gasm-sdk/arch/amd64_ext.go

// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
// SPDX-License-Identifier: BSD-3-Clause

// This file carries the amd64 side of the extended-instruction layer:
// instructions the Go toolchain does not know at all, described as data and
// validated against golden vectors from the Intel SDM rather than against the
// toolchain.  It sits beside the generated table, never inside it:
// arch/amd64_gen.go stays untouched, and asm.Encodable keeps answering false
// for every mnemonic here, so the layer stays out of the main encoders.
//
// The families are AVX512-BF16, AVX512-VP2INTERSECT and AVX512-FP16, the
// latter's scalar core with its imm8-control group and its packed 512-bit
// arithmetic, in their EVEX register forms.  The encodings are transcribed
// from the SDM instruction entries and cross-checked against binutils-gdb's
// assembler testsuite; the golden vectors in amd64_ext_test.go pin the
// bytes.  VPOPCNTD and VPOPCNTQ, the third family of the 2026-09-19 survey,
// no longer belong here: the Go toolchain's assembler knows them today, they
// live in the generated table and the EVEX encoder, and a mnemonic the
// toolchain has is not an extension.
//
// The forms encode the unmasked shapes: register forms throughout, and the
// scalar FP16 memory forms beside them, base-relative operands with the
// ModR/M disp8 and disp32 choices and the SIB byte RSP and R12 demand.  A
// scaled index, write masking ({k1}{z}) and embedded rounding still arrive
// with a later slice.

package arch

import "fmt"

// The features the amd64 layer covers.
const (
	ExtFeatureBF16         ExtFeature = "avx512bf16"
	ExtFeatureVP2INTERSECT ExtFeature = "avx512vp2intersect"
	ExtFeatureFP16         ExtFeature = "avx512fp16"
)

// ExtXmm, ExtYmm and ExtZmm build vector operands of the three EVEX register
// widths, VADDPS ZMM1, ZMM2, ZMM3 style.  The register number runs 0..31,
// XMM16 and above included: EVEX carries five register bits in every
// position, and the golden vectors exercise the high registers on purpose.
func ExtXmm(reg int) ExtOperand { return ExtOperand{Kind: ExtXMM, Reg: reg} }
func ExtYmm(reg int) ExtOperand { return ExtOperand{Kind: ExtYMM, Reg: reg} }
func ExtZmm(reg int) ExtOperand { return ExtOperand{Kind: ExtZMM, Reg: reg} }

// ExtMask builds an opmask operand, VP2INTERSECTD K1, ZMM2, ZMM3 style.  The
// register number runs 0..7.
func ExtMask(reg int) ExtOperand { return ExtOperand{Kind: ExtKReg, Reg: reg} }

// ExtGpr32 and ExtGpr64 build general-register operands, VCVTSI2SH XMM1,
// XMM2, EAX style.  The register number runs 0..15.
func ExtGpr32(reg int) ExtOperand { return ExtOperand{Kind: ExtR32, Reg: reg} }
func ExtGpr64(reg int) ExtOperand { return ExtOperand{Kind: ExtR64, Reg: reg} }

// amd64LengthClass reads the vector length the template encodes out of the
// L'L field of the EVEX byte three and names the register class every vector
// operand of that entry must carry.
func amd64LengthClass(b []byte) ExtOperandKind {
	switch (b[3] >> 5) & 3 {
	case 0:
		return ExtXMM
	case 1:
		return ExtYMM
	default:
		return ExtZMM
	}
}

// amd64HalfClass names the half-width companion of a vector class, the
// destination class of the narrow conversions.  At 128 bits the companion is
// the class itself, which is what the manual gives for the narrowest form.
func amd64HalfClass(k ExtOperandKind) ExtOperandKind {
	switch k {
	case ExtZMM:
		return ExtYMM
	case ExtYMM:
		return ExtXMM
	default:
		return ExtXMM
	}
}

// amd64Encode returns the template with the register-derived bits filled in:
// dest and rm are register numbers for the ModR/M reg and r/m fields, vvvv is
// the third-operand register or -1 when the form leaves it unused.  The EVEX
// plumbing follows the encoder in asm: reg[3] rides R bar and reg[4] R prime
// bar, rm[3] rides B bar, and in a register form rm[4] rides X bar, while
// vvvv[4] rides V prime bar in byte three.
func amd64Encode(b []byte, dest, vvvv, rm int) []byte {
	out := make([]byte, len(b))
	copy(out, b)
	rBar, rPrimeBar := 1, 1
	if dest&8 != 0 {
		rBar = 0
	}
	if dest&16 != 0 {
		rPrimeBar = 0
	}
	xBar, bBar := 1, 1
	if rm&8 != 0 {
		bBar = 0
	}
	if rm&16 != 0 {
		xBar = 0
	}
	out[1] |= byte(rBar<<7 | xBar<<6 | bBar<<5 | rPrimeBar<<4)
	vBar, vPrimeBar := 15, 1
	if vvvv >= 0 {
		vBar = 15 - (vvvv & 15)
		if vvvv&16 != 0 {
			vPrimeBar = 0
		}
	}
	out[2] |= byte(vBar << 3)
	out[3] |= byte(vPrimeBar << 3)
	out[5] |= byte((dest&7)<<3 | rm&7)
	return out
}

// amd64Memory validates a memory operand of an amd64 entry: no arrangement
// and no qualifier, a base general register inside 0-15, a signed 32-bit
// displacement and no shift.  The base number rides the operand's Reg and
// the displacement its Imm.
func (in ExtInstr) amd64Memory(op ExtOperand, pos int) (base int, disp int64, err error) {
	if op.Kind != ExtMem {
		return 0, 0, fmt.Errorf("%s: operand %d wants a memory operand, got %s", in.Name, pos, op.Kind)
	}
	if op.Arr != ExtArrNone {
		return 0, 0, fmt.Errorf("%s: operand %d carries an arrangement suffix, the amd64 layer takes none", in.Name, pos)
	}
	if op.Qual != ExtQualNone {
		return 0, 0, fmt.Errorf("%s: operand %d carries a predicate qualifier, the amd64 layer takes none", in.Name, pos)
	}
	if op.HasShift {
		return 0, 0, fmt.Errorf("%s: operand %d carries a shift, the amd64 memory forms take none", in.Name, pos)
	}
	if op.Reg < 0 || op.Reg > 15 {
		return 0, 0, fmt.Errorf("%s: operand %d names base register %d, outside 0-15", in.Name, pos, op.Reg)
	}
	if op.Imm < -1<<31 || op.Imm >= 1<<31 {
		return 0, 0, fmt.Errorf("%s: operand %d carries displacement %d, outside the signed 32-bit range", in.Name, pos, op.Imm)
	}
	return op.Reg, op.Imm, nil
}

