feat(asm): add the GPR-interchanging conversions, completing the amd64 EVEX set

Assisted-by: Qwen 3.8 Max Preview

asm/assemble.go

@@ -23,7 +23,7 @@ import (
 // operands require relocations and are not yet supported; the SIMD (VEX/AVX2)
 // integer and shuffle/extract/permute/move set is in.
 func Assemble(t *ast.Text) ([]byte, map[string]int, error) {
-	code, _, labels, err := assemble(t, nil)
+	code, _, labels, _, err := assemble(t, nil)
 	return code, labels, err
 }
 
@@ -48,9 +48,19 @@ type sbPatch struct {
 	addend int64
 }
 
+// spadjStep is one stack-adjustment boundary within a function: Value is the
+// SP delta from the entry state (just below the return address) in effect
+// from PC (function-relative) until the next step.  The steps feed the
+// pcsp table of the object-file emitters.
+type spadjStep struct {
+	pc    int
+	value int
+}
+
 // assemble encodes a TEXT body, returning the machine code, the static-symbol
-// patch sites (for the file-level layout to resolve) and the label table.
-func assemble(t *ast.Text, link *linkInfo) ([]byte, []sbPatch, map[string]int, error) {
+// patch sites (for the file-level layout to resolve), the label table and the
+// stack-adjustment boundaries.
+func assemble(t *ast.Text, link *linkInfo) ([]byte, []sbPatch, map[string]int, []spadjStep, error) {
 	fi := computeFrame(t)
 	chain := jumpChain(t)
 	resolve := func(name string) string {
@@ -74,7 +84,7 @@ func assemble(t *ast.Text, link *linkInfo) ([]byte, []sbPatch, map[string]int, e
 			case *ast.Instr:
 				sz, err := instrSize(s, fi, long[i], link)
 				if err != nil {
-					return nil, nil, nil, fmt.Errorf("%s: %w", s.Mnemonic.Text, err)
+					return nil, nil, nil, nil, fmt.Errorf("%s: %w", s.Mnemonic.Text, err)
 				}
 				sizes[i] = sz
 				pcs[i] = pos
@@ -114,24 +124,42 @@ func assemble(t *ast.Text, link *linkInfo) ([]byte, []sbPatch, map[string]int, e
 	// Pass 2: emit.
 	out := append([]byte(nil), fi.prologue...)
 	var patches []sbPatch
+	var steps []spadjStep
+	if fi.useFP {
+		// PUSHQ BP saves the return-address-relative base (+8); the MOVQ
+		// changes nothing; SUBQ $size, SP completes the frame.
+		steps = append(steps,
+			spadjStep{1, 8},
+			spadjStep{len(fi.prologue), 8 + fi.size},
+		)
+	}
 	pos := len(fi.prologue)
 	for i, stmt := range t.Body {
 		s, ok := stmt.(*ast.Instr)
 		if !ok {
 			continue
 		}
+		if strings.ToUpper(s.Mnemonic.Text) == "RET" && fi.useFP {
+			// The RET's epilogue prefix unwinds: ADDQ $size, SP restores
+			// the saved-BP-only stack, POPQ BP the entry state.
+			epi := len(fi.epilogue)
+			steps = append(steps,
+				spadjStep{pos + epi - 1, 8},
+				spadjStep{pos + epi, 0},
+			)
+		}
 		code, ps, err := encodeInstr(s, pos, offsets, fi, long[i], resolve, link)
 		if err != nil {
-			return nil, nil, nil, fmt.Errorf("%s: %w", s.Mnemonic.Text, err)
+			return nil, nil, nil, nil, fmt.Errorf("%s: %w", s.Mnemonic.Text, err)
 		}
 		if len(code) != sizes[i] {
-			return nil, nil, nil, fmt.Errorf("%s: size mismatch (%d vs %d)", s.Mnemonic.Text, len(code), sizes[i])
+			return nil, nil, nil, nil, fmt.Errorf("%s: size mismatch (%d vs %d)", s.Mnemonic.Text, len(code), sizes[i])
 		}
 		patches = append(patches, ps...)
 		out = append(out, code...)
 		pos += len(code)
 	}
-	return out, patches, offsets, nil
+	return out, patches, offsets, steps, nil
 }
 
 // jumpChain precomputes jump-to-jump folding: a label whose first instruction

asm/encode.go

@@ -51,16 +51,17 @@ func (e *enc) encode(mnem string, ops []Operand) error {
 
 	// VEX (AVX/AVX2) and EVEX (AVX-512) instructions: the trailing
 	// B/W/L/Q/D is part of the mnemonic, not a size suffix, so dispatch
-	// before splitSize.  A ".Z" suffix requests EVEX zeroing.
-	base, zeroing, err := stripEvexSuffix(upper)
+	// before splitSize.  EVEX suffixes (.Z, .SAE, rounding, .BCST) split
+	// off the mnemonic too.
+	base, sfx, err := parseEvexSuffix(upper)
 	if err != nil {
 		return err
 	}
-	if isVex(base) || isEvex(base) || base == "KMOVW" {
-		return e.encodeVec(base, ops, zeroing)
+	if isVex(base) || isEvex(base) || isKOp(base) || isGather(base) || isScatter(base) || base == "KMOVW" || base == "KMOVQ" {
+		return e.encodeVec(base, ops, sfx)
 	}
-	if zeroing {
-		return fmt.Errorf("%s: the .Z suffix requires an EVEX instruction", mnem)
+	if sfx.any() {
+		return fmt.Errorf("%s: the suffix requires an EVEX instruction", mnem)
 	}
 
 	// CMOVcc and SETcc carry the condition in the mnemonic (CMOVLGT, SETNE).
@@ -128,20 +129,32 @@ func splitSize(upper string) (base string, size int) {
 // its own direction-dependent opcodes; KTESTW is always VEX; everything else
 // takes EVEX when an operand demands it (a ZMM or K register, or an
 // EVEX-only mnemonic) and VEX otherwise.
-func (e *enc) encodeVec(upper string, ops []Operand, zeroing bool) error {
-	if upper == "KMOVW" {
-		if zeroing {
-			return fmt.Errorf("KMOVW takes no .Z suffix")
-		}
-		return e.encodeKmovw(ops)
+func (e *enc) encodeVec(upper string, ops []Operand, sfx evexSuffix) error {
+	if gs, ok := gatherTable[upper]; ok {
+		return e.encodeGather(upper, gs, ops, sfx)
 	}
-	if upper == "KTESTW" || !evexRequired(upper, ops) {
-		if zeroing {
+	if ss, ok := scatterTable[upper]; ok {
+		return e.encodeScatter(upper, ss, ops, sfx)
+	}
+	if upper == "KMOVW" || upper == "KMOVQ" {
+		if sfx.any() {
+			return fmt.Errorf("%s takes no EVEX suffixes", upper)
+		}
+		return e.encodeKmov(upper, ops)
+	}
+	if isKOp(upper) {
+		if sfx.any() {
+			return fmt.Errorf("%s takes no EVEX suffixes", upper)
+		}
+		return e.encodeKOp(upper, ops)
+	}
+	if upper == "KTESTW" || (!evexRequired(upper, ops) && !sfx.evexOnly()) {
+		if sfx.any() {
 			return fmt.Errorf("%s: the .Z suffix requires an EVEX instruction", upper)
 		}
 		return e.encodeVex(upper, ops)
 	}
-	return e.encodeEvex(upper, ops, zeroing)
+	return e.encodeEvex(upper, ops, sfx)
 }
 
 // --- instruction components -------------------------------------------------

asm/evex_test.go

@@ -230,7 +230,11 @@ func TestEvexMasking(t *testing.T) {
 		{"K0 mask", "VPADDD", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "K0"), vreg(t, "Z3")}},
 		{"two masks", "VPADDD", []Operand{vreg(t, "Z1"), vreg(t, "K1"), vreg(t, "K2"), vreg(t, "Z3")}},
 		{".Z on VEX-only", "VPSHUFD.Z", []Operand{Imm(1), vreg(t, "X0"), vreg(t, "X1")}},
-		{"unsupported suffix", "VPADDD.BCST", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}},
+		{"broadcast unsupported", "VPXORD.BCST", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}},
+		{"rounding unsupported", "VPXORD.RN_SAE", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}},
+		{"bcst with rounding", "VADDPD.BCST.RN_SAE", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}},
+		{"Z not last", "VADDPD.Z.RN_SAE", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}},
+		{"duplicate suffix", "VADDPD.Z.Z", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}},
 		{"KMOVW.Z", "KMOVW.Z", []Operand{vreg(t, "K1"), vreg(t, "K2")}},
 	}
 	for _, c := range bad {
@@ -240,6 +244,387 @@ func TestEvexMasking(t *testing.T) {
 	}
 }
 
