https://gcc.gnu.org/bugzilla/show_bug.cgi?id=82192
--- Comment #15 from Jakub Jelinek <jakub at gcc dot gnu.org> --- #c13 passed bootstrap/regtest on x86_64-linux and i686-linux for me. 71876582 49700491 14917420 5101491 is the grand total of combine.c's total_attempts, total_merges, total_extras and total_successes across those 2 bootstraps/regtests without the patch and 71928009 49744171 14941291 5102988 with the patch (both were profiledbootstraps and the combine.c code was slightly different in between). Out of this there is 4 times a: 35 35 4 3 to 31 31 4 2 change in gcc.c-torture/execute/pr82192.c with -m64, 2x 38605 33326 9261 1923 to 38653 33375 9285 1932 change in combine.c with -m32 and 2x 71226 33524 9201 2720 to 71276 33575 9205 2729 in combine.c with -m64. Comparing the sorted dumps which contain word bitsize, filename and these 4 numbers is a little bit hard, because some filenames keep changing (lto related, or e.g. the target-supports.exp created snippets which have pid appended to them), I see some cases of gcc.dg/compat/generate-random.c having different number of lines (due to make -jN check and randomness coming out of that). After some scripting and eyeballing the results, I believe pr82192.c and combine.c are the only spots that have real changes in those numbers, the rest is just whether some effective target tests or similar has been compiled 20 or 30 times depending on how many runtests have been invoked for the particular *.exp file. Also, stripped gcc/*.o files from the latest stage between non-patched and patched differ just in the expected combine.o and cc1*checksum.o.
