[mlir][bufferization] Add DeallocationSimplification pass

[maerhart]

Summary

Adds a pass that can be run after buffer deallocation to simplify the deallocation operations.
In particular, there are patterns that need alias information and thus cannot be added as a regular canonicalization pattern.
This initial commit moves an incorrect canonicalization pattern from over to this new pass and fixes it by querying the alias analysis for the additional information it needs to be correct (there must not by any potential aliasing memref in the retain list other than the currently mached one).
Also, improves this pattern by considering the extract_strided_metadata operation which is inserted by the deallocation pass by default.

[springerm]

mlir/lib/Dialect/Bufferization/Transforms/BufferDeallocationSimplification.cpp

58

The implementation is doing something with aliasing. Can we capture that in the description here.

67

Let's add an example that adds an arith.ori

97–99

Where is the NoAlias and where is the AlwaysAlias case in the code? There's just a single "if" check below, so it looks like both are handled the same way.

[maerhart]

Address comments. Thanks for reviewing!

mlir/lib/Dialect/Bufferization/Transforms/BufferDeallocationSimplification.cpp

67

This example actually adds an arith.ori, it was just missing. Ultimately, it will be canonicalized away by the other patterns though and all that remains of this initial dealloc is %arg1 (the condition) or %cond0 in the updated code.

[springerm]

mlir/lib/Dialect/Bufferization/Transforms/BufferDeallocationSimplification.cpp

59–60

let's mention that this is needed to compute the result of the op

66

I think this op should have 3 return values?

71

same here

71

Can you add a comment in front of that: bufferization.dealloc without memrefs and conditions returns %true for every retained value.

[maerhart]

Fix description.

mlir/lib/Dialect/Bufferization/Transforms/BufferDeallocationSimplification.cpp

71

Sure! However, it returns %false in that case because that is the identity element of OR and composes nicely in the situations we use dealloc in.
Typically we add all memrefs that have to potentially be deallocated in a given basic block to a single dealloc operation at the end of the block. If we want to split them into multiple dealloc operations, we also need to compute the element-wise disjunction of the results where %false composes nicely because it is the identity element. This also allows us to remove memrefs from the list which have constant false as condition.

[maerhart]

Rebase had messed up something.

mlir/include/mlir/Dialect/Bufferization/Transforms/Passes.h

 
 #define GEN_PASS_DECL
 #include "mlir/Dialect/Bufferization/Transforms/Passes.h.inc"
 
 /// Creates an instance of the BufferDeallocation pass to free all allocated
 /// buffers.
 std::unique_ptr<Pass> createBufferDeallocationPass();
 
+/// Creates a pass that optimizes `bufferization.dealloc` operations. For
+/// example, it reduces the number of alias checks needed at runtime using
+/// static alias analysis.
+std::unique_ptr<Pass> createBufferDeallocationSimplificationPass();
+
 /// Run buffer deallocation.
 LogicalResult deallocateBuffers(Operation *op);
 
 /// Creates a pass that moves allocations upwards to reduce the number of
 /// required copies that are inserted during the BufferDeallocation pass.
 std::unique_ptr<Pass> createBufferHoistingPass();
 
 /// Creates a pass that moves allocations upwards out of loops. This avoids

mlir/include/mlir/Dialect/Bufferization/Transforms/Passes.td

let description = [{
 
     }
     ```
 
   }];
   let constructor = "mlir::bufferization::createBufferDeallocationPass()";
 }
 
+def BufferDeallocationSimplification :
+    Pass<"buffer-deallocation-simplification", "func::FuncOp"> {
+  let summary = "Optimizes `bufferization.dealloc` operation for more "
+                "efficient codegen";
+  let description = [{
+    This pass uses static alias analysis to reduce the number of alias checks
+    required at runtime. Such checks are sometimes necessary to make sure that
+    memrefs aren't deallocated before their last usage (use after free) or that
+    some memref isn't deallocated twice (double free).
+  }];
+
+  let constructor =
+    "mlir::bufferization::createBufferDeallocationSimplificationPass()";
+
+  let dependentDialects = [
+    "mlir::bufferization::BufferizationDialect", "mlir::arith::ArithDialect",
+    "mlir::memref::MemRefDialect"
+  ];
+}
+
 def BufferHoisting : Pass<"buffer-hoisting", "func::FuncOp"> {
   let summary = "Optimizes placement of allocation operations by moving them "
                 "into common dominators and out of nested regions";
   let description = [{
     This pass implements an approach to aggressively move allocations upwards
     into common dominators and out of nested regions.
   }];
   let constructor = "mlir::bufferization::createBufferHoistingPass()";

