[mlir][linalg][bufferize][NFC] Clean up comments and minor code refactorings

[springerm]

Summary

Depends On D116450

[springerm]

update

mlir/include/mlir/Dialect/Linalg/ComprehensiveBufferize/BufferizableOpInterface.h

struct AllocationCallbacks {
   /// A function that copies memory between two allocations.
   MemCpyFn memCpyFn;
 };
 
 /// Return default allocation callbacks.
 std::unique_ptr<AllocationCallbacks> defaultAllocationCallbacks();
 
 /// PostAnalysisSteps can be registered with `BufferizationOptions` and are
-/// executed after the analysis, but before bufferization. They can be used
+/// executed after the analysis, but before bufferization. They can be used to
 /// implement custom dialect-specific optimizations.
 struct PostAnalysisStep {
   virtual ~PostAnalysisStep() {}
 
   /// Run the post analysis step. This function may modify the IR, but must keep
-  /// `aliasInfo` (inside `state`) consistent. Newly created operations and
-  /// operations that should be re-analyzed must be stored in `newOps`.
+  /// `aliasInfo` consistent. Newly created operations and operations that
+  /// should be re-analyzed must be added to `newOps`.
   virtual LogicalResult run(Operation *op, BufferizationState &state,
                             BufferizationAliasInfo &aliasInfo,
                             SmallVector<Operation *> &newOps) = 0;
 };
 
 using PostAnalysisStepList = std::vector<std::unique_ptr<PostAnalysisStep>>;
 
 /// Options for ComprehensiveBufferize.
struct BufferizationOptions {
   /// Return `true` if the op is allowed to be bufferized.
   bool isOpAllowed(Operation *op) const {
     if (!dialectFilter.hasValue())
       return true;
     return dialectFilter->contains(op->getDialect()->getNamespace());
   }
 
   /// Allow-list the given dialects in the dialect filter. Only ops from
-  /// allow-listed dialects will be bufferized.
+  /// allow-listed dialects will be bufferized. If no dialect is added, ops from
+  /// any dialect will be bufferized.
   template <typename... DialectTs>
   void addToDialectFilter() {
     // The following expands a call to addToDialectFilterImpl for each dialect
     // in 'DialectTs'. This magic is necessary due to a limitation in the places
     // that a parameter pack can be expanded in c++11.
     // FIXME: In c++17 this can be simplified by using 'fold expressions'.
     (void)std::initializer_list<int>{
         0, (addToDialectFilterImpl<DialectTs>(), 0)...};
struct DialectBufferizationState {
 
   virtual ~DialectBufferizationState() = default;
 
   // Copying state is forbidden. Always pass as reference.
   DialectBufferizationState(const DialectBufferizationState &) = delete;
 };
 
 /// BufferizationState provides a variety of helper functions for dealing with
-/// tensor values and memref buffers. In particular,
-/// `BufferizableOpInterface::bufferize` implementation should utilize the
-/// following helper functions.
-///
-/// * `createAlloc` / `createDealloc` / `createAllocDeallocPair` creates ops
-///   that allocate and/or deallocate memref buffers.
-/// * `lookupBuffer` returns the memref buffer of a given tensor value.
-/// * `getResultBuffer` returns the memref buffer for a given tensor OpResult.
-///   Based on inplace bufferization decisions of the analysis, it may either
-///   directly return a mapped buffer or allocate a new brand new buffer.
-/// * `replaceOp` replaces an op with new values.
+/// tensor values and memref buffers.
 class BufferizationState {
 public:
   BufferizationState(Operation *op, const BufferizationOptions &options);
 
   // BufferizationState should be passed as a reference.
   BufferizationState(const BufferizationState &) = delete;
 
   /// Determine which OpOperand* will alias with `result` if the op is
public:
   Value lookupBuffer(RewriterBase &rewriter, Value tensor);
 
   /// Return `true` if the given OpResult has been decided to bufferize inplace.
   bool isInPlace(OpResult opResult) const;
 
   /// Return the result buffer (memref) for a given OpResult (tensor). Allocate
   /// a new buffer and copy over data from the existing buffer if out-of-place
   /// bufferization is necessary.
-  Value getResultBuffer(RewriterBase &rewriter, OpResult result);
+  FailureOr<Value> getResultBuffer(RewriterBase &rewriter, OpResult result);
 
   /// Return dialect-specific bufferization state.
   template <typename StateT> StateT &getDialectState(StringRef name) {
     // Create state if it does not exist yet.
     if (!dialectState.count(name))
       dialectState[name] = std::make_unique<StateT>();
     return static_cast<StateT &>(*dialectState[name]);
   }

 
 /// Return a contiguous MemRefType (i.e. with canonical/empty layout map)
 /// with the same shape as `shapedType` and specified `layout` and
 /// `addressSpace`.
 MemRefType getContiguousMemRefType(ShapedType shapedType,
                                    MemRefLayoutAttrInterface layout = {},
                                    Attribute memorySpace = {});
 
-/// Return a contiguous MemRefType (i.e. with canonical/empty layout map)
-/// with the same shape as `shapedType` and specified `layout` and
-/// `addressSpace` or an UnrankedMemRefType otherwise.
-Type getContiguousOrUnrankedMemRefType(Type type,
-                                       MemRefLayoutAttrInterface layout = {},
+/// Return an UnrankedMemRefType with the given element type and memory space.
+UnrankedMemRefType getUnrankedMemRefType(Type elementType,
                                          Attribute memorySpace = {});
 
 /// Return a MemRefType to which the `tensorType` can be bufferized in a
 /// composable fashion. The layout must be the most dynamic possible and
 /// canonicalize away once bufferization is finished.
 MemRefType getDynamicMemRefType(RankedTensorType tensorType,
                                 unsigned addressSpace = 0);
 
 } // namespace comprehensive_bufferize
: public BufferizableOpInterface::ExternalModel<
           AllocationHoistingBarrierOnly<OpTy>, OpTy> {
   bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand,
                               BufferizationState &state) const {
     return true;
   }
 
   bool bufferizesToMemoryWrite(Operation *op, OpOperand &opOperand,
                                BufferizationState &state) const {
-    return false;
+    return true;
   }
 
   SmallVector<OpOperand *>
   getAliasingOpOperand(Operation *op, OpResult opResult,
                        BufferizationState &state) const {
     return {};
   }
 

mlir/include/mlir/Dialect/Linalg/ComprehensiveBufferize/LinalgInterfaceImpl.h

 namespace linalg {
 namespace comprehensive_bufferize {
 
 class BufferizationAliasInfo;
 
 namespace linalg_ext {
 
 struct InitTensorEliminationStep : public PostAnalysisStep {
+  /// A function that matches anchor OpOperands for InitTensorOp elimination.
+  using AnchorMatchFn = std::function<bool(OpOperand &)>;
+
+  /// A function that rewrites matched anchors.
+  using RewriteFn = std::function<Value(OpBuilder &, Location, OpOperand &)>;
+
   /// Try to eliminate InitTensorOps inside `op`.
   ///
   /// * `rewriteFunc` generates the replacement for the InitTensorOp.
   /// * Only InitTensorOps that are anchored on a matching OpOperand as per
   ///   `anchorMatchFunc` are considered. "Anchored" means that there is a path
   ///   on the reverse SSA use-def chain, starting from the OpOperand and always
   ///   following the aliasing  OpOperand, that eventually ends at a single
   ///   InitTensorOp.
   /// * The result of `rewriteFunc` must usually be analyzed for inplacability.
   ///   This analysis can be skipped with `skipAnalysis`.
-  LogicalResult eliminateInitTensors(
-      Operation *op, BufferizationState &state,
+  LogicalResult eliminateInitTensors(Operation *op, BufferizationState &state,
                                      BufferizationAliasInfo &aliasInfo,
-      std::function<bool(OpOperand &)> anchorMatchFunc,
-      std::function<Value(OpBuilder &, Location, OpOperand &)> rewriteFunc,
+                                     AnchorMatchFn anchorMatchFunc,
+                                     RewriteFn rewriteFunc,
                                      SmallVector<Operation *> &newOps);
 };
 