// amd64EncodeMemory returns the register-form template with a base-relative
// memory operand filled in: dest and vvvv keep their register meanings, the
// ModR/M r/m field carries the base, and the high base bit rides EVEX.B as
// amd64Encode lays it.  The ModR/M mod bits and the trailing SIB and
// displacement bytes follow the canonical choices the GNU assembler makes
// for the plain, unscaled SDM displacements: no displacement bytes at
// displacement zero, a disp8 when the value fits a signed byte and a disp32
// otherwise, the SIB byte 0x24 when the base is RSP or R12, whose r/m
// encoding 100 demands it, and a forced displacement on RBP and R13, whose
// mod-00 r/m encoding 101 means RIP-relative.  The operand must have passed
// amd64Memory first.
func amd64EncodeMemory(b []byte, dest, vvvv, base int, disp int64) []byte {
	out := amd64Encode(b, dest, vvvv, base)
	rm := base & 7
	var tail []byte
	mod := byte(0)
	switch {
	case rm == 5 || disp != 0:
		// RBP and R13 cannot drop the displacement: mod 00 with r/m 101
		// addresses RIP-relative, not through the base.
		if disp >= -128 && disp <= 127 {
			mod = 1
			tail = []byte{byte(disp)}
		} else {
			mod = 2
			tail = []byte{byte(disp), byte(disp >> 8), byte(disp >> 16), byte(disp >> 24)}
		}
	}
	if rm == 4 {
		// RSP and R12 need the SIB byte: no index, base 100.
		tail = append([]byte{0x24}, tail...)
	}
	out[5] = out[5]&0x3f | mod<<6
	return append(out, tail...)
}

// amd64PlainReg checks the invariants every amd64 register operand carries:
// no arm64 arrangement, no predicate qualifier, and a register number inside
// the class the instruction encodes.
func (in ExtInstr) amd64PlainReg(op ExtOperand, max, pos int) error {
	if op.Arr != ExtArrNone {
		return fmt.Errorf("%s: operand %d carries an arrangement suffix, the amd64 layer takes none", in.Name, pos)
	}
	if op.Qual != ExtQualNone {
		return fmt.Errorf("%s: operand %d carries a predicate qualifier, the amd64 layer takes none", in.Name, pos)
	}
	if op.Reg < 0 || op.Reg > max {
		return fmt.Errorf("%s: operand %d is register %d, outside 0-%d", in.Name, pos, op.Reg, max)
	}
	return nil
}

// amd64Vector checks one vector operand against the class the entry encodes.
func (in ExtInstr) amd64Vector(op ExtOperand, class ExtOperandKind, pos int) error {
	if op.Kind != class {
		article := "a"
		if class == ExtXMM {
			article = "an"
		}
		return fmt.Errorf("%s: operand %d wants %s %s, got %s", in.Name, pos, article, class, op.Kind)
	}
	return in.amd64PlainReg(op, 31, pos)
}

// amd64Gpr checks the general-register operand against the width the entry
// encodes: the W bit picks 32-bit or 64-bit, unless the entry ignores W, and
// the general registers run 0..15.
func (in ExtInstr) amd64Gpr(op ExtOperand, pos int) error {
	want := ExtR32
	if in.Bytes[2]&0x80 != 0 {
		want = ExtR64
	}
	if in.Wig {
		if op.Kind != ExtR32 && op.Kind != ExtR64 {
			return fmt.Errorf("%s: operand %d wants a 32-bit or 64-bit general register, got %s", in.Name, pos, op.Kind)
		}
	} else if op.Kind != want {
		return fmt.Errorf("%s: operand %d wants a %s, got %s", in.Name, pos, want, op.Kind)
	}
	return in.amd64PlainReg(op, 15, pos)
}

// encodeAmd64 encodes the amd64 forms: it validates the operand list against
// the class the template encodes and fills the register bits.  An operand the
// form cannot carry is an error, never a silent mis-encoding.
func (in ExtInstr) encodeAmd64(ops []ExtOperand) ([]byte, error) {
	switch in.Form {
	case ExtFormAmdVec3:
		return in.encodeAmdVec3(ops)
	case ExtFormAmdVec2, ExtFormAmdVec2Half:
		return in.encodeAmdVec2(ops)
	case ExtFormAmdMask2:
		return in.encodeAmdMask2(ops)
	case ExtFormAmdVecGprVec:
		return in.encodeAmdVecGprVec(ops)
	case ExtFormAmdGprVec, ExtFormAmdVecGpr:
		return in.encodeAmdGprPair(ops)
	case ExtFormAmdMemVec:
		return in.encodeAmdMemVec(ops)
	case ExtFormAmdVecMem:
		return in.encodeAmdVecMem(ops)
	case ExtFormAmdVec3Imm:
		return in.encodeAmdVec3Imm(ops)
	case ExtFormAmdMask2Imm:
		return in.encodeAmdMask2Imm(ops)
	default:
		return nil, fmt.Errorf("%s: unknown form %d", in.Name, in.Form)
	}
}

// encodeAmdVec3 fills the non-destructive three-vector form: src1, src2,
// dest, all under one register class.  An entry with Mem set takes the
// memory shape of that position too: the second source of the scalar
// arithmetic, spelled xmm3/m16 in the manual, may be a base-relative
// operand, which rides the r/m field with its displacement bytes after the
// opcode.
func (in ExtInstr) encodeAmdVec3(ops []ExtOperand) ([]byte, error) {
	class := amd64LengthClass(in.Bytes)
	if err := in.amd64Vector(ops[0], class, 1); err != nil {
		return nil, err
	}
	if in.Mem == 2 && ops[1].Kind == ExtMem {
		base, disp, err := in.amd64Memory(ops[1], 2)
		if err != nil {
			return nil, err
		}
		if err := in.amd64Vector(ops[2], class, 3); err != nil {
			return nil, err
		}
		return amd64EncodeMemory(in.Bytes, ops[2].Reg, ops[0].Reg, base, disp), nil
	}
	for i, op := range ops[1:] {
		if err := in.amd64Vector(op, class, i+2); err != nil {
			return nil, err
		}
	}
	return amd64Encode(in.Bytes, ops[2].Reg, ops[0].Reg, ops[1].Reg), nil
}

