X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;ds=sidebyside;f=src%2Futil%2Fu_atomic.h;h=2a5bbae295b727de41f65cb44933cb27f1b8db19;hb=75968a668e44b3fd7c9b9277937c005366fca116;hp=f326bd1c2c3c83055ccc490d707fec6ef55ac59f;hpb=972f8458f1ebdf805059d9252ae0ebb727934613;p=mesa.git diff --git a/src/util/u_atomic.h b/src/util/u_atomic.h index f326bd1c2c3..2a5bbae295b 100644 --- a/src/util/u_atomic.h +++ b/src/util/u_atomic.h @@ -6,9 +6,13 @@ * */ +#include "no_extern_c.h" + #ifndef U_ATOMIC_H #define U_ATOMIC_H +#include + /* Favor OS-provided implementations. * * Where no OS-provided implementation is available, fall back to @@ -19,188 +23,51 @@ #define PIPE_ATOMIC_OS_SOLARIS #elif defined(_MSC_VER) #define PIPE_ATOMIC_MSVC_INTRINSIC -#elif defined(__GNUC__) && ((__GNUC__ * 100 + __GNUC_MINOR__) >= 401) +#elif defined(__GNUC__) #define PIPE_ATOMIC_GCC_INTRINSIC -#elif (defined(__GNUC__) && defined(__i386__)) -#define PIPE_ATOMIC_ASM_GCC_X86 -#elif (defined(__GNUC__) && defined(__x86_64__)) -#define PIPE_ATOMIC_ASM_GCC_X86_64 #else #error "Unsupported platform" #endif -#if defined(PIPE_ATOMIC_ASM_GCC_X86_64) -#define PIPE_ATOMIC "GCC x86_64 assembly" - -#ifdef __cplusplus -extern "C" { -#endif - -#define p_atomic_set(_v, _i) (*(_v) = (_i)) -#define p_atomic_read(_v) (*(_v)) - -static inline boolean -p_atomic_dec_zero(int32_t *v) -{ - unsigned char c; - - __asm__ __volatile__("lock; decl %0; sete %1":"+m"(*v), "=qm"(c) - ::"memory"); - - return c != 0; -} - -static inline void -p_atomic_inc(int32_t *v) -{ - __asm__ __volatile__("lock; incl %0":"+m"(*v)); -} - -static inline void -p_atomic_dec(int32_t *v) -{ - __asm__ __volatile__("lock; decl %0":"+m"(*v)); -} - -static inline int32_t -p_atomic_inc_return(int32_t *v) -{ - return __sync_add_and_fetch(v, 1); -} - -static inline int32_t -p_atomic_dec_return(int32_t *v) -{ - return __sync_sub_and_fetch(v, 1); -} - -static inline int32_t -p_atomic_cmpxchg(int32_t *v, int32_t old, int32_t _new) -{ - return __sync_val_compare_and_swap(v, old, _new); -} - -#ifdef __cplusplus -} -#endif - -#endif /* PIPE_ATOMIC_ASM_GCC_X86_64 */ - - -#if defined(PIPE_ATOMIC_ASM_GCC_X86) - -#define PIPE_ATOMIC "GCC x86 assembly" - -#ifdef __cplusplus -extern "C" { -#endif - -#define p_atomic_set(_v, _i) (*(_v) = (_i)) -#define p_atomic_read(_v) (*(_v)) - -static inline boolean -p_atomic_dec_zero(int32_t *v) -{ - unsigned char c; - - __asm__ __volatile__("lock; decl %0; sete %1":"+m"(*v), "=qm"(c) - ::"memory"); - - return c != 0; -} - -static inline void -p_atomic_inc(int32_t *v) -{ - __asm__ __volatile__("lock; incl %0":"+m"(*v)); -} - -static inline void -p_atomic_dec(int32_t *v) -{ - __asm__ __volatile__("lock; decl %0":"+m"(*v)); -} - -static inline int32_t -p_atomic_inc_return(int32_t *v) -{ - return __sync_add_and_fetch(v, 1); -} - -static inline int32_t -p_atomic_dec_return(int32_t *v) -{ - return __sync_sub_and_fetch(v, 1); -} - -static inline int32_t -p_atomic_cmpxchg(int32_t *v, int32_t old, int32_t _new) -{ - return __sync_val_compare_and_swap(v, old, _new); -} - -#ifdef __cplusplus -} -#endif - -#endif - - - /* Implementation using GCC-provided synchronization intrinsics */ #if defined(PIPE_ATOMIC_GCC_INTRINSIC) #define PIPE_ATOMIC "GCC Sync Intrinsics" -#ifdef __cplusplus -extern "C" { -#endif +#if defined(USE_GCC_ATOMIC_BUILTINS) + +/* The builtins with explicit memory model are available since GCC 4.7. */ +#define p_atomic_set(_v, _i) __atomic_store_n((_v), (_i), __ATOMIC_RELEASE) +#define p_atomic_read(_v) __atomic_load_n((_v), __ATOMIC_ACQUIRE) +#define p_atomic_dec_zero(v) (__atomic_sub_fetch((v), 1, __ATOMIC_ACQ_REL) == 0) +#define p_atomic_inc(v) (void) __atomic_add_fetch((v), 1, __ATOMIC_ACQ_REL) +#define p_atomic_dec(v) (void) __atomic_sub_fetch((v), 1, __ATOMIC_ACQ_REL) +#define p_atomic_add(v, i) (void) __atomic_add_fetch((v), (i), __ATOMIC_ACQ_REL) +#define p_atomic_inc_return(v) __atomic_add_fetch((v), 1, __ATOMIC_ACQ_REL) +#define p_atomic_dec_return(v) __atomic_sub_fetch((v), 1, __ATOMIC_ACQ_REL) + +#else #define p_atomic_set(_v, _i) (*(_v) = (_i)) #define p_atomic_read(_v) (*(_v)) +#define p_atomic_dec_zero(v) (__sync_sub_and_fetch((v), 1) == 0) +#define p_atomic_inc(v) (void) __sync_add_and_fetch((v), 1) +#define p_atomic_dec(v) (void) __sync_sub_and_fetch((v), 1) +#define p_atomic_add(v, i) (void) __sync_add_and_fetch((v), (i)) +#define p_atomic_inc_return(v) __sync_add_and_fetch((v), 1) +#define p_atomic_dec_return(v) __sync_sub_and_fetch((v), 1) -static inline boolean -p_atomic_dec_zero(int32_t *v) -{ - return (__sync_sub_and_fetch(v, 1) == 0); -} - -static inline void -p_atomic_inc(int32_t *v) -{ - (void) __sync_add_and_fetch(v, 1); -} - -static inline void -p_atomic_dec(int32_t *v) -{ - (void) __sync_sub_and_fetch(v, 1); -} - -static inline int32_t -p_atomic_inc_return(int32_t *v) -{ - return __sync_add_and_fetch(v, 1); -} - -static inline int32_t -p_atomic_dec_return(int32_t *v) -{ - return __sync_sub_and_fetch(v, 1); -} - -static inline int32_t -p_atomic_cmpxchg(int32_t *v, int32_t old, int32_t _new) -{ - return __sync_val_compare_and_swap(v, old, _new); -} - -#ifdef __cplusplus -} #endif +/* There is no __atomic_* compare and exchange that returns the current value. + * Also, GCC 5.4 seems unable to optimize a compound statement expression that + * uses an additional stack variable with __atomic_compare_exchange[_n]. + */ +#define p_atomic_cmpxchg(v, old, _new) \ + __sync_val_compare_and_swap((v), (old), (_new)) + #endif @@ -214,12 +81,13 @@ p_atomic_cmpxchg(int32_t *v, int32_t old, int32_t _new) #define p_atomic_set(_v, _i) (*(_v) = (_i)) #define p_atomic_read(_v) (*(_v)) -#define p_atomic_dec_zero(_v) ((boolean) --(*(_v))) -#define p_atomic_inc(_v) ((void) (*(_v))++) -#define p_atomic_dec(_v) ((void) (*(_v))--) -#define p_atomic_inc_return(_v) ((*(_v))++) -#define p_atomic_dec_return(_v) ((*(_v))--) -#define p_atomic_cmpxchg(_v, old, _new) (*(_v) == old ? *(_v) = (_new) : *(_v)) +#define p_atomic_dec_zero(_v) (p_atomic_dec_return(_v) == 0) +#define p_atomic_inc(_v) ((void) p_atomic_inc_return(_v)) +#define p_atomic_dec(_v) ((void) p_atomic_dec_return(_v)) +#define p_atomic_add(_v, _i) (*(_v) = *(_v) + (_i)) +#define p_atomic_inc_return(_v) (++(*(_v))) +#define p_atomic_dec_return(_v) (--(*(_v))) +#define p_atomic_cmpxchg(_v, _old, _new) (*(_v) == (_old) ? (*(_v) = (_new), (_old)) : *(_v)) #endif @@ -228,58 +96,62 @@ p_atomic_cmpxchg(int32_t *v, int32_t old, int32_t _new) #define PIPE_ATOMIC "MSVC Intrinsics" +/* We use the Windows header's Interlocked*64 functions instead of the + * _Interlocked*64 intrinsics wherever we can, as support for the latter varies + * with target CPU, whereas Windows headers take care of all portability + * issues: using intrinsics where available, falling back to library + * implementations where not. + */ +#ifndef WIN32_LEAN_AND_MEAN +#define WIN32_LEAN_AND_MEAN 1 +#endif +#include #include +#include -#pragma intrinsic(_InterlockedIncrement) -#pragma intrinsic(_InterlockedDecrement) -#pragma intrinsic(_InterlockedCompareExchange) - -#ifdef __cplusplus -extern "C" { -#endif +/* MSVC supports decltype keyword, but it's only supported on C++ and doesn't + * quite work here; and if a C++-only solution is worthwhile, then it would be + * better to use templates / function overloading, instead of decltype magic. + * Therefore, we rely on implicit casting to LONGLONG for the functions that return + */ #define p_atomic_set(_v, _i) (*(_v) = (_i)) #define p_atomic_read(_v) (*(_v)) -static inline boolean -p_atomic_dec_zero(int32_t *v) -{ - return _InterlockedDecrement((long *)v) == 0; -} - -static inline void -p_atomic_inc(int32_t *v) -{ - _InterlockedIncrement((long *)v); -} - -static inline int32_t -p_atomic_inc_return(int32_t *v) -{ - return _InterlockedIncrement((long *)v); -} - -static inline void -p_atomic_dec(int32_t *v) -{ - _InterlockedDecrement((long *)v); -} - -static inline int32_t -p_atomic_dec_return(int32_t *v) -{ - return _InterlockedDecrement((long *)v); -} - -static inline int32_t -p_atomic_cmpxchg(int32_t *v, int32_t old, int32_t _new) -{ - return _InterlockedCompareExchange((long *)v, _new, old); -} - -#ifdef __cplusplus -} -#endif +#define p_atomic_dec_zero(_v) \ + (p_atomic_dec_return(_v) == 0) + +#define p_atomic_inc(_v) \ + ((void) p_atomic_inc_return(_v)) + +#define p_atomic_inc_return(_v) (\ + sizeof *(_v) == sizeof(short) ? _InterlockedIncrement16((short *) (_v)) : \ + sizeof *(_v) == sizeof(long) ? _InterlockedIncrement ((long *) (_v)) : \ + sizeof *(_v) == sizeof(__int64) ? InterlockedIncrement64 ((__int64 *)(_v)) : \ + (assert(!"should not get here"), 0)) + +#define p_atomic_dec(_v) \ + ((void) p_atomic_dec_return(_v)) + +#define p_atomic_dec_return(_v) (\ + sizeof *(_v) == sizeof(short) ? _InterlockedDecrement16((short *) (_v)) : \ + sizeof *(_v) == sizeof(long) ? _InterlockedDecrement ((long *) (_v)) : \ + sizeof *(_v) == sizeof(__int64) ? InterlockedDecrement64 ((__int64 *)(_v)) : \ + (assert(!"should not get here"), 0)) + +#define p_atomic_add(_v, _i) (\ + sizeof *(_v) == sizeof(char) ? _InterlockedExchangeAdd8 ((char *) (_v), (_i)) : \ + sizeof *(_v) == sizeof(short) ? _InterlockedExchangeAdd16((short *) (_v), (_i)) : \ + sizeof *(_v) == sizeof(long) ? _InterlockedExchangeAdd ((long *) (_v), (_i)) : \ + sizeof *(_v) == sizeof(__int64) ? InterlockedExchangeAdd64((__int64 *)(_v), (_i)) : \ + (assert(!"should not get here"), 0)) + +#define p_atomic_cmpxchg(_v, _old, _new) (\ + sizeof *(_v) == sizeof(char) ? _InterlockedCompareExchange8 ((char *) (_v), (char) (_new), (char) (_old)) : \ + sizeof *(_v) == sizeof(short) ? _InterlockedCompareExchange16((short *) (_v), (short) (_new), (short) (_old)) : \ + sizeof *(_v) == sizeof(long) ? _InterlockedCompareExchange ((long *) (_v), (long) (_new), (long) (_old)) : \ + sizeof *(_v) == sizeof(__int64) ? InterlockedCompareExchange64 ((__int64 *)(_v), (__int64)(_new), (__int64)(_old)) : \ + (assert(!"should not get here"), 0)) #endif @@ -288,37 +160,62 @@ p_atomic_cmpxchg(int32_t *v, int32_t old, int32_t _new) #define PIPE_ATOMIC "Solaris OS atomic functions" #include - -#ifdef __cplusplus -extern "C" { -#endif +#include #define p_atomic_set(_v, _i) (*(_v) = (_i)) #define p_atomic_read(_v) (*(_v)) -static inline boolean -p_atomic_dec_zero(int32_t *v) -{ - uint32_t n = atomic_dec_32_nv((uint32_t *) v); +#define p_atomic_dec_zero(v) (\ + sizeof(*v) == sizeof(uint8_t) ? atomic_dec_8_nv ((uint8_t *)(v)) == 0 : \ + sizeof(*v) == sizeof(uint16_t) ? atomic_dec_16_nv((uint16_t *)(v)) == 0 : \ + sizeof(*v) == sizeof(uint32_t) ? atomic_dec_32_nv((uint32_t *)(v)) == 0 : \ + sizeof(*v) == sizeof(uint64_t) ? atomic_dec_64_nv((uint64_t *)(v)) == 0 : \ + (assert(!"should not get here"), 0)) + +#define p_atomic_inc(v) (void) (\ + sizeof(*v) == sizeof(uint8_t) ? atomic_inc_8 ((uint8_t *)(v)) : \ + sizeof(*v) == sizeof(uint16_t) ? atomic_inc_16((uint16_t *)(v)) : \ + sizeof(*v) == sizeof(uint32_t) ? atomic_inc_32((uint32_t *)(v)) : \ + sizeof(*v) == sizeof(uint64_t) ? atomic_inc_64((uint64_t *)(v)) : \ + (assert(!"should not get here"), 0)) + +#define p_atomic_inc_return(v) ((__typeof(*v)) \ + sizeof(*v) == sizeof(uint8_t) ? atomic_inc_8_nv ((uint8_t *)(v)) : \ + sizeof(*v) == sizeof(uint16_t) ? atomic_inc_16_nv((uint16_t *)(v)) : \ + sizeof(*v) == sizeof(uint32_t) ? atomic_inc_32_nv((uint32_t *)(v)) : \ + sizeof(*v) == sizeof(uint64_t) ? atomic_inc_64_nv((uint64_t *)(v)) : \ + (assert(!"should not get here"), 0)) + +#define p_atomic_dec(v) ((void) \ + sizeof(*v) == sizeof(uint8_t) ? atomic_dec_8 ((uint8_t *)(v)) : \ + sizeof(*v) == sizeof(uint16_t) ? atomic_dec_16((uint16_t *)(v)) : \ + sizeof(*v) == sizeof(uint32_t) ? atomic_dec_32((uint32_t *)(v)) : \ + sizeof(*v) == sizeof(uint64_t) ? atomic_dec_64((uint64_t *)(v)) : \ + (assert(!"should not get here"), 0)) + +#define p_atomic_dec_return(v) ((__typeof(*v)) \ + sizeof(*v) == sizeof(uint8_t) ? atomic_dec_8_nv ((uint8_t *)(v)) : \ + sizeof(*v) == sizeof(uint16_t) ? atomic_dec_16_nv((uint16_t *)(v)) : \ + sizeof(*v) == sizeof(uint32_t) ? atomic_dec_32_nv((uint32_t *)(v)) : \ + sizeof(*v) == sizeof(uint64_t) ? atomic_dec_64_nv((uint64_t *)(v)) : \ + (assert(!"should not get here"), 0)) + +#define p_atomic_add(v, i) ((void) \ + sizeof(*v) == sizeof(uint8_t) ? atomic_add_8 ((uint8_t *)(v), (i)) : \ + sizeof(*v) == sizeof(uint16_t) ? atomic_add_16((uint16_t *)(v), (i)) : \ + sizeof(*v) == sizeof(uint32_t) ? atomic_add_32((uint32_t *)(v), (i)) : \ + sizeof(*v) == sizeof(uint64_t) ? atomic_add_64((uint64_t *)(v), (i)) : \ + (assert(!"should not get here"), 0)) + +#define p_atomic_cmpxchg(v, old, _new) ((__typeof(*v)) \ + sizeof(*v) == sizeof(uint8_t) ? atomic_cas_8 ((uint8_t *)(v), (uint8_t )(old), (uint8_t )(_new)) : \ + sizeof(*v) == sizeof(uint16_t) ? atomic_cas_16((uint16_t *)(v), (uint16_t)(old), (uint16_t)(_new)) : \ + sizeof(*v) == sizeof(uint32_t) ? atomic_cas_32((uint32_t *)(v), (uint32_t)(old), (uint32_t)(_new)) : \ + sizeof(*v) == sizeof(uint64_t) ? atomic_cas_64((uint64_t *)(v), (uint64_t)(old), (uint64_t)(_new)) : \ + (assert(!"should not get here"), 0)) - return n != 0; -} - -#define p_atomic_inc(_v) atomic_inc_32((uint32_t *) _v) -#define p_atomic_dec(_v) atomic_dec_32((uint32_t *) _v) -#define p_atomic_inc_return(_v) atomic_inc_32_nv((uint32_t *) _v) -#define p_atomic_dec_return(_v) atomic_dec_32_nv((uint32_t *) _v) - -#define p_atomic_cmpxchg(_v, _old, _new) \ - atomic_cas_32( (uint32_t *) _v, (uint32_t) _old, (uint32_t) _new) - -#ifdef __cplusplus -} #endif -#endif - - #ifndef PIPE_ATOMIC #error "No pipe_atomic implementation selected" #endif