produced+mLen > remaining {
+			mLen = remaining - produced
+			if mLen < 4 {
+				// Not enough room for a match; emit final literals.
+				lLen = remaining
+				block = appendToken(block, lLen, 0)
+				for i := 0; i < lLen; i++ {
+					block = append(block, byte(rng.Intn(256)))
+				}
+				break
+			}
+			mLenRaw = mLen - 4
+		}
+
+		block = appendToken(block, lLen, mLenRaw)
+		for i := 0; i < lLen; i++ {
+			block = append(block, byte(rng.Intn(256)))
+		}
+		produced += lLen
+
+		// Offset: must be <= produced (can't reference before start).
+		maxOff := produced
+		if maxOff > 65535 {
+			maxOff = 65535
+		}
+		offset := 1 + rng.Intn(maxOff)
+		block = append(block, byte(offset), byte(offset>>8))
+		produced += mLen
+	}
+	return block
+}
+
+// appendToken appends a token (and extension bytes if needed) for the given
+// literal and match lengths.
+func appendToken(block []byte, lLen, mLenRaw int) []byte {
+	lit4 := lLen
+	if lit4 > 15 {
+		lit4 = 15
+	}
+	ml4 := mLenRaw
+	if ml4 > 15 {
+		ml4 = 15
+	}
+	block = append(block, byte(lit4<<4|ml4))
+	// Literal extension bytes.
+	rem := lLen - 15
+	for rem >= 255 {
+		block = append(block, 255)
+		rem -= 255
+	}
+	if lLen >= 15 {
+		block = append(block, byte(rem))
+	}
+	// Match extension bytes.
+	rem = mLenRaw - 15
+	for rem >= 255 {
+		block = append(block, 255)
+		rem -= 255
+	}
+	if mLenRaw >= 15 {
+		block = append(block, byte(rem))
+	}
+	return block
+}
+
+func min(a, b int) int {
+	if a < b {
+		return a
+	}
+	return b
+}
+
+// TestDifferentialLZ4Fuzz drives the JIT-assembled decodeBlockAVX2 with
+// random valid LZ4 blocks and compares the output bit-for-bit against the
+// portable Go reference.
+func TestDifferentialLZ4Fuzz(t *testing.T) {
+	k := loadLZ4Kernel(t)
+
+	const iterations = 5000
+	rng := rand.New(rand.NewSource(42))
+
+	for i := 0; i < iterations; i++ {
+		wantSize := 1 + rng.Intn(4096)
+		src := genLZ4Block(rng, wantSize)
+		dstSize := wantSize + 64 // generous destination
+
+		// Go reference.
+		goDst := make([]byte, dstSize)
+		goN, goCode := decodeBlockGo(src, goDst)
+
+		// JIT kernel.
+		jitDst := make([]byte, dstSize)
+		jitN, jitCode := callDecodeBlockAVX2(t, k, src, jitDst)
+
+		if jitCode != goCode {
+			t.Fatalf("iter %d: code mismatch: JIT=%d, Go=%d (src len=%d)",
+				i, jitCode, goCode, len(src))
+		}
+		if jitCode != 0 {
+			continue // both agree it's malformed/zero-offset
+		}
+		if jitN != goN {
+			t.Fatalf("iter %d: n mismatch: JIT=%d, Go=%d (src len=%d)",
+				i, jitN, goN, len(src))
+		}
+		if !bytes.Equal(jitDst[:jitN], goDst[:goN]) {
+			t.Fatalf("iter %d: output mismatch (n=%d, src len=%d)", i, jitN, len(src))
+		}
+	}
+}
+
+// TestDifferentialLZ4Hostile drives the kernel with random garbage to check
+// that error codes agree with the Go reference (no crashes, same classification).
+func TestDifferentialLZ4Hostile(t *testing.T) {
+	k := loadLZ4Kernel(t)
+
+	const iterations = 2000
+	rng := rand.New(rand.NewSource(99))
+
+	for i := 0; i < iterations; i++ {
+		srcLen := rng.Intn(128)
+		src := make([]byte, srcLen)
+		rng.Read(src)
+		dstSize := rng.Intn(512)
+		dst := make([]byte, dstSize)
+
+		// Go reference.
+		goDst := make([]byte, dstSize)
+		copy(goDst, dst)
+		_, goCode := decodeBlockGo(src, goDst)
+
+		// JIT kernel.
+		jitDst := make([]byte, dstSize)
+		copy(jitDst, dst)
+		_, jitCode := callDecodeBlockAVX2(t, k, src, jitDst)
+
+		if jitCode != goCode {
+			t.Fatalf("iter %d: hostile code mismatch: JIT=%d, Go=%d (srcLen=%d, dstSize=%d)",
+				i, jitCode, goCode, srcLen, dstSize)
+		}
+	}
+}
+
+// callDecodeBlockAVX2Raw is like callDecodeBlockAVX2 but accepts explicit
+// dst size (for hostile tests where dst may be smaller than the output).
