|
25 | 25 | #define pg_write_barrier_impl() __asm__ __volatile__ ("lwsync" : : : "memory")
|
26 | 26 | #endif
|
27 | 27 |
|
| 28 | +#define PG_HAVE_ATOMIC_U32_SUPPORT |
| 29 | +typedef struct pg_atomic_uint32 |
| 30 | +{ |
| 31 | + volatile uint32 value; |
| 32 | +} pg_atomic_uint32; |
| 33 | + |
| 34 | +/* 64bit atomics are only supported in 64bit mode */ |
| 35 | +#ifdef __64BIT__ |
| 36 | +#define PG_HAVE_ATOMIC_U64_SUPPORT |
| 37 | +typedef struct pg_atomic_uint64 |
| 38 | +{ |
| 39 | + volatile uint64 value pg_attribute_aligned(8); |
| 40 | +} pg_atomic_uint64; |
| 41 | + |
| 42 | +#endif /* __64BIT__ */ |
| 43 | + |
| 44 | +#define PG_HAVE_ATOMIC_FETCH_ADD_U32 |
| 45 | +static inline uint32 |
| 46 | +pg_atomic_fetch_add_u32_impl(volatile pg_atomic_uint32 *ptr, int32 add_) |
| 47 | +{ |
| 48 | + uint32 _t; |
| 49 | + uint32 res; |
| 50 | + |
| 51 | + /* |
| 52 | + * xlc has a no-longer-documented __fetch_and_add() intrinsic. In xlc |
| 53 | + * 12.01.0000.0000, it emits a leading "sync" and trailing "isync". In |
| 54 | + * xlc 13.01.0003.0004, it emits neither. Hence, using the intrinsic |
| 55 | + * would add redundant syncs on xlc 12. |
| 56 | + */ |
| 57 | +#ifdef HAVE_I_CONSTRAINT__BUILTIN_CONSTANT_P |
| 58 | + if (__builtin_constant_p(add_) && |
| 59 | + add_ <= PG_INT16_MAX && add_ >= PG_INT16_MIN) |
| 60 | + __asm__ __volatile__( |
| 61 | + " sync \n" |
| 62 | + " lwarx %1,0,%4 \n" |
| 63 | + " addi %0,%1,%3 \n" |
| 64 | + " stwcx. %0,0,%4 \n" |
| 65 | + " bne $-12 \n" /* branch to lwarx */ |
| 66 | + " isync \n" |
| 67 | +: "=&r"(_t), "=&r"(res), "+m"(ptr->value) |
| 68 | +: "i"(add_), "r"(&ptr->value) |
| 69 | +: "memory", "cc"); |
| 70 | + else |
| 71 | +#endif |
| 72 | + __asm__ __volatile__( |
| 73 | + " sync \n" |
| 74 | + " lwarx %1,0,%4 \n" |
| 75 | + " add %0,%1,%3 \n" |
| 76 | + " stwcx. %0,0,%4 \n" |
| 77 | + " bne $-12 \n" /* branch to lwarx */ |
| 78 | + " isync \n" |
| 79 | +: "=&r"(_t), "=&r"(res), "+m"(ptr->value) |
| 80 | +: "r"(add_), "r"(&ptr->value) |
| 81 | +: "memory", "cc"); |
| 82 | + |
| 83 | + return res; |
| 84 | +} |
| 85 | + |
| 86 | +#ifdef PG_HAVE_ATOMIC_U64_SUPPORT |
| 87 | +#define PG_HAVE_ATOMIC_FETCH_ADD_U64 |
| 88 | +static inline uint64 |
| 89 | +pg_atomic_fetch_add_u64_impl(volatile pg_atomic_uint64 *ptr, int64 add_) |
| 90 | +{ |
| 91 | + uint64 _t; |
| 92 | + uint64 res; |
| 93 | + |
| 94 | + /* Like u32, but s/lwarx/ldarx/; s/stwcx/stdcx/ */ |
| 95 | +#ifdef HAVE_I_CONSTRAINT__BUILTIN_CONSTANT_P |
| 96 | + if (__builtin_constant_p(add_) && |
| 97 | + add_ <= PG_INT16_MAX && add_ >= PG_INT16_MIN) |
| 98 | + __asm__ __volatile__( |
| 99 | + " sync \n" |
| 100 | + " ldarx %1,0,%4 \n" |
| 101 | + " addi %0,%1,%3 \n" |
| 102 | + " stdcx. %0,0,%4 \n" |
| 103 | + " bne $-12 \n" /* branch to ldarx */ |
| 104 | + " isync \n" |
| 105 | +: "=&r"(_t), "=&r"(res), "+m"(ptr->value) |
| 106 | +: "i"(add_), "r"(&ptr->value) |
| 107 | +: "memory", "cc"); |
| 108 | + else |
| 109 | +#endif |
| 110 | + __asm__ __volatile__( |
| 111 | + " sync \n" |
| 112 | + " ldarx %1,0,%4 \n" |
| 113 | + " add %0,%1,%3 \n" |
| 114 | + " stdcx. %0,0,%4 \n" |
| 115 | + " bne $-12 \n" /* branch to ldarx */ |
| 116 | + " isync \n" |
| 117 | +: "=&r"(_t), "=&r"(res), "+m"(ptr->value) |
| 118 | +: "r"(add_), "r"(&ptr->value) |
| 119 | +: "memory", "cc"); |
| 120 | + |
| 121 | + return res; |
| 122 | +} |
| 123 | + |
| 124 | +#endif /* PG_HAVE_ATOMIC_U64_SUPPORT */ |
| 125 | + |
28 | 126 | /* per architecture manual doubleword accesses have single copy atomicity */
|
29 | 127 | #define PG_HAVE_8BYTE_SINGLE_COPY_ATOMICITY
|
0 commit comments