1// RUN: mlir-opt %s -eliminate-empty-tensors -empty-tensor-to-alloc-tensor -one-shot-bufferize="bufferize-function-boundaries test-analysis-only" -split-input-file | FileCheck %s
2
3// CHECK-LABEL: func @buffer_forwarding_conflict
4func.func @buffer_forwarding_conflict(%arg0: tensor<?xf32> {bufferization.writable = true}, %arg1: index) -> (tensor<?xf32>, tensor<?xf32>) {
5  %cst = arith.constant 0.000000e+00 : f32
6  //      CHECK: tensor.extract_slice
7  // CHECK-SAME: {__inplace_operands_attr__ = ["false", "none"]
8  // Instead of allocating, share buffer with some inplace bufferization?
9  %0 = tensor.empty(%arg1) : tensor<?xf32>
10
11  //      CHECK: linalg.fill
12  // CHECK-SAME: {__inplace_operands_attr__ = ["none", "true"]
13  %1 = linalg.fill ins(%cst : f32) outs(%0 : tensor<?xf32>) -> tensor<?xf32>
14
15  //      CHECK: tensor.insert_slice
16  // CHECK-SAME: {__inplace_operands_attr__ = ["true", "false", "none"]
17  %2 = tensor.insert_slice %1 into %arg0[0] [%arg1] [1] : tensor<?xf32> into tensor<?xf32>
18
19  //      CHECK: tensor.insert_slice
20  // CHECK-SAME: {__inplace_operands_attr__ = ["true", "true", "none"]
21  %3 = tensor.insert_slice %1 into %arg0[42] [%arg1] [1] : tensor<?xf32> into tensor<?xf32>
22
23  //      CHECK: return
24  // CHECK-SAME: __equivalent_func_args__ = [-1, 0]
25  return %2, %3 : tensor<?xf32>, tensor<?xf32>
26}
27
28// -----
29
30// CHECK-LABEL: func @buffer_forwarding_no_conflict
31func.func @buffer_forwarding_no_conflict(%arg0: tensor<?xf32> {bufferization.writable = true}, %arg1: index) -> (tensor<?xf32>, tensor<?xf32>) {
32  %cst = arith.constant 0.000000e+00 : f32
33  //      CHECK: tensor.extract_slice
34  // CHECK-SAME: {__inplace_operands_attr__ = ["true", "none"]
35  // Instead of allocating, share buffer with some inplace bufferization?
36  %0 = tensor.empty(%arg1) : tensor<?xf32>
37
38  //      CHECK: linalg.fill
39  // CHECK-SAME: {__inplace_operands_attr__ = ["none", "true"]
40  %1 = linalg.fill ins(%cst : f32) outs(%0 : tensor<?xf32>) -> tensor<?xf32>
41
42  //      CHECK: tensor.insert_slice
43  // CHECK-SAME: {__inplace_operands_attr__ = ["true", "true", "none"]
44  %2 = tensor.insert_slice %1 into %arg0[42] [%arg1] [1] : tensor<?xf32> into tensor<?xf32>
45
46  //      CHECK: return
47  // CHECK-SAME: __equivalent_func_args__ = [0, 0]
48  return %2, %2 : tensor<?xf32>, tensor<?xf32>
49}
50
51// -----
52
53// CHECK-LABEL: func @buffer_forwarding_conflict_with_different_element_type
54func.func @buffer_forwarding_conflict_with_different_element_type(%arg0: tensor<?xf32> {bufferization.writable = true}, %arg1: index) -> (tensor<?xf32>, tensor<?xf32>) {
55  //      CHECK: tensor.extract_slice
56  // CHECK-SAME: {__inplace_operands_attr__ = ["true", "none"]
57  %cst = arith.constant 0.000000e+00 : f32
58  //      CHECK: bufferization.alloc_tensor(%arg1)
59  %0 = tensor.empty(%arg1) : tensor<?xf32>
60
61  //      CHECK: bufferization.alloc_tensor(%arg1)
62  %1 = tensor.empty(%arg1) : tensor<?xbf16>
63
64  //      CHECK: linalg.copy
65  // CHECK-SAME: {__inplace_operands_attr__ = ["true", "true"]
66  %2 = linalg.copy ins(%0 : tensor<?xf32>) outs(%1 : tensor<?xbf16>) -> tensor<?xbf16>
67
68  //      CHECK: linalg.copy
69  // CHECK-SAME: {__inplace_operands_attr__ = ["true", "true"]
70  %3 = linalg.copy ins(%2 : tensor<?xbf16>) outs(%0 : tensor<?xf32>) -> tensor<?xf32>
71
72  //      CHECK: tensor.insert_slice
73  // CHECK-SAME: {__inplace_operands_attr__ = ["true", "true", "none"]
74  %4 = tensor.insert_slice %3 into %arg0[42] [%arg1] [1] : tensor<?xf32> into tensor<?xf32>
75
76  //      CHECK: return
77  // CHECK-SAME: __equivalent_func_args__ = [0, 0]
78  return %4, %4 : tensor<?xf32>, tensor<?xf32>
79}
80