[x64] Avoid the sign extension in the DoubleToI conversion.

We use the cvttsd2si (float64-to-int32) instruction to avoid that negative numbers get sign extended in the destination register.

@rodolph: Is sign extension also a thing on arm64? I noticed that also on arm64 a float64-to-int64 instruction is used.

Review URL: https://codereview.chromium.org/1877133004

Cr-Commit-Position: refs/heads/master@{#35466}

src/compiler/x64/code-generator-x64.cc

@@ -791,8 +791,8 @@ void CodeGenerator::AssembleArchInstruction(Instruction* instr) {
       auto result = i.OutputRegister();
       auto input = i.InputDoubleRegister(0);
       auto ool = new (zone()) OutOfLineTruncateDoubleToI(this, result, input);
-      __ Cvttsd2siq(result, input);
-      __ cmpq(result, Immediate(1));
+      __ Cvttsd2si(result, input);
+      __ cmpl(result, Immediate(1));
       __ j(overflow, ool->entry());
       __ bind(ool->exit());
       break;

src/x64/macro-assembler-x64.cc

@@ -3744,8 +3744,8 @@ void MacroAssembler::TruncateHeapNumberToI(Register result_reg,
                                            Register input_reg) {
   Label done;
   Movsd(xmm0, FieldOperand(input_reg, HeapNumber::kValueOffset));
-  Cvttsd2siq(result_reg, xmm0);
-  cmpq(result_reg, Immediate(1));
+  Cvttsd2si(result_reg, xmm0);
+  cmpl(result_reg, Immediate(1));
   j(no_overflow, &done, Label::kNear);
 
   // Slow case.
@@ -3767,8 +3767,8 @@ void MacroAssembler::TruncateHeapNumberToI(Register result_reg,
 void MacroAssembler::TruncateDoubleToI(Register result_reg,
                                        XMMRegister input_reg) {
   Label done;
-  Cvttsd2siq(result_reg, input_reg);
-  cmpq(result_reg, Immediate(1));
+  Cvttsd2si(result_reg, input_reg);
+  cmpl(result_reg, Immediate(1));
   j(no_overflow, &done, Label::kNear);
 
   subp(rsp, Immediate(kDoubleSize));