+// TestEvexExtendedGroundTruth covers the wider EVEX/AVX-512 set — ternary
+// logic, lane shuffles/inserts/extracts, compares with a K destination,
+// permutes, the wider integer families, expand/compress, broadcasts,
+// rotates and word shifts, the opmask instructions, the EVEX suffixes
+// (rounding/SAE/broadcast) and the aligned/scalar moves — byte for byte
+// against the Go assembler.
+func TestEvexExtendedGroundTruth(t *testing.T) {
+	mem64 := func(base Reg) Operand { return Ptr(base, 0, 64) }
+	cases := []struct {
+		name string
+		mnem string
+		ops  []Operand
+		want string
+	}{
+		// Ternary logic and lane shuffles (NDS + imm8).
+		{"VPTERNLOGD", "VPTERNLOGD", []Operand{Imm(0xE8), vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f36d4825d9e8"},
+		{"VPTERNLOGQ", "VPTERNLOGQ", []Operand{Imm(0x96), vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f3ed4825d996"},
+		{"VSHUFI32X4", "VSHUFI32X4", []Operand{Imm(0x4E), vreg(t, "Y1"), vreg(t, "Y2"), vreg(t, "Y3")}, "62f36d2843d94e"},
+		{"VSHUFF64X2", "VSHUFF64X2", []Operand{Imm(1), vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f3ed4823d901"},
+		{"VPALIGNR", "VPALIGNR", []Operand{Imm(7), vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f36d480fd907"},
+		// Permutes.
+		{"VPERMB", "VPERMB", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f26d488dd9"},
+		{"VPERMW", "VPERMW", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f2ed488dd9"},
+		{"VPERMI2D", "VPERMI2D", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f26d4876d9"},
+		{"VPERMT2PD", "VPERMT2PD", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f2ed487fd9"},
+		// Compare with a K destination (and an immediate predicate).
+		{"VCMPPD", "VCMPPD", []Operand{Imm(4), vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "K3")}, "62f1ed48c2d904"},
+		{"VCMPPS", "VCMPPS", []Operand{Imm(0), vreg(t, "Y1"), vreg(t, "Y2"), vreg(t, "K4")}, "62f16c28c2e100"},
+		{"VCMPSD", "VCMPSD", []Operand{Imm(17), vreg(t, "X1"), vreg(t, "X2"), vreg(t, "K5")}, "62f1ef08c2e911"},
+		// Rounding / SAE / broadcast suffixes.
+		{"VADDPD.RN_SAE", "VADDPD.RN_SAE", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f1ed1858d9"},
+		{"VMULPD.RZ_SAE.Z", "VMULPD.RZ_SAE.Z", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "K1"), vreg(t, "Z3")}, "62f1edf959d9"},
+		{"VMAXPD.SAE", "VMAXPD.SAE", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f1ed585fd9"},
+		{"VADDPD.BCST", "VADDPD.BCST", []Operand{mem64(AX), vreg(t, "Z1"), vreg(t, "Z2")}, "62f1f5585810"},
+		// Packed single arithmetic (same opcodes, no mandatory prefix) —
+		// ZMM, YMM and XMM widths, rounding and broadcast.
+		{"VADDPS", "VADDPS", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f16c4858d9"},
+		{"VMULPS", "VMULPS", []Operand{vreg(t, "Y1"), vreg(t, "Y2"), vreg(t, "Y3")}, "c5ec59d9"},
+		{"VMAXPS", "VMAXPS", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "c5e85fd9"},
+		{"VDIVPS.RD_SAE", "VDIVPS.RD_SAE", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f16c385ed9"},
+		{"VADDPS.BCST", "VADDPS.BCST", []Operand{mem64(AX), vreg(t, "Z1"), vreg(t, "Z2")}, "62f174585810"},
+		// Compress / expand.
+		{"VCOMPRESSPD", "VCOMPRESSPD", []Operand{vreg(t, "Z1"), mem64(DI)}, "62f2fd488a0f"},
+		{"VEXPANDPS", "VEXPANDPS", []Operand{mem64(SI), vreg(t, "Y2")}, "62f27d288816"},
+		{"VPCOMPRESSD.Z", "VPCOMPRESSD.Z", []Operand{vreg(t, "Z1"), vreg(t, "K2"), mem64(DI)}, "62f27dca8b0f"},
+		// Broadcasts.
+		{"VPBROADCASTB gpr", "VPBROADCASTB", []Operand{BX, vreg(t, "Z1")}, "62f27d487acb"},
+		{"VPBROADCASTW mem", "VPBROADCASTW", []Operand{mem64(AX), vreg(t, "Z2")}, "62f27d487910"},
+		{"VBROADCASTSS", "VBROADCASTSS", []Operand{mem64(AX), vreg(t, "Y3")}, "c4e27d1818"},
+		{"VBROADCASTSD", "VBROADCASTSD", []Operand{mem64(AX), vreg(t, "Z4")}, "62f2fd481920"},
+		// Wider integer families.
+		{"VPMADDWD", "VPMADDWD", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f16d48f5d9"},
+		{"VPMADDUBSW", "VPMADDUBSW", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f26d4804d9"},
+		{"VPMULHUW", "VPMULHUW", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f16d48e4d9"},
+		{"VPSLLVW", "VPSLLVW", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f2ed4812d9"},
+		{"VPACKSSWB", "VPACKSSWB", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f16d4863d9"},
+		{"VPACKUSDW", "VPACKUSDW", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f26d482bd9"},
+		// Absolute values and replicating moves.
+		{"VPABSD", "VPABSD", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f27d481ed1"},
+		{"VPABSQ mem", "VPABSQ", []Operand{mem64(AX), vreg(t, "Z2")}, "62f2fd481f10"},
+		{"VMOVSLDUP", "VMOVSLDUP", []Operand{vreg(t, "X1"), vreg(t, "X2")}, "c5fa12d1"},
+		{"VMOVSHDUP", "VMOVSHDUP", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f17e4816d1"},
+		// Rotates and word/qword shifts.
+		{"VPROLD", "VPROLD", []Operand{Imm(5), vreg(t, "Z1"), vreg(t, "Z2")}, "62f16d4872c905"},
+		{"VPRORQ", "VPRORQ", []Operand{Imm(63), vreg(t, "Z1"), vreg(t, "Z2")}, "62f1ed4872c13f"},
+		{"VPSLLW", "VPSLLW", []Operand{Imm(9), vreg(t, "X1"), vreg(t, "X2")}, "c5e971f109"},
+		{"VPSRLQ", "VPSRLQ", []Operand{Imm(3), vreg(t, "Z1"), vreg(t, "Z2")}, "62f1ed4873d103"},
+		// Opmask instructions (VEX-encoded, the width in the L/W/pp bits).
+		{"KANDW", "KANDW", []Operand{vreg(t, "K1"), vreg(t, "K2"), vreg(t, "K3")}, "c5ec41d9"},
+		{"KORD", "KORD", []Operand{vreg(t, "K4"), vreg(t, "K5"), vreg(t, "K6")}, "c4e1d545f4"},
+		{"KXNORQ", "KXNORQ", []Operand{vreg(t, "K1"), vreg(t, "K2"), vreg(t, "K3")}, "c4e1ec46d9"},
+		{"KNOTB", "KNOTB", []Operand{vreg(t, "K4"), vreg(t, "K5")}, "c5f944ec"},
+		{"KUNPCKBW", "KUNPCKBW", []Operand{vreg(t, "K1"), vreg(t, "K2"), vreg(t, "K3")}, "c5ed4bd9"},
+		{"KSHIFTLW", "KSHIFTLW", []Operand{Imm(2), vreg(t, "K1"), vreg(t, "K2")}, "c4e3f932d102"},
+		{"KADDQ", "KADDQ", []Operand{vreg(t, "K1"), vreg(t, "K2"), vreg(t, "K3")}, "c4e1ec4ad9"},
+		{"KORTESTD", "KORTESTD", []Operand{vreg(t, "K1"), vreg(t, "K2")}, "c4e1f998d1"},
+		{"KMOVQ k,k", "KMOVQ", []Operand{vreg(t, "K1"), vreg(t, "K2")}, "c4e1f890d1"},
+		{"KMOVQ gpr,k", "KMOVQ", []Operand{BX, vreg(t, "K1")}, "c4e1fb92cb"},
+		// Lane extract / insert.
