Re: [PATCH 1/3] protect some clobbered variables with volatile

[Rich Felker <dalias@aerifal.cx>]

On Sun, Feb 10, 2013 at 11:31:41PM +0100, Jens Gustedt wrote:
> When switching optimization to higher levels (-O3) and enable link time
> optimization (-flto) gcc finds two variables that might be clobbered
> accross longjmp (orig_tail in dynlink) or vfork (f in popen):
> 
> src/ldso/dynlink.c:1014:27: warning: variable ‘orig_tail’ might be clobbered by ‘longjmp’ or ‘vfork’ [-Wclobbered]
> src/stdio/popen.c:21:8: warning: variable ‘f’ might be clobbered by ‘longjmp’ or ‘vfork’ [-Wclobbered]
> 
> Trust the analysis of the compiler and protect these variables with
> volatile. Both variables are only loaded once or twice, so this should
> never cause a performance penalty.
> 
> 1	1	src/ldso/dynlink.c
> 1	1	src/stdio/popen.c
> 
> diff --git a/src/ldso/dynlink.c b/src/ldso/dynlink.c
> index efbec8f..e19a21f 100644
> --- a/src/ldso/dynlink.c
> +++ b/src/ldso/dynlink.c
> @@ -1011,7 +1011,7 @@ void __init_ldso_ctors(void)
>  
>  void *dlopen(const char *file, int mode)
>  {
> -	struct dso *volatile p, *orig_tail, *next;
> +	struct dso *volatile p, *volatile orig_tail, *next;

As far as I can tell, this is a false positive. orig_tail is never
modified between setjmp and longjmp. Static analysis is probably
failing due to subsequent modification to orig_tail after the last
possible point at which a longjmp could occur.

> diff --git a/src/stdio/popen.c b/src/stdio/popen.c
> index ed20f5a..e5fbc4f 100644
> --- a/src/stdio/popen.c
> +++ b/src/stdio/popen.c
> @@ -18,7 +18,7 @@ FILE *popen(const char *cmd, const char *mode)
>  {
>  	int p[2], op, i;
>  	pid_t pid;
> -	FILE *f;
> +	FILE *volatile f;
>  	sigset_t old;
>  	const char *modes = "rw", *mi = strchr(modes, *mode);

Could you explain what the issue is here? I'm not following it. I
intend to remove the vfork usage soon anyway, but I'd like to
understand (and commit a patch with a commit-message documenting what
the problem was) if it's wrong right now for reasons other than the
fact that vfork is wrong to begin with. But on the other hand, I don't
want to commit a cargo-cult patch with a message like "because the
compiler warnings said so"...

Rich