On Wed, Aug 19, 2020 at 11:05 AM Richard Sandiford <richard.sandif...@arm.com> wrote: > Sorry for the slow reply.
Not a problem at all. Thank you for responding! > Pip Cet via Gcc-patches <gcc-patches@gcc.gnu.org> writes: > > I'm working on the AVR cc0 -> CCmode conversion (bug#92729). One > > problem is that the cmpelim pass is currently very strict in requiring > > insns of the form > > > > (parallel [(set (reg:SI) (op:SI ... ...)) > > (clobber (reg:CC REG_CC))]) > > > > when in fact AVR's insns often have the form > > > > (parallel [(set (reg:SI) (op:SI ... ...)) > > (clobber (scratch:QI)) > > (clobber (reg:CC REG_CC))]) > > > > The attached patch relaxes checks in the cmpelim code to recognize > > such insns, and makes it attempt to recognize > > > > (parallel [(set (reg:CC REG_CC) (compare:CC ... ...)) > > (set (reg:SI (op:SI ... ...))) > > (clobber (scratch:QI))]) > > > > as a new insn for that example. This appears to work. > > The idea looks good. However, I think it'd be better (or at least > more usual) for the define_insns to list the clobbers the other > way around: > > (parallel [(set (reg:SI) (op:SI ... ...)) > (clobber (reg:CC REG_CC)) > (clobber (scratch:QI))]) That makes sense, thanks for the suggestion. I realized quite quickly that I would have to reproduce the patterns precisely, including order in a parallel, and decided to go with the wrong consistent option by always placing the CC clobber last. My question is whether it makes more sense to recognize either form (i.e. a clobber of targetm.flags_regnum at any position in a parallel) or whether it's okay to assume that the clobber is always the second element in the parallel. I'm leaning towards the latter version. > (clobber (scratch…))s generally come last because any rtl optimisation > pass that uses recog can automatically add any necessary > (clobber (scratch…))s. In contrast, very few passes (probably just > combine) know how to add (clobber (reg:CC REG_CC)) to a pattern that > didn't already have it. This is because adding a REG_CC clobber > requires the pass to “prove” that REG_CC is dead at that point, > whereas there are no restrictions on adding a scratch clobber. Thanks again! Pip
