[PATCH] D11591: Allow PeepholeOptimizer to fold a few more cases

[Quentin Colombet via llvm-commits]

qcolombet added a comment.

Hi Michael,

I believe the current situation is indeed an historical mistake :).

Regarding the load folding, no, I do not believe isPosition should be a barrier, probably the same for inline asm. The thing about inline asm is that if the user forgot the volatile keyword whereas it was doing unmodeled side effect, we may not model the right thing. That being said, I guess this is a mistake from the user side and we do not need to be conservative in such case.

I’d say update your patch to reflect that.

Side note, could you be more specific in the test cases regular expressions?
I.e., if you expect a memory location for a given parameter now, instead of a register, try to reflect the in the expression so that if we regress that we will see it.

Thanks,
-Quentin


http://reviews.llvm.org/D11591