feat(fmt): go-fmt-style recursive formatting and canonical blank-line layout

Assisted-by: Qwen 3.8 Max Preview
fix(lint): calibrate register-clobber to the Go ABI and add legacy SSE moves
2026-07-12 21:24:41 +02:00 · 2026-07-11 17:36:52 +02:00 · 2026-07-10 13:20:49 +02:00 · 2026-07-09 15:56:03 +02:00
21 changed files with 2081 additions and 228 deletions
@@ -23,6 +23,31 @@ 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)
+	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).
+type linkInfo struct {
+	symbols map[string]bool
+}
+
+// sbPatch is a function-relative static-symbol relocation: the disp32 field
+// at off must become the symbol's address minus after, where after is the
+// function-relative address just past the instruction.
+type sbPatch struct {
+	off    int
+	after  int
+	name   string
+	addend int64
+}
+
+// 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.
+func assemble(t *ast.Text, link *linkInfo) ([]byte, []sbPatch, map[string]int, error) {
 	fi := computeFrame(t)
 	chain := jumpChain(t)
 	resolve := func(name string) string {
@@ -44,9 +69,9 @@ func Assemble(t *ast.Text) ([]byte, map[string]int, error) {
 			case *ast.Label:
 				offsets[s.Name.Text] = pos
 			case *ast.Instr:
-				sz, err := instrSize(s, fi, long[i])
+				sz, err := instrSize(s, fi, long[i], link)
 				if err != nil {
-					return nil, nil, fmt.Errorf("%s: %w", s.Mnemonic.Text, err)
+					return nil, nil, nil, fmt.Errorf("%s: %w", s.Mnemonic.Text, err)
 				}
 				sizes[i] = sz
 				pcs[i] = pos
@@ -85,23 +110,25 @@ func Assemble(t *ast.Text) ([]byte, map[string]int, error) {

 	// Pass 2: emit.
 	out := append([]byte(nil), fi.prologue...)
+	var patches []sbPatch
 	pos := len(fi.prologue)
 	for i, stmt := range t.Body {
 		s, ok := stmt.(*ast.Instr)
 		if !ok {
 			continue
 		}
-		code, err := encodeInstr(s, pos, offsets, fi, long[i], resolve)
+		code, ps, err := encodeInstr(s, pos, offsets, fi, long[i], resolve, link)
 		if err != nil {
-			return nil, nil, fmt.Errorf("%s: %w", s.Mnemonic.Text, err)
+			return nil, nil, nil, fmt.Errorf("%s: %w", s.Mnemonic.Text, err)
 		}
 		if len(code) != sizes[i] {
-			return nil, nil, fmt.Errorf("%s: size mismatch (%d vs %d)", s.Mnemonic.Text, len(code), sizes[i])
+			return 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, offsets, nil
+	return out, patches, offsets, nil
 }

 // jumpChain precomputes jump-to-jump folding: a label whose first instruction
@@ -214,12 +241,12 @@ func addSP(size int) []byte { // ADDQ $size, SP
 // instrSize returns the encoded length of an instruction (layout pass).
 // encodeInstr already includes the epilogue for a RET in a frame-pointer
 // function; jumps use their short or long form (never an epilogue).
-func instrSize(s *ast.Instr, fi frameInfo, long bool) (int, error) {
+func instrSize(s *ast.Instr, fi frameInfo, long bool, link *linkInfo) (int, error) {
 	mnem := strings.ToUpper(s.Mnemonic.Text)
 	if isJumpMnemonic(mnem) {
 		return jumpSize(mnem, long), nil
 	}
-	code, err := encodeInstr(s, 0, nil, fi, false, nil)
+	code, _, err := encodeInstr(s, 0, nil, fi, false, nil, link)
 	if err != nil {
 		return 0, err
 	}
@@ -254,7 +281,7 @@ func jumpSize(mnem string, long bool) int {
 // (relative to pc, the instruction's own offset).  A RET in a frame-pointer
 // function is prefixed with the epilogue.  resolve, when non-nil, redirects a
 // jump label through the jump-to-jump chain before the offset lookup.
-func encodeInstr(s *ast.Instr, pc int, offsets map[string]int, fi frameInfo, long bool, resolve func(string) string) ([]byte, error) {
+func encodeInstr(s *ast.Instr, pc int, offsets map[string]int, fi frameInfo, long bool, resolve func(string) string, link *linkInfo) ([]byte, []sbPatch, error) {
 	mnem := strings.ToUpper(s.Mnemonic.Text)

 	var prefix []byte
@@ -263,32 +290,48 @@ func encodeInstr(s *ast.Instr, pc int, offsets map[string]int, fi frameInfo, lon
 	}

 	var code []byte
+	var ps []sbPatch
 	var err error
 	if isJumpMnemonic(mnem) {
 		code, err = encodeJump(s, mnem, pc+len(prefix), offsets, long, resolve)
 	} else {
-		code, err = encodeNormal(s, fi)
+		code, ps, err = encodeNormal(s, fi, link)
 	}
 	if err != nil {
-		return nil, err
+		return nil, nil, err
 	}
-	return append(prefix, code...), nil
+	// Anchor the patch fields at function-relative positions: off indexes the
+	// disp32 field, after is the address just past the instruction.
+	body := pc + len(prefix)
+	for i := range ps {
+		ps[i].off += body
+		ps[i].after = body + len(code)
+	}
+	return append(prefix, code...), ps, nil
 }

-func encodeNormal(s *ast.Instr, fi frameInfo) ([]byte, error) {
+func encodeNormal(s *ast.Instr, fi frameInfo, link *linkInfo) ([]byte, []sbPatch, error) {
 	_, size := splitSize(strings.ToUpper(s.Mnemonic.Text))
 	if size == 0 {
 		size = 8
 	}
 	ops := make([]Operand, len(s.Operands))
 	for i, op := range s.Operands {
-		o, err := operandFromAST(op, size, fi)
+		o, err := operandFromAST(op, size, fi, link)
 		if err != nil {
-			return nil, err
+			return nil, nil, err
 		}
 		ops[i] = o
 	}
-	return Encode(s.Mnemonic.Text, ops...)
+	e := &enc{}
+	if err := e.encode(s.Mnemonic.Text, ops); err != nil {
+		return nil, nil, err
+	}
+	ps := make([]sbPatch, len(e.patches))
+	for i, p := range e.patches {
+		ps[i] = sbPatch{off: p.off, name: p.name, addend: p.addend}
+	}
+	return e.out, ps, nil
 }

 // encodeJump encodes a JMP/CALL/Jcc with a relative offset resolved from the
@@ -345,7 +388,7 @@ var spReg = Reg{idx: 4, size: 8}

 // operandFromAST converts a parsed operand into an encoder Operand, applying
 // the frame translation to FP/SP pseudo-register operands.
-func operandFromAST(op *ast.Operand, size int, fi frameInfo) (Operand, error) {
+func operandFromAST(op *ast.Operand, size int, fi frameInfo, link *linkInfo) (Operand, error) {
 	switch op.Kind {
 	case ast.OpImmediate:
 		if op.Imm.HasVal {
@@ -371,9 +414,20 @@ func operandFromAST(op *ast.Operand, size int, fi frameInfo) (Operand, error) {
 			off := fi.spAdjust + a.Sym.Offset
 			return Mem{Base: spReg, Disp: off, HasBase: true, Size: size}, nil
 		}
-		// SB (global symbol) needs a relocation — not yet supported.
+		// SB (global symbol): a symbol defined in the same file (GLOBL) is
+		// encoded RIP-relative and resolved by the file-level layout;
+		// anything not defined here needs object-file emission.
 		if a.Sym != nil && a.Sym.Pseudo == "SB" {
-			return nil, fmt.Errorf("SB (global symbol) operands need relocation support (pending)")
+			if link == nil || link.symbols == nil {
+				return nil, fmt.Errorf("symbol %q needs file-level assembly (AssembleFile)", a.Sym.Name)
+			}
+			if !link.symbols[a.Sym.Name] {
+				if a.Sym.Static {
+					return nil, fmt.Errorf("undefined symbol %q", a.Sym.Name)
+				}
+				return nil, fmt.Errorf("external symbol %q needs object-file emission", a.Sym.Name)
+			}
+			return sbMem{size: size, name: a.Sym.Name, addend: a.Sym.Offset}, nil
 		}

 		// Memory with a real base register: (base), off(base), (base)(index*scale).
@@ -19,7 +19,16 @@ func Encode(mnemonic string, ops ...Operand) ([]byte, error) {
 }

 type enc struct {
-	out []byte
+	out     []byte
+	patches []encPatch // disp32 fields awaiting static-symbol resolution
+}
+
+// encPatch marks a 4-byte displacement field in enc.out that must receive the
+// RIP-relative offset of a static symbol once the file layout is settled.
+type encPatch struct {
+	off    int
+	name   string
+	addend int64
 }

 func (e *enc) encode(mnem string, ops []Operand) error {
@@ -40,10 +49,11 @@ func (e *enc) encode(mnem string, ops []Operand) error {
 		return e.encodeJcc(cc, ops)
 	}

-	// VEX (AVX/AVX2) instructions: the trailing B/W/L/Q/D is part of the
-	// mnemonic, not a size suffix, so dispatch before splitSize.
-	if isVex(upper) {
-		return e.encodeVex(upper, ops)
+	// VEX (AVX/AVX2) and EVEX (AVX-512) instructions: the trailing
+	// B/W/L/Q/D is part of the mnemonic, not a size suffix, so dispatch
+	// before splitSize.
+	if isVex(upper) || isEvex(upper) || upper == "KMOVW" {
+		return e.encodeVec(upper, ops)
 	}

 	// CMOVcc and SETcc carry the condition in the mnemonic (CMOVLGT, SETNE).
@@ -83,6 +93,8 @@ 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)
 }
@@ -105,6 +117,20 @@ func splitSize(upper string) (base string, size int) {
 	return upper, 0
 }

