Re: [Patch][aarch64] Use IFUNCs to enable LSE instructions in libatomic on aarch64

[Szabolcs Nagy <szabolcs dot nagy at arm dot com>]

On 07/08/17 21:44, Steve Ellcey wrote:
> This patch uses the libatomic IFUNC infrastructure so that aarch64
> machines that support the LSE instructions can use them.  Note that
> aarch64 still isn't enabling IFUNC support by default though I have
> submitted a patch to do that.  You can enable IFUNC support by
> configuring with --enable-gnu-indirect-function.
> 
> Glibc has an environment variable, LD_HWCAP_MASK, that can be used to
> mask out some or all of the bits returned by getauxval(AT_HWCAP) and
> ignore certain hardware capabilities.  I enabled this functionality
> for libatomic by looking at the LD_HWCAP_MASK variable in the IFUNC
> resolver function.  That way, if I had a system that supported LSE but
> did not want to use it for some reason, I could set LD_HWCAP_MASK to
> zero and then the IFUNC selector function would not enable the LSE
> routines.  I could remove this functionality if we thought it was not
> appropriate but I think it is useful, both for testing and for end
> users.
> 

the use of ifunc in gcc target libraries was a mistake
in my opinion, there are several known bugs in the ifunc
design and uclibc/musl/bionic don't plan to support it.
but at this point i dont have a better proposal for doing
runtime selection of optimal atomics code.

however in this patch i don't see why would the ctor run
before ifunc resolvers. how does this work on x86_64?
(there the different 16byte atomics are not even compatible,
so if ifunc resolvers in different dsos return different
result because one ran before the ctor, another after then
the runtime behaviour is broken. this can happen when one
dso is bindnow so ifunc relocation is processed before
ctors and another runs resolvers lazily or dlopened later..
but i didnt test it just looks broken)

note that aarch64 ifunc resolvers get hwcap as an argument
so all this brokenness can be avoided (and if we want to
disable hwcap flags then probably glibc should take care
of that before passing hwcap to the ifunc resolver).

> Tested on aarch64, OK for checkin?
> 
> Steve Ellcey
> sellcey@cavium.com
> 
> 
> 
> 
> 2017-08-07  Steve Ellcey  <sellcey@cavium.com>
> 
> 	* Makefile.am (ARCH_AARCH64_LINUX_LSE): Add IFUNC_OPTIONS and
> 	libatomic_la_LIBADD.
> 	* config/linux/aarch64/host-config.h: New file.
> 	* config/linux/aarch64/init.c: New file.
> 	* configure.ac (AC_CHECK_HEADERS): Check for sys/auxv.h.
> 	(AC_CHECK_FUNCS): Check for getauxval.
> 	(ARCH_AARCH64_LINUX_LSE): New conditional for IFUNC builds.
> 	* configure.tgt (aarch64): Set AARCH and try_ifunc.
> 	(aarch64*-*-linux*) Update config_path.
> 	* Makefile.in: Regenerate.
> 	* auto-config.h.in: Regenerate.
> 	* configure: Regenerate.
>