Is that due to ConvertToLLVMPatternInterface?
Yes
Why all?
I expressed myself in a wrong way, let me explain again. I meant that downstream users should at least check all their patterns converting to LLVM, and mandatorly update those working on symbol / symbol tables.
Failing to do so would result in “symbol-table-caching patterns” (the upstream ones, after my eventual PR) to have outdated information, and thus operate in a wrong way when collected through
In other words, this imposes the constraint to “To LLVM” patterns - both upstream and downstream ones - to always keep the symbol tables in a consistent state. Which is exactly the same I’ve beeing doing for the bufferization pipeline, and of course I can replicate here, but this recurring problem is giving me the feeling that it should be taken into consideration more by the infrastructure rather than users.