feat(verify): add differential fuzz testing against a portable Go reference

Assisted-by: Qwen 3.8 Max Preview

asm/evex.go

@@ -308,6 +308,106 @@ var evexTable = map[string]evexSpec{
 	// EVEX.66.0F3A — half-precision convert back ($imm, src, dst: reg=src,
 	// rm=dst, imm8 — the extract layout).
 	"VCVTPS2PH": {3, 0x1D, 0, 1, -1, vexExtract, [3]int{8, 16, 32}},
+
+	// EVEX — unsigned and truncating conversions.  The PD sources are the
+	// wide operand (the bare names are 512-bit only, the X/Y spellings fix
+	// the length); the PS/UQQ destinations are wide and follow the
+	// destination.
+	"VCVTPD2PS":    {1, 0x5A, 1, 1, -1, vexRMSrcLen, [3]int{0, 0, 64}},
+	"VCVTPD2PSX":   {1, 0x5A, 1, 1, -1, vexRMSrcLen, [3]int{16, 0, 0}},
+	"VCVTPD2PSY":   {1, 0x5A, 1, 1, -1, vexRMSrcLen, [3]int{0, 32, 0}},
+	"VCVTPD2UDQ":   {1, 0x79, 1, 0, -1, vexRMSrcLen, [3]int{0, 0, 64}},
+	"VCVTPD2UDQX":  {1, 0x79, 1, 0, -1, vexRMSrcLen, [3]int{16, 0, 0}},
+	"VCVTPD2UDQY":  {1, 0x79, 1, 0, -1, vexRMSrcLen, [3]int{0, 32, 0}},
+	"VCVTTPD2UDQ":  {1, 0x78, 1, 0, -1, vexRMSrcLen, [3]int{0, 0, 64}},
+	"VCVTTPD2UDQX": {1, 0x78, 1, 0, -1, vexRMSrcLen, [3]int{16, 0, 0}},
+	"VCVTTPD2UDQY": {1, 0x78, 1, 0, -1, vexRMSrcLen, [3]int{0, 32, 0}},
+	"VCVTTPD2UQQ":  {1, 0x78, 1, 1, -1, vexRM, [3]int{16, 32, 64}},
+	"VCVTPS2UDQ":   {1, 0x79, 0, 0, -1, vexRM, [3]int{16, 32, 64}},
+	"VCVTTPS2UDQ":  {1, 0x78, 0, 0, -1, vexRM, [3]int{16, 32, 64}},
+	"VCVTPS2UQQ":   {1, 0x79, 0, 1, -1, vexRM, [3]int{8, 16, 32}},
+	"VCVTTPS2UQQ":  {1, 0x78, 0, 1, -1, vexRM, [3]int{8, 16, 32}},
+	"VCVTTPD2QQ":   {1, 0x7A, 1, 1, -1, vexRM, [3]int{16, 32, 64}},
+	"VCVTTPS2QQ":   {1, 0x7A, 0, 1, -1, vexRM, [3]int{8, 16, 32}},
+	"VCVTUQQ2PD":   {1, 0x7A, 1, 2, -1, vexRM, [3]int{16, 32, 64}},
+	"VCVTUQQ2PS":   {1, 0x7A, 1, 3, -1, vexRMSrcLen, [3]int{0, 0, 64}},
+	"VCVTUQQ2PSX":  {1, 0x7A, 1, 3, -1, vexRMSrcLen, [3]int{16, 0, 0}},
+	"VCVTUQQ2PSY":  {1, 0x7A, 1, 3, -1, vexRMSrcLen, [3]int{0, 32, 0}},
+	"VCVTQQ2PSX":   {1, 0x5B, 1, 0, -1, vexRMSrcLen, [3]int{16, 0, 0}},
+	"VCVTQQ2PSY":   {1, 0x5B, 1, 0, -1, vexRMSrcLen, [3]int{0, 32, 0}},
+
+	// EVEX.66.0F38 — the remaining sign/zero-extending moves (narrow
+	// source; disp8×N follows its size).
+	"VPMOVSXBD": {2, 0x21, 0, 1, -1, vexRM, [3]int{4, 8, 16}},
+	"VPMOVSXBQ": {2, 0x22, 0, 1, -1, vexRM, [3]int{2, 4, 8}},
+	"VPMOVSXWD": {2, 0x23, 0, 1, -1, vexRM, [3]int{8, 16, 32}},
+	"VPMOVSXWQ": {2, 0x24, 0, 1, -1, vexRM, [3]int{4, 8, 16}},
