[mlir][Interface] DestinationStyleOpInterface: Rename hasTensor/BufferSemantics

Rename interface functions as follows:
* `hasTensorSemantics` -> `hasPureTensorSemantics`
* `hasBufferSemantics` -> `hasPureBufferSemantics`

These two functions return "true" if the op has tensor/buffer operands but not buffer/tensor operands.

Add two new interface functions:
* `hasTensorSemantics`: Return "true" if the op has tensor operands. Whether the op has buffer operands or not does not matter.
* `hasBufferSemantics`: Return "true" if the op has buffer operands. Whether the op has tensor operands or not does not matter.

Also drop the "ranked" part from the interface, i.e., do not distinguish between ranked/unranked types.

This change aligns the meaning of "tensor semantics" with the bufferization framework. (An op is supposed to be bufferized if it has tensor operands, and we don't care if it also has memref operands.) This change is in preparation of llvm#75273, which adds `BufferizableOpInterface::hasTensorSemantics`.

Author: matthias-springer

mlir/include/mlir/Interfaces/DestinationStyleOpInterface.td

@@ -17,24 +17,24 @@ def DestinationStyleOpInterface : OpInterface<"DestinationStyleOpInterface"> {
     as initial tensor values for the results of the operation or the init
     buffers to which the results of the op will be written.
 
-    Init operands must be ranked tensors or ranked memrefs. Input operands can
-    have any type. All non-init operands are DPS inputs.
+    Init operands must be tensors or memrefs. Input operands can have any type.
+    All non-init operands are DPS inputs.
 
     The init operands of this op are specified by the MutableOperandRange that
     the `getDpsInitsMutable` interface methods returns. This implies that the
     init operands must be a consecutive range of operands.
 
-    If the op has "tensor semantics", then the input operands are either ranked
-    tensors or other non-tensor/memref types ("scalars"). The init operands are
-    ranked tensors and every tensor init is tied to a corresponding tensor
-    OpResult in a 1-to-1 fashion. The i-th init tensor is tied to the i-th
-    OpResult. The op may not have any additional OpResults. Init operands and
-    their tied OpResults have the same type. Dynamic dimension sizes also match
-    at runtime.
+    Each tensor init operand is tied to a corresponding tensor OpResult in a
+    1-to-1 fashion. The i-th init tensor is tied to the i-th OpResult. The op
+    may not have any additional OpResults. Init operands and their tied
+    OpResults have the same type. Dynamic dimension sizes also match at runtime.
 
-    If the op has "buffer semantics", then the input operands are either ranked
-    memrefs or other non-tensor/memref types ("scalar" types). Furthermore, the
-    init operands are ranked memrefs and the op has no results.
+    Note: This implies that a destination style op without any tensor inits must
+    not have any OpResults.
+
+    An op has "pure tensor semantics" if it has at least one tensor operand and
+    no buffer (memref) operands. It has "pure buffer semantics" if it has at
+    least one buffer (memref) operand and no tensor operands.
 
     Destination-passing style abstraction makes certain transformations easier.
     For example, tiling implementation can extract/insert slices from/into the
@@ -148,7 +148,8 @@ def DestinationStyleOpInterface : OpInterface<"DestinationStyleOpInterface"> {
     /// neither a MemRef nor a tensor value.
     bool isScalar(::mlir::OpOperand *opOperand) {
       assert(opOperand->getOwner() == $_op && "invalid operand");
-      return !::llvm::isa<MemRefType, TensorType>(opOperand->get().getType());
+      return !::llvm::isa<BaseMemRefType, TensorType>(
+          opOperand->get().getType());
     }
 
     /// Return the OpResult that is tied to the given OpOperand.
@@ -169,37 +170,36 @@ def DestinationStyleOpInterface : OpInterface<"DestinationStyleOpInterface"> {
       return $_op.getDpsInitOperand(opResult.getResultNumber());
     }
 
