AuroraMiddleware/v8
1
0
Fork 0
Code Issues 2 Pull Requests Releases Wiki Activity

Fix BinaryOpIC implementation on ARM.

Browse Source 
On a pair of smis HEAP_NUMBERS stub is significantly slower than GENERIC. This slows
down some tests dramatically (crypto-aes from SunSpider).
With this change HEAP_NUMBERS stub switches to GENERIC stub the first time it sees 2 smis
as its operands.

Review URL: http://codereview.chromium.org/1687005

git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@4483 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
This commit is contained in:
kaznacheev@chromium.org 2010-04-23 10:08:24 +00:00
parent 6f34ff44e3
commit 3ca99e722b
1 changed files with 11 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
src/arm/codegen-arm.cc
View File

@ -6468,11 +6468,22 @@ void GenericBinaryOpStub::HandleBinaryOpSlowCases(
__ pop(lr);
}
// HEAP_NUMBERS stub is slower than GENERIC on a pair of smis.
// r0 is known to be a smi. If r1 is also a smi then switch to GENERIC.
Label r1_is_not_smi;
if (runtime_operands_type_ == BinaryOpIC::HEAP_NUMBERS) {
__ tst(r1, Operand(kSmiTagMask));
__ b(ne, &r1_is_not_smi);
GenerateTypeTransition(masm);
__ jmp(&r1_is_smi);
}
__ bind(&finished_loading_r0);
// Move r1 to a double in r0-r1.
__ tst(r1, Operand(kSmiTagMask));
__ b(eq, &r1_is_smi); // It's a Smi so don't check it's a heap number.
__ bind(&r1_is_not_smi);
__ CompareObjectType(r1, r4, r4, HEAP_NUMBER_TYPE);
__ b(ne, &slow);
if (mode_ == OVERWRITE_LEFT) {