mlir/lib/Dialect/Bufferization/IR/BufferizationOps.cpp

SmallVector<Value> replacements(
         llvm::map_range(resultReplacementIdx, [&](unsigned idx) {
           return newDeallocOp.getUpdatedConditions()[idx];
         }));
     rewriter.replaceOp(deallocOp, replacements);
     return success();
   }
 };
 
-/// Remove memrefs to be deallocated that are also present in the retained list
-/// since they will always alias and thus never actually be deallocated.
-/// Example:
-/// ```mlir
-/// %0 = bufferization.dealloc (%arg0 : ...) if (%arg1) retain (%arg0 : ...)
-/// ```
-/// is canonicalized to
-/// ```mlir
-/// %0 = bufferization.dealloc retain (%arg0 : ...)
-/// ```
-struct DeallocRemoveDeallocMemrefsContainedInRetained
-    : public OpRewritePattern<DeallocOp> {
-  using OpRewritePattern<DeallocOp>::OpRewritePattern;
-
-  LogicalResult matchAndRewrite(DeallocOp deallocOp,
-                                PatternRewriter &rewriter) const override {
-    // Unique memrefs to be deallocated.
-    DenseMap<Value, unsigned> retained;
-    for (auto [i, ret] : llvm::enumerate(deallocOp.getRetained()))
-      retained[ret] = i;
-
-    // There must not be any duplicates in the retain list anymore because we
-    // would miss updating one of the result values otherwise.
-    if (retained.size() != deallocOp.getRetained().size())
-      return failure();
-
-    SmallVector<Value> newMemrefs, newConditions;
-    for (auto [memref, cond] :
-         llvm::zip(deallocOp.getMemrefs(), deallocOp.getConditions())) {
-      if (retained.contains(memref)) {
-        rewriter.setInsertionPointAfter(deallocOp);
-        auto orOp = rewriter.create<arith::OrIOp>(
-            deallocOp.getLoc(),
-            deallocOp.getUpdatedConditions()[retained[memref]], cond);
-        rewriter.replaceAllUsesExcept(
-            deallocOp.getUpdatedConditions()[retained[memref]],
-            orOp.getResult(), orOp);
-        continue;
-      }
-
-      newMemrefs.push_back(memref);
-      newConditions.push_back(cond);
-    }
-
-    // Return failure if we don't change anything such that we don't run into an
-    // infinite loop of pattern applications.
-    return updateDeallocIfChanged(deallocOp, newMemrefs, newConditions,
-                                  rewriter);
-  }
-};
-
 /// Erase deallocation operations where the variadic list of memrefs to
 /// deallocate is empty. Example:
 /// ```mlir
 /// %0 = bufferization.dealloc retain (%arg0: memref<2xi32>)
 /// ```
 struct EraseEmptyDealloc : public OpRewritePattern<DeallocOp> {
   using OpRewritePattern<DeallocOp>::OpRewritePattern;
 
struct SkipExtractMetadataOfAlloc : public OpRewritePattern<DeallocOp> {
   }
 };
 
 } // anonymous namespace
 
 void DeallocOp::getCanonicalizationPatterns(RewritePatternSet &results,
                                             MLIRContext *context) {
   results.add<DeallocRemoveDuplicateDeallocMemrefs,
-              DeallocRemoveDuplicateRetainedMemrefs,
-              DeallocRemoveDeallocMemrefsContainedInRetained, EraseEmptyDealloc,
+              DeallocRemoveDuplicateRetainedMemrefs, EraseEmptyDealloc,
               EraseAlwaysFalseDealloc, SkipExtractMetadataOfAlloc>(context);
 }
 
