[Mlir-commits] [mlir] 2298eba - [mlir][MemRef] Add ViewLikeOpInterface to memref.transpose (#215331)

[llvmlistbot at llvm.org]

Author: tridhapuku
Date: 2026-08-25T14:19:52-05:00
New Revision: 2298ebad2fb88aa3767026af3bcb14f6bb953bcf

URL: https://github.com/llvm/llvm-project/commit/2298ebad2fb88aa3767026af3bcb14f6bb953bcf
DIFF: https://github.com/llvm/llvm-project/commit/2298ebad2fb88aa3767026af3bcb14f6bb953bcf.diff

LOG: [mlir][MemRef] Add ViewLikeOpInterface to memref.transpose (#215331)

**memref.transpose** is a view-producing op but does not implement
**ViewLikeOpInterface**, unlike _memref.subview, memref.cast,
memref.expand_shape, memref.collapse_shape_, and
_memref.reinterpret_cast_. This causes generic passes that walk view
chains via ViewLikeOpInterface to silently skip transpose ops, producing
incorrect results.

This patch adds `ViewLikeOpInterface` to `MemRef_TransposeOp` and
implements `getViewSource()` to return the input memref, matching the
behaviour of the other view-like ops.

Also adds a round-trip test covering a transpose that carries a
non-default memory space, which is the case that motivated the fix.

`ninja check-mlir` passes locally with no regressions.

**AI tool usage:** The code change (interface declaration,
`getViewSource()` implementation, and test) was authored by me. An AI
assistant was used to help draft the commit message and this PR
description, and to run the local build and test suite.

Signed-off-by: Abhinav Kumar <abhinav.d.kumar at ericsson.com>

Added: 
    

Modified: 
    mlir/include/mlir/Dialect/MemRef/IR/MemRefOps.td
    mlir/lib/Dialect/MemRef/IR/MemRefOps.cpp
    mlir/test/Dialect/MemRef/ops.mlir

Removed: 
    


################################################################################
diff  --git a/mlir/include/mlir/Dialect/MemRef/IR/MemRefOps.td b/mlir/include/mlir/Dialect/MemRef/IR/MemRefOps.td
index 64c10ead05470..5b2794a8aa540 100644
--- a/mlir/include/mlir/Dialect/MemRef/IR/MemRefOps.td
+++ b/mlir/include/mlir/Dialect/MemRef/IR/MemRefOps.td
@@ -2432,6 +2432,7 @@ def SubViewOp : MemRef_OpWithOffsetSizesAndStrides<"subview", [
 def MemRef_TransposeOp : MemRef_Op<"transpose", [
     DeclareOpInterfaceMethods<OpAsmOpInterface, ["getAsmResultNames"]>,
     DeclareOpInterfaceMethods<MemorySpaceCastConsumerOpInterface>,
+    DeclareOpInterfaceMethods<ViewLikeOpInterface>,
     Pure]>,
     Arguments<(ins AnyStridedMemRef:$in, AffineMapAttr:$permutation)>,
     Results<(outs AnyStridedMemRef)> {

diff  --git a/mlir/lib/Dialect/MemRef/IR/MemRefOps.cpp b/mlir/lib/Dialect/MemRef/IR/MemRefOps.cpp
index 720097cc097a1..c8bba20fc7edf 100644
--- a/mlir/lib/Dialect/MemRef/IR/MemRefOps.cpp
+++ b/mlir/lib/Dialect/MemRef/IR/MemRefOps.cpp
@@ -3818,6 +3818,8 @@ static MemRefType inferTransposeResultType(MemRefType memRefType,
           StridedLayoutAttr::get(memRefType.getContext(), offset, strides));
 }
 
+Value TransposeOp::getViewSource() { return getIn(); }
+
 void TransposeOp::build(OpBuilder &b, OperationState &result, Value in,
                         AffineMapAttr permutation,
                         ArrayRef<NamedAttribute> attrs) {

diff  --git a/mlir/test/Dialect/MemRef/ops.mlir b/mlir/test/Dialect/MemRef/ops.mlir
index 3cde68e1631a3..b09a5b95e663d 100644
--- a/mlir/test/Dialect/MemRef/ops.mlir
+++ b/mlir/test/Dialect/MemRef/ops.mlir
@@ -663,3 +663,9 @@ func.func @memref_transpose_map(%src : memref<?x?xf32>) -> memref<?x?xf32, affin
   %dst = memref.transpose %src (i, j) -> (j, i) : memref<?x?xf32> to memref<?x?xf32, affine_map<(d0, d1)[s0] -> (d1 * s0 + d0)>>
   return %dst : memref<?x?xf32, affine_map<(d0, d1)[s0] -> (d1 * s0 + d0)>>
 }
+
+// CHECK-LABEL: func @memref_transpose_map_with_memory_space
+func.func @memref_transpose_map_with_memory_space(%src : memref<?x?xf32, 1>) -> memref<?x?xf32, affine_map<(d0, d1)[s0] -> (d1 * s0 + d0)>, 1> {
+  %dst = memref.transpose %src (i, j) -> (j, i) : memref<?x?xf32, 1> to memref<?x?xf32, affine_map<(d0, d1)[s0] -> (d1 * s0 + d0)>, 1>
+  return %dst : memref<?x?xf32, affine_map<(d0, d1)[s0] -> (d1 * s0 + d0)>, 1>
+}