 /// Try to eliminate InitTensorOps inside `op` that are anchored on an
 /// InsertSliceOp, i.e., if it is eventually inserted into another tensor
 /// (and some other conditions are met).
 struct InsertSliceAnchoredInitTensorEliminationStep
     : public InitTensorEliminationStep {
   LogicalResult run(Operation *op, BufferizationState &state,

mlir/lib/Dialect/Linalg/ComprehensiveBufferize/AffineInterfaceImpl.cpp

 //===- AffineInterfaceImpl.cpp - Affine Impl. of BufferizableOpInterface --===//
 //
 // 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
 //
 //===----------------------------------------------------------------------===//
 
 #include "mlir/Dialect/Linalg/ComprehensiveBufferize/AffineInterfaceImpl.h"
 
 #include "mlir/Dialect/Affine/IR/AffineOps.h"
 #include "mlir/Dialect/Linalg/ComprehensiveBufferize/BufferizableOpInterface.h"
 
 void mlir::linalg::comprehensive_bufferize::affine_ext::
     registerBufferizableOpInterfaceExternalModels(DialectRegistry &registry) {
+  // AffineParallelOp bufferization not implemented yet. However, never hoist
+  // memref allocations across AffineParallelOp boundaries.
   registry.addOpInterface<AffineParallelOp,
                           AllocationHoistingBarrierOnly<AffineParallelOp>>();
 }

mlir/lib/Dialect/Linalg/ComprehensiveBufferize/ArithInterfaceImpl.cpp

 #include "mlir/IR/Operation.h"
 #include "mlir/Transforms/BufferUtils.h"
 
 namespace mlir {
 namespace linalg {
 namespace comprehensive_bufferize {
 namespace arith_ext {
 
+/// Bufferization of arith.constant. Replace with memref.get_global.
 struct ConstantOpInterface
     : public BufferizableOpInterface::ExternalModel<ConstantOpInterface,
                                                     arith::ConstantOp> {
   LogicalResult bufferize(Operation *op, RewriterBase &rewriter,
                           BufferizationState &state) const {
     auto constantOp = cast<arith::ConstantOp>(op);
-    assert(constantOp.getType().dyn_cast<RankedTensorType>() &&
-           "not a constant ranked tensor");
+
+    // Only ranked tensors are supported.
+    if (!constantOp.getType().isa<RankedTensorType>())
+      return failure();
+
+    // Only constants inside a module are supported.
     auto moduleOp = constantOp->getParentOfType<ModuleOp>();
     if (!moduleOp)
-      return constantOp.emitError(
-          "cannot bufferize constants not within builtin.module op");
+      return failure();
 
+    // Create global memory segment and replace tensor with memref pointing to
+    // that memory segment.
     GlobalCreator globalCreator(moduleOp);
     auto globalMemref = globalCreator.getGlobalFor(constantOp);
     state.replaceOpWithNewBufferizedOp<memref::GetGlobalOp>(
         rewriter, op, globalMemref.type(), globalMemref.getName());
+
     return success();
   }
 
   bool isWritable(Operation *op, Value value, BufferizationState &state) const {
     // Memory locations returned by memref::GetGlobalOp may not be written to.
     assert(value.isa<OpResult>());
     return false;
   }

mlir/lib/Dialect/Linalg/ComprehensiveBufferize/BufferizableOpInterface.cpp

return std::make_unique<AllocationCallbacks>(
       defaultAllocationFn, defaultDeallocationFn, defaultMemCpyFn);
 }
 
 // Default constructor for BufferizationOptions that sets all allocation
 // callbacks to their default functions.
 BufferizationOptions::BufferizationOptions()
     : allocationFns(defaultAllocationCallbacks()) {}
 
+BufferizableOpInterface mlir::linalg::comprehensive_bufferize::
+    BufferizationOptions::dynCastBufferizableOp(Operation *op) const {
+  if (isOpAllowed(op))
+    return dyn_cast<BufferizableOpInterface>(op);
+  return nullptr;
+}
+
+BufferizableOpInterface mlir::linalg::comprehensive_bufferize::
+    BufferizationOptions::dynCastBufferizableOp(Value value) const {
+  if (auto bufferizableOp = value.getDefiningOp<BufferizableOpInterface>())
+    if (isOpAllowed(bufferizableOp.getOperation()))
+      return bufferizableOp;
+  return nullptr;
+}
+
 //===----------------------------------------------------------------------===//
 // BufferizationAliasInfo
 //===----------------------------------------------------------------------===//
 
 BufferizationAliasInfo::BufferizationAliasInfo(Operation *rootOp) {
   rootOp->walk([&](Operation *op) {
     for (Value v : op->getResults())
       if (v.getType().isa<TensorType>())

 static void setInsertionPointAfter(OpBuilder &b, Value value) {
   if (auto bbArg = value.dyn_cast<BlockArgument>()) {
     b.setInsertionPointToStart(bbArg.getOwner());
   } else {
     b.setInsertionPointAfter(value.getDefiningOp());
   }
 }
 
-BufferizableOpInterface mlir::linalg::comprehensive_bufferize::
-    BufferizationOptions::dynCastBufferizableOp(Operation *op) const {
-  if (isOpAllowed(op))
-    return dyn_cast<BufferizableOpInterface>(op);
-  return nullptr;
-}
-
-BufferizableOpInterface mlir::linalg::comprehensive_bufferize::
-    BufferizationOptions::dynCastBufferizableOp(Value value) const {
-  if (auto bufferizableOp = value.getDefiningOp<BufferizableOpInterface>())
-    if (isOpAllowed(bufferizableOp.getOperation()))
-      return bufferizableOp;
-  return nullptr;
-}
-
 /// Determine which OpOperand* will alias with `result` if the op is bufferized
 /// in place. Return an empty vector if the op is not bufferizable.
 SmallVector<OpOperand *>
 mlir::linalg::comprehensive_bufferize::BufferizationState::getAliasingOpOperand(
     OpResult result) {
   if (Operation *op = result.getDefiningOp())
     if (auto bufferizableOp = dyn_cast<BufferizableOpInterface>(op))
       return bufferizableOp.getAliasingOpOperand(result, *this);
op->walk([&](BufferizableOpInterface bufferizableOp) {
     }
     return WalkResult::advance();
   });
 }
 
 /// Return the result buffer (memref) for a given OpResult (tensor). Allocate
 /// a new buffer and copy over data from the existing buffer if out-of-place
 /// bufferization is necessary.
-Value mlir::linalg::comprehensive_bufferize::BufferizationState::
-    getResultBuffer(RewriterBase &rewriter, OpResult result) {
+FailureOr<Value>
+mlir::linalg::comprehensive_bufferize::BufferizationState::getResultBuffer(
+    RewriterBase &rewriter, OpResult result) {
   OpBuilder::InsertionGuard guard(rewriter);
   Operation *op = result.getOwner();
   SmallVector<OpOperand *> aliasingOperands = getAliasingOpOperand(result);
   assert(!aliasingOperands.empty() && "could not get aliasing OpOperand");
   OpOperand *opOperand = aliasingOperands.front();
   Value operand = opOperand->get();
   Value operandBuffer = lookupBuffer(rewriter, operand);
   // Make sure that all OpOperands are the same buffer. If this is not the case,
   // we would have to materialize a memref value.
   // TODO: Should be looking for checking for "equivalent buffers" instead of
   // operator== here, but equivalent buffers for scf.if yield values are not
   // set up yet.
   if (aliasingOperands.size() > 1 &&
       !llvm::all_of(aliasingOperands, [&](OpOperand *o) {
         return lookupBuffer(rewriter, o->get()) == operandBuffer;
-      })) {
-    op->emitError("result buffer is ambiguous");
-    return Value();
-  }
+      }))
+    return FailureOr<Value>(op->emitError("result buffer is ambiguous"));
 
   // If bufferizing out-of-place, allocate a new buffer.
   if (!aliasInfo.isInPlace(result)) {
     // Ops with multiple aliasing operands can currently not bufferize
     // out-of-place.
     assert(
         aliasingOperands.size() == 1 &&
         "ops with multiple aliasing OpOperands cannot bufferize out-of-place");
Value mlir::linalg::comprehensive_bufferize::BufferizationState::lookupBuffer(
 
   // Replace "%t = to_tensor %m" with %m.
   if (auto toTensorOp = tensor.getDefiningOp<bufferization::ToTensorOp>())
     return toTensorOp.memref();
 
