[PATCH 2/3] x32: Properly pass long to syscall [BZ #25810]

[Florian Weimer]

* H. J. Lu via Libc-alpha:

> diff --git a/sysdeps/unix/sysv/linux/x86_64/sysdep.h b/sysdeps/unix/sysv/linux/x86_64/sysdep.h
> index 06e39212ee..93dad97c8f 100644
> --- a/sysdeps/unix/sysv/linux/x86_64/sysdep.h
> +++ b/sysdeps/unix/sysv/linux/x86_64/sysdep.h
> @@ -219,12 +219,14 @@
>  /* Registers clobbered by syscall.  */
>  # define REGISTERS_CLOBBERED_BY_SYSCALL "cc", "r11", "cx"
>  
> -/* Create a variable 'name' based on type 'X' to avoid explicit types.
> -   This is mainly used set use 64-bits arguments in x32.   */
> -#define TYPEFY(X, name) __typeof__ ((X) - (X)) name
> +/* This also works when X is an array.  */
> +#define TYPEFY1(X) __typeof__ ((X) - (X))

I think the comment is now misleading.  For an array X, (X) - (X) is
of type ptrdiff_t, which is signed, so it triggers sign extension, not
zero extension, in the x32 case.  I think this is the reason why my
proposed construct went astray in your experiments.

Is it really necessary to support arrays for system call arguments?

> diff --git a/sysdeps/unix/sysv/linux/x86_64/x32/sysdep.h b/sysdeps/unix/sysv/linux/x86_64/x32/sysdep.h
> index 2b5cf0c0ea..f3bc7e89df 100644
> --- a/sysdeps/unix/sysv/linux/x86_64/x32/sysdep.h
> +++ b/sysdeps/unix/sysv/linux/x86_64/x32/sysdep.h
> @@ -44,6 +44,24 @@
>  # define ZERO_EXTEND_5 movl %r8d, %r8d;
>  # undef ZERO_EXTEND_6
>  # define ZERO_EXTEND_6 movl %r9d, %r9d;
> +#else /*! __ASSEMBLER__ */
> +# undef ARGIFY
> +/* For x32, cast
> +   1. pointer to unsigned long (32 bit).
> +   2. 32-bit unsigned integer to unsigned long long (64 bit).
> +   3. 32-bit integer to long long (64 bit).
> + */
> +# define ARGIFY(X) \
> + ({									\
> +    _Pragma ("GCC diagnostic push");					\
> +    _Pragma ("GCC diagnostic ignored \"-Wpointer-to-int-cast\"");	\
> +    (__builtin_classify_type (X) == 5					\
> +     ? (unsigned long) (X)						\
> +     : ((sizeof (X) <= 4 && (TYPEFY1 (X)) 0 < (TYPEFY1 (X)) -1)		\
> +	? (unsigned long long) (X)					\
> +	: (long long) (X)));						\
> +    _Pragma ("GCC diagnostic pop");					\
> +  })
>  #endif	/* __ASSEMBLER__ */

This should use long int instead long everywhere.
__builtin_classify_type is undocumented.  I'm not sure if its behavior
and the constant 5 is actually stable across GCC versions, but the
powerpc uses it as well, for a slightly difference purpose (to
recognize array arguments, it seems).