[llvm] [SLP] Bail out on store-to-load forwarding hazards (PR #199606)

[Alexey Bataev via llvm-commits]

================

----------------
alexey-bataev wrote:

Unsound and Unreachanble, better to delete.
1. For the misalignment-only predicate, but not once term (b) (LoadElementSize > VectorStoreSize) exists: the widened-load width is VF-independent, so LoadElementSize <= VF*ElementSize does not imply LoadElementSize <= V*ElementSize for a divisor V. 
2. It is also dead code today: a store scalar belongs to exactly one tree entry with exactly one VF, and every buildTree calls deleteTree() (SLPVectorizer.cpp:9100,9110), which clears the cache — so (store, VF) is never queried at two VFs within one cache lifetime.

https://github.com/llvm/llvm-project/pull/199606