[PATCH] D65280: Add a pass to lower is.constant and objectsize intrinsics

[Chandler Carruth via Phabricator via llvm-commits]

chandlerc requested changes to this revision.
chandlerc added a comment.
This revision now requires changes to proceed.

Generally, I do like the approach. Two high level comments:

First, I don't think we can crash if these things reach lower layers. I think we should retain the logic to just fold them to a constant if somehow they show up. Anything else makes things weird -- the IR is valid, passes the verifier, but crashes? We want to support fuzzing and bisecting and such and so it should be possible to rip this pass out and still have things lower successfully. This pass should just be a much more *advanced* way of lowering these intrinsics. And yes, I understand that in some bizarre cases failing to do this "optimization" will result in inline assembly that cannot be emitted or some such. While I find that really frustrating as well, I still think it is better than making even more boring uses of these intrinsics hostile to things like fuzzers.

Second, I don't think you need the fancy approach you're taking to find the branches that need simplification. How about the following approach:

1. Walk every instruction and if it is one of these instrinsics, fold it and then recursively simplify.

2. If we simplify any intrinsics, but after simplifying *all* of them found in the function, walk every terminator in every block and if it is a conditional branch with a constant condition, fold the branch.

I don't think this will be too wasteful, and seems much simpler.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D65280/new/

https://reviews.llvm.org/D65280