[mlir][bufferization] Support custom types (1/N) (llvm#142986)

Following the addition of TensorLike and BufferLike type interfaces (see
00eaff3), introduce minimal changes
required to bufferize a custom tensor operation into a custom buffer
operation.

To achieve this, new interface methods are added to TensorLike type
interface that abstract away the differences between existing (tensor ->
memref) and custom conversions.

The scope of the changes is intentionally limited (for example,
BufferizableOpInterface is untouched) in order to first understand the
basics and reach consensus design-wise.

---
Notable changes:
* mlir::bufferization::getBufferType() returns BufferLikeType (instead
of BaseMemRefType)
* ToTensorOp / ToBufferOp operate on TensorLikeType / BufferLikeType.
Operation argument "memref" renamed to "buffer"
* ToTensorOp's tensor type inferring builder is dropped (users now need
to provide the tensor type explicitly)

Author: andrey-golubev

mlir/include/mlir/Dialect/Bufferization/IR/BufferizableOpInterface.h

@@ -17,6 +17,7 @@
 #include <optional>
 
 #include "mlir/Dialect/Bufferization/IR/BufferizationEnums.h.inc"
+#include "mlir/Dialect/Bufferization/IR/BufferizationTypeInterfaces.h"
 
 namespace mlir {
 class OpBuilder;
@@ -600,7 +601,7 @@ FailureOr<Value> getBuffer(RewriterBase &rewriter, Value value,
 /// IR, this function can be used.
 ///
 /// This function is a wrapper around BufferizableOpInterface::getBufferType.
-FailureOr<BaseMemRefType> getBufferType(Value value,
+FailureOr<BufferLikeType> getBufferType(Value value,
                                         const BufferizationOptions &options);
 
 /// Return the buffer type for a given Value (tensor) after bufferization
@@ -613,7 +614,7 @@ FailureOr<BaseMemRefType> getBufferType(Value value,
 /// IR, this function can be used.
 ///
 /// This function is a wrapper around `BufferizableOpInterface::getBufferType`.
-FailureOr<BaseMemRefType> getBufferType(Value value,
+FailureOr<BufferLikeType> getBufferType(Value value,
                                         const BufferizationOptions &options,
                                         SmallVector<Value> &invocationStack);
 
@@ -721,6 +722,19 @@ AliasingValueList unknownGetAliasingValues(OpOperand &opOperand);
 /// This is the default implementation of
 /// BufferizableOpInterface::hasTensorSemantics
 bool defaultHasTensorSemantics(Operation *op);
+
+/// This is a helper function used when buffer type is guaranteed to be memref.
+/// It performs two actions: failure state checking and an explicit llvm::cast<>
+/// from the buffer-like type interface to a BaseMemRefType. This allows easier
+/// management of differences in C++ types at the API boundaries. Valid buffer
+/// type is casted to the memref type. Otherwise, the failure state is
+/// propagated i.e. asMemRefType(mlir::failure()) returns mlir::failure().
+FailureOr<BaseMemRefType> asMemRefType(FailureOr<BufferLikeType> bufferType);
+
+/// This function is a free-standing helper that relies on
+/// bufferization::TensorLikeTypeInterface to verify the types in tensor and
+/// buffer worlds match.
+bool typesMatchAfterBufferization(Operation &op, Value tensor, Value buffer);
 } // namespace detail
 
 } // namespace bufferization

mlir/include/mlir/Dialect/Bufferization/IR/BufferizationOps.td

@@ -12,6 +12,7 @@
 include "mlir/Dialect/Bufferization/IR/AllocationOpInterface.td"
 include "mlir/Dialect/Bufferization/IR/BufferViewFlowOpInterface.td"
 include "mlir/Dialect/Bufferization/IR/BufferizableOpInterface.td"
+include "mlir/Dialect/Bufferization/IR/BufferizationTypeInterfaces.td"
 include "mlir/Dialect/Bufferization/IR/BufferizationBase.td"
 include "mlir/Interfaces/DestinationStyleOpInterface.td"
 include "mlir/Interfaces/InferTypeOpInterface.td"
@@ -383,20 +384,31 @@ def Bufferization_DeallocTensorOp : Bufferization_Op<"dealloc_tensor",
 // ToTensorOp
 //===----------------------------------------------------------------------===//
 
