[PATCH] D86222: Fix PR46880: Fail CHECK-NOT with undefined variable

[Thomas Preud'homme via Phabricator via llvm-commits]

thopre added a comment.

In D86222#2656914 <https://reviews.llvm.org/D86222#2656914>, @jdenny wrote:

> In D86222#2654396 <https://reviews.llvm.org/D86222#2654396>, @thopre wrote:
>
>> 
>
>
>
>> A lot of the failures (e.g.  `llvm/test/Transforms/GVN/big-endian.ll`, `llvm/test/Transforms/LoopVectorize/pr34681.ll`, `llvm/test/Transforms/LoopVectorize/X86/x86-pr39099.ll`) are tests that use several CHECK-NOT with one line defining a variable and another one using it.
>
> How many tests does this patch break?

60, one of them being the FileCheck unittest which I haven't adapted yet. I've already made 4 patches solving 5 tests:

https://reviews.llvm.org/D99581
https://reviews.llvm.org/D99582
https://reviews.llvm.org/D99583
https://reviews.llvm.org/D99589

>> What do you think? Is adding a new feature first the best way forward? If yes, how would you design it?
>
> Those tests are passing without this patch, right?  That should mean the CHECK-NOT that defines the variable never matches, so it should be impossible for the CHECK-NOT that uses the variable to match even if we had the feature you're suggesting.  A quick fix is then to remove or comment out the CHECK-NOT that uses the variable.  I believe the tests would be just as robust as now.
>
> Before trying to design a new feature, I'd pursue that quick fix and see what the test authors say.

You're right, let's try that.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D86222/new/

https://reviews.llvm.org/D86222