+	"VPMOVZXBD": {2, 0x31, 0, 1, -1, vexRM, [3]int{4, 8, 16}},
+	"VPMOVZXBQ": {2, 0x32, 0, 1, -1, vexRM, [3]int{2, 4, 8}},
+	"VPMOVZXWD": {2, 0x33, 0, 1, -1, vexRM, [3]int{8, 16, 32}},
+	"VPMOVZXWQ": {2, 0x34, 0, 1, -1, vexRM, [3]int{4, 8, 16}},
+	"VPMOVZXDQ": {2, 0x35, 0, 1, -1, vexRM, [3]int{8, 16, 32}},
+
+	// EVEX.F3.0F38 — the remaining narrowing stores (vector source in reg,
+	// narrow destination in r/m): signed, unsigned and the D/Q truncations.
+	"VPMOVSDB":  {2, 0x21, 0, 2, -1, vexRMRev, [3]int{4, 8, 16}},
+	"VPMOVSQB":  {2, 0x22, 0, 2, -1, vexRMRev, [3]int{2, 4, 8}},
+	"VPMOVSDW":  {2, 0x23, 0, 2, -1, vexRMRev, [3]int{8, 16, 32}},
+	"VPMOVSQW":  {2, 0x24, 0, 2, -1, vexRMRev, [3]int{4, 8, 16}},
+	"VPMOVSQD":  {2, 0x25, 0, 2, -1, vexRMRev, [3]int{8, 16, 32}},
+	"VPMOVSWB":  {2, 0x20, 0, 2, -1, vexRMRev, [3]int{8, 16, 32}},
+	"VPMOVUSWB": {2, 0x10, 0, 2, -1, vexRMRev, [3]int{8, 16, 32}},
+	"VPMOVUSDB": {2, 0x11, 0, 2, -1, vexRMRev, [3]int{4, 8, 16}},
+	"VPMOVUSQB": {2, 0x12, 0, 2, -1, vexRMRev, [3]int{2, 4, 8}},
+	"VPMOVUSDW": {2, 0x13, 0, 2, -1, vexRMRev, [3]int{8, 16, 32}},
+	"VPMOVUSQW": {2, 0x14, 0, 2, -1, vexRMRev, [3]int{4, 8, 16}},
+	"VPMOVUSQD": {2, 0x15, 0, 2, -1, vexRMRev, [3]int{8, 16, 32}},
+	"VPMOVDB":   {2, 0x31, 0, 2, -1, vexRMRev, [3]int{4, 8, 16}},
+	"VPMOVQW":   {2, 0x34, 0, 2, -1, vexRMRev, [3]int{4, 8, 16}},
+
+	// EVEX.F3.0F38 — mask/vector conversions: M2* moves an opmask register
+	// into a vector (rm = K source, reg = vector destination), *2M does the
+	// reverse (reg = K destination, rm = vector source, the length follows
+	// the vector).
+	"VPMOVM2B": {2, 0x28, 0, 2, -1, vexRM, [3]int{16, 32, 64}},
+	"VPMOVM2W": {2, 0x28, 1, 2, -1, vexRM, [3]int{16, 32, 64}},
+	"VPMOVM2D": {2, 0x38, 0, 2, -1, vexRM, [3]int{16, 32, 64}},
+	"VPMOVM2Q": {2, 0x38, 1, 2, -1, vexRM, [3]int{16, 32, 64}},
+	"VPMOVB2M": {2, 0x29, 0, 2, -1, vexRM, [3]int{16, 32, 64}},
+	"VPMOVW2M": {2, 0x29, 1, 2, -1, vexRM, [3]int{16, 32, 64}},
+	"VPMOVD2M": {2, 0x39, 0, 2, -1, vexRM, [3]int{16, 32, 64}},
+	"VPMOVQ2M": {2, 0x39, 1, 2, -1, vexRM, [3]int{16, 32, 64}},
+
+	// EVEX — scalar conversions between vector and general-purpose
+	// registers.  Vector to GPR (two operands: vec/mem source, GPR
+	// destination, vvvv unused): the signed and truncated pair, and the
+	// unsigned forms (EVEX only).
+	"VCVTSD2SI":    {1, 0x2D, 0, 3, -1, vexRM, [3]int{8, 8, 8}},
+	"VCVTSD2SIQ":   {1, 0x2D, 1, 3, -1, vexRM, [3]int{8, 8, 8}},
