feat(asm): byte-identical go-flac AVX2 assembly with scalar families and jump relaxation

Assisted-by: Qwen 3.8 Max Preview

asm/vex.go

@@ -133,6 +133,17 @@ var vexTable = map[string]vexSpec{
 	"VZEROUPPER": {1, 0x77, 0, 0, -1, vexZero},
 }
 
+// vexVarShift maps the shift mnemonics to their variable-count opcode — the
+// form whose count comes from an XMM register or memory (VPSRLQ X0, Y8, Y8),
+// an ordinary NDS encoding rather than the /digit immediate form above.
+var vexVarShift = map[string]byte{
+	"VPSLLD": 0xF2,
+	"VPSLLQ": 0xF3,
+	"VPSRAD": 0xE2,
+	"VPSRLD": 0xD2,
+	"VPSRLQ": 0xD3,
+}
+
 // vexMoveSpec describes a VEX move, which takes different opcodes (and
 // sometimes a different VEX.W) per operand direction.  The Go assembler
 // encodes a vector→vector move with the store-form opcode (reg = source,
@@ -180,6 +191,17 @@ func (e *enc) encodeVex(mnemUpper string, ops []Operand) error {
 	if ms, ok := vexMoveTable[mnemUpper]; ok {
 		return e.encodeVexMove(mnemUpper, ms, ops)
 	}
+	// The shifts come in two shapes under one mnemonic: an immediate count
+	// ($imm, src, dst) and a variable count in an XMM register or memory
+	// (count, src, dst), the latter an ordinary NDS form.
+	if op, ok := vexVarShift[mnemUpper]; ok && len(ops) == 3 {
+		if _, isImm := ops[0].(Imm); !isImm {
+			if !vecOrMem(ops[0]) {
+				return fmt.Errorf("%s: shift count must be an immediate, a vector register or memory", mnemUpper)
+			}
+			return e.encodeVexNDS3(vexSpec{mapSel: 1, opcode: op, pp: 1, opdigit: -1, form: vexNDS3}, ops)
+		}
+	}
 	spec := vexTable[mnemUpper]
 	switch spec.form {
 	case vexNDS3:
@@ -483,6 +505,15 @@ func vecReg(op Operand) (Reg, bool) {
 	return r, ok && r.isVec()
 }
 
+// vecOrMem reports whether op is a vector register or a memory reference.
+func vecOrMem(op Operand) bool {
+	if _, ok := op.(Mem); ok {
+		return true
+	}
+	r, ok := op.(Reg)
+	return ok && r.isVec()
+}
+
 // validMoveOther reports whether the non-vector operand of a move is
 // acceptable: memory always is, a GPR only for VMOVD/VMOVQ.
 func validMoveOther(ms vexMoveSpec, op Operand) bool {