-    /// Return whether the op has buffer semantics. That is the case if the op
-    /// has no ranked tensor operands and at least one memref operand.
-    bool hasBufferSemantics() {
+    /// Return whether the op has pure buffer semantics. That is the case if the
+    /// op has no tensor operands and at least one memref operand.
+    bool hasPureBufferSemantics() {
       // No tensors.
       auto isTensor = [](Value v){
-        return ::llvm::isa<::mlir::RankedTensorType>(v.getType());
+        return ::llvm::isa<::mlir::TensorType>(v.getType());
       };
       if (::llvm::any_of($_op->getOperands(), isTensor))
         return false;
       // At least one memref.
       auto isMemref = [](Value v){
-        return ::llvm::isa<::mlir::MemRefType>(v.getType());
+        return ::llvm::isa<::mlir::BaseMemRefType>(v.getType());
       };
       return llvm::any_of($_op->getOperands(), isMemref);
     }
 
-    /// Return whether the op has tensor semantics. That is the case if the op
-    /// has no memref operands and at least one ranked tensor operand.
-    bool hasTensorSemantics() {
+    /// Return whether the op has pure tensor semantics. That is the case if the
+    /// op has no memref operands and at least one tensor operand.
+    bool hasPureTensorSemantics() {
       // No memrefs.
       auto isMemref = [](Value v){
-        return ::llvm::isa<::mlir::MemRefType>(v.getType());
+        return ::llvm::isa<::mlir::BaseMemRefType>(v.getType());
       };
       if (::llvm::any_of($_op->getOperands(), isMemref))
         return false;
       // At least one tensor.
       auto isTensor = [](Value v){
-        return ::llvm::isa<::mlir::RankedTensorType>(v.getType());
+        return ::llvm::isa<::mlir::TensorType>(v.getType());
       };
-      return llvm::any_of($_op->getOperands(), isTensor);
-    }
+      return llvm::any_of($_op->getOperands(), isTensor);    }
   }];
 
   let verify = [{ return detail::verifyDestinationStyleOpInterface($_op); }];

mlir/lib/Dialect/Linalg/IR/LinalgOps.cpp

@@ -550,7 +550,7 @@ struct EraseSelfCopy : OpRewritePattern<CopyOp> {
                                 PatternRewriter &rewriter) const override {
     if (copyOp.getInputs() != copyOp.getOutputs())
       return rewriter.notifyMatchFailure(copyOp, "not a self copy");
-    if (copyOp.hasBufferSemantics())
+    if (copyOp.hasPureBufferSemantics())
       rewriter.eraseOp(copyOp);
     else
       rewriter.replaceOp(copyOp, copyOp.getInputs());
@@ -1112,7 +1112,7 @@ struct EraseIdentityGenericOp : public OpRewritePattern<GenericOp> {
       return failure();
 
     // In the buffer case, we need to check exact buffer equality.
-    if (genericOp.hasBufferSemantics()) {
+    if (genericOp.hasPureBufferSemantics()) {
       if (genericOp.getNumDpsInputs() == 1 && genericOp.getNumDpsInits() == 1 &&
           genericOp.getDpsInputOperand(0)->get() ==
               genericOp.getDpsInitOperand(0)->get()) {
@@ -1123,7 +1123,7 @@ struct EraseIdentityGenericOp : public OpRewritePattern<GenericOp> {
     }
 
     // Mixed semantics is not supported yet.
-    if (!genericOp.hasTensorSemantics())
+    if (!genericOp.hasPureTensorSemantics())
       return failure();
 
     // Get the argument number of the returned values. That is the operand
@@ -2257,7 +2257,7 @@ struct InferStaticShapeOfOperands : public OpInterfaceRewritePattern<LinalgOp> {
 
   LogicalResult matchAndRewrite(LinalgOp linalgOp,
                                 PatternRewriter &rewriter) const override {
-    if (!linalgOp.hasTensorSemantics())
+    if (!linalgOp.hasPureTensorSemantics())
       return failure();
 