+		{"VEXTRACTF32X4", "VEXTRACTF32X4", []Operand{Imm(1), vreg(t, "Y1"), vreg(t, "X2")}, "62f37d2819ca01"},
+		{"VEXTRACTI64X2", "VEXTRACTI64X2", []Operand{Imm(1), vreg(t, "Y1"), vreg(t, "X2")}, "62f3fd2839ca01"},
+		{"VINSERTF32X8", "VINSERTF32X8", []Operand{Imm(1), vreg(t, "Y1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f36d481ad901"},
+		{"VINSERTI64X4", "VINSERTI64X4", []Operand{Imm(1), vreg(t, "Y1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f3ed483ad901"},
+		// Aligned moves and the scalar single move.
+		{"VMOVAPS", "VMOVAPS", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f17c4829ca"},
+		{"VMOVDQA64 mem", "VMOVDQA64", []Operand{mem64(AX), vreg(t, "Z2")}, "62f1fd486f10"},
+		{"VMOVSS mem", "VMOVSS", []Operand{mem64(AX), vreg(t, "X2")}, "c5fa1010"},
+		// Conversions and extending/narrowing moves.
+		{"VCVTPS2DQ", "VCVTPS2DQ", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f17d485bd1"},
+		{"VCVTTPS2DQ", "VCVTTPS2DQ", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f17e485bd1"},
+		{"VPMOVZXBW", "VPMOVZXBW", []Operand{vreg(t, "X1"), vreg(t, "Y2")}, "c4e27d30d1"},
+		{"VPMOVSXBW mem", "VPMOVSXBW", []Operand{mem64(AX), vreg(t, "Z2")}, "62f27d482010"},
+		{"VPMOVWB", "VPMOVWB", []Operand{vreg(t, "Z1"), vreg(t, "Y2")}, "62f27e4830ca"},
+		{"VPMOVQB", "VPMOVQB", []Operand{vreg(t, "Z1"), vreg(t, "X2")}, "62f27e4832ca"},
+	}
+	for _, c := range cases {
+		code, err := Encode(c.mnem, c.ops...)
+		if err != nil {
+			t.Errorf("%s: Encode: %v", c.name, err)
+			continue
+		}
+		if got := hexCompact(code); got != c.want {
+			t.Errorf("%s: bytes %s, want %s", c.name, got, c.want)
+			continue
+		}
+		inst, err := x86asm.Decode(code, 64)
+		if err != nil {
+			t.Errorf("%s: Decode(%x): %v", c.name, code, err)
+			continue
+		}
+		want := c.mnem
+		if i := strings.IndexByte(want, '.'); i > 0 {
+			want = want[:i]
+		}
+		if inst.Op.String() != want {
+			t.Errorf("%s: decoded as %s", c.name, inst.Op.String())
+		}
+	}
+}
+
+// TestEvexHelperGroundTruth covers the floating-point helper and conversion
+// tail of the EVEX set — reciprocals, rsqrt, getexp/getmant, scalef,
+// rndscale, reduce, fixupimm, range, fpclass, the remaining conversions —
+// plus gather/scatter with VSIB addressing, byte for byte against the Go
+// assembler.
+func TestEvexHelperGroundTruth(t *testing.T) {
+	vsib := func(base, idx string, scale int) Operand {
+		return Idx(vreg(t, base), vreg(t, idx), scale, 0, 0)
+	}
+	cases := []struct {
+		name string
+		mnem string
+		ops  []Operand
+		want string
+	}{
+		// Reciprocals and rsqrt (packed RM, scalar NDS).
+		{"VRCP14PD", "VRCP14PD", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f2fd484cd1"},
+		{"VRCP14PS", "VRCP14PS", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f27d484cd1"},
+		{"VRCP14SD", "VRCP14SD", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "62f2ed084dd9"},
+		{"VRCP14SS", "VRCP14SS", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "62f26d084dd9"},
+		{"VRSQRT14PD", "VRSQRT14PD", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f2fd484ed1"},
+		{"VRSQRT14PS", "VRSQRT14PS", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f27d484ed1"},
+		{"VRSQRT14SD", "VRSQRT14SD", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "62f2ed084fd9"},
+		{"VRSQRT14SS", "VRSQRT14SS", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "62f26d084fd9"},
+		// Getexp (packed RM, scalar NDS).
+		{"VGETEXPPD", "VGETEXPPD", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f2fd4842d1"},
+		{"VGETEXPPS", "VGETEXPPS", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f27d4842d1"},
+		{"VGETEXPSD", "VGETEXPSD", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "62f2ed0843d9"},
+		{"VGETEXPSS", "VGETEXPSS", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "62f26d0843d9"},
+		// Scalef (NDS).
+		{"VSCALEFPD", "VSCALEFPD", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f2ed482cd9"},
+		{"VSCALEFPS", "VSCALEFPS", []Operand{vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f26d482cd9"},
+		{"VSCALEFSD", "VSCALEFSD", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "62f2ed082dd9"},
+		{"VSCALEFSS", "VSCALEFSS", []Operand{vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "62f26d082dd9"},
+		// Rndscale / getmant / reduce (packed $imm,src,dst; scalar NDS+imm).
+		{"VRNDSCALEPD", "VRNDSCALEPD", []Operand{Imm(4), vreg(t, "Z1"), vreg(t, "Z2")}, "62f3fd4809d104"},
+		{"VRNDSCALEPS", "VRNDSCALEPS", []Operand{Imm(4), vreg(t, "Z1"), vreg(t, "Z2")}, "62f37d4808d104"},
+		{"VRNDSCALESD", "VRNDSCALESD", []Operand{Imm(4), vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "62f3ed080bd904"},
+		{"VRNDSCALESS", "VRNDSCALESS", []Operand{Imm(4), vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "62f36d080ad904"},
+		{"VGETMANTPD", "VGETMANTPD", []Operand{Imm(3), vreg(t, "Z1"), vreg(t, "Z2")}, "62f3fd4826d103"},
+		{"VGETMANTPS", "VGETMANTPS", []Operand{Imm(3), vreg(t, "Z1"), vreg(t, "Z2")}, "62f37d4826d103"},
+		{"VGETMANTSD", "VGETMANTSD", []Operand{Imm(3), vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "62f3ed0827d903"},
+		{"VGETMANTSS", "VGETMANTSS", []Operand{Imm(3), vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "62f36d0827d903"},
+		{"VREDUCEPD", "VREDUCEPD", []Operand{Imm(4), vreg(t, "Z1"), vreg(t, "Z2")}, "62f3fd4856d104"},
+		{"VREDUCEPS", "VREDUCEPS", []Operand{Imm(4), vreg(t, "Z1"), vreg(t, "Z2")}, "62f37d4856d104"},
+		{"VREDUCESD", "VREDUCESD", []Operand{Imm(4), vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "62f3ed0857d904"},
+		{"VREDUCESS", "VREDUCESS", []Operand{Imm(4), vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "62f36d0857d904"},
+		// Fixupimm / range (NDS + imm8).
+		{"VFIXUPIMMPD", "VFIXUPIMMPD", []Operand{Imm(2), vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f3ed4854d902"},
+		{"VFIXUPIMMPS", "VFIXUPIMMPS", []Operand{Imm(2), vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f36d4854d902"},
+		{"VFIXUPIMMSD", "VFIXUPIMMSD", []Operand{Imm(2), vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "62f3ed0855d902"},
+		{"VFIXUPIMMSS", "VFIXUPIMMSS", []Operand{Imm(2), vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "62f36d0855d902"},
+		{"VRANGEPD", "VRANGEPD", []Operand{Imm(1), vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f3ed4850d901"},
