Bug 46207 – Speed up function.apply(..., arguments)

• Status: RESOLVED
• Resolution: FIXED
• Component: New Bugs

[Oliver Hunt]

Speed up function.apply(..., arguments)

[Oliver Hunt]

Created attachment 68271 [details]
Patch

[Geoffrey Garen]

Comment on attachment 68271 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=68271&action=review

r=me, but I think you should make these changes.

> JavaScriptCore/jit/JITOpcodes.cpp:1533
> +        emitLoad(argsOffset, regT1, regT0);
> +        slowJumps.append(branchTestPtr(NonZero, regT0));
> +        slowJumps.append(branchTestPtr(NonZero, regT1));

To test for JSValue(), test the tag against JSValue::EmptyValueTag.

> JavaScriptCore/jit/JITOpcodes.cpp:1541
> +        slowJumps.append(branch32(Above, regT0, Imm32(Arguments::MaxArguments)));

Do you need this check on the hot path? In theory, the only limit to argument use on the hot path is the size of the register file, which you check below.

> JavaScriptCore/jit/JITOpcodes.cpp:1560
> +        if (sizeof(Register) == 2 * ScalePtr) {
> +            loadPtr(BaseIndex(regT1, regT0, ScalePtr, static_cast<unsigned>(-sizeof(Register) + sizeof(void*))), regT3);
> +            storePtr(regT3, BaseIndex(regT2, regT0, ScalePtr, sizeof(void*)));
> +        }

Might be clearer just to use #if USE(JSVALUE32_64) here instead.

> JavaScriptCore/jit/JITOpcodes.cpp:1561
> +        jump(loopStart);

This will "jump to jump" each time through the loop. Might be faster to do "branchTest32(Zero, regT0)" here, branching back to the sub32 above.

[Oliver Hunt]

Created attachment 68294 [details]
Patch

[Geoffrey Garen]

Comment on attachment 68294 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=68294&action=review

r=me

> JavaScriptCore/jit/JITOpcodes.cpp:1572
> +    JumpList slowJumps;
> +    JumpList endBranches;

These variables are unused.

[Oliver Hunt]

Committed r67990: <http://trac.webkit.org/changeset/67990>