     // Maps must be projected permutations.
@@ -2376,7 +2376,7 @@ SoftmaxOp::getTiledImplementation(OpBuilder &builder,
       getSlice(builder, getLoc(), getOutput(), offsets, sizes, strides));
 
   SmallVector<Type, 4> resultTypes;
-  if (hasTensorSemantics())
+  if (hasPureTensorSemantics())
     resultTypes.push_back(tiledOperands[1].getType());
   Operation *tiledOp =
       mlir::clone(builder, getOperation(), resultTypes, tiledOperands);

mlir/lib/Dialect/Linalg/Transforms/BubbleUpExtractSlice.cpp

@@ -68,7 +68,7 @@ struct BubbleUpExtractSliceOpPattern
                                          "expected single output of linalg op");
     }
 
-    if (!linalgOp.hasTensorSemantics()) {
+    if (!linalgOp.hasPureTensorSemantics()) {
       return rewriter.notifyMatchFailure(sliceOp,
                                          "expected tensor of linalg op");
     }

mlir/lib/Dialect/Linalg/Transforms/BufferizableOpInterfaceImpl.cpp

@@ -32,13 +32,13 @@ bufferizeDestinationStyleOpInterface(RewriterBase &rewriter,
   rewriter.setInsertionPoint(op);
 
   // Nothing to do. This op is already bufferized.
-  if (op.hasBufferSemantics())
+  if (op.hasPureBufferSemantics())
     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";
+  if (!op.hasPureTensorSemantics())
+    return op->emitError() << "op does not have pure tensor semantics";
 
   // New input operands for the cloned op.
   SmallVector<Value> newInputBuffers;

mlir/lib/Dialect/Linalg/Transforms/ConstantFold.cpp

@@ -57,7 +57,7 @@ class FoldConstantBase : public OpRewritePattern<GenericOp> {
   LogicalResult matchAndRewrite(GenericOp genericOp,
                                 PatternRewriter &rewriter) const override {
     // Mixed and buffer sematics aren't supported.
-    if (!genericOp.hasTensorSemantics())
+    if (!genericOp.hasPureTensorSemantics())
       return failure();
 
     // Only support ops generating one output for now.

mlir/lib/Dialect/Linalg/Transforms/DecomposeLinalgOps.cpp

@@ -258,7 +258,7 @@ DecomposeLinalgOp::matchAndRewrite(GenericOp genericOp,
   // TODO: this could be generalized to handle `linalg.generic` with buffer
   // operands too but requires allocation for intermediates. Punt on this for
   // now.
-  if (!genericOp.hasTensorSemantics()) {
+  if (!genericOp.hasPureTensorSemantics()) {
     return rewriter.notifyMatchFailure(
         genericOp, "only operations with tensor semantics are handled");
   }

mlir/lib/Dialect/Linalg/Transforms/DropUnitDims.cpp

@@ -83,7 +83,7 @@ struct MoveInitOperandsToInput : public OpRewritePattern<GenericOp> {
   using OpRewritePattern<GenericOp>::OpRewritePattern;
   LogicalResult matchAndRewrite(GenericOp genericOp,
                                 PatternRewriter &rewriter) const override {
-    if (!genericOp.hasTensorSemantics())
+    if (!genericOp.hasPureTensorSemantics())
       return failure();
     if (genericOp.getNumParallelLoops() != genericOp.getNumLoops())
       return failure();

mlir/lib/Dialect/Linalg/Transforms/ElementwiseOpFusion.cpp

@@ -105,7 +105,7 @@ bool mlir::linalg::areElementwiseOpsFusable(OpOperand *fusedOperand) {
   // Consumer can have mixed semantics, just check operand itself has tensor
   // type. Producer must have full tensor semantics to avoid potential
   // aliasing between producer and consumer memrefs.
-  if (!producer.hasTensorSemantics() ||
+  if (!producer.hasPureTensorSemantics() ||
       !isa<RankedTensorType>(fusedOperand->get().getType()))
     return false;
 