+		{"VRANGEPS", "VRANGEPS", []Operand{Imm(1), vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "Z3")}, "62f36d4850d901"},
+		{"VRANGESD", "VRANGESD", []Operand{Imm(1), vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "62f3ed0851d901"},
+		{"VRANGESS", "VRANGESS", []Operand{Imm(1), vreg(t, "X1"), vreg(t, "X2"), vreg(t, "X3")}, "62f36d0851d901"},
+		// FP class test ($imm, src, kdst; packed forms carry the length in
+		// the X/Y/Z mnemonic suffix the decoder drops).
+		{"VFPCLASSPDZ", "VFPCLASSPDZ", []Operand{Imm(4), vreg(t, "Z1"), vreg(t, "K2")}, "62f3fd4866d104"},
+		{"VFPCLASSPSY", "VFPCLASSPSY", []Operand{Imm(4), vreg(t, "Y1"), vreg(t, "K2")}, "62f37d2866d104"},
+		{"VFPCLASSSD", "VFPCLASSSD", []Operand{Imm(4), vreg(t, "X1"), vreg(t, "K2")}, "62f3fd0867d104"},
+		{"VFPCLASSSS", "VFPCLASSSS", []Operand{Imm(4), vreg(t, "X1"), vreg(t, "K2")}, "62f37d0867d104"},
+		// Gather: VEX spelling (mask register, VSIB, destination) and EVEX
+		// spelling (VSIB, K mask, destination; L'L follows the VSIB index).
+		{"VGATHERDPS vex", "VGATHERDPS", []Operand{vreg(t, "X2"), vsib("SI", "X1", 4), vreg(t, "X3")}, "c4e269921c8e"},
+		{"VPGATHERDD vex", "VPGATHERDD", []Operand{vreg(t, "Y2"), vsib("SI", "Y1", 4), vreg(t, "Y3")}, "c4e26d901c8e"},
+		{"VGATHERDPS evex", "VGATHERDPS", []Operand{vsib("SI", "X1", 4), vreg(t, "K2"), vreg(t, "X3")}, "62f27d0a921c8e"},
+		{"VPGATHERQD evex", "VPGATHERQD", []Operand{vsib("SI", "Z1", 8), vreg(t, "K2"), vreg(t, "Y3")}, "62f27d4a911cce"},
+		// Scatter (EVEX only: source, K mask, VSIB).
+		{"VSCATTERDPS", "VSCATTERDPS", []Operand{vreg(t, "X3"), vreg(t, "K1"), vsib("SI", "X1", 4)}, "62f27d09a21c8e"},
+		{"VSCATTERQPD", "VSCATTERQPD", []Operand{vreg(t, "Z3"), vreg(t, "K1"), vsib("SI", "Z1", 8)}, "62f2fd49a31cce"},
+		// The remaining conversions.
+		{"VCVTDQ2PS", "VCVTDQ2PS", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f17c485bd1"},
+		{"VCVTQQ2PS", "VCVTQQ2PS", []Operand{vreg(t, "Z1"), vreg(t, "Y2")}, "62f1fc485bd1"},
+		{"VCVTPD2QQ", "VCVTPD2QQ", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f1fd487bd1"},
+		{"VCVTPS2QQ", "VCVTPS2QQ", []Operand{vreg(t, "Y1"), vreg(t, "Z2")}, "62f17d487bd1"},
+		{"VCVTUDQ2PD", "VCVTUDQ2PD", []Operand{vreg(t, "X1"), vreg(t, "Y2")}, "62f17e287ad1"},
+		{"VCVTPH2PS", "VCVTPH2PS", []Operand{vreg(t, "Y1"), vreg(t, "Z2")}, "62f27d4813d1"},
+		{"VCVTPS2PH", "VCVTPS2PH", []Operand{Imm(4), vreg(t, "Y1"), vreg(t, "X2")}, "c4e37d1dca04"},
+	}
+	for _, c := range cases {
+		code, err := Encode(c.mnem, c.ops...)
+		if err != nil {
+			t.Errorf("%s: Encode: %v", c.name, err)
+			continue
+		}
+		if got := hexCompact(code); got != c.want {
+			t.Errorf("%s: bytes %s, want %s", c.name, got, c.want)
+			continue
+		}
+		inst, err := x86asm.Decode(code, 64)
+		if err != nil {
+			t.Errorf("%s: Decode(%x): %v", c.name, code, err)
+			continue
+		}
+		want := c.mnem
+		got := inst.Op.String()
+		if got != want && !(len(want) > len(got) && want[:len(got)] == got) {
+			t.Errorf("%s: decoded as %s", c.name, got)
+		}
+	}
+}
+
+// TestEvexGprGroundTruth covers the scalar conversions between vector and
+// general-purpose registers — the signed and truncated VCVT{,T}S{D,S}2SI
+// forms (VEX and EVEX), the unsigned EVEX-only forms, and the GPR-to-vector
+// VCVTSI2*/VCVTUSI2* forms with the preserved vector source in vvvv — byte
+// for byte against the Go assembler, including memory sources and extended
+// GPRs.
+func TestEvexGprGroundTruth(t *testing.T) {
+	mem := func(b Reg) Operand { return Ptr(b, 0, 8) }
+	cases := []struct {
+		name string
+		mnem string
+		ops  []Operand
+		want string
+	}{
+		{"VCVTSD2SI", "VCVTSD2SI", []Operand{vreg(t, "X1"), AX}, "c5fb2dc1"},
+		{"VCVTSD2SIQ", "VCVTSD2SIQ", []Operand{vreg(t, "X1"), AX}, "c4e1fb2dc1"},
+		{"VCVTSS2SI", "VCVTSS2SI", []Operand{vreg(t, "X1"), AX}, "c5fa2dc1"},
+		{"VCVTSS2SIQ", "VCVTSS2SIQ", []Operand{vreg(t, "X1"), AX}, "c4e1fa2dc1"},
+		{"VCVTTSD2SI", "VCVTTSD2SI", []Operand{vreg(t, "X1"), AX}, "c5fb2cc1"},
+		{"VCVTTSD2SIQ", "VCVTTSD2SIQ", []Operand{vreg(t, "X1"), AX}, "c4e1fb2cc1"},
+		{"VCVTTSS2SI", "VCVTTSS2SI", []Operand{vreg(t, "X1"), AX}, "c5fa2cc1"},
+		{"VCVTTSS2SIQ", "VCVTTSS2SIQ", []Operand{vreg(t, "X1"), AX}, "c4e1fa2cc1"},
+		{"VCVTSD2USIL", "VCVTSD2USIL", []Operand{vreg(t, "X1"), AX}, "62f17f0879c1"},
+		{"VCVTSD2USIQ", "VCVTSD2USIQ", []Operand{vreg(t, "X1"), AX}, "62f1ff0879c1"},
+		{"VCVTSS2USIL", "VCVTSS2USIL", []Operand{vreg(t, "X1"), AX}, "62f17e0879c1"},
+		{"VCVTSS2USIQ", "VCVTSS2USIQ", []Operand{vreg(t, "X1"), AX}, "62f1fe0879c1"},
+		{"VCVTTSD2USIL", "VCVTTSD2USIL", []Operand{vreg(t, "X1"), AX}, "62f17f0878c1"},
+		{"VCVTTSD2USIQ", "VCVTTSD2USIQ", []Operand{vreg(t, "X1"), AX}, "62f1ff0878c1"},
+		{"VCVTTSS2USIL", "VCVTTSS2USIL", []Operand{vreg(t, "X1"), AX}, "62f17e0878c1"},
+		{"VCVTTSS2USIQ", "VCVTTSS2USIQ", []Operand{vreg(t, "X1"), AX}, "62f1fe0878c1"},
+		{"VCVTSI2SDL", "VCVTSI2SDL", []Operand{AX, vreg(t, "X1"), vreg(t, "X2")}, "c5f32ad0"},
+		{"VCVTSI2SDQ", "VCVTSI2SDQ", []Operand{AX, vreg(t, "X1"), vreg(t, "X2")}, "c4e1f32ad0"},
+		{"VCVTSI2SSL", "VCVTSI2SSL", []Operand{AX, vreg(t, "X1"), vreg(t, "X2")}, "c5f22ad0"},
+		{"VCVTSI2SSQ", "VCVTSI2SSQ", []Operand{AX, vreg(t, "X1"), vreg(t, "X2")}, "c4e1f22ad0"},
+		{"VCVTUSI2SDL", "VCVTUSI2SDL", []Operand{AX, vreg(t, "X1"), vreg(t, "X2")}, "62f177087bd0"},
+		{"VCVTUSI2SDQ", "VCVTUSI2SDQ", []Operand{AX, vreg(t, "X1"), vreg(t, "X2")}, "62f1f7087bd0"},
+		{"VCVTUSI2SSL", "VCVTUSI2SSL", []Operand{AX, vreg(t, "X1"), vreg(t, "X2")}, "62f176087bd0"},
+		{"VCVTUSI2SSQ", "VCVTUSI2SSQ", []Operand{AX, vreg(t, "X1"), vreg(t, "X2")}, "62f1f6087bd0"},
+		{"VCVTSD2SI mem", "VCVTSD2SI", []Operand{mem(AX), BX}, "c5fb2d18"},
+		{"VCVTSI2SDQ mem", "VCVTSI2SDQ", []Operand{mem(BX), vreg(t, "X1"), vreg(t, "X2")}, "c4e1f32a13"},
+		{"VCVTSD2SIQ hi gpr", "VCVTSD2SIQ", []Operand{vreg(t, "X1"), vreg(t, "R9")}, "c461fb2dc9"},
+		{"VCVTSI2SDQ hi gpr", "VCVTSI2SDQ", []Operand{vreg(t, "R10"), vreg(t, "X1"), vreg(t, "X2")}, "c4c1f32ad2"},
+	}
+	for _, c := range cases {
+		code, err := Encode(c.mnem, c.ops...)
+		if err != nil {
+			t.Errorf("%s: Encode: %v", c.name, err)
+			continue
+		}
+		if got := hexCompact(code); got != c.want {
