https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68173
--- Comment #7 from Vladimir Makarov <vmakarov at gcc dot gnu.org> --- GCC on trunk is 50 times slower than LLVM-3.7 in -O0 mode on this test. In -O2 mode GCC is only 20% slower than LLVM-3.7 in the same mode and faster than LLVM-3.7 with -O0. About 80% of compile time GCC spent in RA, more accurately in bitmap code used by LRA (actually bitmap code takes even more if we take their usage from other passes). Conversion bitmaps to sparsets where memory requirements permit did not help at all. After analyzing RA dumps (working with 8GB dump files in emacs even on machine with 128GB memory is not fun as it reacts to any key press with big delays), I believe the problem is not in RA but somewhere before as we have *5M* allocnos and *10M* program points in -O0 mode and only *6K* allocnos and *20K* program points in -O2 mode. Even if I make RA instant, GCC will be still 10 times slower in -O0 mode than LLVM. I guess solution for the PR would be switching on some optimizations in -O0 mode which can help decreasing number of pseudos but I am not sure it is a right thing to do in general.
