feat(asm): add the EVEX FP helper tail and gather/scatter with VSIB

Assisted-by: Qwen 3.8 Max Preview

asm/encode.go

@@ -57,7 +57,7 @@ func (e *enc) encode(mnem string, ops []Operand) error {
 	if err != nil {
 		return err
 	}
-	if isVex(base) || isEvex(base) || isKOp(base) || base == "KMOVW" || base == "KMOVQ" {
+	if isVex(base) || isEvex(base) || isKOp(base) || isGather(base) || isScatter(base) || base == "KMOVW" || base == "KMOVQ" {
 		return e.encodeVec(base, ops, sfx)
 	}
 	if sfx.any() {
@@ -130,6 +130,12 @@ func splitSize(upper string) (base string, size int) {
 // 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, sfx evexSuffix) error {
+	if gs, ok := gatherTable[upper]; ok {
+		return e.encodeGather(upper, gs, ops, sfx)
+	}
+	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)

asm/evex.go

@@ -234,6 +234,80 @@ var evexTable = map[string]evexSpec{
 	// EVEX W1 qword shifts.
 	"VPSRLQ": {1, 0x73, 1, 1, 2, vexShiftImm, [3]int{16, 32, 64}},
 	"VPSLLQ": {1, 0x73, 1, 1, 6, vexShiftImm, [3]int{16, 32, 64}},
+
+	// EVEX.66.0F38 — floating-point helpers, packed (reg=dst, rm=src).
+	"VRCP14PD":   {2, 0x4C, 1, 1, -1, vexRM, [3]int{16, 32, 64}},
+	"VRCP14PS":   {2, 0x4C, 0, 1, -1, vexRM, [3]int{16, 32, 64}},
+	"VRSQRT14PD": {2, 0x4E, 1, 1, -1, vexRM, [3]int{16, 32, 64}},
+	"VRSQRT14PS": {2, 0x4E, 0, 1, -1, vexRM, [3]int{16, 32, 64}},
+	"VGETEXPPD":  {2, 0x42, 1, 1, -1, vexRM, [3]int{16, 32, 64}},
+	"VGETEXPPS":  {2, 0x42, 0, 1, -1, vexRM, [3]int{16, 32, 64}},
+	// EVEX.66.0F38 — floating-point helpers, scalar (NDS form: src2 is
+	// rm, src1 is vvvv, the XMM destination is reg).  Like the scalar 0F3A
+	// forms, these take the 66 prefix; W selects double/single.
+	"VRCP14SD":   {2, 0x4D, 1, 1, -1, vexNDS3, [3]int{8, 8, 8}},
+	"VRCP14SS":   {2, 0x4D, 0, 1, -1, vexNDS3, [3]int{4, 4, 4}},
+	"VRSQRT14SD": {2, 0x4F, 1, 1, -1, vexNDS3, [3]int{8, 8, 8}},
+	"VRSQRT14SS": {2, 0x4F, 0, 1, -1, vexNDS3, [3]int{4, 4, 4}},
+	"VGETEXPSD":  {2, 0x43, 1, 1, -1, vexNDS3, [3]int{8, 8, 8}},
+	"VGETEXPSS":  {2, 0x43, 0, 1, -1, vexNDS3, [3]int{4, 4, 4}},
+	// EVEX.66.0F38 — scale by a power of two (NDS form).
+	"VSCALEFPD": {2, 0x2C, 1, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VSCALEFPS": {2, 0x2C, 0, 1, -1, vexNDS3, [3]int{16, 32, 64}},
+	"VSCALEFSD": {2, 0x2D, 1, 1, -1, vexNDS3, [3]int{8, 8, 8}},
+	"VSCALEFSS": {2, 0x2D, 0, 1, -1, vexNDS3, [3]int{4, 4, 4}},
+
+	// EVEX.66.0F3A — packed round/getmant/reduce ($imm, src, dst: reg=dst,
+	// rm=src, imm8).
+	"VRNDSCALEPD": {3, 0x09, 1, 1, -1, vexImmRM, [3]int{16, 32, 64}},
+	"VRNDSCALEPS": {3, 0x08, 0, 1, -1, vexImmRM, [3]int{16, 32, 64}},
+	"VGETMANTPD":  {3, 0x26, 1, 1, -1, vexImmRM, [3]int{16, 32, 64}},
+	"VGETMANTPS":  {3, 0x26, 0, 1, -1, vexImmRM, [3]int{16, 32, 64}},
+	"VREDUCEPD":   {3, 0x56, 1, 1, -1, vexImmRM, [3]int{16, 32, 64}},
+	"VREDUCEPS":   {3, 0x56, 0, 1, -1, vexImmRM, [3]int{16, 32, 64}},
+	// EVEX.66.0F3A — scalar round/getmant/reduce and fixup/range (NDS +
