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

Assisted-by: Qwen 3.8 Max Preview
This commit is contained in:
2026-08-02 23:11:30 +02:00
parent db8e3fc160
commit 5af12e15ac
7 changed files with 143 additions and 11 deletions
+46 -5
View File
@@ -377,6 +377,37 @@ var evexTable = map[string]evexSpec{
"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).
@@ -615,6 +646,12 @@ var evexRound = map[string]bool{
"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
@@ -801,19 +838,23 @@ 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. The destination may be an opmask register (the
// *2M mask conversions), in which case the vector length comes from the
// source.
// *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() && !dstReg.mask) {
return fmt.Errorf("EVEX destination must be a vector or mask register")
if !ok {
return fmt.Errorf("EVEX destination must be a register")
}
ll := dstReg.vecLenBit()
if dstReg.mask {
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()
}