[mlir][tensor][bufferize] tensor.empty bufferizes to an allocation (#68080)

Make `tensor.empty` bufferizable, so that the
`-empty-tensor-to-alloc-tensor` pass becomes optional. This makes the
bufferization easier to use. `tensor.empty` used to be non-bufferizable,
so that there two separate ops, one that can be optimized away
(`tensor.empty`) and one that is guaranteed to bufferize to an
allocation (`bufferization.alloc_tensor`). With the recent improvements
of "empty tensor elimination" this is no longer needed and
`bufferization.alloc_tensor` can be phased out.

Author: matthias-springer

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

@@ -253,10 +253,7 @@ struct DimOpInterface
   }
 };
 
-/// Bufferization of tensor.empty. This op does not bufferize, but we need an
-/// interface implementation, so that the result of this op is considered
-/// "writable" (default impl. of `isWritable`). Results of ops that do not
-/// implement `BufferizableOpInterface` are not writable.
+/// Bufferization of "tensor.empty". Replace with "bufferization.alloc_tensor".
 struct EmptyOpInterface
     : public BufferizableOpInterface::ExternalModel<EmptyOpInterface,
                                                     tensor::EmptyOp> {
@@ -268,17 +265,21 @@ struct EmptyOpInterface
 
   LogicalResult bufferize(Operation *op, RewriterBase &rewriter,
                           const BufferizationOptions &options) const {
+    auto emptyOp = cast<tensor::EmptyOp>(op);
+
+    // Optimization: Fold away the op if it has no uses.
     if (op->getUses().empty()) {
       rewriter.eraseOp(op);
       return success();
     }
 
-    // tensor.empty ops are used to indicate the shape of a tensor. They have
-    // no defined contents and cannot be bufferized. However, they can be
-    // converted to bufferization.alloc_tensor ops, which then bufferize to an
-    // allocation (--empty-tensor-to-alloc-tensor).
-    return op->emitOpError("cannot be bufferized, but can be converted to "
-                           "bufferization.alloc_tensor");
+    // Allocate a tensor. This emits a "bufferization.alloc_tensor" op.
+    FailureOr<Value> allocTensor = allocateTensorForShapedValue(
+        rewriter, op->getLoc(), emptyOp.getResult(), options, /*copy=*/false);
+    if (failed(allocTensor))
+      return failure();
+    rewriter.replaceOp(op, *allocTensor);
+    return success();
   }
 };
 

mlir/test/Dialect/Tensor/bufferize.mlir

@@ -1,4 +1,4 @@
-// RUN: mlir-opt %s -tensor-bufferize -cse -split-input-file -verify-diagnostics | FileCheck %s
+// RUN: mlir-opt %s -tensor-bufferize -cse -split-input-file | FileCheck %s
 
 // CHECK-LABEL:   func @dim(
 // CHECK-SAME:              %[[TENSOR:.*]]: tensor<*xf32>,
@@ -62,9 +62,12 @@ func.func @tensor.cast_to_unranked(%arg0: tensor<2xf32>) -> tensor<*xf32> {
 }
 
 // -----
+
+// CHECK-LABEL:   func @tensor.empty(
+// CHECK:           %[[ALLOC:.*]] = memref.alloc() {{.*}} : memref<5xf32>
+// CHECK:           %[[RET:.*]] = bufferization.to_tensor %[[ALLOC]] : memref<5xf32>
+// CHECK:           return %[[RET]] : tensor<5xf32>
 func.func @tensor.empty() -> tensor<5xf32> {
-  // expected-error@+2 {{failed to bufferize op}}
-  // expected-error@+1 {{cannot be bufferized, but can be converted to bufferization.alloc_tensor}}
   %0 = tensor.empty() : tensor<5xf32>
   return %0 : tensor<5xf32>
 }