1104199 - Optimize LUnbox codegen on x64

Status: RESOLVED FIXED

(Core :: JavaScript Engine: JIT, defect)

Description

[Jan de Mooij [:jandem]]

Attachment: Patch

The micro-benchmark in bug 624299 is slower on x64 than on x86, apparently due to Unbox:Int32 instructions for the 3 arguments. For each argument we do:

[MoveGroup]
movq       0x68(%rsp), %rax
[Unbox:Int32]
movl       %eax, %eax

This patch changes useRegisterAtStart to useAtStart for infallible unbox instructions, so that we have just:

[Unbox:Int32]
movl       0x68(%rsp), %eax

With this patch the numbers are about the same as with a 32-bit build.

Comment 1

[Dan Gohman [:sunfish]]

Comment on attachment 8527826 [details] [diff] [review]
Patch

Review of attachment 8527826 [details] [diff] [review]:
-----------------------------------------------------------------

::: js/src/jit/x64/CodeGenerator-x64.cpp
@@ +90,2 @@
>      MUnbox *mir = unbox->mir();
> +    LAllocation *input = unbox->getOperand(LUnbox::Input);

This could be moved down along with the initialization of the result variable. And it could to do the ToOperand, instead of having it done on each use.

Comment 2

[Jan de Mooij [:jandem]]

https://hg.mozilla.org/integration/mozilla-inbound/rev/6d64e1e46eb1

(In reply to Dan Gohman [:sunfish] from comment #1)
> This could be moved down along with the initialization of the result
> variable. And it could to do the ToOperand, instead of having it done on
> each use.

Done.

Comment 3

[Phil Ringnalda (:philor)]

https://hg.mozilla.org/mozilla-central/rev/6d64e1e46eb1