feat(asm): add vpcmp compare, full opmask set, legacy sse integers and bswap

Assisted-by: GLM 5.3

asm/encode.go

@@ -82,8 +82,12 @@ func (e *enc) encode(mnem string, ops []Operand) error {
 	if m, ok := sseShufTable[upper]; ok {
 		return e.encodeSSEShuf(m, ops)
 	}
-	// Legacy SSE packed binary ops and imm8 shuffles (ADDPS/MULPS/
-	// SHUFPS/PSHUFD/...): no size suffix, dispatch on the full name.
+	// Legacy SSE packed binaries dispatch on the full name: the packed
+	// integer mnemonics carry real width suffixes (PADDB/PCMPGTW/...),
+	// which the size split must not eat.
+	if m, ok := sseBinTable[upper]; ok {
+		return e.encodeSSEBin(m, ops)
+	}
 	if m, ok := sseBinTable[base]; ok {
 		return e.encodeSSEBin(m, ops)
 	}
@@ -108,6 +112,8 @@ func (e *enc) encode(mnem string, ops []Operand) error {
 		return e.encodePushPop(ops, false)
 	case "BSF", "BSR", "LZCNT", "TZCNT", "POPCNT":
 		return e.encodeCount(base, ops, size)
+	case "BSWAP":
+		return e.encodeBswap(ops, size)
 	case "MOVBLZX", "MOVBQZX", "MOVWLZX", "MOVWQZX", "MOVWLSX", "MOVLQSX":
 		return e.encodeMovExtend(base, ops)
 	case "CVTSL2SD", "CVTSQ2SD":

asm/evex.go

@@ -166,6 +166,19 @@ var evexTable = map[string]evexSpec{
 	"VCMPSD": {1, 0xC2, 1, 3, -1, vexNDS3Imm, [3]int{8, 8, 8}},
 	"VCMPSS": {1, 0xC2, 0, 2, -1, vexNDS3Imm, [3]int{4, 4, 4}},
 
+	// EVEX.66.0F3A — integer compares with an opmask destination, the same
+	// NDS3Imm-with-k-reg shape as the floating-point compares; W selects the
+	// operand width (byte/word vs dword/qword), the opcode the signedness.
+	// The memory form takes a full vector, so disp8×N is 16/32/64.
+	"VPCMPB":  {3, 0x3F, 0, 1, -1, vexNDS3Imm, [3]int{16, 32, 64}},
+	"VPCMPUB": {3, 0x3E, 0, 1, -1, vexNDS3Imm, [3]int{16, 32, 64}},
+	"VPCMPW":  {3, 0x3F, 1, 1, -1, vexNDS3Imm, [3]int{16, 32, 64}},
+	"VPCMPUW": {3, 0x3E, 1, 1, -1, vexNDS3Imm, [3]int{16, 32, 64}},
+	"VPCMPD":  {3, 0x1F, 0, 1, -1, vexNDS3Imm, [3]int{16, 32, 64}},
+	"VPCMPUD": {3, 0x1E, 0, 1, -1, vexNDS3Imm, [3]int{16, 32, 64}},
+	"VPCMPQ":  {3, 0x1F, 1, 1, -1, vexNDS3Imm, [3]int{16, 32, 64}},
+	"VPCMPUQ": {3, 0x1E, 1, 1, -1, vexNDS3Imm, [3]int{16, 32, 64}},
+
 	// EVEX.66.0F38 — permutes (NDS form).
 	"VPERMB":    {2, 0x8D, 0, 1, -1, vexNDS3, [3]int{16, 32, 64}},
 	"VPERMW":    {2, 0x8D, 1, 1, -1, vexNDS3, [3]int{16, 32, 64}},
@@ -1471,24 +1484,60 @@ type kOpSpec struct {
 
