[mlir] Add option for a cleanup pattern set to SCF tiling helper

The SCF helper for tiling an operation implementing the TilingInterface
and greedily fusing consumers requires an uninterrupted chain of
operations implementing the tiling interface to succeed. There can be
cases with intermediate ops that don't implement the interface but have
producers that could be fused if various canonicalization/simplification
patterns could run in between fusion steps.

This adds an option to SCFTileAndFuseOptions for a pattern set to run
between fusion steps to the ops that result from fusion/tiling. Removed
and newly inserted slices are tracked for continued fusion applications.

See this RFC for more discussion:
https://discourse.llvm.org/t/rfc-split-fusion-portions-of-the-tilinginterface-into-a-new-interface/81155

Author: qedawkins

mlir/include/mlir/Dialect/Linalg/TransformOps/LinalgTransformOps.td

@@ -284,18 +284,23 @@ def FuseOp : Op<Transform_Dialect, "structured.fuse",
   let description = [{
     Tiles the operations pointed to by the target handle and fuses their
     producers greedily using the options provided as attributes.
+
+    If `apply_cleanup` is true then slice canonicalization is applied between
+    fusion steps.
   }];
 
   let arguments =
     (ins TransformHandleTypeInterface:$target,
          DefaultValuedAttr<I64ArrayAttr, "{}">:$tile_sizes,
-         DefaultValuedAttr<I64ArrayAttr, "{}">:$tile_interchange);
+         DefaultValuedAttr<I64ArrayAttr, "{}">:$tile_interchange,
+         DefaultValuedAttr<BoolAttr, "false">:$apply_cleanup);
   let results = (outs TransformHandleTypeInterface:$transformed,
                       Variadic<TransformHandleTypeInterface>:$loops);
 
   let assemblyFormat = [{
     $target ($tile_sizes^)? (`interchange` $tile_interchange^)?
-    attr-dict `:` functional-type(operands, results)
+    (`apply_cleanup` `=` $apply_cleanup^)? attr-dict
+    `:` functional-type(operands, results)
   }];
   let hasVerifier = 1;
 }

mlir/include/mlir/Dialect/SCF/Transforms/TileUsingInterface.h

@@ -15,6 +15,7 @@
 #include "mlir/Interfaces/LoopLikeInterface.h"
 #include "mlir/Interfaces/TilingInterface.h"
 #include "mlir/Interfaces/ViewLikeInterface.h"
+#include "mlir/Rewrite/FrozenRewritePatternSet.h"
 
 #include <deque>
 
@@ -153,6 +154,11 @@ struct SCFTileAndFuseOptions {
     fusionControlFn = controlFn;
     return *this;
   }
+
+  /// An optional set of rewrite patterns to apply to the results of tiling
+  /// before fusion. This will track deleted and newly inserted
+  /// `tensor.extract_slice` ops and update the worklist.
+  std::optional<FrozenRewritePatternSet> cleanupPatterns = std::nullopt;
 };
 
 /// Fuse the producer of the source of `candidateSliceOp` by computing the

mlir/lib/Dialect/Linalg/TransformOps/LinalgTransformOps.cpp

@@ -557,6 +557,15 @@ transform::FuseOp::apply(transform::TransformRewriter &rewriter,
   tilingOptions = tilingOptions.setTileSizes(tileSizesOfr);
   scf::SCFTileAndFuseOptions tileAndFuseOptions;
   tileAndFuseOptions.tilingOptions = tilingOptions;
+
+  if (getApplyCleanup()) {
+    MLIRContext *context = rewriter.getContext();
+    RewritePatternSet patterns(context);
+    tensor::ExtractSliceOp::getCanonicalizationPatterns(patterns, context);
+    tensor::populateMergeConsecutiveInsertExtractSlicePatterns(patterns);
+    tileAndFuseOptions.cleanupPatterns = std::move(patterns);
+  }
+
   LogicalResult result = applyTilingToAll(
       rewriter, getOperation(), state.getPayloadOps(getTarget()),
       tileSizes.size() - llvm::count(tileSizes, 0), transformResults,

mlir/lib/Dialect/SCF/Transforms/TileUsingInterface.cpp

@@ -24,6 +24,8 @@
 #include "mlir/IR/PatternMatch.h"
 #include "mlir/Interfaces/DestinationStyleOpInterface.h"
 #include "mlir/Interfaces/TilingInterface.h"
+#include "mlir/Rewrite/FrozenRewritePatternSet.h"
+#include "mlir/Transforms/GreedyPatternRewriteDriver.h"
 #include "llvm/ADT/TypeSwitch.h"
 #include "llvm/Support/Debug.h"
 #include <optional>
@@ -1315,6 +1317,172 @@ FailureOr<SmallVector<Operation *>> mlir::scf::yieldReplacementForFusedProducer(
   return generatedSlices;
 }
 
