[Mlir-commits] [mlir] [MLIR][MemRef] Validate linear size before lowering allocs (PR #179155)

[Matthias Springer]

================
@@ -164,8 +164,11 @@ class AllocOpLowering : public ConvertOpToLLVMPattern<memref::AllocOp> {
     SmallVector<Value, 4> strides;
     Value sizeBytes;
 
-    this->getMemRefDescriptorSizes(loc, memRefType, adaptor.getOperands(),
-                                   rewriter, sizes, strides, sizeBytes, true);
+    if (failed(this->getMemRefDescriptorSizes(loc, memRefType,
+                                              adaptor.getOperands(), rewriter,
+                                              sizes, strides, sizeBytes, true)))
+      return rewriter.notifyMatchFailure(
----------------
matthias-springer wrote:

Here and in the other patterns: `matchAndRewrite` must not return "failure" when the IR has already been modified. This "if" check must be before `getNotalignedAllocFn`.

However, `getNotalignedAllocFn` can also fail, so it must be before `getMemRefDescriptorSizes`.

The size check must be in a separate function that returns `LogicalResult`. `getMemRefDescriptorSizes` should keep returning `void`.


https://github.com/llvm/llvm-project/pull/179155