Re: [PATCH v2] tools/nolibc: add support for Alpha

On 2026-08-02 09:06:59+0200, Magnus Lindholm wrote:
> On Wed, Jul 22, 2026 at 3:28 AM Thomas Weißschuh <[email protected]> wrote:
> 
> > + * Syscalls for Alpha:
> > + *   - registers are 64-bit
> > + *   - syscall number is passed in $0/v0
> > + *   - the system call is performed by calling callsys
> > + *   - syscall return comes in $0/v0, error flag in $19/a4
> 
> typo in comment ? should be:
>  syscall return comes in $0/v0, error flag in $19/a3

Ack.

> > + *   - arguments are passed in $16/a0 to $21/a5
> > + *   - GCC does not support symbol register names
> > + */
> 
> > +#define __nolibc_syscall5(num, arg1, arg2, arg3, arg4, arg5)                  \
> > +({                                                                            \
> > +       register long _num __asm__ ("$0") = (num);                            \
> > +       register long _err __asm__ ("$19");                                   \
> > +       register long _arg1 __asm__ ("$16") = (long)(arg1);                   \
> > +       register long _arg2 __asm__ ("$17") = (long)(arg2);                   \
> > +       register long _arg3 __asm__ ("$18") = (long)(arg3);                   \
> > +       register long _arg4 __asm__ ("$19") = (long)(arg4);                   \
> > +       register long _arg5 __asm__ ("$20") = (long)(arg5);                   \
> > +                                                                              \
> > +       __asm__ volatile (                                                    \
> > +               "callsys"                                                     \
> > +               : "+r"(_num), "=r"(_err)                                      \
> > +               : "r"(_arg1), "r"(_arg2), "r"(_arg3), "r"(_arg4), "r"(_arg5)  \
> > +               : _NOLIBC_SYSCALL_CLOBBERLIST, "$21"                          \
> > +       );                                                                    \
> > +       _err ? -_num : _num;                                                  \
> > +})
> 
> One observation, although I am not an expert on GCC inline-assembly constraints:
> in the four- through six-argument wrappers, $19/a3 contains the fourth
> argument on
> entry and the error indicator on return. Would using a single read/write operand
> for $19, similar to the handling of $0, make that transition clearer?

Good catch. We had miscompilations from clang in this case before.
So I'll fix this up to be on the safe side.

> Michael’s successful testing on both XP1000 and ES45 also shows that the current
> implementation works correctly on real hardware, so I do not consider
> changing the
> operand representation a requirement for v3. From my side, addressing the $19/a4
> comment typo, together with the "TASK_SIZEi" typo already noted by Willy,
> would be sufficient.
> 
> Michael’s testing should be sufficient, but I also have access to Alpha hardware
> and would be happy to assist with any additional testing if useful.

Thanks for the offer. For me Michael's Tested-by would be sufficient.


Thomas