+namespace {
+
+//===----------------------------------------------------------------------===//
+// SliceWorklist
+//===----------------------------------------------------------------------===//
+
+/// Struct for tracking the number of stale entries on the worklist and whether
+/// there is a remaining valid entry.
+struct EntryCount {
+  bool isValid = true;
+  unsigned count = 0;
+};
+
+/// A FIFO worklist of operations with efficient removal and set semantics.
+///
+/// This class maintains a queue of operations and a mapping of operations to
+/// positions in the vector, so that operations can be removed efficiently at
+/// random. When an operation is removed, it is replaced with nullptr. Such
+/// nullptr are skipped when pop'ing elements.
+///
+/// This is similar to the worklist used by the GreedyPatternRewriteDriver,
+/// except instead FIFO so that slices for fusion can be processed breadth
+/// first.
+class SliceWorklist {
+public:
+  SliceWorklist() = default;
+
+  /// Push an operation to the end of the worklist. This assumes that
+  /// the given operation is not already on the worklist.
+  void push(Operation *op);
+
+  /// Pop the an operation from the end of the worklist. Returns nullptr if
+  /// there are no remaining valid operations.
+  Operation *pop();
+
+  /// Remove an operation from the worklist.
+  void remove(Operation *op);
+
+protected:
+  /// The queue of operations.
+  std::deque<Operation *> list;
+
+  /// A mapping of operations to the number of stale copies in the queue.
+  DenseMap<Operation *, EntryCount> map;
+};
+
+void SliceWorklist::push(Operation *op) {
+  assert(op && "cannot push nullptr to worklist");
+  list.push_back(op);
+  EntryCount newCount = map.lookup(op);
+  // Because operations are only pushed on creation, valid duplicates are
+  // never added.
+  assert((!map.contains(op) || !newCount.isValid) &&
+         "cannot push a duplicate operation");
+  map[op] = {/*isValid=*/true, newCount.count + 1};
+}
+
+Operation *SliceWorklist::pop() {
+  // Pop the front of the queue until we hit a valid entry.
+  while (!list.empty()) {
+    Operation *op = list.front();
+    list.pop_front();
+
+    EntryCount e = map.lookup(op);
+    // If the entry count is greater than 1 or there is no valid entry,
+    // this must be a stale entry. Decrement the map entry by one and continue.
+    if (e.count > 1 || !e.isValid) {
+      int64_t newCount = e.count - 1;
+      if (newCount <= 0)
+        map.erase(op);
+      else
+        map[op] = {e.isValid, static_cast<unsigned int>(newCount)};
+      continue;
+    }
+
+    map.erase(op);
+    return op;
+  }
+  return nullptr;
+}
+
+// Mark the operation as invalid if present. Removal from the map will
+// happen later when popping from the worklist.
+void SliceWorklist::remove(Operation *op) {
+  if (!map.contains(op))
+    return;
+
+  EntryCount e = map.lookup(op);
+  map[op] = {/*isValid=*/false, e.count};
+}
+
+//===----------------------------------------------------------------------===//
+// SliceTrackingListener
+//===----------------------------------------------------------------------===//
+
+/// This class is a listener for tracking the insertion and removal of
+/// `tensor.extract_slice` ops in a worklist. This can be used in a greedy
+/// fusion algorithm to apply cleanup patterns in between fusion steps.
+class SliceTrackingListener : public RewriterBase::Listener {
+public:
+  explicit SliceTrackingListener(
+      std::optional<FrozenRewritePatternSet> patterns);
+  SliceTrackingListener() = default;
+
+  /// Adds the given list of operations to the worklist, and if present, applies
+  /// the list of `patterns` to the newly added operations. This only processes
+  /// the given operations and any newly inserted ones by the pattern set.
+  LogicalResult insertAndApplyPatterns(ArrayRef<Operation *> newOps);
+
+  /// Add to the new operation worklist if it is an extract_slice.
+  void notifyOperationInserted(Operation *op,
+                               OpBuilder::InsertPoint previous) override;
+
+  /// Remove the operation from the worklist.
+  void notifyOperationErased(Operation *op) override;
+
+  /// Remove the operation from the worklist.
+  void notifyOperationReplaced(Operation *op, ValueRange replacement) override;
+
+  /// The worklist for this transformation keeps track of the operations that
+  /// need to be (re)visited.
+  SliceWorklist worklist;
+
+private:
+  /// Optional pattern set to apply when adding new operations to the worklist.
+  std::optional<FrozenRewritePatternSet> patterns = std::nullopt;
+};
+
+SliceTrackingListener::SliceTrackingListener(
+    std::optional<FrozenRewritePatternSet> p) {
+  patterns = std::move(p);
+}
+
+LogicalResult
+SliceTrackingListener::insertAndApplyPatterns(ArrayRef<Operation *> ops) {
+  for (Operation *op : ops) {
+    if (isa<tensor::ExtractSliceOp>(op))
+      worklist.push(op);
+  }
+
+  if (!patterns)
+    return success();
+
+  GreedyRewriteConfig config;
+  config.listener = this;
+  config.strictMode = GreedyRewriteStrictness::ExistingAndNewOps;
+  return applyOpPatternsAndFold(ops, patterns.value(), config);
+}
+
+void SliceTrackingListener::notifyOperationInserted(
+    Operation *op, OpBuilder::InsertPoint previous) {
+  if (!isa<tensor::ExtractSliceOp>(op))
+    return;
+  worklist.push(op);
+}
+
+void SliceTrackingListener::notifyOperationErased(Operation *op) {
+  worklist.remove(op);
+}
+
+void SliceTrackingListener::notifyOperationReplaced(Operation *op,
+                                                    ValueRange replacement) {
+  worklist.remove(op);
+}
+} // namespace
+
 /// Implementation of tile consumer and fuse producer greedily.
 FailureOr<scf::SCFTileAndFuseResult>
 mlir::scf::tileConsumerAndFuseProducersUsingSCF(
@@ -1370,33 +1538,33 @@ mlir::scf::tileConsumerAndFuseProducersUsingSCF(
     tensor::ExtractSliceOp candidateSlice;
     SCFTileAndFuseOptions::ControlFnResult controlFnResult;
   };
-  std::deque<WorklistItem> worklist;
-  auto addCandidateSlices = [&worklist, &options,
-                             &loops](ArrayRef<Operation *> candidates) {
-    for (auto candidate : candidates) {
-      auto sliceOp = dyn_cast<tensor::ExtractSliceOp>(candidate);
-      if (!sliceOp || sliceOp.use_empty())
-        continue;
 
