[mlir][vector] Take dim sizes into account in DropInnerMostUnitDims. (llvm#71752)

The `stride == 1` does not imply that we can drop it. Because it could
load more than 1 elements. We should also take source sizes and vector
sizes into account. Otherwise it generates invalid IRs. E.g.,

```mlir
func.func @foo(%arg0: memref<1x1xf32>) -> vector<4x8xf32> {
  %c0 = arith.constant 0 : index
  %cst = arith.constant 0.000000e+00 : f32
  %0 = vector.transfer_read %arg0[%c0, %c0], %cst : memref<1x1xf32>, vector<4x8xf32>
  return %0 : vector<4x8xf32>
}
```

Fixes iree-org/iree#15493

Author: hanhanW

mlir/lib/Dialect/Vector/Transforms/VectorTransforms.cpp

@@ -1184,10 +1184,19 @@ class DropInnerMostUnitDims : public OpRewritePattern<vector::TransferReadOp> {
     if (failed(getStridesAndOffset(srcType, srcStrides, srcOffset)))
       return failure();
 
+    // According to vector.transfer_read semantics, the result can be a slice.
+    // It pads the indices with `1` starting from beginning. Thus, we have to
+    // offset the check index with `rankDiff` in `srcStrides` and source dim
+    // sizes.
     size_t dimsToDrop = 0;
-    for (size_t i = 1; i < srcStrides.size(); ++i) {
-      int dim = srcType.getRank() - i - 1;
-      if (srcStrides[dim] == 1) {
+    int rankDiff = srcType.getRank() - targetType.getRank();
+    for (int64_t i = 0, e = targetType.getRank(); i < e; ++i) {
+      // Check that the inner dim size is 1 for both memref/tensor type and
+      // vector slice. It can be folded only if they are 1 and the stride is 1.
+      int dim = targetType.getRank() - i - 1;
+      if (srcStrides[dim + rankDiff] == 1 &&
+          srcType.getDimSize(dim + rankDiff) == 1 &&
+          targetType.getDimSize(dim) == 1) {
         dimsToDrop++;
       } else {
         break;

mlir/test/Dialect/Vector/vector-transfer-collapse-inner-most-dims.mlir

@@ -60,3 +60,19 @@ func.func @contiguous_inner_most_dim_bounds_2d(%A: memref<1000x1x1xf32>, %i:inde
 //      CHECK:   %[[V:.+]] = vector.transfer_read %[[SRC_1]]
 // CHECK-SAME:       {in_bounds = [true]}
 // CHECK-SAME:       vector<4xf32>
+
+// -----
+
+func.func @contiguous_inner_most_dim_out_of_bounds_2d(%arg0: memref<1x1xf32>) -> vector<4x8xf32> {
+  %c0 = arith.constant 0 : index
+  %cst = arith.constant 0.000000e+00 : f32
+  %0 = vector.transfer_read %arg0[%c0, %c0], %cst : memref<1x1xf32>, vector<4x8xf32>
+  return %0 : vector<4x8xf32>
+}
+// The inner most unit dim can not be dropped. In this context, we do not
+// generate rank-reduced memref.subview ops.
+//      CHECK: func.func @contiguous_inner_most_dim_out_of_bounds_2d
+// CHECK-SAME:   %[[SRC:[a-zA-Z0-9]+]]
+//  CHECK-NOT:   memref.subview
+//      CHECK:   %[[READ:.+]] = vector.transfer_read %[[SRC]]
+//      CHECK:   return %[[READ]] : vector<4x8xf32>