+	"VCVTSS2SI":    {1, 0x2D, 0, 2, -1, vexRM, [3]int{4, 4, 4}},
+	"VCVTSS2SIQ":   {1, 0x2D, 1, 2, -1, vexRM, [3]int{4, 4, 4}},
+	"VCVTTSD2SI":   {1, 0x2C, 0, 3, -1, vexRM, [3]int{8, 8, 8}},
+	"VCVTTSD2SIQ":  {1, 0x2C, 1, 3, -1, vexRM, [3]int{8, 8, 8}},
+	"VCVTTSS2SI":   {1, 0x2C, 0, 2, -1, vexRM, [3]int{4, 4, 4}},
+	"VCVTTSS2SIQ":  {1, 0x2C, 1, 2, -1, vexRM, [3]int{4, 4, 4}},
+	"VCVTSD2USIL":  {1, 0x79, 0, 3, -1, vexRM, [3]int{8, 8, 8}},
+	"VCVTSD2USIQ":  {1, 0x79, 1, 3, -1, vexRM, [3]int{8, 8, 8}},
+	"VCVTSS2USIL":  {1, 0x79, 0, 2, -1, vexRM, [3]int{4, 4, 4}},
+	"VCVTSS2USIQ":  {1, 0x79, 1, 2, -1, vexRM, [3]int{4, 4, 4}},
+	"VCVTTSD2USIL": {1, 0x78, 0, 3, -1, vexRM, [3]int{8, 8, 8}},
+	"VCVTTSD2USIQ": {1, 0x78, 1, 3, -1, vexRM, [3]int{8, 8, 8}},
+	"VCVTTSS2USIL": {1, 0x78, 0, 2, -1, vexRM, [3]int{4, 4, 4}},
+	"VCVTTSS2USIQ": {1, 0x78, 1, 2, -1, vexRM, [3]int{4, 4, 4}},
+	// 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, [3]int{4, 4, 4}},
+	"VCVTSI2SDQ":  {1, 0x2A, 1, 3, -1, vexNDS3, [3]int{8, 8, 8}},
+	"VCVTSI2SSL":  {1, 0x2A, 0, 2, -1, vexNDS3, [3]int{4, 4, 4}},
+	"VCVTSI2SSQ":  {1, 0x2A, 1, 2, -1, vexNDS3, [3]int{8, 8, 8}},
+	"VCVTUSI2SDL": {1, 0x7B, 0, 3, -1, vexNDS3, [3]int{4, 4, 4}},
+	"VCVTUSI2SDQ": {1, 0x7B, 1, 3, -1, vexNDS3, [3]int{8, 8, 8}},
+	"VCVTUSI2SSL": {1, 0x7B, 0, 2, -1, vexNDS3, [3]int{4, 4, 4}},
+	"VCVTUSI2SSQ": {1, 0x7B, 1, 2, -1, vexNDS3, [3]int{8, 8, 8}},
 	// 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).
@@ -543,6 +643,15 @@ var evexRound = map[string]bool{
 	"VSCALEFPD": true, "VSCALEFPS": true, "VSCALEFSD": true, "VSCALEFSS": true,
 	"VGETEXPPD": true, "VGETEXPPS": true, "VGETEXPSD": true, "VGETEXPSS": true,
 	"VCVTDQ2PS": true, "VCVTPS2QQ": true, "VCVTQQ2PS": true, "VCVTPD2UQQ": true,
+	"VCVTPD2PS": true, "VCVTPD2UDQ": true, "VCVTTPD2UDQ": true, "VCVTTPD2UQQ": true,
+	"VCVTPS2UDQ": true, "VCVTTPS2UDQ": true, "VCVTPS2UQQ": true, "VCVTTPS2UQQ": true,
+	"VCVTTPD2QQ": true, "VCVTTPS2QQ": true, "VCVTUQQ2PD": true, "VCVTUQQ2PS": true,
+	"VCVTSD2SI": true, "VCVTSD2SIQ": true, "VCVTSS2SI": true, "VCVTSS2SIQ": true,
+	"VCVTSD2USIL": true, "VCVTSD2USIQ": true, "VCVTSS2USIL": true, "VCVTSS2USIQ": true,
+	"VCVTTSD2SI": true, "VCVTTSD2SIQ": true, "VCVTTSS2SI": true, "VCVTTSS2SIQ": true,
+	"VCVTTSD2USIL": true, "VCVTTSD2USIQ": true, "VCVTTSS2USIL": true, "VCVTTSS2USIQ": true,
+	"VCVTSI2SDQ": true, "VCVTSI2SSL": true, "VCVTSI2SSQ": true,
+	"VCVTUSI2SDQ": true, "VCVTUSI2SSL": true, "VCVTUSI2SSQ": true,
 }
 