+func callDecodeBlockAVX2Raw(t *testing.T, k *Kernel, src, dst []byte) (int, int) {
+	t.Helper()
+	args := make([]byte, 64)
+	if len(src) > 0 {
+		PutPtr(args, 0, unsafe.Pointer(&src[0]))
+	}
+	PutUint64(args, 8, uint64(len(src)))
+	PutUint64(args, 16, uint64(cap(src)))
+	if len(dst) > 0 {
+		PutPtr(args, 24, unsafe.Pointer(&dst[0]))
+	}
+	PutUint64(args, 32, uint64(len(dst)))
+	PutUint64(args, 40, uint64(cap(dst)))
+
+	out, err := k.CallFunc("decodeBlockAVX2", args)
+	if err != nil {
+		t.Fatalf("CallFunc(decodeBlockAVX2): %v", err)
+	}
+	return int(GetUint64(out, 48)), int(GetUint64(out, 56))
+}

verify/flac_test.go

@@ -0,0 +1,330 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package verify
+
+import (
+	"bytes"
+	"math/rand"
+	"os"
+	"testing"
+	"unsafe"
+)
+
+const flacKernelPath = "../../go-libraries/go-flac/avx2_amd64.s"
+
+func loadFLACKernel(t *testing.T) *Kernel {
+	t.Helper()
+	if _, err := os.Stat(flacKernelPath); err != nil {
+		t.Skipf("sibling kernel not available: %v", err)
+	}
+	k, err := Load(flacKernelPath)
+	if err != nil {
+		t.Fatalf("Load(%s): %v", flacKernelPath, err)
+	}
+	t.Cleanup(k.Close)
+	return k
+}
+
+// --- Portable Go references (from go-flac/simd.go) ---
+
+func decodeMono16Go(src []byte, dst []int32) {
+	for i := 0; i < len(dst); i++ {
+		dst[i] = int32(int16(uint16(src[2*i]) | uint16(src[2*i+1])<<8))
+	}
+}
+
+func pack16Go(dst []byte, src []int32) {
+	for i, v := range src {
+		dst[2*i] = byte(v)
+		dst[2*i+1] = byte(v >> 8)
+	}
+}
+
+func decorrelateLeftSideGo(left, right, out []int32) {
+	for i := range left {
+		l := left[i]
+		out[2*i] = l
+		out[2*i+1] = l - right[i]
+	}
+}
+
+func decorrelateSideRightGo(left, right, out []int32) {
+	for i := range left {
+		side := left[i]
+		rch := right[i]
+		out[2*i] = rch + side
+		out[2*i+1] = rch
+	}
+}
+
+func decorrelateMidSideGo(left, right, out []int32) {
+	for i := range left {
+		mid := left[i]
+		side := right[i]
+		mid2 := mid<<1 | (side & 1)
+		out[2*i] = (mid2 + side) >> 1
+		out[2*i+1] = (mid2 - side) >> 1
+	}
+}
+
+func decorrelateInterleaveGo(left, right, out []int32) {
+	for i := range left {
+		out[2*i] = left[i]
+		out[2*i+1] = right[i]
+	}
+}
+
+func analyzeO1RangeGo(swin []int32, dstP []uint32, hist *[32]uint16) (partSum uint64, overflow bool) {
+	swin = swin[:len(dstP)+1]
+	for j := 0; j+1 < len(swin); j++ {
+		r := swin[j+1] - swin[j]
+		if r == -2147483648 { // math.MinInt32
+			overflow = true
+		}
+		f := uint32(r<<1) ^ uint32(r>>31)
+		dstP[j] = f
+		partSum += uint64(f)
+		bl := 0
+		for v := f; v > 0; v >>= 1 {
+			bl++
+		}
+		if bl > 31 {
+			bl = 31
+		}
+		hist[bl]++
+	}
+	return
+}
+
+// --- Differential tests ---
+
+func TestFLACDecodeMono16(t *testing.T) {
+	k := loadFLACKernel(t)
+	rng := rand.New(rand.NewSource(7))
+
+	for iter := 0; iter < 500; iter++ {
+		n := rng.Intn(256)
+		src := make([]byte, 2*n)
+		rng.Read(src)
+
+		goDst := make([]int32, n)
+		decodeMono16Go(src, goDst)
+
+		jitDst := make([]int32, n)
+		args := make([]byte, 48)
+		if len(src) > 0 {
+			PutPtr(args, 0, unsafe.Pointer(&src[0]))
+		}
+		PutUint64(args, 8, uint64(len(src)))
+		PutUint64(args, 16, uint64(cap(src)))
+		if n > 0 {
+			PutPtr(args, 24, unsafe.Pointer(&jitDst[0]))
+		}
+		PutUint64(args, 32, uint64(n))
+		PutUint64(args, 40, uint64(cap(jitDst)))
+
+		_, err := k.CallFunc("decodeMono16AVX2", args)
+		if err != nil {
+			t.Fatalf("iter %d: %v", iter, err)
+		}
+		for i := range goDst {
+			if jitDst[i] != goDst[i] {
+				t.Fatalf("iter %d: mismatch at [%d]: JIT=%d Go=%d", iter, i, jitDst[i], goDst[i])
