[PATCH] D90529: Allow nonnull attribute to accept poison

[Juneyoung Lee via Phabricator via llvm-commits]

aqjune added a comment.

Should we have clarification about the semantics of align/nonnull bundle for the assume intrinsic?
When a function call is inlined, the attributes of the parameters may remain as assume's bundles:

  call f(nonnull %p)
  ->
  llvm.assume [ "nonnull"(%p) ]

Here are two possible semantics:

1. When `"nonnull"(%p)` is at the operand bundle, there also should be `"noundef"(%p)`, otherwise it is a no-op.

  ; A correct transformation
  call f(nonnull noundef %p)
  ->
  llvm.assume [ "nonnull"(%p), "noundef"(%p) ] ; since noundef and nonnull are both in the bundle, it is UB if %p is null



2. `"nonnull"(%p)` diverges from `nonnull`, and immediately raises UB if `%p` is null. This means that the lowering should be done only when `noundef` existed.

  ; A correct transformation
  call f(nonnull noundef %p)
  ->
  llvm.assume [ "nonnull"(%p) ] ; this is UB if %p is null

The former makes attribute -> op bundle conversion simpler, whereas the latter makes reasoning from op bundle simpler (well, maybe both are simple enough). :)
I think either option is fine.


Repository:
  rG LLVM Github Monorepo

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

https://reviews.llvm.org/D90529