   // Insert to_memref op.
   OpBuilder::InsertionGuard g(rewriter);
   setInsertionPointAfter(rewriter, tensor);
-  Type memrefType =
-      tensor.getType().isa<RankedTensorType>()
-          ? getDynamicMemRefType(tensor.getType().cast<RankedTensorType>())
-          : getContiguousOrUnrankedMemRefType(tensor.getType());
+  Type memrefType;
+  if (auto rankedTensorType = tensor.getType().dyn_cast<RankedTensorType>()) {
+    memrefType = getDynamicMemRefType(rankedTensorType);
+  } else {
+    memrefType = getUnrankedMemRefType(
+        tensor.getType().cast<TensorType>().getElementType());
+  }
   return rewriter.create<bufferization::ToMemrefOp>(tensor.getLoc(), memrefType,
                                                     tensor);
 }
 
 bool mlir::linalg::comprehensive_bufferize::BufferizationState::isInPlace(
     OpResult opResult) const {
   return aliasInfo.isInPlace(opResult);
 }
 
 MemRefType mlir::linalg::comprehensive_bufferize::getContiguousMemRefType(
     ShapedType shapedType, MemRefLayoutAttrInterface layout,
     Attribute memorySpace) {
   return MemRefType::get(shapedType.getShape(), shapedType.getElementType(),
                          layout, memorySpace);
 }
 
-Type mlir::linalg::comprehensive_bufferize::getContiguousOrUnrankedMemRefType(
-    Type type, MemRefLayoutAttrInterface layout, Attribute memorySpace) {
-  if (type.isa<RankedTensorType, MemRefType>())
-    return getContiguousMemRefType(type.cast<ShapedType>(), layout,
-                                   memorySpace);
-  assert(!layout && "expected empty layout with UnrankedMemRefType");
-  return UnrankedMemRefType::get(getElementTypeOrSelf(type), memorySpace);
+UnrankedMemRefType mlir::linalg::comprehensive_bufferize::getUnrankedMemRefType(
+    Type elementType, Attribute memorySpace) {
+  return UnrankedMemRefType::get(elementType, memorySpace);
 }
 
 MemRefType mlir::linalg::comprehensive_bufferize::getDynamicMemRefType(
     RankedTensorType tensorType, unsigned addressSpace) {
   // TODO: address space decisions to connect with the actual alloc.
   int64_t dynamicOffset = ShapedType::kDynamicStrideOrOffset;
   SmallVector<int64_t> dynamicStrides(tensorType.getRank(),
                                       ShapedType::kDynamicStrideOrOffset);
   AffineMap stridedLayout = makeStridedLinearLayoutMap(
       dynamicStrides, dynamicOffset, tensorType.getContext());
   return MemRefType::get(tensorType.getShape(), tensorType.getElementType(),
                          stridedLayout, addressSpace);
 }

mlir/lib/Dialect/Linalg/ComprehensiveBufferize/BufferizationInterfaceImpl.cpp

 namespace mlir {
 namespace linalg {
 namespace comprehensive_bufferize {
 namespace bufferization_ext {
 
 // TODO: These ops should implement BufferizableOpInterface directly when moved
 // to the Bufferization dialect.
 
+/// Bufferization of bufferization.to_memref. to_memref(to_tensor(x)) is folded
+/// to x. Other to_memref ops are ignored during bufferization.
+///
 /// ToMemrefOp casts a tensor into a memref. The resulting memref is the memory
 /// location of the incoming tensor once it will be bufferized. In the anlysis,
 /// the incoming tensor is assumed to bufferize to a memory read and to an
 /// inplace memory write, since it is unknown what will happen to the resulting
 /// memref.
 ///
 /// Note: ToMemrefOp / ToTensorOp are temporary ops that are inserted at the
 /// bufferization boundary. When bufferization is complete, there should be no
 /// such ops left over. If `allowUnknownOps`, such ops may be part of the
 /// resulting IR, but such IR may no longer be bufferizable by Comprehensive
 /// Bufferize.
 struct ToMemrefOpInterface
     : public BufferizableOpInterface::ExternalModel<ToMemrefOpInterface,
                                                     bufferization::ToMemrefOp> {
   bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand,
                               BufferizationState &state) const {
-    // It is unknown whether the resulting MemRef will be read or not.
+    // It is unknown whether the resulting memref will be read or not.
     return true;
   }
 
   OpResult getAliasingOpResult(Operation *op, OpOperand &opOperand,
                                BufferizationState &state) const {
     return OpResult();
   }
 
   LogicalResult bufferize(Operation *op, RewriterBase &rewriter,
                           BufferizationState &state) const {
     auto toMemrefOp = cast<bufferization::ToMemrefOp>(op);
 
     // Fold to_memref(to_tensor(x)) to x. Insert a cast if necessary.
     if (auto toTensorOp =
             toMemrefOp.tensor().getDefiningOp<bufferization::ToTensorOp>()) {
       Value buffer = toTensorOp.memref();
+
+      // Insert cast in case to_memref(to_tensor(x))'s type is different from
+      // x's type.
       if (toTensorOp.memref().getType() != toMemrefOp.getType())
         buffer = rewriter.create<memref::CastOp>(toMemrefOp.getLoc(), buffer,
                                                  toMemrefOp.getType());
-      rewriter.replaceOp(toMemrefOp, buffer);
+      state.replaceOpWithBufferizedValues(rewriter, toMemrefOp, buffer);
       return success();
     }
 
     return failure();
   }
 };
 
-/// ToTensorOp conceptually loads a tensor from a memory location. Such ops do
-/// not lower any further, and they should have disappeared by the time the
-/// input is fully bufferized.
+/// Bufferization of bufferization.to_tensor. Such ops cannot be bufferized.
+/// However, other ops that are using to_tensor's result will eventually be
+/// bufferized. At that point, they will start using to_tensor's memref operand.
+/// Once all users of to_tensor are bufferized, the op will not have any users
+/// anymore and DCE away.
 ///
-/// The analysis has no information about the memref that is loaded from by the
-/// ToTensorOp. We have to assume that the loaded tensor may after bufferization
-/// potentially alias with any other bufferized tensor. Since ToTensorOp and
-/// ToMemrefOp have no aliasing OpOperand/OpResult pairs, this cannot be encoded
-/// directly in the analysis. However, declaring ToTensorOp results as not
-/// writable also enforces a buffer copy and has the same effect.
+/// ToTensorOp conceptually loads a tensor from a memory location. The analysis
+/// has no information about the memref that is loaded from by ToTensorOp. We
+/// have to assume that the loaded tensor may after bufferization potentially
+/// alias with any other bufferized tensor. Since ToTensorOp and ToMemrefOp have
+/// no aliasing OpOperand/OpResult pairs, this cannot be encoded directly in the
+/// analysis. However, declaring ToTensorOp results as not writable enforces a
+/// buffer copy and has the same effect.
 struct ToTensorOpInterface
     : public BufferizableOpInterface::ExternalModel<ToTensorOpInterface,
                                                     bufferization::ToTensorOp> {
   LogicalResult bufferize(Operation *op, RewriterBase &rewriter,
                           BufferizationState &state) const {
     return failure();
   }
 
   bool isWritable(Operation *op, Value value, BufferizationState &state) const {
-    // It is unknown whether the MemRef operand is writable or not.
+    // It is unknown whether the memref operand is writable or not.
     return false;
   }
 };
 
