MIPS: Inline Math.fround in optimized code.

src/mips64/lithium-codegen-mips64.cc

Index: src/mips64/lithium-codegen-mips64.cc
diff --git a/src/mips64/lithium-codegen-mips64.cc b/src/mips64/lithium-codegen-mips64.cc
index 4a625610d879295c86a0dbd00ab0bb0462c48483..7f98a62a72506040ddcecca03b00e59a8c44fcbb 100644
--- a/src/mips64/lithium-codegen-mips64.cc
+++ b/src/mips64/lithium-codegen-mips64.cc
@@ -3854,6 +3854,14 @@ void LCodeGen::DoMathRound(LMathRound* instr) {
 }
 
 
+void LCodeGen::DoMathFround(LMathFround* instr) {
+  DoubleRegister input = ToDoubleRegister(instr->value());
+  DoubleRegister result = ToDoubleRegister(instr->result());
+  __ cvt_s_d(result.low(), input);
+  __ cvt_d_s(result, result.low());

[paul.l..., 2014/07/29 16:54:26]

Please remove use of result.low() in both instructions, and just use 'result'.
In fp64 mode, we pass full 64-bit FP regs that can be used as target of either
single or double-precision ops. We no longer use pairs, and should then no
longer use .low() or .high(). (They will be removed in a future mips64 cleanup
CL)

[balazs.kilvady, 2014/07/29 18:36:11]

Done.

+}
+
+
 void LCodeGen::DoMathSqrt(LMathSqrt* instr) {
   DoubleRegister input = ToDoubleRegister(instr->value());
   DoubleRegister result = ToDoubleRegister(instr->result());