On Sun, Sep 1, 2013 at 6:57 AM, Jan Hubicka <hubi...@ucw.cz> wrote: > Hi, > this patch implement speculative devirtualization. It is a trivial pass that > asks for targets of every polymorphic call in a program and if the list > contains one likely target, it produces an speculative call. No context > sensitive analysis is done at the moment. This call may or may not survive > into final program depending if we do somehting useful about the direct call. > > The pass is currently disabled for LTO because > http://gcc.gnu.org/ml/gcc-patches/2013-08/msg01007.html is needed to properly > build type inheritance graph. > > With LTO it is supposed to be effective on a premise that most types are not > escaping LTO unit and thus we see all possible targets. Without LTO it makes > stronger assumption that you usually call only targets defined in current > unit, > if any. > > Path is suprisingly effective on Firefox: > 105306 polymorphic calls, 0 devirtualized, 34258 speculatively devirtualized, > 4056 cold > 66875 have multiple targets, 0 overwritable, 0 already speculated (0 agree, 0 > disagree), 0 not defined > > So about 32% of calls are devirutalized. By random checking, these can be > tracked to real occurences of code where virtual is used in a silly way. > I plan to introduce warning for that (I have code for that already since it > makes it easier to analyze what changes are really made and why). > > Martin Liska rebuilt with FDO based indirect call resolution. Here we get: > 23928 indirect calls trained. > 12837 (53.65%) have common target. > 342 (1.43%) targets was not found. > 8378 (35.01%) speculations seems useless. > 4117 (17.21%) speculations produced. > > I compared the overlap that is devirtualized by both techniques. There is > almost 100% match, except that FDO code is not dealing well with thunks and > those 342 calls that seem to go out of libxul into plugins. I will fix the > thunk issue later. > > I also tested QT, where the numbers are smaller - only about 20% of > devirtualized > calls, but largery things seems similar. > > For non-LTO build, the devirtualization also seems sane, there seems to be > about 8% of miss rate on GCC bootstrap that seems acceptable. I tracked most > of > those down into randomly included headers that do define derived types of a > given class that are unused in the current unit. I think we can track this by > computing reachable functions in current unit and looking for vtables actually > used by construction. One of such actually triggers undefined reference in > build of libstdc++ and therefore I added the check disabling devirtualization > to DECL_EXTERNAL for now. It is because the libstdc++ header seems to have > explicit instantiation of a template that is never linked with. > > I currently enabled the pass by default at -O2. Based on the experience about > missrate, we may want to disable it for -O2 non-LTO if it shows to be too > risky > on some codebases. > > Bootstrapped/regtested x86_64-linux and ppc64-linux, also tested with lto > bootstrap > with the LTO ODR code and tested on Firefox and QT builds. Will commit the > patch > later today. > > Comments are welcome. > Honza > > * common.opt (fdevirtualize-speculatively): New function. > * invoke.texi (fdevirtualize-speculatively): Document. > * ipa-devirt.c: Include ipa-inline.h > (likely_target_p): New function. > (ipa_devirt): New function. > (gate_ipa_devirt): New function. > (pass_data_ipa_devirt): New static var. > (pass_ipa_devirt): Likewise. > (make_pass_ipa_devirt): New function. > * opts.c (default_options): Add OPT_fdevirtualize_speculatively. > (common_handle_option): Disable devirtualization when > value range profiling is available. > * passes.def (pass_ipa_devirt): Add. > * timever.def (TV_IPA_DEVIRT): New timevar. > * tree-pass.h (make_pass_ipa_devirt): >
This triggered: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63814 H.J.
