wazevo(arm64): fixes VBitselect translation (#1718)

Signed-off-by: Takeshi Yoneda <t.y.mathetake@gmail.com>

internal/engine/wazevo/backend/isa/arm64/lower_instr.go

@@ -320,18 +320,18 @@ func (m *machine) LowerInstr(instr *ssa.Instruction) {
 		m.lowerVecRRR(vecOpBic, x, y, instr.Return(), vecArrangement16B)
 	case ssa.OpcodeVbitselect:
 		c, x, y := instr.SelectData()
-		ins := m.allocateInstr()
 		rn := m.getOperand_NR(m.compiler.ValueDefinition(x), extModeNone)
 		rm := m.getOperand_NR(m.compiler.ValueDefinition(y), extModeNone)
 		creg := m.getOperand_NR(m.compiler.ValueDefinition(c), extModeNone)
-		ins.asVecRRR(vecOpBsl, creg, rn, rm, vecArrangement16B)
-		m.insert(ins)
-
+		// creg is overwritten by BSL, so we need to move it to the result register before the instruction
+		// in case when it is used somewhere else.
 		rd := m.compiler.VRegOf(instr.Return())
 		mov := m.allocateInstr()
 		mov.asFpuMov128(rd, creg.nr())
-
 		m.insert(mov)
+		ins := m.allocateInstr()
+		ins.asVecRRR(vecOpBsl, operandNR(rd), rn, rm, vecArrangement16B)
+		m.insert(ins)
 	case ssa.OpcodeVIadd:
 		x, y, lane := instr.Arg2WithLane()
 		arr := ssaLaneToArrangement(lane)

internal/engine/wazevo/e2e_test.go

@@ -261,7 +261,6 @@ func TestE2E(t *testing.T) {
 				{params: []uint64{uint64(wasm.MemoryPageSize) - 3, 0}, expErr: "out of bounds memory access"},
 			},
 		},
-
 		{
 			name: "memory_load_basic",
 			m:    testcases.MemoryLoadBasic.Module,
@@ -379,6 +378,13 @@ func TestE2E(t *testing.T) {
 				{params: []uint64{1, 200}, expResults: []uint64{1}},
 			},
 		},
+		{
+			name: "vector_bit_select",
+			m:    testcases.VecBitSelect.Module,
+			calls: []callCase{
+				{params: []uint64{1, 2, 3, 4, 5, 6}, expResults: []uint64{0x3, 0x2, 0x5, 0x6}},
+			},
+		},
 	} {
 		tc := tc
 		t.Run(tc.name, func(t *testing.T) {

internal/engine/wazevo/testcases/testcases.go

@@ -1512,6 +1512,26 @@ var (
 			}}},
 		},
 	}
+
+	VecBitSelect = TestCase{
+		Name: "vector_bit_select",
+		Module: &wasm.Module{
+			TypeSection:     []wasm.FunctionType{{Params: []wasm.ValueType{v128, v128, v128}, Results: []wasm.ValueType{v128, v128}}},
+			ExportSection:   []wasm.Export{{Name: ExportedFunctionName, Type: wasm.ExternTypeFunc, Index: 0}},
+			FunctionSection: []wasm.Index{0},
+			CodeSection: []wasm.Code{{Body: []byte{
+				// In arm64, the "C" of BSL instruction is overwritten by the result,
+				// so this case can be used to ensure that it is still alive if 'c' is used somewhere else,
+				// which in this case is as a return value.
+				wasm.OpcodeLocalGet, 0,
+				wasm.OpcodeLocalGet, 1,
+				wasm.OpcodeLocalGet, 2,
+				wasm.OpcodeVecPrefix, wasm.OpcodeVecV128Bitselect,
+				wasm.OpcodeLocalGet, 2, // Returns the 'c' as-is.
+				wasm.OpcodeEnd,
+			}}},
+		},
+	}
 )
 
 type TestCase struct {
@@ -1546,10 +1566,11 @@ var (
 )
 
 const (
-	i32 = wasm.ValueTypeI32
-	i64 = wasm.ValueTypeI64
-	f32 = wasm.ValueTypeF32
-	f64 = wasm.ValueTypeF64
+	i32  = wasm.ValueTypeI32
+	i64  = wasm.ValueTypeI64
+	f32  = wasm.ValueTypeF32
+	f64  = wasm.ValueTypeF64
+	v128 = wasm.ValueTypeV128
 
 	blockSignature_vv = 0x40 // 0x40 is the v_v signature in 33-bit signed. See wasm.DecodeBlockType.
 )