Bug 411639 turns this off for image decoders, pixman, xptcall, and cairo
Correction, that's bug 411369. I'm filing this so I have something to point FIXME comments at.
Now that libpng no longer includes any assembly code, let alone MMX, can the -GL- work-around be removed from the build? ~/modules/libimg/png/Makefile.in: ifeq ($(OS_ARCH),WINNT) # FIXME: bug 413019 OS_COMPILE_CFLAGS += -GL- endif
Also, from looking at the cairo/libpixman code in Gecko v18.104.22.168, it seems that there's no longer any MMX in assembly there either. The remaining ASM code is either GCC-specific or just used to get the x86 feature flags.
We could certainly try it, write a patch?
Created attachment 389828 [details] [diff] [review] Partial removal of PGO breakage work-around This patch removes the -GL- work-around in the Cairo and PNG libraries. I saw no adverse reactions to this in a PGO build of Firefox v3.5.1 (on a Pentium4-based Win2K/SP4 system). The work-around is left intact in the libpixman library because removal caused crashes on Firefix start-up.
Comment on attachment 389828 [details] [diff] [review] Partial removal of PGO breakage work-around If you could run your PGO build through the unittests I'd feel even better about this, but it's definitely worth a shot.
Is this fix ready to check in?
I'd say it could stand a pass by the try server, but I'm pretty sure that doesn't do PGO builds. We have unittests on PGO builds on mozilla-central, so I think it should be pretty safe to checkin there. If it causes any bustage we can back it out.
Dietrich, possible perf win here, but needs testing. Maybe via the Places branch?
Created attachment 444630 [details] [diff] [review] Further reduction in scope of work-around Restored PGO optimization to all libpixman sources files except pixman-mmx.c and pixman-sse2.c. Only these 2 files are adversely affected by VS2005/SP1 PGO. This patch applies cleanly to Firefox v3.6.4 Release Candidate Build #3 Tested under WinXP/SP3 on Pentium3 and Core2 Duo CPUs (to exercise the MMX and SSE2 code paths). As expected, the complete removal of the PGO disabling causes the browser to crash on start-up. The crashing is cured on on a given system by disabling PGO for the SIMD code path that that system uses at runtime. The other 24 source files can safely be optimized. It is possible that the disabling of PGO is also required for source file pixman-vmx.c, but I can't test that.
In the time since comment 8 was made, the try server now does PGO builds, so a spin past the try server should be sufficient for testing.
Created attachment 543049 [details] [diff] [review] fix
passed on try server http://tbpl.mozilla.org/?tree=Try&rev=b60bd9b73fd0
Comment on attachment 543049 [details] [diff] [review] fix Now this is passed on try server. Should I retry this after changed to VS2010 (after next merge cycle)? Also, I remove PGO workaround into jpeg. Because now, we switch to libjepg-turbo. (Original libjpeg + our SSE2 code had inline assembler.)
Comment on attachment 543049 [details] [diff] [review] fix Review of attachment 543049 [details] [diff] [review]: ----------------------------------------------------------------- This seems fine. I wouldn't worry about re-testing with Visual C++ 2010, we should be able to see if our crash-stats change when we switch.
Mozilla/5.0 (Windows NT 5.1; rv:7.0) Gecko/20100101 Firefox/7.0 I visually inspected the code changes. Marking as Verified Fixed as per bug 659997 Comment 7.