[PATCH] D106447: [DAGCombine] DAGTypeLegalizer::GenWidenVectorLoads(): make use of dereferenceability knowledge

[Sebastian Neubauer via Phabricator via llvm-commits]

sebastian-ne added a comment.

Hi,
we talked a bit about this internally and found some potential problems.
There were some concerns that assuming that dereferenceable bytes >= alignment is incorrect. Memory protection may work at 4-byte or even byte granularity and objects may be over-aligned in memory. In these cases, loading padding bytes does not work.

Even if it can be proven that enough bytes are dereferenceable, we do not want to widen loads for amdgpu (at least not in cases where the load gets a lot larger, i.e. loading 64 Bytes instead of 32). Widening a load poses more restrictions on the register allocator, as a larger consecutive set of registers needs to be allocated. And, in the case that a widened load hits more cache lines than before, it also consumes more memory bandwidth.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D106447/new/

https://reviews.llvm.org/D106447