[Mlir-commits] [mlir] [mlir][vector] Additional transpose folding (PR #138347)

[James Newling]

https://github.com/newling created https://github.com/llvm/llvm-project/pull/138347

Fold transpose with unit-dimensions. Seen in the wild:
```
 %0 = vector.transpose %arg, [0, 2, 1, 3] : vector<6x1x1x4xi8> to vector<6x1x1x4xi8>
```

This transpose can be folded because (1) it preserves the shape and (2) the shuffled dims are unit extent. 

Also addresses comment about static vs anonymous namespace: https://github.com/llvm/llvm-project/pull/135841#discussion_r2071869067

>From 457afd9a319a5b19585e52acab33eae879a07e10 Mon Sep 17 00:00:00 2001
From: James Newling <james.newling at gmail.com>
Date: Fri, 2 May 2025 14:11:29 -0700
Subject: [PATCH] catch additional foldable case

---
 mlir/lib/Dialect/Vector/IR/VectorOps.cpp      | 32 +++++++++---------
 mlir/test/Dialect/Vector/canonicalize.mlir    | 22 -------------
 .../Vector/canonicalize/vector-transpose.mlir | 33 +++++++++++++++++++
 3 files changed, 49 insertions(+), 38 deletions(-)

diff --git a/mlir/lib/Dialect/Vector/IR/VectorOps.cpp b/mlir/lib/Dialect/Vector/IR/VectorOps.cpp
index 0f96442bc3756..dc30741daccd4 100644
--- a/mlir/lib/Dialect/Vector/IR/VectorOps.cpp
+++ b/mlir/lib/Dialect/Vector/IR/VectorOps.cpp
@@ -5575,13 +5575,11 @@ LogicalResult ShapeCastOp::verify() {
   return success();
 }
 
-namespace {
-
 /// Return true if `transpose` does not permute a pair of non-unit dims.
 /// By `order preserving` we mean that the flattened versions of the input and
 /// output vectors are (numerically) identical. In other words `transpose` is
 /// effectively a shape cast.
-bool isOrderPreserving(TransposeOp transpose) {
+static bool isOrderPreserving(TransposeOp transpose) {
   ArrayRef<int64_t> permutation = transpose.getPermutation();
   VectorType sourceType = transpose.getSourceVectorType();
   ArrayRef<int64_t> inShape = sourceType.getShape();
@@ -5601,8 +5599,6 @@ bool isOrderPreserving(TransposeOp transpose) {
   return true;
 }
 
-} // namespace
-
 OpFoldResult ShapeCastOp::fold(FoldAdaptor adaptor) {
 
   VectorType resultType = getType();
@@ -5999,18 +5995,22 @@ OpFoldResult vector::TransposeOp::fold(FoldAdaptor adaptor) {
   if (llvm::dyn_cast_if_present<ub::PoisonAttr>(adaptor.getVector()))
     return ub::PoisonAttr::get(getContext());
 
-  // Eliminate identity transpose ops. This happens when the dimensions of the
-  // input vector remain in their original order after the transpose operation.
-  ArrayRef<int64_t> perm = getPermutation();
-
-  // Check if the permutation of the dimensions contains sequential values:
-  // {0, 1, 2, ...}.
-  for (int64_t i = 0, e = perm.size(); i < e; i++) {
-    if (perm[i] != i)
-      return {};
-  }
+  // Eliminate identity transposes, and more generally any transposes that
+  // preserves the shape without permuting elements.
+  //
+  // Examples of what to fold:
+  // %0 = vector.transpose %arg, [0, 1] : vector<1x1xi8> to vector<1x1xi8>
+  // %0 = vector.transpose %arg, [0, 1] : vector<2x2xi8> to vector<2x2xi8>
+  // %0 = vector.transpose %arg, [1, 0] : vector<1x1xi8> to vector<1x1xi8>
+  //
+  // Example of what NOT to fold:
+  // %0 = vector.transpose %arg, [1, 0] : vector<2x2xi8> to vector<2x2xi8>
+  //
+  if (getSourceVectorType() == getResultVectorType() &&
+      isOrderPreserving(*this))
+    return getVector();
 