+// encodeVec dispatches a VEX/EVEX mnemonic to the right encoding: KMOVW has
+// 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 {
+	if upper == "KMOVW" {
+		return e.encodeKmovw(ops)
+	}
+	if upper == "KTESTW" || !evexRequired(upper, ops) {
+		return e.encodeVex(upper, ops)
+	}
+	return e.encodeEvex(upper, ops)
+}
+
 // --- instruction components -------------------------------------------------

 type instr struct {
@@ -120,6 +146,14 @@ type instr struct {
 	sib       int // -1 if absent
 	disp      []byte
 	imm       []byte
+	sb        *sbRef // static-symbol displacement in disp, awaiting resolution
+}
+
+// sbRef records that an instruction's displacement refers to a static symbol
+// rather than holding a literal value.
+type sbRef struct {
+	name   string
+	addend int64
 }

 func (e *enc) emit(i *instr) error {
@@ -152,6 +186,9 @@ func (e *enc) emit(i *instr) error {
 	if i.sib >= 0 {
 		e.out = append(e.out, byte(i.sib))
 	}
+	if i.sb != nil {
+		e.patches = append(e.patches, encPatch{off: len(e.out), name: i.sb.name, addend: i.sb.addend})
+	}
 	e.out = append(e.out, i.disp...)
 	e.out = append(e.out, i.imm...)
 	return nil
@@ -199,6 +236,13 @@ func setRMReg(i *instr, regField int, rexR, regForced bool, rm Operand, opSize i
 		return nil
 	case Mem:
 		return setMem(i, regField, r)
+	case sbMem:
+		// RIP-relative reference; the displacement is patched once the static
+		// symbol's address is known.
+		i.modrm = regField<<3 | 0x05 // mod=00, rm=101 → (RIP)+disp32
+		i.disp = le32(0)
+		i.sb = &sbRef{name: r.name, addend: r.addend}
+		return nil
 	default:
 		return fmt.Errorf("invalid r/m operand %T", rm)
 	}
@@ -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) {
@@ -0,0 +1,558 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package asm
+
+import "fmt"
+
+// This file implements EVEX (AVX-512) instruction encoding: the four-byte
+// EVEX prefix with 5-bit vector register fields (Z0–Z31, X/Y 16–31), the
+// compressed disp8×N displacement, and the operand shapes the go-flac
+// AVX-512 kernels use.  Masking ({k}) and zeroing ({z}) are not supported —
+// the kernels do not use them.  K-register operands (mask destinations,
+// KMOVW, KTESTW) are.
+
+// evexSpec describes one EVEX instruction's encoding parameters.  The form
+// field reuses the vexForm shapes, which carry over unchanged.
+type evexSpec struct {
+	mapSel  int // 1 = 0F, 2 = 0F38, 3 = 0F3A
+	opcode  byte
+	w       int
+	pp      int     // 0 = none, 1 = 66, 2 = F3, 3 = F2
+	opdigit int     // ModRM.reg /digit, or -1 when reg is a register
+	form    vexForm // vexNDS3, vexRM, vexShiftImm, vexNDS3Imm, vexExtract
+	n       [3]int  // disp8×N multiplier per vector length (128/256/512)
+}
+
+// evexTable maps an upper-case mnemonic to its EVEX encoding.  Mnemonics
+// that also have a VEX form (VPADDD, VMOVUPD, …) are dispatched here only
+// when an operand demands EVEX (a ZMM or K register); EVEX-only mnemonics
+// (VPXORD, VALIGND, …) always encode through this table.  The N multipliers
+// are taken from the Go assembler's opcode tables, which are authoritative
+// for byte-for-byte agreement.
+var evexTable = map[string]evexSpec{
+	// EVEX.128/256/512.66.0F — integer arithmetic / logic, NDS form.
+	"VPADDD":      {1, 0xFE, 0, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VPADDQ":      {1, 0xD4, 1, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VPSUBD":      {1, 0xFA, 0, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VPSUBQ":      {1, 0xFB, 1, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VPUNPCKLDQ":  {1, 0x62, 0, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VPUNPCKHDQ":  {1, 0x6A, 0, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VPXORD":      {1, 0xEF, 0, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VPXORQ":      {1, 0xEF, 1, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VPCMPEQD":    {1, 0x76, 0, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VFMADD231PD": {2, 0xB8, 1, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+
+	// EVEX.128/256/512.66.0F.W1 — packed double arithmetic.
+	"VADDPD": {1, 0x58, 1, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VMULPD": {1, 0x59, 1, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+
+	// EVEX.512.66.0F3A — align (NDS + imm8).
+	"VALIGND": {3, 0x03, 0, 1, -1, vexNDS3Imm, [3]int{16, 32, 64}},
+
+	// EVEX.128/256/512.66.0F — immediate shift (VPSRAD /4).
+	"VPSRAD": {1, 0x72, 0, 1, 4, vexShiftImm, [3]int{16, 32, 64}},
+	// EVEX.128/256/512.66.0F.W1 — variable shift with an XMM count (VPSRAQ;
+	// the W bit distinguishes it from VPSRAD's E2 form).
+	"VPSRAQ": {1, 0xE2, 1, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+
+	// EVEX.128/256/512.F3.0F.W1 — signed qword to packed double (reg=dst,
+	// rm=src, no vvvv).
+	"VCVTQQ2PD": {1, 0xE6, 1, 2, -1, vexRM, [3]int{16, 32, 64}},
+	// EVEX.128/256/512.66.0F38.W0 — sign-extend dwords to qwords; the memory
+	// operand is the narrow source, so disp8×N follows its size (8/16/32 for
+	// the xmm/ymm/zmm destination lengths).
+	"VPMOVSXDQ": {2, 0x25, 0, 1, -1, vexRM, [3]int{8, 16, 32}},
+
+	// EVEX.512.66.0F3A.W1 — lane extract (reg=ZMM source, rm=YMM/memory
+	// destination, imm8).
+	"VEXTRACTI64X4": {3, 0x3B, 1, 1, -1, vexExtract, [3]int{0, 0, 32}},
+	"VEXTRACTF64X4": {3, 0x1B, 1, 1, -1, vexExtract, [3]int{0, 0, 32}},
+
+	// EVEX.66.0F38 — more integer NDS forms (W distinguishes D/Q).
+	"VPMULLD": {2, 0x40, 0, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VPMULLQ": {2, 0x40, 1, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VPERMD":  {2, 0x36, 0, 1, -1, vexNDS3, [3]int{0, 32, 64}},
+
+	// EVEX.66.0F — immediate shift (VPSLLD /6).
+	"VPSLLD": {1, 0x72, 0, 1, 6, vexShiftImm, [3]int{16, 32, 64}},
+
+	// EVEX.F3.0F38.W0 — narrowing stores: reg = wide source, rm = narrow
+	// destination (VPMOVDW dword→word, VPMOVQD qword→dword).
+	"VPMOVDW": {2, 0x33, 0, 2, -1, vexRMRev, [3]int{8, 16, 32}},
+	"VPMOVQD": {2, 0x35, 0, 2, -1, vexRMRev, [3]int{8, 16, 32}},
+}
+
+// evexBcastSpec describes an EVEX broadcast (VPBROADCASTD/Q): the opcode
+// depends on the source kind — a GPR source uses opReg, a memory source uses
+// opMem with a disp8×N of n.
+type evexBcastSpec struct {
+	mapSel int
+	opReg  byte
+	opMem  byte
+	w      int
+	n      int
+}
+
+var evexBcastTable = map[string]evexBcastSpec{
+	// EVEX.128/256/512.66.0F38 — broadcast a dword/qword to all lanes.
+	"VPBROADCASTD": {2, 0x7C, 0x58, 0, 4},
+	"VPBROADCASTQ": {2, 0x7C, 0x59, 1, 8},
+}
+
+// evexMoveSpec describes an EVEX move (load and store opcodes, like the VEX
+// move table).
+type evexMoveSpec struct {
+	mapSel int
+	pp     int
+	load   byte // r/m → vector
+	store  byte // vector → r/m
+	w      int
+	n      [3]int
+}
+
+// evexMoveTable maps an upper-case EVEX move mnemonic to its encoding.
+var evexMoveTable = map[string]evexMoveSpec{
+	// EVEX.128/256/512.F3.0F.W0 — unaligned integer move.
+	"VMOVDQU32": {1, 2, 0x6F, 0x7F, 0, [3]int{16, 32, 64}},
+	// EVEX.128/256/512.F3.0F.W1 — unaligned qword move.
+	"VMOVDQU64": {1, 2, 0x6F, 0x7F, 1, [3]int{16, 32, 64}},
+	// EVEX.128/256/512.66.0F.W1 — unaligned packed double move.
+	"VMOVUPD": {1, 1, 0x10, 0x11, 1, [3]int{16, 32, 64}},
+}
+
+// isEvex reports whether the mnemonic has an EVEX encoding we handle.
+func isEvex(mnemUpper string) bool {
+	if _, ok := evexTable[mnemUpper]; ok {
+		return true
+	}
+	if _, ok := evexBcastTable[mnemUpper]; ok {
+		return true
+	}
+	_, ok := evexMoveTable[mnemUpper]
+	return ok
+}
+
+// evexRequired reports whether the operands force the EVEX encoding of a
+// mnemonic that also has a VEX form: ZMM and K registers do, and so do
+// register indices 16–31, which only EVEX can represent (X16–Y31 exist
+// solely under AVX-512).
+func evexRequired(upper string, ops []Operand) bool {
+	_, inVex := vexTable[upper]
+	_, inVexMove := vexMoveTable[upper]
+	if !inVex && !inVexMove {
+		return true // EVEX-only mnemonic
+	}
+	for _, op := range ops {
+		if r, ok := op.(Reg); ok && (r.size == 64 || r.mask || (r.isVec() && r.idx >= 16)) {
+			return true
+		}
+	}
+	return false
+}
+
+// encodeEvex encodes an EVEX instruction with operands in Plan 9 order.
+func (e *enc) encodeEvex(mnemUpper string, ops []Operand) error {
+	if bs, ok := evexBcastTable[mnemUpper]; ok {
+		return e.encodeEvexBcast(bs, ops)
+	}
+	if ms, ok := evexMoveTable[mnemUpper]; ok {
+		return e.encodeEvexMove(mnemUpper, ms, ops)
+	}
+	spec, ok := evexTable[mnemUpper]
+	if !ok {
+		return fmt.Errorf("unsupported instruction %q for ZMM/K operands", mnemUpper)
+	}
+	switch spec.form {
+	case vexNDS3:
+		return e.encodeEvexNDS3(spec, ops)
+	case vexRM:
+		return e.encodeEvexRM(spec, ops)
+	case vexRMRev:
+		return e.encodeEvexRMRev(spec, ops)
+	case vexShiftImm:
+		return e.encodeEvexShiftImm(spec, ops)
+	case vexNDS3Imm:
+		return e.encodeEvexNDS3Imm(spec, ops)
+	case vexExtract:
+		return e.encodeEvexExtract(spec, ops)
+	}
+	return fmt.Errorf("unhandled EVEX form for %s", mnemUpper)
+}
+
+// encodeEvexNDS3 encodes the three-operand NDS form: OP src2, src1, dst.  The
+// destination may be an opmask register (VPCMPEQD), in which case the vector
+// length comes from the sources.
+func (e *enc) encodeEvexNDS3(spec evexSpec, ops []Operand) error {
+	if len(ops) != 3 {
+		return fmt.Errorf("EVEX NDS instruction expects 3 operands, got %d", len(ops))
+	}
+	src2, src1, dst := ops[0], ops[1], ops[2]
+	dstReg, ok := dst.(Reg)
+	if !ok || (!dstReg.isVec() && !dstReg.mask) {
+		return fmt.Errorf("EVEX destination must be a vector or mask register")
+	}
+	vvvvReg, ok := src1.(Reg)
+	if !ok || !vvvvReg.isVec() {
+		return fmt.Errorf("EVEX vvvv operand must be a vector register")
+	}
+	ll := dstReg.vecLenBit()
+	if dstReg.mask {
+		ll = vvvvReg.vecLenBit()
+		if r, ok := src2.(Reg); ok && r.isVec() {
+			ll = r.vecLenBit()
+		}
+	}
+	return e.emitEvexFields(spec, ll, dstReg.idx, vvvvReg.idx, src2)
+}
+
+// encodeEvexRM encodes the two-operand form: OP src, dst (reg=dst, rm=src,
+// no vvvv), e.g. VCVTQQ2PD.
+func (e *enc) encodeEvexRM(spec evexSpec, ops []Operand) error {
+	if len(ops) != 2 {
+		return fmt.Errorf("EVEX two-operand instruction expects 2 operands, got %d", len(ops))
+	}
+	src, dst := ops[0], ops[1]
+	dstReg, ok := dst.(Reg)
