On Wed, May 3, 2017 at 12:16 AM, Daniel Santos <daniel.san...@pobox.com> wrote:
>> I'd recommend that in the predicate, you check match_parallel from the >> bottom up, since subexpressions on the top are already matched, and >> you can have different number of subexpressions at the top. >> >> Uros. > > > OK, this all makes sense then. I was using const_int tags to differentiate > the insns, but omitting the insns for the leave (for example) are what > ambiguates them in the first place. > > So one question I never had resolved is rather or not the order the insns in > a parallel matters. From your suggestions, I'm guessing that it would be > perfectly OK for the leave insns to follow use symbol and then have the > register restores after that, even though that wouldn't make sense from an > assembly level because we would be writing beyond the stack pointer. Am I > reading this correctly? If so, then putting the register save/restore > matching in the predicate (checking top down) and having all others in the > pattern make great sense and should both simplify it and make it more clear. The order of subexpressions of parallel in general does not matter. > Also, I'm wondering if there's anything wrong with calling ix86_gen_leave () > and plucking the insns out of the generated parallel insn and moving that > into my own parallel rather than generating them in my own function. I > guess all the matters is what is cleanest. Hm... I'd rather see subexpressions generated "by hand". Uros.
