[mlir]Fix compose subview (#80551)
I found a bug in `test-compose-subview`,You can see the example I gave.
```
#map = affine_map<() -> ()>
module {
func.func private @fun(%arg0: memref<10x10xf32>, %arg1: memref<5x5xf32>) -> memref<5x5xf32> {
%c0 = arith.constant 0 : index
%c5 = arith.constant 5 : index
%c1 = arith.constant 1 : index
%subview = memref.subview %arg0[0, 0] [5, 5] [1, 1] : memref<10x10xf32> to memref<5x5xf32, strided<[10, 1]>>
%alloc = memref.alloc() : memref<5x5xf32>
scf.for %arg2 = %c0 to %c5 step %c1 {
scf.for %arg3 = %c0 to %c5 step %c1 {
%subview_0 = memref.subview %subview[%arg2, %arg3] [1, 1] [1, 1] : memref<5x5xf32, strided<[10, 1]>> to memref<f32, strided<[], offset: ?>>
%subview_1 = memref.subview %arg1[%arg2, %arg3] [1, 1] [1, 1] : memref<5x5xf32> to memref<f32, strided<[], offset: ?>>
%alloc_2 = memref.alloc() : memref<f32>
linalg.generic {indexing_maps = [#map, #map, #map], iterator_types = []} ins(%subview_0, %subview_1 : memref<f32, strided<[], offset: ?>>, memref<f32, strided<[], offset: ?>>) outs(%alloc_2 : memref<f32>) {
^bb0(%in: f32, %in_4: f32, %out: f32):
%0 = arith.addf %in, %in_4 : f32
linalg.yield %0 : f32
}
%subview_3 = memref.subview %alloc[%arg2, %arg3] [1, 1] [1, 1] : memref<5x5xf32> to memref<f32, strided<[], offset: ?>>
memref.copy %alloc_2, %subview_3 : memref<f32> to memref<f32, strided<[], offset: ?>>
}
}
return %alloc : memref<5x5xf32>
}
func.func @test(%arg0: memref<10x10xf32>, %arg1: memref<5x5xf32>) -> memref<5x5xf32> {
%0 = call @fun(%arg0, %arg1) : (memref<10x10xf32>, memref<5x5xf32>) -> memref<5x5xf32>
return %0 : memref<5x5xf32>
}
}
```
When I run `mlir-opt test.mlir ---test-compose-subview`.
```
test.mlir:14:9: error: 'linalg.generic' op expected operand rank (2) to match the result rank of indexing_map #0 (0)
linalg.generic {indexing_maps = [#map, #map, #map], iterator_types = []} ins(%subview_0, %subview_1 : memref<f32, strided<[], offset: ?>>, memref<f32, strided<[], offset: ?>>) outs(%alloc_2 : memref<f32>) {
^
test1.mlir:14:9: note: see current operation:
"linalg.generic"(%4, %5, %6) <{indexing_maps = [affine_map<() -> ()>, affine_map<() -> ()>, affine_map<() -> ()>], iterator_types = [], operandSegmentSizes = array<i32: 2, 1>}> ({
^bb0(%arg4: f32, %arg5: f32, %arg6: f32):
%8 = "arith.addf"(%arg4, %arg5) <{fastmath = #arith.fastmath<none>}> : (f32, f32) -> f32
"linalg.yield"(%8) : (f32) -> ()
}) : (memref<1x1xf32, strided<[10, 1], offset: ?>>, memref<f32, strided<[], offset: ?>>, memref<f32>) -> ()
```
This PR fixes that.In the meantime I've extended this PR to handle cases
where stride is greater than 1.
```
func.func private @Unknown0(%arg0: memref<10x10xf32>, %arg1: memref<5x5xf32>) -> memref<5x5xf32> {
%c0 = arith.constant 0 : index
%c5 = arith.constant 5 : index
%c1 = arith.constant 1 : index
%subview = memref.subview %arg0[0, 0] [5, 5] [2, 2] : memref<10x10xf32> to memref<5x5xf32, strided<[20, 2]>>
%alloc = memref.alloc() : memref<5x5xf32>
scf.for %arg2 = %c0 to %c5 step %c1 {
scf.for %arg3 = %c0 to %c5 step %c1 {
%subview_0 = memref.subview %subview[%arg2, %arg3] [1, 1] [1, 1] : memref<5x5xf32, strided<[20, 2]>> to memref<f32, strided<[], offset: ?>>
%subview_1 = memref.subview %arg1[%arg2, %arg3] [1, 1] [1, 1] : memref<5x5xf32> to memref<f32, strided<[], offset: ?>>
%alloc_2 = memref.alloc() : memref<f32>
linalg.generic {indexing_maps = [affine_map<() -> ()>, affine_map<() -> ()>, affine_map<() -> ()>], iterator_types = []} ins(%subview_0, %subview_1 : memref<f32, strided<[], offset: ?>>, memref<f32, strided<[], offset: ?>>) outs(%alloc_2 : memref<f32>) {
^bb0(%in: f32, %in_4: f32, %out: f32):
%0 = arith.addf %in, %in_4 : f32
linalg.yield %0 : f32
}
%subview_3 = memref.subview %alloc[%arg2, %arg3] [1, 1] [1, 1] : memref<5x5xf32> to memref<f32, strided<[], offset: ?>>
memref.copy %alloc_2, %subview_3 : memref<f32> to memref<f32, strided<[], offset: ?>>
}
}
return %alloc : memref<5x5xf32>
}
$ mlir-opt test.mlir -test-compose-subview
#map = affine_map<()[s0] -> (s0 * 2)>
#map1 = affine_map<() -> ()>
module {
func.func private @Unknown0(%arg0: memref<10x10xf32>, %arg1: memref<5x5xf32>) -> memref<5x5xf32> {
%c0 = arith.constant 0 : index
%c5 = arith.constant 5 : index
%c1 = arith.constant 1 : index
%alloc = memref.alloc() : memref<5x5xf32>
scf.for %arg2 = %c0 to %c5 step %c1 {
scf.for %arg3 = %c0 to %c5 step %c1 {
%0 = affine.apply #map()[%arg2]
%1 = affine.apply #map()[%arg3]
%subview = memref.subview %arg0[%0, %1] [1, 1] [2, 2] : memref<10x10xf32> to memref<f32, strided<[], offset: ?>>
%subview_0 = memref.subview %arg1[%arg2, %arg3] [1, 1] [1, 1] : memref<5x5xf32> to memref<f32, strided<[], offset: ?>>
%alloc_1 = memref.alloc() : memref<f32>
linalg.generic {indexing_maps = [#map1, #map1, #map1], iterator_types = []} ins(%subview, %subview_0 : memref<f32, strided<[], offset: ?>>, memref<f32, strided<[], offset: ?>>) outs(%alloc_1 : memref<f32>) {
^bb0(%in: f32, %in_3: f32, %out: f32):
%2 = arith.addf %in, %in_3 : f32
linalg.yield %2 : f32
}
%subview_2 = memref.subview %alloc[%arg2, %arg3] [1, 1] [1, 1] : memref<5x5xf32> to memref<f32, strided<[], offset: ?>>
memref.copy %alloc_1, %subview_2 : memref<f32> to memref<f32, strided<[], offset: ?>>
}
}
return %alloc : memref<5x5xf32>
}
}
```