 // evexBcstN maps an instruction accepting .BCST to the broadcast element
@@ -562,6 +671,9 @@ var evexBcstN = map[string]int{
 	"VRANGEPD": 8, "VRANGEPS": 4,
 	"VCVTDQ2PS": 4, "VCVTPS2QQ": 4, "VCVTQQ2PS": 8,
 	"VCVTUDQ2PD": 4, "VCVTUDQ2PS": 4,
+	"VCVTPD2PS": 8, "VCVTPD2UDQ": 8, "VCVTTPD2UDQ": 8, "VCVTTPD2UQQ": 8,
+	"VCVTPS2UDQ": 4, "VCVTTPS2UDQ": 4, "VCVTPS2UQQ": 4, "VCVTTPS2UQQ": 4,
+	"VCVTTPD2QQ": 8, "VCVTTPS2QQ": 4, "VCVTUQQ2PD": 8, "VCVTUQQ2PS": 8,
 }
 
 // splitMask extracts an explicit mask register (K1–K7) from the operand list,
@@ -591,6 +703,18 @@ func splitMask(ops []Operand) ([]Operand, int, error) {
 // operands; the mnemonic suffix carries zeroing, rounding/SAE and
 // broadcast.
 func (e *enc) encodeEvex(mnemUpper string, ops []Operand, sfx evexSuffix) error {
+	// The mask/vector conversions take the K register as a genuine operand
+	// (source or destination), not as a mask, and accept no suffixes.
+	if evexKOperand[mnemUpper] {
+		if sfx.any() {
+			return fmt.Errorf("%s takes no EVEX suffixes", mnemUpper)
+		}
+		spec, ok := evexTable[mnemUpper]
+		if !ok {
+			return fmt.Errorf("unsupported instruction %q", mnemUpper)
+		}
+		return e.encodeEvexRM(spec, ops, 0, sfx)
+	}
 	spec, inTable := evexTable[mnemUpper]
 	if inTable {
 		if (sfx.rounding >= 0 || sfx.sae) && !evexRound[mnemUpper] {
@@ -713,17 +837,29 @@ func (e *enc) encodeEvexNDS3(spec evexSpec, ops []Operand, mask int, sfx evexSuf
 }
 
 // encodeEvexRM encodes the two-operand form: OP src, dst (reg=dst, rm=src,
-// no vvvv), e.g. VCVTQQ2PD.
+// no vvvv), e.g. VCVTQQ2PD.  The destination may be an opmask register (the
+// *2M mask conversions) or a general-purpose register (the scalar
+// vector-to-GPR conversions); in both cases the vector length comes from
+// the source.
 func (e *enc) encodeEvexRM(spec evexSpec, ops []Operand, mask int, sfx evexSuffix) 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() {
+	if !ok {
-		return fmt.Errorf("EVEX destination must be a vector register")
+		return fmt.Errorf("EVEX destination must be a register")
 	}
-	return e.emitEvexFields(spec, dstReg.vecLenBit(), dstReg.idx, -1, src, mask, sfx)
+	ll := dstReg.vecLenBit()
+	if !dstReg.isVec() {
+		// Mask or GPR destination: the length follows the vector source
+		// (128 for a memory source).
+		ll = 0
+		if r, ok := src.(Reg); ok && r.isVec() {
+			ll = r.vecLenBit()
+		}
+	}
+	return e.emitEvexFields(spec, ll, dstReg.idx, -1, src, mask, sfx)
 }
 