+	if !ok || !dstReg.isVec() {
+		return fmt.Errorf("EVEX destination must be a vector register")
+	}
+	return e.emitEvexFields(spec, dstReg.vecLenBit(), dstReg.idx, -1, src)
+}
+
+// encodeEvexShiftImm encodes an immediate shift: OP $imm, src, dst
+// (ModRM.reg = /digit, vvvv = dst, rm = src, imm8), e.g. VPSRAD $31, Z3, Z5.
+func (e *enc) encodeEvexShiftImm(spec evexSpec, ops []Operand) error {
+	if len(ops) != 3 {
+		return fmt.Errorf("EVEX shift expects 3 operands ($imm, src, dst), got %d", len(ops))
+	}
+	imm, src, dst := ops[0], ops[1], ops[2]
+	immVal, ok := imm.(Imm)
+	if !ok {
+		return fmt.Errorf("shift count must be an immediate")
+	}
+	srcReg, ok := src.(Reg)
+	if !ok || !srcReg.isVec() {
+		return fmt.Errorf("shift source must be a vector register")
+	}
+	dstReg, ok := dst.(Reg)
+	if !ok || !dstReg.isVec() {
+		return fmt.Errorf("shift destination must be a vector register")
+	}
+	immByte, err := imm8(int64(immVal))
+	if err != nil {
+		return err
+	}
+	if err := e.emitEvexFields(spec, dstReg.vecLenBit(), spec.opdigit, dstReg.idx, srcReg); err != nil {
+		return err
+	}
+	e.out = append(e.out, immByte)
+	return nil
+}
+
+// encodeEvexNDS3Imm encodes OP $imm, src2, src1, dst (reg=dst, vvvv=src1,
+// rm=src2, imm8), e.g. VALIGND.
+func (e *enc) encodeEvexNDS3Imm(spec evexSpec, ops []Operand) error {
+	if len(ops) != 4 {
+		return fmt.Errorf("instruction expects 4 operands ($imm, src2, src1, dst), got %d", len(ops))
+	}
+	imm, src2, src1, dst := ops[0], ops[1], ops[2], ops[3]
+	immVal, ok := imm.(Imm)
+	if !ok {
+		return fmt.Errorf("shuffle control must be an immediate")
+	}
+	dstReg, ok := dst.(Reg)
+	if !ok || !dstReg.isVec() {
+		return fmt.Errorf("destination must be a vector register")
+	}
+	vvvvReg, ok := src1.(Reg)
+	if !ok || !vvvvReg.isVec() {
+		return fmt.Errorf("second source must be a vector register")
+	}
+	immByte, err := imm8(int64(immVal))
+	if err != nil {
+		return err
+	}
+	if err := e.emitEvexFields(spec, dstReg.vecLenBit(), dstReg.idx, vvvvReg.idx, src2); err != nil {
+		return err
+	}
+	e.out = append(e.out, immByte)
+	return nil
+}
+
+// encodeEvexExtract encodes OP $imm, zsrc, ydst (reg=ZMM source, rm=YMM/memory
+// destination, imm8), e.g. VEXTRACTI64X4.
+func (e *enc) encodeEvexExtract(spec evexSpec, ops []Operand) error {
+	if len(ops) != 3 {
+		return fmt.Errorf("extract expects 3 operands ($imm, zsrc, ydst), got %d", len(ops))
+	}
+	imm, src, dst := ops[0], ops[1], ops[2]
+	immVal, ok := imm.(Imm)
+	if !ok {
+		return fmt.Errorf("extract lane must be an immediate")
+	}
+	srcReg, ok := src.(Reg)
+	if !ok || !srcReg.isVec() {
+		return fmt.Errorf("extract source must be a vector register")
+	}
+	immByte, err := imm8(int64(immVal))
+	if err != nil {
+		return err
+	}
+	if err := e.emitEvexFields(spec, srcReg.vecLenBit(), srcReg.idx, -1, dst); err != nil {
+		return err
+	}
+	e.out = append(e.out, immByte)
+	return nil
+}
+
+// encodeEvexMove encodes a two-operand EVEX move; a vector→vector move uses
+// the store-form opcode (reg = source, rm = destination), matching the Go
+// assembler.
+func (e *enc) encodeEvexMove(mnem string, ms evexMoveSpec, ops []Operand) error {
+	if len(ops) != 2 {
+		return fmt.Errorf("EVEX move expects 2 operands, got %d", len(ops))
+	}
+	src, dst := ops[0], ops[1]
+	srcReg, srcIsVec := vecReg(src)
+	dstReg, dstIsVec := vecReg(dst)
+
+	op := ms.store
+	var reg Reg
+	var rm Operand
+	switch {
+	case srcIsVec && dstIsVec:
+		reg, rm = srcReg, dst
+	case srcIsVec:
+		if !memOperand(dst) {
+			return fmt.Errorf("%s: invalid destination operand", mnem)
+		}
+		reg, rm = srcReg, dst
+	case dstIsVec:
+		if !memOperand(src) {
+			return fmt.Errorf("%s: invalid source operand", mnem)
+		}
+		op = ms.load
+		reg, rm = dstReg, src
+	default:
+		return fmt.Errorf("%s needs a vector register operand", mnem)
+	}
+	spec := evexSpec{mapSel: ms.mapSel, opcode: op, w: ms.w, pp: ms.pp, opdigit: -1, n: ms.n}
+	return e.emitEvexFields(spec, reg.vecLenBit(), reg.idx, -1, rm)
+}
+
+// memOperand reports whether op is a memory reference (including a
+// static-symbol reference).
+func memOperand(op Operand) bool {
+	switch op.(type) {
+	case Mem, sbMem:
+		return true
+	}
+	return false
+}
+
+// encodeEvexRMRev encodes the narrowing-store form: OP src, dst with the wide
+// source in the reg field and the narrow destination in r/m (VPMOVDW/QD).
+func (e *enc) encodeEvexRMRev(spec evexSpec, ops []Operand) error {
+	if len(ops) != 2 {
+		return fmt.Errorf("EVEX store instruction expects 2 operands, got %d", len(ops))
+	}
+	src, dst := ops[0], ops[1]
+	srcReg, ok := src.(Reg)
+	if !ok || !srcReg.isVec() {
+		return fmt.Errorf("EVEX source must be a vector register")
+	}
+	return e.emitEvexFields(spec, srcReg.vecLenBit(), srcReg.idx, -1, dst)
+}
+
+// encodeEvexBcast encodes VPBROADCASTD/Q: OP src, dst with the GPR or memory
+// source broadcast to every lane of the vector destination.
+func (e *enc) encodeEvexBcast(bs evexBcastSpec, ops []Operand) error {
+	if len(ops) != 2 {
+		return fmt.Errorf("broadcast expects 2 operands, got %d", len(ops))
+	}
+	src, dst := ops[0], ops[1]
+	dstReg, ok := dst.(Reg)
+	if !ok || !dstReg.isVec() {
+		return fmt.Errorf("broadcast destination must be a vector register")
+	}
+	spec := evexSpec{mapSel: bs.mapSel, w: bs.w, pp: 1, opdigit: -1}
+	switch src.(type) {
+	case Mem, sbMem:
+		spec.opcode = bs.opMem
+		spec.n = [3]int{bs.n, bs.n, bs.n}
+	case Reg:
+		spec.opcode = bs.opReg
+	default:
+		return fmt.Errorf("broadcast source must be a register or memory")
+	}
+	return e.emitEvexFields(spec, dstReg.vecLenBit(), dstReg.idx, -1, src)
+}
+
+// emitEvexFields emits the EVEX prefix, opcode, ModR/M, SIB and displacement
+// (disp8×N compressed) for the given precomputed fields.  regIdx is the
+// unextended reg-field register index, or a /digit (0–7); vvvvIdx is the
+// vvvv register index, or -1 when unused.
+func (e *enc) emitEvexFields(spec evexSpec, ll, regIdx, vvvvIdx int, rm Operand) error {
+	if ll > 2 {
+		return fmt.Errorf("invalid vector length")
+	}
+	// reg-field extension bits (R̄, R'̄), inverted.
+	rBar, rPrimeBar := 1, 1
+	if regIdx&8 != 0 {
+		rBar = 0
+	}
+	if regIdx&16 != 0 {
+		rPrimeBar = 0
+	}
+	// vvvv (inverted) and its extension bit V'̄.
+	vBar, vPrimeBar := 15, 1
+	if vvvvIdx >= 0 {
+		vBar = 15 - (vvvvIdx & 15)
+		if vvvvIdx&16 != 0 {
+			vPrimeBar = 0
+		}
+	}
+
+	var modrm, sib int
+	var disp []byte
+	xBar, bBar := 1, 1
+	var sb *sbRef
+	switch r := rm.(type) {
+	case Reg:
+		// ModRM.mod = 11: rm[3] extends via B̄, rm[4] via X̄.
+		modrm = 0xC0 | (regIdx&7)<<3 | (r.idx & 7)
+		sib = -1
+		if r.idx&8 != 0 {
+			bBar = 0
+		}
+		if r.idx&16 != 0 {
+			xBar = 0
+		}
+	case Mem:
+		var err error
+		modrm, sib, disp, xBar, bBar, err = memComponentsEvex(regIdx&7, r, spec.n[ll])
+		if err != nil {
+			return err
+		}
+		// An indexed memory operand carries index[4] in V'̄ (Go folds it
+		// together with vvvv[4] into the same bit).
+		if r.HasIndex && r.Index.idx&16 != 0 {
+			vPrimeBar = 0
+		}
+	case sbMem:
+		// RIP-relative static-symbol reference; disp32 patched at link time
+		// (no disp8 scaling for RIP-relative addressing).
+		modrm = (regIdx&7)<<3 | 0x05
+		sib = -1
+		disp = le32(0)
+		sb = &sbRef{name: r.name, addend: r.addend}
+	default:
+		return fmt.Errorf("invalid EVEX r/m operand")
+	}
+
+	p0 := byte(rBar<<7 | xBar<<6 | bBar<<5 | rPrimeBar<<4 | spec.mapSel)
+	p1 := byte(spec.w<<7 | vBar<<3 | 1<<2 | spec.pp)
+	p2 := byte(ll<<5 | vPrimeBar<<3) // z = 0, b = 0, aaa = 0
+	e.out = append(e.out, 0x62, p0, p1, p2, spec.opcode, byte(modrm))
+	if sib >= 0 {
+		e.out = append(e.out, byte(sib))
+	}
+	if sb != nil {
+		e.patches = append(e.patches, encPatch{off: len(e.out), name: sb.name, addend: sb.addend})
+	}
+	e.out = append(e.out, disp...)
+	return nil
+}
+
+// memComponentsEvex computes the ModR/M byte (with the given reg field), the
+// SIB byte (-1 if none), the displacement bytes and the (inverted sense)
+// index/base extension bits for an EVEX memory operand.  The displacement is
+// compressed to disp8×N when it is a multiple of n and the quotient fits a
+// signed byte; otherwise a full disp32 is used.
+func memComponentsEvex(regField int, m Mem, n int) (modrm, sib int, disp []byte, xBar, bBar int, err error) {
+	sib = -1
+	xBar, bBar = 1, 1 // inverted bits: 1 = no extension
+	if !m.HasBase && !m.HasIndex {
+		return regField<<3 | 0x05, -1, le32(m.Disp), 1, 1, nil // RIP-relative
+	}
+
+	needSIB := m.HasIndex || (m.HasBase && m.Base.idx&7 == 4)
+
+	var mod int
+	switch {
+	case !m.HasBase:
+		mod = 0
+		disp = le32(m.Disp)
+	case m.Base.idx&7 == 5 && m.Disp == 0:
+		mod = 1
+		disp = []byte{0}
+	case m.Disp == 0:
+		mod = 0
+	case n > 0 && m.Disp%int64(n) == 0 && m.Disp/int64(n) >= -128 && m.Disp/int64(n) <= 127:
+		mod = 1
+		disp = []byte{byte(int8(m.Disp / int64(n)))}
+	default:
+		mod = 2
+		disp = le32(m.Disp)
+	}
+
+	if needSIB {
+		idxField := 4 // 100 = no index
+		if m.HasIndex {
+			idxField = m.Index.idx & 7
+			if m.Index.idx&8 != 0 {
+				xBar = 0
+			}
+		}
+		baseField := 5 // 101 = no base (with mod=00 → disp32)
+		if m.HasBase {
+			baseField = m.Base.idx & 7
+			if m.Base.idx&8 != 0 {
+				bBar = 0
+			}
+		}
+		return mod<<6 | regField<<3 | 0x04, scaleBits(m.Scale)<<6 | idxField<<3 | baseField, disp, xBar, bBar, nil
+	}
+
+	if m.Base.idx&8 != 0 {
+		bBar = 0
+	}
+	return mod<<6 | regField<<3 | (m.Base.idx & 7), -1, disp, 1, bBar, nil
+}
+
+// encodeKmovw encodes KMOVW, whose opcode depends on the operand direction:
+// 90 (k/mem → K), 91 (K → mem), 92 (GPR → K), 93 (K → GPR); k → k uses 90.
+func (e *enc) encodeKmovw(ops []Operand) error {
+	if len(ops) != 2 {
+		return fmt.Errorf("KMOVW expects 2 operands, got %d", len(ops))
+	}
+	src, dst := ops[0], ops[1]
+	srcReg, srcIsReg := src.(Reg)
+	dstReg, dstIsReg := dst.(Reg)
+	srcK := srcIsReg && srcReg.mask
+	dstK := dstIsReg && dstReg.mask
+	spec := vexSpec{mapSel: 1, w: 0, pp: 0, opdigit: -1}
+	switch {
+	case srcK && dstK:
+		spec.opcode = 0x90 // k ← k: reg = dst, rm = src
+		return e.emitVexFields(spec, 0, dstReg.idx&7, 0, 15, src)
+	case srcK && dstIsReg:
+		spec.opcode = 0x93 // GPR ← k: reg = dst, rm = src
+		rBit := 0
+		if dstReg.idx >= 8 {
+			rBit = 1
+		}
