mozilla :: #ionmonkey

8 Aug 2017
10:24nbpjandem: https://crash-stats.mozilla.com/signature/?product=Firefox&product=Thunderbird&address=~0xfff&address=%21~0xffffffff&signature=jit%20%7C%20NOT_JIT_CODE&date=%3E%3D2017-02-08T10%3A18%3A32.000Z&date=%3C2017-08-08T10%3A18%3A32.000Z&_columns=date&_columns=product&_columns=version&_columns=build_id&_columns=platform&_columns=reason&_columns=address&_columns=install_time&_sort=-address&_sort=-date&page=1
10:25nbpjandem: if you sort by addresses the other way around, it does not sounds like an AMD bug.
10:26jandemnbp: yeah probably a number of different things :/
10:27nbpjandem: "AuthenticAMD family 20 model 2 stepping 0 | 2" stands out of the group.
10:33jandemyeah that's the infamous one
17:33nbpjandem: r? https://pastebin.mozilla.org/9029146
17:34jandemnbp: r+
17:34jandemnbp: i guess not an issue until now because DEBUG implies JS_GC_ZEAL?
17:35nbpjandem: oh? I never noticed.
17:35tcampbelloh, nice. I ran into that last week
17:38nbpwhat surprize me is that we did not noticed that on treeherder, which makes me wonder if we do not have gc-zeal enabled by default in opt builds.
17:38tcampbellwas building release with --enable-oom-breakpoint
17:44nbprunning valgrind + debug-build + --ion-eager takes ~30 minutes to report a reference error on line 49.
17:59nbpseriously, gcc generates a prolog and epilog for a virtual function which literaly does "mov 0x8c,eax ret" ?!
17:59nbp^ MFoo::op() functions
17:59tcampbellin an opt build?
18:00nbptcampbell: yes
18:02tcampbellnbp: do we have -fomit-frame-pointer off perfhaps?
18:03tcampbellalthough, optimizations should enable be default
18:03nbptcampbell: sounds more than likely, I guess this is something we should remove for a few tiny frames.
18:04nbpI guess the problem might be that profilers won't like the stack they are on, and probably the reason why we have it.
18:04nbpin which case, it sounds like we should at least do it for the tons of tiny accessors that we have in Ion.
18:05tcampbellhttps://searchfox.org/mozilla-central/source/build/autoconf/frameptr.m4#35
18:05tcampbellyou seem to be right.
18:07nbpWe spend roughtly 1/6 of the cost of our register allocator calling the MFoo::op() functions.
18:07nbpSounds like something that should be easy to fix.
18:07nbpI will open a bug and work on it.
18:07tcampbellsorry, whats an example of a MFoo::op(). I'm not following the substitution
18:08nbptcampbell: ins->isConstant()
18:08tcampbellgot it
21:25nbptcampbell: r? ping
21:25tcampbellnbp: i'm going thing it right now. sorry about delay
21:25nbptcampbell: ok ;)
9 Aug 2017
No messages
   
Last message: 69 days and 23 hours ago