[PATCH] D88880: [LoopFlatten] Initial support for different types

[Oliver Stannard (Linaro) via Phabricator via llvm-commits]

ostannard added inline comments.


================
Comment at: llvm/lib/Transforms/Scalar/LoopFlatten.cpp:386
+      //
+      // For the overflow check, we first calculate an 'effective' GEP
+      // bitwidth. I.e., we look through a ZExt instruction which widens the
----------------
I don't think this logic is correct. The check below is proving that the add and mul cannot overflow because the GEP would overflow before they do, and that would be UB. However, in the example in this comment, the add or mul could overflow, but the GEP would be fine, because it never sees an offset greater than 2*32-1. Since there's no UB, we would need to preserve the wrapping behaviour, which we can't do with a single, simple loop.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D88880/new/

https://reviews.llvm.org/D88880