[Mlir-commits] [mlir] [mlir] Make symbol-user type verification proportional to participating IR (PR #212354)

[Jared Hoberock]

jaredhoberock wrote:

Correct me if I'm wrong, but there appears to be nothing like a public `walkSymbolsIn(op|attr|type)` that offers a straightforward way to visit the symbols in an object. If there was something like that, I think the complexity of the symbol table verifier could collapse and any acceleration/pruning could live in a single place behind that function.

The kind-level bit sounds interesting as a partial solution for statically-classifiable objects. There's still the gap left by instance-sensitive containers like DictionaryAttr.

In any case, these bits born as early as possible and stored close to the object of interest is probably the most principled shape, if intrusive. It eliminates all the complexity and performance concerns of caching and locking.

So I'm for this direction if maintainers feel like it's a worthwhile trade-off.

https://github.com/llvm/llvm-project/pull/212354