// encodeAmdMemVec fills the memory-load form: mem, dest.  VMOVSH X30,
// 4660(R8) shape, the manual's xmm1, m16 lines beside the register form.
// The form reads one value from memory, so the third register slot stays
// unused, which the encoding spells as vvvv 1111.
func (in ExtInstr) encodeAmdMemVec(ops []ExtOperand) ([]byte, error) {
	class := amd64LengthClass(in.Bytes)
	base, disp, err := in.amd64Memory(ops[0], 1)
	if err != nil {
		return nil, err
	}
	if err := in.amd64Vector(ops[1], class, 2); err != nil {
		return nil, err
	}
	return amd64EncodeMemory(in.Bytes, ops[1].Reg, -1, base, disp), nil
}

// encodeAmdVecMem fills the memory-store form: src, mem.  VMOVSH 4660(R9),
// X29 shape, the manual's m16, xmm1 lines.  The register source sits in the
// ModR/M reg field and the memory destination in r/m, and vvvv stays
// unused.
func (in ExtInstr) encodeAmdVecMem(ops []ExtOperand) ([]byte, error) {
	class := amd64LengthClass(in.Bytes)
	if err := in.amd64Vector(ops[0], class, 1); err != nil {
		return nil, err
	}
	base, disp, err := in.amd64Memory(ops[1], 2)
	if err != nil {
		return nil, err
	}
	return amd64EncodeMemory(in.Bytes, ops[0].Reg, -1, base, disp), nil
}

// encodeAmdVec2 fills the two-vector form: src, dest.  The half form narrows
// the destination: VCVTNEPS2BF16 converts 512 bits of source into 256 bits
// of destination, and at 128 bits the companion stays the class itself.  An
// entry with Mem set takes the memory shape of that position too: the
// compares and the packed square root read their source from memory, and
// the narrow BF16 convert reads its full-width source there.
func (in ExtInstr) encodeAmdVec2(ops []ExtOperand) ([]byte, error) {
	class := amd64LengthClass(in.Bytes)
	destClass := class
	if in.Form == ExtFormAmdVec2Half {
		destClass = amd64HalfClass(class)
	}
	if in.Mem == 1 && ops[0].Kind == ExtMem {
		base, disp, err := in.amd64Memory(ops[0], 1)
		if err != nil {
			return nil, err
		}
		if err := in.amd64Vector(ops[1], destClass, 2); err != nil {
			return nil, err
		}
		return amd64EncodeMemory(in.Bytes, ops[1].Reg, -1, base, disp), nil
	}
	if in.Mem == 2 && ops[1].Kind == ExtMem {
		if err := in.amd64Vector(ops[0], class, 1); err != nil {
			return nil, err
		}
		base, disp, err := in.amd64Memory(ops[1], 2)
		if err != nil {
			return nil, err
		}
		return amd64EncodeMemory(in.Bytes, ops[0].Reg, -1, base, disp), nil
	}
	if err := in.amd64Vector(ops[0], class, 1); err != nil {
		return nil, err
	}
	if err := in.amd64Vector(ops[1], destClass, 2); err != nil {
		return nil, err
	}
	return amd64Encode(in.Bytes, ops[1].Reg, -1, ops[0].Reg), nil
}

// encodeAmdMask2 fills the mask-destination form: src1, src2, dest, where the
// destination is an opmask register and both sources share the class.
func (in ExtInstr) encodeAmdMask2(ops []ExtOperand) ([]byte, error) {
	class := amd64LengthClass(in.Bytes)
	if err := in.amd64Vector(ops[0], class, 1); err != nil {
		return nil, err
	}
	if err := in.amd64Vector(ops[1], class, 2); err != nil {
		return nil, err
	}
	if ops[2].Kind != ExtKReg {
		return nil, fmt.Errorf("%s: operand 3 wants an opmask register, got %s", in.Name, ops[2].Kind)
	}
	if err := in.amd64PlainReg(ops[2], 7, 3); err != nil {
		return nil, err
	}
	return amd64Encode(in.Bytes, ops[2].Reg, ops[0].Reg, ops[1].Reg), nil
}

// amd64Imm8 validates the leading immediate operand of an imm8-control form:
// an ExtImm with no shift, inside the unsigned byte range, and free of the
// bits the entry's control layout reserves.  The reserved upper nibble of
// the VGETMANTSH control must encode as zero; the SDM marks every other
// layout here fully defined, and the VCMPSH hardware masks its predicate to
// five bits.
func (in ExtInstr) amd64Imm8(op ExtOperand, pos int) (byte, error) {
	if op.Kind != ExtImm {
		return 0, fmt.Errorf("%s: operand %d wants an immediate control byte, got %s", in.Name, pos, op.Kind)
	}
	if op.Arr != ExtArrNone {
		return 0, fmt.Errorf("%s: operand %d carries an arrangement suffix, the amd64 layer takes none", in.Name, pos)
	}
	if op.HasShift {
		return 0, fmt.Errorf("%s: operand %d carries a shift, the amd64 imm8 forms take none", in.Name, pos)
	}
	if op.Imm < 0 || op.Imm > 255 {
		return 0, fmt.Errorf("%s: operand %d is immediate %d, outside the unsigned byte range 0-255", in.Name, pos, op.Imm)
	}
	if in.Imm8 == ExtImm8GetMant && op.Imm > 15 {
		return 0, fmt.Errorf("%s: operand %d is immediate %d, the upper nibble of the mantissa control is reserved and must be zero", in.Name, pos, op.Imm)
	}
	return byte(op.Imm), nil
}

// encodeAmdVec3Imm fills the three-vector form with a control immediate:
// imm, src1, src2, dest, the order the reference listings write it in.
func (in ExtInstr) encodeAmdVec3Imm(ops []ExtOperand) ([]byte, error) {
	class := amd64LengthClass(in.Bytes)
	imm, err := in.amd64Imm8(ops[0], 1)
	if err != nil {
		return nil, err
	}
	for i, op := range ops[1:] {
		if err := in.amd64Vector(op, class, i+2); err != nil {
			return nil, err
		}
	}
	out := amd64Encode(in.Bytes, ops[3].Reg, ops[1].Reg, ops[2].Reg)
	return append(out, imm), nil
}