 //===----------------------------------------------------------------------===//
 // TableGen'd op method definitions
 //===----------------------------------------------------------------------===//
 
 #define GET_OP_CLASSES
 #include "mlir/Dialect/Bufferization/IR/BufferizationOps.cpp.inc"

mlir/lib/Dialect/Bufferization/Transforms/BufferDeallocationSimplification.cpp

+//===- BufferDeallocationSimplification.cpp -------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements logic for optimizing `bufferization.dealloc` operations
+// that requires more analysis than what can be supported by regular
+// canonicalization patterns.
+//
+//===----------------------------------------------------------------------===//
+
+#include "mlir/Analysis/AliasAnalysis.h"
+#include "mlir/Dialect/Bufferization/IR/Bufferization.h"
+#include "mlir/Dialect/Bufferization/Transforms/Passes.h"
+#include "mlir/Dialect/Func/IR/FuncOps.h"
+#include "mlir/Dialect/MemRef/IR/MemRef.h"
+#include "mlir/Transforms/GreedyPatternRewriteDriver.h"
+
+namespace mlir {
+namespace bufferization {
+#define GEN_PASS_DEF_BUFFERDEALLOCATIONSIMPLIFICATION
+#include "mlir/Dialect/Bufferization/Transforms/Passes.h.inc"
+} // namespace bufferization
+} // namespace mlir
+
+using namespace mlir;
+using namespace mlir::bufferization;
+
+//===----------------------------------------------------------------------===//
+// Helpers
+//===----------------------------------------------------------------------===//
+
+static LogicalResult updateDeallocIfChanged(DeallocOp deallocOp,
+                                            ValueRange memrefs,
+                                            ValueRange conditions,
+                                            PatternRewriter &rewriter) {
+  if (deallocOp.getMemrefs() == memrefs &&
+      deallocOp.getConditions() == conditions)
+    return failure();
+
+  rewriter.updateRootInPlace(deallocOp, [&]() {
+    deallocOp.getMemrefsMutable().assign(memrefs);
+    deallocOp.getConditionsMutable().assign(conditions);
+  });
+  return success();
+}
+
+//===----------------------------------------------------------------------===//
+// Patterns
+//===----------------------------------------------------------------------===//
+
+namespace {
+
+/// Remove values from the `memref` operand list that are also present in the
+/// `retained` list since they will always alias and thus never actually be
+/// deallocated. However, we also need to be certain that no other value in the
+/// `retained` list can alias, for which we use a static alias analysis and, as
+/// a result, this pattern cannot be applied as a regular canonicalization
+/// pattern.
+///
+/// Example:
+/// ```mlir
+/// %0 = bufferization.dealloc (%m0 : ...) if (%cond0)
+///                     retain (%m0, %r0, %r1 : ...)
+/// ```
+/// is canonicalized to
+/// ```mlir
+/// %0 = bufferization.dealloc retain (%m0, %r0, %r1 : ...)
+/// %1 = arith.ori %0, %cond0 : i1
+/// ```
+/// given that `%r0` and `%r1` may not alias with `%m0`.
+struct DeallocRemoveDeallocMemrefsContainedInRetained
+    : public OpRewritePattern<DeallocOp> {
+  DeallocRemoveDeallocMemrefsContainedInRetained(MLIRContext *context,
+                                                 AliasAnalysis &aliasAnalysis)
+      : OpRewritePattern<DeallocOp>(context), aliasAnalysis(aliasAnalysis) {}
+
+  LogicalResult matchAndRewrite(DeallocOp deallocOp,
+                                PatternRewriter &rewriter) const override {
+    // Unique memrefs to be deallocated.
+    DenseMap<Value, unsigned> retained;
+    for (auto [i, ret] : llvm::enumerate(deallocOp.getRetained()))
+      retained[ret] = i;
+