+	// imm8: $imm, src2, src1, dst).  The scalar 0F3A forms all take the 66
+	// prefix; W selects double/single.
+	"VRNDSCALESD": {3, 0x0B, 1, 1, -1, vexNDS3Imm, [3]int{8, 8, 8}},
+	"VRNDSCALESS": {3, 0x0A, 0, 1, -1, vexNDS3Imm, [3]int{4, 4, 4}},
+	"VGETMANTSD":  {3, 0x27, 1, 1, -1, vexNDS3Imm, [3]int{8, 8, 8}},
+	"VGETMANTSS":  {3, 0x27, 0, 1, -1, vexNDS3Imm, [3]int{4, 4, 4}},
+	"VREDUCESD":   {3, 0x57, 1, 1, -1, vexNDS3Imm, [3]int{8, 8, 8}},
+	"VREDUCESS":   {3, 0x57, 0, 1, -1, vexNDS3Imm, [3]int{4, 4, 4}},
+	"VFIXUPIMMPD": {3, 0x54, 1, 1, -1, vexNDS3Imm, [3]int{16, 32, 64}},
+	"VFIXUPIMMPS": {3, 0x54, 0, 1, -1, vexNDS3Imm, [3]int{16, 32, 64}},
+	"VFIXUPIMMSD": {3, 0x55, 1, 1, -1, vexNDS3Imm, [3]int{8, 8, 8}},
+	"VFIXUPIMMSS": {3, 0x55, 0, 1, -1, vexNDS3Imm, [3]int{4, 4, 4}},
+	"VRANGEPD":    {3, 0x50, 1, 1, -1, vexNDS3Imm, [3]int{16, 32, 64}},
+	"VRANGEPS":    {3, 0x50, 0, 1, -1, vexNDS3Imm, [3]int{16, 32, 64}},
+	"VRANGESD":    {3, 0x51, 1, 1, -1, vexNDS3Imm, [3]int{8, 8, 8}},
+	"VRANGESS":    {3, 0x51, 0, 1, -1, vexNDS3Imm, [3]int{4, 4, 4}},
+
+	// EVEX.66.0F3A — floating-point class test ($imm, src, kdst): the
+	// reg field carries the opmask destination.  The packed forms carry an
+	// explicit length in the mnemonic (X/Y/Z).
+	"VFPCLASSPDX": {3, 0x66, 1, 1, -1, vexImmRM, [3]int{16, 0, 0}},
+	"VFPCLASSPDY": {3, 0x66, 1, 1, -1, vexImmRM, [3]int{0, 32, 0}},
+	"VFPCLASSPDZ": {3, 0x66, 1, 1, -1, vexImmRM, [3]int{0, 0, 64}},
+	"VFPCLASSPSX": {3, 0x66, 0, 1, -1, vexImmRM, [3]int{16, 0, 0}},
+	"VFPCLASSPSY": {3, 0x66, 0, 1, -1, vexImmRM, [3]int{0, 32, 0}},
+	"VFPCLASSPSZ": {3, 0x66, 0, 1, -1, vexImmRM, [3]int{0, 0, 64}},
+	"VFPCLASSSD":  {3, 0x67, 1, 1, -1, vexImmRM, [3]int{8, 0, 0}},
+	"VFPCLASSSS":  {3, 0x67, 0, 1, -1, vexImmRM, [3]int{4, 0, 0}},
+
+	// EVEX — the remaining conversions.  VCVTQQ2PS narrows (the 512-bit
+	// source sets the length); the rest follow the destination.
+	"VCVTQQ2PS":  {1, 0x5B, 1, 0, -1, vexRMSrcLen, [3]int{0, 0, 64}},
+	"VCVTPD2QQ":  {1, 0x7B, 1, 1, -1, vexRM, [3]int{16, 32, 64}},
+	"VCVTPD2UQQ": {1, 0x79, 1, 1, -1, vexRM, [3]int{16, 32, 64}},
+	"VCVTPS2QQ":  {1, 0x7B, 0, 1, -1, vexRM, [3]int{8, 16, 32}},
+	"VCVTUDQ2PD": {1, 0x7A, 0, 2, -1, vexRM, [3]int{8, 16, 32}},
+	"VCVTUDQ2PS": {1, 0x7A, 0, 0, -1, vexRM, [3]int{8, 16, 32}},
+	// EVEX.66.0F38 — half-precision convert (half-width source).
+	"VCVTPH2PS": {2, 0x13, 0, 1, -1, vexRM, [3]int{8, 16, 32}},
+	// 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.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).
@@ -466,6 +540,9 @@ var evexRound = map[string]bool{
 	"VMINSD": true, "VMAXSD": true,
 	"VADDSS": true, "VSUBSS": true, "VMULSS": true, "VDIVSS": true,
 	"VMINSS": true, "VMAXSS": true,
+	"VSCALEFPD": true, "VSCALEFPS": true, "VSCALEFSD": true, "VSCALEFSS": true,
+	"VGETEXPPD": true, "VGETEXPPS": true, "VGETEXPSD": true, "VGETEXPSS": true,
+	"VCVTDQ2PS": true, "VCVTPS2QQ": true, "VCVTQQ2PS": true, "VCVTPD2UQQ": true,
 }
 
