Fixing MathPowHalf on ARM.

src/arm/lithium-codegen-arm.cc

Index: src/arm/lithium-codegen-arm.cc
diff --git a/src/arm/lithium-codegen-arm.cc b/src/arm/lithium-codegen-arm.cc
index ac10793b46f56e7d088227233861c8e2aec06387..35dcc17b0da3ed6e6dc47a1f162598abe3dd53d0 100644
--- a/src/arm/lithium-codegen-arm.cc
+++ b/src/arm/lithium-codegen-arm.cc
@@ -3097,9 +3097,20 @@ void LCodeGen::DoMathSqrt(LUnaryMathOperation* instr) {
 void LCodeGen::DoMathPowHalf(LUnaryMathOperation* instr) {
   DoubleRegister input = ToDoubleRegister(instr->InputAt(0));
   DoubleRegister result = ToDoubleRegister(instr->result());
+
+  // Note that according to ECMA-262 15.8.2.13:
+  // Math.pow(-Infinity, 0.5) == Infinity
+  // Math.sqrt(-Infinity) == NaN
+  Label done;
+  __ vmov(result, -INFINITY);
+  __ vcmp(input, result);

[Rodolph Perfetta, 2011/12/05 17:38:36]

Confusingly on ARM vcmp does no set the flags! you need a vmrs into pc
afterwards.
There is a macro in the macro-assembler for this:
VPCompareAndSetFlags

+  __ vneg(result, input, eq);
+  __ b(&done, eq);
+
   // Add +0 to convert -0 to +0.
   __ vadd(result, input, kDoubleRegZero);
   __ vsqrt(result, result);
+  __ bind(&done);
 }