+			}
+		}
+	}
+}
+
+func TestFLACPack16(t *testing.T) {
+	k := loadFLACKernel(t)
+	rng := rand.New(rand.NewSource(13))
+
+	for iter := 0; iter < 500; iter++ {
+		n := rng.Intn(256)
+		src := make([]int32, n)
+		for i := range src {
+			src[i] = int32(rng.Intn(65536) - 32768)
+		}
+
+		goDst := make([]byte, 2*n)
+		pack16Go(goDst, src)
+
+		jitDst := make([]byte, 2*n)
+		args := make([]byte, 48)
+		if len(jitDst) > 0 {
+			PutPtr(args, 0, unsafe.Pointer(&jitDst[0]))
+		}
+		PutUint64(args, 8, uint64(len(jitDst)))
+		PutUint64(args, 16, uint64(cap(jitDst)))
+		if n > 0 {
+			PutPtr(args, 24, unsafe.Pointer(&src[0]))
+		}
+		PutUint64(args, 32, uint64(n))
+		PutUint64(args, 40, uint64(cap(src)))
+
+		_, err := k.CallFunc("pack16AVX2", args)
+		if err != nil {
+			t.Fatalf("iter %d: %v", iter, err)
+		}
+		if !bytes.Equal(jitDst, goDst) {
+			t.Fatalf("iter %d: output mismatch (n=%d)", iter, n)
+		}
+	}
+}
+
+func TestFLACDecorrelate(t *testing.T) {
+	k := loadFLACKernel(t)
+	rng := rand.New(rand.NewSource(21))
+
+	kernels := []struct {
+		name string
+		ref  func(left, right, out []int32)
+	}{
+		{"decorrelateLeftSideAVX2", decorrelateLeftSideGo},
+		{"decorrelateSideRightAVX2", decorrelateSideRightGo},
+		{"decorrelateMidSideAVX2", decorrelateMidSideGo},
+		{"decorrelateInterleaveAVX2", decorrelateInterleaveGo},
+	}
+
+	for _, kk := range kernels {
+		t.Run(kk.name, func(t *testing.T) {
+			for iter := 0; iter < 200; iter++ {
+				n := rng.Intn(128)
+				left := make([]int32, n)
+				right := make([]int32, n)
+				for i := range left {
+					left[i] = int32(rng.Intn(1<<24) - 1<<23)
+					right[i] = int32(rng.Intn(1<<24) - 1<<23)
+				}
+
+				goOut := make([]int32, 2*n)
+				kk.ref(left, right, goOut)
+
+				jitOut := make([]int32, 2*n)
+				args := make([]byte, 72)
+				if n > 0 {
+					PutPtr(args, 0, unsafe.Pointer(&left[0]))
+					PutPtr(args, 24, unsafe.Pointer(&right[0]))
+					PutPtr(args, 48, unsafe.Pointer(&jitOut[0]))
+				}
+				PutUint64(args, 8, uint64(n))
+				PutUint64(args, 16, uint64(cap(left)))
+				PutUint64(args, 32, uint64(n))
+				PutUint64(args, 40, uint64(cap(right)))
+				PutUint64(args, 56, uint64(2*n))
+				PutUint64(args, 64, uint64(cap(jitOut)))
+
+				_, err := k.CallFunc(kk.name, args)
+				if err != nil {
+					t.Fatalf("iter %d: %v", iter, err)
+				}
+				for i := range goOut {
+					if jitOut[i] != goOut[i] {
+						t.Fatalf("iter %d: mismatch at [%d]: JIT=%d Go=%d", iter, i, jitOut[i], goOut[i])
+					}
+				}
+			}
+		})
+	}
+}
+
+func TestFLACAnalyzeO1Range(t *testing.T) {
+	k := loadFLACKernel(t)
+	rng := rand.New(rand.NewSource(33))
+
+	for iter := 0; iter < 300; iter++ {
+		n := 1 + rng.Intn(128) // partition size
+		swin := make([]int32, n+1)
+		for i := range swin {
+			swin[i] = int32(rng.Intn(1<<20) - 1<<19)
+		}
+
+		goDstP := make([]uint32, n)
+		var goHist [32]uint16
+		goSum, goOvf := analyzeO1RangeGo(swin, goDstP, &goHist)
+
+		jitDstP := make([]uint32, n)
+		var jitHist [32]uint16
+		args := make([]byte, 72) // 65 rounded up
+		PutPtr(args, 0, unsafe.Pointer(&swin[0]))
+		PutUint64(args, 8, uint64(len(swin)))
+		PutUint64(args, 16, uint64(cap(swin)))
+		PutPtr(args, 24, unsafe.Pointer(&jitDstP[0]))
+		PutUint64(args, 32, uint64(n))
+		PutUint64(args, 40, uint64(cap(jitDstP)))
+		PutPtr(args, 48, unsafe.Pointer(&jitHist[0]))
+
+		out, err := k.CallFunc("analyzeO1RangeAVX2", args)
+		if err != nil {
+			t.Fatalf("iter %d: %v", iter, err)
+		}
+		jitSum := GetUint64(out, 56)
+		jitOvf := out[64] != 0
+
+		if jitSum != goSum {