+    // There must not be any duplicates in the retain list anymore because we
+    // would miss updating one of the result values otherwise.
+    if (retained.size() != deallocOp.getRetained().size())
+      return failure();
+
+    SmallVector<Value> newMemrefs, newConditions;
+    for (auto memrefAndCond :
+         llvm::zip(deallocOp.getMemrefs(), deallocOp.getConditions())) {
+      Value memref = std::get<0>(memrefAndCond);
+      Value cond = std::get<1>(memrefAndCond);
+
+      auto replaceResultsIfNoInvalidAliasing = [&](Value memref) -> bool {
+        Value retainedMemref = deallocOp.getRetained()[retained[memref]];
+        // The current memref must not have a may-alias relation to any retained
+        // memref, and exactly one must-alias relation.
+        // TODO: it is possible to extend this pattern to allow an arbitrary
+        // number of must-alias relations as long as there is no may-alias. If
+        // it's no-alias, then just proceed (only supported case as of now), if
+        // it's must-alias, we also need to update the condition for that alias.
+        if (llvm::all_of(deallocOp.getRetained(), [&](Value mr) {
+              return aliasAnalysis.alias(mr, memref).isNo() ||
+                     mr == retainedMemref;
+            })) {
+          rewriter.setInsertionPointAfter(deallocOp);
+          auto orOp = rewriter.create<arith::OrIOp>(
+              deallocOp.getLoc(),
+              deallocOp.getUpdatedConditions()[retained[memref]], cond);
+          rewriter.replaceAllUsesExcept(
+              deallocOp.getUpdatedConditions()[retained[memref]],
+              orOp.getResult(), orOp);
+          return true;
+        }
+        return false;
+      };
+
+      if (retained.contains(memref) &&
+          replaceResultsIfNoInvalidAliasing(memref))
+        continue;
+
+      auto extractOp = memref.getDefiningOp<memref::ExtractStridedMetadataOp>();
+      if (extractOp && retained.contains(extractOp.getOperand()) &&
+          replaceResultsIfNoInvalidAliasing(extractOp.getOperand()))
+        continue;
+
+      newMemrefs.push_back(memref);
+      newConditions.push_back(cond);
+    }
+
+    // Return failure if we don't change anything such that we don't run into an
+    // infinite loop of pattern applications.
+    return updateDeallocIfChanged(deallocOp, newMemrefs, newConditions,
+                                  rewriter);
+  }
+
+private:
+  AliasAnalysis &aliasAnalysis;
+};
+
+} // namespace
+
+//===----------------------------------------------------------------------===//
+// BufferDeallocationSimplificationPass
+//===----------------------------------------------------------------------===//
+
+namespace {
+
+/// The actual buffer deallocation pass that inserts and moves dealloc nodes
+/// into the right positions. Furthermore, it inserts additional clones if
+/// necessary. It uses the algorithm described at the top of the file.
+struct BufferDeallocationSimplificationPass
+    : public bufferization::impl::BufferDeallocationSimplificationBase<
+          BufferDeallocationSimplificationPass> {
+  void runOnOperation() override {
+    AliasAnalysis &aliasAnalysis = getAnalysis<AliasAnalysis>();
+    RewritePatternSet patterns(&getContext());
+    patterns.add<DeallocRemoveDeallocMemrefsContainedInRetained>(&getContext(),
+                                                                 aliasAnalysis);
+
+    if (failed(
+            applyPatternsAndFoldGreedily(getOperation(), std::move(patterns))))
+      signalPassFailure();
+  }
+};
+
+} // namespace
+
+std::unique_ptr<Pass>
+mlir::bufferization::createBufferDeallocationSimplificationPass() {
+  return std::make_unique<BufferDeallocationSimplificationPass>();
+}