@@ -530,7 +530,7 @@ static bool isFusableWithReshapeByDimExpansion(GenericOp genericOp,
   //   permutations.
   // - The fused tensor is not a scalar.
   // - All the loops are parallel loops.
-  return genericOp.hasTensorSemantics() &&
+  return genericOp.hasPureTensorSemantics() &&
          llvm::all_of(genericOp.getIndexingMaps().getValue(),
                       [](Attribute attr) {
                         return cast<AffineMapAttr>(attr)
@@ -1124,7 +1124,7 @@ static SmallVector<ReassociationIndices>
 getCollapsableIterationSpaceDims(GenericOp genericOp, OpOperand *fusableOperand,
                                  ArrayRef<ReassociationIndices> reassociation) {
   // Some basic checks for this fusion to be valid.
-  if (!genericOp.hasTensorSemantics() || genericOp.getNumDpsInits() != 1)
+  if (!genericOp.hasPureTensorSemantics() || genericOp.getNumDpsInits() != 1)
     return {};
 
   if (!llvm::all_of(genericOp.getIndexingMapsArray(), [](AffineMap map) {
@@ -1476,7 +1476,7 @@ Operation *createCollapsedOp(LinalgType op,
     outputOperands.push_back(newOutput);
     // If the op has "buffer semantics", then the init operands are ranked
     // memrefs and the op has no results.
-    if (!op.hasBufferSemantics())
+    if (!op.hasPureBufferSemantics())
       resultTypes.push_back(newOutput.getType());
   }
 
@@ -1521,8 +1521,8 @@ FailureOr<SmallVector<Value>> mlir::linalg::collapseOpIterationDims(
       }))
     return failure();
 
-  bool hasBufferSemantics = op.hasBufferSemantics();
-  if (hasBufferSemantics &&
+  bool hasPureBufferSemantics = op.hasPureBufferSemantics();
+  if (hasPureBufferSemantics &&
       !llvm::all_of(op->getOperands(), [&](Value operand) -> bool {
         MemRefType memRefToCollapse = dyn_cast<MemRefType>(operand.getType());
         if (!memRefToCollapse)
@@ -1705,7 +1705,7 @@ class FoldScalarOrSplatConstant : public OpRewritePattern<GenericOp> {
 
   LogicalResult matchAndRewrite(GenericOp genericOp,
                                 PatternRewriter &rewriter) const override {
-    if (!genericOp.hasTensorSemantics())
+    if (!genericOp.hasPureTensorSemantics())
       return failure();
     for (OpOperand *opOperand : genericOp.getDpsInputOperands()) {
       Operation *def = opOperand->get().getDefiningOp();
@@ -1857,7 +1857,7 @@ struct FoldFillWithGenericOp : public OpRewritePattern<GenericOp> {
 
   LogicalResult matchAndRewrite(GenericOp genericOp,
                                 PatternRewriter &rewriter) const override {
-    if (!genericOp.hasTensorSemantics())
+    if (!genericOp.hasPureTensorSemantics())
       return failure();
     bool fillFound = false;
     Block &payload = genericOp.getRegion().front();

mlir/lib/Dialect/Linalg/Transforms/EraseUnusedOperandsAndResults.cpp

@@ -183,7 +183,7 @@ struct DeduplicateAndRemoveDeadOperandsAndResults
         dedupedOutpts;
     // If the op doesn't have tensor semantics or outputs should not be removed,
     // keep all the outputs as preserved.
-    if (!genericOp.hasTensorSemantics() || !removeOutputs) {
+    if (!genericOp.hasPureTensorSemantics() || !removeOutputs) {
       for (const auto &en : llvm::enumerate(genericOp.getDpsInitsMutable())) {
         origToNewPos[en.index()] = newOutputOperands.size();
         newOutputOperands.push_back(en.value().get());
@@ -317,7 +317,7 @@ struct RemoveUnusedCycleInGenericOp : public OpRewritePattern<GenericOp> {
                                 PatternRewriter &rewriter) const override {
 
