[Mlir-commits] [mlir] [MLIR][Affine] Normalize memref.alloc ops with non trivial layout map (PR #129875)

[Uday Bondhugula]

================
@@ -1819,8 +1817,40 @@ LogicalResult mlir::affine::normalizeMemRef(AllocLikeOp *allocOp) {
         b.create<AllocLikeOp>(allocOp->getLoc(), newMemRefType, newDynamicSizes,
                               allocOp->getAlignmentAttr());
   } else {
-    newAlloc = b.create<AllocLikeOp>(allocOp->getLoc(), newMemRefType,
-                                     allocOp->getAlignmentAttr());
+    mlir::ValueRange dynamicSizes = allocOp->getDynamicSizes();
+    mlir::ValueRange symbolOperands = allocOp->getSymbolOperands();
+    ArrayRef<int64_t> newShape = newMemRefType.getShape();
+    ArrayRef<int64_t> oldShape = memrefType.getShape();
+    SmallVector<Value> mapOperands(oldShape.size() + symbolOperands.size());
+    SmallVector<Value> dimensionOperands;
+    unsigned dimId = 0, symId = 0;
+    // Collect all the map operands of `allocOp` (both dynamic sizes and symbol
+    // operands), which will help us to compute the dynamic sizes of the new
+    // alloc op we are going to create.
+    for (unsigned i = 0, e = oldShape.size(); i < e; i++) {
+      if (oldShape[i] == ShapedType::kDynamic)
----------------
bondhugula wrote:

Use isDynamicDim

https://github.com/llvm/llvm-project/pull/129875