 } // namespace bufferization_ext
 } // namespace comprehensive_bufferize
 } // namespace linalg
 } // namespace mlir
 
 void mlir::linalg::comprehensive_bufferize::bufferization_ext::
     registerBufferizableOpInterfaceExternalModels(DialectRegistry &registry) {
   registry.addOpInterface<bufferization::ToMemrefOp,
                           bufferization_ext::ToMemrefOpInterface>();
   registry.addOpInterface<bufferization::ToTensorOp,
                           bufferization_ext::ToTensorOpInterface>();
 }

mlir/lib/Dialect/Linalg/ComprehensiveBufferize/ComprehensiveBufferize.cpp

 //===- ComprehensiveBufferize.cpp - Single pass bufferization -------------===//
 //
 // 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
 //
 //===----------------------------------------------------------------------===//
 //
-// Perform inplace bufferization within function boundaries.
-// This is a specialized pass that supports inplace analysis for a fixed subset
-// of ops that have well-defined inplace semantics.
+// Comprehensive Bufferize bufferizes function bodies. Function boundaries
+// (FuncOp bbArgs, CallOps, ReturnOps) are treated as "unknown" ops.
+// ModuleBufferization.cpp is an extension of Comprehensive Bufferize for simple
+// call graphs.
+//
+// Comprehensive Bufferize consists of two phases.
+//
+// 1. Analyze ops to decide which OpResults can bufferize inplace, i.e., without
+//    inserting buffer copies. The analysis queries op bufferization semantics
+//    via `BufferizableOpInterface`.
+// 2. Bufferize ops by calling `BufferizableOpInterface::bufferize`. This
+//    function does not generate buffer copies for OpResults that were decided
+//    to bufferize inplace during the analysis phase.
+//
+// Inplace bufferization decisions are passed from the analysis to the
+// bufferization phase via `BufferizationState` and `BufferizationAliasInfo`.
+// They can be printed for debugging purposes with `testAnalysisOnly`.
+//
+// Ops that do not implement `BufferizableOpInterface` can be analyzed but are
+// treated conservatively. E.g., the analysis has to assume that their
+// OpOperands bufferize to memory writes. While such ops can be analyzed, they
+// are not bufferized and remain in the IR. to_tensor and to_memref ops are
+// inserted at the bufferization boundary.
+//
+// Note: If `allowUnknownOps` is set to false, bufferization fails when an
+// unknown op (that does not implement `BufferizableOpInterface`) is found. No
+// to_tensor/to_memref ops are inserted.
+//
 // This pass caters to high-performance codegen where buffer reuse is deemed
 // critical: the pass should fail if the bufferized form of the function needs
-// to return any buffer.
-// Generic control-flow and branching are unsupported.
-// Composability with extensible set of ops is not a first-class concern.
-//
-// Bufferization occurs by:
-//  a. performing an inPlace analysis `inPlaceAnalysis` which marks each
-//     operation within the op with the `kInPlaceResultsAttrName` attribute.
-//  b. traversing each operation in the op and rewriting it in
-//     buffer form and keeping a BlockAndValueMapping mapping of the
-//     rewrites. New allocations are introduced during this step.
-//     TODO: Allocation + depending op hoisting to outermost enclosing
-//     sequential scope.
-//  c. at the end of this bufferization, 3 cases may occur:
-//     i. inplaceable function arguments may be reused in place after the
-//        function itself has been bufferized. This is encoded by IR resembling:
-//
-//        ```
-//          #map = affine_map<(d0)[s0, s1] -> (d0 * s1 + s0)>
-//           func @foo(%A: tensor<?xf32> {linalg.inplaceable = true})
-//              -> tensor<?xf32> {
-//            %0 = bufferization.to_memref %A : memref<?xf32, #map>
-//            // ... uses of %0
-//            %res = bufferization.to_tensor %0 : memref<?xf32, #map>
-//            return %res : tensor<?xf32>
-//          }
-//        ```
-//
-//        this is the cue for the bufferization of the function foo (and calls
-//        to it) may bufferize to `func @foo(%A: memref<?xf32, some_layout>)`.
-//        To fully achieve bufferization, an additional analysis is needed to
-//        determine whether function argument/operand pairs bufferize to a
-//        single inplace buffer argument (i.e. functions may return tensors in
-//        arbitrary order that may not match argument numbers).
-//
-//    ii. results that don't map to an inplaceable function argument are
-//        generally allocated. Since memref semantics wrt ownership of the
-//        underlying memory region are not well-defined, comprehensive
-//        bufferization chooses to perform allocations in a scoped fashion:
-//        returning memrefs is always considered illegal.
-//        Such scenarios are encoded by IR resembling:
-//
-//        ```
-//          #map = affine_map<(d0)[s0, s1] -> (d0 * s1 + s0)>
-//          func @foo(%A: tensor<?xf32> {linalg.inplaceable = true})
-//              -> tensor<?xf32> {
-//            %0 = bufferization.to_memref %A : memref<?xf32, #map>
-//            %1 = memref.dim %0, %c0 : memref<?xf32, #map>
-//            %2 = memref.alloc(%1) : memref<?xf32>
-//            %3 = memref.cast %2 : memref<?xf32> to memref<?xf32, #map>
-//            // ... uses of %3
-//            memref.dealloc %2 : memref<?xf32, #map>
-//            %res = bufferization.to_tensor %3 : memref<?xf32, #map>
-//            return %res : tensor<?xf32>
-//          }
-//       ```
-//
-//        this is the cue for the bufferization of the function foo (and calls
-//        to it) that it must bufferize to `func @foo(%A: memref<?xf32,
-//        some_layout>,
-//                   %B: memref<?xf32, some_layout>)` (i.e. make a cloned
-//        allocation of the result tensor)
-//        To fully achieve bufferization, the alloc/dealloc pair must be lifted
-//        out of the function at each call site.
-//
-//   iii. as an optimization over ii., it may be possible to reuse an argument
-//        and only want to return a slice.
-//        This may forego allocation by letting *all* callers decide whether to
-//        pass a new *aliasing* memref function argument (i.e. a subview).
-//        Without loss of generality, callers may agree to allocate a new buffer
-//        to avoid this aliasing. Such scenarios are encoded by IR resembling:
-//
-//        ```
-//          #map = affine_map<(d0)[s0, s1] -> (d0 * s1 + s0)>
-//          func @foo(%arg0: tensor<?xf32> {linalg.inplaceable = true})
-//              -> tensor<4xf32> {
-//            %0 = bufferization.to_memref %arg0 : memref<?xf32, #map>
-//            %1 = memref.subview %0[0] [4] [1] : memref<?xf32, #map> to
-//                                                memref<4xf32, #map>
-//            // ... inplace computes into %1
-//            %3 = bufferization.to_tensor %1 : memref<4xf32, #map>
-//            return %3 : tensor<4xf32>
-//          }
-//        ```
-//
-//  Note: In the future, it may be worthwhile to design special bufferization
-//  ops to encode the desired semantics at function boundaries for i., ii. and
-//  iii.
+// to return any buffer, unless `allowReturnMemref` is enabled.
 //
 //  Lastly, note that layout map chosen to bufferize is the most dynamic
 //  canonical strided layout of the proper rank. This ensures compatibility with
 //  expected layouts after transformations. Combinations of memref.cast +
 //  canonicalization are responsible for clean ups.
 
 #include "mlir/Dialect/Linalg/ComprehensiveBufferize/ComprehensiveBufferize.h"
 

mlir/lib/Dialect/Linalg/ComprehensiveBufferize/LinalgInterfaceImpl.cpp

static LogicalResult bufferizeLinalgOp(RewriterBase &rewriter, LinalgOp op,
   if (op.hasBufferSemantics())
     return success();
 
   // Ensure op has only tensors. Allow mixed tensor-buffer mode on a per-need
   // basis.
   if (!op.hasTensorSemantics())
     return op->emitError() << "op does not have tensor semantics";
 
+  // New input operands for the cloned op.
   SmallVector<Value> newInputBuffers;
   newInputBuffers.reserve(op.getNumInputs());
   for (OpOperand *opOperand : op.getInputOperands()) {
     if (op.isScalar(opOperand)) {
       newInputBuffers.push_back(opOperand->get());
       continue;
     }
     newInputBuffers.push_back(state.lookupBuffer(rewriter, opOperand->get()));
   }
 
+  // New output operands for the cloned op.
   SmallVector<Value> newOutputBuffers;
   for (OpOperand *opOperand : op.getOutputOperands()) {
     OpResult opResult = op.getTiedOpResult(opOperand);
     assert(opResult && "could not find correspond OpResult");
-    Value resultBuffer = state.getResultBuffer(rewriter, opResult);
-    if (!resultBuffer)
-      return failure();
-    newOutputBuffers.push_back(resultBuffer);
+    FailureOr<Value> resultBuffer = state.getResultBuffer(rewriter, opResult);
+    newOutputBuffers.push_back(*resultBuffer);
   }
 
-  // Clone the newly bufferized op.
+  // Merge input/output operands.
   SmallVector<Value> newOperands = newInputBuffers;
   newOperands.append(newOutputBuffers.begin(), newOutputBuffers.end());
 
