Re: [PATCH] inline llsc atomics when compiling for sh4a

[Rich Felker <dalias@libc.org>]

On Mon, May 18, 2015 at 08:30:45PM -0400, Rich Felker wrote:
> [...]
> static inline int __sh_cas_llsc(volatile int *p, int t, int s)
> {
> 	do old = llsc_start(p);
> 	while (!llsc_end(p, old==t ? s : old));
> 	return old;
> }
> 
> This version is structurally analogous to the non-CAS atomics, but
> perhaps more costly in the old!=t case.
> 
> Anyway at this point I don't see an efficient way do to the
> conditional, so it's mostly a theoretical topic at this point.

I have a partial solution but it's utterly hideous:

static inline int llsc_end(volatile int *p, int v)
{
	__asm__ __volatile__ goto ( "mov.co %1, @%0 ; bf %l[fail]" : :
	"r"(p), "z"(v) : "memory" : fail );
	return 1;
fail:
	return 0;
}

This doesn't eliminate the branch in the asm, but it does allow gcc to
eliminate the branch outside the asm. The resulting code is:

a_cas:
        .align 2
.L2:
#APP
        synco ; mov.li @r4, r0
#NO_APP
        cmp/eq  r0,r5
        bf      .L4
        mov     r6,r0
#APP
        mov.co r0, @r4 ; bf .L2
#NO_APP
.L6:
        rts
        mov     r5,r0
        .align 1
.L4:
        bra     .L6
        mov     r0,r5

Of course asm goto is utterly hideous, so I don't propose actually
using this. I doubt any of the non-GCC compilers we support would
accept it.

Rich