     // If the op doesnt have tensor semantics, preserve the outputs as is.
-    if (!genericOp.hasTensorSemantics())
+    if (!genericOp.hasPureTensorSemantics())
       return failure();
 
     bool hasRemovedCycles = false;

mlir/lib/Dialect/Linalg/Transforms/Generalization.cpp

@@ -59,7 +59,7 @@ FailureOr<GenericOp> mlir::linalg::generalizeNamedOp(RewriterBase &rewriter,
   ValueRange outputs = linalgOp.getDpsInits();
   SmallVector<AffineMap> indexingMaps = linalgOp.getIndexingMapsArray();
   SmallVector<utils::IteratorType> iterators = linalgOp.getIteratorTypesArray();
-  SmallVector<Type> resultTypes = linalgOp.hasTensorSemantics()
+  SmallVector<Type> resultTypes = linalgOp.hasPureTensorSemantics()
                                       ? TypeRange(ValueRange(outputs))
                                       : TypeRange{};
 

mlir/lib/Dialect/Linalg/Transforms/InlineScalarOperands.cpp

@@ -35,7 +35,7 @@ struct InlineScalarOperands : public OpRewritePattern<GenericOp> {
   using OpRewritePattern<GenericOp>::OpRewritePattern;
   LogicalResult matchAndRewrite(GenericOp genericOp,
                                 PatternRewriter &rewriter) const override {
-    if (!genericOp.hasTensorSemantics())
+    if (!genericOp.hasPureTensorSemantics())
       return failure();
 
     SmallVector<size_t> scalarOperands;

mlir/lib/Dialect/Linalg/Transforms/Loops.cpp

@@ -128,7 +128,7 @@ template <typename LoadOpTy, typename StoreOpTy>
 static void emitScalarImplementation(OpBuilder &b, Location loc,
                                      ArrayRef<Value> allIvs,
                                      LinalgOp linalgOp) {
-  assert(linalgOp.hasBufferSemantics() &&
+  assert(linalgOp.hasPureBufferSemantics() &&
          "expected linalg op with buffer semantics");
   SmallVector<Value> indexedValues;
   indexedValues.reserve(linalgOp->getNumOperands());
@@ -218,7 +218,7 @@ static FailureOr<LinalgLoops> linalgOpToLoopsImpl(RewriterBase &rewriter,
 
   // The flattened loopToOperandRangesMaps is expected to be an invertible
   // permutation map (which is asserted in the inverse calculation).
-  assert(linalgOp.hasBufferSemantics() &&
+  assert(linalgOp.hasPureBufferSemantics() &&
          "expected linalg op with buffer semantics");
 
   auto loopRanges = linalgOp.createLoopRanges(rewriter, linalgOp.getLoc());
@@ -264,7 +264,7 @@ class LinalgRewritePattern : public RewritePattern {
   LogicalResult matchAndRewrite(Operation *op,
                                 PatternRewriter &rewriter) const override {
     auto linalgOp = dyn_cast<LinalgOp>(op);
-    if (!isa<LinalgOp>(op) || !linalgOp.hasBufferSemantics()) {
+    if (!isa<LinalgOp>(op) || !linalgOp.hasPureBufferSemantics()) {
       return rewriter.notifyMatchFailure(
           op, "expected linalg op with buffer semantics");
     }

mlir/lib/Dialect/Linalg/Transforms/NamedOpConversions.cpp

@@ -39,7 +39,7 @@ matchAndReplaceDepthwiseConv(Operation *operation, Value input, Value kernel,
   Location loc = operation->getLoc();
   auto linalgOp = dyn_cast<LinalgOp>(operation);
   // Exit out on the memref version of this operation.
-  if (!linalgOp || !linalgOp.hasTensorSemantics())
+  if (!linalgOp || !linalgOp.hasPureTensorSemantics())
     return failure();
 
   auto result = operation->getResult(0);

mlir/lib/Dialect/Linalg/Transforms/Padding.cpp

@@ -168,7 +168,7 @@ linalg::rewriteAsPaddedOp(RewriterBase &rewriter, LinalgOp opToPad,
   }
 