   // Set insertion point now that potential alloc/dealloc are introduced.
   rewriter.setInsertionPoint(op);
+  // Clone the op, but use the new operands. Since the new op does not have any
+  // tensor results, it does not return anything.
   op.clone(rewriter, op.getLoc(), /*resultTypes=*/TypeRange{}, newOperands);
 
   // Replace the results of the old op with the new output buffers.
   state.replaceOpWithBufferizedValues(rewriter, op, newOutputBuffers);
 
   return success();
 }
 
for (OpResult opResult : op->getOpResults()) {
     if (!chosenOperand)
       chosenOperand = tiedOperand;
 
     mapping[chosenOperand] = opResult;
   }
   return mapping;
 }
 
+/// Bufferization of linalg.generic. Replace with a new linalg.generic that
+/// operates entirely on memrefs.
 template <typename OpTy>
 struct LinalgOpInterface
     : public BufferizableOpInterface::ExternalModel<LinalgOpInterface<OpTy>,
                                                     OpTy> {
   bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand,
                               BufferizationState &state) const {
+    // Operand is read if it is used in the computation.
     auto genericOp = cast<linalg::LinalgOp>(op);
     return genericOp.payloadUsesValueFromOperand(&opOperand);
   }
 
   bool bufferizesToMemoryWrite(Operation *op, OpOperand &opOperand,
                                BufferizationState &state) const {
+    // Operand is written to if it has an aliasing OpResult. For more details,
+    // see `computeAliasingPairs`.
     auto bufferizableOp = cast<BufferizableOpInterface>(op);
     return static_cast<bool>(
         bufferizableOp.getAliasingOpResult(opOperand, state));
   }
 
   SmallVector<OpOperand *>
   getAliasingOpOperand(Operation *op, OpResult opResult,
                        BufferizationState &state) const {
     auto genericOp = cast<linalg::LinalgOp>(op);
+
+    // Aliasing OpOperand/OpResult pairs are computed by `computeAliasingPairs`.
     DenseMap<OpOperand *, OpResult> pairs = computeAliasingPairs(genericOp);
     for (OpOperand *opOperand : genericOp.getInputAndOutputOperands())
       if (pairs[opOperand] == opResult)
         return {opOperand};
     return {};
   }
 
   OpResult getAliasingOpResult(Operation *op, OpOperand &opOperand,
                                BufferizationState &state) const {
     auto genericOp = cast<linalg::LinalgOp>(op);
+
+    // Aliasing OpOperand/OpResult pairs are computed by `computeAliasingPairs`.
     DenseMap<OpOperand *, OpResult> pairs = computeAliasingPairs(genericOp);
     return pairs[&opOperand];
   }
 
   BufferRelation bufferRelation(Operation *op, OpResult opResult,
                                 const BufferizationAliasInfo &aliasInfo,
                                 BufferizationState &state) const {
     return BufferRelation::Equivalent;
LogicalResult bufferize(Operation *op, RewriterBase &rewriter,
     Value alloc = state.createAlloc(rewriter, initTensorOp->getLoc(),
                                     initTensorOp.result(),
                                     state.getOptions().createDeallocs);
     state.replaceOpWithBufferizedValues(rewriter, op, alloc);
     return success();
   }
 };
 
+/// Bufferization of linalg.tiled_loop. Replace with a new linalg.tiled_loop
+/// that operates entirely on memrefs.
 struct TiledLoopOpInterface
     : public BufferizableOpInterface::ExternalModel<TiledLoopOpInterface,
                                                     linalg::TiledLoopOp> {
   bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand,
                               BufferizationState &state) const {
-    // TiledLoop alone doesn't bufferize to a memory read, one of the uses of
-    // its matching bbArg may.
     auto tiledLoopOp = cast<linalg::TiledLoopOp>(op);
+
+    // linalg.tiled_loop operands alone do not bufferize to a memory read, but
+    // one of the uses of their matching bbArgs may.
     return state.isValueRead(tiledLoopOp.getTiedBlockArgument(opOperand));
   }
 
   bool bufferizesToMemoryWrite(Operation *op, OpOperand &opOperand,
                                BufferizationState &state) const {
-    // TiledLoop alone doesn't bufferize to a memory write, one of the uses of
-    // its matching bbArg may.
     auto bufferizableOp = cast<BufferizableOpInterface>(op);
+
+    // Only operands with an aliasing OpResult (i.e., output operands) bufferize
+    // to a memory write.
     return static_cast<bool>(
         bufferizableOp.getAliasingOpResult(opOperand, state));
   }
 
   OpResult getAliasingOpResult(Operation *op, OpOperand &opOperand,
                                BufferizationState &state) const {
     auto tiledLoopOp = cast<linalg::TiledLoopOp>(op);
+
+    // Output operands are tied to their corresponding OpResults.
     return tiledLoopOp.getTiedOpResult(opOperand);
   }
 
   BufferRelation bufferRelation(Operation *op, OpResult opResult,
                                 const BufferizationAliasInfo &aliasInfo,
                                 BufferizationState &state) const {
     return BufferRelation::Equivalent;
   }
 
   bool isWritable(Operation *op, Value value, BufferizationState &state) const {
-    // Interestingly, linalg::TiledLoopOp's bbArg can **always** be viewed
-    // inplace from the perspective of ops nested under:
+    // Interestingly, linalg::TiledLoopOp's bbArgs can **always** be viewed
+    // inplace from the perspective of nested ops:
     //   1. Either the matching iter operand is not bufferized inplace and an
     //      alloc + optional copy makes the bbArg itself inplaceable.
     //   2. Or the matching iter operand is bufferized inplace and bbArg just
     //      bufferizes to that too.
     return true;
   }
 
   bool isAllocationHoistingBarrier(Operation *op) const { return true; }
for (Value value : tiledLoopOp.inputs()) {
         newInputs.push_back(state.lookupBuffer(rewriter, value));
       } else {
         newInputs.push_back(value);
       }
     }
     int nextResultNum = 0;
     for (Value value : tiledLoopOp.outputs()) {
       if (value.getType().isa<TensorType>()) {
-        Value buffer = state.getResultBuffer(
+        FailureOr<Value> buffer = state.getResultBuffer(
             rewriter, tiledLoopOp->getResult(nextResultNum++));
-        newOutputs.push_back(buffer);
-        newResults.push_back(buffer);
+        newOutputs.push_back(*buffer);
+        newResults.push_back(*buffer);
       } else {
         newOutputs.push_back(value);
       }
     }
 
     // Create new TiledLoopOp.
     auto newTiledLoopOp = rewriter.create<TiledLoopOp>(
         tiledLoopOp.getLoc(), tiledLoopOp.lowerBound(),
LogicalResult bufferize(Operation *op, RewriterBase &rewriter,
 
     // Replace results and delete old op.
     state.replaceOpWithBufferizedValues(rewriter, op, newResults);
 
     return success();
   }
 };
 
+/// Bufferization of linalg.yield. Bufferized as part of linalg.tiled_loop's
+/// bufferization.
 struct YieldOpInterface
     : public BufferizableOpInterface::ExternalModel<YieldOpInterface,
                                                     linalg::YieldOp> {
   bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand,
                               BufferizationState &state) const {
     return true;
   }
 

 
 } // namespace
 
 /// Try to eliminate InitTensorOps inside `op`. An InitTensorOp is replaced
 /// with the the result of `rewriteFunc` if it is anchored on a matching
 /// OpOperand. "Anchored" means that there is a path on the reverse SSA use-def
 /// chain, starting from the OpOperand and always following the aliasing
 /// OpOperand, that eventually ends at a single InitTensorOp.
-LogicalResult mlir::linalg::comprehensive_bufferize::linalg_ext::
-    InitTensorEliminationStep::eliminateInitTensors(
-        Operation *op, BufferizationState &state,
+LogicalResult
+mlir::linalg::comprehensive_bufferize::linalg_ext::InitTensorEliminationStep::
+    eliminateInitTensors(Operation *op, BufferizationState &state,
                          BufferizationAliasInfo &aliasInfo,
-        std::function<bool(OpOperand &)> anchorMatchFunc,
-        std::function<Value(OpBuilder &, Location, OpOperand &)> rewriteFunc,
+                         AnchorMatchFn anchorMatchFunc, RewriteFn rewriteFunc,
                          SmallVector<Operation *> &newOps) {
   OpBuilder b(op->getContext());
 
