Reapply "[mlir][vector] Drop inner unit dims for transfer ops on dynamic shapes." (#80712)

[dcaballe]

This reverts commit b4c7152.

Downstream regression due to another issue that this PR exposes. We have identified the work-items to fix the new issue here: iree-org/iree#16406

[llvmbot]

@llvm/pr-subscribers-mlir-vector

@llvm/pr-subscribers-mlir

Author: Diego Caballero (dcaballe)

Changes

This reverts commit b4c7152.

Downstream regression due to another issue that this PR exposes. We have identified the work-items to fix the new issue here: iree-org/iree#16406

---

Full diff: https://github.com/llvm/llvm-project/pull/81778.diff

2 Files Affected:

• (modified) mlir/lib/Dialect/Vector/Transforms/VectorTransforms.cpp (+8-6)
• (modified) mlir/test/Dialect/Vector/vector-transfer-collapse-inner-most-dims.mlir (+19)

diff --git a/mlir/lib/Dialect/Vector/Transforms/VectorTransforms.cpp b/mlir/lib/Dialect/Vector/Transforms/VectorTransforms.cpp
index 53ae138d1e43a0..74dd1dfaca0da8 100644
--- a/mlir/lib/Dialect/Vector/Transforms/VectorTransforms.cpp
+++ b/mlir/lib/Dialect/Vector/Transforms/VectorTransforms.cpp
@@ -1237,7 +1237,7 @@ class DropInnerMostUnitDimsTransferRead
       return failure();
 
     auto srcType = dyn_cast<MemRefType>(readOp.getSource().getType());
-    if (!srcType || !srcType.hasStaticShape())
+    if (!srcType)
       return failure();
 
     if (!readOp.getPermutationMap().isMinorIdentity())
@@ -1261,19 +1261,21 @@ class DropInnerMostUnitDimsTransferRead
                         targetType.getElementType());
 
     auto loc = readOp.getLoc();
+    SmallVector<OpFoldResult> sizes =
+        memref::getMixedSizes(rewriter, loc, readOp.getSource());
+    SmallVector<OpFoldResult> offsets(srcType.getRank(),
+                                      rewriter.getIndexAttr(0));
+    SmallVector<OpFoldResult> strides(srcType.getRank(),
+                                      rewriter.getIndexAttr(1));
     MemRefType resultMemrefType =
         getMemRefTypeWithDroppingInnerDims(rewriter, srcType, dimsToDrop);
-    SmallVector<int64_t> offsets(srcType.getRank(), 0);
-    SmallVector<int64_t> strides(srcType.getRank(), 1);
-
     ArrayAttr inBoundsAttr =
         readOp.getInBounds()
             ? rewriter.getArrayAttr(
                   readOp.getInBoundsAttr().getValue().drop_back(dimsToDrop))
             : ArrayAttr();
     Value rankedReducedView = rewriter.create<memref::SubViewOp>(
-        loc, resultMemrefType, readOp.getSource(), offsets, srcType.getShape(),
-        strides);
+        loc, resultMemrefType, readOp.getSource(), offsets, sizes, strides);
     auto permMap = getTransferMinorIdentityMap(
         cast<ShapedType>(rankedReducedView.getType()), resultTargetVecType);
     Value result = rewriter.create<vector::TransferReadOp>(
diff --git a/mlir/test/Dialect/Vector/vector-transfer-collapse-inner-most-dims.mlir b/mlir/test/Dialect/Vector/vector-transfer-collapse-inner-most-dims.mlir
index 750879df129b14..3984f17f9e8cdb 100644
--- a/mlir/test/Dialect/Vector/vector-transfer-collapse-inner-most-dims.mlir
+++ b/mlir/test/Dialect/Vector/vector-transfer-collapse-inner-most-dims.mlir
@@ -16,6 +16,25 @@ func.func @contiguous_inner_most_view(%in: memref<1x1x8x1xf32, strided<[3072, 8,
 
 // -----
 
+func.func @contiguous_outer_dyn_inner_most_view(%in: memref<?x1x8x1xf32, strided<[3072, 8, 1, 1], offset: ?>>) -> vector<1x8x1xf32>{
+  %c0 = arith.constant 0 : index
+  %cst = arith.constant 0.0 : f32
+  %0 = vector.transfer_read %in[%c0, %c0, %c0, %c0], %cst {in_bounds = [true, true, true]} : memref<?x1x8x1xf32, strided<[3072, 8, 1, 1], offset: ?>>, vector<1x8x1xf32>
+  return %0 : vector<1x8x1xf32>
+}
+//      CHECK: func @contiguous_outer_dyn_inner_most_view(
+// CHECK-SAME:   %[[SRC:[a-zA-Z0-9]+]]
+//  CHECK-DAG:   %[[C0:.+]] = arith.constant 0 : index
+//  CHECK-DAG:   %[[D0:.+]] = memref.dim %[[SRC]], %[[C0]]
+//      CHECK:   %[[SRC_0:.+]] = memref.subview %[[SRC]][0, 0, 0, 0] [%[[D0]], 1, 8, 1] [1, 1, 1, 1]
+// CHECK-SAME:    memref<?x1x8x1xf32, strided<[3072, 8, 1, 1], offset: ?>> to memref<?x1x8xf32, strided<[3072, 8, 1], offset: ?>>
+//      CHECK:   %[[VEC:.+]] = vector.transfer_read %[[SRC_0]]
+// CHECK-SAME:    memref<?x1x8xf32, strided<[3072, 8, 1], offset: ?>>, vector<1x8xf32>
+//      CHECK:   %[[RESULT:.+]] = vector.shape_cast %[[VEC]]
+//      CHECK:   return %[[RESULT]]
+
+// -----
+
 func.func @contiguous_inner_most_dim(%A: memref<16x1xf32>, %i:index, %j:index) -> (vector<8x1xf32>) {
   %c0 = arith.constant 0 : index
   %f0 = arith.constant 0.0 : f32

[MaheshRavishankar]

Thanks for tracking this down!!

[MacDue]

Hey, this is breaking building dynamically sized matmuls for neon with IREE (see iree-org/iree#16475).

Reduced to this case that produces incorrect IR: https://godbolt.org/z/36Er81G7G