Improve the assembly code for power function with integer exponential.

The change removes one unused multiply and reschedules
the shift, multiply and jump instructions to reduce
stall. Experiment shows it improve about 20% performance
on x86 for exponetials from about 100 to 2000.

Review URL: https://chromiumcodereview.appspot.com/10916311

git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12521 ce2b1a6d-e550-0410-aec6-3dcde31c8c00

diff --git a/src/ia32/code-stubs-ia32.cc b/src/ia32/code-stubs-ia32.cc
index d75d5cd..1fdbaf1 100644 (file)
--- a/src/ia32/code-stubs-ia32.cc
+++ b/src/ia32/code-stubs-ia32.cc
@@ -3213,21 +3213,28 @@ void MathPowStub::Generate(MacroAssembler* masm) {
   __ movsd(double_scratch2, double_result);  // Load double_exponent with 1.
 
   // Get absolute value of exponent.
-  Label no_neg, while_true, no_multiply;
+  Label no_neg, while_true, no_multiply, while_false;
   __ test(scratch, scratch);
   __ j(positive, &no_neg, Label::kNear);
   __ neg(scratch);
   __ bind(&no_neg);
 
-  __ bind(&while_true);
+  __ j(zero, &while_false, Label::kNear);
   __ shr(scratch, 1);
-  __ j(not_carry, &no_multiply, Label::kNear);
-  __ mulsd(double_result, double_scratch);
-  __ bind(&no_multiply);
+  // Above condition means CF==0 && ZF==0.  This means that the
+  // bit that has been shifted out is 0 and the result is not 0.
+  __ j(above, &while_true, Label::kNear);
+  __ movsd(double_result, double_scratch);
+  __ j(zero, &while_false, Label::kNear);
 
+  __ bind(&while_true);
+  __ shr(scratch, 1);
   __ mulsd(double_scratch, double_scratch);
+  __ j(above, &while_true, Label::kNear);
+  __ mulsd(double_result, double_scratch);
   __ j(not_zero, &while_true);
 
+  __ bind(&while_false);
   // scratch has the original value of the exponent - if the exponent is
   // negative, return 1/result.
   __ test(exponent, exponent);