 var kOpsTable = map[string]kOpSpec{
 	// k ← k OP k: reg = dst, vvvv = src1, rm = src2 (three opmask
-	// registers).
+	// registers).  Byte/word widths share W0 and differ by the 66 prefix;
+	// dword/qword share the W bit selection the Go assembler emits.
 	"KANDB":    {1, 0x41, 0, 1, 1, vexNDS3},
 	"KANDW":    {1, 0x41, 0, 0, 1, vexNDS3},
+	"KANDD":    {1, 0x41, 1, 1, 1, vexNDS3},
 	"KANDQ":    {1, 0x41, 1, 0, 1, vexNDS3},
+	"KANDNB":   {1, 0x42, 0, 1, 1, vexNDS3},
+	"KANDNW":   {1, 0x42, 0, 0, 1, vexNDS3},
+	"KANDND":   {1, 0x42, 1, 1, 1, vexNDS3},
+	"KANDNQ":   {1, 0x42, 1, 0, 1, vexNDS3},
 	"KORB":     {1, 0x45, 0, 1, 1, vexNDS3},
+	"KORW":     {1, 0x45, 0, 0, 1, vexNDS3},
 	"KORD":     {1, 0x45, 1, 1, 1, vexNDS3},
+	"KORQ":     {1, 0x45, 1, 0, 1, vexNDS3},
+	"KXNORB":   {1, 0x46, 0, 1, 1, vexNDS3},
 	"KXNORW":   {1, 0x46, 0, 0, 1, vexNDS3},
+	"KXNORD":   {1, 0x46, 1, 1, 1, vexNDS3},
 	"KXNORQ":   {1, 0x46, 1, 0, 1, vexNDS3},
+	"KXORB":    {1, 0x47, 0, 1, 1, vexNDS3},
+	"KXORW":    {1, 0x47, 0, 0, 1, vexNDS3},
+	"KXORD":    {1, 0x47, 1, 1, 1, vexNDS3},
+	"KXORQ":    {1, 0x47, 1, 0, 1, vexNDS3},
 	"KUNPCKBW": {1, 0x4B, 0, 1, 1, vexNDS3},
 	"KUNPCKDQ": {1, 0x4B, 1, 0, 1, vexNDS3},
 	"KADDB":    {1, 0x4A, 0, 1, 1, vexNDS3},
 	"KADDW":    {1, 0x4A, 0, 0, 1, vexNDS3},
+	"KADDD":    {1, 0x4A, 1, 1, 1, vexNDS3},
 	"KADDQ":    {1, 0x4A, 1, 0, 1, vexNDS3},
-	// k ← OP k (KNOT) and flags ← k OP k (KORTEST): reg = dst, rm = src.
+	// k ← OP k (KNOT), k ← k AND~ k (KTEST-style RM) and flags ← k OP k
+	// (KORTEST): reg = dst, rm = src.
 	"KNOTB":    {1, 0x44, 0, 1, 0, vexRM},
+	"KNOTW":    {1, 0x44, 0, 0, 0, vexRM},
+	"KNOTD":    {1, 0x44, 1, 1, 0, vexRM},
+	"KNOTQ":    {1, 0x44, 1, 0, 0, vexRM},
+	"KORTESTB": {1, 0x98, 0, 1, 0, vexRM},
+	"KORTESTW": {1, 0x98, 0, 0, 0, vexRM},
 	"KORTESTD": {1, 0x98, 1, 1, 0, vexRM},
-	// OP $imm, src, dst: reg = dst, rm = src, imm8.
+	"KORTESTQ": {1, 0x98, 1, 0, 0, vexRM},
+	"KTESTB":   {1, 0x99, 0, 1, 0, vexRM},
+	"KTESTW":   {1, 0x99, 0, 0, 0, vexRM},
+	"KTESTD":   {1, 0x99, 1, 1, 0, vexRM},
+	"KTESTQ":   {1, 0x99, 1, 0, 0, vexRM},
+	// OP $imm, src, dst: reg = dst, rm = src, imm8.  The opcodes split by
+	// direction (0x32/0x33 left, 0x30/0x31 right) and within each by
+	// element half (0x32 byte/word, 0x33 dword/qword); W picks byte/dword
+	// (W0) against word/qword (W1).
+	"KSHIFTLB": {3, 0x32, 0, 1, 0, vexImmRM},
 	"KSHIFTLW": {3, 0x32, 1, 1, 0, vexImmRM},
+	"KSHIFTLD": {3, 0x33, 0, 1, 0, vexImmRM},
+	"KSHIFTLQ": {3, 0x33, 1, 1, 0, vexImmRM},
+	"KSHIFTRB": {3, 0x30, 0, 1, 0, vexImmRM},
+	"KSHIFTRW": {3, 0x30, 1, 1, 0, vexImmRM},
+	"KSHIFTRD": {3, 0x31, 0, 1, 0, vexImmRM},
+	"KSHIFTRQ": {3, 0x31, 1, 1, 0, vexImmRM},
 }
 
