126204 - segfault on start -funroll-loops

Status: RESOLVED WONTFIX

(Core :: XPCOM, defect)

Description

[Ervin Németh]

Mozilla makes a segfault on startup.  The 2002-02-15 4:00 (GMT) build was ok,
but the 2002-02-18 4:00 (and after) build makes a segfault.  TestXPTCInvoke is
still working.

gcc-3.0.3 CFLAGS="-O3 -march=i686 -ffast-math -funroll-loops
-maccumulate-outgoing-args" LDFLAGS="-Wl,-O1,-z,combreloc"

Deleting ~/.mozilla doesn't help, either.

No trace available (yet), it is a non-debug build.

Comment 1

[Doug Turner (:dougt)]

try rebuilding with default options.  see http://www.mozilla.org/build/unix.html
for details.

Comment 2

[Ervin Németh]

I am reporting a successfull build using CFLAGS="-O2 -march=i686"

I was there as "-fomit-frame-pointer" started to cause a wrong build, now I am
here as something breaks an other flag.  As I have time I will research which
flag is unusable this time.

Now some philosophical fuming.  My Mozilla build is now 144 Megabytes, mainly of
C and C++ code.  Then somebody comes and does some assembly wizardry which makes
a single object-file a little bit faster, but at the same time prevents the
other 143M of code from being optimized making the program slower.  Call me a
freak but I still believe that code optimization is cheaper than buying a faster
processor.

Should we close this bug when I've figured out the compiler-flag or does
somebody agree with me?

Comment 3

[Doug Turner (:dougt)]

I agree with you in theory.  although, I do not totally grok the need for it.  

I think that there are bugs with this -fomit-frame-pointer discussion.  Find
those discussions and lets close out this bug.

Comment 4

[Ervin Németh]

Finally I've got it.  It is the "-funroll-loops" flag.  With gcc-3.0.4, too.

I'm really sad because this flag could give a good performance boost.

Doug, should we mark this bug WONTFIX or LATER?  :)

Comment 5

[Doug Turner (:dougt)]

post a message to the builds newsgroup regarding this.  Maybe someone would know
how or want to try to fix this.  As for the bug, afaiac, WONTFIX.  

Comment 6

[timeless]

fwiw later should not be used. and if you intend to get some help with this bug please find out precisely which file is causing the problem

--enable-optimize="-O1 -g", crash, get a stack trace, figure out which file doesn't work with  -funroll-loops, reuild that file without the flag, rinse and repeat until mozilla runs, then provide the list here (or try to isolate the affected functions). when you get stuck, as for help.

Comment 7

[Ferdinand]

'-O2 -march=i686 -funroll-loops' works ok with the gcc version included in
redhat 7.2

Comment 8

[Ervin Németh]

gcc-3.1 seems not to be affected either.  Try out the new MMX and SSE
optimization, too.