 // encodeEvexImmRM encodes the immediate shuffle form: OP $imm, src, dst
@@ -1260,6 +1396,14 @@ func (e *enc) encodeScatter(upper string, ss gatherSpec, ops []Operand, sfx evex
 	return e.emitEvexFields(evex, ll, src.idx, -1, vsib, mask, sfx)
 }
 
+// evexKOperand lists the instructions whose K register is a genuine operand
+// (the source or destination of a mask/vector conversion) rather than a
+// mask modifier — the M2 and 2M conversions.  They take no masking.
+var evexKOperand = map[string]bool{
+	"VPMOVM2B": true, "VPMOVM2W": true, "VPMOVM2D": true, "VPMOVM2Q": true,
+	"VPMOVB2M": true, "VPMOVW2M": true, "VPMOVD2M": true, "VPMOVQ2M": true,
+}
+
 // kmovSpec describes a KMOV width: the opcode depends on the operand
 // direction — kk (k/mem → K is 90, k → k uses the same), kmem (K → mem),
 // gprk (GPR/mem → K), kgpr (K → GPR) — and the GPR forms carry a mandatory

asm/evex_test.go

@@ -467,6 +467,164 @@ func TestEvexHelperGroundTruth(t *testing.T) {
 	}
 }
 
+// TestEvexGprGroundTruth covers the scalar conversions between vector and
+// general-purpose registers — the signed and truncated VCVT{,T}S{D,S}2SI
+// 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/vex.go

@@ -130,9 +130,15 @@ var vexTable = map[string]vexSpec{
 	// no vvvv).
 	"VPMOVSXWD":    {2, 0x23, 0, 1, -1, vexRM},
 	"VPMOVSXDQ":    {2, 0x25, 0, 1, -1, vexRM},
-	"VPMOVSXBW":    {2, 0x20, 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
@@ -198,6 +204,29 @@ var vexTable = map[string]vexSpec{
 	// 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},
@@ -222,6 +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

asm/vex_test.go

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

cmd/gasm/main.go

@@ -24,11 +24,12 @@ import (
 	"sourcedock.dev/petrbalvin/gasm-devkit/lint"
 	"sourcedock.dev/petrbalvin/gasm-devkit/lsp"
 	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+	"sourcedock.dev/petrbalvin/gasm-devkit/verify"
 )
 
 // version is the release version, stamped at build time via
 // -ldflags "-X main.version=…" (defaulting to the current release).
-var version = "0.14.0"
+var version = "0.18.0"
 
 func main() {
 	if len(os.Args) < 2 {
@@ -46,6 +47,8 @@ func main() {
 		os.Exit(cmdLint(os.Args[2:]))
 	case "asm":
 		os.Exit(cmdAsm(os.Args[2:]))
+	case "verify":
+		os.Exit(cmdVerify(os.Args[2:]))
 	case "lsp":
 		os.Exit(cmdLSP(os.Args[2:]))
 	case "version", "--version", "-V":
@@ -80,6 +83,7 @@ Commands:
   fmt       canonicalise formatting (gofmt for assembly)
   lint      run static checks
   asm       assemble .s files to machine code (amd64)
+  verify    JIT-assemble and run dynamic checks (amd64)
   lsp       run the language server over stdio
   version   print the version (same as --version)
 