mlir/lib/Dialect/Bufferization/Transforms/CMakeLists.txt

 add_mlir_dialect_library(MLIRBufferizationTransforms
   Bufferize.cpp
   BufferDeallocation.cpp
+  BufferDeallocationSimplification.cpp
   BufferOptimizations.cpp
   BufferResultsToOutParams.cpp
   BufferUtils.cpp
   BufferViewFlowAnalysis.cpp
   DropEquivalentBufferResults.cpp
   EmptyTensorElimination.cpp
   EmptyTensorToAllocTensor.cpp
   FuncBufferizableOpInterfaceImpl.cpp

mlir/test/Dialect/Bufferization/Transforms/buffer-deallocation-simplification.mlir

+// RUN: mlir-opt %s --buffer-deallocation-simplification --split-input-file | FileCheck %s
+
+func.func @dealloc_deallocated_in_retained(%arg0: memref<2xi32>, %arg1: i1, %arg2: memref<2xi32>) -> (i1, i1, i1, i1) {
+  %0 = bufferization.dealloc (%arg0 : memref<2xi32>) if (%arg1) retain (%arg0 : memref<2xi32>)
+  %1 = bufferization.dealloc (%arg0, %arg2 : memref<2xi32>, memref<2xi32>) if (%arg1, %arg1) retain (%arg0 : memref<2xi32>)
+  %2:2 = bufferization.dealloc (%arg0 : memref<2xi32>) if (%arg1) retain (%arg0, %arg2 : memref<2xi32>, memref<2xi32>)
+  return %0, %1, %2#0, %2#1 : i1, i1, i1, i1
+}
+
+// CHECK-LABEL: func @dealloc_deallocated_in_retained
+//  CHECK-SAME: ([[ARG0:%.+]]: memref<2xi32>, [[ARG1:%.+]]: i1, [[ARG2:%.+]]: memref<2xi32>)
+//  CHECK-NEXT: [[V0:%.+]] = bufferization.dealloc retain ([[ARG0]] : memref<2xi32>)
+//  CHECK-NEXT: [[O0:%.+]] = arith.ori [[V0]], [[ARG1]]
+//  CHECK-NEXT: [[V1:%.+]] = bufferization.dealloc ([[ARG2]] : memref<2xi32>) if ([[ARG1]]) retain ([[ARG0]] : memref<2xi32>)
+//  CHECK-NEXT: [[O1:%.+]] = arith.ori [[V1]], [[ARG1]]
+//  CHECK-NEXT: [[V2:%.+]]:2 = bufferization.dealloc ([[ARG0]] : memref<2xi32>) if ([[ARG1]]) retain ([[ARG0]], [[ARG2]] : memref<2xi32>, memref<2xi32>)
+//  CHECK-NEXT: return [[O0]], [[O1]], [[V2]]#0, [[V2]]#1 :
+
+// -----
+
+func.func @dealloc_deallocated_in_retained_extract_base_memref(%arg0: memref<2xi32>, %arg1: i1, %arg2: memref<2xi32>) -> (i1, i1, i1, i1) {
+  %base_buffer, %offset, %size, %stride = memref.extract_strided_metadata %arg0 : memref<2xi32> -> memref<i32>, index, index, index
+  %base_buffer0, %offset0, %size0, %stride0 = memref.extract_strided_metadata %arg2 : memref<2xi32> -> memref<i32>, index, index, index
+  %0 = bufferization.dealloc (%base_buffer : memref<i32>) if (%arg1) retain (%arg0 : memref<2xi32>)
+  %1 = bufferization.dealloc (%base_buffer, %base_buffer0 : memref<i32>, memref<i32>) if (%arg1, %arg1) retain (%arg0 : memref<2xi32>)
+  %2:2 = bufferization.dealloc (%base_buffer : memref<i32>) if (%arg1) retain (%arg0, %arg2 : memref<2xi32>, memref<2xi32>)
+  return %0, %1, %2#0, %2#1 : i1, i1, i1, i1
+}
+
+// CHECK-LABEL: func @dealloc_deallocated_in_retained_extract_base_memref
+//  CHECK-SAME: ([[ARG0:%.+]]: memref<2xi32>, [[ARG1:%.+]]: i1, [[ARG2:%.+]]: memref<2xi32>)
+//  CHECK-NEXT: [[BASE0:%[a-zA-Z0-9_]+]]{{.*}} = memref.extract_strided_metadata [[ARG0]] :
+//  CHECK-NEXT: [[BASE1:%[a-zA-Z0-9_]+]]{{.*}} = memref.extract_strided_metadata [[ARG2]] :
+//  CHECK-NEXT: [[V0:%.+]] = bufferization.dealloc retain ([[ARG0]] : memref<2xi32>)
+//  CHECK-NEXT: [[O0:%.+]] = arith.ori [[V0]], [[ARG1]]
+//  CHECK-NEXT: [[V1:%.+]] = bufferization.dealloc ([[BASE1]] : memref<i32>) if ([[ARG1]]) retain ([[ARG0]] : memref<2xi32>)
+//  CHECK-NEXT: [[O1:%.+]] = arith.ori [[V1]], [[ARG1]]
+//  CHECK-NEXT: [[V2:%.+]]:2 = bufferization.dealloc ([[BASE0]] : memref<i32>) if ([[ARG1]]) retain ([[ARG0]], [[ARG2]] : memref<2xi32>, memref<2xi32>)
+//  CHECK-NEXT: return [[O0]], [[O1]], [[V2]]#0, [[V2]]#1 :