+		return e.emitVexFields(spec, 0, dstReg.idx&7, rBit, 15, src)
+	case srcK:
+		if _, ok := dst.(Mem); !ok {
+			return fmt.Errorf("KMOVW: invalid destination operand")
+		}
+		spec.opcode = 0x91 // mem ← k: reg = src, rm = dst
+		return e.emitVexFields(spec, 0, srcReg.idx&7, 0, 15, dst)
+	case dstK:
+		spec.opcode = 0x92 // k ← GPR/mem: reg = dst, rm = src
+		return e.emitVexFields(spec, 0, dstReg.idx&7, 0, 15, src)
+	}
+	return fmt.Errorf("KMOVW requires a K register operand")
+}
@@ -0,0 +1,225 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package asm
+
+import (
+	"os"
+	"strings"
+	"testing"
+
+	"golang.org/x/arch/x86/x86asm"
+
+	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+)
+
+// TestEvexGroundTruth checks the EVEX (AVX-512) encodings byte for byte
+// against machine code extracted from the Go toolchain's assembly of the
+// same instructions, covering every operand shape the go-flac AVX-512
+// kernels use: NDS arithmetic, immediate and variable shifts, shuffles with
+// an immediate, lane extracts, narrowing stores, broadcasts from a GPR or
+// memory, mask destinations, mask moves, disp8×N compression and the 5-bit
+// register fields (X/Y 16–31, Z 0–31).
+func TestEvexGroundTruth(t *testing.T) {
+	cases := []struct {
+		name string
+		mnem string
+		ops  []Operand
+		want string
+	}{
+		// NDS integer arithmetic / logic.
+		{"VPXORD Z12,Z12,Z12", "VPXORD", []Operand{vreg(t, "Z12"), vreg(t, "Z12"), vreg(t, "Z12")}, "62511d48efe4"},
+		{"VPXORQ Z8,Z9,Z10", "VPXORQ", []Operand{vreg(t, "Z8"), vreg(t, "Z9"), vreg(t, "Z10")}, "6251b548efd0"},
+		{"VPADDD Z1,Z0,Z0", "VPADDD", []Operand{vreg(t, "Z1"), vreg(t, "Z0"), vreg(t, "Z0")}, "62f17d48fec1"},
+		{"VPSUBQ Z8,Z11,Z11", "VPSUBQ", []Operand{vreg(t, "Z8"), vreg(t, "Z11"), vreg(t, "Z11")}, "6251a548fbd8"},
+		{"VPUNPCKLDQ Z5,Z3,Z6", "VPUNPCKLDQ", []Operand{vreg(t, "Z5"), vreg(t, "Z3"), vreg(t, "Z6")}, "62f1654862f5"},
+		{"VPUNPCKHDQ Z5,Z3,Z7", "VPUNPCKHDQ", []Operand{vreg(t, "Z5"), vreg(t, "Z3"), vreg(t, "Z7")}, "62f165486afd"},
+		{"VPMULLQ Z9,Z10,Z10", "VPMULLQ", []Operand{vreg(t, "Z9"), vreg(t, "Z10"), vreg(t, "Z10")}, "6252ad4840d1"},
+		{"VPMULLD Z13,Z11,Z2", "VPMULLD", []Operand{vreg(t, "Z13"), vreg(t, "Z11"), vreg(t, "Z2")}, "62d2254840d5"},
+		{"VPERMD Z0,Z15,Z8", "VPERMD", []Operand{vreg(t, "Z0"), vreg(t, "Z15"), vreg(t, "Z8")}, "6272054836c0"},
+		// Packed-double arithmetic (EVEX forms carry W=1).
+		{"VADDPD Z11,Z10,Z10", "VADDPD", []Operand{vreg(t, "Z11"), vreg(t, "Z10"), vreg(t, "Z10")}, "6251ad4858d3"},
+		{"VMULPD Z13,Z12,Z12", "VMULPD", []Operand{vreg(t, "Z13"), vreg(t, "Z12"), vreg(t, "Z12")}, "62519d4859e5"},
+		{"VFMADD231PD Z14,Z12,Z10", "VFMADD231PD", []Operand{vreg(t, "Z14"), vreg(t, "Z12"), vreg(t, "Z10")}, "62529d48b8d6"},
+		// Align (NDS + imm8).
+		{"VALIGND $12,Z12,Z0,Z1", "VALIGND", []Operand{Imm(12), vreg(t, "Z12"), vreg(t, "Z0"), vreg(t, "Z1")}, "62d37d4803cc0c"},
+		{"VALIGND $15,Z9,Z0,Z1", "VALIGND", []Operand{Imm(15), vreg(t, "Z9"), vreg(t, "Z0"), vreg(t, "Z1")}, "62d37d4803c90f"},
+		// Shifts: immediate (/digit) and variable (XMM count).
+		{"VPSRAD $31,Z3,Z5", "VPSRAD", []Operand{Imm(31), vreg(t, "Z3"), vreg(t, "Z5")}, "62f1554872e31f"},
+		{"VPSLLD $1,Z3,Z4", "VPSLLD", []Operand{Imm(1), vreg(t, "Z3"), vreg(t, "Z4")}, "62f15d4872f301"},
+		{"VPSRAQ X31,Z8,Z8", "VPSRAQ", []Operand{vreg(t, "X31"), vreg(t, "Z8"), vreg(t, "Z8")}, "6211bd48e2c7"},
+		// Mask destinations (the K register occupies the reg field).
+		{"VPCMPEQD Z0,Z3,K1", "VPCMPEQD", []Operand{vreg(t, "Z0"), vreg(t, "Z3"), vreg(t, "K1")}, "62f1654876c8"},
+		{"VPCMPEQD Y30,Y11,K1", "VPCMPEQD", []Operand{vreg(t, "Y30"), vreg(t, "Y11"), vreg(t, "K1")}, "6291252876ce"},
+		// Mask moves and test (VEX-encoded).
+		{"KMOVW K1,CX", "KMOVW", []Operand{vreg(t, "K1"), CX}, "c5f893c9"},
+		{"KMOVW K1,R12", "KMOVW", []Operand{vreg(t, "K1"), vreg(t, "R12")}, "c57893e1"},
+		{"KTESTW K1,K1", "KTESTW", []Operand{vreg(t, "K1"), vreg(t, "K1")}, "c5f899c9"},
+		// Moves, incl. disp8×N (64 for a 512-bit operand).
+		{"VMOVDQU32 (SI)(R15*4),Z3", "VMOVDQU32", []Operand{Idx(SI, vreg(t, "R15"), 4, 0, 64), vreg(t, "Z3")}, "62b17e486f1cbe"},
+		{"VMOVDQU32 4(SI)(AX*1),Z4", "VMOVDQU32", []Operand{Idx(SI, AX, 1, 4, 64), vreg(t, "Z4")}, "62f17e486fa40604000000"},
+		{"VMOVDQU32 16(SI)(R15*4),Z4", "VMOVDQU32", []Operand{Idx(SI, vreg(t, "R15"), 4, 16, 64), vreg(t, "Z4")}, "62b17e486fa4be10000000"},
+		{"VMOVDQU32 Z0,4(SI)(AX*1)", "VMOVDQU32", []Operand{vreg(t, "Z0"), Idx(SI, AX, 1, 4, 64)}, "62f17e487f840604000000"},
+		{"VMOVDQU32 Z3,(DI)(R15*4)", "VMOVDQU32", []Operand{vreg(t, "Z3"), Idx(DI, vreg(t, "R15"), 4, 0, 64)}, "62b17e487f1cbf"},
+		// VMOVDQU64 — the W1 qword variant.
+		{"VMOVDQU64 (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"},
+		{"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).
+		{"VCVTQQ2PD Z12,Z12", "VCVTQQ2PD", []Operand{vreg(t, "Z12"), vreg(t, "Z12")}, "6251fe48e6e4"},
+		{"VCVTQQ2PD X13,X13", "VCVTQQ2PD", []Operand{vreg(t, "X13"), vreg(t, "X13")}, "6251fe08e6ed"},
+		{"VPMOVSXDQ 32(SI),Z12", "VPMOVSXDQ", []Operand{Ptr(SI, 32, 32), vreg(t, "Z12")}, "62727d48256601"},
+		{"VPMOVDW Z0,Y0", "VPMOVDW", []Operand{vreg(t, "Z0"), vreg(t, "Y0")}, "62f27e4833c0"},
+		{"VPMOVQD Z11,Y11", "VPMOVQD", []Operand{vreg(t, "Z11"), vreg(t, "Y11")}, "62527e4835db"},
+		// Lane extracts.
+		{"VEXTRACTI64X4 $1,Z8,Y9", "VEXTRACTI64X4", []Operand{Imm(1), vreg(t, "Z8"), vreg(t, "Y9")}, "6253fd483bc101"},
+		{"VEXTRACTF64X4 $1,Z10,Y11", "VEXTRACTF64X4", []Operand{Imm(1), vreg(t, "Z10"), vreg(t, "Y11")}, "6253fd481bd301"},
+		// Broadcasts: GPR source (0x7C) vs memory source (0x58/0x59, disp8×4/8).
+		{"VPBROADCASTD AX,Z15", "VPBROADCASTD", []Operand{AX, vreg(t, "Z15")}, "62727d487cf8"},
+		{"VPBROADCASTD (SI),Z8", "VPBROADCASTD", []Operand{Ptr(SI, 0, 4), vreg(t, "Z8")}, "62727d485806"},
+		{"VPBROADCASTD 4(SI),Z10", "VPBROADCASTD", []Operand{Ptr(SI, 4, 4), vreg(t, "Z10")}, "62727d48585601"},
+		{"VPBROADCASTQ R8,X31", "VPBROADCASTQ", []Operand{vreg(t, "R8"), vreg(t, "X31")}, "6242fd087cf8"},
+		{"VPBROADCASTQ AX,Z9", "VPBROADCASTQ", []Operand{AX, vreg(t, "Z9")}, "6272fd487cc8"},
+		// Register indices 16–31 exist only in EVEX encodings.
+		{"VPBROADCASTD AX,Y30", "VPBROADCASTD", []Operand{AX, vreg(t, "Y30")}, "62627d287cf0"},
+	}
+	for _, c := range cases {
+		want := strings.ReplaceAll(c.want, " ", "")
+		code, err := Encode(c.mnem, c.ops...)
+		if err != nil {
+			t.Errorf("%s: Encode: %v", c.name, err)
+			continue
+		}
+		if got := hexCompact(code); got != want {
+			t.Errorf("%s: bytes %s, want %s", c.name, got, want)
+			continue
+		}
+		inst, err := x86asm.Decode(code, 64)
+		if err != nil {
+			t.Errorf("%s: Decode(%x): %v", c.name, code, err)
+			continue
+		}
+		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 {
+			t.Errorf("%s: decoded as %s", c.name, inst.Op.String())
+		}
+	}
+}
+
+// TestEvexErrors checks the EVEX-specific error paths.
+func TestEvexErrors(t *testing.T) {
+	cases := []struct {
+		name string
+		mnem string
+		ops  []Operand
+	}{
+		{"NDS arity", "VPXORD", []Operand{vreg(t, "Z0"), vreg(t, "Z1")}},
+		{"KMOVW arity", "KMOVW", []Operand{vreg(t, "K1")}},
+		{"KMOVW no K", "KMOVW", []Operand{AX, CX}},
+		{"VMOVUPD Z gpr", "VMOVUPD", []Operand{AX, vreg(t, "Z1")}},
+		{"broadcast src", "VPBROADCASTD", []Operand{Imm(1), vreg(t, "Z1")}},
+		{"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")}},
+	}
+	for _, c := range cases {
+		if _, err := Encode(c.mnem, c.ops...); err == nil {
+			t.Errorf("%s: expected an error, got none", c.name)
+		}
+	}
+}
+
+// TestAssembleGoFlacAVX512Kernel assembles the whole production AVX-512
+// kernel — all functions plus the file-global idx16 constant — and checks
+// that the static-symbol load resolves to the right bytes in the image.
+// Skipped when the sibling repository is not checked out.
+func TestAssembleGoFlacAVX512Kernel(t *testing.T) {
+	path := "../../go-libraries/go-flac/avx512_amd64.s"
+	if _, err := os.Stat(path); err != nil {
+		t.Skip("go-libraries repository not present next to gasm-devkit")
+	}
+	src, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatal(err)
+	}
+	f, errs := parser.Parse(path, string(src))
+	if len(errs) > 0 {
+		t.Fatalf("parse: %v", errs)
+	}
+	img, err := AssembleFile(f)
+	if err != nil {
+		t.Fatalf("AssembleFile: %v", err)
+	}
+	if len(img.Funcs) != 10 {
+		t.Errorf("functions = %d, want 10", len(img.Funcs))
+	}
+
+	// idx16 as the DATA directives define it: dwords 1..16.
+	idx := make([]byte, 0, 64)
+	for i := 1; i <= 16; i++ {
+		idx = append(idx, byte(i), 0, 0, 0)
+	}
+	image := img.Bytes()
+	base := img.Symbols["idx16"]
+	if base == 0 {
+		t.Fatal("idx16 not laid out")
+	}
+	if got := image[base : base+64]; hexCompact(got) != hexCompact(idx) {
+		t.Errorf("idx16 contents %x, want %x", got, idx)
+	}
+
+	// The VMOVDQU32 idx16(SB), Z13 load (62 71 7e 48 6f 2d + rel32) must
+	// resolve to idx16 within the image.
+	loads := 0
+	for _, fn := range img.Funcs {
+		code := img.Code[fn.Offset : fn.Offset+fn.Size]
+		pat := []byte{0x62, 0x71, 0x7e, 0x48, 0x6f, 0x2d}
+		for pos := 0; ; {
+			i := indexOf(code[pos:], pat)
+			if i < 0 {
+				break
+			}
+			i += pos
+			rel := int32(uint32(code[i+6]) | uint32(code[i+7])<<8 | uint32(code[i+8])<<16 | uint32(code[i+9])<<24)
+			target := fn.Offset + i + 10 + int(rel)
+			if target != base {
+				t.Errorf("%s: idx16 load at +%d targets 0x%x, want 0x%x", fn.Name, i, target, base)
+			}
+			loads++
+			pos = i + 10
+		}
+	}
+	if loads != 1 {
+		t.Errorf("idx16 loads found = %d, want 1", loads)
+	}
+}
+
+// hexCompact renders bytes as a lowercase hex string without separators.
+func hexCompact(b []byte) string {
+	const hexdig = "0123456789abcdef"
+	out := make([]byte, len(b)*2)
+	for i, c := range b {
+		out[i*2] = hexdig[c>>4]
+		out[i*2+1] = hexdig[c&0xf]
+	}
+	return string(out)
+}
+
+// indexOf returns the index of the first occurrence of pat in b, or -1.
+func indexOf(b, pat []byte) int {
+	for i := 0; i+len(pat) <= len(b); i++ {
+		j := 0
+		for j < len(pat) && b[i+j] == pat[j] {
+			j++
+		}
+		if j == len(pat) {
+			return i
+		}
+	}
+	return -1
+}
@@ -78,6 +78,17 @@ func (e *enc) encodeMov(ops []Operand, size int) error {
 		}
 		return e.emit(i)

