[Mlir-commits] [mlir] [MLIR] Implement emulation of static indexing subbyte type vector stores (PR #115922)

[llvmlistbot at llvm.org]

MaheshRavishankar wrote:

> 
> Feel like in such a non-competing case we can do strength reduction and use masked store instead of atomic for the unaligned parts (the beginning and the end byte if it is unaligned).

You cant always tell from the program analysis that the atomic isnt needed. This depends on the caller's context. For example if the caller knows that, even dynamically, the stores never overlap. It is hard to recover that information at this late stage. So I think it would be good to have an option where atomics are avoided (cause it is expensive) and let caller opt-in/out?

https://github.com/llvm/llvm-project/pull/115922