[llvm] [polly] [IR] Enable dereferenceable-at-point semantics (PR #204795)
Nikita Popov via llvm-commits
llvm-commits at lists.llvm.org
Tue Jun 30 01:09:39 PDT 2026
nikic wrote:
> Overall great to close this hole.
>
> > It's expected that this will regress optimization of C++ code in some cases, because we can no longer use dereferenceability facts that are not, in fact, guaranteed to hold.
>
> Indeed, from a large IR corpus of C/C++ workloads there seems to be quite quite a few cases that regress with the change, e.g. missed LICM on something like https://llvm.godbolt.org/z/1zWonrqvo and in turn missed vectorization. For most cases I spotted, we won't be able to infer `nofree` (either for argument or calls) unfortunately.
Unfortunately, I don't think there is anything we can do about cases like that (unless it's possible to improve nofree inference in some way).
https://github.com/llvm/llvm-project/pull/204795
More information about the llvm-commits
mailing list