+	case sbMem:
+		if !dstIsReg {
+			return fmt.Errorf("MOV: two memory operands")
+		}
+		// MOV r, r/m: reg=dst, rm=src(static symbol).
+		i := newInstr(size, []byte{movRR(size)})
+		if err := setRM(i, dstReg, src, size); err != nil {
+			return err
+		}
+		return e.emit(i)
+
 	case Imm:
 		if dstIsReg {
 			// MOV r, imm: 0xB0+reg (8-bit) / 0xB8+reg (16/32/64, imm64 for Q).
@@ -280,12 +291,13 @@ func (e *enc) encodeLea(ops []Operand, size int) error {
 	if !ok {
 		return fmt.Errorf("LEA: destination must be a register")
 	}
-	mem, ok := src.(Mem)
-	if !ok {
+	switch src.(type) {
+	case Mem, sbMem:
+	default:
 		return fmt.Errorf("LEA: source must be a memory operand")
 	}
 	i := newInstr(size, []byte{0x8D})
-	if err := setRM(i, dstReg, mem, size); err != nil {
+	if err := setRM(i, dstReg, src, size); err != nil {
 		return err
 	}
 	return e.emit(i)
@@ -649,6 +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
@@ -0,0 +1,162 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package asm
+
+import (
+	"fmt"
+
+	"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
+// references are encoded RIP-relative and resolved within the image, so the
+// bytes are self-consistent and executable at any base address.
+type Image struct {
+	Code    []byte         // concatenated function bodies
+	Data    []byte         // static data section
+	Funcs   []FuncLayout   // function positions, in source order
+	Symbols map[string]int // static symbol → byte offset within the image
+}
+
+// FuncLayout describes one assembled function within an Image.
+type FuncLayout struct {
+	Name   string
+	Offset int // start offset within the image (== offset within Code)
+	Size   int
+	Labels map[string]int // local labels, function-relative
+}
+
+// Bytes returns the whole image: code, then data.
+func (img *Image) Bytes() []byte {
+	out := make([]byte, 0, len(img.Code)+len(img.Data))
+	out = append(out, img.Code...)
+	return append(out, img.Data...)
+}
+
+// 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
+// resolved against that layout.  External (non-file-local) symbol references
+// are rejected: they need object-file emission.
+func AssembleFile(f *ast.File) (*Image, error) {
+	syms, order, err := collectData(f)
+	if err != nil {
+		return nil, err
+	}
+	known := make(map[string]bool, len(syms))
+	for name := range syms {
+		known[name] = true
+	}
+	link := &linkInfo{symbols: known}
+
+	img := &Image{Symbols: map[string]int{}}
+	type asmFunc struct {
+		name    string
+		patches []sbPatch
+	}
+	var funcs []asmFunc
+	for _, d := range f.Decls {
+		t, ok := d.(*ast.Text)
+		if !ok {
+			continue
+		}
+		code, patches, labels, err := assemble(t, link)
+		if err != nil {
+			return nil, fmt.Errorf("%s: %w", t.Name.Name, err)
+		}
+		img.Funcs = append(img.Funcs, FuncLayout{
+			Name:   t.Name.Name,
+			Offset: len(img.Code),
+			Size:   len(code),
+			Labels: labels,
+		})
+		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 {
+		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.Data = append(img.Data, syms[name]...)
+	}
+
+	// Resolve the RIP-relative displacements now that every address is known.
+	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)
+			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))
+		}
+	}
+	return img, nil
+}
+
+// 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) {
+	syms := map[string][]byte{}
+	var order []string
+	for _, d := range f.Decls {
+		switch dd := d.(type) {
+		case *ast.Globl:
+			if dd.Name == nil || dd.Name.Pseudo != "SB" {
+				continue
+			}
+			name := dd.Name.Name
+			if _, dup := syms[name]; dup {
+				return nil, 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)
+			order = append(order, name)
+
+		case *ast.Data:
+			if dd.Name == nil || dd.Name.Pseudo != "SB" {
+				continue
+			}
+			buf, ok := syms[dd.Name.Name]
+			if !ok {
+				return nil, 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)
+			}
+			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)
+			}
+			off := dd.Name.Offset
+			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))
+			}
+			v := dd.Value.Imm.Val
+			if dd.Value.Imm.Neg {
+				v = -v
+			}
+			for i := 0; i < w; i++ {
+				buf[off+int64(i)] = byte(v >> (8 * i))
+			}
+		}
+	}
+	return syms, order, nil
+}
+
+// align16 rounds n up to the next multiple of 16.
+func align16(n int) int {
+	return (n + 15) &^ 15
+}
@@ -0,0 +1,196 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package asm
+
+import (
+	"bytes"
+	"os"
+	"strings"
+	"testing"
+
+	"golang.org/x/arch/x86/x86asm"
+
+	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+)
+
+// TestAssembleFileStaticData checks the whole-image layout — code, padding
+// and the data section — and that the RIP-relative displacements of static
+// symbol loads resolve to the right bytes.
+func TestAssembleFileStaticData(t *testing.T) {
+	f, errs := parser.Parse("d_amd64.s", `
+#include "textflag.h"
+TEXT ·load(SB), NOSPLIT, $0
+	VMOVDQU mask<>(SB), X15
+	MOVL small<>(SB), AX
+	RET
+GLOBL mask<>(SB), RODATA, $16
+DATA mask<>+0(SB)/4, $0x80020100
+DATA mask<>+4(SB)/4, $0x80050403
+DATA mask<>+8(SB)/4, $0x80080706
+DATA mask<>+12(SB)/4, $0x800B0A09
+GLOBL small<>(SB), RODATA, $4
+DATA small<>+0(SB)/4, $0x1234
+`)
+	if len(errs) > 0 {
+		t.Fatalf("parse: %v", errs)
+	}
+	img, err := AssembleFile(f)
+	if err != nil {
+		t.Fatalf("AssembleFile: %v", err)
+	}
+
+	// Code (15 bytes) + 1 pad byte to align the data section to 16:
+	//   VMOVDQU mask<>(SB), X15   c5 7a 6f 3d 08 00 00 00   (disp = 16 − 8)
+	//   MOVL small<>(SB), AX      8b 05 12 00 00 00         (disp = 32 − 14)
+	//   RET                       c3
+	// Data: pad, mask (16 bytes), small (4 bytes).
+	want := "c57a6f3d080000008b0512000000c300" +
+		"000102800304058006070880090a0b80" +
+		"34120000"
+	if got := strings.ReplaceAll(hexBytes(img.Bytes()), " ", ""); got != want {
+		t.Errorf("image bytes:\n got  %s\n want %s", got, want)
+	}
+	if img.Symbols["mask"] != 16 || img.Symbols["small"] != 32 {
+		t.Errorf("symbol offsets = %v, want mask=16 small=32", img.Symbols)
+	}
+	if len(img.Funcs) != 1 || img.Funcs[0].Name != "load" || img.Funcs[0].Size != 15 {
+		t.Errorf("funcs = %+v", img.Funcs)
+	}
+}
+
+// TestAssembleFileErrors checks the static-symbol error paths.
+func TestAssembleFileErrors(t *testing.T) {
+	cases := []struct {
+		name string
+		src  string
+		want string // substring of the error
+	}{
+		{
+			"undefined symbol",
+			`
+#include "textflag.h"
+TEXT ·f(SB), NOSPLIT, $0
+	VMOVDQU nope<>(SB), X0
+	RET
+`,
+			"undefined symbol",
+		},
+		{
+			"DATA without GLOBL",
+			`
+#include "textflag.h"
+TEXT ·f(SB), NOSPLIT, $0
+	RET
+DATA orphan<>+0(SB)/4, $1
+`,
+			"no matching GLOBL",
+		},
+		{
+			"DATA exceeds size",
+			`
+#include "textflag.h"
+TEXT ·f(SB), NOSPLIT, $0
+	RET
+GLOBL tiny<>(SB), RODATA, $4
+DATA tiny<>+0(SB)/8, $1
+`,
+			"exceeds GLOBL size",
+		},
+		{
+			"DATA bad width",
+			`
+#include "textflag.h"
+TEXT ·f(SB), NOSPLIT, $0
+	RET
+GLOBL odd<>(SB), RODATA, $4
+DATA odd<>+0(SB)/3, $1
+`,
+			"invalid width",
+		},
+	}
+	for _, c := range cases {
+		f, errs := parser.Parse("e_amd64.s", c.src)
+		if len(errs) > 0 {
+			t.Fatalf("%s: parse: %v", c.name, errs)
+		}
+		if _, err := AssembleFile(f); err == nil || !strings.Contains(err.Error(), c.want) {
+			t.Errorf("%s: error %v, want substring %q", c.name, err, c.want)
+		}
+	}
+
+	// A static-symbol operand is unresolvable in single-function assembly.
+	fn := firstText(t, `
+#include "textflag.h"
+TEXT ·f(SB), NOSPLIT, $0
+	MOVQ x<>(SB), AX
+	RET
+GLOBL x<>(SB), RODATA, $8
+DATA x<>+0(SB)/4, $1
+`)
+	if _, _, err := Assemble(fn); err == nil || !strings.Contains(err.Error(), "file-level assembly") {
+		t.Errorf("single-function SB: error %v, want a file-level-assembly error", err)
+	}
+}
+
+// TestAssembleGoFlacAVX2Kernel assembles the whole production AVX2 kernel —
+// all functions plus the file-local mask24 constant — and checks that every
+// static-symbol load resolves to the right bytes in the image.  Skipped when
+// the sibling repository is not checked out.
+func TestAssembleGoFlacAVX2Kernel(t *testing.T) {
+	path := "../../go-libraries/go-flac/avx2_amd64.s"
+	if _, err := os.Stat(path); err != nil {
+		t.Skip("go-libraries repository not present next to gasm-devkit")
+	}
+	src, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatal(err)
+	}
+	f, errs := parser.Parse(path, string(src))
+	if len(errs) > 0 {
+		t.Fatalf("parse: %v", errs)
+	}
+	img, err := AssembleFile(f)
+	if err != nil {
+		t.Fatalf("AssembleFile: %v", err)
+	}
+	if len(img.Funcs) != 17 {
+		t.Errorf("functions = %d, want 17", len(img.Funcs))
+	}
+
+	// mask24 as the DATA directives define it.
+	mask := []byte{
+		0x00, 0x01, 0x02, 0x80, 0x03, 0x04, 0x05, 0x80,
+		0x06, 0x07, 0x08, 0x80, 0x09, 0x0a, 0x0b, 0x80,
+	}
+	image := img.Bytes()
+	if got := image[img.Symbols["mask24"] : img.Symbols["mask24"]+16]; !bytes.Equal(got, mask) {
+		t.Errorf("mask24 contents %x, want %x", got, mask)
+	}
+
+	// Every VMOVDQU mask24<>(SB), X15 (c5 7a 6f 3d + rel32, i.e. a VMOVDQU
+	// with a RIP-relative r/m) must land on the mask bytes within the image.
+	loads := 0
+	for _, fn := range img.Funcs {
+		code := img.Code[fn.Offset : fn.Offset+fn.Size]
+		for pc := 0; pc < len(code); {
+			inst, err := x86asm.Decode(code[pc:], 64)
+			if err != nil {
+				t.Fatalf("%s: decode at +%d: %v", fn.Name, pc, err)
+			}
+			// mod=00, rm=101 → RIP-relative.
+			if inst.Op == x86asm.VMOVDQU && inst.Len == 8 && code[pc+3]&0xC7 == 0x05 {
+				rel := int32(uint32(code[pc+4]) | uint32(code[pc+5])<<8 | uint32(code[pc+6])<<16 | uint32(code[pc+7])<<24)
+				target := fn.Offset + pc + 8 + int(rel)
+				if !bytes.Equal(image[target:target+16], mask) {
+					t.Errorf("%s: mask load at +%d lands on %x, want %x", fn.Name, pc, image[target:target+16], mask)
+				}
+				loads++
+			}
+			pc += inst.Len
+		}
+	}
+	if loads != 2 {
+		t.Errorf("mask loads found = %d, want 2", loads)
+	}
+}
@@ -41,3 +41,15 @@ func Idx(base, index Reg, scale int, disp int64, size int) Mem {
 func Rip(disp int64, size int) Mem {
 	return Mem{Disp: disp, Size: size}
 }
+
+// sbMem is a memory operand that references a static (SB) symbol.  It encodes
+// as a RIP-relative reference with a placeholder displacement; the encoder
+// records a patch site so the file-level layout can fill in the true rel32
+// once the symbol's address is known.
+type sbMem struct {
+	size   int
+	name   string // static symbol name (the GLOBL identifier)
+	addend int64  // byte offset within the symbol
+}
+
+func (sbMem) isOperand() {}
@@ -14,19 +14,24 @@ import "strings"
 // so the encoder keys off the register's index and lets the mnemonic supply the
 // size.  The high flag marks the legacy high-byte registers AH/CH/DH/BH, which
 // occupy indices 4–7 yet take no REX prefix, unlike SPL/BPL/SIL/DIL that share
-// those indices but require one.
+// those indices but require one.  The mask flag marks the AVX-512 opmask
+// registers K0–K7.
 type Reg struct {
 	idx  int
 	size int  // informational width implied by the name; the mnemonic decides
 	high bool // AH/CH/DH/BH
+	mask bool // K0–K7 opmask register
 }

