[InstSimplify] Fold out-of-bounds shift to poison

[nikic]

Summary

Make InstSimplify return poison rather than undef for out-of-bounds shifts, as specified by LandRef:

If op2 is (statically or dynamically) equal to or larger than the number of bits in op1, this instruction returns a poison value.

[nikic]

llvm/test/Transforms/InstCombine/shift.ll

1723

A peculiar regression. Not sure how this folded before, but we're clearly missing a fold here: https://llvm.godbolt.org/z/89bah8 Poison can fold to poison: https://alive2.llvm.org/ce/z/7Nwdri Undef can fold to base pointer: https://alive2.llvm.org/ce/z/y6NLvJ Undef can't fold to undef: https://alive2.llvm.org/ce/z/emLp_H

[aqjune]

llvm/test/Transforms/InstCombine/shift.ll

1723

The source program looks really interesting - it has gep with i1 index..!
I guess %G18 should have been folded like this:
gep %A2, i1 %C8 -> gep %A2, i1 undef -> gep %A2, i64 0 (by constfold sext undef) -> gep %A2
With poison, constfold sext poison is poison, hence it resulted in here.
I'll make a patch that makes gep poison -> poison.

[aqjune]

llvm/test/Transforms/InstCombine/shift.ll

1723

It should work now: https://github.com/llvm/llvm-project/commit/f665a8c5b8b42d88f89e1a3594b7d410ef206c32

[nikic]

Rebase over new gep poison fold.

[aqjune]

LGTM