https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67463
--- Comment #3 from Shlomi Fish <shlomif at shlomifish dot org> --- Hi Martin, thanks for returning to me. (In reply to Martin Liška from comment #2) > Hi. > > Sorry for waiting for some time. I tested your benchmark, where I had to > disable tcmalloc as I can't link it with GCC 4.9, caused by C++ abi change. > I also enhanced dump output to print total time spent (when function > FCS_PRINT_FINISHED is called): > > 4.9.4: > Finished in 3.607 > Finished in 3.613 > Finished in 3.621 > Finished in 3.662 > > 5.4.0: > Finished in 3.643 > Finished in 3.636 > Finished in 3.630 > Finished in 3.708 > > 6.3.0: > Finished in 3.618 > Finished in 3.618 > Finished in 3.627 > Finished in 3.652 > > Tested on my desktop machine with Intel(R) Core(TM) i7-4770 CPU @ 3.40GHz. > That mentioned, I cannot see the regression, all numbers look within noise > level. > The regression you reported is very small (~3%) and it would be very hard to > compare generated assembly to find different decisions made by GCC. Please > try to > test GCC 6 and possible GCC 7 (which will be released in couple of weeks) > and if seen > a significant regress, please reopen the issue. Thanks! I'll try.