--- libeio/ecb.h 2012/05/28 17:00:19 1.12 +++ libeio/ecb.h 2012/06/17 17:07:25 1.14 @@ -81,16 +81,22 @@ #endif #endif +#define ECB_C (__STDC__+0) /* this assumes that __STDC__ is either empty or a number */ +#define ECB_C99 (__STDC_VERSION__ >= 199901L) +#define ECB_C11 (__STDC_VERSION__ >= 201112L) +#define ECB_CPP (__cplusplus+0) +#define ECB_CPP11 (__cplusplus >= 201103L) + /*****************************************************************************/ /* ECB_NO_THREADS - ecb is not used by multiple threads, ever */ /* ECB_NO_SMP - ecb might be used in multiple threads, but only on a single cpu */ #if ECB_NO_THREADS -# define ECB_NO_SMP 1 + #define ECB_NO_SMP 1 #endif -#if ECB_NO_THREADS || ECB_NO_SMP +#if ECB_NO_SMP #define ECB_MEMORY_FENCE do { } while (0) #endif @@ -98,39 +104,48 @@ #if ECB_GCC_VERSION(2,5) || defined __INTEL_COMPILER || (__llvm__ && __GNUC__) || __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110 #if __i386 || __i386__ #define ECB_MEMORY_FENCE __asm__ __volatile__ ("lock; orb $0, -1(%%esp)" : : : "memory") - #define ECB_MEMORY_FENCE_ACQUIRE ECB_MEMORY_FENCE /* non-lock xchg might be enough */ - #define ECB_MEMORY_FENCE_RELEASE do { } while (0) /* unlikely to change in future cpus */ + #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory") + #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") #elif __amd64 || __amd64__ || __x86_64 || __x86_64__ - #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory") - #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("lfence" : : : "memory") - #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("sfence") /* play safe - not needed in any current cpu */ + #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory") + #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory") + #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") #elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ - #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory") + #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory") #elif defined __ARM_ARCH_6__ || defined __ARM_ARCH_6J__ \ || defined __ARM_ARCH_6K__ || defined __ARM_ARCH_6ZK__ #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mcr p15,0,%0,c7,c10,5" : : "r" (0) : "memory") #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \ - || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__ - #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory") + || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__ + #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory") #elif __sparc || __sparc__ - #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad | " : : : "memory") - #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory") - #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore") + #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory") + #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory") + #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore") #elif defined __s390__ || defined __s390x__ #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory") #elif defined __mips__ - #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory") + #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory") #elif defined __alpha__ - #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mb" : : : "memory") + #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mb" : : : "memory") + #elif defined __hppa__ + #define ECB_MEMORY_FENCE __asm__ __volatile__ ("" : : : "memory") + #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") + #elif defined __ia64__ + #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mf" : : : "memory") #endif #endif #endif #ifndef ECB_MEMORY_FENCE - #if ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__ + #if ECB_GCC_VERSION(4,7) + /* see comment below (stdatomic.h) about the C11 memory model. */ + #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST) + #elif defined __clang && __has_feature (cxx_atomic) + /* see comment below (stdatomic.h) about the C11 memory model. */ + #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST) + #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__ #define ECB_MEMORY_FENCE __sync_synchronize () - /*#define ECB_MEMORY_FENCE_ACQUIRE ({ char dummy = 0; __sync_lock_test_and_set (&dummy, 1); }) */ - /*#define ECB_MEMORY_FENCE_RELEASE ({ char dummy = 1; __sync_lock_release (&dummy ); }) */ #elif _MSC_VER >= 1400 /* VC++ 2005 */ #pragma intrinsic(_ReadBarrier,_WriteBarrier,_ReadWriteBarrier) #define ECB_MEMORY_FENCE _ReadWriteBarrier () @@ -150,6 +165,21 @@ #endif #ifndef ECB_MEMORY_FENCE + #if ECB_C11 && !defined __STDC_NO_ATOMICS__ + /* we assume that these memory fences work on all variables/all memory accesses, */ + /* not just C11 atomics and atomic accesses */ + #include + /* Unfortunately, neither gcc 4.7 nor clang 3.1 generate any instructions for */ + /* any fence other than seq_cst, which isn't very efficient for us. */ + /* Why that is, we don't know - either the C11 memory model is quite useless */ + /* for most usages, or gcc and clang have a bug */ + /* I *currently* lean towards the latter, and inefficiently implement */ + /* all three of ecb's fences as a seq_cst fence */ + #define ECB_MEMORY_FENCE atomic_thread_fence (memory_order_seq_cst) + #endif +#endif + +#ifndef ECB_MEMORY_FENCE #if !ECB_AVOID_PTHREADS /* * if you get undefined symbol references to pthread_mutex_lock, @@ -177,8 +207,6 @@ /*****************************************************************************/ -#define ECB_C99 (__STDC_VERSION__ >= 199901L) - #if __cplusplus #define ecb_inline static inline #elif ECB_GCC_VERSION(2,5) @@ -226,11 +254,16 @@ #endif #define ecb_noinline ecb_attribute ((__noinline__)) -#define ecb_noreturn ecb_attribute ((__noreturn__)) #define ecb_unused ecb_attribute ((__unused__)) #define ecb_const ecb_attribute ((__const__)) #define ecb_pure ecb_attribute ((__pure__)) +#if ECB_C11 + #define ecb_noreturn _Noreturn +#else + #define ecb_noreturn ecb_attribute ((__noreturn__)) +#endif + #if ECB_GCC_VERSION(4,3) #define ecb_artificial ecb_attribute ((__artificial__)) #define ecb_hot ecb_attribute ((__hot__)) @@ -428,14 +461,32 @@ #endif /* try to tell the compiler that some condition is definitely true */ -#define ecb_assume(cond) do { if (!(cond)) ecb_unreachable (); } while (0) +#define ecb_assume(cond) if (!(cond)) ecb_unreachable (); else 0 ecb_inline unsigned char ecb_byteorder_helper (void) ecb_const; ecb_inline unsigned char ecb_byteorder_helper (void) { - const uint32_t u = 0x11223344; - return *(unsigned char *)&u; + /* the union code still generates code under pressure in gcc, */ + /* but less than using pointers, and always seem to */ + /* successfully return a constant. */ + /* the reason why we have this horrible preprocessor mess */ + /* is to avoid it in all cases, at least on common architectures */ + /* and yes, gcc defines __BYTE_ORDER__, g++ does not */ +#if __i386 || __i386__ || _M_X86 || __amd64 || __amd64__ || _M_X64 + return 0x44; +#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ + return 0x44; +#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ + retrurn 0x11; +#else + union + { + uint32_t i; + uint8_t c; + } u = { 0x11223344 }; + return u.c; +#endif } ecb_inline ecb_bool ecb_big_endian (void) ecb_const;