From: Jeff King Date: Tue, 14 Apr 2026 22:52:06 GMT Subject: Re: [PATCH 06/12] t: prepare execution of potentially failing commands for `set -e` Message-ID: <20260414225206.GA3486072@coredump.intra.peff.net> In-Reply-To: <20260414220347.GA3475127@coredump.intra.peff.net> On Tue, Apr 14, 2026 at 06:03:47PM -0400, Jeff King wrote: > > -perl -MHTTP::Date -e 0 >/dev/null 2>&1 && test_set_prereq DATE_PARSER > > -perl -MTime::ParseDate -e 0 >/dev/null 2>&1 && test_set_prereq DATE_PARSER > > +test_lazy_prereq DATE_PARSER ' > > + perl -MHTTP::Date -e 0 || > > + perl -MTime::ParseDate -e 0 > > +' > > Maybe I am being dense, but I don't see how the original would have > problems with "set -e". The perl command is on the left-hand side of an > &&, so "set -e" will be suppressed. Joy of joys, it looks like a bash bug. Using "command" re-enables "set -e", even on the left-hand side of an &&. With bash 5.3.9 on my Debian system: $ bash -ec 'command false && echo one; echo two' two On bash 3.2.57, tmate'd into the GitHub Actions macOS image: $ bash -ec 'command false && echo one; echo two' [no output, we exited after command failed] It triggers in this case because perl in our test suite is a shell function which runs "command $PERL_PATH". But I don't think switching to an if/then conditional helps. Doing: bash -ec 'if command false; then echo one; fi; echo two' likewise exits early (but prints "two" on modern versions). Sadly I do not think we'll see a fixed version anytime soon. Apple is sticking with ancient bash because of licensing, IIRC. I don't know if they're backporting any fixes (and even if they wanted to, there are probably license complications). Short of requiring a third-party shell, the only workaround I can think of is to manually "set +e" before using "command", and then restore it with "set -e". Gross. -Peff