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

[Rich Felker <dalias@libc.org>]

On Sun, May 17, 2015 at 10:34:02PM -0400, Rich Felker wrote:
> Another idea is letting the compiler simplify, with something like the
> following, which could actually be used cross-platform for all
> llsc-type archs:
> 
> static inline int __sh_cas_llsc(volatile int *p, int t, int s)
> {
> 	do old = llsc_start(p);
> 	while (*p == t && !llsc_end(p, s));
> 	return old;
> }
> 
> Here llsc_start and llsc_end would be inline functions using asm with
> appropriate constraints. Unfortunately I don't see a way to model
> using the value of the truth flag "t" as the output of the asm for
> llsc_end, though. I suspect this would be a problem on a number of
> other archs too; the asm would have to waste an instruction (or
> several) converting the flag to an integer. Unless there's a solution
> to that problem, it makes an approach like this less appealing.

Indeed, I can't find a way to make it work without two useless
instructions. This is what I get (for an extern function a_cas based
on the above approach):

a_cas:
        .align 2
.L4:
#APP
        synco ; mov.li @r4, r0
#NO_APP
        cmp/eq  r0,r5
        bf/s    .L7
        mov     r0,r1
        mov     r6,r0
#APP
        mov.co r0, @r4 ; movt r2
#NO_APP
        tst     r2,r2
        bt      .L4
.L7:
        rts
        mov     r1,r0

Both the movt (which I had to write) and the tst (which the compiler
generated) are useless. It's a shame, because otherwise this approach
is really clean and elegant.

Rich