-      auto [fusableProducer, destinationInitArg] =
-          getUntiledProducerFromSliceSource(&sliceOp.getSourceMutable(), loops);
-      if (!fusableProducer)
-        continue;
-      std::optional<SCFTileAndFuseOptions::ControlFnResult> controlFnResult =
-          options.fusionControlFn(sliceOp, fusableProducer,
-                                  destinationInitArg.has_value());
-      if (!controlFnResult)
-        continue;
-      worklist.emplace_back(WorklistItem{sliceOp, controlFnResult.value()});
-    }
-  };
+  SliceTrackingListener sliceTracker =
+      SliceTrackingListener(options.cleanupPatterns);
 
-  addCandidateSlices(tilingResult->generatedSlices);
+  if (failed(
+          sliceTracker.insertAndApplyPatterns(tilingResult->generatedSlices))) {
+    return rewriter.notifyMatchFailure(consumer, "cleanup patterns failed");
+  }
   OpBuilder::InsertionGuard g(rewriter);
-  while (!worklist.empty()) {
-    // Traverse the slices in BFS fashion.
-    WorklistItem worklistItem = worklist.front();
-    worklist.pop_front();
+  while (Operation *next = sliceTracker.worklist.pop()) {
+    auto candidateSlice = dyn_cast<tensor::ExtractSliceOp>(next);
+    if (!candidateSlice)
+      continue;
+
+    auto [fusableProducer, destinationInitArg] =
+        getUntiledProducerFromSliceSource(&candidateSlice.getSourceMutable(),
+                                          loops);
+    if (!fusableProducer)
+      continue;
+
+    std::optional<SCFTileAndFuseOptions::ControlFnResult> controlFnResult =
+        options.fusionControlFn(candidateSlice, fusableProducer,
+                                destinationInitArg.has_value());
+    if (!controlFnResult)
+      continue;
+
+    WorklistItem worklistItem = {candidateSlice, controlFnResult.value()};
 
     // The operands of the fused producer might themselved be slices of
     // values produced by operations that implement the `TilingInterface`.
@@ -1407,6 +1575,8 @@ mlir::scf::tileConsumerAndFuseProducersUsingSCF(
     if (!fusedResult)
       continue;
 
+    SmallVector<Operation *> worklistCandidates = fusedResult->generatedSlices;
+
     if (worklistItem.controlFnResult.yieldProducerReplacement) {
       // Reconstruct and yield all opResult of fusableProducerOp by default. The
       // caller can specific which one to yield by designating optional argument
@@ -1421,20 +1591,23 @@ mlir::scf::tileConsumerAndFuseProducersUsingSCF(
             fusableProducerOp, "failed to replacement value for this "
                                "operation from within the tiled loop");
       }
-      addCandidateSlices(newSlices.value());
+      worklistCandidates.append(newSlices.value());
       for (auto [index, result] :
            llvm::enumerate(fusableProducerOp->getResults())) {
         origValToResultNumber[result] = loops.front()->getNumResults() -
                                         fusableProducerOp->getNumResults() +
                                         index;
       }
     }
-    addCandidateSlices(fusedResult->generatedSlices);
     if (Operation *tiledAndFusedOp =
             fusedResult->tiledAndFusedProducer.getDefiningOp()) {
       fusedProducers.insert(fusedResult->origProducer.getDefiningOp());
       tiledAndFusedOps.insert(tiledAndFusedOp);
     }
+
+    if (failed(sliceTracker.insertAndApplyPatterns(worklistCandidates))) {
+      return rewriter.notifyMatchFailure(consumer, "cleanup patterns failed");
+    }
   }
 
   DenseMap<Value, Value> replacements;