+			t.Fatalf("iter %d: partSum mismatch: JIT=%d Go=%d", iter, jitSum, goSum)
+		}
+		if jitOvf != goOvf {
+			t.Fatalf("iter %d: overflow mismatch: JIT=%v Go=%v", iter, jitOvf, goOvf)
+		}
+		for i := range goDstP {
+			if jitDstP[i] != goDstP[i] {
+				t.Fatalf("iter %d: dstP[%d] mismatch: JIT=%d Go=%d", iter, i, jitDstP[i], goDstP[i])
+			}
+		}
+		if jitHist != goHist {
+			t.Fatalf("iter %d: hist mismatch: JIT=%v Go=%v", iter, jitHist, goHist)
+		}
+	}
+}
+
+func TestFLACFastStereoSums(t *testing.T) {
+	k := loadFLACKernel(t)
+	rng := rand.New(rand.NewSource(44))
+
+	for iter := 0; iter < 300; iter++ {
+		n := 1 + rng.Intn(256)
+		left := make([]int32, n)
+		right := make([]int32, n)
+		for i := range left {
+			left[i] = int32(rng.Intn(1<<24) - 1<<23)
+			right[i] = int32(rng.Intn(1<<24) - 1<<23)
+		}
+
+		// Go reference: compute the four sums.
+		var goSums [4]uint64
+		for i := 0; i < n; i++ {
+			l := left[i]
+			r := right[i]
+			side := l - r
+			mid := (l + r) >> 1
+			goSums[0] += foldAbs(l) + foldAbs(r)
+			goSums[1] += foldAbs(l) + foldAbs(side)
+			goSums[2] += foldAbs(side) + foldAbs(r)
+			goSums[3] += foldAbs(mid) + foldAbs(side)
+		}
+
+		var jitSums [4]uint64
+		args := make([]byte, 56)
+		PutPtr(args, 0, unsafe.Pointer(&left[0]))
+		PutUint64(args, 8, uint64(n))
+		PutUint64(args, 16, uint64(cap(left)))
+		PutPtr(args, 24, unsafe.Pointer(&right[0]))
+		PutUint64(args, 32, uint64(n))
+		PutUint64(args, 40, uint64(cap(right)))
+		PutPtr(args, 48, unsafe.Pointer(&jitSums[0]))
+
+		_, err := k.CallFunc("fastStereoSumsAVX2", args)
+		if err != nil {
+			t.Fatalf("iter %d: %v", iter, err)
+		}
+		if jitSums != goSums {
+			t.Fatalf("iter %d: sums mismatch:\n  JIT=%v\n  Go =%v", iter, jitSums, goSums)
+		}
+	}
+}
+
+func foldAbs(v int32) uint64 {
+	return uint64(uint32(v<<1) ^ uint32(v>>31))
+}

verify/jit.go

@@ -0,0 +1,88 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+// Package verify provides the dynamic-analysis substrate for gasm: it
+// JIT-assembles Plan 9 amd64 kernels into executable memory and calls them
+// directly, enabling differential testing against portable Go references,
+// runtime ABI checks and basic-block coverage profiling.
+//
+// The execution model is pure Go (stdlib only): machine code is mapped with
+// syscall.Mmap and invoked through an assembly trampoline that switches to a
+// prepared ABI0 stack.  No cgo, no external toolchain.
+package verify
+
+import (
+	"encoding/binary"
+	"fmt"
+	"syscall"
+	"unsafe"
+)
+
+// Executable maps a copy of code into a read-execute memory region suitable
+// for direct invocation.  The mapping is anonymous and private; the original
+// slice is not retained.  Call Unmap to release the region.
+type Executable struct {
+	addr uintptr // base address of the mapping
+	size int
+	mem  []byte // the mmap'd slice (for Unmap)
+}
+
+// Map copies code into a freshly allocated RX region and returns it.
+// The mapping is PROT_READ|PROT_EXEC; writes are not permitted after the
+// copy, matching W^X policy.
+func Map(code []byte) (*Executable, error) {
+	size := len(code)
+	if size == 0 {
+		return nil, fmt.Errorf("verify: cannot map zero-length code")
+	}
+	// Round up to the page size.
+	const pageSize = 4096
+	mapSize := (size + pageSize - 1) &^ (pageSize - 1)
+
+	mem, err := syscall.Mmap(-1, 0, mapSize,
+		syscall.PROT_READ|syscall.PROT_WRITE, syscall.MAP_PRIVATE|syscall.MAP_ANON)
+	if err != nil {
+		return nil, fmt.Errorf("verify: mmap: %w", err)
+	}
+	copy(mem, code)
+
+	// Remove write permission (W^X).