+			t.Errorf("%s: bytes %s, want %s", c.name, got, c.want)
+			continue
+		}
+		inst, err := x86asm.Decode(code, 64)
+		if err != nil {
+			t.Errorf("%s: Decode(%x): %v", c.name, code, err)
+			continue
+		}
+		// The decoder does not distinguish the Plan 9 SIQ spelling (the
+		// 64-bit GPR destination) from the base name; the W bit carries it.
+		want := c.mnem
+		got := inst.Op.String()
+		if got != want && !(len(want) > len(got) && want[:len(got)] == got) {
+			t.Errorf("%s: decoded as %s", c.name, got)
+		}
+	}
+}
+
+// TestEvexConversionGroundTruth covers the unsigned and truncating VCVT*
+// conversions, the remaining sign/zero-extending moves, the signed/unsigned
+// narrowing stores and the mask/vector conversions, byte for byte against
+// the Go assembler.
+func TestEvexConversionGroundTruth(t *testing.T) {
+	cases := []struct {
+		name string
+		mnem string
+		ops  []Operand
+		want string
+	}{
+		// Unsigned and truncating conversions.
+		{"VCVTPD2PS", "VCVTPD2PS", []Operand{vreg(t, "Z1"), vreg(t, "Y2")}, "62f1fd485ad1"},
+		{"VCVTPD2PSX", "VCVTPD2PSX", []Operand{vreg(t, "X1"), vreg(t, "X2")}, "c5f95ad1"},
+		{"VCVTPD2PSY", "VCVTPD2PSY", []Operand{vreg(t, "Y1"), vreg(t, "X2")}, "c5fd5ad1"},
+		{"VCVTPD2UDQ", "VCVTPD2UDQ", []Operand{vreg(t, "Z1"), vreg(t, "Y2")}, "62f1fc4879d1"},
+		{"VCVTPD2UDQX", "VCVTPD2UDQX", []Operand{vreg(t, "X1"), vreg(t, "X2")}, "62f1fc0879d1"},
+		{"VCVTTPD2UDQ", "VCVTTPD2UDQ", []Operand{vreg(t, "Z1"), vreg(t, "Y2")}, "62f1fc4878d1"},
+		{"VCVTTPD2UDQY", "VCVTTPD2UDQY", []Operand{vreg(t, "Y1"), vreg(t, "X2")}, "62f1fc2878d1"},
+		{"VCVTTPD2UQQ", "VCVTTPD2UQQ", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f1fd4878d1"},
+		{"VCVTPS2UDQ", "VCVTPS2UDQ", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f17c4879d1"},
+		{"VCVTTPS2UDQ", "VCVTTPS2UDQ", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f17c4878d1"},
+		{"VCVTPS2UQQ", "VCVTPS2UQQ", []Operand{vreg(t, "Y1"), vreg(t, "Z2")}, "62f17d4879d1"},
+		{"VCVTTPS2UQQ", "VCVTTPS2UQQ", []Operand{vreg(t, "Y1"), vreg(t, "Z2")}, "62f17d4878d1"},
+		{"VCVTTPD2QQ", "VCVTTPD2QQ", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f1fd487ad1"},
+		{"VCVTTPS2QQ", "VCVTTPS2QQ", []Operand{vreg(t, "Y1"), vreg(t, "Z2")}, "62f17d487ad1"},
+		{"VCVTUQQ2PD", "VCVTUQQ2PD", []Operand{vreg(t, "Z1"), vreg(t, "Z2")}, "62f1fe487ad1"},
+		{"VCVTUQQ2PS", "VCVTUQQ2PS", []Operand{vreg(t, "Z1"), vreg(t, "Y2")}, "62f1ff487ad1"},
+		{"VCVTUQQ2PSX", "VCVTUQQ2PSX", []Operand{vreg(t, "X1"), vreg(t, "X2")}, "62f1ff087ad1"},
+		{"VCVTQQ2PSX", "VCVTQQ2PSX", []Operand{vreg(t, "X1"), vreg(t, "X2")}, "62f1fc085bd1"},
+		{"VCVTQQ2PSY", "VCVTQQ2PSY", []Operand{vreg(t, "Y1"), vreg(t, "X2")}, "62f1fc285bd1"},
+		// The remaining sign/zero-extending moves.
+		{"VPMOVSXBD", "VPMOVSXBD", []Operand{vreg(t, "X1"), vreg(t, "Y2")}, "c4e27d21d1"},
+		{"VPMOVSXBQ evex", "VPMOVSXBQ", []Operand{vreg(t, "X1"), vreg(t, "Z2")}, "62f27d4822d1"},
+		{"VPMOVSXWQ", "VPMOVSXWQ", []Operand{vreg(t, "X1"), vreg(t, "Y2")}, "c4e27d24d1"},
+		{"VPMOVSXWD", "VPMOVSXWD", []Operand{vreg(t, "X1"), vreg(t, "Y2")}, "c4e27d23d1"},
+		{"VPMOVZXBD", "VPMOVZXBD", []Operand{vreg(t, "X1"), vreg(t, "Y2")}, "c4e27d31d1"},
+		{"VPMOVZXBQ evex", "VPMOVZXBQ", []Operand{vreg(t, "X1"), vreg(t, "Z2")}, "62f27d4832d1"},
+		{"VPMOVZXWD", "VPMOVZXWD", []Operand{vreg(t, "X1"), vreg(t, "Y2")}, "c4e27d33d1"},
+		{"VPMOVZXWQ", "VPMOVZXWQ", []Operand{vreg(t, "X1"), vreg(t, "Y2")}, "c4e27d34d1"},
+		// Signed narrowing stores.
+		{"VPMOVSDB", "VPMOVSDB", []Operand{vreg(t, "Z1"), vreg(t, "X2")}, "62f27e4821ca"},
+		{"VPMOVSDW", "VPMOVSDW", []Operand{vreg(t, "Z1"), vreg(t, "Y2")}, "62f27e4823ca"},
+		{"VPMOVSQB", "VPMOVSQB", []Operand{vreg(t, "Z1"), vreg(t, "X2")}, "62f27e4822ca"},
+		{"VPMOVSQD", "VPMOVSQD", []Operand{vreg(t, "Z1"), vreg(t, "Y2")}, "62f27e4825ca"},
+		{"VPMOVSQW", "VPMOVSQW", []Operand{vreg(t, "Z1"), vreg(t, "X2")}, "62f27e4824ca"},
+		{"VPMOVSWB", "VPMOVSWB", []Operand{vreg(t, "Z1"), vreg(t, "Y2")}, "62f27e4820ca"},
+		// Unsigned narrowing stores.
+		{"VPMOVUSDB", "VPMOVUSDB", []Operand{vreg(t, "Z1"), vreg(t, "X2")}, "62f27e4811ca"},
+		{"VPMOVUSDW", "VPMOVUSDW", []Operand{vreg(t, "Z1"), vreg(t, "Y2")}, "62f27e4813ca"},
+		{"VPMOVUSQB", "VPMOVUSQB", []Operand{vreg(t, "Z1"), vreg(t, "X2")}, "62f27e4812ca"},
+		{"VPMOVUSQD", "VPMOVUSQD", []Operand{vreg(t, "Z1"), vreg(t, "Y2")}, "62f27e4815ca"},
+		{"VPMOVUSQW", "VPMOVUSQW", []Operand{vreg(t, "Z1"), vreg(t, "X2")}, "62f27e4814ca"},
+		{"VPMOVUSWB", "VPMOVUSWB", []Operand{vreg(t, "Z1"), vreg(t, "Y2")}, "62f27e4810ca"},
+		{"VPMOVDB", "VPMOVDB", []Operand{vreg(t, "Z1"), vreg(t, "X2")}, "62f27e4831ca"},
+		{"VPMOVQW", "VPMOVQW", []Operand{vreg(t, "Z1"), vreg(t, "X2")}, "62f27e4834ca"},
+		// Mask/vector conversions (the K register is an operand, not a
+		// mask).
+		{"VPMOVM2B", "VPMOVM2B", []Operand{vreg(t, "K1"), vreg(t, "X2")}, "62f27e0828d1"},
+		{"VPMOVM2W", "VPMOVM2W", []Operand{vreg(t, "K1"), vreg(t, "X2")}, "62f2fe0828d1"},
+		{"VPMOVM2D", "VPMOVM2D", []Operand{vreg(t, "K1"), vreg(t, "X2")}, "62f27e0838d1"},
+		{"VPMOVM2Q", "VPMOVM2Q", []Operand{vreg(t, "K1"), vreg(t, "Z2")}, "62f2fe4838d1"},
+		{"VPMOVB2M", "VPMOVB2M", []Operand{vreg(t, "X1"), vreg(t, "K2")}, "62f27e0829d1"},
+		{"VPMOVW2M", "VPMOVW2M", []Operand{vreg(t, "X1"), vreg(t, "K2")}, "62f2fe0829d1"},
+		{"VPMOVD2M", "VPMOVD2M", []Operand{vreg(t, "Z1"), vreg(t, "K2")}, "62f27e4839d1"},
+		{"VPMOVQ2M", "VPMOVQ2M", []Operand{vreg(t, "Z1"), vreg(t, "K2")}, "62f2fe4839d1"},
+	}
+	for _, c := range cases {
+		code, err := Encode(c.mnem, c.ops...)
+		if err != nil {
+			t.Errorf("%s: Encode: %v", c.name, err)
+			continue
+		}
+		if got := hexCompact(code); got != c.want {
+			t.Errorf("%s: bytes %s, want %s", c.name, got, c.want)
+			continue
+		}
+		inst, err := x86asm.Decode(code, 64)
+		if err != nil {
+			t.Errorf("%s: Decode(%x): %v", c.name, code, err)
+			continue
+		}
+		want := c.mnem
+		got := inst.Op.String()
+		if got != want && !(len(want) > len(got) && want[:len(got)] == got) {
+			t.Errorf("%s: decoded as %s", c.name, got)
+		}
+	}
+}
+
 // TestEvexErrors checks the EVEX-specific error paths.
 func TestEvexErrors(t *testing.T) {
 	cases := []struct {

