[llvm] Zext sext undef in sdag (PR #115451)

[Craig Topper via llvm-commits]

topperc wrote:

> > This optimization is not correct for 'undef'. I think it would be correct for 'poison', but SelectionDAG doesn't have 'poison' today.
> > See also https://blog.regehr.org/archives/1837
> 
> I see. Thanks for your explanation, so zext undef to a zero is by designed, is this right?

For (zext undef), the zero extended bits of the result must be 0. The bits that started out as undef can be any value. We choose 0 for simplicity.

For (sext undef), the sign extended bits must have the same value as the original sign bit. Because of undef, the original sign bit can be either 0 or 1. Folding to undef wouldn't guarantee the extended bits all match, so we fold to all 0s.

https://github.com/llvm/llvm-project/pull/115451