952286 - AWFY: 18% regression on x86 asmjs-ubench-skinning

Status: RESOLVED INVALID

(Core :: JavaScript Engine: JIT, defect)

Description

[Luke Wagner [:luke]]

The regression is only on 32-bit range is:
  http://hg.mozilla.org/integration/mozilla-inbound/pushloghtml?fromchange=7450803f061c&tochange=afe7a6c1c9b2
which seems to suggest bug 950703.

Comment 1

[Dan Gohman [:sunfish]]

I'm unable to reproduce this so far. IONFLAGS=codegen output between a2a5e25445e5 and afe7a6c2c9b2 contains identical instructions on 32-bit x86 Linux.

Comment 2

[Hannes Verschore [:h4writer]]

(In reply to Dan Gohman [:sunfish] from comment #1)
> I'm unable to reproduce this so far. IONFLAGS=codegen output between
> a2a5e25445e5 and afe7a6c2c9b2 contains identical instructions on 32-bit x86
> Linux.

You did run with --no-asmjs right?

Comment 3

[Dan Gohman [:sunfish]]

(In reply to Hannes Verschore [:h4writer] from comment #2)
> (In reply to Dan Gohman [:sunfish] from comment #1)
> > I'm unable to reproduce this so far. IONFLAGS=codegen output between
> > a2a5e25445e5 and afe7a6c2c9b2 contains identical instructions on 32-bit x86
> > Linux.
> 
> You did run with --no-asmjs right?

No, I didn't think this bug was about --no-asmjs mode. But now I have, and I don't see a codegen difference between those revisions with --no-asmjs either.

Comment 4

[Luke Wagner [:luke]]

I should have been more explicit, but this was an Odin (--asmjs) regression; --no-asmjs was unaffected.

Comment 5

[Luke Wagner [:luke]]

I am not able to reproduce the regression locally either.  Looking again, GGC and the default build basically switched scores starting with the cset range in question.  Normally, GGC and the default build get the exact same score on asm.js code since practically 100% of the time is in generated code.  This all strongly smells like some artifact of code/data alignment or caching or something else not at all related to Dan's patches.  Thus, I think we can resolve this particular bug INVALID unless someone thinks something else is afoot.  Sorry for the noise Dan!