[llvm] 6102310 - [InstSimplify][EarlyCSE] Try to CSE PHI nodes in the same basic block

[Roman Lebedev via llvm-commits]

Thank you Nikita, Philip and Owen!

With iterative process it was deduced that the main culprit is actually
the Instruction::isIdenticalToWhenDefined() change,
because e.g. EliminateDuplicatePHINodes() expects it do be dumb,
but (unlike EarlyCSE) did not verify it's invariants.

I've added that missing invariant checking
in 961483a5ea7c0e628c025187287d1658457ffcb3,
and verified that it would have caught
the Instruction::isIdenticalToWhenDefined() change.

With InstCombine-based PHI CSE still in place,
but Instruction::isIdenticalToWhenDefined() change reverted,
the build is finally green again:
http://lab.llvm.org:8011/builders/clang-with-thin-lto-ubuntu/builds/24100

Since InstCombine-based PHI CSE, indeed, catches all the motivational cases,
i'm *NOT* going to re-attempt InstSimpify-based approach.

Roman.

On Sat, Aug 29, 2020 at 2:35 AM Owen Anderson <resistor at mac.com> wrote:
>
>
>
> On Aug 28, 2020, at 2:21 PM, Nikita Popov via llvm-commits <llvm-commits at lists.llvm.org> wrote:
>
> As for actual breakage: I have a strong suspicion that this change has introduced compiler non-determinism. Looking at https://llvm-compile-time-tracker.com/index.php?config=ReleaseThinLTO&stat=size-text text size fluctuations have started occurring directly after this commit.
>
> Regards,
> Nikita
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at lists.llvm.org
> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits
>
>
>
> I’ve just finished bisecting a stage3 mismatch on x86 linux and tracked it down to this diff as well.  I’m going to move ahead with a revert.
>
> —Owen