   // TODO: there are cases where we may still want to pad to larger sizes.
-  if (!opToPad.hasTensorSemantics())
+  if (!opToPad.hasPureTensorSemantics())
     return rewriter.notifyMatchFailure(opToPad,
                                        "expected operation on tensors");
 
@@ -265,7 +265,7 @@ mlir::linalg::padAndHoistLinalgOp(RewriterBase &rewriter, LinalgOp linalgOp,
   assert(options.copyBackOp == LinalgPaddingOptions::CopyBackOp::None &&
          "invalid options");
 
-  if (!linalgOp.hasTensorSemantics())
+  if (!linalgOp.hasPureTensorSemantics())
     return rewriter.notifyMatchFailure(
         linalgOp, "only applies to Linalg ops with tensor semantics");
 

mlir/lib/Dialect/Linalg/Transforms/Promotion.cpp

@@ -164,7 +164,8 @@ struct LinalgOpInstancePromotionOptions {
 LinalgOpInstancePromotionOptions::LinalgOpInstancePromotionOptions(
     LinalgOp linalgOp, const LinalgPromotionOptions &options)
     : subViews(), alignment(options.alignment) {
-  assert(linalgOp.hasBufferSemantics() && "revisit usage of shaped operand");
+  assert(linalgOp.hasPureBufferSemantics() &&
+         "revisit usage of shaped operand");
   auto vUseFullTileBuffers =
       options.useFullTileBuffers.value_or(llvm::SmallBitVector());
   vUseFullTileBuffers.resize(linalgOp->getNumOperands(),
@@ -346,7 +347,8 @@ promoteSubViews(ImplicitLocOpBuilder &b,
 static FailureOr<LinalgOp>
 promoteSubViews(ImplicitLocOpBuilder &b, LinalgOp op,
                 LinalgOpInstancePromotionOptions options, DataLayout &layout) {
-  assert(op.hasBufferSemantics() && "expected linalg op with buffer semantics");
+  assert(op.hasPureBufferSemantics() &&
+         "expected linalg op with buffer semantics");
 
   // 1. Promote the specified views and use them in the new op.
   auto promotedBuffersAndViews = promoteSubViews(b, options, layout);
@@ -400,7 +402,7 @@ mlir::linalg::promoteSubviewsPrecondition(Operation *op,
                                           LinalgPromotionOptions options) {
   LinalgOp linalgOp = dyn_cast<LinalgOp>(op);
   // Transformation applies to buffers only.
-  if (!linalgOp || !linalgOp.hasBufferSemantics())
+  if (!linalgOp || !linalgOp.hasPureBufferSemantics())
     return failure();
   // Check that at least one of the requested operands is indeed a subview.
   for (OpOperand &opOperand : linalgOp->getOpOperands()) {

mlir/lib/Dialect/Linalg/Transforms/TilingInterfaceImpl.cpp

@@ -212,7 +212,7 @@ struct LinalgOpTilingInterface
                                              Location loc,
                                              ValueRange ivs) const {
     auto linalgOp = cast<LinalgOp>(op);
-    if (!linalgOp.hasBufferSemantics())
+    if (!linalgOp.hasPureBufferSemantics())
       return op->emitOpError("expected operation to have buffer semantics");
 
     SmallVector<Value> indexedValues;
@@ -256,7 +256,7 @@ struct LinalgOpPartialReductionInterface
     auto linalgOp = cast<LinalgOp>(op);
     OpBuilder::InsertionGuard guard(b);
 
-    if (linalgOp.hasBufferSemantics())
+    if (linalgOp.hasPureBufferSemantics())
       return op->emitOpError("expected operation to have tensor semantics");
     // Insert the new parallel dimension based on the index of the reduction
     // loops. This could be controlled by user for more flexibility.