// encodeAmdMask2Imm fills the opmask-destination form with a control
// immediate: imm, src1, src2, dest.
func (in ExtInstr) encodeAmdMask2Imm(ops []ExtOperand) ([]byte, error) {
	class := amd64LengthClass(in.Bytes)
	imm, err := in.amd64Imm8(ops[0], 1)
	if err != nil {
		return nil, err
	}
	if err := in.amd64Vector(ops[1], class, 2); err != nil {
		return nil, err
	}
	if err := in.amd64Vector(ops[2], class, 3); err != nil {
		return nil, err
	}
	if ops[3].Kind != ExtKReg {
		return nil, fmt.Errorf("%s: operand 4 wants an opmask register, got %s", in.Name, ops[3].Kind)
	}
	if err := in.amd64PlainReg(ops[3], 7, 4); err != nil {
		return nil, err
	}
	out := amd64Encode(in.Bytes, ops[3].Reg, ops[1].Reg, ops[2].Reg)
	return append(out, imm), nil
}

// ExtFP16RoundingModes names the two-bit rounding mode the round control of
// VRNDSCALESH and VREDUCESH carries, indexed by imm8[1:0], the SDM's RC
// field encoding.
var ExtFP16RoundingModes = [4]string{
	"round to nearest (even)",
	"round down (toward -infinity)",
	"round up (toward +infinity)",
	"round toward zero (truncate)",
}

// ExtFP16GetMantSigns names the sign control imm8[3:2] of the VGETMANTSH
// immediate, indexed by the field: the source's own sign, a forced positive,
// and the two encodings that yield the indefinite NaN on a negative source.
var ExtFP16GetMantSigns = [4]string{
	"the sign of the source",
	"positive",
	"the indefinite NaN when the source is negative",
	"the indefinite NaN when the source is negative",
}

// ExtFP16CmpPredicates names the 32 comparison predicates the VCMPSH
// immediate carries in imm8[4:0], in encoding order.  The SDM's own
// spellings are the fixed vocabulary of the predicate suffixes.
var ExtFP16CmpPredicates = [32]string{
	"EQ_OQ", "LT_OS", "LE_OS", "UNORD_Q", "NEQ_UQ", "NLT_US", "NLE_US", "ORD_Q",
	"EQ_UQ", "NGE_US", "NGT_US", "FALSE_OQ", "NEQ_OQ", "GE_OS", "GT_OS", "TRUE_UQ",
	"EQ_OS", "LT_OQ", "LE_OQ", "UNORD_S", "NEQ_US", "NLT_UQ", "NLE_UQ", "ORD_S",
	"EQ_US", "NGE_UQ", "NGT_UQ", "FALSE_OS", "NEQ_OS", "GE_OQ", "GT_OQ", "TRUE_US",
}

// encodeAmdVecGprVec fills the conversion form with a general-register
// source: src1, gpr, dest.  VCVTSI2SH XMM1, XMM2, EAX style.
func (in ExtInstr) encodeAmdVecGprVec(ops []ExtOperand) ([]byte, error) {
	class := amd64LengthClass(in.Bytes)
	if err := in.amd64Vector(ops[0], class, 1); err != nil {
		return nil, err
	}
	if err := in.amd64Gpr(ops[1], 2); err != nil {
		return nil, err
	}
	if err := in.amd64Vector(ops[2], class, 3); err != nil {
		return nil, err
	}
	return amd64Encode(in.Bytes, ops[2].Reg, ops[0].Reg, ops[1].Reg), nil
}

// encodeAmdGprPair fills the two-operand general-register forms: gpr, vec
// (the move into a vector register and the integer conversions) and vec, gpr
// (the move out of one).  In both orders the second operand is the
// destination in the reg field and the first the r/m source; the vector
// changes position with the form.
func (in ExtInstr) encodeAmdGprPair(ops []ExtOperand) ([]byte, error) {
	class := amd64LengthClass(in.Bytes)
	vecPos := 1
	if in.Form == ExtFormAmdVecGpr {
		vecPos = 0
	}
	if err := in.amd64Vector(ops[vecPos], class, vecPos+1); err != nil {
		return nil, err
	}
	if err := in.amd64Gpr(ops[1-vecPos], 2-vecPos); err != nil {
		return nil, err
	}
	return amd64Encode(in.Bytes, ops[1].Reg, -1, ops[0].Reg), nil
}

// --- the amd64 AVX512-BF16 and VP2INTERSECT table -----------------------------

