Add linalg.fill canonicalization pattern for tensor.extract

[suryajasper]

Added the initial canonicalization pattern (mozphab issues with first submission)

[antiagainst]

mlir/lib/Dialect/Linalg/IR/LinalgOps.cpp

639 ↗

(On Diff #543129)

Can we add some tests in test/Dialect/Linalg/canonicalize.mlir for this pattern?

647 ↗

(On Diff #543129)

No wrapping { } per LLVM coding style.

655 ↗

(On Diff #543129)

We don't need to perform extension here as fill then extract shouldn't change the element type anyway. Also, it's not guaranteed that the element types would be floating point.

[antiagainst]

mlir/lib/Dialect/Linalg/IR/LinalgOps.cpp

639 ↗

(On Diff #544570)

We need to register this pattern to getCanonicalizationPatterns.

645 ↗

(On Diff #544570)

Could you capitalize each sentence and close with a .? Also for other comments.

mlir/test/Dialect/Linalg/canonicalize.mlir

341

Can we make this as an input to the function?

353

We need to check that we are returning the scalar value.

[antiagainst]

mlir/test/Dialect/Linalg/canonicalize.mlir

344

Have you tried to run the tests? I don't think we will see this in the transformed IR and so this won't pass?

347

For this line too.

[suryajasper]

Requesting review after addressing revision comments

mlir/lib/Dialect/Linalg/IR/LinalgOps.cpp

639 ↗

(On Diff #544570)

It has been registered already. That change was in the previous diff, so it's included in the full expanded source here

[suryajasper]

Fixed FileCheck prompts for test mlir

[antiagainst]

Ah, I think you only uploaded the last commit?? Can we squash all your local commits into one and reupload it so I can see the full patch? Just make sure the commit message contains "Differential Revision: https://reviews.llvm.org/D156008" when you upload

[suryajasper]

Squashed prior commits for review

[suryajasper]

Squashed prior commits for review (fixed commit message)

mlir/test/Dialect/Linalg/canonicalize.mlir

func.func @fold_fill_reshape_dynamic(%arg0 : tensor<?x?x?x?x?xf32>) -> tensor<?x?xf32> {
   // CHECK: %[[RESULT:.+]] = linalg.fill ins(%{{.+}}{{.*}}outs(%[[RESHAPE]]
   %1 = tensor.collapse_shape %0 [[0, 1, 2], [3, 4]]
       : tensor<?x?x?x?x?xf32> into tensor<?x?xf32>
   // CHECK: return %[[RESULT]]
   return %1 : tensor<?x?xf32>
 }
 
 // -----
-// CHECK-LABEL: func @fold_fill_extract()
+//       CHECK: func @fold_fill_extract
 //  CHECK-SAME:   %[[ARG0:.+]]: i1
 func.func @fold_fill_extract(%arg0 : i1) -> i1 {
   %c0 = arith.constant 0 : index
   %c1 = arith.constant 1 : index
   
-  //  CHECK:   %[[FILL:.+]] = linalg.fill ins(%[[ARG0]]outs(%{{.+}}{{.*}}
   %empty_dynamic = tensor.empty(%c1) : tensor<1x2x3x?xi1>
   %filled = linalg.fill ins(%arg0 : i1) outs(%empty_dynamic : tensor<1x2x3x?xi1>) -> tensor<1x2x3x?xi1>
   
-  //  CHECK:   %[[EXTRACTED:.+]] = tensor.extract %[[FILL]]
   %extracted = tensor.extract %filled[%c0, %c0, %c0, %c0] : tensor<1x2x3x?xi1>
 
   //  CHECK:   return %[[ARG0]]
   return %extracted : i1
 }
 
 // -----
 
 // CHECK: func @fold_self_copy
 func.func @fold_self_copy(%0 : memref<4x16xf32>) {
 // CHECK-NEXT: return
   linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>,
                                    affine_map<(d0, d1) -> (d0, d1)>],
                   iterator_types = ["parallel", "parallel"]}