|
* Use of __atomic builtins when available.
* Improved configure test that checks for missing memory
barrier in __sync_synchronize(). The old approach was to
verify known working gcc versions and check gcc version at
compile time. Besides not being very safe, the old approach
often unnecessarily caused usage of the very expensive
workaround.
* Introduced (no overhead) workaround for missing clobber in
__sync_synchronize() when using buggy LLVM implementation of
__sync_synchronize().
* Implement native memory barriers for ARM processors supporting
the DMB instruction.
* Use of volatile store on Alpha as atomic set operation if no
__atomic_store_n() is available (already used on x86/x86_64
Sparc V9, PowerPC, and MIPS). Fallback used when not using
volatile store is typically very expensive.
* Use volatile load on Alpha and ARM as atomic read operation
if no __atomic_load_n() is available (already used on
x86/x86_64 Sparc V9, PowerPC, and MIPS). Fallback when not
using volatile load is typically very expensive.
|
|
The ethread atomics API now also provide double word size atomics.
Double word size atomics are implemented using native atomic
instructions on x86 (when the cmpxchg8b instruction is available)
and on x86_64 (when the cmpxchg16b instruction is available). On
other hardware where 32-bit atomics or word size atomics are
available, an optimized fallback is used; otherwise, a spinlock,
or a mutex based fallback is used.
The ethread library now performs runtime tests for presence of
hardware features, such as for example SSE2 instructions, instead
of requiring this to be determined at compile time.
There are now functions implementing each atomic operation with the
following implied memory barrier semantics: none, read, write,
acquire, release, and full. Some of the operation-barrier
combinations aren't especially useful. But instead of filtering
useful ones out, and potentially miss a useful one, we implement
them all.
A much smaller set of functionality for native atomics are required
to be implemented than before. More or less only cmpxchg and a
membar macro are required to be implemented for each atomic size.
Other functions will automatically be constructed from these. It is,
of course, often wise to implement more that this if possible from a
performance perspective.
|