 // evexBcstN maps an instruction accepting .BCST to the broadcast element
@@ -475,6 +552,16 @@ var evexBcstN = map[string]int{
 	"VMINPD": 8, "VMAXPD": 8,
 	"VADDPS": 4, "VSUBPS": 4, "VMULPS": 4, "VDIVPS": 4,
 	"VMINPS": 4, "VMAXPS": 4,
+	"VRCP14PD": 8, "VRCP14PS": 4, "VRSQRT14PD": 8, "VRSQRT14PS": 4,
+	"VGETEXPPD": 8, "VGETEXPPS": 4,
+	"VSCALEFPD": 8, "VSCALEFPS": 4,
+	"VRNDSCALEPD": 8, "VRNDSCALEPS": 4,
+	"VGETMANTPD": 8, "VGETMANTPS": 4,
+	"VREDUCEPD": 8, "VREDUCEPS": 4,
+	"VFIXUPIMMPD": 8, "VFIXUPIMMPS": 4,
+	"VRANGEPD": 8, "VRANGEPS": 4,
+	"VCVTDQ2PS": 4, "VCVTPS2QQ": 4, "VCVTQQ2PS": 8,
+	"VCVTUDQ2PD": 4, "VCVTUDQ2PS": 4,
 }
 
 // splitMask extracts an explicit mask register (K1–K7) from the operand list,
@@ -547,6 +634,10 @@ func (e *enc) encodeEvex(mnemUpper string, ops []Operand, sfx evexSuffix) error
 			if dst, ok := ops[len(ops)-1].(Reg); ok && dst.mask {
 				return kdst(e.encodeEvexNDS3Imm)
 			}
+		case vexImmRM:
+			if dst, ok := ops[len(ops)-1].(Reg); ok && dst.mask {
+				return kdst(e.encodeEvexImmRM)
+			}
 		}
 	}
 
@@ -636,7 +727,9 @@ func (e *enc) encodeEvexRM(spec evexSpec, ops []Operand, mask int, sfx evexSuffi
 }
 