 // isKOp reports whether the mnemonic is an opmask-register instruction.

asm/evex_test.go

@@ -319,6 +319,43 @@ func TestEvexExtendedGroundTruth(t *testing.T) {
 		{"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"},
+		// Completed opmask families (ANDN, NOT, OR/XOR word+qword, TEST,
+		// word-width shifts; byte-exact against go tool asm).
+		{"KANDNW", "KANDNW", []Operand{vreg(t, "K1"), vreg(t, "K2"), vreg(t, "K3")}, "c5ec42d9"},
+		{"KANDNB", "KANDNB", []Operand{vreg(t, "K4"), vreg(t, "K5"), vreg(t, "K6")}, "c5d542f4"},
+		{"KANDND", "KANDND", []Operand{vreg(t, "K1"), vreg(t, "K2"), vreg(t, "K3")}, "c4e1ed42d9"},
+		{"KANDNQ", "KANDNQ", []Operand{vreg(t, "K4"), vreg(t, "K5"), vreg(t, "K6")}, "c4e1d442f4"},
+		{"KANDD", "KANDD", []Operand{vreg(t, "K1"), vreg(t, "K2"), vreg(t, "K3")}, "c4e1ed41d9"},
+		{"KADDD", "KADDD", []Operand{vreg(t, "K4"), vreg(t, "K5"), vreg(t, "K6")}, "c4e1d54af4"},
+		{"KNOTW", "KNOTW", []Operand{vreg(t, "K1"), vreg(t, "K2")}, "c5f844d1"},
+		{"KNOTD", "KNOTD", []Operand{vreg(t, "K3"), vreg(t, "K4")}, "c4e1f944e3"},
+		{"KNOTQ", "KNOTQ", []Operand{vreg(t, "K5"), vreg(t, "K6")}, "c4e1f844f5"},
+		{"KORW", "KORW", []Operand{vreg(t, "K1"), vreg(t, "K2"), vreg(t, "K3")}, "c5ec45d9"},
+		{"KORQ", "KORQ", []Operand{vreg(t, "K4"), vreg(t, "K5"), vreg(t, "K6")}, "c4e1d445f4"},
+		{"KXNORB", "KXNORB", []Operand{vreg(t, "K1"), vreg(t, "K2"), vreg(t, "K3")}, "c5ed46d9"},
+		{"KXORW", "KXORW", []Operand{vreg(t, "K1"), vreg(t, "K2"), vreg(t, "K3")}, "c5ec47d9"},
+		{"KXORQ", "KXORQ", []Operand{vreg(t, "K4"), vreg(t, "K5"), vreg(t, "K6")}, "c4e1d447f4"},
+		{"KORTESTW", "KORTESTW", []Operand{vreg(t, "K1"), vreg(t, "K2")}, "c5f898d1"},
+		{"KORTESTB", "KORTESTB", []Operand{vreg(t, "K3"), vreg(t, "K4")}, "c5f998e3"},
+		{"KORTESTQ", "KORTESTQ", []Operand{vreg(t, "K5"), vreg(t, "K6")}, "c4e1f898f5"},
+		{"KTESTW", "KTESTW", []Operand{vreg(t, "K1"), vreg(t, "K2")}, "c5f899d1"},
+		{"KTESTD", "KTESTD", []Operand{vreg(t, "K3"), vreg(t, "K4")}, "c4e1f999e3"},
+		{"KSHIFTLB", "KSHIFTLB", []Operand{Imm(1), vreg(t, "K1"), vreg(t, "K2")}, "c4e37932d101"},
+		{"KSHIFTLD", "KSHIFTLD", []Operand{Imm(2), vreg(t, "K3"), vreg(t, "K4")}, "c4e37933e302"},
+		{"KSHIFTLQ", "KSHIFTLQ", []Operand{Imm(3), vreg(t, "K5"), vreg(t, "K6")}, "c4e3f933f503"},
+		{"KSHIFTRB", "KSHIFTRB", []Operand{Imm(4), vreg(t, "K1"), vreg(t, "K2")}, "c4e37930d104"},
+		{"KSHIFTRW", "KSHIFTRW", []Operand{Imm(5), vreg(t, "K3"), vreg(t, "K4")}, "c4e3f930e305"},
+		{"KSHIFTRQ", "KSHIFTRQ", []Operand{Imm(6), vreg(t, "K5"), vreg(t, "K6")}, "c4e3f931f506"},
+		// Integer compares with an opmask destination (0F3A map, the
+		// go-bzip2 partition kernel's classify instructions).
+		{"VPCMPUB", "VPCMPUB", []Operand{Imm(1), vreg(t, "X1"), vreg(t, "X0"), vreg(t, "K1")}, "62f37d083ec901"},
+		{"VPCMPB", "VPCMPB", []Operand{Imm(2), vreg(t, "Y2"), vreg(t, "Y3"), vreg(t, "K2")}, "62f365283fd202"},
+		{"VPCMPUW", "VPCMPUW", []Operand{Imm(5), vreg(t, "Z1"), vreg(t, "Z2"), vreg(t, "K3")}, "62f3ed483ed905"},
+		{"VPCMPW", "VPCMPW", []Operand{Imm(6), vreg(t, "X3"), vreg(t, "X4"), vreg(t, "K4")}, "62f3dd083fe306"},
+		{"VPCMPD", "VPCMPD", []Operand{Imm(0), vreg(t, "Y1"), vreg(t, "Y2"), vreg(t, "K1")}, "62f36d281fc900"},
+		{"VPCMPUD", "VPCMPUD", []Operand{Imm(1), vreg(t, "Z2"), vreg(t, "Z3"), vreg(t, "K2")}, "62f365481ed201"},
+		{"VPCMPQ", "VPCMPQ", []Operand{Imm(2), vreg(t, "X1"), vreg(t, "X2"), vreg(t, "K3")}, "62f3ed081fd902"},
+		{"VPCMPUQ", "VPCMPUQ", []Operand{Imm(3), vreg(t, "Y3"), vreg(t, "Y4"), vreg(t, "K4")}, "62f3dd281ee303"},
 		// 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"},

