Re: First cut of PowerPC support in NPTL

[Paul Mackerras <paulus samba org>]

Jakub Jelinek writes:

> Is there any gain on PPC (or s390) to have separate lll_mutex_* and lll_*
> implementations instead of using just one (lll_mutex_*) for both?
> 
> On IA-32 lll_lock is one insn faster than lll_mutex_lock, which is why I
> believe Ulrich did this, but on s390 with compare_and_swap it
> really doesn't matter if you increment or decrement and what the
> unlocked value is.

My implementation on PPC has the same number of instructions for
lll_mutex_lock and lll_lock.  The underlying __lll_mutex_add uses one
more register than __lll_add, but we have plenty of registers on PPC
so that shouldn't matter.  I'll change the lll_lock/unlock
implementation to be the same as lll_mutex_lock/unlock.

Paul.