asm/goobj.go

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

asm/goobj_test.go

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

asm/link.go

@@ -28,13 +28,26 @@ type Image struct {
 
 // FuncLayout describes one assembled function within an Image.
 type FuncLayout struct {
-	Name   string
-	Pkg    string // explicit package prefix ("" = the current package)
-	Static bool   // the <> marker: file-local, not exported
-	Offset int    // start offset within the image (== offset within Code)
-	Size   int
-	Labels map[string]int // local labels, function-relative
-	Relocs []Reloc        // static-symbol references, in emission order
+	Name    string
+	Pkg     string // explicit package prefix ("" = the current package)
+	Static  bool   // the <> marker: file-local, not exported
+	Offset  int    // start offset within the image (== offset within Code)
+	Size    int
+	Args    int            // declared argument/result area (the TEXT size suffix)
+	Frame   int            // local frame size (the TEXT $framesize)
+	NoSplit bool           // the NOSPLIT flag
+	SPWrite bool           // the SPWRITE flag: writes an arbitrary value to SP
+	Line    int            // source line of the TEXT directive
+	Labels  map[string]int // local labels, function-relative
+	Relocs  []Reloc        // static-symbol references, in emission order
+	Spadj   []SpadjStep    // stack-adjustment boundaries, ascending by PC
+}
+
+// SpadjStep is one stack-adjustment boundary: Value is the SP delta from the
+// entry state in effect from PC (function-relative) until the next step.
+type SpadjStep struct {
+	PC    int
+	Value int
 }
 
 // Reloc is one static-symbol reference within a function body: the disp32
@@ -57,6 +70,8 @@ type DataSymbol struct {
 	Offset int    // byte offset within Data
 	Size   int
 	Static bool // the <> marker: file-local, not exported
+	Rodata bool // the RODATA flag: read-only data
+	Dupok  bool // the DUPOK flag: duplicate-OK
 }
 
 // Bytes returns the whole image: code, then data.
@@ -95,18 +110,33 @@ func AssembleFile(f *ast.File) (*Image, error) {
 		if !ok {
 			continue
 		}
-		code, patches, labels, err := assemble(t, link)
+		code, patches, labels, steps, err := assemble(t, link)
 		if err != nil {
 			return nil, fmt.Errorf("%s: %w", t.Name.Name, err)
 		}
-		img.Funcs = append(img.Funcs, FuncLayout{
+		fl := FuncLayout{
 			Name:   t.Name.Name,
 			Pkg:    t.Name.Pkg,
 			Static: t.Name.Static,
 			Offset: len(img.Code),
 			Size:   len(code),
+			Frame:  frameSize(t),
+			Args:   argsSize(t),
+			Line:   t.Pos().Line,
 			Labels: labels,
-		})
+		}
+		for _, f := range t.Flags {
+			switch f {
+			case "NOSPLIT":
+				fl.NoSplit = true
+			case "SPWRITE":
+				fl.SPWrite = true
+			}
+		}
+		for _, s := range steps {
+			fl.Spadj = append(fl.Spadj, SpadjStep{PC: s.pc, Value: s.value})
+		}
+		img.Funcs = append(img.Funcs, fl)
 		img.Code = append(img.Code, code...)
 		funcs = append(funcs, asmFunc{name: t.Name.Name, patches: patches})
 	}
@@ -124,6 +154,8 @@ func AssembleFile(f *ast.File) (*Image, error) {
 			Offset: len(img.Data),
 			Size:   len(d.buf),
 			Static: d.static,
+			Rodata: d.rodata,
+			Dupok:  d.dupok,
 		})
 		img.Data = append(img.Data, d.buf...)
 	}
@@ -163,6 +195,8 @@ type dataSym struct {
 	pkg    string
 	buf    []byte
 	static bool
+	rodata bool
+	dupok  bool
 }
 
 // collectData gathers the file's static symbols (GLOBL) and their initial
@@ -185,12 +219,28 @@ func collectData(f *ast.File) ([]dataSym, error) {
 				size = int(dd.Size.Imm.Val)
 			}
 			index[name] = len(syms)
-			syms = append(syms, dataSym{
+			ds := dataSym{
 				name:   name,
 				pkg:    dd.Name.Pkg,
 				buf:    make([]byte, size),
 				static: dd.Name.Static,
-			})
+			}
+			for _, f := range dd.Flags {
+				switch f {
+				case "RODATA":
+					ds.rodata = true
+				case "DUPOK":
+					ds.dupok = true
+				case "1":
+					ds.dupok = true
+				case "8":
+					ds.rodata = true
+				case "9":
+					ds.dupok = true
+					ds.rodata = true
+				}
+			}
+			syms = append(syms, ds)
 
 		case *ast.Data:
 			if dd.Name == nil || dd.Name.Pseudo != "SB" {
@@ -230,3 +280,19 @@ func collectData(f *ast.File) ([]dataSym, error) {
 func align16(n int) int {
 	return (n + 15) &^ 15
 }
+
+// frameSize returns the local frame size declared on the TEXT directive.
+func frameSize(t *ast.Text) int {
+	if t.Frame != nil && t.Frame.Imm.HasVal {
+		return int(t.Frame.Imm.Val)
+	}
+	return 0
+}
+
+// argsSize returns the argument/result area declared on the TEXT directive.
+func argsSize(t *ast.Text) int {
+	if t.Args != nil && t.Args.Imm.HasVal {
+		return int(t.Args.Imm.Val)
+	}
+	return 0
+}

asm/vex.go

@@ -91,12 +91,19 @@ var vexTable = map[string]vexSpec{
 	"VPCMPGTQ": {2, 0x37, 0, 1, -1, vexNDS3},
 
