Re: Question regarding C code generation for a compound condition

[Andrew Haley <aph at redhat dot com>]

On 06/07/2011 02:38 PM, Albert ARIBAUD wrote:

> First of all, apologies if this is not the right list, or the correct
> way, to ask -- just point me in the right direction.
> 
> We have encountered a weird case of gcc C code generation when
> compiling an if statement with a compound condition under -Os -O2; we
> suspect that the generated code is wrong, and we would like to make
> sure before we report a compiler bug...
> 
> The code is as follows:
> 
> -------------------------------------------
> #include <stdio.h>
> 
> struct fields {
>   char        a;
>   char        b;
> };
> 
> 
> struct fields *g;
> 
> void test_function(void)
> {
> 
>   if ( (g->a==0) || (g==NULL) )
>   {
>     g->b = 0;
>     return;
>   }
> 
>   g->b = 10;
> 
>   return;
> 
> }
> -------------------------------------------
> 
> Granted, the way the condition is written is not valid, as the
> evaluation will be done left to right, causing a dereference of g
> before testing it against NULL -- the right order should be "(g==NULL)
> || (g->a==0)", with short-circuit ensuring g->a is only evaluated if g
> != NULL.
> 
> However, even with this admittedly bad order, we would expect both
> sides of the '||' expression to be generated; however, only the left
> side is, both with an ARM or and x86 backend as an objdump extract
> shows:

That's correct.  The dereference of g in (g->a==0) with g == NULL
invokes undefined behaviour.  In the case of undefined behaviour, the
compiler is entitled to do anything, which, obviously, includes not
testing g==NULL.

Andrew.