Sandro Wenzel

Hi. I understand the motivation behind this transformation (reading many files from CVMFS over the network or local file system is not ideal) and thanks for reacting to my initial...

@mconcas : Yes, I agree we'll need more configurability in any case. The present PR is a first step/investigation that was motivated from the goal to not load upgrade geometry...

Could you please summarize shortly what is the idea behind this change? Will it be backward compatible/optional or will it lead to deep changes?

@Barthelemy: I think the issue should stay open here in parallel. Otherwise, it is hard to track progress and know when it was done ...

Agreed.

I agree that there is room for improvement here and I appreciate the initiative. I remember it simply did not work with just `T*`. Some limitations come in my understanding...

[callgrind_generator.zip](https://github.com/root-project/root/files/15387569/callgrind_generator.zip) Attaching callgrind output files showing that the slowdown comes from: ``` (a) ROOT spends much more time/instructions (roughly a factor 9x) in libCore with ROOT6.30.01. Notably the function TClassEdit::CleanType...

So far we had this patch in our local fork but we would prefer to have it integrated into the mainstream repository.

> LIBRARY_PATH They don't. Just tested it. One of the reasons is certainly that TUnixSystem is not part of Clang.

Just to add, I think the actual timing gain depends heavily on the use-case and software environment. In the end, for us, this is not even the decisive factor. What...