 	// VEX.128/256.66.0F.WIG — packed double-precision arithmetic / logic.
-	"VADDPD":    {1, 0x58, 0, 1, -1, vexNDS3},
-	"VMULPD":    {1, 0x59, 0, 1, -1, vexNDS3},
-	"VSUBPD":    {1, 0x5C, 0, 1, -1, vexNDS3},
-	"VDIVPD":    {1, 0x5E, 0, 1, -1, vexNDS3},
-	"VMINPD":    {1, 0x5D, 0, 1, -1, vexNDS3},
-	"VMAXPD":    {1, 0x5F, 0, 1, -1, vexNDS3},
+	"VADDPD": {1, 0x58, 0, 1, -1, vexNDS3},
+	"VMULPD": {1, 0x59, 0, 1, -1, vexNDS3},
+	"VSUBPD": {1, 0x5C, 0, 1, -1, vexNDS3},
+	"VDIVPD": {1, 0x5E, 0, 1, -1, vexNDS3},
+	"VMINPD": {1, 0x5D, 0, 1, -1, vexNDS3},
+	"VMAXPD": {1, 0x5F, 0, 1, -1, vexNDS3},
+	// VEX.128/256.0F.WIG — packed single-precision arithmetic.
+	"VADDPS":    {1, 0x58, 0, 0, -1, vexNDS3},
+	"VMULPS":    {1, 0x59, 0, 0, -1, vexNDS3},
+	"VSUBPS":    {1, 0x5C, 0, 0, -1, vexNDS3},
+	"VDIVPS":    {1, 0x5E, 0, 0, -1, vexNDS3},
+	"VMINPS":    {1, 0x5D, 0, 0, -1, vexNDS3},
+	"VMAXPS":    {1, 0x5F, 0, 0, -1, vexNDS3},
 	"VXORPD":    {1, 0x57, 0, 1, -1, vexNDS3},
 	"VUNPCKHPD": {1, 0x15, 0, 1, -1, vexNDS3},
 	"VUNPCKLPD": {1, 0x14, 0, 1, -1, vexNDS3},
@@ -123,7 +130,15 @@ var vexTable = map[string]vexSpec{
 	// no vvvv).
 	"VPMOVSXWD":    {2, 0x23, 0, 1, -1, vexRM},
 	"VPMOVSXDQ":    {2, 0x25, 0, 1, -1, vexRM},
+	"VPMOVSXBD":    {2, 0x21, 0, 1, -1, vexRM},
+	"VPMOVSXBQ":    {2, 0x22, 0, 1, -1, vexRM},
+	"VPMOVSXWQ":    {2, 0x24, 0, 1, -1, vexRM},
 	"VPMOVZXDQ":    {2, 0x35, 0, 1, -1, vexRM},
+	"VPMOVZXBW":    {2, 0x30, 0, 1, -1, vexRM},
+	"VPMOVZXBD":    {2, 0x31, 0, 1, -1, vexRM},
+	"VPMOVZXBQ":    {2, 0x32, 0, 1, -1, vexRM},
+	"VPMOVZXWD":    {2, 0x33, 0, 1, -1, vexRM},
+	"VPMOVZXWQ":    {2, 0x34, 0, 1, -1, vexRM},
 	"VPBROADCASTD": {2, 0x58, 0, 1, -1, vexRM},
 	"VPBROADCASTQ": {2, 0x59, 0, 1, -1, vexRM},
 	// VEX.128/256.F3.0F.WIG — signed dword to packed double conversion
@@ -169,6 +184,9 @@ var vexTable = map[string]vexSpec{
 	// VEX.256.66.0F3A.W0 — lane extract (reg=YMM src, rm=XMM/memory dst, imm8).
 	"VEXTRACTI128": {3, 0x39, 0, 1, -1, vexExtract},
 	"VEXTRACTF128": {3, 0x19, 0, 1, -1, vexExtract},
+	// VEX.128/256.66.0F3A.W0 — half-precision convert back ($imm, src, dst:
+	// reg=src, rm=XMM/memory dst, imm8 — the extract layout).
+	"VCVTPS2PH": {3, 0x1D, 0, 1, -1, vexExtract},
 
 	// VEX.128.0F.W0 — no operands.
 	"VZEROUPPER": {1, 0x77, 0, 0, -1, vexZero},
@@ -176,6 +194,45 @@ var vexTable = map[string]vexSpec{
 	// VEX.128.0F.W0 — mask-register test (KTESTW k1, k2: reg = dst, rm = src).
 	"KTESTW": {1, 0x99, 0, 0, -1, vexRM},
 
+	// VEX.66.0F38.W0 — broadcast a single/double to all lanes (reg=dst,
+	// rm=scalar memory; SD is 256-bit only).
+	"VBROADCASTSS": {2, 0x18, 0, 1, -1, vexRM},
+	"VBROADCASTSD": {2, 0x19, 0, 1, -1, vexRM},
+	// VEX.66.0F38.W0 — half-precision convert (reg=dst, rm=half-width
+	// source).
+	"VCVTPH2PS": {2, 0x13, 0, 1, -1, vexRM},
+	// VEX.F3.0F.WIG — replicate even/odd singles (reg=dst, rm=src).
+	"VMOVSLDUP": {1, 0x12, 0, 2, -1, vexRM},
+	"VMOVSHDUP": {1, 0x16, 0, 2, -1, vexRM},
+	// VEX.66.0F.WIG — packed double to packed single conversion, the X/Y
+	// spellings: the destination is always XMM and the spelling fixes the
+	// source length (X = 128, Y = 256).
+	"VCVTPD2PSX": {1, 0x5A, 0, 1, -1, vexRMSrcLen},
+	"VCVTPD2PSY": {1, 0x5A, 0, 1, -1, vexRMSrcLen},
+
+	// VEX scalar conversions between vector and general-purpose registers.
+	// Vector to GPR (two operands: vec/mem source, GPR destination, vvvv
+	// unused; the length follows the source).
+	"VCVTSD2SI":   {1, 0x2D, 0, 3, -1, vexRM},
+	"VCVTSD2SIQ":  {1, 0x2D, 1, 3, -1, vexRM},
+	"VCVTSS2SI":   {1, 0x2D, 0, 2, -1, vexRM},
+	"VCVTSS2SIQ":  {1, 0x2D, 1, 2, -1, vexRM},
+	"VCVTTSD2SI":  {1, 0x2C, 0, 3, -1, vexRM},
+	"VCVTTSD2SIQ": {1, 0x2C, 1, 3, -1, vexRM},
+	"VCVTTSS2SI":  {1, 0x2C, 0, 2, -1, vexRM},
+	"VCVTTSS2SIQ": {1, 0x2C, 1, 2, -1, vexRM},
+	// GPR to vector (three operands: GPR/mem source in r/m, the preserved
+	// vector source in vvvv, vector destination in reg).
+	"VCVTSI2SDL": {1, 0x2A, 0, 3, -1, vexNDS3},
+	"VCVTSI2SDQ": {1, 0x2A, 1, 3, -1, vexNDS3},
+	"VCVTSI2SSL": {1, 0x2A, 0, 2, -1, vexNDS3},
+	"VCVTSI2SSQ": {1, 0x2A, 1, 2, -1, vexNDS3},
+
+	// VEX.128/256.66.0F.WIG — word shifts (opdigit selects the shift).
+	"VPSRLW": {1, 0x71, 0, 1, 2, vexShiftImm},
+	"VPSRAW": {1, 0x71, 0, 1, 4, vexShiftImm},
+	"VPSLLW": {1, 0x71, 0, 1, 6, vexShiftImm},
+
 	// VEX.F2.0F — packed double to packed dword conversions, truncating and
 	// non-truncating.  The destination is always XMM; the X/Y spellings fix
 	// the source length (XMM/YMM), and VEX.L follows it — see vexSrcLen.
@@ -194,6 +251,8 @@ var vexSrcLen = map[string]int{
 	"VCVTPD2DQY":  1,
 	"VCVTTPD2DQX": 0,
 	"VCVTTPD2DQY": 1,
+	"VCVTPD2PSX":  0,
+	"VCVTPD2PSY":  1,
 }
 
 // vexVarShift maps the shift mnemonics to their variable-count opcode — the
@@ -238,6 +297,11 @@ var vexMoveTable = map[string]vexMoveSpec{
 	// VEX.128.F2.0F.WIG — scalar double move, memory operands only (the
 	// register form takes three operands and is not supported yet).
 	"VMOVSD": {1, 3, 0x10, 0x11, 0, 0, 0, 0, false, false, true},
+	// VEX.128.F3.0F.WIG — scalar single move, memory operands only.
+	"VMOVSS": {1, 2, 0x10, 0x11, 0, 0, 0, 0, false, false, true},
+	// VEX.128/256 — aligned packed moves.
+	"VMOVAPS": {1, 0, 0x28, 0x29, 0, 0, 0, 0, true, false, false},
+	"VMOVAPD": {1, 1, 0x28, 0x29, 0, 0, 0, 0, true, false, false},
 }
 
 // isVex reports whether the mnemonic is a VEX-encoded instruction we handle.

