Introduce orps for IA32/X64

src/ia32/lithium-codegen-ia32.cc

 // Copyright 2012 the V8 project authors. All rights reserved.
 // Redistribution and use in source and binary forms, with or without
 // modification, are permitted provided that the following conditions are
 // met:
 //
 //     * Redistributions of source code must retain the above copyright
 //       notice, this list of conditions and the following disclaimer.
 //     * Redistributions in binary form must reproduce the above
 //       copyright notice, this list of conditions and the following
 //       disclaimer in the documentation and/or other materials provided
@@ skipping 1940 matching lines @@
           __ pinsrd(res, Operand(temp), 1);
         }
       } else {
         __ Set(temp, Immediate(upper));
         __ movd(res, Operand(temp));
         __ psllq(res, 32);
         if (lower != 0) {
           XMMRegister xmm_scratch = double_scratch0();
           __ Set(temp, Immediate(lower));
           __ movd(xmm_scratch, Operand(temp));
-          __ por(res, xmm_scratch);
+          __ orps(res, xmm_scratch);
         }
       }
     }
   }
 }
 
 
 void LCodeGen::DoConstantE(LConstantE* instr) {
   __ lea(ToRegister(instr->result()), Operand::StaticVariable(instr->value()));
 }
@@ skipping 205 matching lines @@
     } else {
       // Since we operate on +0 and/or -0, addsd and andsd have the same effect.
       __ addsd(left_reg, right_reg);
     }
     __ jmp(&return_left, Label::kNear);
 
     __ bind(&check_nan_left);
     __ 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);
   }
 }
 
 
 void LCodeGen::DoArithmeticD(LArithmeticD* instr) {
   if (CpuFeatures::IsSafeForSnapshot(SSE2)) {
     CpuFeatureScope scope(masm(), SSE2);
     XMMRegister left = ToDoubleRegister(instr->left());
@@ skipping 1790 matching lines @@
   DeoptimizeIf(equal, instr->environment());
   __ jmp(&done);
 
   __ bind(&below_one_half);
   __ movsd(xmm_scratch, Operand::StaticVariable(minus_one_half));
   __ ucomisd(xmm_scratch, input_reg);
   __ j(below_equal, &round_to_zero);
 
   // 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.
   __ cmp(output_reg, 0x80000000u);
   __ RecordComment("D2I conversion overflow");
   DeoptimizeIf(equal, instr->environment());
 
   __ Cvtsi2sd(xmm_scratch, output_reg);
   __ ucomisd(xmm_scratch, input_temp);
   __ j(equal, &done);
@@ skipping 2369 matching lines @@
                               FixedArray::kHeaderSize - kPointerSize));
   __ bind(&done);
 }
 
 
 #undef __
 
 } }  // namespace v8::internal
 
 #endif  // V8_TARGET_ARCH_IA32