+	if err := syscall.Mprotect(mem, syscall.PROT_READ|syscall.PROT_EXEC); err != nil {
+		syscall.Munmap(mem)
+		return nil, fmt.Errorf("verify: mprotect: %w", err)
+	}
+	return &Executable{
+		addr: uintptr(unsafe.Pointer(&mem[0])),
+		size: size,
+		mem:  mem,
+	}, nil
+}
+
+// Unmap releases the executable region.
+func (e *Executable) Unmap() {
+	if e.mem != nil {
+		syscall.Munmap(e.mem)
+		e.mem = nil
+	}
+}
+
+// FuncAddr returns the absolute address of a function at the given offset
+// within the mapped image.
+func (e *Executable) FuncAddr(offset int) uintptr {
+	return e.addr + uintptr(offset)
+}
+
+// PutUint64 writes v into buf at byte offset off (little-endian).
+func PutUint64(buf []byte, off int, v uint64) {
+	binary.LittleEndian.PutUint64(buf[off:off+8], v)
+}
+
+// GetUint64 reads a little-endian uint64 from buf at byte offset off.
+func GetUint64(buf []byte, off int) uint64 {
+	return binary.LittleEndian.Uint64(buf[off : off+8])
+}
+
+// PutPtr writes a pointer value into buf at byte offset off.
+func PutPtr(buf []byte, off int, p unsafe.Pointer) {
+	binary.LittleEndian.PutUint64(buf[off:off+8], uint64(uintptr(p)))
+}

verify/jit_test.go

@@ -0,0 +1,215 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package verify
+
+import (
+	"bytes"
+	"testing"
+	"unsafe"
+)
+
+func loadBasic(t *testing.T) *Kernel {
+	t.Helper()
+	k, err := Load("../testdata/verify/basic_amd64.s")
+	if err != nil {
+		t.Fatalf("Load: %v", err)
+	}
+	t.Cleanup(k.Close)
+	return k
+}
+
+func TestJITAdd(t *testing.T) {
+	k := loadBasic(t)
+
+	tests := []struct {
+		a, b, want int64
+	}{
+		{0, 0, 0},
+		{1, 2, 3},
+		{-1, 1, 0},
+		{1 << 62, 1 << 62, -9223372036854775808}, // overflow wraps (MinInt64)
+		{-100, -200, -300},
+	}
+	for _, tt := range tests {
+		args := make([]byte, 24)
+		PutUint64(args, 0, uint64(tt.a))
+		PutUint64(args, 8, uint64(tt.b))
+
+		out, err := k.CallFunc("add", args)
+		if err != nil {
+			t.Fatalf("CallFunc(add, %d, %d): %v", tt.a, tt.b, err)
+		}
+		got := int64(GetUint64(out, 16))
+		if got != tt.want {
+			t.Errorf("add(%d, %d) = %d, want %d", tt.a, tt.b, got, tt.want)
+		}
+	}
+}
+
+func TestJITSum(t *testing.T) {
+	k := loadBasic(t)
+
+	tests := []struct {
+		data []int64
+		want int64
+	}{
+		{nil, 0},
+		{[]int64{1}, 1},
+		{[]int64{1, 2, 3, 4, 5}, 15},
+		{[]int64{-10, 20, -30, 40}, 20},
+	}
+	for _, tt := range tests {
+		args := make([]byte, 32)
+		if len(tt.data) > 0 {
+			PutPtr(args, 0, unsafe.Pointer(&tt.data[0]))
+		}
+		PutUint64(args, 8, uint64(len(tt.data)))
+		PutUint64(args, 16, uint64(cap(tt.data)))
+
+		out, err := k.CallFunc("sum", args)
+		if err != nil {
+			t.Fatalf("CallFunc(sum, %v): %v", tt.data, err)
+		}
+		got := int64(GetUint64(out, 24))
+		if got != tt.want {
+			t.Errorf("sum(%v) = %d, want %d", tt.data, got, tt.want)
+		}
+	}
+}
+
+func TestJITWideCopy(t *testing.T) {
+	k := loadBasic(t)
+
+	tests := []struct {
+		name string
+		n    int
+	}{
+		{"empty", 0},
+		{"tiny", 7},
+		{"exact32", 32},
+		{"overlap_range", 48},
+		{"exact64", 64},
+		{"unaligned", 45},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			src := make([]byte, tt.n)
+			for i := range src {
+				src[i] = byte(i * 7)
+			}
+			dst := make([]byte, tt.n)
+
+			args := make([]byte, 48)
+			if tt.n > 0 {
+				PutPtr(args, 0, unsafe.Pointer(&dst[0]))
+				PutPtr(args, 24, unsafe.Pointer(&src[0]))
+			}
+			PutUint64(args, 8, uint64(tt.n))  // dst_len
+			PutUint64(args, 16, uint64(tt.n)) // dst_cap
