On Mon, Dec 20 2021, Junio C Hamano wrote: > Ævar Arnfjörð Bjarmason <avarab@xxxxxxxxx> writes: > >> I think for this and 1/2 it would be really nice to pick up a version of >> Hamza's CI changes: >> https://lore.kernel.org/git/20211118084143.279174-2-someguy@xxxxxxxxxxxxxxxxxxx/ > > Are there so many incompatible versions of pcre2 library whose usage > are subtly different that we need to protect ourselves with multiple > variants in CI from breakages? > > I doubt pcre2 library was _that_ bad. It's really not, but: * We have an optional >=10.34 feature we use (albeit trivial) * We have an optional >=10.36 feature we use (major, and directly related) * We might be targeting JIT or not, and the error handling isn't the same (known PCRE caveat) * We might be targeting a PCRE that knows about Unicode, or not * We use it in a mode where we might feed UTF-8 invalid data into the UTF-8 mode Any update to the relevant code really needs to test the combination of those, so it's a perfect target for CI to make that less tedious. > Adding a special task that builds with the minimum version we > support may not be too bad, but the library should be stable enough > to allow us to declare it sufficient to test the most common version > with the most common build options in our ordinary build job(s). That's a nice idea, but not the reality of the situation. Unless we're willing to bump the version requirement & insist on JIT && Unicode support before using it. The CI itself should be realtively cheap, and just runs the few tests that would spot any breakages with the above.
