[PATCH] D90529: Allow nonnull attribute to accept poison

[Johannes Doerfert via Phabricator via llvm-commits]

jdoerfert requested changes to this revision.
jdoerfert added a comment.
This revision now requires changes to proceed.

I think we have two choices here:

1. Don't raise UB when "value attributes" are passed a "wrong value", e.g., `null` for a `nonnull` attribute, but make the value poison. Use `nonull` + `noundef` to make it UB.
2. Make all "value attributes" accept poison without raising UB.

I was in the past pushing for 1), I don't have the link handy but we can probably find it. I think the last time I brought this up was the `noundef` discussion actually.
One of my examples was the `gep` one shown in the commit message,
I vaguely remember one where the user "broke the contract" but in a way they would assume to be harmless, e.g., they did not cause any side-effect, maybe something like:

  void foo(bool valid, X& x) {
    if (!valid) return;
    ...
  }
  
  obj_ptr = null;
  foo(obj_ptr != null, *obj_ptr);

I remember @efriedma was not a fan of 1) at the time, unsure if that is still the case with `noundef` in place.

If we don't do 1), we should talk about 2) before we make `nonnull` special. I fail to see the reason it is different from any other (or at least most) "value attributes".


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D90529/new/

https://reviews.llvm.org/D90529