xref: /llvm-project/mlir/test/Dialect/Linalg/detensorize_if.mlir (revision 441b672bbdc68ad88036f3e258759854c8283adb)
1// RUN: mlir-opt %s -split-input-file -allow-unregistered-dialect -pass-pipeline="builtin.module(func.func(linalg-detensorize))" | FileCheck %s
2
3#map0 = affine_map<() -> ()>
4
5#attrs = {
6  indexing_maps = [#map0, #map0, #map0],
7  iterator_types = []
8}
9
10func.func @main() -> (tensor<i32>) attributes {} {
11  %c0 = arith.constant 0 : i32
12  %0 = tensor.from_elements %c0 : tensor<i32>
13  %c10 = arith.constant 10 : i32
14  %1 = tensor.from_elements %c10 : tensor<i32>
15  cf.br ^bb1(%0 : tensor<i32>)
16
17^bb1(%2: tensor<i32>):  // 2 preds: ^bb0, ^bb2
18  %3 = tensor.empty() : tensor<i1>
19  %4 = linalg.generic #attrs
20    ins(%2, %1 : tensor<i32>, tensor<i32>)
21    outs(%3 : tensor<i1>) {
22    ^bb0(%arg0: i32, %arg1: i32, %arg2: i1):
23      %8 = arith.cmpi slt, %arg0, %arg1 : i32
24      linalg.yield %8 : i1
25  } -> tensor<i1>
26  %5 = tensor.extract %4[] : tensor<i1>
27  cf.cond_br %5, ^bb2(%2 : tensor<i32>), ^bb3(%2 : tensor<i32>)
28
29^bb2(%6: tensor<i32>):  // pred: ^bb1
30  %7 = tensor.empty() : tensor<i32>
31  %8 = linalg.generic #attrs
32    ins(%6, %6 : tensor<i32>, tensor<i32>)
33    outs(%7 : tensor<i32>) {
34    ^bb0(%arg0: i32, %arg1: i32, %arg2: i32):
35      %9 = arith.addi %arg0, %arg1 : i32
36      linalg.yield %9 : i32
37  } -> tensor<i32>
38  cf.br ^bb3(%8 : tensor<i32>)
39
40^bb3(%10: tensor<i32>):  // pred: ^bb1
41  return %10 : tensor<i32>
42}
43
44// CHECK-LABEL:  func @main()
45// CHECK-DAG:     %[[cst:.*]] = arith.constant dense<0>
46// CHECK-DAG:     arith.constant true
47// CHECK:         cf.br
48// CHECK-NEXT:   ^[[bb1:.*]]:
49// CHECK-NEXT:     cf.cond_br %{{.*}}, ^[[bb2:.*]], ^bb3
50// CHECK-NEXT:   ^[[bb2]]
51// CHECK-NEXT:     cf.br ^[[bb3:.*]]
52// CHECK-NEXT:   ^[[bb3]]
53// CHECK-NEXT:     return %[[cst]]
54// CHECK-NEXT:   }
55
56// -----
57
58// Similar to the above test with one change: one of the block after the
59// if-condition passes/forwards its tensor argument to another block.
60
61#map0 = affine_map<() -> ()>
62
63#attrs = {
64  indexing_maps = [#map0, #map0, #map0],
65  iterator_types = []
66}
67
68func.func @main() -> (tensor<i32>) attributes {} {
69  %c0 = arith.constant 0 : i32
70  %0 = tensor.from_elements %c0 : tensor<i32>
71  %c10 = arith.constant 10 : i32
72  %1 = tensor.from_elements %c10 : tensor<i32>
73  cf.br ^bb1(%0 : tensor<i32>)
74
75^bb1(%2: tensor<i32>):  // 2 preds: ^bb0, ^bb2
76  %3 = tensor.empty() : tensor<i1>
77  %4 = linalg.generic #attrs
78    ins(%2, %1 : tensor<i32>, tensor<i32>)
79    outs(%3 : tensor<i1>) {
80    ^bb0(%arg0: i32, %arg1: i32, %arg2: i1):
81      %8 = arith.cmpi slt, %arg0, %arg1 : i32
82      linalg.yield %8 : i1
83  } -> tensor<i1>
84  %5 = tensor.extract %4[] : tensor<i1>
85  cf.cond_br %5, ^bb2(%2 : tensor<i32>), ^bb3(%2 : tensor<i32>)
86
87^bb2(%6: tensor<i32>):  // pred: ^bb1
88  %7 = tensor.empty() : tensor<i32>
