On Fri, Jan 15, 2021 at 12:35:15AM +0100, Ævar Arnfjörð Bjarmason wrote:

Makes it useful for something like the git test suite. With vanilla
bash and GIT_TEST_PIPEFAIL=true we'll fail 4 tests in my one-off test.

With my patched bash the only tests we need to skip are those that are
explicitly testing that a piped command returned SIGPIPE.

As Jeff noted in [3] that count isn't reliable, as more will fail in a
way that's hard to reproduce due to the racy nature of vanilla "set -o
pipefail"

Yeah, the count is IMHO not important. Without a way to globally ignore
sigpipe for pipefail, we're left with annotating callers. Which means it
has now become much easier for people to introduce tests which racily
fail. That is much worse than the problem you are trying to solve here. ;)

So it does not matter much how many cases we have. The fact that it
would be easy to introduce new ones makes it unworkable IMHO (without
the bash patch, I mean).

-Peff