perf: Avoid iterating over already processed files with gcov

[krlmlr]

This is the reason that covr is running so slow in igraph. Reduced the check time from over one hour (timeout) to a whopping 8 minutes. Other projects with many C/C++ files will also benefit.

https://github.com/igraph/rigraph/actions/runs/15930556545/job/44938446998

[jimhester]

I think the problem with this change is this comment, if headers are included in multiple files they get tracked only once if you try to process all the object files at once, so you need to run gcov for each object file separately if you want accurate counting. This means you can have tests which exercise a header via one object file, and it might show up as uncovered depending on the order of the object files.

This is also why the coverage counts have now changed in the test.

[krlmlr]

I believe gcov is still run separately for each file, it's only that parse_gcov() (in R) is run once for each file, and not n times for the first file and once for the last file (giving a total of O(n^2) runs). I admit, the patch is a little difficult to read. What am I missing?

[jimhester]

For each compilation unit you need to run gcov, then parse the results, then remove the output files. You can't just run gcov multiple times, then parse all the output files after the fact because the output files get overwritten by each individual run of gcov and you lose coverage information.

[jimhester]

It is possible there is another better way to do this, but this was the only way I found back when I wrote this code originally.

[krlmlr]

Thanks: I thought I saw increasing runtime with each file: https://github.com/igraph/rigraph/actions/runs/15911464511/job/44880531901 . Let me do some experiments and get back to you.

[krlmlr]

It seems that with clean = FALSE we're processing more and more files with each new object. Possible solutions:

• Collect gcov artifacts in separate subdirectories
• Only look for changed files (could be unreliable with some file systems)

[krlmlr]

Now:

• Entirely running inside src/, shorter path names
• Moving generated files out of sight with clean = FALSE
• More verbosity

[krlmlr]

• No more changes to test files
• Example run (12 minutes instead of >1 hour): https://github.com/igraph/rigraph/actions/runs/15950882297/job/44990920986#step:9:8982

[mcol]

This seems unused?