[all-commits] [llvm/llvm-project] 50826f: [mlir][MemRef] Add position-based matching heurist...

[Abhishek Varma via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: 50826f9c3b9c31bcc77846ba00ef106d0e0abc83
      https://github.com/llvm/llvm-project/commit/50826f9c3b9c31bcc77846ba00ef106d0e0abc83
  Author: Abhishek Varma <avarma094 at gmail.com>
  Date:   2026-03-05 (Thu, 05 Mar 2026)

  Changed paths:
    M mlir/lib/Dialect/MemRef/IR/MemRefOps.cpp
    M mlir/test/Dialect/MemRef/canonicalize.mlir
    M mlir/test/Dialect/MemRef/fold-memref-alias-ops.mlir

  Log Message:
  -----------
  [mlir][MemRef] Add position-based matching heuristics for rank-reduction with dynamic strides (#184334)

When multiple source dimensions have multiple unit dimensions,
stride-based disambiguation can be wrong with dynamic strides. Add
position-based matching: for each result dimension in order, pick the
leftmost unmatched source dimension with the same size; unmatched source
dims are dropped.

Example: subview from memref<1x8x1x3> to memref<1x8x3>. Both dim 0 and
dim 2 have size 1. Stride-based logic cannot distinguish when strides
are dynamic. Position-based matching correctly drops dim 2 (middle unit
dim) instead of dim 0.

When we have non-trivial static strides, we make use of the stride-based
logic, else we fall back to position-based logic as introduced by this
patch.

INPUT :-
```
func.func @fold_rank_reducing_subview_1x8x1x3_to_1x8x3_drop_middle_unit_dim(
    %arg0 : memref<?x?x?x?xf32, strided<[?, ?, ?, ?], offset: ?>>,
    %arg1 : index, %arg2 : index, %arg3 : index, %arg4 : index) -> f32 {
  %c0 = arith.constant 0 : index
  %0 = memref.subview %arg0[0, 0, 0, 0][1, 8, 1, 3][1, 1, 1, 1]
      : memref<?x?x?x?xf32, strided<[?, ?, ?, ?], offset: ?>> to
        memref<1x8x3xf32, strided<[?, ?, ?], offset: ?>>
  %1 = memref.load %0[%c0, %arg1, %arg2] : memref<1x8x3xf32, strided<[?, ?, ?], offset: ?>>
  return %1 : f32
}
```

WITHOUT this patch we get :-
```
memref.load %0[%c0, %c0, %arg1, %arg2]
```

WITH this patch we get :-
```
memref.load %0[%c0, %arg1, %c0, %arg2]
```

Signed-off-by: Abhishek Varma <abhvarma at amd.com>



To unsubscribe from these emails, change your notification settings at https://github.com/llvm/llvm-project/settings/notifications