[Mlir-commits] [mlir] [mlir][AMDGPU] Add a scheduling barrier guard around inlineAsm lds.barrier (PR #109678)

[Krzysztof Drewniak]

krzysz00 wrote:

And re "asm should not be used" - this is the "trust me, we specifically don't want to wait on global memory to complete at the barrier because we're doing software pipelining" barrier where we're by-hand mitigating the effects of `BackOffBarrier` not being set (even though it breaks debugging, hence the warning).

This rewrite only uses ASM on the architectures where you need to use inline ASM to get that effect

https://github.com/llvm/llvm-project/pull/109678