 // encodeEvexImmRM encodes the immediate shuffle form: OP $imm, src, dst
-// (reg = dst, rm = src, imm8), e.g. VPSHUFD.
+// (reg = dst, rm = src, imm8), e.g. VPSHUFD.  The destination may be an
+// opmask register (VFPCLASS*), in which case the vector length comes from
+// the source.
 func (e *enc) encodeEvexImmRM(spec evexSpec, ops []Operand, mask int, sfx evexSuffix) error {
 	if len(ops) != 3 {
 		return fmt.Errorf("shuffle expects 3 operands ($imm, src, dst), got %d", len(ops))
@@ -647,13 +740,17 @@ func (e *enc) encodeEvexImmRM(spec evexSpec, ops []Operand, mask int, sfx evexSu
 		return fmt.Errorf("shuffle control must be an immediate")
 	}
 	dstReg, ok := dst.(Reg)
-	if !ok || !dstReg.isVec() {
+	if !ok || (!dstReg.isVec() && !dstReg.mask) {
-		return fmt.Errorf("shuffle destination must be a vector register")
+		return fmt.Errorf("shuffle destination must be a vector or mask register")
 	}
 	ll := dstReg.vecLenBit()
+	if dstReg.mask {
 		if r, ok := src.(Reg); ok && r.isVec() {
 			ll = r.vecLenBit()
 		}
+	} else if r, ok := src.(Reg); ok && r.isVec() {
+		ll = r.vecLenBit()
+	}
 	immByte, err := imm8(int64(immVal))
 	if err != nil {
 		return err
@@ -1034,6 +1131,135 @@ func memComponentsEvex(regField int, m Mem, n int) (modrm, sib int, disp []byte,
 	return mod<<6 | regField<<3 | (m.Base.idx & 7), -1, disp, 1, bBar, nil
 }
 
