[Mlir-commits] [mlir] [mlir][bufferization] Introduce reconcileBufferTypeMismatchFn hook (PR #202667)

[Matthias Springer]

================
@@ -83,28 +81,35 @@ struct OpWithUnstructuredControlFlowBufferizableOpInterfaceExternalModel
       if (bufferType == callerType)
         continue;
 
-      // If the computed buffer type does not match the computed buffer type
-      // of the earlier forwarded operands, fall back to a buffer type with a
-      // fully dynamic layout map.
+      // If the computed buffer type does not match the computed buffer type of
+      // the earlier forwarded operands, fall back to a reconciled buffer type.
 #ifndef NDEBUG
-      if (auto rankedTensorType = dyn_cast<RankedTensorType>(tensorType)) {
-        assert(bufferType.hasRank() && callerType.hasRank() &&
-               "expected ranked memrefs");
-        assert(llvm::all_equal({bufferType.getShape(), callerType.getShape(),
-                                rankedTensorType.getShape()}) &&
-               "expected same shape");
-      } else {
-        assert(!bufferType.hasRank() && !callerType.hasRank() &&
-               "expected unranked memrefs");
+      auto tensorType = value.getType();
----------------
matthias-springer wrote:

The code in this `NDEBUG` block looks outdated. We should probably not hard-code tensor / memref types here. I believe with the current design it's not even guaranteed that a tensor type bufferizes to a memref type (?).

I'm not sure how exactly to fix this. Maybe the entire `NDEBUG` section should just be deleted.


https://github.com/llvm/llvm-project/pull/202667