-// Index returns the register number (0–15).
+// Index returns the register number (0–15 for GPRs, 0–31 for vectors).
 func (r Reg) Index() int { return r.idx }

 // Size returns the width in bytes implied by the register's name.
 func (r Reg) Size() int { return r.size }

+// IsMask reports whether r is an AVX-512 opmask register (K0–K7).
+func (r Reg) IsMask() bool { return r.mask }
+
 func (r Reg) isOperand() {}

 // needsREX reports whether this register forces a REX prefix at the given
@@ -41,45 +46,45 @@ func (r Reg) needsREX(opSize int) bool {

 // Register constants (the size is the width the name implies).
 var (
-	AL  = Reg{0, 1, false}
-	CL  = Reg{1, 1, false}
-	DL  = Reg{2, 1, false}
-	BL  = Reg{3, 1, false}
-	AH  = Reg{4, 1, true}
-	CH  = Reg{5, 1, true}
-	DH  = Reg{6, 1, true}
-	BH  = Reg{7, 1, true}
-	SPL = Reg{4, 1, false}
-	BPL = Reg{5, 1, false}
-	SIL = Reg{6, 1, false}
-	DIL = Reg{7, 1, false}
+	AL  = Reg{idx: 0, size: 1}
+	CL  = Reg{idx: 1, size: 1}
+	DL  = Reg{idx: 2, size: 1}
+	BL  = Reg{idx: 3, size: 1}
+	AH  = Reg{idx: 4, size: 1, high: true}
+	CH  = Reg{idx: 5, size: 1, high: true}
+	DH  = Reg{idx: 6, size: 1, high: true}
+	BH  = Reg{idx: 7, size: 1, high: true}
+	SPL = Reg{idx: 4, size: 1}
+	BPL = Reg{idx: 5, size: 1}
+	SIL = Reg{idx: 6, size: 1}
+	DIL = Reg{idx: 7, size: 1}

-	AX = Reg{0, 2, false}
-	CX = Reg{1, 2, false}
-	DX = Reg{2, 2, false}
-	BX = Reg{3, 2, false}
-	SP = Reg{4, 2, false}
-	BP = Reg{5, 2, false}
-	SI = Reg{6, 2, false}
-	DI = Reg{7, 2, false}
+	AX = Reg{idx: 0, size: 2}
+	CX = Reg{idx: 1, size: 2}
+	DX = Reg{idx: 2, size: 2}
+	BX = Reg{idx: 3, size: 2}
+	SP = Reg{idx: 4, size: 2}
+	BP = Reg{idx: 5, size: 2}
+	SI = Reg{idx: 6, size: 2}
+	DI = Reg{idx: 7, size: 2}

-	EAX = Reg{0, 4, false}
-	ECX = Reg{1, 4, false}
-	EDX = Reg{2, 4, false}
-	EBX = Reg{3, 4, false}
-	ESP = Reg{4, 4, false}
-	EBP = Reg{5, 4, false}
-	ESI = Reg{6, 4, false}
-	EDI = Reg{7, 4, false}
+	EAX = Reg{idx: 0, size: 4}
+	ECX = Reg{idx: 1, size: 4}
+	EDX = Reg{idx: 2, size: 4}
+	EBX = Reg{idx: 3, size: 4}
+	ESP = Reg{idx: 4, size: 4}
+	EBP = Reg{idx: 5, size: 4}
+	ESI = Reg{idx: 6, size: 4}
+	EDI = Reg{idx: 7, size: 4}

-	RAX = Reg{0, 8, false}
-	RCX = Reg{1, 8, false}
-	RDX = Reg{2, 8, false}
-	RBX = Reg{3, 8, false}
-	RSP = Reg{4, 8, false}
-	RBP = Reg{5, 8, false}
-	RSI = Reg{6, 8, false}
-	RDI = Reg{7, 8, false}
+	RAX = Reg{idx: 0, size: 8}
+	RCX = Reg{idx: 1, size: 8}
+	RDX = Reg{idx: 2, size: 8}
+	RBX = Reg{idx: 3, size: 8}
+	RSP = Reg{idx: 4, size: 8}
+	RBP = Reg{idx: 5, size: 8}
+	RSI = Reg{idx: 6, size: 8}
+	RDI = Reg{idx: 7, size: 8}
 )

