[mlir][vector][bufferize] Better analysis for vector.transfer_write

[springerm]

Summary

The destination operand does not bufferize to a memory read if it is completely overwritten.

[nicolasvasilache]

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

91

you probably also want to bail out if there is a mask for now

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

 /// Bufferization of vector.transfer_write. Replace with a new
 /// vector.transfer_write that operates on a memref.
 ///
 /// Note: DstBufferizableOpInterfaceExternalModel provides many default method
 /// implementations for DestinationStyle ops.
 struct TransferWriteOpInterface
     : public DstBufferizableOpInterfaceExternalModel<TransferWriteOpInterface,
                                                      vector::TransferWriteOp> {
+  bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand,
+                              const AnalysisState &state) const {
+    auto writeOp = cast<vector::TransferWriteOp>(op);
+
+    // Does not bufferize to a memory read if the vector completely overwrites
+    // the buffer.
+
+    // Destination must have static shape.
+    if (!writeOp.getShapedType().hasStaticShape())
+      return true;
+
+    // All offsets must be 0.
+    for (Value offset : writeOp.getIndices()) {
+      if (getConstantIntValue(offset) != 0)
+        return true;
+    }
+
+    // There is no mask.
+    if (writeOp.isMasked())
+      return true;
+
+    // Must write at least the full dimension size.
+    for (auto [d0, d1] : llvm::zip(writeOp.getShapedType().getShape(),
+                                   writeOp.getVectorType().getShape())) {
+      if (d0 > d1)
+        return true;
+    }
+
+    return false;
+  }
+
   LogicalResult bufferize(Operation *op, RewriterBase &rewriter,
                           const BufferizationOptions &options) const {
     auto writeOp = cast<vector::TransferWriteOp>(op);
     assert(isa<TensorType>(writeOp.getShapedType()) &&
            "only tensor types expected");
 
     // Create a new transfer_write on buffer that doesn't have a return value.
     FailureOr<Value> resultBuffer =

mlir/test/Dialect/Vector/one-shot-bufferize.mlir

 // RUN: mlir-opt %s -one-shot-bufferize="bufferize-function-boundaries" -split-input-file | FileCheck %s
+// RUN: mlir-opt %s -one-shot-bufferize="bufferize-function-boundaries test-analysis-only" -split-input-file | FileCheck %s -check-prefix=CHECK-ANALYSIS
 
 // CHECK-LABEL: func @mask(
 //  CHECK-SAME:     %[[t0:.*]]: memref<?xf32, strided<[?], offset: ?>>
 func.func @mask(%t0: tensor<?xf32>, %val: vector<16xf32>, %idx: index, %m0: vector<16xi1>) -> tensor<?xf32> {
   // CHECK-NOT: alloc
   // CHECK-NOT: copy
   //     CHECK: vector.mask %{{.*}} { vector.transfer_write %{{.*}}, %[[t0]][%{{.*}}] : vector<16xf32>, memref<?xf32, strided<[?], offset: ?>> } : vector<16xi1>
   %0 = vector.mask %m0 { vector.transfer_write %val, %t0[%idx] : vector<16xf32>, tensor<?xf32> } : vector<16xi1> -> tensor<?xf32>
   //     CHECK: return %[[t0]]
   return %0 : tensor<?xf32>
 }
+
+// -----
+
+// CHECK-ANALYSIS-LABEL: func @non_reading_xfer_write(
+//  CHECK-ANALYSIS-SAME: tensor<5x10xf32> {bufferization.access = "write"}
+func.func @non_reading_xfer_write(%t: tensor<5x10xf32>, %v: vector<6x11xf32>) -> tensor<5x10xf32> {
+  %c0 = arith.constant 0 : index
+  %1 = vector.transfer_write %v, %t[%c0, %c0] : vector<6x11xf32>, tensor<5x10xf32>
+  return %1 : tensor<5x10xf32>
+}
+// -----
+
+// CHECK-ANALYSIS-LABEL: func @reading_xfer_write(
+//  CHECK-ANALYSIS-SAME: tensor<5x10xf32> {bufferization.access = "read-write"}
+func.func @reading_xfer_write(%t: tensor<5x10xf32>, %v: vector<4x11xf32>) -> tensor<5x10xf32> {
+  %c0 = arith.constant 0 : index
+  %1 = vector.transfer_write %v, %t[%c0, %c0] : vector<4x11xf32>, tensor<5x10xf32>
+  return %1 : tensor<5x10xf32>
+}