+			PutUint64(args, 32, uint64(tt.n)) // src_len
+			PutUint64(args, 40, uint64(tt.n)) // src_cap
+
+			_, err := k.CallFunc("wideCopy", args)
+			if err != nil {
+				t.Fatalf("CallFunc(wideCopy): %v", err)
+			}
+			if !bytes.Equal(dst, src) {
+				t.Errorf("wideCopy: dst ≠ src\n  got  %x\n  want %x", dst, src)
+			}
+		})
+	}
+}
+
+func TestKernelFuncNames(t *testing.T) {
+	k := loadBasic(t)
+	names := k.FuncNames()
+	want := []string{"add", "sum", "wideCopy"}
+	if len(names) != len(want) {
+		t.Fatalf("FuncNames() = %v, want %v", names, want)
+	}
+	for i, n := range names {
+		if n != want[i] {
+			t.Errorf("FuncNames()[%d] = %q, want %q", i, n, want[i])
+		}
+	}
+}
+
+func TestKernelFuncNotFound(t *testing.T) {
+	k := loadBasic(t)
+	_, err := k.CallFunc("nonexistent", make([]byte, 8))
+	if err == nil {
+		t.Fatal("expected error for nonexistent function")
+	}
+}
+
+func TestKernelArgTooSmall(t *testing.T) {
+	k := loadBasic(t)
+	_, err := k.CallFunc("add", make([]byte, 8)) // needs 24
+	if err == nil {
+		t.Fatal("expected error for too-small arg block")
+	}
+}
+
+func TestMapZeroLength(t *testing.T) {
+	_, err := Map(nil)
+	if err == nil {
+		t.Fatal("expected error for zero-length code")
+	}
+}
+
+func TestLoadSourceError(t *testing.T) {
+	_, err := LoadSource("bad.s", "TEXT ·f(SB), NOSPLIT\n\tBADINSTRUCTION\n")
+	// The parser may or may not error on unknown instructions (it's
+	// error-tolerant), but the assembler will reject it.
+	if err == nil {
+		t.Log("LoadSource succeeded unexpectedly (parser is error-tolerant)")
+	}
+}
+
+func TestLoadSourceParseError(t *testing.T) {
+	// A completely invalid file that the parser rejects.
+	_, err := LoadSource("empty.s", "")
+	if err != nil {
+		t.Logf("expected: %v", err)
+	}
+}
+
+func TestFuncLookup(t *testing.T) {
+	k := loadBasic(t)
+	fl, err := k.Func("add")
+	if err != nil {
+		t.Fatalf("Func(add): %v", err)
+	}
+	if fl.Name != "add" {
+		t.Errorf("Func(add).Name = %q, want %q", fl.Name, "add")
+	}
+	if fl.Args != 24 {
+		t.Errorf("Func(add).Args = %d, want 24", fl.Args)
+	}
+	_, err = k.Func("nonexistent")
+	if err == nil {
+		t.Fatal("expected error for nonexistent function")
+	}
+}
+
+func TestABIReportString(t *testing.T) {
+	r := ABIReport{}
+	if r.String() != "ABI clean" {
+		t.Errorf("clean report = %q", r.String())
+	}
+	r.BPClobbered = true
+	if r.OK() {
+		t.Error("expected not OK with BP clobbered")
+	}
+	s := r.String()
+	if s == "ABI clean" {
+		t.Error("expected violation string, got clean")
+	}
+	r.R14Clobbered = true
+	r.RedZoneHit = true
+	s = r.String()
+	if s == "ABI clean" {
+		t.Error("expected violation string for all flags")
+	}
+}

verify/lz4_test.go

@@ -0,0 +1,160 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package verify
+
+import (
+	"bytes"
+	"os"
+	"testing"
+	"unsafe"
+)
+
+// lz4KernelPath is the sibling repository's AVX2 kernel, used for
+// integration testing.  The test is skipped when the file is absent
+// (e.g. in CI without the sibling checkout).
+const lz4KernelPath = "../../go-libraries/go-lz4/avx2_amd64.s"
+
+func loadLZ4Kernel(t *testing.T) *Kernel {
+	t.Helper()
+	if _, err := os.Stat(lz4KernelPath); err != nil {
+		t.Skipf("sibling kernel not available: %v", err)
+	}
+	k, err := Load(lz4KernelPath)
+	if err != nil {
+		t.Fatalf("Load(%s): %v", lz4KernelPath, err)
+	}
+	t.Cleanup(k.Close)
+	return k
+}
+
+// callDecodeBlockAVX2 invokes the JIT-assembled decodeBlockAVX2 with the
+// given src and dst buffers, returning (n, code).
