[uclibc-ng-devel] [PATCH] arm: simplify handling of Thumb related options

[Rich Felker]

On Sun, Mar 20, 2016 at 06:47:04PM +0100, Thomas Petazzoni wrote:
> Hello,
> 
> Thanks for your feedback!
> 
> On Sun, 20 Mar 2016 13:40:40 -0400, Rich Felker wrote:
> > On Sun, Mar 20, 2016 at 05:58:35PM +0100, Thomas Petazzoni wrote:
> > > Currently, the Thumb support on ARM has three related Config.in
> > > options, which are not trivial for users to understand, and are in
> > > fact not needed:
> > > 
> > >  - The USE_BX option is not needed: knowing whether BX is available or
> > >    not is easy. If you have an ARM > v4 or ARMv4T, then BX is
> > >    available, otherwise it's not. This is the logic used in glibc.
> > 
> > This sounds correct.
> > 
> > >  - The USE_LDREXSTREX option is not needed: whenever Thumb2 is
> > >    available, ldrex/strex are available, so we can simply rely on
> > >    __thumb2__ to determine whether ldrex/strex should be used, without
> > >    requiring a Config.in option.
> > 
> > __thumb2__ is sufficient for ldrex/strex but not sufficient; v6 has
> > them in the arm isa.
> 
> I'm not sure what you mean between:
> 
>  (1) ldrex/strex are also available on plain ARMv6 (i.e not Thumb2).
> 
>  (2) ldrex/strex are not available on ARMv6, while some ARMv6 support
>      Thumb2.

(1) is what I mean.

> If you mean (1), then before my patch, the ldrex/strex code was only
> enabled when USE_LDREXSTREX was enabled, and this option had a "depends
> on COMPILE_IN_THUMB". Hence my patch doesn't change the fact that
> the ldrex/strex logic is only used on Thumb. Whether it can/should be
> enabled in non-Thumb configurations is a separate matter from what my
> patch is doing.

OK, then the logic was wrong before, too. :-)

> If you mean (2), then there is indeed a problem. Though, according to
> http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.faqs/14979.html,
> "On a v6 or later core (ARM11 onwards) you should use the LDREX/STREX
> instructions instead", which would suggest that such instructions are
> available.

Indeed, and it's further complicated by the fact that none of these
instructions are proper memory barriers so they're basically useless
without a separate barrier instruction, which does not exist on
pre-v6, and on v6, only as a coprocessor operation that's not
available in thumb1 or on uc models. Not until v7 did arm have proper
working atomics. :(

Rich