@@ -466,3 +470,57 @@ requires -p, the package path, and the installed Go toolchain).
 	}
 	return 0
 }
+
+func cmdVerify(args []string) int {
+	fs := newCommand("verify", "gasm verify <file.s>", `
+Assemble FILE (amd64), map it into executable memory and report the available
+functions.  This confirms the assembled image is self-consistent (no
+unresolved external symbols) and executable — the prerequisite for dynamic
+testing.
+
+With -smoke, each NOSPLIT function is called with a zeroed argument block to
+confirm the JIT trampoline works end-to-end.  This is safe only for functions
+that tolerate nil pointers and zero lengths in their arguments.
+`)
+	smoke := fs.Bool("smoke", false, "call each NOSPLIT function with zeroed args")
+	fs.Parse(args)
+	if fs.NArg() != 1 {
+		fmt.Fprintln(os.Stderr, "usage: gasm verify [-smoke] <file.s>")
+		return 2
+	}
+	path := fs.Arg(0)
+	if arch.FromFilename(path) != arch.AMD64 {
+		fmt.Fprintln(os.Stderr, "gasm verify: only amd64 is supported")
+		return 1
+	}
+
+	k, err := verify.Load(path)
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "gasm verify: %v\n", err)
+		return 1
+	}
+	defer k.Close()
+
+	names := k.FuncNames()
+	fmt.Printf("%s: %d functions JIT-loaded\n", path, len(names))
+	rc := 0
+	for _, name := range names {
+		fl, _ := k.Func(name)
+		flags := ""
+		if fl.NoSplit {
+			flags = " NOSPLIT"
+		}
+		fmt.Printf("  %s: %d bytes, args=%d, frame=%d%s\n", name, fl.Size, fl.Args, fl.Frame, flags)
+		if *smoke && fl.NoSplit {
+			args := make([]byte, fl.Args)
+			_, err := k.CallFunc(name, args)
+			if err != nil {
+				fmt.Printf("    smoke: FAIL — %v\n", err)
+				rc = 1
+			} else {
+				fmt.Printf("    smoke: OK\n")
+			}
+		}
+	}
+	return rc
+}

docs/ARCHITECTURE.md

@@ -240,8 +240,11 @@ expand/compress family, the broadcasts, the opmask-register instructions
 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 remaining VCVT* conversions including the
+opmask destination, and the VCVT* conversions — signed, unsigned and
-half-precision pair), and gather/scatter with VSIB addressing — both the
+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
@@ -281,6 +284,31 @@ references and the implicit funcdata/DWARF symbols remain future work (the
 linker fills the latter's defaults); the rest of Phase 2 is those, the
 remaining EVEX forms and the other architectures.
 
+### `verify`
+
+The dynamic-analysis substrate (Phase 3).  It JIT-loads assembled images into
+executable memory and invokes them directly, enabling differential testing,
+runtime ABI checks and coverage profiling.
+
+The execution model is pure Go (stdlib only).  `Map` copies machine code into
+an anonymous `syscall.Mmap` mapping and enforces W^X (write the bytes, then
+`mprotect` to read-execute).  `Call` prepares a stack whose first word is the
+address of an assembly trampoline (`leaveJIT`), lays the ABI0 argument
+block after it, switches to that stack via `enterJIT` (which saves the Go
+stack pointer in a package global and jumps to the target), and recovers
+control when the function RETs into `leaveJIT` (which restores the Go stack
+and returns).  A 64-byte pad below the return address accommodates the
+ABIInternal wrapper that the Go runtime interposes on assembly functions.
+
+`Load` / `LoadSource` / `LoadAST` parse, assemble and map a `.s` file in one
+step, returning a `Kernel` whose `CallFunc` method marshals the argument block
+by name.  The image must be self-contained (no external relocations); the
+assembler’s `Image.Bytes()` provides the code-and-data concatenation.
+
+The `gasm verify` CLI subcommand exposes this: it loads a file, reports the
+available functions and (with `-smoke`) calls each NOSPLIT function with zeroed
+arguments to confirm the trampoline round-trips.
+
 ## Extension points
 
 - **New architecture:** add an entry to the generator in `_gen`, run

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 := "0.14.0"
+version := "0.18.0"
 
 default:
     @just --list

testdata/verify/basic_amd64.s

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

verify/call_amd64.go

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

verify/call_other.go

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

verify/differential_test.go

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

verify/jit.go

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

verify/jit_test.go

@@ -0,0 +1,159 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package verify
+
+import (
+	"bytes"
+	"testing"
+	"unsafe"
+)
+
+func loadBasic(t *testing.T) *Kernel {
+	t.Helper()
+	k, err := Load("../testdata/verify/basic_amd64.s")
+	if err != nil {
+		t.Fatalf("Load: %v", err)
+	}
+	t.Cleanup(k.Close)
+	return k
+}
+
+func TestJITAdd(t *testing.T) {
+	k := loadBasic(t)
+
+	tests := []struct {
+		a, b, want int64
+	}{
+		{0, 0, 0},
+		{1, 2, 3},
+		{-1, 1, 0},
+		{1 << 62, 1 << 62, -9223372036854775808}, // overflow wraps (MinInt64)
+		{-100, -200, -300},
+	}
+	for _, tt := range tests {
+		args := make([]byte, 24)
+		PutUint64(args, 0, uint64(tt.a))
+		PutUint64(args, 8, uint64(tt.b))
+
+		out, err := k.CallFunc("add", args)
+		if err != nil {
+			t.Fatalf("CallFunc(add, %d, %d): %v", tt.a, tt.b, err)
+		}
+		got := int64(GetUint64(out, 16))
+		if got != tt.want {
+			t.Errorf("add(%d, %d) = %d, want %d", tt.a, tt.b, got, tt.want)
+		}
+	}
+}
+
+func TestJITSum(t *testing.T) {
+	k := loadBasic(t)
+
+	tests := []struct {
+		data []int64
+		want int64
+	}{
+		{nil, 0},
+		{[]int64{1}, 1},
+		{[]int64{1, 2, 3, 4, 5}, 15},
+		{[]int64{-10, 20, -30, 40}, 20},
+	}
+	for _, tt := range tests {
+		args := make([]byte, 32)
+		if len(tt.data) > 0 {
+			PutPtr(args, 0, unsafe.Pointer(&tt.data[0]))
+		}
+		PutUint64(args, 8, uint64(len(tt.data)))
+		PutUint64(args, 16, uint64(cap(tt.data)))
+
+		out, err := k.CallFunc("sum", args)
+		if err != nil {
+			t.Fatalf("CallFunc(sum, %v): %v", tt.data, err)
+		}
+		got := int64(GetUint64(out, 24))
+		if got != tt.want {
+			t.Errorf("sum(%v) = %d, want %d", tt.data, got, tt.want)
+		}
+	}
+}
+
+func TestJITWideCopy(t *testing.T) {
+	k := loadBasic(t)
+
+	tests := []struct {
+		name string
+		n    int
+	}{
+		{"empty", 0},
+		{"tiny", 7},
+		{"exact32", 32},
+		{"overlap_range", 48},
+		{"exact64", 64},
+		{"unaligned", 45},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			src := make([]byte, tt.n)
+			for i := range src {
+				src[i] = byte(i * 7)
+			}
+			dst := make([]byte, tt.n)
+
+			args := make([]byte, 48)
+			if tt.n > 0 {
+				PutPtr(args, 0, unsafe.Pointer(&dst[0]))
+				PutPtr(args, 24, unsafe.Pointer(&src[0]))
+			}
+			PutUint64(args, 8, uint64(tt.n))  // dst_len
+			PutUint64(args, 16, uint64(tt.n)) // dst_cap
+			PutUint64(args, 32, uint64(tt.n)) // src_len
+			PutUint64(args, 40, uint64(tt.n)) // src_cap
+
+			_, err := k.CallFunc("wideCopy", args)
+			if err != nil {
+				t.Fatalf("CallFunc(wideCopy): %v", err)
+			}
+			if !bytes.Equal(dst, src) {
+				t.Errorf("wideCopy: dst ≠ src\n  got  %x\n  want %x", dst, src)
+			}
+		})
+	}
+}
+
+func TestKernelFuncNames(t *testing.T) {
+	k := loadBasic(t)
+	names := k.FuncNames()
+	want := []string{"add", "sum", "wideCopy"}
+	if len(names) != len(want) {
+		t.Fatalf("FuncNames() = %v, want %v", names, want)
+	}
+	for i, n := range names {
+		if n != want[i] {
+			t.Errorf("FuncNames()[%d] = %q, want %q", i, n, want[i])
+		}
+	}
+}
+
+func TestKernelFuncNotFound(t *testing.T) {
+	k := loadBasic(t)
+	_, err := k.CallFunc("nonexistent", make([]byte, 8))
+	if err == nil {
+		t.Fatal("expected error for nonexistent function")
+	}
+}
+
+func TestKernelArgTooSmall(t *testing.T) {
+	k := loadBasic(t)
+	_, err := k.CallFunc("add", make([]byte, 8)) // needs 24
+	if err == nil {
+		t.Fatal("expected error for too-small arg block")
+	}
+}
+
+func TestMapZeroLength(t *testing.T) {
+	_, err := Map(nil)
+	if err == nil {
+		t.Fatal("expected error for zero-length code")
+	}
+}

