Factor out i64x2_mul

titzer · Jul 22, 2023 · 35b631f · 35b631f
1 parent ca4a8b8
commit 35b631f
Show file tree

Hide file tree

Showing 2 changed files with 2 additions and 16 deletions.
diff --git a/src/engine/x86-64/X86_64Interpreter.v3 b/src/engine/x86-64/X86_64Interpreter.v3
@@ -2423,6 +2423,7 @@ class X86_64InterpreterGen(ic: X86_64InterpreterCode, w: DataWriter) {
 		genSimdBinop(Opcode.I16X8_MUL, asm.pmullw_s_s);
 		genSimdBinop(Opcode.I32X4_MUL, asm.pmulld_s_s);
 		genSimdBinop(Opcode.I32X4_DOT_I16X8_S, asm.pmaddwd_s_s);
+		genSimdBinop(Opcode.I64X2_MUL, masm.emit_i64x2_mul(_, _, r_xmm2, r_xmm3));
 		genSimdBinop(Opcode.F32X4_ADD, asm.addps_s_s);
 		genSimdBinop(Opcode.F64X2_ADD, asm.addpd_s_s);
 		genSimdBinop(Opcode.F32X4_SUB, asm.subps_s_s);
@@ -2491,14 +2492,6 @@ class X86_64InterpreterGen(ic: X86_64InterpreterCode, w: DataWriter) {
 		genSimdBinop(Opcode.F32X4_MAX, masm.emit_f32x4_max(_, _, r_xmm2));
 		genSimdBinop(Opcode.F64X2_MAX, masm.emit_f64x2_max(_, _, r_xmm2));
 
-		bindHandler(Opcode.I64X2_MUL); {
-			asm.movdqu_s_m(r_xmm0, vsph[-2].value); // lhs, dst
-			asm.movdqu_s_m(r_xmm1, vsph[-1].value); // rhs
-			masm.emit_i64x2_mul(r_xmm0, r_xmm1, r_xmm2, r_xmm3);
-			asm.movdqu_m_s(vsph[-2].value, r_xmm0);
-			decrementVsp();
-			endHandler();
-		}
 
 		genSimdBinop(Opcode.I8X16_GT_S, asm.pcmpgtb_s_s);
 		genSimdBinop(Opcode.I16X8_GT_S, asm.pcmpgtw_s_s);

diff --git a/src/engine/x86-64/X86_64SinglePassCompiler.v3 b/src/engine/x86-64/X86_64SinglePassCompiler.v3
@@ -507,14 +507,7 @@ class X86_64SinglePassCompiler extends SinglePassCompiler {
 
 	def visit_I64X2_ADD() { do_op2_x_x(ValueKind.V128, asm.paddq_s_s); }
 	def visit_I64X2_SUB() { do_op2_x_x(ValueKind.V128, asm.psubq_s_s); }
-	def visit_I64X2_MUL() {
-		var b = popReg();
-		var a = popRegToOverwrite();
-		var t1 = allocTmp(ValueKind.V128);
-		var t2 = allocTmp(ValueKind.V128);
-		mmasm.emit_i64x2_mul(X(a.reg), X(b.reg), X(t1), X(t2));
-		state.push(a.kindFlagsMatching(ValueKind.V128, IN_REG), a.reg, 0);
-	}
+	def visit_I64X2_MUL() { do_op2_x_x(ValueKind.V128, mmasm.emit_i64x2_mul(_, _, X(allocTmp(ValueKind.V128)), X(allocTmp(ValueKind.V128)))); }
 	def visit_I64X2_NEG() { visit_V128_I_NEG(mmasm.emit_i64x2_neg); }
 	def visit_I64X2_EQ() { do_op2_x_x(ValueKind.V128, asm.pcmpeqq_s_s); }
 	def visit_I64X2_NE() { do_op2_x_x(ValueKind.V128, mmasm.emit_i64x2_ne); }