[LLVMdev] [RFC] Switching make check to use 'set -o pipefail'

[Rafael Espíndola]

> Hmm, I don't know LLVM's Makefile system well enough to know the
> easiest way to implement an option; if it's non-trivial then maybe
> it's not worth it.

That is my impression at least. These errors are somewhat easy to
introduce, but also easy to fix.

> I also don't know the workflow of most people doing out-of-tree work,
> so I'm not sure how much impact this might have. It can obviously be
> temporarily reverted locally pretty easily, but it assumes people are
> paying attention to LLVMDev/llvm-commits and know what's going on.
> Also, the commit causing all the new failures might not be as obvious
> down the line to someone updating their tree irregularly (which is
> probably true for a lot of academic LLVM users, I'm guessing.)
>
> Just curious, does using pipefail give any information about where in
> the pipe the failure actually comes from? Some kind of message would
> be useful for debugging purposes, in addition to explaining what's
> going on to someone who wasn't watching dev lists and commit messages
> carefully.

With the external (shell based) testing we don't get anything. With
the internal (python based) one we get the last failing command line.
It doesn't say which component of it failed, but the vast majority of
the pipes we have are of the form "command | FileCheck", so it is
clear that "command" must be the one failing since the pipe was
passing before enabling pipefail.

Cheers,
Rafael