554364 - use -fomit-frame-pointer on Linux

Status: RESOLVED FIXED

(Toolkit :: Crash Reporting, defect, P1)

Description

[Jim Blandy :jimb]

+++ This bug was initially created as a clone of Bug #492688 +++

We get a 5.5% (50ms) speedup with gcc on MacOSX when using -O3 and -fomit-frame-pointer instead of -Os. Peak speedup is 16% for md5. Everything seems to get faster. spectral-norm is probably a measurement outlier.

We should do this on Linux as well.

Comment 1

[Justin Dolske [:Dolske]]

FWIW, the last compiler option tuning I remember being done for Linux was in bug 409803...

Comment 2

[[e-mail]]

Yeah, linux is different, in weird and wonderful ways, and needs to be tested separately. But hopefully gcc is saner than before, and the gains translate as hoped. :)

Comment 3

[(not currently active) Ted Mielczarek]

Note that we're using GCC 4.3.3 nowadays, to start.

Comment 4

[(not currently active) Ted Mielczarek]

Scoping this down slightly, someone else can fiddle with the optimization flags further if they want.

Comment 5

[(not currently active) Ted Mielczarek]

Attachment: add -fomit-frame-pointer to MOZ_OPTIMIZE_FLAGS [Checkin: Comment 8]

Comment 6

[Jim Blandy :jimb]

Okay. Let's give it a spin.

Comment 7

[Jim Blandy :jimb]

For what it's worth, here are some directories that set MODULE_OPTIMIZE_FLAGS that could enable -fomit-frame-pointer:

./db/sqlite3/src/Makefile.in
./gfx/cairo/cairo/src/Makefile.in
./gfx/cairo/libpixman/src/Makefile.in
./memory/jemalloc/Makefile.in
./netwerk/protocol/ftp/src/Makefile.in
./xpcom/io/Makefile.in

Comment 8

[(not currently active) Ted Mielczarek]

Pushed to m-c:
http://hg.mozilla.org/mozilla-central/rev/6c003a9165b9

Yeah, we can twiddle those in a followup if we want to. Might be a little more perf to be gained. The original bug was more concerned with JS perf, which we've covered here.

Comment 9

[Serge Gautherie (:sgautherie)]

Attachment: (Bv1-CC) Copy it to comm-central [Checkin: Comment 12]

Comment 10

[Serge Gautherie (:sgautherie)]

(In reply to comment #7)
> For what it's worth, here are some directories that set MODULE_OPTIMIZE_FLAGS
> that could enable -fomit-frame-pointer:

Fwiw,
http://mxr.mozilla.org/mozilla-central/search?string=MODULE_OPTIMIZE_FLAGS&case=on

Comment 11

[Justin Wood (:Callek)]

Comment on attachment 435399 [details] [diff] [review]
(Bv1-CC) Copy it to comm-central
[Checkin: Comment 12]

fyi, c-c has none of the MOZ_OPTIMIZE_FLAGS override problems [except in the backport of sqlite in our top level; but I'll leave that to Mail team to care about]

Comment 12

[Serge Gautherie (:sgautherie)]

Comment on attachment 435399 [details] [diff] [review]
(Bv1-CC) Copy it to comm-central
[Checkin: Comment 12]


http://hg.mozilla.org/comm-central/rev/cbe96b862e08

Comment 13

[Gábor Stefanik]

Out of curiosity, why are we using -Os (optimize for code size), rather than e.g. -O2 (optimize for execution speed)? Does -Os actually run faster than the real speed optimization, or is code size still more important than speed on today's high-bandwidth connections?

Comment 14

[(not currently active) Ted Mielczarek]

Previous tests showed that using -Os for most of our codebase actually produced code that ran faster. It's possible that it has to do with cache sizes, etc. We are compiling our JavaScript library at -O3, since the optimizations are a definite win there:
http://mxr.mozilla.org/mozilla-central/source/js/src/Makefile.in#91