+func callDecodeBlockAVX2(t *testing.T, k *Kernel, src, dst []byte) (int, int) {
+	t.Helper()
+	args := make([]byte, 64)
+	if len(src) > 0 {
+		PutPtr(args, 0, unsafe.Pointer(&src[0]))
+	}
+	PutUint64(args, 8, uint64(len(src)))
+	PutUint64(args, 16, uint64(cap(src)))
+	if len(dst) > 0 {
+		PutPtr(args, 24, unsafe.Pointer(&dst[0]))
+	}
+	PutUint64(args, 32, uint64(len(dst)))
+	PutUint64(args, 40, uint64(cap(dst)))
+
+	out, err := k.CallFunc("decodeBlockAVX2", args)
+	if err != nil {
+		t.Fatalf("CallFunc(decodeBlockAVX2): %v", err)
+	}
+	return int(GetUint64(out, 48)), int(GetUint64(out, 56))
+}
+
+func TestLZ4DecodeKnownAnswers(t *testing.T) {
+	k := loadLZ4Kernel(t)
+
+	tests := []struct {
+		name     string
+		src      []byte
+		dstSize  int
+		wantDst  []byte
+		wantN    int
+		wantCode int
+	}{
+		{
+			name:     "literals_only",
+			src:      []byte{0x50, 'H', 'e', 'l', 'l', 'o'},
+			dstSize:  16,
+			wantDst:  []byte("Hello"),
+			wantN:    5,
+			wantCode: 0,
+		},
+		{
+			name:     "literals_and_match",
+			src:      []byte{0x54, 'A', 'A', 'A', 'A', 'A', 0x05, 0x00, 0x30, 'B', 'B', 'B'},
+			dstSize:  32,
+			wantDst:  []byte("AAAAAAAAAAAAABBB"),
+			wantN:    16,
+			wantCode: 0,
+		},
+		{
+			name: "overlapping_match",
+			// 1 literal 'X', then match offset=1 length=4+4=8 → "XXXXXXXXX",
+			// then final 1 literal 'Y'.
+			src:      []byte{0x14, 'X', 0x01, 0x00, 0x10, 'Y'},
+			dstSize:  16,
+			wantDst:  []byte("XXXXXXXXXY"),
+			wantN:    10,
+			wantCode: 0,
+		},
+		{
+			name:     "malformed_truncated",
+			src:      []byte{0x50, 'H', 'e'}, // claims 5 literals, has 2
+			dstSize:  16,
+			wantN:    0,
+			wantCode: 1,
+		},
+		{
+			name:     "zero_offset",
+			src:      []byte{0x14, 'X', 0x00, 0x00},
+			dstSize:  16,
+			wantN:    0,
+			wantCode: 2,
+		},
+		{
+			name:     "empty_token",
+			src:      []byte{0x00}, // 0 literals, end of block
+			dstSize:  16,
+			wantDst:  nil,
+			wantN:    0,
+			wantCode: 0,
+		},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			dst := make([]byte, tt.dstSize)
+			n, code := callDecodeBlockAVX2(t, k, tt.src, dst)
+			if n != tt.wantN || code != tt.wantCode {
+				t.Fatalf("decodeBlockAVX2: got (n=%d, code=%d), want (n=%d, code=%d)",
+					n, code, tt.wantN, tt.wantCode)
+			}
+			if tt.wantCode == 0 && tt.wantDst != nil {
+				if !bytes.Equal(dst[:n], tt.wantDst) {
+					t.Errorf("output mismatch:\n  got  %q\n  want %q", dst[:n], tt.wantDst)
+				}
+			}
+		})
+	}
+}
+
+func TestLZ4WideCopyAVX2(t *testing.T) {
+	k := loadLZ4Kernel(t)
+
+	sizes := []int{0, 1, 15, 16, 31, 32, 33, 63, 64, 100, 256, 1024}
+	for _, n := range sizes {
+		src := make([]byte, n)
+		for i := range src {
+			src[i] = byte(i*13 + 7)
+		}
+		dst := make([]byte, n)
+
+		args := make([]byte, 48)
+		if n > 0 {
+			PutPtr(args, 0, unsafe.Pointer(&dst[0]))
+			PutPtr(args, 24, unsafe.Pointer(&src[0]))
+		}
+		PutUint64(args, 8, uint64(n))
+		PutUint64(args, 16, uint64(n))
+		PutUint64(args, 32, uint64(n))
+		PutUint64(args, 40, uint64(n))
+
+		_, err := k.CallFunc("wideCopyAVX2", args)
+		if err != nil {
+			t.Fatalf("wideCopyAVX2(n=%d): %v", n, err)
+		}
+		if !bytes.Equal(dst, src) {
+			t.Errorf("wideCopyAVX2(n=%d): output mismatch", n)
+		}
+	}
+}

verify/trampoline_amd64.s

@@ -0,0 +1,30 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+#include "textflag.h"
+
+// ABI0 JIT trampoline.  enterJIT switches from the Go stack to a prepared
+// stack and jumps to the assembled function; when the function RETs, control
+// lands in leaveJIT, which restores the Go stack and returns to the Go caller.
+//
+// The prepared stack must begin with the address of leaveJIT (the return
+// address the JIT function will pop), followed by the function's ABI0
+// argument area.
+//
+// Single-threaded: savedSP is a package global, so only one JIT call may be
+// in flight at a time.  gasm verify runs sequentially.
+
+// func enterJIT(fn uintptr, stack uintptr)
+// Switches to the prepared stack and jumps to fn.  Does not return normally;
+// the JIT function's RET transfers control to leaveJIT.
