felix314159 wrote:

> Hi there, Before I'd look at the patch, let me ask a couple of clarifying 
> questions.
> 
>     1. Why did you work on this issue? I'd like to understand the motivation 
> first.
> 
>     2. Have you used some AI and to what degree?
> 
>     3. Have you explored the llvm algorithms and datastructures that could 
> help you with an implementation?

Hi, it is my first contribution to LLVM so I am sorry if I messed up sth 
process-related.

1. I grep'd for TODOs in production code and found this. I then felt challenged 
in regards to `CallDescriptionMap` because the existing comment specifically 
proposed a potential performance improvement and I wanted to know whether that 
would be significant or not.

2. Yes, Codex assisted me during the initial investigation when constructing 
the benchmark. I reviewed the resulting changes and test results, and I now 
have updated the PR description with this disclosure.

3. I considered LLVM's data structures and selected `llvm::StringMap` because 
the index is keyed by function name, with `llvm::SmallVector` storing the 
uncommon case of multiple descriptions sharing a name. But I retained the 
original vector because some descriptions require fuzzy or special matching. I 
didn't look into every potentially applicable LLVM algorithm or container, so I 
would appreciate suggestions if anyone has a better idea.

https://github.com/llvm/llvm-project/pull/219441
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits

Reply via email to