516348 - nanojit: add findSpecificRegForUnallocated()

Status: RESOLVED FIXED

(Core :: JavaScript Engine, defect)

Description

[Nicholas Nethercote [:njn] (On Long Service Leave)]

Attachment: patch

This patch:
- Adds findSpecificRegForUnallocated(), which is like findSpecificRegFor() but
  requires that the instruction under consideration not have a register
  allocated to it, a case which occurs surprisingly often.
- Applies it throughout the i386 back-end.  
- Adds RegAlloc::removeFree().

I originally did this change just for comprehensibility's sake -- ie.
findSpecificRegForUnallocated() is *much* simpler than findSpecificRegFor()
-- but it also looks like it improves SunSpider times by about 2ms.  I'm not
sure I trust that, but it's certainly fewer instructions executed so can't
hurt.

Comment 1

[Nicholas Nethercote [:njn] (On Long Service Leave)]

http://hg.mozilla.org/projects/nanojit-central/rev/0e40a8977b80

Comment 2

[Graydon Hoare :graydon]

http://hg.mozilla.org/tracemonkey/rev/6b96e43a6dd3

Comment 3

[Robert Sayre]

http://hg.mozilla.org/tracemonkey/rev/6b96e43a6dd3

Comment 4

[Robert Sayre]

http://hg.mozilla.org/mozilla-central/rev/6b96e43a6dd3

Comment 5

[walter sheets]

Hi guys.  My x86 machine is consistently hitting this added nanoassert:
NanoAssert(ins->isUnusedOrHasUnknownReg());  (Assembler.cpp:432)

OTOH, my amd64 machine works normally.  I'm wondering if this could be a compiler bug, because I'm seeing this problem in my own builds (gcc-4.3.4) but not in the nightlies from mozilla.org (gcc-4.1.1, I believe).

Any thoughts?  I can post the backtrace if it would be useful.

Thanks.

Comment 6

[Nicholas Nethercote [:njn] (On Long Service Leave)]

Walter, 34692:515d7e584de6 is a follow-up changeset that fixes the assertion problem.

Comment 7

[walter sheets]

Thanks for the quick reply.  I just pulled again from mozilla-central and I don't see that commit anywhere.  (I searched 'hg log' for 515d7 and got nada, and
"tip" is now at 34912:168606de059d)

Am I looking in the wrong branch of the hg repository?  Or the wrong repository?

Thanks.

Comment 8

[Nicholas Nethercote [:njn] (On Long Service Leave)]

It's in the tracemonkey repo, it hasn't made it to the mozilla-central repo yet.  The relevant bug is bug 527874.