[Mlir-commits] [mlir] [mlir][bufferization] Empty tensor elimination for materialize_in_destination (PR #65468)

[Matthias Springer]

================
@@ -0,0 +1,159 @@
+//===-- SubsetOpInterface.td - Tensor subsets --------------*- tablegen -*-===//
+//
+// 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
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef SUBSET_OP_INTERFACE
+#define SUBSET_OP_INTERFACE
+
+include "mlir/IR/OpBase.td"
+
+def SubsetOpInterface : OpInterface<"SubsetOpInterface"> {
+  let description = [{
+    This interface can be implemented by ops that insert a source tensor into
+    a destination tensor.
+
+    The elements in the destination tensor that are overwritten by this
+    insertion are called the "subset". How the subset is defined is up to the
+    op. E.g., "tensor.insert_slice" defines the subset via a hyperrectangular
+    slice. A scatter operation could define the subset via a list of indices.
+
+    Ops that deal with tensor subsets come in two flavours:
+    - Insertion flavor: Ops that insert a source tensor into a destination
+      tensor at the specified subset. Such ops usuallt return a new destination
+      tensor and implement the `DestinationStyleOpInterface`. Insertion ops can
+      implement the `SubsetOpInterface`. Example: "tensor.insert_slice"
+    - Extraction flavor: Ops that define a tensor subset. They extract a
+      specified subset from a tensor. There is currently no op interface for
+      such ops. Example: "tensor.extract_slice"
+
+    This interface provides helper methods for efficient bufferization of
+    subset-based tensor IR. Tensor subsets can bufferize to buffer "views"/
+    "aliases" (in contrast to one or multiple less efficient buffer allocation).
+
+    This interface is queried by One-Shot Bufferize to detect cases where a
+    seeming read-after-write is not actually a conflict because the respective
+    ops are operating on equivalent subsets. More details can be found in the
+    documentation of One-Shot Analysis.
+
+    Note: This interface currently assumes that a subset op inserts a single
+    tensor (source) into a destination tensor at a single subset.
+  }];
+  let cppNamespace = "::mlir::bufferization";
+  let methods = [
+      InterfaceMethod<
+        /*desc=*/[{
+          Return the source tensor operand.
+        }],
+        /*retType=*/"::mlir::OpOperand &",
+        /*methodName=*/"getSourceOperand",
+        /*args=*/(ins)
+      >,
+      InterfaceMethod<
+        /*desc=*/[{
+          Return the destination tensor operand.
+        }],
+        /*retType=*/"::mlir::OpOperand &",
+        /*methodName=*/"getDestinationOperand",
+        /*args=*/(ins),
+        /*methodBody=*/"",
+        /*defaultImplementation=*/[{
+          return ::mlir::bufferization::detail::defaultGetDestinationOperand(
+              $_op.getOperation());
+        }]
+      >,
+      InterfaceMethod<
+        /*desc=*/[{
+          Return "true" if this operation inserts into a subset that is
+          equivalent to the subset defined by `candidate`.
+
+          Two subsets are "equivalent" and "same" if they can bufferize to the
+          same buffer views/aliases. If they are "equivalent", the tensor IR
+          may be expressed in terms of different SSA values.
+
+          Example:
+          ```
+          // The subset of the SubsetOpInterface op %1 is equivalent to the
+          // subset defined by %2 (but not "same"):
+          %0 = arith.select %c, %t, %t : tensor<?xf32>
+          %1 = tensor.insert_slice %x into %0[0][5][1]
+              : tensor<5xf32> into tensor<?xf32>
+          %2 = tensor.extract_slice %t[0][5][1] : tensor<?xf32> to tensor<5xf32>
+
+          // The subset of the SubsetOpInterface op %1 is equivalent to and
+          // "same" as the subset defined by %2.
+          %1 = tensor.insert_slice %x into %t[0][5][1]
+              : tensor<5xf32> into tensor<?xf32>
+          %2 = tensor.extract_slice %t[0][5][1] : tensor<?xf32> to tensor<5xf32>
+          ```
+        }],
+        /*retType=*/"bool",
+        /*methodName=*/"isEquivalentSubset",
+        /*args=*/(ins
+            "::mlir::Value":$candidate,
+            "::llvm::function_ref<bool(Value, Value)>":$equivalenceFn)
----------------
matthias-springer wrote:

This would work for `bufferization.materialize_in_destination` because the subset is simply the destination tensor. But `getSubset` may build new IR. E.g., for `tensor.insert_slice`, it will build a new `tensor.extract_slice` op and return that one. I'm going to rename `getSubset` to `buildSubset`.

https://github.com/llvm/llvm-project/pull/65468