[PATCH] D49003: [Debugify] Export per-pass debug info loss statistics

[Greg Bedwell via Phabricator via llvm-commits]

gbedwell added a comment.

In https://reviews.llvm.org/D49003#1154631, @vsk wrote:

> At a high-level is it fine to use CSV here? I've found it really convenient because Google Sheets, Numbers, python scripts etc. can ingest the data easily.


It works for me personally as we have plenty of tools and scripts internally that can easily operate on CSVs for tracking and graphing.  If the plan is to monitor loss stats over time, would we want to integrate it into LNT or similar for providing continuous results on an upstream bot?  I'm not sure what works best for that use-case if that's the direction we want to go in.

>> Maybe a compelling argument for taking some approach along the lines of the suggested one of attaching undef dbg.values during DCE?
> 
> I'm not sure this is the right tradeoff between improved stats tracking and compile-time overhead.

I agree, although I am still concerned that we lose a lot of the value of debugify tracking through false positives.  If anything, it would have to be on a flag, but that comes with its own set of tradeoffs in terms of the additional code-complexity that it would create and the fact we'd be testing different code paths compared to what our users are actually using, which is never ideal.


https://reviews.llvm.org/D49003