[PATCH] D50491: [DAGCombiner][AMDGPU][Mips] Fold bitcast with volatile loads if the resulting load is legal for the target.

[Eli Friedman via Phabricator via llvm-commits]

efriedma added a comment.

For volatile loads/stores, we should probably try to preserve the number and addresses of memory operations.  If the source and destination types would lower to equivalent memory operations, like i32 vs. f32 on x86, the transform is fine.  If they wouldn't, like f64 vs. i64 on x86-32, the transform is dubious; memory-mapped I/O can respond in strange ways if you don't use the right access width.


Repository:
  rL LLVM

https://reviews.llvm.org/D50491