+class Bufferization_TensorAndBufferMatch<string tensor, string buffer> : PredOpTrait<
+  "specified tensor and buffer types match",
+  CPred<
+    "::mlir::bufferization::detail::typesMatchAfterBufferization("
+        "$_op, $" # tensor # ", $" # buffer #")"
+  >
+>;
+
 def Bufferization_ToTensorOp : Bufferization_Op<"to_tensor", [
     BufferizableOpInterface,
     SameOperandsAndResultShape,
     SameOperandsAndResultElementType,
-    AllElementTypesMatch<["memref", "result"]>
+    Bufferization_TensorAndBufferMatch<"result", "buffer">
   ]> {
-  let summary = "create a tensor from a `memref`";
+  let summary = "create a buffer-like type from a tensor-like type";
   let description = [{
-    An operation that creates a tensor from a `memref`. The result value is a
-    tensor whose shape and element type match the memref operand.
+    An operation that creates a tensor from a buffer. The result value is a
+    tensor-like type that must match the corresponding buffer-like operand as
+    per TensorLikeType::verifyCompatibleBufferType(). For builtins (TensorType
+    and BaseMemRefType), this means that shapes and element types match between
+    the tensor and the buffer.
 
     The opposite of this op is `to_buffer`. Together, these two ops are
     useful for source/target materializations when doing type conversions
-    involving tensors and memrefs.
+    involving tensors and buffers.
 
     Example:
 
@@ -438,19 +450,16 @@ def Bufferization_ToTensorOp : Bufferization_Op<"to_tensor", [
     away. However, such IR is no longer bufferizable with One-Shot Bufferize.
   }];
 
-  let arguments = (ins Arg<AnyRankedOrUnrankedMemRef,
+  let arguments = (ins Arg<Bufferization_BufferLikeTypeInterface,
                            "the reference to load from",
-                           [MemReadAt<0, FullEffect>]>:$memref,
+                           [MemReadAt<0, FullEffect>]>:$buffer,
                        UnitAttr:$restrict, UnitAttr:$writable);
-  let results = (outs AnyTensor:$result);
+  let results = (outs Bufferization_TensorLikeTypeInterface:$result);
 
   let extraClassDeclaration = [{
     /// The result of a to_tensor is always a tensor.
-    TensorType getType() {
-      Type resultType = getResult().getType();
-      if (::llvm::isa<TensorType>(resultType))
-        return ::llvm::cast<TensorType>(resultType);
-      return {};
+    ::mlir::bufferization::TensorLikeType getType() {
+      return getResult().getType();
     }
 
     //===------------------------------------------------------------------===//
@@ -468,22 +477,15 @@ def Bufferization_ToTensorOp : Bufferization_Op<"to_tensor", [
     FailureOr<BaseMemRefType> getBufferType(
         Value value, const BufferizationOptions &options,
         SmallVector<Value> &invocationStack) {
-      return ::llvm::cast<BaseMemRefType>(getMemref().getType());
+      return ::llvm::cast<BaseMemRefType>(getBuffer().getType());
     }
   }];
 
   let assemblyFormat = [{
-    $memref (`restrict` $restrict^)? (`writable` $writable^)? attr-dict
-      `:` type($memref) `to` type($result)
+    $buffer (`restrict` $restrict^)? (`writable` $writable^)? attr-dict
+      `:` type($buffer) `to` type($result)
   }];
 
-  let builders = [
-    OpBuilder<(ins "Value":$memref, CArg<"bool", "false">:$restrict, CArg<"bool", "false">:$writeable), [{
-      auto rtt = memref::getTensorTypeFromMemRefType(memref.getType());
-      build($_builder, $_state, rtt, memref, restrict, writeable);
-    }]>
-  ];
-
   let hasCanonicalizer = 1;
   let hasFolder = 1;
 }
@@ -498,10 +500,9 @@ def Bufferization_ToBufferOp : Bufferization_Op<"to_buffer", [
     SameOperandsAndResultShape,
     SameOperandsAndResultElementType,
     Pure,
-    AllShapesMatch<["memref", "tensor"]>,
-    AllElementTypesMatch<["memref", "tensor"]>
+    Bufferization_TensorAndBufferMatch<"tensor", "buffer">
   ]> {
-  let summary = "cast a tensor to memref";
+  let summary = "cast a tensor-like type to buffer-like type";
   let description = [{
     An operation that returns the future buffer of a `tensor`.
 
@@ -519,8 +520,8 @@ def Bufferization_ToBufferOp : Bufferization_Op<"to_buffer", [
     the returned buffer) will not be written to.
   }];
 
-  let arguments = (ins AnyTensor:$tensor, UnitAttr:$read_only);
-  let results = (outs AnyRankedOrUnrankedMemRef:$memref);
+  let arguments = (ins Bufferization_TensorLikeTypeInterface:$tensor, UnitAttr:$read_only);
+  let results = (outs Bufferization_BufferLikeTypeInterface:$buffer);
 
   let extraClassDeclaration = [{
     //===------------------------------------------------------------------===//
@@ -554,7 +555,7 @@ def Bufferization_ToBufferOp : Bufferization_Op<"to_buffer", [
   }];
 
   let assemblyFormat = [{
-    $tensor (`read_only` $read_only^)? attr-dict `:` type($tensor) `to` type($memref)
+    $tensor (`read_only` $read_only^)? attr-dict `:` type($tensor) `to` type($buffer)
   }];
 
   let hasFolder = 1;

mlir/include/mlir/Dialect/Bufferization/IR/BufferizationTypeInterfaces.h

@@ -13,6 +13,15 @@
 // Bufferization Type Interfaces
 //===----------------------------------------------------------------------===//
 
