[llvm] 3794cc0 - [InstCombine] invert canonicalization for cast of signbit test

[Sanjay Patel via llvm-commits]

Thanks for letting me know. Reverted:
https://reviews.llvm.org/rG07d549bce94f

If you can get the IR for that file, that would be great. Otherwise, I'll
try to set up a system to extract that.

On Mon, May 16, 2022 at 5:12 PM Florian Hahn <florian_hahn at apple.com> wrote:

> Hi
>
> On May 16, 2022, at 17:56, Sanjay Patel via llvm-commits <
> llvm-commits at lists.llvm.org> wrote:
>
>
> Author: Sanjay Patel
> Date: 2022-05-16T12:55:52-04:00
> New Revision: 3794cc0e996481e10307b67c8436aa44e0d65d22
>
> URL:
> https://github.com/llvm/llvm-project/commit/3794cc0e996481e10307b67c8436aa44e0d65d22
> DIFF:
> https://github.com/llvm/llvm-project/commit/3794cc0e996481e10307b67c8436aa44e0d65d22.diff
>
> LOG: [InstCombine] invert canonicalization for cast of signbit test
>
> The existing transform was wrong in 3 ways:
> 1. It created an extra instruction when the source and dest types don't
> match.
> 2. It did not account for an extra use of the icmp, so could create 2
> extra insts.
> 3. It favored bit hacks over icmp (icmp generally has better analysis).
>
> This fixes #54692 (modeled by the PhaseOrdering tests).
>
> This is a minimal step to fix the bug, but we should likely invert
> the sibling transform for the "is negative" pattern too.
>
> The backend should be able to invert this back to a shift if that
> leads to better codegen.
>
>
>
> I believe this change may be causing a hang in bootstrap builds when
> building llvm-project/clang/unittests/Frontend/CompilerInvocationTest.cpp
> during stage2. With this patch reverted, building that file completes
> within 5 seconds. With the patch, Clang spins in instcombine.
>
> It looks like the following bots are timing out because of this:
>
> https://lab.llvm.org/buildbot#builders/179/builds/3688
> https://lab.llvm.org/buildbot#builders/85/builds/8641
> https://lab.llvm.org/buildbot#builders/168/builds/6486
> https://lab.llvm.org/buildbot#builders/74/builds/11107
>
> It might be good to revert the change if it cannot be fixed quickly, to
> get the bots back to green.
>
> Cheers,
> Florian
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20220516/874861bf/attachment.html>