 // regByName maps an assembly register name (case-insensitive) to a Reg.
@@ -91,28 +96,28 @@ func buildRegByName() map[string]Reg {
 	// 64-bit: RAX..RDI, R8..R15.
 	r64 := []string{"RAX", "RCX", "RDX", "RBX", "RSP", "RBP", "RSI", "RDI"}
 	for i, n := range r64 {
-		m[n] = Reg{i, 8, false}
+		m[n] = Reg{idx: i, size: 8}
 	}
 	for i := 8; i <= 15; i++ {
-		m["R"+itoa(i)] = Reg{i, 8, false}
+		m["R"+itoa(i)] = Reg{idx: i, size: 8}
 	}

 	// 32-bit: EAX..EDI, R8D..R15D.
 	e32 := []string{"EAX", "ECX", "EDX", "EBX", "ESP", "EBP", "ESI", "EDI"}
 	for i, n := range e32 {
-		m[n] = Reg{i, 4, false}
+		m[n] = Reg{idx: i, size: 4}
 	}
 	for i := 8; i <= 15; i++ {
-		m["R"+itoa(i)+"D"] = Reg{i, 4, false}
+		m["R"+itoa(i)+"D"] = Reg{idx: i, size: 4}
 	}

 	// 16-bit: AX..DI, R8W..R15W.
 	w16 := []string{"AX", "CX", "DX", "BX", "SP", "BP", "SI", "DI"}
 	for i, n := range w16 {
-		m[n] = Reg{i, 2, false}
+		m[n] = Reg{idx: i, size: 2}
 	}
 	for i := 8; i <= 15; i++ {
-		m["R"+itoa(i)+"W"] = Reg{i, 2, false}
+		m["R"+itoa(i)+"W"] = Reg{idx: i, size: 2}
 	}

 	// 8-bit: AL..BH, SPL..DIL, R8B..R15B.
@@ -124,25 +129,34 @@ func buildRegByName() map[string]Reg {
 		m[n] = r
 	}
 	for i := 8; i <= 15; i++ {
-		m["R"+itoa(i)+"B"] = Reg{i, 1, false}
+		m["R"+itoa(i)+"B"] = Reg{idx: i, size: 1}
 	}

-	// Vector: X0..X15 (128-bit, encoded size 16), Y0..Y15 (256-bit, size 32).
-	// Z (512-bit) and K (mask) registers arrive with EVEX/AVX-512 support.
-	for i := 0; i <= 15; i++ {
-		m["X"+itoa(i)] = Reg{i, 16, false}
-		m["Y"+itoa(i)] = Reg{i, 32, false}
+	// Vector: X0..X31 (128-bit, size 16), Y0..Y31 (256-bit, size 32),
+	// Z0..Z31 (512-bit, size 64).  Indices 16–31 are only encodable in EVEX
+	// (AVX-512) instructions; the encoder validates that through its tables.
+	for i := 0; i <= 31; i++ {
+		m["X"+itoa(i)] = Reg{idx: i, size: 16}
+		m["Y"+itoa(i)] = Reg{idx: i, size: 32}
+		m["Z"+itoa(i)] = Reg{idx: i, size: 64}
+	}
+	// Opmask: K0..K7.
+	for i := 0; i <= 7; i++ {
+		m["K"+itoa(i)] = Reg{idx: i, size: 8, mask: true}
 	}
 	return m
 }

-// isVec reports whether r is an XMM/YMM vector register.
-func (r Reg) isVec() bool { return r.size == 16 || r.size == 32 }
+// isVec reports whether r is an XMM/YMM/ZMM vector register.
+func (r Reg) isVec() bool { return r.size == 16 || r.size == 32 || r.size == 64 }

-// vecLenBit returns the VEX.L bit for a vector register (X=0/128-bit,
-// Y=1/256-bit).
+// vecLenBit returns the vector-length field for a vector register:
+// 0 (128-bit, VEX.L / EVEX.L'L=00), 1 (256-bit) or 2 (512-bit, EVEX only).
 func (r Reg) vecLenBit() int {
-	if r.size == 32 {
+	switch r.size {
+	case 64:
+		return 2
+	case 32:
 		return 1
 	}
 	return 0
@@ -39,6 +39,10 @@ const (
 	// source lives in the reg field, the destination in r/m — the PEXTR-style
 	// layout.  VEXTRACTI128 and VEXTRACTF128 use this shape.
 	vexExtract
+	// vexRMRev is the reversed two-operand form `OP src, dst` with the source
+	// in ModRM.reg and the destination in r/m — the layout of the EVEX
+	// narrowing stores (VPMOVDW, VPMOVQD).
+	vexRMRev
 	// vexZero is the no-operand form (VZEROUPPER).
 	vexZero
 )
@@ -131,6 +135,9 @@ var vexTable = map[string]vexSpec{

 	// 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},
 }

 // vexVarShift maps the shift mnemonics to their variable-count opcode — the
@@ -188,6 +195,13 @@ func isVex(mnemUpper string) bool {

 // encodeVex encodes a VEX instruction with operands in Plan 9 order.
 func (e *enc) encodeVex(mnemUpper string, ops []Operand) error {
+	// Vector register indices 16–31 exist only in EVEX encodings; fail
+	// loudly rather than silently truncating the index.
+	for _, op := range ops {
+		if r, ok := op.(Reg); ok && r.isVec() && r.idx >= 16 {
+			return fmt.Errorf("%s: vector register index %d needs an EVEX (AVX-512) instruction", mnemUpper, r.idx)
+		}
+	}
 	if ms, ok := vexMoveTable[mnemUpper]; ok {
 		return e.encodeVexMove(mnemUpper, ms, ops)
 	}
@@ -507,7 +521,8 @@ func vecReg(op Operand) (Reg, bool) {

 // vecOrMem reports whether op is a vector register or a memory reference.
 func vecOrMem(op Operand) bool {
-	if _, ok := op.(Mem); ok {
+	switch op.(type) {
+	case Mem, sbMem:
 		return true
 	}
 	r, ok := op.(Reg)
@@ -518,7 +533,7 @@ func vecOrMem(op Operand) bool {
 // acceptable: memory always is, a GPR only for VMOVD/VMOVQ.
 func validMoveOther(ms vexMoveSpec, op Operand) bool {
 	switch o := op.(type) {
-	case Mem:
+	case Mem, sbMem:
 		return true
 	case Reg:
 		return ms.gprOK && !o.isVec()
@@ -530,9 +545,13 @@ func validMoveOther(ms vexMoveSpec, op Operand) bool {
 // the given precomputed fields.  It is shared by every register/rm VEX form;
 // immediate bytes are appended by the caller.
 func (e *enc) emitVexFields(spec vexSpec, l, regField, rBit, vvvvBar int, rm Operand) error {
+	if l > 1 {
+		return fmt.Errorf("ZMM operand requires an EVEX instruction")
+	}
 	var modrm, sib int
 	var disp []byte
 	var xBit, bBit int
+	var sb *sbRef
 	switch r := rm.(type) {
 	case Reg:
 		modrm = 0xC0 | regField<<3 | (r.idx & 7)
@@ -546,6 +565,12 @@ func (e *enc) emitVexFields(spec vexSpec, l, regField, rBit, vvvvBar int, rm Ope
 		if err != nil {
 			return err
 		}
+	case sbMem:
+		// RIP-relative static-symbol reference; disp32 patched at link time.
+		modrm = regField<<3 | 0x05
+		sib = -1
+		disp = le32(0)
+		sb = &sbRef{name: r.name, addend: r.addend}
 	default:
 		return fmt.Errorf("invalid VEX r/m operand")
 	}
@@ -561,6 +586,9 @@ func (e *enc) emitVexFields(spec vexSpec, l, regField, rBit, vvvvBar int, rm Ope
 	if sib >= 0 {
 		e.out = append(e.out, byte(sib))
 	}
+	if sb != nil {
+		e.patches = append(e.patches, encPatch{off: len(e.out), name: sb.name, addend: sb.addend})
+	}
 	e.out = append(e.out, disp...)
 	return nil
 }
@@ -11,7 +11,9 @@ import (
 	"flag"
 	"fmt"
 	"io"
+	"io/fs"
 	"os"
+	"path/filepath"
 	"strings"

 	"sourcedock.dev/petrbalvin/gasm-devkit/arch"
@@ -26,7 +28,7 @@ import (

 // version is the release version, stamped at build time via
 // -ldflags "-X main.version=…" (defaulting to the current release).
-var version = "0.3.0"
+var version = "0.7.0"

 func main() {
 	if len(os.Args) < 2 {
@@ -63,7 +65,8 @@ func usage(w io.Writer) {
 Usage:
  gasm tokens <file>          print the lexical token stream
  gasm parse  <file>          parse and report syntax errors
-  gasm fmt    [-w] <file...>  canonicalise formatting (-w writes in place)
+  gasm fmt    [-w] [path...]  canonicalise formatting (no path or a directory:
+                              reformat every .s below it in place, like go fmt)
  gasm lint   <file...>       run static checks
  gasm asm    [-o out.bin] <file>  assemble to machine code (amd64, Phase 2)
  gasm lsp                    run the language server over stdio
@@ -130,15 +133,39 @@ func cmdParse(args []string) int {
 }

 func cmdFmt(args []string) int {
-	fs := flag.NewFlagSet("fmt", flag.ExitOnError)
-	write := fs.Bool("w", false, "write result to the source file")
-	fs.Parse(args)
-	if fs.NArg() == 0 {
-		fmt.Fprintln(os.Stderr, "usage: gasm fmt [-w] <file...>")
-		return 2
+	fs_ := flag.NewFlagSet("fmt", flag.ExitOnError)
+	write := fs_.Bool("w", false, "write result to the source file")
+	fs_.Parse(args)
+	// Like go fmt: with no arguments, or with a directory argument, every .s
+	// file below the directory is formatted in place and the names of the
+	// 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 +173,15 @@ 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
@@ -160,6 +191,29 @@ 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)
 	disable := fs.String("disable", "", "comma-separated rule codes to disable")
@@ -238,20 +292,18 @@ func cmdAsm(args []string) int {
 		return 1
 	}