+#include "mlir/IR/Diagnostics.h"
+#include "mlir/IR/Types.h"
+
+namespace mlir::bufferization {
+struct BufferizationOptions;
+class BufferizationState;
+class BufferLikeType;
+} // namespace mlir::bufferization
+
 #include "mlir/Dialect/Bufferization/IR/BufferizationTypeInterfaces.h.inc"
 
 #endif // MLIR_DIALECT_BUFFERIZATION_IR_BUFFERIZATIONTYPEINTERFACES_H_

mlir/include/mlir/Dialect/Bufferization/IR/BufferizationTypeInterfaces.td

@@ -21,10 +21,30 @@ def Bufferization_TensorLikeTypeInterface
   let description = [{
     Indicates that this type is a tensor type (similarly to a MLIR builtin
     tensor) for bufferization purposes.
-
-    The interface currently has no methods as it is used by types to opt into
-    being supported by the bufferization procedures.
   }];
+
+  let methods = [
+    InterfaceMethod<[{
+        Returns a BufferLike type for this TensorLike type.
+      }],
+      /*retTy=*/"::mlir::FailureOr<::mlir::bufferization::BufferLikeType>",
+      /*methodName=*/"getBufferType",
+      /*args=*/(ins
+        "const ::mlir::bufferization::BufferizationOptions &":$options,
+        "::llvm::function_ref<::mlir::InFlightDiagnostic()>":$emitError
+      )
+    >,
+    InterfaceMethod<[{
+        Returns whether a BufferLike type is compatible to this TensorLike type.
+        The BufferLike type is assumed to be created by getBufferType().
+      }],
+      /*retTy=*/"::mlir::LogicalResult",
+      /*methodName=*/"verifyCompatibleBufferType",
+      /*args=*/(ins
+        "::mlir::bufferization::BufferLikeType":$bufferType,
+        "::llvm::function_ref<::mlir::InFlightDiagnostic()>":$emitError)
+    >
+  ];
 }
 
 def Bufferization_BufferLikeTypeInterface

mlir/include/mlir/Dialect/Bufferization/IR/UnstructuredControlFlow.h

@@ -65,12 +65,13 @@ struct OpWithUnstructuredControlFlowBufferizableOpInterfaceExternalModel
         // The operand was already bufferized. Take its type directly.
         callerType = memrefType;
       } else {
-        FailureOr<BaseMemRefType> maybeCallerType =
+        FailureOr<BufferLikeType> maybeCallerType =
             bufferization::getBufferType(opOperand->get(), options,
                                          invocationStack);
         if (failed(maybeCallerType))
           return failure();
-        callerType = *maybeCallerType;
+        assert(isa<BaseMemRefType>(*maybeCallerType) && "expected memref type");
+        callerType = cast<BaseMemRefType>(*maybeCallerType);
       }
 
       if (!bufferType) {

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

@@ -159,8 +159,8 @@ struct SelectOpInterface
     // buffers have different types, they differ only in their layout map. Cast
     // both of them to the most dynamic MemRef type.
     if (trueBuffer.getType() != falseBuffer.getType()) {
-      auto targetType =
-          bufferization::getBufferType(selectOp.getResult(), options);
+      auto targetType = bufferization::detail::asMemRefType(
+          bufferization::getBufferType(selectOp.getResult(), options));
       if (failed(targetType))
         return failure();
       if (trueBuffer.getType() != *targetType)
@@ -181,10 +181,12 @@ struct SelectOpInterface
                 SmallVector<Value> &invocationStack) const {
     auto selectOp = cast<arith::SelectOp>(op);
     assert(value == selectOp.getResult() && "invalid value");
-    auto trueType = bufferization::getBufferType(selectOp.getTrueValue(),
-                                                 options, invocationStack);
-    auto falseType = bufferization::getBufferType(selectOp.getFalseValue(),
-                                                  options, invocationStack);
+    auto trueType =
+        bufferization::detail::asMemRefType(bufferization::getBufferType(
+            selectOp.getTrueValue(), options, invocationStack));
+    auto falseType =
+        bufferization::detail::asMemRefType(bufferization::getBufferType(
+            selectOp.getFalseValue(), options, invocationStack));
     if (failed(trueType) || failed(falseType))
       return failure();
     if (*trueType == *falseType)