[Mlir-commits] [mlir] [MLIR][SROA] Replace pattern based approach with a one-shot one (PR #85437)

[Christian Ulmann]

================
@@ -223,12 +215,35 @@ struct SROA : public impl::SROABase<SROA> {
     SROAStatistics statistics{&destructuredAmount, &slotsWithMemoryBenefit,
                               &maxSubelementAmount};
 
-    RewritePatternSet rewritePatterns(&getContext());
-    rewritePatterns.add<SROAPattern>(&getContext(), statistics);
-    FrozenRewritePatternSet frozen(std::move(rewritePatterns));
+    bool changed = false;
+
+    for (Region &region : scopeOp->getRegions()) {
+      if (region.getBlocks().empty())
+        continue;
 
-    if (failed(applyPatternsAndFoldGreedily(scopeOp, frozen)))
-      signalPassFailure();
+      OpBuilder builder(&region.front(), region.front().begin());
+      IRRewriter rewriter(builder);
+
+      // Destructuring a slot can allow for further destructuring of other
+      // slots, destructuring is tried until no destructuring succeeds.
+      while (true) {
+        SmallVector<DestructurableAllocationOpInterface> allocators;
+        // Build a list of allocators to attempt to destructure the slots of.
+        for (Block &block : region)
+          for (Operation &op : block.getOperations())
+            if (auto allocator =
+                    dyn_cast<DestructurableAllocationOpInterface>(op))
+              allocators.emplace_back(allocator);
----------------
Dinistro wrote:

Good catch regarding the recursion difference. Will fix accordingly.

> Other thing: instead of reconstructing the allocators list at each iteration, can you modify tryToDestructureMemorySlots to actually preserve it up-to-date as we go?

We have substantial SROA improvements in the pipeline, I'll try to address this as part of the upcoming changes, if that is fine. 

https://github.com/llvm/llvm-project/pull/85437