Resend in plain text mode.
On Wed, Jul 27, 2016 at 9:07 AM, Xinliang David Li <davi...@google.com> wrote: > Our experience is that non-atomic counter update (the current > implementation) rarely result in corrupted profile (in heavily threaded > environment) -- it usually results in some profile insanity which can be > corrected with -fprofile-correction -- otherwise we would have been forced > to go the TLS route. > > The profile corruption usually happens when server program crashed during > dumping (server program usually do not exit, so forcing it to exit to dump > profile can cause problem). Our solution is to introduce profile runtime API > to be invoked by the user. > > We added the atomic support mostly for linux kernel FDO. The support is in > google/gcc_49 branch. The option description is: > > ; fprofile-generate-atomic=0: disable aotimically update. > ; fprofile-generate-atomic=1: aotimically update edge profile counters. > ; fprofile-generate-atomic=2: aotimically update value profile counters. > ; fprofile-generate-atomic=3: aotimically update edge and value profile > counters. > ; other values will be ignored (fall back to the default of 0). > fprofile-generate-atomic= > Common Joined UInteger Report Var(flag_profile_gen_atomic) Init(0) > Optimization > fprofile-generate-atomic=[0..3] Atomically increments for profile counters. > > thanks, > > David > > > On Wed, Jul 27, 2016 at 5:05 AM, Martin Liška <mli...@suse.cz> wrote: >> >> On 07/26/2016 01:15 AM, Andi Kleen wrote: >> > You definitely need a new flag: atomic or per thread instrumentation >> > will almost certainly have significant overhead (either at run time >> > or in memory). Just making an existing facility a lot of slower >> > without a way around it is not a good idea. >> >> Hi >> >> Agree with that! >> >> > >> > BTW iirc there were patches from google on this a few years back. >> > May be still in some branch. >> >> That's great, I'm CCing Google people who worked on PGO, hope they will >> help me to find patches/discussion about the problem. >> >> Martin >> >> > >> > -Andi >> >
