531324 - nanojit: improve int32 constant generation in i386 backend

Status: RESOLVED FIXED

(Core :: JavaScript Engine, defect)

Description

[Nicholas Nethercote [inactive]]

Attachment: patch

In some places the i386 back-end will use xor(r,r) to generate a zero constant, which is more compact than loading a zero immediate.  But it doesn't look for this opportunity everywhere it could.

This patch creates a new function for generating constants that always tries to use xor(r,r) if possible.  There's a complication that xor(r,r) clobbers the condition codes, which causes problems in some places, so we allow for that.

The net result is that xor(r,r) is used in more places than before.

Comment 1

[Graydon Hoare :graydon]

Neat. Does it speed anything up? Also curious if the same can be done via a trailing default-false bool arg on the asm_int function, rather than an overload. I don't have a strong style preference, just erring on the side of "fewer functions".

Comment 2

[Nicholas Nethercote [inactive]]

(In reply to comment #1)
> Neat. Does it speed anything up? 

I might be seeing roughly a 3ms speedup for SunSpider -- I've tried it on a couple of revisions -- but it's being too noisy to tell for sure.

> Also curious if the same can be done via a
> trailing default-false bool arg on the asm_int function, rather than an
> overload.

The first asm_int() allocates a register, the second just emits the code, so that doesn't really work.  You could argue that the name shouldn't be reused, I was just following the X64 backend which has a similar overloading of asm_quad().

Are you happy with the patch?

Comment 3

[Graydon Hoare :graydon]

Comment on attachment 414786 [details] [diff] [review]
patch

Yeah, sure. Thanks!

Comment 4

[Nicholas Nethercote [inactive]]

http://hg.mozilla.org/projects/nanojit-central/rev/3b04b5fdd47f

Comment 5

[Nicholas Nethercote [inactive]]

http://hg.mozilla.org/tracemonkey/rev/d4547fa92c0f

Comment 6

[Robert Sayre]

http://hg.mozilla.org/mozilla-central/rev/d4547fa92c0f