asm/instrs.go

@@ -48,15 +48,21 @@ func (e *enc) encodeMov(ops []Operand, size int) error {
 	}
 	src, dst := ops[0], ops[1]
 
-	// MOVQ with an XMM operand is the SSE2 packed-quadword move, NOT a
-	// GPR move: mem→xmm and xmm↔xmm encode as F3 0F 7E (reg = dst),
-	// xmm→mem as 66 0F D6 (rm = xmm).  A GPR-move fallback would
-	// silently emit REX.W 8B with the wrong operand meaning.
+	// Integer scalar XMM moves: MOVQ with an XMM operand is the SSE2
+	// packed-quadword move, NOT a GPR move: mem→xmm encodes as F3 0F 7E
+	// (reg = dst, no REX.W — the Go assembler's form), xmm→mem as
+	// 66 0F D6 (rm = xmm).  MOVL is the packed-dword move instead:
+	// 66 0F 6E load, 66 0F 7E store.  A GPR-move fallback would silently
+	// emit REX.W 8B with the wrong operand meaning.
 	_, srcVec := vecReg(src)
 	dstReg, dstVec := vecReg(dst)
 	if srcVec || dstVec {
 		if dstVec {
 			i := &instr{prefix: 0xF3, opcode: []byte{0x0F, 0x7E}, modrm: -1, sib: -1}
+			if size == 4 {
+				i.prefix = 0x66
+				i.opcode = []byte{0x0F, 0x6E}
+			}
 			if err := setRM(i, dstReg, src, 8); err != nil {
 				return err
 			}
@@ -64,9 +70,12 @@ func (e *enc) encodeMov(ops []Operand, size int) error {
 		}
 		srcXMM, srcIsXMM := src.(Reg)
 		if !srcIsXMM || !srcXMM.isVec() {
-			return fmt.Errorf("MOVQ: store needs an XMM source")
+			return fmt.Errorf("MOV: store needs an XMM source")
 		}
 		i := &instr{prefix: 0x66, opcode: []byte{0x0F, 0xD6}, modrm: -1, sib: -1}
+		if size == 4 {
+			i.opcode = []byte{0x0F, 0x7E}
+		}
 		if err := setRM(i, srcXMM, dst, 8); err != nil {
 			return err
 		}
@@ -682,6 +691,21 @@ func (e *enc) encodeCount(base string, ops []Operand, size int) error {
 	return e.emit(i)
 }
 