verify/lz4_test.go

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

verify/trampoline_amd64.s

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

verify/verify.go

@@ -0,0 +1,106 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package verify
+
+import (
+	"fmt"
+	"os"
+
+	"sourcedock.dev/petrbalvin/gasm-devkit/asm"
+	"sourcedock.dev/petrbalvin/gasm-devkit/ast"
+	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+)
+
+// Kernel is a JIT-loaded assembly image ready for direct invocation.
+// It wraps an executable memory mapping and the function layout metadata
+// needed to marshal ABI0 calls.
+type Kernel struct {
+	exec  *Executable
+	img   *asm.Image
+	funcs map[string]int // function name → index into img.Funcs
+}
+
+// Load parses, assembles and maps a .s file into executable memory.
+// The returned Kernel is ready for Call.  The caller must call Close to
+// release the mapping.
+func Load(path string) (*Kernel, error) {
+	src, err := os.ReadFile(path)
+	if err != nil {
+		return nil, fmt.Errorf("verify: %w", err)
+	}
+	return LoadSource(path, string(src))
+}
+
+// LoadSource parses, assembles and maps assembly source into executable memory.
+func LoadSource(filename, src string) (*Kernel, error) {
+	file, errs := parser.Parse(filename, src)
+	if len(errs) > 0 {
+		return nil, fmt.Errorf("verify: parse %s: %v", filename, errs[0])
+	}
+	return LoadAST(file)
+}
+
+// LoadAST assembles a parsed AST file and maps the result into executable
+// memory.
+func LoadAST(file *ast.File) (*Kernel, error) {
+	img, err := asm.AssembleFile(file)
+	if err != nil {
+		return nil, fmt.Errorf("verify: assemble: %w", err)
+	}
+	if len(img.Externals) > 0 {
+		return nil, fmt.Errorf("verify: unresolved external symbols: %v", img.Externals)
+	}
+	code := img.Bytes()
+	exec, err := Map(code)
+	if err != nil {
+		return nil, err
+	}
+	funcs := make(map[string]int, len(img.Funcs))
+	for i, f := range img.Funcs {
+		funcs[f.Name] = i
+	}
+	return &Kernel{exec: exec, img: img, funcs: funcs}, nil
+}
+
+// Func returns the layout metadata for the named function.
+func (k *Kernel) Func(name string) (asm.FuncLayout, error) {
+	idx, ok := k.funcs[name]
+	if !ok {
+		return asm.FuncLayout{}, fmt.Errorf("verify: function %q not found", name)
+	}
+	return k.img.Funcs[idx], nil
+}
+
+// FuncNames returns the names of all functions in the kernel, in source order.
+func (k *Kernel) FuncNames() []string {
+	names := make([]string, len(k.img.Funcs))
+	for i, f := range k.img.Funcs {
+		names[i] = f.Name
+	}
+	return names
+}
+
+// CallFunc invokes the named function with the given ABI0 argument block.
+// The arg block is the raw bytes of the function's argument/result area
+// (as declared by the TEXT $frame-args suffix).  Returns the arg block
+// after the call (with any results written back by the function).
+func (k *Kernel) CallFunc(name string, args []byte) ([]byte, error) {
+	idx, ok := k.funcs[name]
+	if !ok {
+		return nil, fmt.Errorf("verify: function %q not found", name)
+	}
+	fl := k.img.Funcs[idx]
+	if len(args) < fl.Args {
+		return nil, fmt.Errorf("verify: %s: arg block too small: got %d, need %d", name, len(args), fl.Args)
+	}
+	fnAddr := k.exec.FuncAddr(fl.Offset)
+	return Call(fnAddr, args)
+}
+
+// Close releases the executable mapping.
+func (k *Kernel) Close() {
+	if k.exec != nil {
+		k.exec.Unmap()
+	}
+}