https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116896
--- Comment #30 from GCC Commits <cvs-commit at gcc dot gnu.org> --- The master branch has been updated by Jakub Jelinek <ja...@gcc.gnu.org>: https://gcc.gnu.org/g:ff889b35935d5e796cf308fb2368d4e319c60ece commit r15-4167-gff889b35935d5e796cf308fb2368d4e319c60ece Author: Jakub Jelinek <ja...@redhat.com> Date: Tue Oct 8 10:40:29 2024 +0200 ssa-math-opts, i386: Handle most unordered values rather than just 2 [PR116896] On Mon, Oct 07, 2024 at 10:32:57AM +0200, Richard Biener wrote: > > They are implementation defined, -1, 0, 1, 2 is defined by libstdc++: > > using type = signed char; > > enum class _Ord : type { equivalent = 0, less = -1, greater = 1 }; > > enum class _Ncmp : type { _Unordered = 2 }; > > https://eel.is/c++draft/cmp#categories.pre-1 documents them as > > enum class ord { equal = 0, equivalent = equal, less = -1, greater = 1 }; // exposition only > > enum class ncmp { unordered = -127 }; // exposition only > > and now looking at it, LLVM's libc++ takes that literally and uses > > -1, 0, 1, -127. One can't use <=> operator without including <compare> > > which provides the enums, so I think if all we care about is libstdc++, > > then just hardcoding -1, 0, 1, 2 is fine, if we want to also optimize > > libc++ when used with gcc, we could support -1, 0, 1, -127 as another > > option. > > Supporting arbitrary 4 values doesn't make sense, at least on x86 the > > only reason to do the conversion to int in an optab is a good sequence > > to turn the flag comparisons to -1, 0, 1. So, either we do nothing > > more than the patch, or add handle both 2 and -127 for unordered, > > or add support for arbitrary value for the unordered case except > > -1, 0, 1 (then -1 could mean signed int, 1 unsigned int, 0 do the jumps > > and any other value what should be returned for unordered. Here is an incremental patch which adds support for (almost) arbitrary unordered constant value. It changes the .SPACESHIP and spaceship<mode>4 optab conventions, so 0 means use branches, floating point, -1, 0, 1, 2 results consumed by tree-ssa-math-opts.cc emitted comparisons, -1 means signed int comparisons, -1, 0, 1 results, 1 means unsigned int comparisons, -1, 0, 1 results, and for constant other than -1, 0, 1 which fit into [-128, 127] converted to the PHI type are otherwise specified as the last argument (then it is -1, 0, 1, C results). 2024-10-08 Jakub Jelinek <ja...@redhat.com> PR middle-end/116896 * tree-ssa-math-opts.cc (optimize_spaceship): Handle unordered values other than 2, but they still need to be signed char range possibly converted to the PHI result and can't be in [-1, 1] range. Use last .SPACESHIP argument of 1 for unsigned int comparisons, -1 for signed int, 0 for floating point branches and any other for floating point with that value as unordered. * config/i386/i386-expand.cc (ix86_expand_fp_spaceship): Use op2 rather const2_rtx if op2 is not const0_rtx for unordered result. (ix86_expand_int_spaceship): Change INTVAL (op2) == 1 tests to INTVAL (op2) != -1. * doc/md.texi (spaceship@var{m}4): Document the above changes. * gcc.target/i386/pr116896.c: New test.