+// encodeBswap encodes BSWAP: the single register operand is encoded in the
+// opcode byte (0F C8+r), with REX.B for R8-R15 and REX.W for the quad form.
+func (e *enc) encodeBswap(ops []Operand, size int) error {
+	if len(ops) != 1 {
+		return fmt.Errorf("BSWAP expects 1 operand, got %d", len(ops))
+	}
+	reg, ok := ops[0].(Reg)
+	if !ok {
+		return fmt.Errorf("BSWAP operand must be a register")
+	}
+	i := newInstr(size, []byte{0x0F, 0xC8 + byte(reg.idx&7)})
+	i.rexB = reg.idx >= 8
+	return e.emit(i)
+}
+
 // --- mixed-width sign/zero-extending moves -----------------------------------
 
 // movExtendOp maps Go's mixed-width move names to their opcode and destination
@@ -784,33 +808,49 @@ func (e *enc) encodeSSEMove(m sseMove, ops []Operand) error {
 // --- legacy SSE packed binary and shuffles -----------------------------------
 
 // sseBin describes a legacy (non-VEX) SSE packed/scalar binary op: an
-// optional mandatory prefix plus the 0F-prefixed opcode.  Plan 9 asm
-// lists the source operand first, so MULPS X0, X1 computes X1 = X1 * X0.
+// optional mandatory prefix plus the 0F-prefixed opcode (0F38 for the
+// SSSE3 integer shuffles).  Plan 9 asm lists the source operand first, so
+// MULPS X0, X1 computes X1 = X1 * X0.
 type sseBin struct {
 	prefix byte // 0, 0x66, 0xF2 or 0xF3
 	op     byte
+	map38  bool // opcode lives under 0F38 instead of 0F
 }
 
