[Mlir-commits] [mlir] [MLIR][Affine] Fix fusion across ops with unknown memory effects (PR #203231)

[Aditya Pradhan]

================
@@ -83,7 +122,7 @@ unsigned Node::getLoadOpCount(Value memref) const {
     if (auto affineLoad = dyn_cast<AffineReadOpInterface>(loadOp)) {
       if (memref == affineLoad.getMemRef())
         ++loadOpCount;
-    } else if (hasEffect<MemoryEffects::Read>(loadOp, memref)) {
+    } else if (mayHaveEffect<MemoryEffects::Read>(loadOp, memref)) {
       ++loadOpCount;
     }
   }
----------------
pradhanaditya wrote:

Note that the `Node` here is just `MemRefDependenceGraph::Node` (see the using declaration). I think Node already has what it needs here --  it owns its loads/stores (and memrefLoads/…) lists, so it can answer these without reaching outside itself. Passing the MemRefDependenceGraph (which owns the Nodes) back into a Node method couples them in the wrong direction, and if we did memoize, a Node-local field would be the natural place, not the MDG. But either way it would need invalidation/update as fusion mutates the graph, which isn't worth it for a negligible recompute. So, I would rather not add more complexity to the code.

Thanks for thinking it through.

https://github.com/llvm/llvm-project/pull/203231