+// gatherSpec describes a gather/scatter family member: all live in
+// 66.0F38; the opcode and W select the index and data element widths, and n
+// is the data element size (the EVEX disp8×N multiplier).
+type gatherSpec struct {
+	opcode byte
+	w      int
+	n      int
+}
+
+var gatherTable = map[string]gatherSpec{
+	"VGATHERDPS": {0x92, 0, 4},
+	"VGATHERDPD": {0x92, 1, 8},
+	"VGATHERQPS": {0x93, 0, 4},
+	"VGATHERQPD": {0x93, 1, 8},
+	"VPGATHERDD": {0x90, 0, 4},
+	"VPGATHERDQ": {0x90, 1, 8},
+	"VPGATHERQD": {0x91, 0, 4},
+	"VPGATHERQQ": {0x91, 1, 8},
+}
+
+var scatterTable = map[string]gatherSpec{
+	"VSCATTERDPS": {0xA2, 0, 4},
+	"VSCATTERDPD": {0xA2, 1, 8},
+	"VSCATTERQPS": {0xA3, 0, 4},
+	"VSCATTERQPD": {0xA3, 1, 8},
+	"VPSCATTERDD": {0xA0, 0, 4},
+	"VPSCATTERDQ": {0xA0, 1, 8},
+	"VPSCATTERQD": {0xA1, 0, 4},
+	"VPSCATTERQQ": {0xA1, 1, 8},
+}
+
+// isGather reports whether the mnemonic is a gather instruction.
+func isGather(upper string) bool {
+	_, ok := gatherTable[upper]
+	return ok
+}
+
+// isScatter reports whether the mnemonic is a scatter instruction.
+func isScatter(upper string) bool {
+	_, ok := scatterTable[upper]
+	return ok
+}
+
+// vsibLen validates a VSIB memory operand (the index must be a vector
+// register) and returns it with the vector length the index selects — the
+// EVEX L'L field follows the index register, not the data register.
+func vsibLen(op Operand, what string) (Mem, int, error) {
+	m, ok := op.(Mem)
+	if !ok || !m.HasIndex || !m.Index.isVec() {
+		return Mem{}, 0, fmt.Errorf("%s: operand must be a VSIB memory reference with a vector index", what)
+	}
+	return m, m.Index.vecLenBit(), nil
+}
+
+// encodeGather encodes a gather.  The VEX spelling carries the mask in a
+// vector register (OP mask, vsib, dst: vvvv = mask, rm = vsib, reg = dst,
+// L follows the data register); the EVEX spelling carries it in aaa (OP
+// vsib, K, dst: rm = vsib, reg = dst, L follows the VSIB index).
+func (e *enc) encodeGather(upper string, gs gatherSpec, ops []Operand, sfx evexSuffix) error {
+	rest, mask, err := splitMask(ops)
+	if err != nil {
+		return err
+	}
+	if mask != 0 || sfx.any() {
+		// EVEX form: OP vsib, K, dst.
+		if len(rest) != 2 {
+			return fmt.Errorf("%s expects 3 operands (vsib, K, dst), got %d", upper, len(ops))
+		}
+		vsib, ll, err := vsibLen(rest[0], upper)
+		if err != nil {
+			return err
+		}
+		dst, ok := rest[1].(Reg)
+		if !ok || !dst.isVec() {
+			return fmt.Errorf("%s: destination must be a vector register", upper)
+		}
+		evex := evexSpec{mapSel: 2, opcode: gs.opcode, w: gs.w, pp: 1, opdigit: -1, n: [3]int{gs.n, gs.n, gs.n}}
+		return e.emitEvexFields(evex, ll, dst.idx, -1, vsib, mask, sfx)
+	}
+	// VEX form: OP mask, vsib, dst.
+	if len(rest) != 3 {
+		return fmt.Errorf("%s expects 3 operands (mask, vsib, dst), got %d", upper, len(rest))
+	}
+	maskReg, ok := rest[0].(Reg)
+	if !ok || !maskReg.isVec() {
+		return fmt.Errorf("%s: mask must be a vector register", upper)
+	}
+	vsib, _, err := vsibLen(rest[1], upper)
+	if err != nil {
+		return err
+	}
+	dst, ok := rest[2].(Reg)
+	if !ok || !dst.isVec() {
+		return fmt.Errorf("%s: destination must be a vector register", upper)
+	}
+	spec := vexSpec{mapSel: 2, opcode: gs.opcode, w: gs.w, pp: 1, opdigit: -1}
+	rBit := 0
+	if dst.idx >= 8 {
+		rBit = 1
+	}
+	return e.emitVexFields(spec, dst.vecLenBit(), dst.idx&7, rBit, 15-maskReg.idx, vsib)
+}
+
+// encodeScatter encodes a scatter (EVEX only): OP src, K, vsib — reg = src,
+// rm = the VSIB memory operand, the K mask in aaa and L following the VSIB
+// index.
+func (e *enc) encodeScatter(upper string, ss gatherSpec, ops []Operand, sfx evexSuffix) error {
+	rest, mask, err := splitMask(ops)
+	if err != nil {
+		return err
+	}
+	if mask == 0 {
+		return fmt.Errorf("%s requires a K mask register", upper)
+	}
+	if len(rest) != 2 {
+		return fmt.Errorf("%s expects 3 operands (src, K, vsib), got %d", upper, len(ops))
+	}
+	src, ok := rest[0].(Reg)
+	if !ok || !src.isVec() {
+		return fmt.Errorf("%s: source must be a vector register", upper)
+	}
+	vsib, ll, err := vsibLen(rest[1], upper)
+	if err != nil {
+		return err
+	}
+	evex := evexSpec{mapSel: 2, opcode: ss.opcode, w: ss.w, pp: 1, opdigit: -1, n: [3]int{ss.n, ss.n, ss.n}}
+	return e.emitEvexFields(evex, ll, src.idx, -1, vsib, mask, sfx)
+}
+
 // 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

@@ -364,6 +364,109 @@ func TestEvexExtendedGroundTruth(t *testing.T) {
 	}
 }
 
+// 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)
+		}
+	}
+}
+
 // TestEvexErrors checks the EVEX-specific error paths.
 func TestEvexErrors(t *testing.T) {
 	cases := []struct {

asm/vex.go

@@ -178,6 +178,9 @@ 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},
@@ -189,6 +192,9 @@ var vexTable = map[string]vexSpec{
 	// 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},

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 = "0.13.0"
+var version = "0.14.0"
 
 func main() {
 	if len(os.Args) < 2 {

docs/ARCHITECTURE.md

@@ -237,7 +237,14 @@ 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 EVEX mnemonic
+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
+half-precision pair), 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

justfile

@@ -3,7 +3,7 @@
 
 # gasm-devkit — developer tooling for Go's Plan 9 assembler (GAsm).
 
-version := "0.13.0"
+version := "0.14.0"
 
 default:
     @just --list