[Diffusion] rGb53d44fe4741: [EarlyCSE][ConstantFolding] add tests for atan/atan2; NFC

[Sanjay Patel via Phabricator via llvm-commits]

spatel added a comment.

In rGb53d44fe47413c87f619baf9e057f2a2e49d3998#1114307 <https://reviews.llvm.org/rGb53d44fe47413c87f619baf9e057f2a2e49d3998#1114307>, @hubert.reinterpretcast wrote:

> In rGb53d44fe47413c87f619baf9e057f2a2e49d3998#1114285 <https://reviews.llvm.org/rGb53d44fe47413c87f619baf9e057f2a2e49d3998#1114285>, @spatel wrote:
>
>> Note that the planned code change corresponding to these tests landed with 30abc1a6a18e <https://reviews.llvm.org/rG30abc1a6a18e843fd0470157d4d2ec285aaac8cd> (before I saw the comments here), but I don't think that resolves the discrepancy.
>>
>> Should we revert to get the bot back to passing?
>
> If the only value difference is this one (+/- 0.0), maybe the CHECK line can be adjusted to accept either?

Yes, we can do that. To be clear (please correct me if I'm wrong), the discrepancy existed before D127964 <https://reviews.llvm.org/D127964>, so this commit to add tests was just exposing a potential mathlib difference.

It does make me nervous that we were too permissive in D127964 <https://reviews.llvm.org/D127964>. I think there's a general problem that we're doing the constant folding using the host mathlib, but that can produce different results than the target mathlib.


BRANCHES
  main

Users:
  mohammed-nurulhoque (Author)
  spatel (Committer)

https://reviews.llvm.org/rGb53d44fe4741