-  return getVector();
+  return {};
 }
 
 LogicalResult vector::TransposeOp::verify() {
diff --git a/mlir/test/Dialect/Vector/canonicalize.mlir b/mlir/test/Dialect/Vector/canonicalize.mlir
index 99f0850000a16..974f4506a2ef0 100644
--- a/mlir/test/Dialect/Vector/canonicalize.mlir
+++ b/mlir/test/Dialect/Vector/canonicalize.mlir
@@ -450,28 +450,6 @@ func.func @extract_strided_fold_insert(%a: vector<2x8xf32>, %b: vector<1x4xf32>,
 
 // -----
 
-// CHECK-LABEL: transpose_1D_identity
-// CHECK-SAME: ([[ARG:%.*]]: vector<4xf32>)
-func.func @transpose_1D_identity(%arg : vector<4xf32>) -> vector<4xf32> {
-  // CHECK-NOT: transpose
-  %0 = vector.transpose %arg, [0] : vector<4xf32> to vector<4xf32>
-  // CHECK-NEXT: return [[ARG]]
-  return %0 : vector<4xf32>
-}
-
-// -----
-
-// CHECK-LABEL: transpose_2D_identity
-// CHECK-SAME: ([[ARG:%.*]]: vector<4x3xf32>)
-func.func @transpose_2D_identity(%arg : vector<4x3xf32>) -> vector<4x3xf32> {
-  // CHECK-NOT: transpose
-  %0 = vector.transpose %arg, [0, 1] : vector<4x3xf32> to vector<4x3xf32>
-  // CHECK-NEXT: return [[ARG]]
-  return %0 : vector<4x3xf32>
-}
-
-// -----
-
 // CHECK-LABEL: transpose_3D_identity
 // CHECK-SAME: ([[ARG:%.*]]: vector<4x3x2xf32>)
 func.func @transpose_3D_identity(%arg : vector<4x3x2xf32>) -> vector<4x3x2xf32> {
diff --git a/mlir/test/Dialect/Vector/canonicalize/vector-transpose.mlir b/mlir/test/Dialect/Vector/canonicalize/vector-transpose.mlir
index 91ee0d335ecca..aa227c4628d42 100644
--- a/mlir/test/Dialect/Vector/canonicalize/vector-transpose.mlir
+++ b/mlir/test/Dialect/Vector/canonicalize/vector-transpose.mlir
@@ -247,3 +247,36 @@ func.func @negative_shape_cast_transpose(%arg : vector<6xi8>) -> vector<2x3xi8>
   %1 = vector.transpose %0, [1, 0] : vector<3x2xi8> to vector<2x3xi8>
   return %1 : vector<2x3xi8>
 }
+
+// -----
+
+// Test of transpose folding
+//  CHECK-LABEL: transpose_1D_identity
+//   CHECK-SAME:   [[ARG:%.*]]: vector<4xf32>
+//   CHECK-NEXT:   return [[ARG]]
+func.func @transpose_1D_identity(%arg : vector<4xf32>) -> vector<4xf32> {
+  %0 = vector.transpose %arg, [0] : vector<4xf32> to vector<4xf32>
+  return %0 : vector<4xf32>
+}
+
+// -----
+
+// Test of transpose folding
+// CHECK-LABEL: transpose_2D_identity
+//  CHECK-SAME:    [[ARG:%.*]]: vector<4x3xf32>
+//  CHECK-NEXT:    return [[ARG]]
+func.func @transpose_2D_identity(%arg : vector<4x3xf32>) -> vector<4x3xf32> {
+  %0 = vector.transpose %arg, [0, 1] : vector<4x3xf32> to vector<4x3xf32>
+  return %0 : vector<4x3xf32>
+}
+
+// -----
+
+// Test of transpose folding
+// CHECK-LABEL: transpose_shape_and_order_preserving
+//  CHECK-SAME:    [[ARG:%.*]]: vector<6x1x1x4xi8>
+//  CHECK-NEXT:    return [[ARG]]
+func.func @transpose_shape_and_order_preserving(%arg : vector<6x1x1x4xi8>) -> vector<6x1x1x4xi8> {
+  %0 = vector.transpose %arg, [0, 2, 1, 3] : vector<6x1x1x4xi8> to vector<6x1x1x4xi8>
+  return %0 : vector<6x1x1x4xi8>
+}