Re: .ONESHELL enhancement?

[David Boyce]

I'm going to try to cut the Gordian knot on at least my little corner
of this (the .ONESHELL suggestion) by clarifying and deepening the
background. This does not necessarily address the bigger questions of
set -e and POSIX and so on, though there seems a good chance of
finding synergy. So, two points:

1. The thing I'm offering to work on is specifically the feature
whereby lines get joined together with an implicit && (or newline in
the presence of -e) without unduly changing makefile semantics. I
referred to this feature as .ONESHELL because it seemed to me clear
that they're the same thing, and it made sense to use the suggested
name.

However, if there's legitimate debate about the intended meaning of
.ONESHELL I'd rather change the name than join the debate. I'm not
offering to implement .ONESHELL regardless of its semantics, I'm
offering to implement certain semantics regardless of the name, so I'm
happy to call the feature
.PASS_ALL_LINES_TO_A_SINGLE_SHELL_WITH_AMPERSANDS if need be. I think
there is a legitimate reason for the feature I propose, as elucidated
below.

2. Perhaps explaining some background will help focus the discussion.
I'm developing a build-auditing tool which I hope to release under GPL
quite soon. It's working nicely already, and in particular it works
well with GNU make. But there's a misfeature of make programs in
general which causes it to behave sub-optimally. I can't address this
problem everywhere but I'd like to be able to address it in the most
widespread version: GNU make.

The following is a common make idiom (taken from a build of bash-3.2 FWIW):

charset.alias:
        /bin/sh ./config.charset 'i386-pc-solaris2.10' > t-charset.alias
        mv t-charset.alias charset.alias

Looking at the makefile, as a human would do, it's obvious that
"t-charset.alias" is a temp file and that the result is the single
file "charset.alias". Or to put it differently, it's clear from the
makefile that the two lines are part of the same logical recipe.

But when looking only at the process tree, as the build auditor does,
there's no indication that the two commands are related at all. The
process tree comes with no marking indicating "recipe boundaries", so
we have to take these as two separate commands which are not known to
be related. This doesn't break the auditing, by the way, it just
causes the auditor to have a less intelligent understanding of what's
going on.

I've spent a little time trying to work out a solution for this. For
instance, I considered trying to have make export $@ to the
environment, so that each command in a recipe would know its final
goal and their results could be collated based on that. But it seemed
more elegant and less intrusive to implement (my conception of)
.ONESHELL. For one thing, it provides benefits to the larger
community, not just that subset which might use the auditor.

The main point here is that though .ONESHELL may provide measurable
performance benefits, my motivation is in the atomicity of treating a
recipe as a logical unit. When the auditor sees in the process tree
something like

        sh -c "touch foo && mv foo bar"

it can figure out exactly what's going on. There's no magic in the &&;
what matters is having a common and exclusive parent process. Of
course a workaround is to "fix" your makefile:

charset.alias:
        /bin/sh ./config.charset 'i386-pc-solaris2.10' > t-charset.alias &&\
        mv t-charset.alias charset.alias

which IMHO is almost always better anyway, but getting people to fix
their makefiles is like bailing with a thimble.

Anyway, I hope this clarifies my proposal. From what I see now, it
seems likely that when the dust settles on the POSIX/set -e issue my
feature will be trivial to implement. The question would be whether to
use the .ONESHELL name or something else. So my work here may devolve
to helping Paul implement the -e stuff.

DSB