optimization/2917: mozilla doesn't run with "-O2 -march=pentiumpro"

[dbaron@fas.harvard.edu]

>Number:         2917
>Category:       optimization
>Synopsis:       mozilla doesn't run with "-O2 -march=pentiumpro"
>Confidential:   no
>Severity:       serious
>Priority:       medium
>Responsible:    unassigned
>State:          open
>Class:          wrong-code
>Submitter-Id:   net
>Arrival-Date:   Wed May 23 19:26:00 PDT 2001
>Closed-Date:
>Last-Modified:
>Originator:     dbaron@fas.harvard.edu
>Release:        gcc 3.0 prerelease, various dates (see below)
>Organization:
>Environment:
RedHat 7.0+errata, dual P-733.
>Description:
See http://bugzilla.mozilla.org/show_bug.cgi?id=76604 for more information.

Sorry this bug doesn't have all the details it should have -- it may be very difficult to reduce down to something simple and I don't have time to reduce it.  (I haven't even had time to narrow down further when it started and which optimization levels it happens on, or keep up with testing that it still occurs on current snapshots.)

What I do know is this:
On gcc 3.0 prerelease snapshots from March 12 and earlier, mozilla compiled both without optimization (and with -g) and with optimization options "-O2 -march=pentiumpro" (and without -g).

On gcc 3.0 snapshots from March 26, April 2, April 9, April 16, and I think I tested April 30 (I didn't get a chance to try newer snapshots yet, although I'm building the May 21 snapshot now), the build without optimization still works but the build with optimization crashes.

I've found some slight changes to code that fix the crash, that are described in the URL above (the bugzilla bug report), but I'm not sure how helpful that is.

At one point I looked at the differences between the generated assembly between the March 12 and March 26 compilers, and some of the major differences seemed to be related to changes allowing inlining of static functions (changes that I think were made due to Mozilla's extensive use of them for retrieving IIDs).

Sorry for the lack of information here, but I decided to just file this bug rather than keep putting it off for lack of information that I'm only slowly finding the time to gather.  I really would like to see Mozilla be able to build with gcc3.
>How-To-Repeat:

>Fix:

>Release-Note:
>Audit-Trail:
>Unformatted: