Thread: flexible array members

gcc 4.6 has now arrived as the default compiler on my desktop, and as
previously reported, it throws a bunch of warnings, foiling my life-long
plan of compiling PostgreSQL with -Werror.

So looking more aggressively into fixing some of these, let's look at
this case:

gistutil.c: In function ‘gistMakeUnionKey’:
gistutil.c:263:16: warning: array subscript is above array bounds [-Warray-bounds]
gistutil.c:268:16: warning: array subscript is above array bounds [-Warray-bounds]
gistutil.c:273:16: warning: array subscript is above array bounds [-Warray-bounds]

The code in question is this:

typedef struct
{
    int32       n;              /* number of elements */
    GISTENTRY   vector[1];      /* variable-length array */
} GistEntryVector;

Not sure why the new gcc is confused about this when -Warray-bounds has
existed for a while.  But thinking a bit further, the "proper" fix for
this would be to use flexible array members like this:

typedef struct
{
    int32       n;              /* number of elements */
    GISTENTRY   vector[];
} GistEntryVector;

This is C99, but with some gentle standard autoconf seasoning, it can be
made transparent.  See attached patch.

Is this a route we want to go down?

It looks as though other compilers could also benefit from this.  clang
throws even more warnings of this kind, and the clang static analyzer
even more.

One thing that is a bit concerning is that throwing more flexible array
members around the code wherever variable-length arrays are used results
in crash and burn.  Probably some places are using sizeof or offsetof on
these structures in incompatible ways.  So each place would have to be
examined separately.

Peter Eisentraut <peter_e@gmx.net> writes:
> Is this a route we want to go down?

> -    GISTENTRY    vector[1];        /* variable-length array */
> +    GISTENTRY    vector[FLEXIBLE_ARRAY_MEMBER];

Yes, I was thinking about the same trick after noting these warnings on
Fedora 15, although personally I'd name the macro VARIABLE_LENGTH_ARRAY.

> One thing that is a bit concerning is that throwing more flexible array
> members around the code wherever variable-length arrays are used results
> in crash and burn.  Probably some places are using sizeof or offsetof on
> these structures in incompatible ways.  So each place would have to be
> examined separately.

Hmm, that's nasty.  But from a code-documentation standpoint I think
this is a useful improvement, so it seems worth doing the work to clean
things up.

(I do recall a number of places that assume that sizeof() includes a
single array element ...)
        regards, tom lane

On ons, 2011-06-15 at 18:19 -0400, Tom Lane wrote:
> Peter Eisentraut <peter_e@gmx.net> writes:
> > Is this a route we want to go down?
> 
> > -     GISTENTRY       vector[1];              /* variable-length
> array */
> > +     GISTENTRY       vector[FLEXIBLE_ARRAY_MEMBER];
> 
> Yes, I was thinking about the same trick after noting these warnings
> on Fedora 15, although personally I'd name the macro
> VARIABLE_LENGTH_ARRAY.

This macro is provided by Autoconf and it appears to be using the
standard's terminology.

Actually, the term "variable-length array" appears to refer to another
C99 feature, namely this one:

void
foo(int n)
{   bar int[n];
   do_something();
}

<span id="IDstID">On Thu, 16 Jun 2011 22:49:45 +0300, Peter Eisentraut <a class="moz-txt-link-rfc2396E"
href="mailto:peter_e@gmx.net"><peter_e@gmx.net></a> wrote:</span><br/><blockquote
cite="mid:1308253785.6721.2.camel@vanquo.pezone.net"type="cite"><pre wrap="">This macro is provided by Autoconf and it
appearsto be using the
 
standard's terminology.
</pre></blockquote><br /> commit dbbba5279f66f95805c1e084e6f646d174931e56 refs/heads/master<br /> Author: Peter
Eisentraut<a class="moz-txt-link-rfc2396E" href="mailto:peter_e@gmx.net"><peter_e@gmx.net></a><br /> Date:   Thu
Jun16 22:39:09 2011 +0300<br /><br /> Periodically checking my VS2010 patch I noticed that this one broke Visual Studio
builds.<br/><br /> At least mine and "<a
href="http://buildfarm.postgresql.org/cgi-bin/show_history.pl?nm=chough&br=HEAD"title="History">chough</a>" in the
buildfarm - I expect others to follow once they rebuild.<br /><br /><pre>error C2065: 'FLEXIBLE_ARRAY_MEMBER' :
undeclaredidentifier
 
error C2057: expected constant expression

Regards,

Brar
</pre><br /><br /><br />