Testcase for strict-aliasing wanted

Testcase for strict-aliasing wanted

[Osku Salerma]

Does anybody have a simple testcase available that breaks with strict
aliasing enabled? I tried to construct my own, but I can't get either
2.95.2 or the latest CVS version to break intentionally. Either I'm
misunderstanding some of the aliasing rules or gcc isn't optimizing as
aggressively in this case as would be needed for the aliasing issues
to surface.

For reference, here's my (not-working) effort:

#include <stdio.h>

void f(char* p1, int* p2)
{
    char c1,c2;

    c1 = *p1;
    *p2 = 0;
    c2 = *p1;

    printf("(%d,%d)\n", c1,c2);
}

int main(void)
{
    char a[4];

    a[0] = 1;

    f(a, (int*)a);

    return 0;
}

--
Osku Salerma - osku@iki.fi - http://www.iki.fi/osku/
Information wants to be free.

Re: Testcase for strict-aliasing wanted

[Geoff Keating]

Osku Salerma <osku@iki.fi> writes:

> Does anybody have a simple testcase available that breaks with strict
> aliasing enabled? I tried to construct my own, but I can't get either
> 2.95.2 or the latest CVS version to break intentionally. Either I'm
> misunderstanding some of the aliasing rules or gcc isn't optimizing as
> aggressively in this case as would be needed for the aliasing issues
> to surface.
> 
> For reference, here's my (not-working) effort:
> 
> #include <stdio.h>
> 
> void f(char* p1, int* p2)
> {
>     char c1,c2;
> 
>     c1 = *p1;
>     *p2 = 0;
>     c2 = *p1;
> 
>     printf("(%d,%d)\n", c1,c2);
> }
> 
> int main(void)
> {
>     char a[4];
> 
>     a[0] = 1;
> 
>     f(a, (int*)a);
> 
>     return 0;
> }

`char' is a special case; ISO C requires that it be allowed to alias
with anything.  Try the same testcase with `short' instead of `char'.

-- 
- Geoffrey Keating <geoffk@geoffk.org> <geoffk@redhat.com>

Re: Testcase for strict-aliasing wanted

[Osku Salerma]

On 18 Jan 2002, Geoff Keating wrote:
> Osku Salerma <osku@iki.fi> writes:
> > [invalid test case]
> 
> `char' is a special case; ISO C requires that it be allowed to alias
> with anything.  Try the same testcase with `short' instead of `char'.

Yep, it works as expected with 'char' changed to 'short'. I did know
about the fact that char can alias anything, but for some reason I
kept thinking that since I was accessing "char data" through an int
pointer, the compiler would be free to assume such things wouldn't
happen. Totally wrong, of course, for several reasons.

Thanks for the help.

--
Osku Salerma - osku@iki.fi - http://www.iki.fi/osku/
Information wants to be free.

Re: Testcase for strict-aliasing wanted

[mike stump]

> Date: Fri, 18 Jan 2002 21:39:38 +0200 (EET)
> From: Osku Salerma <osku@iki.fi>
> To: gcc@gcc.gnu.org

> Does anybody have a simple testcase available that breaks with strict
> aliasing enabled?

Ok, here is a quick quiz to determine if you know C.  Does the
following program always print:

a)	  12345678	 or	12345678	(depending on endian)
	  12340000		5678

or

b)	12345678
	12345678

or c) neither.

?  Answer at the end.

#include <stdio.h>
 
int
main ()
{
    int a = 0x12345678;
    unsigned short *b = (unsigned short *)&a;
 
    printf ("%x\n", a);
    b[1] = 0;
    printf ("%x\n", a);
 
    return 0;
}

If you said a, you'd be wrong, if you said b, you'd be wrong.  This is
non-portable in ANSI/ISO C.  With the older compilers, those without
type based aliasing, one would always get a, with newer compilers with
more powerful optimizers, it is more common to get b if one optimizes.

[ ... ]


Isn't this in the C programming FAQ yet?  If not, please go over there
and get it added.