[llvm] [VectorCombine] Ignore frees when checking dereferenceability for load widening (PR #218657)

[Antonio Frighetto via llvm-commits]

https://github.com/antoniofrighetto approved this pull request.

LGTM. Haven't looked deeply at the regression in ocio/MatrixOpCPU, might that presumably be a cost-model issue (or a missed coercion/forwarding opportunity)?

https://github.com/llvm/llvm-project/pull/218657