[Gc] Re: [libatomic_ops] On Linux,
ARMv5tel does not have all functions available (#3)
ivmai at mail.ru
Sat Mar 3 08:39:39 PST 2012
Thank you for pointing us to this patch but IMHO this is a logically incorrect patch.
By libatomic_ops contract:
1. if some operation is not available then AO_HAVE_x just not defined;
2. for CAS ops, if we strictly need it then we should define AO_REQUIRE_CAS causing automatic fall-back on demand.
3. fetch_and_add is derived from CAS in the worth case, so you should define AO_REQUIRE_CAS as well if you need atomic inc (this seems to be not explicitly specified in the docs but should be).
Looking into arm.h (of master branch), the deference between, roughly speaking, ARMv5 and ARMv6 is in presence of CAS.
Please point me at if I misunderstand something of your post.
03 03 2012, 19:02 Niels de Vos <reply+i-3489920-a6264862991435ec2c55a219c5150534b26c055f-460469 at reply.github.com>:
> It seems that libatomic_ops provides an optimized implementation for ARMv6 and newer architectures. Unfortunately there are no asm functions for ARMv5, which should fallback automatically to the pthread implementation.
> This [proposed patch](http://pkgs.fedoraproject.org/gitweb/?p=gc.git;a=blob;f=gc-7.2alpha6-libatomic_ops_arm5.patch;h=84e9bba6287093d106ecfbbdc083e0e320af5be1;hb=73b5e99c72ab0dd1248cc8326139893cfd20408d) is already available in the Fedora package. This change should provide the pthread implementation when atomic_ops.h is included on ARMv5 and earlier architectures.
> Please review and include this in the next release if it make sense. Thanks!
> Reply to this email directly or view it on GitHub:
More information about the Gc