mlir/test/Dialect/Bufferization/canonicalize.mlir

 //  CHECK-SAME:  ([[ARG0:%.+]]: memref<2xi32>, [[ARG1:%.+]]: i1, [[ARG2:%.+]]: i1, [[ARG3:%.+]]: memref<2xi32>, [[ARG4:%.+]]: memref<2xi32>, [[ARG5:%.+]]: memref<2xi32>)
 //  CHECK-NEXT:   [[V0:%.+]]:2 = bufferization.dealloc ([[ARG4]], [[ARG0]] : memref<2xi32>, memref<2xi32>) if ([[ARG1]], [[ARG1]]) retain ([[ARG3]], [[ARG5]] : memref<2xi32>, memref<2xi32>)
 //  CHECK-NEXT:   [[NEW_COND:%.+]] = arith.ori [[ARG1]], [[ARG2]] : i1
 //  CHECK-NEXT:   bufferization.dealloc ([[ARG0]] : memref<2xi32>) if ([[NEW_COND]])
 //  CHECK-NEXT:   return [[V0]]#0, [[V0]]#1, [[V0]]#0 :
 
 // -----
 
-func.func @dealloc_canonicalize_retained_and_deallocated(%arg0: memref<2xi32>, %arg1: i1, %arg2: memref<2xi32>) -> (i1, i1) {
-  %0 = bufferization.dealloc (%arg0 : memref<2xi32>) if (%arg1) retain (%arg0 : memref<2xi32>)
-  %1 = bufferization.dealloc (%arg0, %arg2 : memref<2xi32>, memref<2xi32>) if (%arg1, %arg1) retain (%arg0 : memref<2xi32>)
+func.func @dealloc_erase_empty(%arg0: memref<2xi32>, %arg1: i1, %arg2: memref<2xi32>) -> i1 {
   bufferization.dealloc
-  bufferization.dealloc retain (%arg0 : memref<2xi32>)
-  return %0, %1 : i1, i1
+  %0 = bufferization.dealloc retain (%arg0 : memref<2xi32>)
+  return %0 : i1
 }
 
-// CHECK-LABEL: func @dealloc_canonicalize_retained_and_deallocated
+// CHECK-LABEL: func @dealloc_erase_empty
 //  CHECK-SAME: ([[ARG0:%.+]]: memref<2xi32>, [[ARG1:%.+]]: i1, [[ARG2:%.+]]: memref<2xi32>)
-//  CHECK-NEXT: [[V0:%.+]] = bufferization.dealloc ([[ARG2]] : memref<2xi32>) if ([[ARG1]]) retain ([[ARG0]] : memref<2xi32>)
-//  CHECK-NEXT: [[V1:%.+]] = arith.ori [[V0]], [[ARG1]]
-//  CHECK-NEXT: return [[ARG1]], [[V1]] :
+//  CHECK-NEXT: [[FALSE:%.+]] = arith.constant false
+//  CHECK-NEXT: return [[FALSE]] :
 
 // -----
 
 func.func @dealloc_always_false_condition(%arg0: memref<2xi32>, %arg1: memref<2xi32>, %arg2: i1) {
   %false = arith.constant false
   bufferization.dealloc (%arg0, %arg1 : memref<2xi32>, memref<2xi32>) if (%false, %arg2)
   return
 }