[InstructionSimplify] InstructionSimplify error deleted `phi`

[dianqk]

I tried this IR:

define i32 @src(i1 %c0, i1 %c1, i32 %v1) {
start:
  br i1 %c0, label %bb0, label %bb1

bb0:
  br label %bb1

bb1:
  %x = phi i32 [ %v1, %bb0 ], [ undef, %start ]
  br i1 %c1, label %bb0, label %bb2

bb2:
  ret i32 %x
}

phi should not be deleted by InstructionSimplify.

alive2: https://alive2.llvm.org/ce/z/LNXeGp

[nikic]

I don't think this is related to CVP? This is InstSimplify behavior.

This poison-propagation hole (and a number of similar ones) has been left intentionally until the "poison on uninit load" semantics come into effect. Though it's possible that the relatively wide use of noundef nowadays allows us to fix this independently. We'd have to evaluate codegen impact.

[dianqk]

I don't think this is related to CVP? This is InstSimplify behavior.

Yes. The problem is in the

llvm-project/llvm/lib/Analysis/InstructionSimplify.cpp

Line 5284 in a5686c2

return valueDominatesPHI(CommonValue, PN, Q.DT) ? CommonValue : nullptr;

.

This poison-propagation hole (and a number of similar ones) has been left intentionally until the "poison on uninit load" semantics come into effect. Though it's possible that the relatively wide use of noundef nowadays allows us to fix this independently. We'd have to evaluate codegen impact.

Do you have any suggestions? It looks like this is a rather large change/impact.

[dtcxzyw]

It looks like this is a rather large change/impact.

I agree. This has a huge impact on Rust applications due to the lack of noundef on arguments/retvals.