Improve the assembly code for power function with integer exponential on x64

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 x64 for exponetials from about 100 to 2000.

Review URL: https://chromiumcodereview.appspot.com/10939013
Patch from Xi Qian <xi.qian@intel.com>.

git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12535 ce2b1a6d-e550-0410-aec6-3dcde31c8c00

diff --git a/src/ia32/code-stubs-ia32.cc b/src/ia32/code-stubs-ia32.cc
index 1fdbaf11fd31ccdd4362e9ebe3500c18866b4565..96d45d319eeb8c4e13870e3b7ee636a9a67648a8 100644 (file)
--- a/src/ia32/code-stubs-ia32.cc
+++ b/src/ia32/code-stubs-ia32.cc
@@ -3213,7 +3213,7 @@ 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, while_false;
+  Label no_neg, while_true, while_false;
   __ test(scratch, scratch);
   __ j(positive, &no_neg, Label::kNear);
   __ neg(scratch);

diff --git a/src/x64/code-stubs-x64.cc b/src/x64/code-stubs-x64.cc
index 62a89b83b294efa1d747fb30a81cacfa51f10fb0..c5025cf81c3c878e9d4d260ccae310688a3ba7a3 100644 (file)
--- a/src/x64/code-stubs-x64.cc
+++ b/src/x64/code-stubs-x64.cc
@@ -2262,21 +2262,28 @@ 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, while_false;
   __ testl(scratch, scratch);
   __ j(positive, &no_neg, Label::kNear);
   __ negl(scratch);
   __ bind(&no_neg);
 
-  __ bind(&while_true);
+  __ j(zero, &while_false, Label::kNear);
   __ shrl(scratch, Immediate(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);
+  __ shrl(scratch, Immediate(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);
   // If the exponent is negative, return 1/result.
   __ testl(exponent, exponent);
   __ j(greater, &done);