[llvm] [AMDGPU] Extend DS loop wait optimization with flush point tracking (PR #175658)

[via llvm-commits]

================
@@ -3475,7 +3515,15 @@ SIInsertWaitcnts::getPreheaderFlushFlags(MachineLoop *ML,
   // are not used in the loop.
   // DSInvalidated is pre-set to true on non-GFX12+ targets where DS_CNT
   // is LGKM_CNT which also tracks FLAT/SMEM.
-  if (!DSInvalidated && UsesVgprLoadedOutsideDS)
+  // Pure prefetch: no same-iteration DS load use, no DS stores
+  bool PurePrefetchFlush = !DSInvalidated && UsesVgprLoadedOutsideDS;
+  // Prefetch with flush points: some DS loads used in same iteration,
+  // but unflushed loads remain at backedge
+  bool HasUnflushedDSLoads = (DSLoadCount > LastFlushedPosition);
+  bool FlushPointPrefetchFlush = !FlushPointTrackingInvalid &&
----------------
hidekisaito wrote:

Fair enough. Now that I got rid of "barrier in latch block" code, difference between the two modes narrowed quite a bit. Still, neither is a subset of the other. 1) PurePrefetch can work for multiple basic blocks. FlushPointPrefetch does not. 2) For a single block loop, PurePrefetch is a subset of FlushPointPrefetch, but the latter analysis is more expensive.

>FlushPointTracking is valid only if DS is valid

That is not the case. Same iteration use of a load flags DSInvalidated. FlushPoint analysis continues.

Anyway, your comment makes me think that I need to add more clarity to those two modes.

https://github.com/llvm/llvm-project/pull/175658