   WalkResult status = op->walk([&](Operation *op) {
     for (OpOperand &operand : op->getOpOperands()) {
       // Is this a matching OpOperand?
       if (!anchorMatchFunc(operand))
         continue;
 

 /// Note that the newly inserted ExtractSliceOp may have to bufferize
 /// out-of-place due to RaW conflicts.
 LogicalResult mlir::linalg::comprehensive_bufferize::linalg_ext::
     InsertSliceAnchoredInitTensorEliminationStep::run(
         Operation *op, BufferizationState &state,
         BufferizationAliasInfo &aliasInfo, SmallVector<Operation *> &newOps) {
   return eliminateInitTensors(
       op, state, aliasInfo,
+      /*anchorMatchFunc=*/
       [&](OpOperand &operand) {
         auto insertSliceOp =
             dyn_cast<tensor::InsertSliceOp>(operand.getOwner());
         if (!insertSliceOp)
           return false;
         // Only inplace bufferized InsertSliceOps are eligible.
         if (!aliasInfo.isInPlace(insertSliceOp->getOpResult(0)))
           return false;
         return &operand == &insertSliceOp->getOpOperand(0) /*source*/;
       },
+      /*rewriteFunc=*/
       [](OpBuilder &b, Location loc, OpOperand &operand) {
         auto insertSliceOp = cast<tensor::InsertSliceOp>(operand.getOwner());
         auto extractOp = b.create<tensor::ExtractSliceOp>(
             loc, insertSliceOp.dest(), insertSliceOp.getMixedOffsets(),
             insertSliceOp.getMixedSizes(), insertSliceOp.getMixedStrides());
         return extractOp.result();
       },
       newOps);

mlir/lib/Dialect/Linalg/ComprehensiveBufferize/ModuleBufferization.cpp

 //===- ModuleBufferization.cpp - Bufferization across Func. Boundaries ----===//
 //
 // 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
 //
 //===----------------------------------------------------------------------===//
+//
+// Module bufferization is an extension of Comprehensive Bufferize that
+// bufferizes function boundaries. It provides `BufferizableOpInterface`
+// implementations for FuncOp, CallOp and ReturnOp, along with a few helper
+// functions that control the order in which functions are bufferized.
+//
+// Three cases can occur during bufferization of FuncOps.
+//
+//     i. inplaceable function arguments may be reused in place after the
+//        function itself has been bufferized. This is encoded by IR resembling:
+//
+//        ```
+//          #map = affine_map<(d0)[s0, s1] -> (d0 * s1 + s0)>
+//           func @foo(%A: tensor<?xf32> {linalg.inplaceable = true})
+//              -> tensor<?xf32> {
+//            %0 = bufferization.to_memref %A : memref<?xf32, #map>
+//            // ... uses of %0
+//            %res = bufferization.to_tensor %0 : memref<?xf32, #map>
+//            return %res : tensor<?xf32>
+//          }
+//        ```
+//
+//        this is the cue for the bufferization of the function foo (and calls
+//        to it) may bufferize to `func @foo(%A: memref<?xf32, some_layout>)`.
+//        To fully achieve bufferization, an additional analysis is needed to
+//        determine whether function argument/operand pairs bufferize to a
+//        single inplace buffer argument (i.e. functions may return tensors in
+//        arbitrary order that may not match argument numbers).
+//
+//    ii. results that don't map to an inplaceable function argument are
+//        generally allocated. Since memref semantics wrt ownership of the
+//        underlying memory region are not well-defined, comprehensive
+//        bufferization chooses to perform allocations in a scoped fashion:
+//        returning memrefs is always considered illegal.
+//        Such scenarios are encoded by IR resembling:
+//
+//        ```
+//          #map = affine_map<(d0)[s0, s1] -> (d0 * s1 + s0)>
+//          func @foo(%A: tensor<?xf32> {linalg.inplaceable = true})
+//              -> tensor<?xf32> {
+//            %0 = bufferization.to_memref %A : memref<?xf32, #map>
+//            %1 = memref.dim %0, %c0 : memref<?xf32, #map>
+//            %2 = memref.alloc(%1) : memref<?xf32>
+//            %3 = memref.cast %2 : memref<?xf32> to memref<?xf32, #map>
+//            // ... uses of %3
+//            memref.dealloc %2 : memref<?xf32, #map>
+//            %res = bufferization.to_tensor %3 : memref<?xf32, #map>
+//            return %res : tensor<?xf32>
+//          }
+//       ```
+//
+//        this is the cue for the bufferization of the function foo (and calls
+//        to it) that it must bufferize to `func @foo(%A: memref<?xf32,
+//        some_layout>,
+//                   %B: memref<?xf32, some_layout>)` (i.e. make a cloned
+//        allocation of the result tensor)
+//        To fully achieve bufferization, the alloc/dealloc pair must be lifted
+//        out of the function at each call site.
+//
+//   iii. as an optimization over ii., it may be possible to reuse an argument
+//        and only want to return a slice.
+//        This may forego allocation by letting *all* callers decide whether to
+//        pass a new *aliasing* memref function argument (i.e. a subview).
+//        Without loss of generality, callers may agree to allocate a new buffer
+//        to avoid this aliasing. Such scenarios are encoded by IR resembling:
+//
+//        ```
+//          #map = affine_map<(d0)[s0, s1] -> (d0 * s1 + s0)>
+//          func @foo(%arg0: tensor<?xf32> {linalg.inplaceable = true})
+//              -> tensor<4xf32> {
+//            %0 = bufferization.to_memref %arg0 : memref<?xf32, #map>
+//            %1 = memref.subview %0[0] [4] [1] : memref<?xf32, #map> to
+//                                                memref<4xf32, #map>
+//            // ... inplace computes into %1
+//            %3 = bufferization.to_tensor %1 : memref<4xf32, #map>
+//            return %3 : tensor<4xf32>
+//          }
+//        ```
+//
+//  Note: In the future, it may be worthwhile to design special bufferization
+//  ops to encode the desired semantics at function boundaries for i., ii. and
+//  iii.
 
 #include "mlir/Dialect/Linalg/ComprehensiveBufferize/ModuleBufferization.h"
 
 #include "mlir/Dialect/Bufferization/IR/Bufferization.h"
 #include "mlir/Dialect/Linalg/ComprehensiveBufferize/BufferizableOpInterface.h"
 #include "mlir/Dialect/Linalg/ComprehensiveBufferize/ComprehensiveBufferize.h"
 #include "mlir/Dialect/MemRef/IR/MemRef.h"
 #include "mlir/Dialect/StandardOps/IR/Ops.h"
static FunctionType getBufferizedFunctionType(MLIRContext *ctx,
                                               TypeRange argumentTypes,
                                               TypeRange resultTypes) {
   auto rewrite = [](Type t) -> Type {
     // TODO: non-zero address space.
     // TODO: layout information if relevant.
     if (auto rankedTensorType = t.dyn_cast<RankedTensorType>())
       return getDynamicMemRefType(rankedTensorType);
     if (auto tensorType = t.dyn_cast<TensorType>())
-      return getContiguousOrUnrankedMemRefType(tensorType);
+      return getUnrankedMemRefType(tensorType.getElementType());
     return t;
   };
   auto argTypes = llvm::to_vector<4>(llvm::map_range(argumentTypes, rewrite));
   auto retTypes = llvm::to_vector<4>(llvm::map_range(resultTypes, rewrite));
   return FunctionType::get(ctx, argTypes, retTypes);
 }
 
 /// If an entry for `funcOp` is available in `bufferizedFunctionTypes`, return

mlir/lib/Dialect/Linalg/ComprehensiveBufferize/SCFInterfaceImpl.cpp

 #include "mlir/IR/Operation.h"
 #include "mlir/IR/PatternMatch.h"
 
 namespace mlir {
 namespace linalg {
 namespace comprehensive_bufferize {
 namespace scf_ext {
 
+/// Bufferization of scf.execute_region. Can be analyzed, but bufferization not
+/// fully implemented at the moment.
 struct ExecuteRegionOpInterface
     : public BufferizableOpInterface::ExternalModel<ExecuteRegionOpInterface,
                                                     scf::ExecuteRegionOp> {
   SmallVector<OpOperand *>
   getAliasingOpOperand(Operation *op, OpResult opResult,
                        BufferizationState &state) const {
     // ExecuteRegionOps do not have tensor OpOperands. The yielded value can be
     // any SSA value that is in scope. To allow for use-def chain traversal
struct ExecuteRegionOpInterface
 
