lebedev.ri added a comment. In D67122#1663619 <https://reviews.llvm.org/D67122#1663619>, @vsk wrote:
> > In D67122#1659721 <https://reviews.llvm.org/D67122#1659721>, @vsk wrote: > > > >> Still think this looks good. Have you tried running this on the llvm test > >> suite, or some other interesting corpus? Would be curious to see any > >> pre/post patch numbers. > > > > > > There were 1.5 occurrences in test-suite. > > > > I'd prefer not to try running it over llvm stage-2/3 (i can, but it's gonna > > be slow..) > > > > Though i guess you were talking about *performance* numbers? > > Yes, I'm primarily interested in run-time and compile-time performance (at > {-O0 -g, -Os -g}). Uh, feel free to measure _that_ yourself :D > The number of occurrences of new diagnostics is also interesting. What does > the .5 mean? There were two occurrences - rL371219 <https://reviews.llvm.org/rL371219> and rL371220 <https://reviews.llvm.org/rL371220>, and only one of those is new. (as per @rupprecht it catches something in the original miscompile test, so there's that too) >> I'm not good with using test-suite for perf, so i'm yet again going to use >> my own bench. >> F9934220: rawspeed-pointer-overflow-0-baseline.json >> <https://reviews.llvm.org/F9934220> F9934221: >> rawspeed-pointer-overflow-1-old.json <https://reviews.llvm.org/F9934221> >> F9934222: rawspeed-pointer-overflow-2-new.json >> <https://reviews.llvm.org/F9934222> > > I see, I'll make some time to collect & share stats from it soon. Cool > On that note, I've never had any problems with running the suite in the past, > and the results have been very helpful (if nothing else, > it's nice to have a mid-size test corpus everyone can share). > What sorts of issues are you running into? > Were you able to get interesting results with -DTEST_SUITE_RUN_BENCHMARKS=0 > set? Note that i wasn't looking for compile time stats, let alone at -O0, there is nothing odd in the diff that could have an impact other than the usual "death by a thousand cuts". As for benchmarking, it has pretty long run-time, multiply that by the number of repetitions, and the results are noisy still. >> TLDR: (all measurements done with llvm ToT, the sanitizer never fired.) >> >> - no sanitization vs. existing check: average `+20.75%` slowdown >> - existing check vs. check after this patch: average `+26.36%` slowdown >> - no sanitization vs. this patch: average `+52.58%` slowdown ^ so i provide my own numbers. I'm chirping away at missed folds, already handled two, two more coming, so i suspect runtime overhead is already lower and will be lower still. Repository: rG LLVM Github Monorepo CHANGES SINCE LAST ACTION https://reviews.llvm.org/D67122/new/ https://reviews.llvm.org/D67122 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits