Usage: 0) If possible, do this on a multiprocessor, especially if you are planning on modifying or enhancing the package. It will work on a uniprocessor, but the tests are much more likely to pass in the presence of serious problems. 1) Type ./configure --prefix=; make; make check in the directory containing unpacked source. The usual GNU build machinery is used, except that only static, but position-independent, libraries are normally built. On Windows, read README_win32.txt instead. 2) Applications should include atomic_ops.h. Nearly all operations are implemented by header files included from it. It is sometimes necessary, and always recommended to also link against libatomic_ops.a. To use the almost non-blocking stack or malloc implementations, see the corresponding README files, and also link against libatomic_gpl.a before linking against libatomic_ops.a. OVERVIEW: Atomic_ops.h defines a large collection of operations, each one of which is a combination of an (optional) atomic memory operation, and a memory barrier. Also defines associated feature-test macros to determine whether a particular operation is available on the current target hardware (either directly or by synthesis). This is an attempt to replace various existing files with similar goals, since they usually do not handle differences in memory barrier styles with sufficient generality. If this is included after defining AO_REQUIRE_CAS, then the package will make an attempt to emulate compare-and-swap in a way that (at least on Linux) should still be async-signal-safe. As a result, most other atomic operations will then be defined using the compare-and-swap emulation. This emulation is slow, since it needs to disable signals. And it needs to block in case of contention. If you care about performance on a platform that can't directly provide compare-and-swap, there are probably better alternatives. But this allows easy ports to some such platforms (e.g. PA_RISC). The option is ignored if compare-and-swap can be implemented directly. If atomic_ops.h is included after defining AO_USE_PTHREAD_DEFS, then all atomic operations will be emulated with pthread locking. This is NOT async-signal-safe. And it is slow. It is intended primarily for debugging of the atomic_ops package itself. Note that the implementation reflects our understanding of real processor behavior. This occasionally diverges from the documented behavior. (E.g. the documented X86 behavior seems to be weak enough that it is impractical to use. Current real implementations appear to be much better behaved.) We of course are in no position to guarantee that future processors (even HPs) will continue to behave this way, though we hope they will. This is a work in progress. Corrections/additions for other platforms are greatly appreciated. It passes rudimentary tests on X86, Itanium, and Alpha. OPERATIONS: Most operations operate on values of type AO_t, which are unsigned integers whose size matches that of pointers on the given architecture. Exceptions are: - AO_test_and_set operates on AO_TS_t, which is whatever size the hardware supports with good performance. In some cases this is the length of a cache line. In some cases it is a byte. In many cases it is equivalent to AO_t. - A few operations are implemented on smaller or larger size integers. Such operations are indicated by the appropriate prefix: AO_char_... Operates on unsigned char values. AO_short_... Operates on unsigned short values. AO_int_... Operates on unsigned int values. (Currently a very limited selection of these is implemented. We're working on it.) The defined operations are all of the form AO_[_](). The component specifies an atomic memory operation. It may be one of the following, where the corresponding argument and result types are also specified: void nop() No atomic operation. The barrier may still be useful. AO_t load(const volatile AO_t * addr) Atomic load of *addr. void store(volatile AO_t * addr, AO_t new_val) Atomically store new_val to *addr. AO_t fetch_and_add(volatile AO_t *addr, AO_t incr) Atomically add incr to *addr, and return the original value of *addr. AO_t fetch_and_add1(volatile AO_t *addr) Equivalent to AO_fetch_and_add(addr, 1). AO_t fetch_and_sub1(volatile AO_t *addr) Equivalent to AO_fetch_and_add(addr, (AO_t)(-1)). void or(volatile AO_t *addr, AO_t incr) Atomically or incr into *addr. int compare_and_swap(volatile AO_t * addr, AO_t old_val, AO_t new_val) Atomically compare *addr to old_val, and replace *addr by new_val if the first comparison succeeds. Returns nonzero if the comparison succeeded and *addr was updated. AO_TS_VAL_t test_and_set(volatile AO_TS_t * addr) Atomically read the binary value at *addr, and set it. AO_TS_VAL_t is an enumeration type which includes two values AO_TS_SET and AO_TS_CLEAR. An AO_TS_t location is capable of holding an AO_TS_VAL_t, but may be much larger, as dictated by hardware constraints. Test_and_set logically sets the value to AO_TS_SET. It may be reset to AO_TS_CLEAR with the AO_CLEAR(AO_TS_t *) macro. AO_TS_t locations should be initialized to AO_TS_INITIALIZER. The values of AO_TS_SET and AO_TS_CLEAR are hardware dependent. (On PA-RISC, AO_TS_SET is zero!) Test_and_set is a more limited version of compare_and_swap. Its only advantage is that it is more easily implementable on some hardware. It should thus be used if only binary test-and-set functionality is needed. If available, we also provide compare_and_swap operations that operate on wider values. Since standard data types for double width values may not be available, these explicitly take pairs of arguments for the new and/or old value. Unfortunately, there are two common variants, neither of which can easily and efficiently emulate the other. The first performs a comparison against the entire value being replaced, where the second replaces a double-width replacement, but performs a single-width comparison: int compare_double_and_swap_double(volatile AO_double_t * addr, AO_t old_val1, AO_t old_val2, AO_t new_val1, AO_t new_val2); int compare_and_swap_double(volatile AO_double_t * addr, AO_t old_val1, AO_t new_val1, AO_t new_val2); where AO_double_t is a structure containing AO_val1 and AO_val2 fields, both of type AO_t. For compare_and_swap_double, we compare against the val1 field. AO_double_t exists only if AO_HAVE_double_t is defined. ORDERING CONSTRAINTS: Each operation name also includes a suffix that specifies the associated ordering semantics. The ordering constraint limits reordering of this operation with respect to other atomic operations and ordinary memory references. The current implementation assumes that all memory references are to ordinary cacheable memory; the ordering guarantee is with respect to other threads or processes, not I/O devices. (Whether or not this distinction is important is platform-dependent.) Ordering suffixes are one of the following: : No memory barrier. A plain AO_nop() really does nothing. _release: Earlier operations must become visible to other threads before the atomic operation. _acquire: Later operations must become visible after this operation. _read: Subsequent reads must become visible after reads included in the atomic operation or preceding it. Rarely useful for clients? _write: Earlier writes become visible before writes during or after the atomic operation. Rarely useful for clients? _full: Ordered with respect to both earlier and later memory ops. AO_store_full or AO_nop_full are the normal ways to force a store to be ordered with respect to a later load. _release_write: Ordered with respect to earlier writes. This is normally implemented as either a _write or _release barrier. _dd_acquire_read: Ordered with respect to later reads that are data dependent on this one. This is needed on a pointer read, which is later dereferenced to read a second value, with the expectation that the second read is ordered after the first one. On most architectures, this is equivalent to no barrier. (This is very hard to define precisely. It should probably be avoided. A major problem is that optimizers tend to try to eliminate dependencies from the generated code, since dependencies force the hardware to execute the code serially.) _release_read: Ordered with respect to earlier reads. Useful for implementing read locks. Can be implemented as _release, but not as _read, since _read groups the current operation with the earlier ones. We assume that if a store is data-dependent on an a previous load, then the two are always implicitly ordered. It is possible to test whether AO_ is available on the current platform by checking whether AO_HAVE__ is defined as a macro. Note that we generally don't implement operations that are either meaningless (e.g. AO_nop_acquire, AO_nop_release) or which appear to have no clear use (e.g. AO_load_release, AO_store_acquire, AO_load_write, AO_store_read). On some platforms (e.g. PA-RISC) many operations will remain undefined unless AO_REQUIRE_CAS is defined before including the package. When typed in the package build directory, the following command will print operations that are unimplemented on the platform: make test_atomic; ./test_atomic The following command generates a file "list_atomic.i" containing the macro expansions of all implemented operations on the platform: make list_atomic.i Future directions: It currently appears that something roughly analogous to this is very likely to become part of the C++0x standard. That effort has pointed out a number of issues that we expect to address there. Since some of the solutions really require compiler support, they may not be completely addressed here. Known issues include: We should be more precise in defining the semantics of the ordering constraints, and if and how we can guarantee sequential consistency. Dd_acquire_read is very hard or impossible to define in a way that cannot be invalidated by reasonably standard compiler transformations. There is probably no good reason to provide operations on standard integer types, since those may have the wrong alignment constraints. Example: If you want to initialize an object, and then "publish" a pointer to it in a global location p, such that other threads reading the new value of p are guaranteed to see an initialized object, it suffices to use AO_release_write(p, ...) to write the pointer to the object, and to retrieve it in other threads with AO_acquire_read(p). Platform notes: All X86: We quietly assume 486 or better. Microsoft compilers: Define AO_ASSUME_WINDOWS98 to get access to hardware compare-and-swap functionality. This relies on the InterlockedCompareExchange() function which was apparently not supported in Windows95. (There may be a better way to get access to this.) Gcc on x86: Define AO_USE_PENTIUM4_INSTRS to use the Pentium 4 mfence instruction. Currently this is appears to be of marginal benefit.