   BufferRelation bufferRelation(Operation *op, OpResult opResult,
                                 const BufferizationAliasInfo &aliasInfo,
                                 BufferizationState &state) const {
     return BufferRelation::Equivalent;
   }
 };
 
+/// Bufferization of scf.if. Replace with a new scf.if that yields memrefs.
 struct IfOpInterface
     : public BufferizableOpInterface::ExternalModel<IfOpInterface, scf::IfOp> {
   SmallVector<OpOperand *>
   getAliasingOpOperand(Operation *op, OpResult opResult,
                        BufferizationState &state) const {
     // IfOps do not have tensor OpOperands. The yielded value can be any SSA
     // value that is in scope. To allow for use-def chain traversal through
     // IfOps in the analysis, both corresponding yield values from the then/else
SmallVector<OpOperand *> yieldValues =
         bufferizableOp.getAliasingOpOperand(opResult, state);
     assert(yieldValues.size() == 2 && "expected 2 yield values");
     bool equivalentYields = aliasInfo.areEquivalentBufferizedValues(
         yieldValues[0]->get(), yieldValues[1]->get());
     return equivalentYields ? BufferRelation::Equivalent : BufferRelation::None;
   }
 };
 
+/// Bufferization of scf.for. Replace with a new scf.for that operates on
+/// memrefs.
 struct ForOpInterface
     : public BufferizableOpInterface::ExternalModel<ForOpInterface,
                                                     scf::ForOp> {
   bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand,
                               BufferizationState &state) const {
     // scf::ForOp alone doesn't bufferize to a memory read, one of the uses of
     // its matching bbArg may.
     auto forOp = cast<scf::ForOp>(op);
auto convert = [&](ValueRange values,
         result.push_back(indices.contains(idx) ? func(val, idx) : val);
       }
       return result;
     };
 
     // Construct a new scf.for op with memref instead of tensor values.
     SmallVector<Value> initArgs =
         convert(forOp.getInitArgs(), [&](Value val, int64_t index) {
-          return state.getResultBuffer(rewriter, forOp->getOpResult(index));
+          return *state.getResultBuffer(rewriter, forOp->getOpResult(index));
         });
     auto newForOp = rewriter.create<scf::ForOp>(
         forOp.getLoc(), forOp.getLowerBound(), forOp.getUpperBound(),
         forOp.getStep(), initArgs);
     Block *loopBody = &newForOp.getLoopBody().front();
 
     // Set up new iter_args. The loop body uses tensors, so wrap the (memref)
     // iter_args of the new loop in ToTensorOps.
if (auto ifOp = dyn_cast<scf::IfOp>(yieldOp->getParentOp())) {
       }
     }
 
     return WalkResult::advance();
   });
   return status;
 }
 
+/// Bufferization of scf.yield. Bufferized as part of their enclosing ops, so
+/// this is for analysis only.
 struct YieldOpInterface
     : public BufferizableOpInterface::ExternalModel<YieldOpInterface,
                                                     scf::YieldOp> {
   bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand,
                               BufferizationState &state) const {
     return true;
   }
 

mlir/lib/Dialect/Linalg/ComprehensiveBufferize/TensorInterfaceImpl.cpp

BufferRelation bufferRelation(Operation *op, OpResult opResult,
                                 BufferizationState &state) const {
     return BufferRelation::Equivalent;
   }
 
   LogicalResult bufferize(Operation *op, RewriterBase &rewriter,
                           BufferizationState &state) const {
     auto castOp = cast<tensor::CastOp>(op);
 
-    Value resultBuffer = state.getResultBuffer(rewriter, castOp->getResult(0));
-    if (!resultBuffer)
-      return failure();
-    Type sourceType = resultBuffer.getType();
-    auto rankedMemRefType = sourceType.dyn_cast<MemRefType>();
-    auto unrankedMemRefType = sourceType.dyn_cast<UnrankedMemRefType>();
-    assert(rankedMemRefType || unrankedMemRefType);
-    Attribute memorySpace = rankedMemRefType
-                                ? rankedMemRefType.getMemorySpace()
-                                : unrankedMemRefType.getMemorySpace();
-    TensorType tensorType = castOp.getResult().getType().cast<TensorType>();
-    MemRefLayoutAttrInterface layout =
-        rankedMemRefType && tensorType.isa<RankedTensorType>()
-            ? rankedMemRefType.getLayout()
-            : MemRefLayoutAttrInterface();
-    Type memRefType = getContiguousOrUnrankedMemRefType(
-        castOp.getResult().getType(), layout, memorySpace);
-    state.replaceOpWithNewBufferizedOp<memref::CastOp>(rewriter, op, memRefType,
-                                                       resultBuffer);
+    // The result buffer still has the old (pre-cast) type.
+    FailureOr<Value> resultBuffer =
+        state.getResultBuffer(rewriter, castOp->getResult(0));
+    auto sourceMemRefType = resultBuffer->getType().cast<BaseMemRefType>();
+    Attribute memorySpace = sourceMemRefType.getMemorySpace();
+    TensorType resultTensorType =
+        castOp.getResult().getType().cast<TensorType>();
+    MemRefLayoutAttrInterface layout;
+
+    if (auto rankedMemRefType = sourceMemRefType.dyn_cast<MemRefType>())
+      if (resultTensorType.isa<RankedTensorType>())
+        layout = rankedMemRefType.getLayout();
+
+    // Compute the new memref type.
+    Type resultMemRefType;
+    if (auto rankedTensorType = resultTensorType.isa<RankedTensorType>()) {
+      resultMemRefType =
+          getContiguousMemRefType(resultTensorType, layout, memorySpace);
+    } else {
+      resultMemRefType =
+          getUnrankedMemRefType(resultTensorType.getElementType(), memorySpace);
+    }
+
+    // Replace the op with a memref.cast.
+    state.replaceOpWithNewBufferizedOp<memref::CastOp>(
+        rewriter, op, resultMemRefType, *resultBuffer);
+
     return success();
   }
 };
 
+/// Bufferization of tensor.dim. Replace with memref.dim.
 struct DimOpInterface
     : public BufferizableOpInterface::ExternalModel<DimOpInterface,
                                                     tensor::DimOp> {
   bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand,
                               BufferizationState &state) const {
     return true;
   }
 
   bool bufferizesToMemoryWrite(Operation *op, OpOperand &opOperand,
                                BufferizationState &state) const {
     return false;
   }
 
   OpResult getAliasingOpResult(Operation *op, OpOperand &opOperand,
                                BufferizationState &state) const {
     return OpResult();
   }
 
   LogicalResult bufferize(Operation *op, RewriterBase &rewriter,
                           BufferizationState &state) const {
     auto dimOp = cast<tensor::DimOp>(op);
-    if (!dimOp.source().getType().isa<RankedTensorType>())
-      return dimOp.emitError("unranked tensor not supported");
     Value v = state.lookupBuffer(rewriter, dimOp.source());
     state.replaceOpWithNewBufferizedOp<memref::DimOp>(rewriter, op, v,
                                                       dimOp.index());
     return success();
   }
 };
 
+/// Bufferization of tensor.extract_slice. Replace with memref.subview.
 struct ExtractSliceOpInterface
     : public BufferizableOpInterface::ExternalModel<ExtractSliceOpInterface,
                                                     tensor::ExtractSliceOp> {
   bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand,
                               BufferizationState &state) const {
     return false;
   }
 
auto subviewMemRefType =
             dstTensorType.getRank(), srcMemrefType,
             extractSliceOp.getMixedOffsets(), extractSliceOp.getMixedSizes(),
             extractSliceOp.getMixedStrides())
             .cast<MemRefType>();
     Value subView = rewriter.create<memref::SubViewOp>(
         loc, subviewMemRefType, srcMemref, extractSliceOp.getMixedOffsets(),
         extractSliceOp.getMixedSizes(), extractSliceOp.getMixedStrides());
 