89  %8 = linalg.generic #attrs
90    ins(%6, %6 : tensor<i32>, tensor<i32>)
91    outs(%7 : tensor<i32>) {
92    ^bb0(%arg0: i32, %arg1: i32, %arg2: i32):
93      %9 = arith.addi %arg0, %arg1 : i32
94      linalg.yield %9 : i32
95  } -> tensor<i32>
96  cf.br ^bb3(%8 : tensor<i32>)
97
98^bb3(%10: tensor<i32>):  // pred: ^bb1
99  cf.br ^bb4(%10 : tensor<i32>)
100
101^bb4(%11: tensor<i32>):  // pred: ^bb1
102  return %11 : tensor<i32>
103}
104
105// CHECK-LABEL:  func @main()
106// CHECK-DAG:     %[[cst:.*]] = arith.constant dense<0>
107// CHECK-DAG:     arith.constant true
108// CHECK:         cf.br ^[[bb1:.*]]
109// CHECK-NEXT:   ^[[bb1:.*]]:
110// CHECK-NEXT:     cf.cond_br %{{.*}}, ^[[bb2:.*]], ^bb3
111// CHECK-NEXT:   ^[[bb2]]:
112// CHECK-NEXT:     cf.br ^[[bb3:.*]]
113// CHECK-NEXT:   ^[[bb3]]:
114// CHECK-NEXT:     cf.br ^[[bb4:.*]]
115// CHECK-NEXT:   ^[[bb4]]:
116// CHECK-NEXT:     return %[[cst]]
117// CHECK-NEXT:   }
118
119// -----
120
121#map0 = affine_map<() -> ()>
122
123#attrs = {
124  indexing_maps = [#map0, #map0, #map0],
125  iterator_types = []
126}
127
128func.func @main() -> (tensor<i32>) attributes {} {
129  %c0 = arith.constant 0 : i32
130  %0 = tensor.from_elements %c0 : tensor<i32>
131  %c10 = arith.constant 10 : i32
132  %1 = tensor.from_elements %c10 : tensor<i32>
133  cf.br ^bb1(%0 : tensor<i32>)
134
135^bb1(%2: tensor<i32>):  // 2 preds: ^bb0, ^bb2
136  %3 = tensor.empty() : tensor<i1>
137  %4 = linalg.generic #attrs
138    ins(%2, %1 : tensor<i32>, tensor<i32>)
139    outs(%3 : tensor<i1>) {
140    ^bb0(%arg0: i32, %arg1: i32, %arg2: i1):
141      %8 = arith.cmpi slt, %arg0, %arg1 : i32
142      linalg.yield %8 : i1
143  } -> tensor<i1>
144  %5 = tensor.extract %4[] : tensor<i1>
145  // This cf.cond_br intentionally has bb2 as it's target for both branches. This
146  // is to make sure that the "forward phase" of the cost-model correctly adds
147  // the users of a block argument (in this case bb2's argument) to the work
148  // list.
149  cf.cond_br %5, ^bb2(%2 : tensor<i32>), ^bb2(%2 : tensor<i32>)
150
151^bb2(%6: tensor<i32>):  // pred: ^bb1
152  %12 = tensor.from_elements %c10 : tensor<i32>
153  %7 = tensor.empty() : tensor<i32>
154  %8 = linalg.generic #attrs
155    ins(%6, %12 : tensor<i32>, tensor<i32>)
156    outs(%7 : tensor<i32>) {
157    ^bb0(%arg0: i32, %arg1: i32, %arg2: i32):
158      %9 = arith.addi %arg0, %arg1 : i32
159      linalg.yield %9 : i32
160  } -> tensor<i32>
161  cf.br ^bb3(%8 : tensor<i32>)
162
163^bb3(%10: tensor<i32>):  // pred: ^bb1
164  return %10 : tensor<i32>
165}
166
167// CHECK-LABEL:  func @main()
168// CHECK-DAG:     %[[cst:.*]] = arith.constant dense<10>
169// CHECK-DAG:     arith.constant true
170// CHECK:         cf.br ^[[bb1:.*]]
171// CHECK-NEXT:   ^[[bb1]]:
172// CHECK-NEXT:     cf.cond_br %{{.*}}, ^[[bb2:.*]], ^bb2
173// CHECK-NEXT:   ^[[bb2]]
174// CHECK-NEXT:     cf.br ^[[bb3:.*]]
175// CHECK-NEXT:   ^[[bb3]]
176// CHECK-NEXT:     return %[[cst]]
177// CHECK-NEXT:   }
178