fix(asm): compress movq immediates to the go-tool-asm imm32 forms

Assisted-by: GLM 5.3

asm/encode_test.go

@@ -57,8 +57,11 @@ func TestMov(t *testing.T) {
 	checkSyntax(t, "mov qword ptr [rbx], rax", "MOVQ", AX, Ptr(BX, 0, 8))
 	checkSyntax(t, "mov rbx, qword ptr [rax+0x10]", "MOVQ", Ptr(AX, 0x10, 8), BX)
 	checkSyntax(t, "mov rbx, qword ptr [rsi+4*rbx]", "MOVQ", Idx(SI, BX, 4, 0, 8), BX)
-	checkSyntax(t, "mov rax, 0x5", "MOVQ", Imm(5), AX)
+	// A small positive immediate compresses to the 32-bit zero-extending
-	checkSyntax(t, "mov r8, 0x5", "MOVQ", Imm(5), Reg{idx: 8, size: 8})
+	// form (matching go tool asm), so the disassembler renders the 32-bit
+	// register name even for MOVQ.
+	checkSyntax(t, "mov eax, 0x5", "MOVQ", Imm(5), AX)
+	checkSyntax(t, "mov r8d, 0x5", "MOVQ", Imm(5), Reg{idx: 8, size: 8})
 	checkSyntax(t, "mov qword ptr [rax], 0x5", "MOVQ", Imm(5), Ptr(AX, 0, 8))
 	checkSyntax(t, "mov r12, r13", "MOVQ", Reg{idx: 13, size: 8}, Reg{idx: 12, size: 8})
 }
@@ -211,6 +214,13 @@ func TestScalarGroundTruth(t *testing.T) {
 		{"BSRQ AX,CX", "BSRQ", []Operand{AX, CX}, "480fbdc8", "BSR"},
 		{"POPCNTL AX,CX", "POPCNTL", []Operand{AX, CX}, "f30fb8c8", "POPCNT"},
 		{"POPCNTQ R8,R9", "POPCNTQ", []Operand{r8, r9}, "f34d0fb8c8", "POPCNT"},
+		// A 64-bit immediate that fits a signed int32 is compressed exactly
+		// as the Go assembler does: positive via B8+rd without REX.W
+		// (zero-extended), negative via REX.W C7 /0 (sign-extended).
+		{"MOVQ $4,BX", "MOVQ", []Operand{Imm(4), BX}, "bb04000000", "MOV"},
+		{"MOVQ $4,R8", "MOVQ", []Operand{Imm(4), r8}, "41b804000000", "MOV"},
+		{"MOVQ $-1,BX", "MOVQ", []Operand{Imm(-1), BX}, "48c7c3ffffffff", "MOV"},
+		{"MOVQ big,BX", "MOVQ", []Operand{Imm(0x1122334455667788), BX}, "48bb8877665544332211", "MOV"},
 		{"CMOVLGT CX,AX", "CMOVLGT", []Operand{CX, AX}, "0f4fc1", "CMOVG"},
 		{"CMOVLEQ CX,AX", "CMOVLEQ", []Operand{CX, AX}, "0f44c1", "CMOVE"},
 		{"CMOVQGT R9,R8", "CMOVQGT", []Operand{r9, r8}, "4d0f4fc1", "CMOVG"},

asm/instrs.go

@@ -91,7 +91,28 @@ func (e *enc) encodeMov(ops []Operand, size int) error {
 
 	case Imm:
 		if dstIsReg {
-			// MOV r, imm: 0xB0+reg (8-bit) / 0xB8+reg (16/32/64, imm64 for Q).
+			v := int64(src)
+			// The Go assembler compresses 64-bit moves whose immediate fits
+			// a signed int32, choosing per sign:
+			//   v >= 0: B8+rd imm32 without REX.W (zero-extended by the
+			//           hardware, REX.B still emitted for R8-R15);
+			//   v <  0: REX.W C7 /0 imm32 (sign-extended — the plain B8+rd
+			//           form would zero-extend and corrupt the value).
+			// Out-of-range immediates keep the B8+rd imm64 form.
+			if size == 8 && v >= 0 && v <= (1<<31)-1 {
+				i := newInstr(4, []byte{0xB8 + byte(dstReg.idx&7)})
+				i.rexB = dstReg.idx >= 8
+				i.imm = le32(v)
+				return e.emit(i)
+			}
+			if size == 8 && v < 0 && v >= -(1<<31) {
+				i := newInstr(8, []byte{0xC7})
+				if err := setRMDigit(i, 0, dstReg, 8); err != nil {
+					return err
+				}
+				i.imm = le32(v)
+				return e.emit(i)
+			}
 			opBase := byte(0xB8)
 			if size == 1 {
 				opBase = 0xB0
@@ -101,7 +122,7 @@ func (e *enc) encodeMov(ops []Operand, size int) error {
 			if dstReg.needsREX(size) {
 				i.rexForced = true
 			}
-			i.imm = immediate(int64(src), size, true)
+			i.imm = immediate(v, size, true)
 			return e.emit(i)
 		}
 		// MOV r/m, imm: 0xC6 (8-bit) / 0xC7 /0.