Re: RE: [PATCH] MIPS64 atomic_arch.h Clang complains about input type

[Szabolcs Nagy <nsz@port70.net>]

* Rich Felker (dalias@libc.org) <dalias@libc.org> [2016-03-08 12:26:50 -0500]:
> On Tue, Mar 08, 2016 at 12:26:13PM +0100, Szabolcs Nagy wrote:
> > * Jaydeep Patil <Jaydeep.Patil@imgtec.com> [2016-03-08 10:09:43 +0000]:
> > > 
> > > ./arch/mips64/atomic_arch.h:37:29: error: unsupported inline asm: input with type 'long *' matching output with type 'int'
> > > 
> > 
> > clang probably should not complain about this
> > 
> > (it's possible to have the output in the low 32bit of a 64bit input reg
> > and then you do want mismatching input/output types for the same reg)
> 
> I can imagine it being problematic on other archs where the register
> names are different for 32-bit and 64-bit versions -- %0 couldn't
> expand to both. Isn't there a similar ugly issue on aarch64's
> atomic_arch.h?
> 

in aarch64 asm wN is the 32bit part of gpr N and xN is the whole 64bit.
(width is not part of the mnemonics, but the operands in most cases).

in gcc inline asm %w0 prints wN and %x0 prints xN for gpr operands,
plain %0 is the same as %x0 for gprs but works for other operands too.

aarch64 gcc does not complain if you pass int for a 64bit operand or long
for 32bit one (that's why we had a bug in atomic_arch.h, i forgot the 'w'),
but this sdc input-output operand would work.

> > > #define a_sc_p a_sc_p
> > > -static inline int a_sc_p(volatile long *p, void *v)
> > > +static inline void *a_sc_p(volatile long *p, void *v)
> > > {
> > > -       int r;
> > > +       void *r;
> > >         __asm__ __volatile__ (
> > >                 "scd %0, %1"
> > >                 : "=r"(r), "=m"(*p) : "0"(v) : "memory");
> > 
> > the return type should be int (1 on success, 0 on failure)
> > 
> > e.g. you can use 'long r' instead of 'int r'
> > 
> > if clang still complains then maybe
> > 
> > 	union {void *v; long r;} u = {v};
> > 	__asm__ __volatile__ (
> > 		"scd %0, %1" : "+r"(u.v), "=m"(*p) :: "memory");
> > 	return u.r;
> 
> If simply using 'long' works, I think that's the cleanest/simplest
> solution. The union is uglier and less obvious what it's doing so I'd
> rather avoid it if we can.
> 
> Rich