svn commit: r353296 - head/sys/powerpc/include

[Konstantin Belousov]

On Tue, Oct 08, 2019 at 09:14:14AM -0500, Justin Hibbits wrote:
> On Tue, 8 Oct 2019 11:16:04 +0300
> Konstantin Belousov <kostikbel at gmail.com> wrote:
> 
> > On Tue, Oct 08, 2019 at 01:36:34AM +0000, Justin Hibbits wrote:
> > > Author: jhibbits
> > > Date: Tue Oct  8 01:36:34 2019
> > > New Revision: 353296
> > > URL: https://svnweb.freebsd.org/changeset/base/353296
> > > 
> > > Log:
> > >   powerpc: Implement atomic_(f)cmpset_ for short and char
> > >   |
> > >   This adds two implementations for each atomic_fcmpset_ and
> > > atomic_cmpset_ short and char functions, selectable at compile time
> > > for the target architecture.  By default, it uses a generic
> > > shift-and-mask to perform atomic updates to sub-components of
> > > 32-bit words from <sys/_atomic_subword.h>. However, if
> > > ISA_206_ATOMICS is defined it uses the ll/sc instructions for
> > > halfword and bytes, introduced in PowerISA 2.06.  These
> > > instructions are supported by all IBM processors from POWER7 on, as
> > > well as the Freescale/NXP e6500 core.  Although the e5500 and
> > > e500mc both implement PowerISA 2.06 they do not implement these
> > > instructions. As part of this, clean up the atomic_(f)cmpset_acq
> > > and _rel wrappers, by using macros to reduce code duplication.
> > >   
> > >   ISA_206_ATOMICS requires clang or newer binutils (2.20 or later).
> > >  
> > Why don't you use normal word-sized ll/sc tlwarx/stwcx, and only
> > modifying the part of the register as needed ?  This would work on
> > all supported CPUs, right ?
> > 
> > When kevans did the _atomic_subword.h, one of the arches involved was
> > sparc64, which does not have ll/sc.  Also for MIPS there are some fine
> > details which might mean that C implementation is less work than using
> > word-sized ll/sc.  But why for power ?
> 
> No real significant reason.  In fact, the review's diff has exactly
> what you're asking for.  The only reason I modified it for commit with
> Kyle's work was purely readability, I thought using the C wrapper with
> atomic_fcmpset_() was just marginally cleaner. I haven't checked, but I
> don't think the inline code difference is too great, but I'll have to do
> another review of it to be sure.  It's easy enough to commit the
> original diff over top instead, if that's the better way to go.

If the generated code difference is not significant, it is a strong
argument to keep the committed version.  But I find it quite surprising.