 var sseBinTable = map[string]sseBin{
-	"ADDPS": {0, 0x58}, "ADDPD": {0x66, 0x58},
-	"MULPS": {0, 0x59}, "MULPD": {0x66, 0x59},
-	"SUBPS": {0, 0x5C}, "SUBPD": {0x66, 0x5C},
-	"DIVPS": {0, 0x5E}, "DIVPD": {0x66, 0x5E},
-	"ANDPS": {0, 0x54}, "ANDPD": {0x66, 0x54},
-	"ORPS": {0, 0x56}, "ORPD": {0x66, 0x56},
-	"XORPS": {0, 0x57}, "XORPD": {0x66, 0x57},
-	"MINPS": {0, 0x5D}, "MINPD": {0x66, 0x5D},
-	"MAXPS": {0, 0x5F}, "MAXPD": {0x66, 0x5F},
-	"ADDSS": {0xF3, 0x58}, "ADDSD": {0xF2, 0x58},
-	"MULSS": {0xF3, 0x59}, "MULSD": {0xF2, 0x59},
-	"SUBSS": {0xF3, 0x5C}, "SUBSD": {0xF2, 0x5C},
-	"DIVSS": {0xF3, 0x5E}, "DIVSD": {0xF2, 0x5E},
-	"MINSS": {0xF3, 0x5D}, "MINSD": {0xF2, 0x5D},
-	"MAXSS": {0xF3, 0x5F}, "MAXSD": {0xF2, 0x5F},
-	"UNPCKLPS": {0, 0x14}, "UNPCKHPS": {0, 0x15},
-	"UNPCKLPD": {0x66, 0x14}, "UNPCKHPD": {0x66, 0x15},
-	"CVTSS2SD": {0xF3, 0x5A}, "CVTSD2SS": {0xF2, 0x5A},
-	"CVTPS2PD": {0, 0x5A}, "CVTPD2PS": {0x66, 0x5A},
+	"ADDPS": {0, 0x58, false}, "ADDPD": {0x66, 0x58, false},
+	"MULPS": {0, 0x59, false}, "MULPD": {0x66, 0x59, false},
+	"SUBPS": {0, 0x5C, false}, "SUBPD": {0x66, 0x5C, false},
+	"DIVPS": {0, 0x5E, false}, "DIVPD": {0x66, 0x5E, false},
+	"ANDPS": {0, 0x54, false}, "ANDPD": {0x66, 0x54, false},
+	"ORPS": {0, 0x56, false}, "ORPD": {0x66, 0x56, false},
+	"XORPS": {0, 0x57, false}, "XORPD": {0x66, 0x57, false},
+	"MINPS": {0, 0x5D, false}, "MINPD": {0x66, 0x5D, false},
+	"MAXPS": {0, 0x5F, false}, "MAXPD": {0x66, 0x5F, false},
+	"ADDSS": {0xF3, 0x58, false}, "ADDSD": {0xF2, 0x58, false},
+	"MULSS": {0xF3, 0x59, false}, "MULSD": {0xF2, 0x59, false},
+	"SUBSS": {0xF3, 0x5C, false}, "SUBSD": {0xF2, 0x5C, false},
+	"DIVSS": {0xF3, 0x5E, false}, "DIVSD": {0xF2, 0x5E, false},
+	"MINSS": {0xF3, 0x5D, false}, "MINSD": {0xF2, 0x5D, false},
+	"MAXSS": {0xF3, 0x5F, false}, "MAXSD": {0xF2, 0x5F, false},
+	"UNPCKLPS": {0, 0x14, false}, "UNPCKHPS": {0, 0x15, false},
+	"UNPCKLPD": {0x66, 0x14, false}, "UNPCKHPD": {0x66, 0x15, false},
+	"CVTSS2SD": {0xF3, 0x5A, false}, "CVTSD2SS": {0xF2, 0x5A, false},
+	"CVTPS2PD": {0, 0x5A, false}, "CVTPD2PS": {0x66, 0x5A, false},
+	// SSE2 packed integers (reg = reg op rm) and the SSSE3 byte shuffle.
+	"PXOR":  {0x66, 0xEF, false},
+	"POR":   {0x66, 0xEB, false},
+	"PAND":  {0x66, 0xDB, false},
+	"PANDN": {0x66, 0xDF, false},
+	"PADDB": {0x66, 0xFC, false}, "PADDW": {0x66, 0xFD, false},
+	"PADDD": {0x66, 0xFE, false}, "PADDQ": {0x66, 0xD4, false},
+	"PSUBB": {0x66, 0xF8, false}, "PSUBW": {0x66, 0xF9, false},
+	"PSUBD": {0x66, 0xFA, false}, "PSUBQ": {0x66, 0xFB, false},
+	"PCMPEQB": {0x66, 0x74, false}, "PCMPEQW": {0x66, 0x75, false},
+	"PCMPEQD": {0x66, 0x76, false},
+	"PCMPGTB": {0x66, 0x64, false}, "PCMPGTW": {0x66, 0x65, false},
+	"PCMPGTD": {0x66, 0x66, false},
+	"PSHUFB":  {0x66, 0x00, true},
 }
 
 // sseShuf describes a legacy SSE shuffle taking a trailing imm8
@@ -835,7 +875,11 @@ func (e *enc) encodeSSEBin(m sseBin, ops []Operand) error {
 	if !ok || !dstReg.isVec() {
 		return fmt.Errorf("SSE binary destination must be a vector register")
 	}
-	i := &instr{prefix: m.prefix, opcode: []byte{0x0F, m.op}, modrm: -1, sib: -1}
+	opcode := []byte{0x0F, m.op}
+	if m.map38 {
+		opcode = []byte{0x0F, 0x38, m.op}
+	}
+	i := &instr{prefix: m.prefix, opcode: opcode, modrm: -1, sib: -1}
 	if err := setRM(i, dstReg, src, 8); err != nil {
 		return err
 	}

asm/vex.go

@@ -141,6 +141,8 @@ var vexTable = map[string]vexSpec{
 	"VPMOVZXWQ":    {2, 0x34, 0, 1, -1, vexRM},
 	"VPBROADCASTD": {2, 0x58, 0, 1, -1, vexRM},
 	"VPBROADCASTQ": {2, 0x59, 0, 1, -1, vexRM},
+	"VPBROADCASTB": {2, 0x78, 0, 1, -1, vexRM},
+	"VPBROADCASTW": {2, 0x79, 0, 1, -1, vexRM},
 	// VEX.128/256.F3.0F.WIG — signed dword to packed double conversion
 	// (reg=dst, rm=src, no vvvv; the length follows the destination).
 	"VCVTDQ2PD": {1, 0xE6, 0, 2, -1, vexRM},