http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53100
--- Comment #1 from Marc Glisse <marc.glisse at normalesup dot org> 2012-04-29 08:05:59 UTC --- (In reply to comment #0) > It would be convenient if I > could just write the whole code with __int128 and let the compiler do the > optimization by tracking the range of numbers. The transformation from an __int128 to a pair of long happens extremely late (optabs.c), so we can't count on tree-vrp to notice that one of them is always zero (and actually it is either 0 or -1, as a sign extension, which would make this hard). On the other hand, tree-vrp does have the information that the differences are in [-4294967295, 4294967295], which comfortably fits in a type half the size of __int128. It seems a possible strategy would be to have tree-vrp mark variables that fit in a type half their size (only for TImode?), try and preserve that information along the way, and finally use it in expand_doubleword_mult. But that seems to imply storing the information in an rtx, and rtx seems a bit too densely packed to add this. Better ideas?
