[PATCH] D86737: [Attributor] Fix AANoUndef identification

[Shinji Okumura via Phabricator via llvm-commits]

okura added inline comments.
Herald added a subscriber: danielkiss.


================
Comment at: llvm/test/Transforms/Attributor/IPConstantProp/PR16052.ll:20
 ; IS__CGSCC____-NEXT:  entry:
-; IS__CGSCC____-NEXT:    [[CALL2:%.*]] = call i64 @fn1(i64 undef)
+; IS__CGSCC____-NEXT:    [[CALL2:%.*]] = call i64 @fn1(i64 noundef undef)
 ; IS__CGSCC____-NEXT:    ret i64 [[CALL2]]
----------------
jdoerfert wrote:
> Hm,... I think this exposes a problem. Could you take a look why we get a noundef here?
I overlooked this problem, thanks.
In this case, I expected that we indicate a pessimistic fixpoint for `%div = sdiv ...` in `genericValueTraversal` because we cannot strip value even once.
A simplified value for `%div` does not have value (i.e. `llvm::None`) because the assumed range is empty. (we can intentionally reduce undef to 0 there.)
Therefore, a callback is not called in `genericValueTraversal` and it returns `true`. That's why we got a noundef here.
We can avoid this by changing `manifest` not to manifest for such positions.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D86737/new/

https://reviews.llvm.org/D86737