// amd64Extensions is the extended-instruction layer of amd64.  The encodings
// are transcribed from the Intel SDM instruction entries and cross-checked
// against binutils-gdb's assembler testsuite (gas/testsuite/gas/i386/
// avx512_bf16.d, avx512_bf16_vl.d and x86-64-vp2intersect.d), whose register
// forms the golden vectors in amd64_ext_test.go quote byte for byte.  Each
// template carries the fixed bits of one encoding with every register-derived
// bit zero: the map selection in byte one, the W bit, the mandatory prefix
// and the reserved one-bit in byte two, the vector length in byte three, and
// the ModR/M mod bits.
var amd64Extensions = []ExtInstr{
	// AVX512-BF16: the two-way packed single to BF16 conversion and the
	// dot product accumulate.  The prefixes differ inside the family, the
	// three-register convert carries F2 while the narrow convert and the dot
	// product carry F3, which the golden vectors pin byte for byte.
	{Name: "VCVTNE2PS2BF16", Summary: "Convert two packed single-precision vectors to packed BF16, truncating",
		Bytes: []byte{0x62, 0x02, 0x07, 0x40, 0x72, 0xC0}, Form: ExtFormAmdVec3, Feature: ExtFeatureBF16,
		Ref: "Intel SDM Vol. 2C, VCVTNE2PS2BF16 (EVEX.NDS.512.F2.0F38.W0 72 /r)"},
	{Name: "VCVTNE2PS2BF16", Summary: "Convert two packed single-precision vectors to packed BF16, truncating",
		Bytes: []byte{0x62, 0x02, 0x07, 0x20, 0x72, 0xC0}, Form: ExtFormAmdVec3, Feature: ExtFeatureBF16,
		Ref: "Intel SDM Vol. 2C, VCVTNE2PS2BF16 (EVEX.NDS.256.F2.0F38.W0 72 /r)"},
	{Name: "VCVTNE2PS2BF16", Summary: "Convert two packed single-precision vectors to packed BF16, truncating",
		Bytes: []byte{0x62, 0x02, 0x07, 0x00, 0x72, 0xC0}, Form: ExtFormAmdVec3, Feature: ExtFeatureBF16,
		Ref: "Intel SDM Vol. 2C, VCVTNE2PS2BF16 (EVEX.NDS.128.F2.0F38.W0 72 /r)"},
	{Name: "VCVTNEPS2BF16", Summary: "Convert packed single precision to packed BF16, truncating, half-width destination",
		Bytes: []byte{0x62, 0x02, 0x06, 0x40, 0x72, 0xC0}, Form: ExtFormAmdVec2Half, Mem: 1, Feature: ExtFeatureBF16,
		Ref: "Intel SDM Vol. 2C, VCVTNEPS2BF16 (EVEX.512.F3.0F38.W0 72 /r, YMM destination)"},
	{Name: "VCVTNEPS2BF16", Summary: "Convert packed single precision to packed BF16, truncating, half-width destination",
		Bytes: []byte{0x62, 0x02, 0x06, 0x20, 0x72, 0xC0}, Form: ExtFormAmdVec2Half, Mem: 1, Feature: ExtFeatureBF16,
		Ref: "Intel SDM Vol. 2C, VCVTNEPS2BF16 (EVEX.256.F3.0F38.W0 72 /r, XMM destination)"},
	{Name: "VCVTNEPS2BF16", Summary: "Convert packed single precision to packed BF16, truncating, half-width destination",
		Bytes: []byte{0x62, 0x02, 0x06, 0x00, 0x72, 0xC0}, Form: ExtFormAmdVec2Half, Mem: 1, Feature: ExtFeatureBF16,
		Ref: "Intel SDM Vol. 2C, VCVTNEPS2BF16 (EVEX.128.F3.0F38.W0 72 /r, XMM destination)"},
	{Name: "VDPBF16PS", Summary: "Multiply BF16 pairs and accumulate the dot product into single precision",
		Bytes: []byte{0x62, 0x02, 0x06, 0x40, 0x52, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureBF16,
		Ref: "Intel SDM Vol. 2C, VDPBF16PS (EVEX.NDS.512.F3.0F38.W0 52 /r)"},
	{Name: "VDPBF16PS", Summary: "Multiply BF16 pairs and accumulate the dot product into single precision",
		Bytes: []byte{0x62, 0x02, 0x06, 0x20, 0x52, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureBF16,
		Ref: "Intel SDM Vol. 2C, VDPBF16PS (EVEX.NDS.256.F3.0F38.W0 52 /r)"},
	{Name: "VDPBF16PS", Summary: "Multiply BF16 pairs and accumulate the dot product into single precision",
		Bytes: []byte{0x62, 0x02, 0x06, 0x00, 0x52, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureBF16,
		Ref: "Intel SDM Vol. 2C, VDPBF16PS (EVEX.NDS.128.F3.0F38.W0 52 /r)"},

	// AVX512-VP2INTERSECT: the pairwise intersection indices, one opmask
	// destination and two vector sources, EVEX.NDS.66.0F38.  The instruction
	// takes no write mask of its own.
	{Name: "VP2INTERSECTD", Summary: "Store the indices of the first pairwise intersections of two dword vectors",
		Bytes: []byte{0x62, 0x02, 0x07, 0x40, 0x68, 0xC0}, Form: ExtFormAmdMask2, Feature: ExtFeatureVP2INTERSECT,
		Ref: "Intel SDM Vol. 2C, VP2INTERSECTD/VP2INTERSECTQ (EVEX.NDS.512.F2.0F38.W0 68 /r)"},
	{Name: "VP2INTERSECTD", Summary: "Store the indices of the first pairwise intersections of two dword vectors",
		Bytes: []byte{0x62, 0x02, 0x07, 0x20, 0x68, 0xC0}, Form: ExtFormAmdMask2, Feature: ExtFeatureVP2INTERSECT,
		Ref: "Intel SDM Vol. 2C, VP2INTERSECTD/VP2INTERSECTQ (EVEX.NDS.256.F2.0F38.W0 68 /r)"},
	{Name: "VP2INTERSECTD", Summary: "Store the indices of the first pairwise intersections of two dword vectors",
		Bytes: []byte{0x62, 0x02, 0x07, 0x00, 0x68, 0xC0}, Form: ExtFormAmdMask2, Feature: ExtFeatureVP2INTERSECT,
		Ref: "Intel SDM Vol. 2C, VP2INTERSECTD/VP2INTERSECTQ (EVEX.NDS.128.F2.0F38.W0 68 /r)"},
	{Name: "VP2INTERSECTQ", Summary: "Store the indices of the first pairwise intersections of two qword vectors",
		Bytes: []byte{0x62, 0x02, 0x87, 0x40, 0x68, 0xC0}, Form: ExtFormAmdMask2, Feature: ExtFeatureVP2INTERSECT,
		Ref: "Intel SDM Vol. 2C, VP2INTERSECTD/VP2INTERSECTQ (EVEX.NDS.512.F2.0F38.W1 68 /r)"},
	{Name: "VP2INTERSECTQ", Summary: "Store the indices of the first pairwise intersections of two qword vectors",
		Bytes: []byte{0x62, 0x02, 0x87, 0x20, 0x68, 0xC0}, Form: ExtFormAmdMask2, Feature: ExtFeatureVP2INTERSECT,
		Ref: "Intel SDM Vol. 2C, VP2INTERSECTD/VP2INTERSECTQ (EVEX.NDS.256.F2.0F38.W1 68 /r)"},
	{Name: "VP2INTERSECTQ", Summary: "Store the indices of the first pairwise intersections of two qword vectors",
		Bytes: []byte{0x62, 0x02, 0x87, 0x00, 0x68, 0xC0}, Form: ExtFormAmdMask2, Feature: ExtFeatureVP2INTERSECT,
		Ref: "Intel SDM Vol. 2C, VP2INTERSECTD/VP2INTERSECTQ (EVEX.NDS.128.F2.0F38.W1 68 /r)"},

