ObjFW  History of src/OFAtomic.h of 39639cd987e8dc1a7d086ff9a2eaed68a9282c3fcd0daa5d5edb61b27f8d4d9f

It's just a less awkward name and also what the BSDs use. file: [7b3ed1a98e] check-in: [b7a4e61b88] user: js, branch: trunk, size: 3646 [annotate] [blame] [check-ins using] [diff]

Duplicating defines for every CPU architecture is not very useful. file: [6778d083ff] check-in: [c859e2230c] user: js, branch: trunk, size: 1261 [annotate] [blame] [check-ins using] [diff]

The assembler on AIX is too limited. file: [e53591a99c] check-in: [4a6d9cd210] user: js, branch: trunk, size: 1220 [annotate] [blame] [check-ins using] [diff]

Apple GCC uses register 0 for "r" and there seems to be no way to avoid
this, while modern GCC does not seem to consider register 0 to be a GPR. file: [72942f053d] check-in: [636533772e] user: js, branch: trunk, size: 1171 [annotate] [blame] [check-ins using] [diff]

Forgot to add 2017, even though I already did quite some changes in
2017. file: [c91cb23822] check-in: [44f45c2e35] user: js, branch: trunk, size: 1145 [annotate] [blame] [check-ins using] [diff]

Having one file per implementation reduces the #ifdef hell and makes it
much more readable. file: [f5eb774718] check-in: [7115c55ef0] user: js, branch: trunk, size: 1071 [annotate] [blame] [check-ins using] [diff]

Also changes the naming slightly (leave -> exit). This is more similar
to the membar_*() API NetBSD and OpenBSD provides. file: [0145f59683] check-in: [24796a2dc1] user: js, branch: trunk, size: 20414 [annotate] [blame] [check-ins using] [diff]

This matches what the assembly versions does. However, the __sync_*
versions still use the barrier, but unfortunately, no version without it
is provided. The only way around this would be to use the new __atomic_*
that has been added in GCC 4.7. file: [50acc02363] check-in: [0c3e3da576] user: js, branch: trunk, size: 20052 [annotate] [blame] [check-ins using] [diff]

This is only available on x86 with SSE2, while it's always available on
x86_64. However, checking if SSE2 is available here would be too slow,
therefore let the compiler decide what do do instead (which will depend
on the selected target CPU). file: [23c4a049f4] check-in: [9feaa90358] user: js, branch: trunk, size: 20199 [annotate] [blame] [check-ins using] [diff]

Replace of_memory_read_barrier() and of_memory_write_barrier() - which
are quite unspecific - with of_memory_enter_barrier() and
of_memory_leave_barrier().

Also add an assembly implementation for ARM and ARM64. file: [c6a5af4a20] check-in: [96a128f954] user: js, branch: trunk, size: 20222 [annotate] [blame] [check-ins using] [diff]

While at it, also update the mail address. file: [4b9a633048] check-in: [cec0f072f8] user: js, branch: 0.8, size: 20345 [annotate] [blame] [check-ins using] [diff]

While at it, also update the mail address. file: [17afd94567] check-in: [2a27cf3000] user: js, branch: trunk, size: 20353 [annotate] [blame] [check-ins using] [diff]

Now that Clang switched to use _Nullable and _Nonnull instead of
__nullable / __nonnull, there is no longer a conflict with glibc, which
means we can just define _Nullable / _Nonnull to nothing if they are not
understood by the compiler (which did not work with __nullable /
__nonnull due to this conflict).

This also defines _Null_unspecified to nothing if unavailable. file: [5acf14775c] check-in: [d526d938d7] user: js, branch: trunk, size: 20349 [annotate] [blame] [check-ins using] [diff]

Using __nonnull directly doesn't work on systems using glibc, as glibc
defines __nonnull as a parameterized define. While this does not fix the
problem of Clang introducing __nonnull even though it conflicts with
glibc, this at least means it's possible again to compile things with
versions of Clang that don't support __nonnull on systems with glibc. file: [cc8d24b757] check-in: [b7097a67b6] user: js, branch: trunk, size: 20341 [annotate] [blame] [check-ins using] [diff]

The compiler can optimize the builtins better. For example, if the
result is not used, the compiler can use an instruction that does not
fetch the result and thus might generate more efficient code. file: [53694a578a] check-in: [a199313c53] user: js, branch: trunk, size: 15396 [annotate] [blame] [check-ins using] [diff]

* of_atomic_cmpswap_*: Clobber eax by adding it as an (unused) output,
so the compiler resets eax (e.g. when used in a loop). (Adding it
to the input and clobber list is lot legal!)
* Don't dereference the pointer passed using "m"(*p) twice.
* Fix some code for ILP64 systems (it seems there is no architecture
using it, though, as it even contained syntax errors).
* Use __asm__ __volatile__ just to be sure.

Thanks to Florian Zeitz for pointing out most of these issues! file: [309e66c1ae] check-in: [6b15bb229e] user: js, branch: trunk, size: 16407 [annotate] [blame] [check-ins using] [diff]

The only places where BOOL is left are those where they are required by
the ABI. file: [a4f95ba0da] check-in: [c5ef582958] user: js, branch: trunk, size: 15322 [annotate] [blame] [check-ins using] [diff]

Changesets:
* b95fcaa6d694
* 804c68d222b4
* da8cd738da3d
* 1109d5ce3419
* d03f5c1ca95d
* 2a7017722165
* 4fccdc79eeb7
* d228149fbc04
* 8782d412a4a6 file: [3969906906] check-in: [b91a8283fe] user: js, branch: 0.2, size: 2842 [annotate] [blame] [check-ins using] [diff]