Introduce orps for IA32/X64

src/ia32/lithium-codegen-ia32.cc

Index: src/ia32/lithium-codegen-ia32.cc
diff --git a/src/ia32/lithium-codegen-ia32.cc b/src/ia32/lithium-codegen-ia32.cc
index 042a47080895f340c98feabba695b346f1f548a5..ee1995b7773ec584faa61a3738a0df937eaaf365 100644
--- a/src/ia32/lithium-codegen-ia32.cc
+++ b/src/ia32/lithium-codegen-ia32.cc
@@ -1958,7 +1958,7 @@ void LCodeGen::DoConstantD(LConstantD* instr) {
           XMMRegister xmm_scratch = double_scratch0();
           __ Set(temp, Immediate(lower));
           __ movd(xmm_scratch, Operand(temp));
-          __ por(res, xmm_scratch);
+          __ orps(res, xmm_scratch);
         }
       }
     }
@@ -2184,7 +2184,7 @@ void LCodeGen::DoMathMinMax(LMathMinMax* instr) {
     __ ucomisd(left_reg, left_reg);  // NaN check.
     __ j(parity_even, &return_left, Label::kNear);  // left == NaN.
     __ bind(&return_right);
-    __ movsd(left_reg, right_reg);
+    __ movaps(left_reg, right_reg);

[Sven Panne, 2013/11/04 14:05:45]

If I see things correctly, this doesn't really solve the false dependency
problem in general: For even wider registers (Haswell etc.?) we still have
useless dependencies on the upper part. Perhaps this should somehow be handled
in a macro assembler instruction, but we have to be careful about snapshots
then.

Anyway, I don't think it hurts things, so this might be OK for now.

 
     __ bind(&return_left);
   }
@@ -3995,7 +3995,7 @@ void LCodeGen::DoMathRound(LMathRound* instr) {
 
   // CVTTSD2SI rounds towards zero, we use ceil(x - (-0.5)) and then
   // compare and compensate.
-  __ movsd(input_temp, input_reg);  // Do not alter input_reg.
+  __ movaps(input_temp, input_reg);  // Do not alter input_reg.
   __ subsd(input_temp, xmm_scratch);
   __ cvttsd2si(output_reg, Operand(input_temp));
   // Catch minint due to overflow, and to prevent overflow when compensating.