On Wed, Jan 25, 2017 at 06:36:04PM +0100, Dominik Vogt wrote: > What's the matter with paradoxical subregs of memory? So, as far > as I understand, Recog refuses to match them unless they are > explicitly expressed as a pattern.
If INSN_SCHEDULING is defined recog does not allow paradoxical subregs of memory as a general_operand (memory_operand does not allow any subregs of memory). > On the other hand, Combine > does not know that they are "outlawed" and happily generates > them. combine should not generate things that can never match. Of course it sometimes does. This should be improved; please open a PR. > (set (reg/i:DI 2 %r2) > (and:DI (subreg:DI (mem:QI (plus:DI (reg:DI 2 %r2 [ f ]) > ^^^^^^^^^^^^^^^^^ > (const_int 2 [0x2])) [1 *f_5(D)+2 S1 A16]) 0) > (const_int 1 [0x1]))) > > Which doesn't match because of the paradoxical subreg of memory (I > think). > > Wouldn't it be better to allow that inside combine, then ann a > pass right after it that forces a reaload into a pseudo for such > memory references (for the targets that don't want them)? (Not > sure if theat would result in a noticeable gain or not.) combine already does this and then tries to split it at that mem; but in this case (a combination of only two insns) it won't even try to split. Segher