-	var all []byte
-	functions := 0
-	for _, d := range f.Decls {
-		txt, ok := d.(*ast.Text)
-		if !ok {
-			continue
-		}
-		code, _, err := asm.Assemble(txt)
-		if err != nil {
-			fmt.Fprintf(os.Stderr, "%s: %s: %v\n", path, txt.Name.Name, err)
-			return 1
-		}
-		functions++
-		fmt.Printf("%s: %d bytes\n", txt.Name.Name, len(code))
+	img, err := asm.AssembleFile(f)
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "%s: %v\n", path, err)
+		return 1
+	}
+	if len(img.Funcs) == 0 {
+		fmt.Fprintln(os.Stderr, "gasm asm: no assemblable TEXT functions found")
+		return 1
+	}
+	for _, fn := range img.Funcs {
+		code := img.Code[fn.Offset : fn.Offset+fn.Size]
+		fmt.Printf("%s: %d bytes\n", fn.Name, fn.Size)
 		for i := 0; i < len(code); i += 16 {
 			end := i + 16
 			if end > len(code) {
@@ -263,13 +315,34 @@ func cmdAsm(args []string) int {
 			}
 			fmt.Println()
 		}
-		all = append(all, code...)
 	}
-	if functions == 0 {
-		fmt.Fprintln(os.Stderr, "gasm asm: no assemblable TEXT functions found")
-		return 1
+	if len(img.Data) > 0 {
+		fmt.Printf("data: %d bytes at 0x%x\n", len(img.Data), len(img.Code))
+		for _, d := range f.Decls {
+			g, ok := d.(*ast.Globl)
+			if !ok || g.Name == nil || g.Name.Pseudo != "SB" {
+				continue
+			}
+			size := 0
+			if g.Size != nil && g.Size.Imm.HasVal {
+				size = int(g.Size.Imm.Val)
+			}
+			fmt.Printf("  %s: %d bytes at 0x%x\n", g.Name.Name, size, img.Symbols[g.Name.Name])
+		}
+		for i := 0; i < len(img.Data); i += 16 {
+			end := i + 16
+			if end > len(img.Data) {
+				end = len(img.Data)
+			}
+			fmt.Printf("  %04x:", len(img.Code)+i)
+			for _, b := range img.Data[i:end] {
+				fmt.Printf(" %02x", b)
+			}
+			fmt.Println()
+		}
 	}
 	if *out != "" {
+		all := img.Bytes()
 		if err := os.WriteFile(*out, all, 0o644); err != nil {
 			fmt.Fprintln(os.Stderr, "gasm asm:", err)
 			return 1
@@ -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}) })
@@ -158,9 +206,10 @@ func TestUsage(t *testing.T) {
 }

 func TestCmdArgErrors(t *testing.T) {
-	// Missing file arguments produce a usage error (code 2).
-	if _, _, code := capture(func() int { return cmdFmt(nil) }); code != 2 {
-		t.Errorf("cmdFmt() code = %d, want 2", code)
+	// A missing path is an error (code 1); cmdFmt with no arguments is the
+	// recursive mode now, covered by TestCmdFmtRecursive.
+	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)
@@ -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
-  never saved and restored** — the per-architecture callee-saved set is amd64
-  `BX/BP/R12–R15`, arm64 `R19–R30`, riscv64 `X1/X8/X9/X18–X27`, loong64
-  `R1/R22–R31`.  This is an *audit*: the runtime's own assembly clobbers these
-  registers freely (it controls both sides of the call), so the rule is
-  advisory there, but in hand-written kernels called from ordinary Go code a
-  clobber is a genuine ABI violation.  It runs only on macro-free files, where
+  point.  On top of that it flags writes to the registers the **Go ABI** fixes
+  across calls that are never saved and restored — calibrated from
+  `cmd/compile/abi-internal.md`, *not* the platform ABI: Go's stack-based ABI0
+  has no System V style callee-saved registers (amd64 `BX`, `R12`–`R15` and
+  the like are caller-saved or permanent scratch, and hand-written kernels may
+  clobber them freely).  The audited set is the frame pointer and 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,
-operand spacing and per-function mnemonic alignment.  It is idempotent and its
-output always round-trips through the parser.
+every line — comments and blanks included.  It normalises indentation, operand
+spacing, per-function mnemonic alignment and blank-line layout: a new block
+(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`

@@ -206,15 +217,24 @@ memory destination r/m), the direction-sensitive moves (`VMOVDQU`, `VMOVUPD`,
 `VMULPD`, `VXORPD`, `VUNPCKHPD`, the scalar `VADDSD`/`VMULSD`, `VCVTDQ2PD`,
 `VFMADD231PD`) and the no-operand `VZEROUPPER` — together with `VPERMD` and
 the scalar families (`CMOVcc`, `SETcc`, `LZCNT`/`TZCNT`, the extending moves,
-`CVTSx2SD`, `IMUL3`), covering every instruction the go-flac AVX2 kernels use
-apart from global-symbol loads.  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 now holds for
-whole functions: every kernel function that avoids `SB` operands assembles to
-exactly the Go toolchain's bytes.  This increment covers register / memory /
-immediate / FP-frame operands, local-label jumps and these VEX SIMD forms;
-EVEX / AVX-512, `SB` (global symbol) operands (relocations) and object-file
-emission are the rest of Phase 2.
+`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
+mask destinations, and the compressed disp8×N displacement, whose multiplier
+follows the memory operand's size — covering every instruction the go-flac
+AVX2 and AVX-512 kernels use.  Every encoding is validated two ways: by
+round-trip decoding through `golang.org/x/arch`, and byte-for-byte against
+the machine code the real Go assembler emits — a comparison that holds for
+whole functions: all 27 functions of both kernels assemble to exactly the Go
+toolchain's bytes, the lone exception being the displacements of the
+static-constant loads, which the Go linker fills at link time.
+
+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)
+symbols are rejected: they need object-file emission, which — together with
+EVEX masking/zeroing and the other architectures — is the rest of Phase 2.

 ## Extension points

@@ -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.
-	var b strings.Builder
+	// Second pass: render each line.
+	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"))
-		b.WriteByte('\n')
+		outs = append(outs, outLine{kind: inf.kind, text: strings.TrimRight(out, " \t")})
 	}
-	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
@@ -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)",
@@ -3,7 +3,7 @@

 # gasm-devkit — developer tooling for Go's Plan 9 assembler (GAsm).

-version := "0.3.0"
+version := "0.7.0"

 default:
    @just --list
@@ -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
-	// saved and restored is clobbered across the call.  The check runs over the
-	// control-flow graph and is skipped for macro-using files, where an opaque
-	// macro may perform the save/restore.
+	// Register liveness: a register the Go ABI fixes across calls that is
+	// written but never saved and restored is clobbered.  The check runs over
+	// the control-flow graph and is skipped for macro-using files, where an
+	// 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.
@@ -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
-	// (BX, R13) without saving them; the register-clobber audit flags that by
-	// design.  This test targets the other rules, so the audit is disabled here
-	// (it is covered by TestRegisterClobber).
-	diags := File(f, Config{Arch: arch.AMD64, Disable: map[string]bool{CodeRegisterClobber: true}})
+	// The fixture mirrors the go-flac kernels, which write the Go ABI0
+	// scratch registers (BX, R13) without saving them — legal under Go's
+	// stack-based ABI, so the register-clobber audit stays silent and the
+	// fixture must lint entirely clean.
+	diags := File(f, Config{Arch: arch.AMD64})
 	if len(diags) != 0 {
 		t.Fatalf("expected no diagnostics on the fixture, got %+v", diags)
 	}
@@ -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
-// Plan 9 (amd64) spelling, first for arm64/riscv64/loong64.
-func dstIndex(in *ast.Instr, a arch.Arch) int {
-	if a == arch.AMD64 {
-		return len(in.Operands) - 1
-	}
-	return 0
+// dstIndex returns the operand index of the destination register: in Plan 9
+// notation the destination is the last operand on every architecture Go
+// supports (amd64, arm64, riscv64 and loong64 alike).
+func dstIndex(in *ast.Instr) int {
+	return len(in.Operands) - 1
 }

 // 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
-// must preserve for its caller, using the register names the assembler accepts
-// for each architecture.
-func calleeSavedGPRs(a arch.Arch) map[string]bool {
+// goFixedGPRs returns the general-purpose registers the Go ABI designates as
+// fixed across calls — the ones hand-written assembly must not permanently
+// clobber.  This follows 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 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
-		for i := 19; i <= 28; i++ {
-			names = append(names, fmt.Sprintf("R%d", i))
-		}
-		return gprSet(names...)
+		// R18 is reserved for the OS on some platforms, R28 holds the current
+		// goroutine, R29 is the frame pointer.
+		return gprSet("R18", "R28", "R29"), nil
 	case arch.RISCV:
-		// RA (X1) and the S registers (X8, X9, X18–X27) are callee-saved.
-		names := []string{"X1", "RA", "X8", "X9", "S0", "S1", "FP"}
-		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...)
+		// X27 holds the current goroutine.
+		return gprSet("X27"), nil
 	case arch.LOONG64:
-		// RA (R1), FP (R22) and S0–S8 (R23–R31) are callee-saved.
-		names := []string{"R1", "RA", "R22", "FP"}
-		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...)
+		// R22 holds the current goroutine.
+		return gprSet("R22"), nil
 	}
-	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
-// without also saving and restoring them — i.e. registers whose caller-owned
-// value is lost across the call.  It walks the blocks of the liveness analysis
-// (so the control-flow graph is what supplies the instruction set) and
-// aggregates each instruction's register effects.
-func clobberedCalleeSaved(l *liveness, a arch.Arch) []string {
-	callee := calleeSavedGPRs(a)
-	if len(callee) == 0 {
-		return nil
+// clobberedGoFixed returns the Go-ABI-fixed registers a function writes
+// without also saving and restoring them.  The first result lists registers
+// whose loss is never safe; the second lists the goroutine-pointer class,
+// whose loss is reported only when reachesRuntime is true (a non-NOSPLIT
+// function, or one that makes calls — the ABI0 transition machinery restores
+// the g pointer only on such paths).
+func clobberedGoFixed(l *liveness, a arch.Arch, reachesRuntime bool) (always, runtime []string) {
+	alwaysSet, runtimeSet := goFixedGPRs(a)
+	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
-	for r := range callee {
-		if def[r] && !(saved[r] && restored[r]) {
-			out = append(out, r)
+	clobbered := func(set map[string]bool) []string {
+		var out []string
+		for r := range set {
+			if def[r] && !(saved[r] && restored[r]) {
+				out = append(out, r)
+			}
 		}
+		sort.Strings(out)
+		return out
 	}
-	sort.Strings(out)
-	return out
+	always = clobbered(alwaysSet)
+	if reachesRuntime {
+		runtime = clobbered(runtimeSet)
+	}
+	return always, runtime
 }
@@ -5,36 +5,132 @@ package lint

 import "testing"

-// TestRegisterClobber detects writes to callee-saved registers that are not
-// saved and restored.
+// TestRegisterClobber checks the register-clobber audit is calibrated to the
+// 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.
-	clob := lintSrc(t, "#include \"textflag.h\"\n"+
+	// amd64: BX, R12, R13 and R15 are argument/permanent-scratch registers in
+	// 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 {
-		t.Fatalf("unsaved callee-saved write should be flagged: %+v", clob)
+	if codes(scratch)[CodeRegisterClobber] != 0 {
+		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"+
-		"\tMOVQ CX, BX\n"+
-		"\tPOPQ BX\n"+
+		"\tPUSHQ R14\n"+
+		"\tXORL R14, R14\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.
-	caller := lintSrc(t, "#include \"textflag.h\"\n"+
+	// amd64: BP maintains the frame chain and is always audited.
+	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 {
-		t.Fatalf("caller-saved register must not be flagged: %+v", caller)
+	if codes(bp)[CodeRegisterClobber] != 1 {
+		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)
 	}
 }
Author	SHA1	Message	Date
petrbalvin	e98680597d	feat(fmt): go-fmt-style recursive formatting and canonical blank-line layout Assisted-by: Qwen 3.8 Max Preview	2026-07-12 21:24:41 +02:00
petrbalvin	1a01870695	fix(lint): calibrate register-clobber to the Go ABI and add legacy SSE moves Assisted-by: Qwen 3.8 Max Preview	2026-07-11 17:36:52 +02:00
petrbalvin	458cfb626e	feat(asm): add EVEX/AVX-512 encoding and assemble the AVX-512 kernel byte-identically Assisted-by: Qwen 3.8 Max Preview	2026-07-10 13:20:49 +02:00
petrbalvin	56ecc39539	feat(asm): assemble static symbols and the whole go-flac AVX2 kernel byte-identically Assisted-by: Qwen 3.8 Max Preview	2026-07-09 15:56:03 +02:00