Re: [PATCH 13/15] x86: add cmpxchg_flag() variant

[Jeremy Fitzhardinge <jeremy@goop.org>]

On 08/23/2011 01:45 PM, H. Peter Anvin wrote:
> On 08/23/2011 12:53 PM, Jeremy Fitzhardinge wrote:
>> Yes, that would ideal.  The closest you can get is asm goto(), but the
>> syntax for that would be awful; something like:
>>
>> #define cmpxchg_jump(ptr, old, new, fail)\
>> 	asm goto (...)
>>
>>
>> :
>> :
>>
>> again:
>> 	old = *thingp;
>> 	new = frobulate(old);
>> 	cmpxchg_jump(thingp, old, new, again);
>> 	/* worked */
>>
>> Would this be useful enough?
>>
> Actually there is a trick:
>
> static inline bool ....
> {
> 	asm goto(... yes);
> no:
> 	return false;
> yes:
> 	return true;
> }
>
> ... which makes syntax a heckuva lot less awkward.

Yeah, but you'd need to define an inline for each type, since the
function isn't polymorphic.  But it can be done with a macro.

However, having prototyped it, I dunno, it doesn't really seem like much
of a win for all the extra code it adds.  I just can't get too excited
about an extra test instruction adjacent to a monster like a locked
cmpxchg.  The jump variant avoids the test, but gcc still generates some
pretty bogus stuff:

        lock; cmpxchgq %rbx,(%rcx); jne .L88    # D.24853, MEM[(volatile u64 *)top_p_26],
# 0 "" 2
#NO_APP
        jmp     .L87    #
.L88:
        xorl    %esi, %esi      #
        movq    %rbx, %rdi      # D.24853,
        call    free_pages      #
.L87:
        addq    p2m_top_mfn(%rip), %r13 # p2m_top_mfn, D.24896
        movq    p2m_top_mfn_p(%rip), %rax       # p2m_top_mfn_p, p2m_top_mfn_p


and adding unlikely()s doesn't help at all.

    J