Re: attribute pure and gcse

[Roger Sayle <roger at eyesopen dot com>]

On Sat, 17 Apr 2004, Dara Hazeghi wrote:
> (resent after no response on gcc-help@)
>
> according to the gcc 3.3 manual, functions marked pure
> should be candidates for common subexpression
> elimination.
>
> In the following case (which seems a good candidate),
> the second call to 'square' is not eliminated by gcse
> (leastaways not with 3.3 on ppc). Is this expected, or
> should cse be doing better here?
>
> static int square(int a) __attribute__ ((pure));
>
> static int square(int a)
> {
>         int j=1, k=1;
>         while(j!=a)
>         {
>                 k = j*k;
>                 j++;
>         }
>         return k;
> }
>
> int main()
> {
>         return square(100) * square(100);
> }


The problem is that in older versions of GCC, including 3.3, the "pure"
attribute wasn't correctly being preserved between the function prototype
and the function definition.

This was fixed by

>> 2003-04-13  Roger Sayle  <roger@eyesopen.com>
>>
>>        * c-decl.c (duplicate_decls): Preserve pure and malloc attributes.


Indeed, your code is correctly optimized by mainline GCC.  I haven't
checked whether this patch made it in time for gcc 3.4, but I can
confirm that it was broken in 3.3.1.


A workaround is to repeat the function prototype including the pure
attribute *after* the function declaration, i.e.

static int square(int a) __attribute__ ((pure));

static int square(int a)
{
        int j=1, k=1;
        while(j!=a)
        {
                k = j*k;
                j++;
        }
        return k;
}

static int square(int a) __attribute__ ((pure));

int main()
{
        return square(100) * square(100);
}


Which allows the second call to square to be optimized away.

I hope this helps.

Roger
--