	// AVX512-FP16, the scalar core: move, arithmetic, compare and
	// conversion on one half-precision value in the low XMM lane, the
	// register forms of the manual's scalar entries.  The family lives in
	// the EVEX maps five and six the toolchain has never emitted, with the
	// mandatory prefixes the manual gives each entry; the golden vectors
	// pin every prefix byte for byte.  LIG encodes as L'L = 00, the XMM
	// class alone.
	{Name: "VMOVSH", Summary: "Move a scalar FP16 value",
		Bytes: []byte{0x62, 0x05, 0x06, 0x00, 0x10, 0xC0}, Form: ExtFormAmdVec3, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VMOVSH (EVEX.NDS.LIG.F3.MAP5.W0 10 /r)"},
	{Name: "VMOVSH", Summary: "Move a scalar FP16 value from memory into an XMM register",
		Bytes: []byte{0x62, 0x05, 0x06, 0x00, 0x10, 0xC0}, Form: ExtFormAmdMemVec, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VMOVSH (EVEX.LIG.F3.MAP5.W0 10 /r, m16 source)"},
	{Name: "VMOVSH", Summary: "Move a scalar FP16 value from an XMM register to memory",
		Bytes: []byte{0x62, 0x05, 0x06, 0x00, 0x11, 0xC0}, Form: ExtFormAmdVecMem, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VMOVSH (EVEX.LIG.F3.MAP5.W0 11 /r, m16 destination)"},
	{Name: "VMOVW", Summary: "Move a word between a general register and an XMM register",
		Bytes: []byte{0x62, 0x05, 0x05, 0x00, 0x6E, 0xC0}, Form: ExtFormAmdGprVec, Feature: ExtFeatureFP16, Wig: true,
		Ref: "Intel SDM Vol. 2C, VMOVW (EVEX.128.66.MAP5.WIG 6E /r)"},
	{Name: "VMOVW", Summary: "Move a word between an XMM register and a general register",
		Bytes: []byte{0x62, 0x05, 0x05, 0x00, 0x7E, 0xC0}, Form: ExtFormAmdVecGpr, Feature: ExtFeatureFP16, Wig: true,
		Ref: "Intel SDM Vol. 2C, VMOVW (EVEX.128.66.MAP5.WIG 7E /r)"},
	{Name: "VMOVW", Summary: "Move a word from memory into an XMM register",
		Bytes: []byte{0x62, 0x05, 0x05, 0x00, 0x6E, 0xC0}, Form: ExtFormAmdMemVec, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VMOVW (EVEX.128.66.MAP5.WIG 6E /r, m16 source)"},
	{Name: "VMOVW", Summary: "Move a word from an XMM register to memory",
		Bytes: []byte{0x62, 0x05, 0x05, 0x00, 0x7E, 0xC0}, Form: ExtFormAmdVecMem, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VMOVW (EVEX.128.66.MAP5.WIG 7E /r, m16 destination)"},
	{Name: "VADDSH", Summary: "Add scalar FP16 values",
		Bytes: []byte{0x62, 0x05, 0x06, 0x00, 0x58, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VADDSH (EVEX.NDS.LIG.F3.MAP5.W0 58 /r)"},
	{Name: "VSUBSH", Summary: "Subtract scalar FP16 values",
		Bytes: []byte{0x62, 0x05, 0x06, 0x00, 0x5C, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VSUBSH (EVEX.NDS.LIG.F3.MAP5.W0 5C /r)"},
	{Name: "VMULSH", Summary: "Multiply scalar FP16 values",
		Bytes: []byte{0x62, 0x05, 0x06, 0x00, 0x59, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VMULSH (EVEX.NDS.LIG.F3.MAP5.W0 59 /r)"},
	{Name: "VDIVSH", Summary: "Divide scalar FP16 values",
		Bytes: []byte{0x62, 0x05, 0x06, 0x00, 0x5E, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VDIVSH (EVEX.NDS.LIG.F3.MAP5.W0 5E /r)"},
	{Name: "VMINSH", Summary: "Return the minimum of scalar FP16 values",
		Bytes: []byte{0x62, 0x05, 0x06, 0x00, 0x5D, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VMINSH (EVEX.NDS.LIG.F3.MAP5.W0 5D /r)"},
	{Name: "VMAXSH", Summary: "Return the maximum of scalar FP16 values",
		Bytes: []byte{0x62, 0x05, 0x06, 0x00, 0x5F, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VMAXSH (EVEX.NDS.LIG.F3.MAP5.W0 5F /r)"},
	{Name: "VSQRTSH", Summary: "Compute the square root of a scalar FP16 value",
		Bytes: []byte{0x62, 0x05, 0x06, 0x00, 0x51, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VSQRTSH (EVEX.NDS.LIG.F3.MAP5.W0 51 /r)"},
	{Name: "VSCALEFSH", Summary: "Scale a scalar FP16 value by the ratio of two others",
		Bytes: []byte{0x62, 0x06, 0x05, 0x00, 0x2D, 0xC0}, Form: ExtFormAmdVec3, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VSCALEFSH (EVEX.NDS.LIG.66.MAP6.W0 2D /r)"},
	{Name: "VGETEXPSH", Summary: "Convert the exponent of a scalar FP16 value to an FP16 value",
		Bytes: []byte{0x62, 0x06, 0x05, 0x00, 0x43, 0xC0}, Form: ExtFormAmdVec3, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VGETEXPSH (EVEX.NDS.LIG.66.MAP6.W0 43 /r)"},
	{Name: "VCOMISH", Summary: "Compare a scalar FP16 value and set EFLAGS",
		Bytes: []byte{0x62, 0x05, 0x04, 0x00, 0x2F, 0xC0}, Form: ExtFormAmdVec2, Mem: 1, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VCOMISH (EVEX.LIG.MAP5.W0 2F /r)"},
	{Name: "VUCOMISH", Summary: "Unordered-compare a scalar FP16 value and set EFLAGS",
		Bytes: []byte{0x62, 0x05, 0x04, 0x00, 0x2E, 0xC0}, Form: ExtFormAmdVec2, Mem: 1, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VUCOMISH (EVEX.LIG.MAP5.W0 2E /r)"},
	{Name: "VCVTSS2SH", Summary: "Convert one FP32 value to one FP16 value",
		Bytes: []byte{0x62, 0x05, 0x04, 0x00, 0x1D, 0xC0}, Form: ExtFormAmdVec3, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VCVTSS2SH (EVEX.NDS.LIG.MAP5.W0 1D /r)"},
	{Name: "VCVTSH2SS", Summary: "Convert a low FP16 value to an FP32 value",
		Bytes: []byte{0x62, 0x06, 0x04, 0x00, 0x13, 0xC0}, Form: ExtFormAmdVec3, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VCVTSH2SS (EVEX.NDS.LIG.MAP6.W0 13 /r)"},
	{Name: "VCVTSH2SD", Summary: "Convert a low FP16 value to an FP64 value",
		Bytes: []byte{0x62, 0x05, 0x06, 0x00, 0x5A, 0xC0}, Form: ExtFormAmdVec3, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VCVTSH2SD (EVEX.NDS.LIG.F3.MAP5.W0 5A /r)"},
	{Name: "VCVTSD2SH", Summary: "Convert one FP64 value to one FP16 value",
		Bytes: []byte{0x62, 0x05, 0x87, 0x00, 0x5A, 0xC0}, Form: ExtFormAmdVec3, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VCVTSD2SH (EVEX.NDS.LIG.F2.MAP5.W1 5A /r)"},
	{Name: "VCVTSI2SH", Summary: "Convert one signed 32-bit integer to one FP16 value",
		Bytes: []byte{0x62, 0x05, 0x06, 0x00, 0x2A, 0xC0}, Form: ExtFormAmdVecGprVec, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VCVTSI2SH (EVEX.NDS.LIG.F3.MAP5.W0 2A /r)"},
	{Name: "VCVTSI2SH", Summary: "Convert one signed 64-bit integer to one FP16 value",
		Bytes: []byte{0x62, 0x05, 0x86, 0x00, 0x2A, 0xC0}, Form: ExtFormAmdVecGprVec, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VCVTSI2SH (EVEX.NDS.LIG.F3.MAP5.W1 2A /r)"},
	{Name: "VCVTUSI2SH", Summary: "Convert one unsigned 32-bit integer to one FP16 value",
		Bytes: []byte{0x62, 0x05, 0x06, 0x00, 0x7B, 0xC0}, Form: ExtFormAmdVecGprVec, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VCVTUSI2SH (EVEX.NDS.LIG.F3.MAP5.W0 7B /r)"},
	{Name: "VCVTUSI2SH", Summary: "Convert one unsigned 64-bit integer to one FP16 value",
		Bytes: []byte{0x62, 0x05, 0x86, 0x00, 0x7B, 0xC0}, Form: ExtFormAmdVecGprVec, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VCVTUSI2SH (EVEX.NDS.LIG.F3.MAP5.W1 7B /r)"},
	{Name: "VCVTSH2SI", Summary: "Convert a low FP16 value to a signed 32-bit integer",
		Bytes: []byte{0x62, 0x05, 0x06, 0x00, 0x2D, 0xC0}, Form: ExtFormAmdVecGpr, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VCVTSH2SI (EVEX.LIG.F3.MAP5.W0 2D /r)"},
	{Name: "VCVTSH2SI", Summary: "Convert a low FP16 value to a signed 64-bit integer",
		Bytes: []byte{0x62, 0x05, 0x86, 0x00, 0x2D, 0xC0}, Form: ExtFormAmdVecGpr, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VCVTSH2SI (EVEX.LIG.F3.MAP5.W1 2D /r)"},
	{Name: "VCVTSH2USI", Summary: "Convert a low FP16 value to an unsigned 32-bit integer",
		Bytes: []byte{0x62, 0x05, 0x06, 0x00, 0x79, 0xC0}, Form: ExtFormAmdVecGpr, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VCVTSH2USI (EVEX.LIG.F3.MAP5.W0 79 /r)"},
	{Name: "VCVTSH2USI", Summary: "Convert a low FP16 value to an unsigned 64-bit integer",
		Bytes: []byte{0x62, 0x05, 0x86, 0x00, 0x79, 0xC0}, Form: ExtFormAmdVecGpr, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VCVTSH2USI (EVEX.LIG.F3.MAP5.W1 79 /r)"},

