topperc wrote:
> Here's a simple-ish example:
>
> ```llvm
> ; Transforms/InstCombine/add.ll
>
> define i5 @zext_sext_not(i4 %x) {
> %zx = zext i4 %x to i5
> %notx = xor i4 %x, 15
> %snotx = sext i4 %notx to i5
> %r = add i5 %zx, %snotx
> ret i5 %r
> }
> =>
> define i5 @zext_sext_not(i4 %x) {
> %zx = zext i4 %x to i5
> %notx = xor i4 %x, 15
> %snotx = sext i4 %notx to i5
> %r = or disjoint i5 %zx, %snotx
> ret i5 %r
> }
> Transformation doesn't verify! (unsound)
> ERROR: Target is more poisonous than source
>
> Example:
> i4 %x = undef
>
> Source:
> i5 %zx = #x00 (0) [based on undef value]
> i4 %notx = #xf (15, -1) [based on undef value]
> i5 %snotx = #x1f (31, -1)
> i5 %r = #x1f (31, -1)
>
> Target:
> i5 %zx = #x0f (15)
> i4 %notx = #xf (15, -1)
> i5 %snotx = #x1f (31, -1)
> i5 %r = poison
> Source value: #x1f (31, -1)
> Target value: poison
> ```
>
> Essentially the code doesn't take into consideration that a same register may
> have different values when it is undef.
Looks like this is caused by the checks in `haveNoCommonBitsSetSpecialCases`.
https://github.com/llvm/llvm-project/pull/72912
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