+TEXT ·enterJIT(SB), NOSPLIT, $0-16
+	MOVQ fn+0(FP), AX      // target function address (before SP switch)
+	MOVQ SP, ·savedSP(SB)  // preserve the Go stack pointer
+	MOVQ stack+8(FP), SP   // switch to the prepared stack
+	JMP  AX
+
+// func leaveJIT()
+// Restores the Go stack pointer and returns to enterJIT's caller.
+TEXT ·leaveJIT(SB), NOSPLIT, $0-0
+	MOVQ ·savedSP(SB), SP
+	RET

verify/verify.go

@@ -0,0 +1,122 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package verify
+
+import (
+	"fmt"
+	"os"
+
+	"sourcedock.dev/petrbalvin/gasm-devkit/asm"
+	"sourcedock.dev/petrbalvin/gasm-devkit/ast"
+	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+)
+
+// Kernel is a JIT-loaded assembly image ready for direct invocation.
+// It wraps an executable memory mapping and the function layout metadata
+// needed to marshal ABI0 calls.
+type Kernel struct {
+	exec  *Executable
+	img   *asm.Image
+	funcs map[string]int // function name → index into img.Funcs
+}
+
+// Load parses, assembles and maps a .s file into executable memory.
+// The returned Kernel is ready for Call.  The caller must call Close to
+// release the mapping.
+func Load(path string) (*Kernel, error) {
+	src, err := os.ReadFile(path)
+	if err != nil {
+		return nil, fmt.Errorf("verify: %w", err)
+	}
+	return LoadSource(path, string(src))
+}
+
+// LoadSource parses, assembles and maps assembly source into executable memory.
+func LoadSource(filename, src string) (*Kernel, error) {
+	file, errs := parser.Parse(filename, src)
+	if len(errs) > 0 {
+		return nil, fmt.Errorf("verify: parse %s: %v", filename, errs[0])
+	}
+	return LoadAST(file)
+}
+
+// LoadAST assembles a parsed AST file and maps the result into executable
+// memory.
+func LoadAST(file *ast.File) (*Kernel, error) {
+	img, err := asm.AssembleFile(file)
+	if err != nil {
+		return nil, fmt.Errorf("verify: assemble: %w", err)
+	}
+	if len(img.Externals) > 0 {
+		return nil, fmt.Errorf("verify: unresolved external symbols: %v", img.Externals)
+	}
+	code := img.Bytes()
+	exec, err := Map(code)
+	if err != nil {
+		return nil, err
+	}
+	funcs := make(map[string]int, len(img.Funcs))
+	for i, f := range img.Funcs {
+		funcs[f.Name] = i
+	}
+	return &Kernel{exec: exec, img: img, funcs: funcs}, nil
+}
+
+// Func returns the layout metadata for the named function.
+func (k *Kernel) Func(name string) (asm.FuncLayout, error) {
+	idx, ok := k.funcs[name]
+	if !ok {
+		return asm.FuncLayout{}, fmt.Errorf("verify: function %q not found", name)
+	}
+	return k.img.Funcs[idx], nil
+}
+
+// FuncNames returns the names of all functions in the kernel, in source order.
+func (k *Kernel) FuncNames() []string {
+	names := make([]string, len(k.img.Funcs))
+	for i, f := range k.img.Funcs {
+		names[i] = f.Name
+	}
+	return names
+}
+
+// CallFunc invokes the named function with the given ABI0 argument block.
+// The arg block is the raw bytes of the function's argument/result area
+// (as declared by the TEXT $frame-args suffix).  Returns the arg block
+// after the call (with any results written back by the function).
+func (k *Kernel) CallFunc(name string, args []byte) ([]byte, error) {
+	idx, ok := k.funcs[name]
+	if !ok {
+		return nil, fmt.Errorf("verify: function %q not found", name)
+	}
+	fl := k.img.Funcs[idx]
+	if len(args) < fl.Args {
+		return nil, fmt.Errorf("verify: %s: arg block too small: got %d, need %d", name, len(args), fl.Args)
+	}
+	fnAddr := k.exec.FuncAddr(fl.Offset)
+	return Call(fnAddr, args)
+}
+
+// CallFuncChecked invokes the named function with ABI sentinels and a
+// red-zone canary, returning the argument block and an ABIReport that
+// records any callee-saved register or red-zone violations.
+func (k *Kernel) CallFuncChecked(name string, args []byte) ([]byte, ABIReport, error) {
+	idx, ok := k.funcs[name]
+	if !ok {
+		return nil, ABIReport{}, fmt.Errorf("verify: function %q not found", name)
+	}
+	fl := k.img.Funcs[idx]
+	if len(args) < fl.Args {
+		return nil, ABIReport{}, fmt.Errorf("verify: %s: arg block too small: got %d, need %d", name, len(args), fl.Args)
+	}
+	fnAddr := k.exec.FuncAddr(fl.Offset)
+	return CallChecked(fnAddr, args)
+}
+
+// Close releases the executable mapping.
+func (k *Kernel) Close() {
+	if k.exec != nil {
+		k.exec.Unmap()
+	}
+}