	// AVX512-FP16 packed arithmetic: the full ZMM lanes the scalar core
	// mirrors plus the VL forms, EVEX.NDS.MAP5 with no mandatory prefix,
	// rounding control left to MXCSR.  The 512-bit register forms are
	// quoted from x86-64-avx512_fp16.d, the 256- and 128-bit ones from
	// avx512_fp16_vl.d, on the same low registers the suite uses.
	{Name: "VADDPH", Summary: "Add packed FP16 values",
		Bytes: []byte{0x62, 0x05, 0x04, 0x40, 0x58, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VADDPH (EVEX.NDS.512.MAP5.W0 58 /r)"},
	{Name: "VADDPH", Summary: "Add packed FP16 values",
		Bytes: []byte{0x62, 0x05, 0x04, 0x20, 0x58, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VADDPH (EVEX.NDS.256.MAP5.W0 58 /r)"},
	{Name: "VADDPH", Summary: "Add packed FP16 values",
		Bytes: []byte{0x62, 0x05, 0x04, 0x00, 0x58, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VADDPH (EVEX.NDS.128.MAP5.W0 58 /r)"},
	{Name: "VSUBPH", Summary: "Subtract packed FP16 values",
		Bytes: []byte{0x62, 0x05, 0x04, 0x40, 0x5C, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VSUBPH (EVEX.NDS.512.MAP5.W0 5C /r)"},
	{Name: "VSUBPH", Summary: "Subtract packed FP16 values",
		Bytes: []byte{0x62, 0x05, 0x04, 0x20, 0x5C, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VSUBPH (EVEX.NDS.256.MAP5.W0 5C /r)"},
	{Name: "VSUBPH", Summary: "Subtract packed FP16 values",
		Bytes: []byte{0x62, 0x05, 0x04, 0x00, 0x5C, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VSUBPH (EVEX.NDS.128.MAP5.W0 5C /r)"},
	{Name: "VMULPH", Summary: "Multiply packed FP16 values",
		Bytes: []byte{0x62, 0x05, 0x04, 0x40, 0x59, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VMULPH (EVEX.NDS.512.MAP5.W0 59 /r)"},
	{Name: "VMULPH", Summary: "Multiply packed FP16 values",
		Bytes: []byte{0x62, 0x05, 0x04, 0x20, 0x59, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VMULPH (EVEX.NDS.256.MAP5.W0 59 /r)"},
	{Name: "VMULPH", Summary: "Multiply packed FP16 values",
		Bytes: []byte{0x62, 0x05, 0x04, 0x00, 0x59, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VMULPH (EVEX.NDS.128.MAP5.W0 59 /r)"},
	{Name: "VDIVPH", Summary: "Divide packed FP16 values",
		Bytes: []byte{0x62, 0x05, 0x04, 0x40, 0x5E, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VDIVPH (EVEX.NDS.512.MAP5.W0 5E /r)"},
	{Name: "VDIVPH", Summary: "Divide packed FP16 values",
		Bytes: []byte{0x62, 0x05, 0x04, 0x20, 0x5E, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VDIVPH (EVEX.NDS.256.MAP5.W0 5E /r)"},
	{Name: "VDIVPH", Summary: "Divide packed FP16 values",
		Bytes: []byte{0x62, 0x05, 0x04, 0x00, 0x5E, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VDIVPH (EVEX.NDS.128.MAP5.W0 5E /r)"},
	{Name: "VMINPH", Summary: "Return the minimum of packed FP16 values",
		Bytes: []byte{0x62, 0x05, 0x04, 0x40, 0x5D, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VMINPH (EVEX.NDS.512.MAP5.W0 5D /r)"},
	{Name: "VMINPH", Summary: "Return the minimum of packed FP16 values",
		Bytes: []byte{0x62, 0x05, 0x04, 0x20, 0x5D, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VMINPH (EVEX.NDS.256.MAP5.W0 5D /r)"},
	{Name: "VMINPH", Summary: "Return the minimum of packed FP16 values",
		Bytes: []byte{0x62, 0x05, 0x04, 0x00, 0x5D, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VMINPH (EVEX.NDS.128.MAP5.W0 5D /r)"},
	{Name: "VMAXPH", Summary: "Return the maximum of packed FP16 values",
		Bytes: []byte{0x62, 0x05, 0x04, 0x40, 0x5F, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VMAXPH (EVEX.NDS.512.MAP5.W0 5F /r)"},
	{Name: "VMAXPH", Summary: "Return the maximum of packed FP16 values",
		Bytes: []byte{0x62, 0x05, 0x04, 0x20, 0x5F, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VMAXPH (EVEX.NDS.256.MAP5.W0 5F /r)"},
	{Name: "VMAXPH", Summary: "Return the maximum of packed FP16 values",
		Bytes: []byte{0x62, 0x05, 0x04, 0x00, 0x5F, 0xC0}, Form: ExtFormAmdVec3, Mem: 2, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VMAXPH (EVEX.NDS.128.MAP5.W0 5F /r)"},
	{Name: "VSQRTPH", Summary: "Compute the square root of packed FP16 values",
		Bytes: []byte{0x62, 0x05, 0x04, 0x40, 0x51, 0xC0}, Form: ExtFormAmdVec2, Mem: 1, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VSQRTPH (EVEX.512.MAP5.W0 51 /r)"},
	{Name: "VSQRTPH", Summary: "Compute the square root of packed FP16 values",
		Bytes: []byte{0x62, 0x05, 0x04, 0x20, 0x51, 0xC0}, Form: ExtFormAmdVec2, Mem: 1, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VSQRTPH (EVEX.256.MAP5.W0 51 /r)"},
	{Name: "VSQRTPH", Summary: "Compute the square root of packed FP16 values",
		Bytes: []byte{0x62, 0x05, 0x04, 0x00, 0x51, 0xC0}, Form: ExtFormAmdVec2, Mem: 1, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VSQRTPH (EVEX.128.MAP5.W0 51 /r)"},

