Re: [PATCH v2 2/4] compat/regex: include alloca.h before undef it

From: Danh Doan <congdanhqx@gmail.com>
To: Ramsay Jones <ramsay@ramsayjones.plus.com>
Cc: git@vger.kernel.org
Subject: Re: [PATCH v2 2/4] compat/regex: include alloca.h before undef it
Date: Mon, 27 Apr 2020 08:08:21 +0700	[thread overview]
Message-ID: <20200427010821.GD14800@danh.dev> (raw)
In-Reply-To: <e9163e0d-2572-e7e0-6aa3-09ce04750b22@ramsayjones.plus.com>

On 2020-04-26 17:17:56+0100, Ramsay Jones <ramsay@ramsayjones.plus.com> wrote:
> 
> 
> On 26/04/2020 01:54, Danh Doan wrote:
> > On 2020-04-25 21:28:05+0100, Ramsay Jones <ramsay@ramsayjones.plus.com> wrote:
> [snip]
> >> OK, I had a quick look at the <alloca.h> header file on a glibc
> >> system (linux) and new-lib system (cygwin) and they both do
> >> (more or less) the same thing: first #undef alloca, and then
> >> if being compiled by gcc, define alloca(size) to be __builtin_alloca(size).
> > 
> > musl people don't do that.
> > They just go ahead define it, if any other header file requires
> > alloca, they will include alloca.h
> > 
> >> So, even if <alloca.h> is #included after regex.c:66, it wouldn't
> >> be a problem. Since I don't have access to a musl based system,
> >> I don't know what that system header is doing.
> > 
> > musl's alloca.h is available here:
> > 
> > https://git.musl-libc.org/cgit/musl/tree/include/alloca.h
> 
> Hmm, OK, so that partly explains the problem. I wonder if the
> musl guys would accept a bug report?

I don't think they have a policy of no `#undef` whatsoever.
But, I think they're picky when come to C-correctly and
POSIX-correctly.
Does C or POSIX define alloca(3) at all?

And, I /think/ they'll likely ignore this one, [musl-faq] says:

	Assuming that including one header will cause symbols from
	another unrelated header to be exposed. This is an application
	bug, and fixing it is as simple as adding the missing #include
	directives. 

I guess they meant, if we have `alloca` defined, we must have included
`alloca.h` somewhere.

> >> However, I said *even if* above, because I don't see why it is trying
> >> to #include <alloca.h> in the first place! ;-)
> > 
> > I looked into my system again. The inclusion chain is:
> > 
> > compat/regex/regex.c:77
> > `-> compat/regex/regex_internal.h:26
> >     `-> /usr/include/stdlib.h:138 [*1*]
> > 
> > [*1*]: https://git.musl-libc.org/cgit/musl/tree/include/stdlib.h#n137
> > 
> > I don't know why _GNU_SOURCE and/or _BSD_SOURCE is defined.
> 
> ... and this explains the main cause. Hmm, as you say, why are
> one (or both) of those pp variables set. :(

Okay, I tracked it down.

compat/regex/regex.c:63
`-> /usr/include/limits.h:4
   `-> /usr/include/features.h:15

https://git.musl-libc.org/cgit/musl/tree/include/features.h?id=8e452abae67db445fb6c3e37cd566c4788c2e8f3#n14
musl defined `_BSD_SOURCE` if none of those below macro was defined:

- _BSD_SOURCE
- _POSIX_SOURCE
- _XOPEN_SOURCE
- _GNU_SOURCE
- __STRICT_ANSI__

I don't think we have any business to define which one of those macros
should be defined in the compat code.

> 
> [snip] 
> >> BTW, why are you compiling with NO_REGEX set anyway?
> > 
> > Because I use musl-libc, and musl-libc doesn't have StartEnd
> 
> Ah, OK. ;-)
> 
> Well, even if the musl project accepted a PR and provided a fix, that
> will not help you in the short term! :D
> 
> Hmm, whatever patch you decide to send (even the original one) I think
> you need to add an explanation of the problem to the commit message,
> including why the patch provides a solution. (You don't have to type
> a novel - see commit bd8f005583 :-P ).
> 
> I haven't thought about this too much, but some options:
> 
>   - iff the musl library sets some kind of identifying pp variable
>     (_MUSL_LIBC_ or somesuch - I haven't looked), then you could
>     make the '#include <alloca.h>' conditional on that variable.
>     This has the benefit of making it obvious to people reading the
>     code that this is specific to musl-libc.

musl's wiki's faq [musl-wiki-faq]:

	Q: Why is there no __MUSL__ macro?

>   - you could simply remove that '#ifdef GAWK' block completely (Lines
>     64->67). We set GAWK and NO_MBSUPPORT  unconditionally in the Makefile
>     so that it compiles (see commit a997bf423d), but these particular
>     lines simply reflect the gawk projects dislike of alloca (along with
>     the desire to catch any attempts to add new calls which are not protected
>     by HAVE_ALLOCA). Given that we are very unlikely to add new calls ...
> 
>   - change the conditional on this block to (totally untested, just typing
>     into my email client) '#if defined(GAWK) && defined(HAVE_ALLOCA)'.
>     This should work, but it does disable the 'catch any attempts to add
>     new _unintentional_ calls' aspect of that block; so you may as well
>     remove it ...

I'll go with your patch in the next email.
<1a0c2b25-e283-9936-1fa2-ce51df1404dc@ramsayjones.plus.com>

[musl-faq]: https://www.musl-libc.org/faq.html
[musl-wiki-faq]: https://wiki.musl-libc.org/faq.html

-- 
Danh