Relevant code is

The code is also wrong in that !isNullValue doesn't mean that the index can't be zero, just that it's not known to be zero.

Maybe this fold can be dropped entirely? It seems pretty useless to me.

Here's another one, which folds (x-1) ugt x -> true:

define i1 @​global_gep_ugt_global_neg_offset() {
%gep = gep * @​g, 8 x i64 -1
%cmp = icmp ugt * %gep, @​g
ret i1 %cmp
}
=>
define i1 @​global_gep_ugt_global_neg_offset() {
ret i1 1
}
Transformation doesn't verify!

ERROR: Value mismatch

Example:

Source:

• %gep = pointer(non-local, block_id=0, offset=-8)
  i1 %cmp = #x0 (0)

Target:
Source value: #x0 (0)
Target value: #x1 (1)

I believe the right way to fix all these folds is https://reviews.llvm.org/D97665, which limits them to inbounds only. While some of them are also valid without inbounds in limited circumstances, I don't think it's worth bothering with the non-inbounds case here.

I believe the right way to fix all these folds is
https://reviews.llvm.org/D97665, which limits them to inbounds only. While
some of them are also valid without inbounds in limited circumstances, I
don't think it's worth bothering with the non-inbounds case here.

I agree; it's not worth it to spend a lot of time with non-inbounds geps. Plus these folds probably don't kick in that often.

But, the last transformation I pasted is completely wrong (likely a copy-paste mistake), as it's trying to prove that (x-1) > x. Since x is non-null, that isn't possible.

I went on a bit of a crusade and deleted a lot of the folds in this function, because they were all subtly incorrect. The one relevant to this report is gone in 6c03178.