asm/vex_test.go

@@ -39,11 +39,14 @@ func TestVexNDS3(t *testing.T) {
 		}
 		inst, err := x86asm.Decode(code, 64)
 		if err != nil {
-			t.Errorf("%s: Decode(% x): %v", mnem, code, err)
+			t.Errorf("%s: Decode(% x): %v", mnem, err, code)
 			continue
 		}
-		if inst.Op.String() != mnem {
-			t.Errorf("%s: decoded as %s (% x)", mnem, inst.Op.String(), code)
+		// The decoder folds the Plan 9 L/Q GPR-width spellings (VCVTSI2SDL/
+		// SDQ, SSL/SSQ) onto the base name; the W bit carries the width.
+		got := inst.Op.String()
+		if got != mnem && !(len(mnem) > len(got) && mnem[:len(got)] == got) {
+			t.Errorf("%s: decoded as %s (% x)", mnem, got, code)
 		}
 	}
 }

cmd/gasm/main.go

@@ -28,7 +28,7 @@ import (
 
 // version is the release version, stamped at build time via
 // -ldflags "-X main.version=…" (defaulting to the current release).
-var version = "0.11.0"
+var version = "0.16.0"
 
 func main() {
 	if len(os.Args) < 2 {
@@ -94,6 +94,7 @@ Examples:
   gasm lint go-flac/*.s     run static checks over the kernels
   gasm asm -o k.bin kern_amd64.s
   gasm asm --format elf -o k.o kern_amd64.s
+  gasm asm --format goobj -p pkg/path -o k.o kern_amd64.s
 `, version)
 }
 
@@ -340,7 +341,7 @@ hover, document symbols, diagnostics and semantic-token highlighting.
 }
 
 func cmdAsm(args []string) int {
-	fs := newCommand("asm", "gasm asm [--format raw|elf|macho] [-o out] <file>", `
+	fs := newCommand("asm", "gasm asm [--format raw|elf|macho|goobj] [-p pkg] [-o out] <file>", `
 Assemble FILE (amd64) without the Go toolchain: every TEXT function is
 encoded to machine code — scalar, VEX/AVX2 and EVEX/AVX-512 instructions,
 FP/SP frame mapping, local labels and file-local static symbols (GLOBL/DATA)
@@ -350,14 +351,16 @@ With -o the output is written to a file instead.  The --format flag selects
 what is written: raw (the default) concatenates the functions and the data
 section into one self-consistent image; elf and macho emit a relocatable
 object (.text/.data sections, a symbol table and one PC32 relocation per
-static-symbol reference) that links with the system toolchain — references
-to symbols no GLOBL in the file defines become undefined external symbols.
+static-symbol reference) that links with the system toolchain; goobj emits
+the Go toolchain's own object format, which cmd/link consumes directly (it
+requires -p, the package path, and the installed Go toolchain).
 `)
 	out := fs.String("o", "", "write the output to this file")
-	format := fs.String("format", "raw", "output format: raw (concatenated image), elf or macho (relocatable object)")
+	format := fs.String("format", "raw", "output format: raw (concatenated image), elf, macho or goobj (Go object)")
+	pkg := fs.String("p", "", "package path for --format goobj (qualifies the exported symbols)")
 	fs.Parse(args)
 	if fs.NArg() != 1 {
-		fmt.Fprintln(os.Stderr, "usage: gasm asm [--format raw|elf|macho] [-o out] <file>")
+		fmt.Fprintln(os.Stderr, "usage: gasm asm [--format raw|elf|macho|goobj] [-p pkg] [-o out] <file>")
 		return 2
 	}
 	path := fs.Arg(0)
@@ -444,8 +447,11 @@ to symbols no GLOBL in the file defines become undefined external symbols.
 		case "macho":
 			obj, err = img.MachOObject()
 			kind = "Mach-O object"
+		case "goobj":
+			obj, err = img.GOObject(*pkg, path)
+			kind = "Go object"
 		default:
-			fmt.Fprintf(os.Stderr, "gasm asm: unknown format %q (want raw, elf or macho)\n", *format)
+			fmt.Fprintf(os.Stderr, "gasm asm: unknown format %q (want raw, elf, macho or goobj)\n", *format)
 			return 2
 		}
 		if err != nil {

docs/ARCHITECTURE.md

@@ -228,11 +228,31 @@ explicit merging/zeroing masks — written the way Go writes them, as a K
 operand among the operands plus a `.Z` mnemonic suffix), and the compressed
 disp8×N displacement, whose multiplier follows the memory operand's size —
 covering every instruction the go-flac and go-lz4 AVX2/AVX-512 kernels use,
-plus the common AVX-512 F/BW integer set and the floating-point and
-conversion set (the packed double arithmetic, the scalar SD/SS forms —
-whose EVEX encodings serve masked and zeroing use — `VMOVDDUP`, and the
-width-changing conversions, including the `VCVTPD2DQ`/`VCVTTPD2DQ` family
-whose length follows the wider source operand).  Every encoding is validated two ways: by
+plus the common AVX-512 F/BW integer set, the floating-point and conversion
+set (the packed double and single arithmetic, the scalar SD/SS forms —
+whose EVEX encodings serve masked and zeroing use — `VMOVDDUP`, the
+replicating moves, and the width-changing conversions, including the
+`VCVTPD2DQ`/`VCVTTPD2DQ` family whose length follows the wider source
+operand), and the wider AVX-512 set: ternary logic, lane shuffles, inserts
+and extracts, compares with an opmask destination, the permutes, the
+expand/compress family, the broadcasts, the opmask-register instructions
+(KAND/KOR/KXNOR/KADD/KUNPCK/KNOT/KSHIFTL/KORTEST and KMOVQ), the aligned
+moves and the remaining extending/narrowing moves, the floating-point
+helper and conversion tail (VRCP14*, VRSQRT14*, VGETEXP*, VGETMANT*,
+VSCALEF*, VRNDSCALE*, VREDUCE*, VFIXUPIMM*, VRANGE*, VFPCLASS* with an
+opmask destination, and the VCVT* conversions — signed, unsigned and
+truncating, including the length-suffixed X/Y spellings and the
+mask/vector conversions VPMOVM2*/VPMOV*2M, and the scalar conversions
+between vector and general-purpose registers (VCVT{,T}S{D,S}2SI{,Q} and
+the unsigned forms, VCVTSI2*/VCVTUSI2*), and gather/scatter with VSIB addressing — both the
+VEX spelling with a vector mask register and the EVEX spelling with an
+explicit K mask, where the EVEX length follows the VSIB index register,
+not the data register.  The EVEX mnemonic
+suffixes — rounding modes (.RN_SAE/.RD_SAE/.RU_SAE/.RZ_SAE),
+suppress-all-exceptions (.SAE) and memory broadcast (.BCST) — set the EVEX
+b bit and the L'L rounding-control field (broadcast keeps the vector length
+and scales disp8 by the element size), and combine with the .Z zeroing
+suffix.  Every encoding is validated two ways: by
 round-trip decoding through `golang.org/x/arch`, and byte-for-byte against
 the machine code the real Go assembler emits — a comparison that holds for
 whole functions: all 27 functions of both kernels assemble to exactly the Go
@@ -250,10 +270,19 @@ and Mach-O writers (`gasm asm --format elf|macho`) lay the code and data out
 as `.text`/`.data` (or `__text`/`__data`) sections, export a symbol per
 `TEXT` and `GLOBL` (the `<>` ones local, the rest global) and emit one
 PC-relative relocation per static-symbol reference — undefined external
-symbols included, so the output links with the system toolchain.  GOOBJ
-emission, the format the Go linker consumes directly, is the remaining
-piece of Phase 2 (together with the rest of the EVEX set and the other
-architectures).
+symbols included, so the output links with the system toolchain.  The GOOBJ
+emitter (`gasm asm --format goobj`) writes the format the Go linker consumes
+directly: the functions as non-package symbols (the way `cmd/asm` records
+assembly symbols), the `GLOBL` data, one `FuncInfo` per function and the
+pc-value tables — `pcsp` built from the prologue and epilogue stack
+boundaries, plus flat `pcfile`, `pcline` and `pcinline` tables — so a
+gasm-assembled object drops into a `go build` in place of the toolchain's.
+The object preamble (the version-and-experiment header the linker compares
+verbatim) is captured from the installed `go tool asm`, so the output is
+always consistent with the toolchain that links it.  External cross-package
+references and the implicit funcdata/DWARF symbols remain future work (the
+linker fills the latter's defaults); the rest of Phase 2 is those, the
+remaining EVEX forms and the other architectures.
 
 ## Extension points
 

docs/DEFERRED.md

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

justfile

@@ -3,7 +3,7 @@
 
 # gasm-devkit — developer tooling for Go's Plan 9 assembler (GAsm).
 
-version := "0.11.0"
+version := "0.16.0"
 
 default:
     @just --list

parser/parser.go

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