1 //===-- sanitizer_atomic_clang_x86.h ----------------------------*- C++ -*-===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file is a part of ThreadSanitizer/AddressSanitizer runtime. 11 // Not intended for direct inclusion. Include sanitizer_atomic.h. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #ifndef SANITIZER_ATOMIC_CLANG_X86_H 16 #define SANITIZER_ATOMIC_CLANG_X86_H 17 18 namespace __sanitizer { 19 20 INLINE void proc_yield(int cnt) { 21 __asm__ __volatile__("" ::: "memory"); 22 for (int i = 0; i < cnt; i++) 23 __asm__ __volatile__("pause"); 24 __asm__ __volatile__("" ::: "memory"); 25 } 26 27 template<typename T> 28 INLINE typename T::Type atomic_load( 29 const volatile T *a, memory_order mo) { 30 DCHECK(mo & (memory_order_relaxed | memory_order_consume 31 | memory_order_acquire | memory_order_seq_cst)); 32 DCHECK(!((uptr)a % sizeof(*a))); 33 typename T::Type v; 34 35 if (sizeof(*a) < 8 || sizeof(void*) == 8) { 36 // Assume that aligned loads are atomic. 37 if (mo == memory_order_relaxed) { 38 v = a->val_dont_use; 39 } else if (mo == memory_order_consume) { 40 // Assume that processor respects data dependencies 41 // (and that compiler won't break them). 42 __asm__ __volatile__("" ::: "memory"); 43 v = a->val_dont_use; 44 __asm__ __volatile__("" ::: "memory"); 45 } else if (mo == memory_order_acquire) { 46 __asm__ __volatile__("" ::: "memory"); 47 v = a->val_dont_use; 48 // On x86 loads are implicitly acquire. 49 __asm__ __volatile__("" ::: "memory"); 50 } else { // seq_cst 51 // On x86 plain MOV is enough for seq_cst store. 52 __asm__ __volatile__("" ::: "memory"); 53 v = a->val_dont_use; 54 __asm__ __volatile__("" ::: "memory"); 55 } 56 } else { 57 // 64-bit load on 32-bit platform. 58 __asm__ __volatile__( 59 "movq %1, %%mm0;" // Use mmx reg for 64-bit atomic moves 60 "movq %%mm0, %0;" // (ptr could be read-only) 61 "emms;" // Empty mmx state/Reset FP regs 62 : "=m" (v) 63 : "m" (a->val_dont_use) 64 : // mark the FP stack and mmx registers as clobbered 65 "st", "st(1)", "st(2)", "st(3)", "st(4)", "st(5)", "st(6)", "st(7)", 66 #ifdef __MMX__ 67 "mm0", "mm1", "mm2", "mm3", "mm4", "mm5", "mm6", "mm7", 68 #endif // #ifdef __MMX__ 69 "memory"); 70 } 71 return v; 72 } 73 74 template<typename T> 75 INLINE void atomic_store(volatile T *a, typename T::Type v, memory_order mo) { 76 DCHECK(mo & (memory_order_relaxed | memory_order_release 77 | memory_order_seq_cst)); 78 DCHECK(!((uptr)a % sizeof(*a))); 79 80 if (sizeof(*a) < 8 || sizeof(void*) == 8) { 81 // Assume that aligned loads are atomic. 82 if (mo == memory_order_relaxed) { 83 a->val_dont_use = v; 84 } else if (mo == memory_order_release) { 85 // On x86 stores are implicitly release. 86 __asm__ __volatile__("" ::: "memory"); 87 a->val_dont_use = v; 88 __asm__ __volatile__("" ::: "memory"); 89 } else { // seq_cst 90 // On x86 stores are implicitly release. 91 __asm__ __volatile__("" ::: "memory"); 92 a->val_dont_use = v; 93 __sync_synchronize(); 94 } 95 } else { 96 // 64-bit store on 32-bit platform. 97 __asm__ __volatile__( 98 "movq %1, %%mm0;" // Use mmx reg for 64-bit atomic moves 99 "movq %%mm0, %0;" 100 "emms;" // Empty mmx state/Reset FP regs 101 : "=m" (a->val_dont_use) 102 : "m" (v) 103 : // mark the FP stack and mmx registers as clobbered 104 "st", "st(1)", "st(2)", "st(3)", "st(4)", "st(5)", "st(6)", "st(7)", 105 #ifdef __MMX__ 106 "mm0", "mm1", "mm2", "mm3", "mm4", "mm5", "mm6", "mm7", 107 #endif // #ifdef __MMX__ 108 "memory"); 109 if (mo == memory_order_seq_cst) 110 __sync_synchronize(); 111 } 112 } 113 114 } // namespace __sanitizer 115 116 #endif // #ifndef SANITIZER_ATOMIC_CLANG_X86_H 117