[llvm] [X86] Remove single use assumption in combineVectorSizedSetCCEquality (PR #182200)

[Paul Kirth via llvm-commits]

ilovepi wrote:

Hi, we started seeing a test on Fuchsia time out after this patch. I'm still investigating, but this is the only x86 specific change in the blame list, and the test is still passing on Aarch64 and RISC-V.  Unfortunately, our CI had other problems in that had to be addressed in the Fuchsia build/upstream fixes, so we're a bit late in reporting/investigating.

Looking at the patch, it seems rather benign. Do you know how/why your PR could cause any noticeable performance difference? When I talked with test maintainers, this test hasn't been particularly close to the threshold, an usually finished in under a few seconds. 

As I'm investigating, is there anything specific that I should be looking for when comparing object files or intermediate IR? I'm trying to reproduce this locally, and test just a revert of your PR, but if there's something specific I should look for that may help me at least get you a better reproducer/bug report if it does end up being your patch.

Bot: https://ci.chromium.org/ui/p/fuchsia/builders/global.ci/clang_toolchain.ci.core.x64-release/b8687435582408623809/overview

https://github.com/llvm/llvm-project/pull/182200