-    /// If not inplaceable, copy.
+    // If not inplaceable, copy.
     if (!inplace) {
       // Do not copy if the copied data is never read.
       if (state.isValueRead(extractSliceOp.result()))
         state.createMemCpy(rewriter, extractSliceOp.getLoc(), subView, alloc);
       subView = alloc;
     }
 
     state.replaceOpWithBufferizedValues(rewriter, op, subView);
     return success();
   }
 };
 
+/// Bufferization of tensor.extract. Replace with memref.load.
 struct ExtractOpInterface
     : public BufferizableOpInterface::ExternalModel<ExtractOpInterface,
                                                     tensor::ExtractOp> {
   bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand,
                               BufferizationState &state) const {
     return true;
   }
 
LogicalResult bufferize(Operation *op, RewriterBase &rewriter,
     auto extractOp = cast<tensor::ExtractOp>(op);
     Value srcMemref = state.lookupBuffer(rewriter, extractOp.tensor());
     state.replaceOpWithNewBufferizedOp<memref::LoadOp>(rewriter, op, srcMemref,
                                                        extractOp.indices());
     return success();
   }
 };
 
+/// Bufferization of tensor.insert. Replace with memref.store.
 struct InsertOpInterface
     : public BufferizableOpInterface::ExternalModel<InsertOpInterface,
                                                     tensor::InsertOp> {
   bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand,
                               BufferizationState &state) const {
     return true;
   }
 
struct InsertOpInterface
   getAliasingOpOperand(Operation *op, OpResult opResult,
                        BufferizationState &state) const {
     return {&op->getOpOperand(1) /*dest*/};
   }
 
   LogicalResult bufferize(Operation *op, RewriterBase &rewriter,
                           BufferizationState &state) const {
     auto insertOp = cast<tensor::InsertOp>(op);
-    Location loc = insertOp.getLoc();
-    Value destMemref =
+    FailureOr<Value> destMemref =
         state.getResultBuffer(rewriter, insertOp->getOpResult(0));
-    rewriter.create<memref::StoreOp>(loc, insertOp.scalar(), destMemref,
-                                     insertOp.indices());
-    state.replaceOpWithBufferizedValues(rewriter, op, destMemref);
+    rewriter.create<memref::StoreOp>(insertOp.getLoc(), insertOp.scalar(),
+                                     *destMemref, insertOp.indices());
+    state.replaceOpWithBufferizedValues(rewriter, op, *destMemref);
     return success();
   }
 
   BufferRelation bufferRelation(Operation *op, OpResult opResult,
                                 const BufferizationAliasInfo &aliasInfo,
                                 BufferizationState &state) const {
     return BufferRelation::Equivalent;
   }
if (auto extractOp = val.getDefiningOp<ExtractSliceOp>())
         return true;
     return false;
   };
 
   return llvm::all_of(state.findValueInReverseUseDefChain(value, condition),
                       condition);
 }
 
+/// Bufferization of tensor.insert_slice. Replace with a memory copy. Under
+/// certain circumstances, this op can also be a no-op.
 struct InsertSliceOpInterface
     : public BufferizableOpInterface::ExternalModel<InsertSliceOpInterface,
                                                     tensor::InsertSliceOp> {
   bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand,
                               BufferizationState &state) const {
     return true;
   }
 
LogicalResult bufferize(Operation *op, RewriterBase &rewriter,
     // generally a deal breaker. When used with loops, this ends up cloning the
     // whole tensor on every single iteration and is a symptom of a
     // catastrophically bad scheduling decision.
     // TODO: be very loud about it or even consider failing the pass.
     auto insertSliceOp = cast<tensor::InsertSliceOp>(op);
     Location loc = insertSliceOp.getLoc();
 
     // When bufferizing out-of-place, `getResultBuffer` allocates.
-    Value dstMemref =
+    FailureOr<Value> dstMemref =
         state.getResultBuffer(rewriter, insertSliceOp->getResult(0));
-    if (!dstMemref)
-      return failure();
 
     // Take a subview of the dst.
-    auto dstMemrefType = dstMemref.getType().cast<MemRefType>();
+    auto dstMemrefType = dstMemref->getType().cast<MemRefType>();
     auto subviewMemRefType =
         memref::SubViewOp::inferRankReducedResultType(
             insertSliceOp.getSourceType().getRank(), dstMemrefType,
             insertSliceOp.getMixedOffsets(), insertSliceOp.getMixedSizes(),
             insertSliceOp.getMixedStrides())
             .cast<MemRefType>();
     Value subView = rewriter.create<memref::SubViewOp>(
-        loc, subviewMemRefType, dstMemref, insertSliceOp.getMixedOffsets(),
+        loc, subviewMemRefType, *dstMemref, insertSliceOp.getMixedOffsets(),
         insertSliceOp.getMixedSizes(), insertSliceOp.getMixedStrides());
 
     // Copy tensor. If this tensor.insert_slice has a matching
     // tensor.extract_slice, the copy operation will eventually fold away.
     Value srcMemref = state.lookupBuffer(rewriter, insertSliceOp.source());
-    state.createMemCpy(rewriter, insertSliceOp.getLoc(), srcMemref, subView);
+    state.createMemCpy(rewriter, loc, srcMemref, subView);
 
-    state.replaceOpWithBufferizedValues(rewriter, op, dstMemref);
+    state.replaceOpWithBufferizedValues(rewriter, op, *dstMemref);
     return success();
   }
 };
 
 } // namespace tensor_ext
 } // namespace comprehensive_bufferize
 } // namespace linalg
 } // namespace mlir

mlir/lib/Dialect/Linalg/ComprehensiveBufferize/VectorInterfaceImpl.cpp

 #include "mlir/IR/Dialect.h"
 #include "mlir/IR/Operation.h"
 
 namespace mlir {
 namespace linalg {
 namespace comprehensive_bufferize {
 namespace vector_ext {
 
+/// Bufferization of vector.transfer_read. Replaced with a new
+/// vector.transfer_read that operates on a memref.
 struct TransferReadOpInterface
     : public BufferizableOpInterface::ExternalModel<TransferReadOpInterface,
                                                     vector::TransferReadOp> {
   bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand,
                               BufferizationState &state) const {
     assert(opOperand.get().getType().isa<RankedTensorType>() &&
            "only tensor types expected");
     return true;
LogicalResult bufferize(Operation *op, RewriterBase &rewriter,
     state.replaceOpWithNewBufferizedOp<vector::TransferReadOp>(
         rewriter, readOp, readOp.getVectorType(), buffer, readOp.indices(),
         readOp.permutation_map(), readOp.padding(), readOp.mask(),
         readOp.in_boundsAttr());
     return success();
   }
 };
 
+/// Bufferization of vector.transfer_write. Replace with a new
+/// vector.transfer_write that operates on a memref.
 struct TransferWriteOpInterface
     : public BufferizableOpInterface::ExternalModel<TransferWriteOpInterface,
                                                     vector::TransferWriteOp> {
   bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand,
                               BufferizationState &state) const {
     assert(opOperand.get().getType().isa<TensorType>() &&
            "only tensor types expected");
     return true;
LogicalResult bufferize(Operation *op, RewriterBase &rewriter,
                           BufferizationState &state) const {
     auto writeOp = cast<vector::TransferWriteOp>(op);
     assert(writeOp.getShapedType().isa<TensorType>() &&
            "only tensor types expected");
 
     // Create a new transfer_write on buffer that doesn't have a return value.
     // Leave the previous transfer_write to dead code as it still has uses at
     // this point.
-    Value resultBuffer = state.getResultBuffer(rewriter, op->getResult(0));
-    if (!resultBuffer)
-      return failure();
+    FailureOr<Value> resultBuffer =
+        state.getResultBuffer(rewriter, op->getResult(0));
     rewriter.create<vector::TransferWriteOp>(
-        writeOp.getLoc(), writeOp.vector(), resultBuffer, writeOp.indices(),
+        writeOp.getLoc(), writeOp.vector(), *resultBuffer, writeOp.indices(),
         writeOp.permutation_mapAttr(), writeOp.in_boundsAttr());
-    state.replaceOpWithBufferizedValues(rewriter, op, resultBuffer);
+    state.replaceOpWithBufferizedValues(rewriter, op, *resultBuffer);
 
     return success();
   }
 };
 
 } // namespace vector_ext
 } // namespace comprehensive_bufferize
 } // namespace linalg