	// AVX512-FP16 scalar, the imm8-control group: mantissa extraction,
	// reduction, rounding to fraction bits and the compare into an opmask.
	// Each carries its control byte as the leading immediate operand, the
	// order the reference listings write it in.  The controls live in map
	// 0F3A: the compare with the F3 prefix the manual gives the compare
	// family, the other three unprefixed.  The immediate layouts and their
	// tables are ExtFP16RoundingModes, ExtFP16GetMantSigns and
	// ExtFP16CmpPredicates above; the reserved upper nibble of the mantissa
	// control is refused rather than encoded.
	{Name: "VCMPSH", Summary: "Compare scalar FP16 values into an opmask under an imm8 predicate",
		Bytes: []byte{0x62, 0x03, 0x06, 0x00, 0xC2, 0xC0}, Form: ExtFormAmdMask2Imm, Imm8: ExtImm8CmpPredicate, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VCMPSH (EVEX.LLIG.F3.0F3A.W0 C2 /r /ib)"},
	{Name: "VGETMANTSH", Summary: "Extract the normalised mantissa of a scalar FP16 value under an imm8 control",
		Bytes: []byte{0x62, 0x03, 0x04, 0x00, 0x27, 0xC0}, Form: ExtFormAmdVec3Imm, Imm8: ExtImm8GetMant, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VGETMANTSH (EVEX.LLIG.NP.0F3A.W0 27 /r /ib)"},
	{Name: "VREDUCESH", Summary: "Reduce a scalar FP16 value by imm8 fraction bits under an imm8 round control",
		Bytes: []byte{0x62, 0x03, 0x04, 0x00, 0x57, 0xC0}, Form: ExtFormAmdVec3Imm, Imm8: ExtImm8ScaleRound, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VREDUCESH (EVEX.LLIG.NP.0F3A.W0 57 /r /ib)"},
	{Name: "VRNDSCALESH", Summary: "Round a scalar FP16 value to imm8 fraction bits under an imm8 round control",
		Bytes: []byte{0x62, 0x03, 0x04, 0x00, 0x0A, 0xC0}, Form: ExtFormAmdVec3Imm, Imm8: ExtImm8ScaleRound, Feature: ExtFeatureFP16,
		Ref: "Intel SDM Vol. 2C, VRNDSCALESH (EVEX.LLIG.NP.0F3A.W0 0A /r /ib)"},
}