[Instcombine] Fix uses of undef (PR46940)

[kazu]

Summary

Without this patch, we attempt to distribute And over Xor even in
unsafe circumstances like so:

undef & (true ^ true)  ==>  (undef & true) ^ (undef & true)

and evaluate it to undef instead of false. Note that "true ^ true"
may show up implicitly with one true being part of a PHI node.

This patch fixes the problem by teaching SimplifyAndInst and
SimplifyUsingDistributiveLaws to not use undef as part of
simplifications when they distribute And over Xor.

[nikic]

Oops, looks like we were working on this in parallel. I just submitted D85684 for the InstSimplify part of the problem.

[kazu]

In D85687#2208099, @nikic wrote:

Oops, looks like we were working on this in parallel. I just submitted D85684 for the InstSimplify part of the problem.

Oh, OK. In that case, once your patch lands. I will drop the InstructionSimplify.cpp portion of my patch. Note that for my testcase in this patch to pass, I need both changes, the one in InstructionSimplify.cpp and the other in InstructionCombining.cpp.

[nikic]

D85684 has landed now. After rebasing, you can also use SQ.getWithoutUndef() to reduce the boilerplate for creating a new SimplifyQuery.

[kazu]

I've rebased this patch on top of https://reviews.llvm.org/D85684.

[kazu]

Please thank a look. Thanks!

[nikic]

LGTM

llvm/test/Transforms/InstCombine/dont-distribute-phi.ll

2

Please use llvm/util/update_test_checks.py to generate CHECK lines.

[kazu]

I've llvm/util/update_test_checks.py to update CHECK lines.

[jdoerfert]

In D85687#2211569, @kazu wrote:

I've llvm/util/update_test_checks.py to update CHECK lines.

I don't think this worked though.