Re: [PATCH] Fix for PR 37809 and 37807

[Uros Bizjak <ubizjak@gmail.com>]

Hello!

> diff --git a/gcc/testsuite/gcc.target/i386/mmx-8.c b/gcc/testsuite/gcc.target/i386/mmx-8.c
> new file mode 100644
> index 0000000..9d665f0
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/mmx-8.c
> @@ -0,0 +1,136 @@
> +/* PR middle-end/37809 */
> +
> +/* { dg-do run } */
> +/* { dg-options "-O2 -mmmx" } */
> +
> +#include <mmintrin.h>
> +
> +// Various tests of cases where it is incorrect to optimise vectors as if they
> +// were integers of the same width.
> +
> +extern void abort (void);
> +
> +/* #include <stdio.h> */
> +/* #define FAIL(...) printf (__VA_ARGS__); */
> +#define FAIL(...) abort();
>   

Please remove the definitions above and simply call abort() at FAIL 
site. Alternatively, you can check for DEBUG definition to enable 
additional debug information (see i.e. gcc.target/i386/mmx-4.c)

> +
> +
> +void Sshift()
> +{
> +    volatile __m64 y = (__m64) 0xffffffffll;
> +    __m64 x = y & (__m64) 0xffffffffll;
> +    x = _m_psradi (x, 1);
> +    x &= (__m64) 0x80000000ll;
> +    if (0 == (long long) x)
>   

Aliasing violation on x variable. You should use union here and through 
the source. You should check your testcase with -Wstrict-aliasing {,=n}.

> +int main()
> +{
> +    Sshift();
> +    shiftU1();
> +    shiftU2();
> +    shiftU3();
> +    shiftU4();
> +    add();
> +    add2();
> +
> +    mult1();
> +    mult2();
> +    mult3();
> +    div();
> +    return 0;
> +}
>   

Please note, that since this is MMX runtime testcase, you should check 
for MMX support at runtime. Please look into 
gcc.target/i386/builtin-apply-mmx.c or gcc.target/i386/mmx-4.c how to 
use support for runtime check, as present in mmx-check.h header.

Thanks,
Uros.