Overcounting unreachable lines

[pvdrz]

Hi, we noticed that the a24dcf8 commit had an unintended effect: Some lines that were uncovered before are being marked as covered now even though it is impossible to run them.

Right now I don't have a minimal example but we've noticed that the precondition_check here is being marked as fully covered even though there is a non-unwinding panic on it. This panic cannot be run via the test suite as it is not possible to catch the panic with #[should_panic] nor any other mechanism.

I'll try working on a minimal example in the meantime.

[xd009642]

Hmm interesting. I did have to soften the condition in this check. I wonder how limiting the size to 1 and just exiting the parsing function once it parses a second one would impact the results a24dcf8#diff-20296ee746a64b120f685ceb6daecd09a66a46e6c905a07954beb7bdae166c7eL134

For what it's worth (and I mentioned it to Jyn at the time), the example profdata you shared is the only one I've seen where the number of results != 1. So given that's an anomaly it feels a sensible place to start looking into

[pvdrz]

The particular case I'm working on doesn't have results != 1. I'll generate the profdata file for it.

[xd009642]

One question, why are you folks using the profdata format? With llvm profparsers you can parse the profraws and then do the report gen from that in one go. No need for a merge and then report gen like the LLVM CLI tools.

Not complaining as it benefits the profdata testing, but you might find things work better going off the profraws (at least then any bugs I ran into from tarpaulin users are fixed)