x86平台读取cpu支持sse2指令集的代码，以及原子操作的代码

// This module gets enough CPU information to optimize the
// atomicops module on x86.

#include <string.h>

#include "base/atomicops.h"
#include "base/basictypes.h"

// This file only makes sense with atomicops_internals_x86_gcc.h -- it
// depends on structs that are defined in that file.  If atomicops.h
// doesn't sub-include that file, then we aren't needed, and shouldn't
// try to do anything.
#ifdef BASE_ATOMICOPS_INTERNALS_X86_GCC_H_

// Inline cpuid instruction.  In PIC compilations, %ebx contains the address
// of the global offset table.  To avoid breaking such executables, this code
// must preserve that register's value across cpuid instructions.
#if defined(__i386__)
#define cpuid(a, b, c, d, inp) \
  asm ("mov %%ebx, %%edi\n"    \
       "cpuid\n"               \
       "xchg %%edi, %%ebx\n"   \
       : "=a" (a), "=D" (b), "=c" (c), "=d" (d) : "a" (inp))
#elif defined (__x86_64__)
#define cpuid(a, b, c, d, inp) \
  asm ("mov %%rbx, %%rdi\n"    \
       "cpuid\n"               \
       "xchg %%rdi, %%rbx\n"   \
       : "=a" (a), "=D" (b), "=c" (c), "=d" (d) : "a" (inp))
#endif

#if defined(cpuid)        // initialize the struct only on x86

// Set the flags so that code will run correctly and conservatively, so even
// if we haven't been initialized yet, we're probably single threaded, and our
// default values should hopefully be pretty safe.
struct AtomicOps_x86CPUFeatureStruct AtomicOps_Internalx86CPUFeatures = {
  false,          // bug can't exist before process spawns multiple threads
  false,          // no SSE2
};

// Initialize the AtomicOps_Internalx86CPUFeatures struct.
static void AtomicOps_Internalx86CPUFeaturesInit()
{
  uint32 eax;
  uint32 ebx;
  uint32 ecx;
  uint32 edx;

  // Get vendor string (issue CPUID with eax = 0)
  cpuid(eax, ebx, ecx, edx, 0);
  char vendor[13];
  memcpy(vendor, &ebx, 4);
  memcpy(vendor + 4, &edx, 4);
  memcpy(vendor + 8, &ecx, 4);
  vendor[12] = 0;

  // get feature flags in ecx/edx, and family/model in eax
  cpuid(eax, ebx, ecx, edx, 1);

  int family = (eax >> 8) & 0xf;        // family and model fields
  int model = (eax >> 4) & 0xf;
  if (family == 0xf) {                  // use extended family and model fields
    family += (eax >> 20) & 0xff;
    model += ((eax >> 16) & 0xf) << 4;
  }

  // Opteron Rev E has a bug in which on very rare occasions a locked
  // instruction doesn't act as a read-acquire barrier if followed by a
  // non-locked read-modify-write instruction.  Rev F has this bug in
  // pre-release versions, but not in versions released to customers,
  // so we test only for Rev E, which is family 15, model 32..63 inclusive.
  if (strcmp(vendor, "AuthenticAMD") == 0 &&       // AMD
      family == 15 &&
      32 <= model && model <= 63) {
    AtomicOps_Internalx86CPUFeatures.has_amd_lock_mb_bug = true;
  } else {
    AtomicOps_Internalx86CPUFeatures.has_amd_lock_mb_bug = false;
  }

  // edx bit 26 is SSE2 which we use to tell use whether we can use mfence
  AtomicOps_Internalx86CPUFeatures.has_sse2 = ((edx >> 26) & 1);
}

namespace
{

class AtomicOpsx86Initializer
{

 public:
  
  AtomicOpsx86Initializer()
  {
    AtomicOps_Internalx86CPUFeaturesInit();
  }
 
};

// A global to get use initialized on startup via static initialization :/
AtomicOpsx86Initializer g_initer;

}  // namespace

#endif  // if x86

#endif  // ifdef BASE_ATOMICOPS_INTERNALS_X86_GCC_H_

 

 

 

 

// This file is an internal atomic implementation, use base/atomicops.h instead.

#ifndef BASE_ATOMICOPS_INTERNALS_X86_GCC_H_
#define BASE_ATOMICOPS_INTERNALS_X86_GCC_H_

#include "base/base_export.h"

// This struct is not part of the public API of this module; clients may not
// use it.  (However, it's exported via BASE_EXPORT because clients implicitly
// do use it at link time by inlining these functions.)
// Features of this x86.  Values may not be correct before main() is run,
// but are set conservatively.
struct AtomicOps_x86CPUFeatureStruct
{
  bool has_amd_lock_mb_bug; // Processor has AMD memory-barrier bug; do lfence
                            // after acquire compare-and-swap.
  bool has_sse2;            // Processor has SSE2.
};
BASE_EXPORT extern struct AtomicOps_x86CPUFeatureStruct
    AtomicOps_Internalx86CPUFeatures;

#define ATOMICOPS_COMPILER_BARRIER() __asm__ __volatile__("" : : : "memory")

namespace base
{

namespace subtle
{

// 32-bit low-level operations on any platform.

inline Atomic32 NoBarrier_CompareAndSwap(volatile Atomic32* ptr,
                                         Atomic32 old_value,
                                         Atomic32 new_value)
{
  Atomic32 prev;
  __asm__ __volatile__("lock; cmpxchgl %1,%2"
                       : "=a" (prev)
                       : "q" (new_value), "m" (*ptr), "0" (old_value)
                       : "memory");
  return prev;
}

inline Atomic32 NoBarrier_AtomicExchange(volatile Atomic32* ptr,
                                         Atomic32 new_value) {
  __asm__ __volatile__("xchgl %1,%0"  // The lock prefix is implicit for xchg.
                       : "=r" (new_value)
                       : "m" (*ptr), "0" (new_value)
                       : "memory");
  return new_value;  // Now it's the previous value.
}

inline Atomic32 NoBarrier_AtomicIncrement(volatile Atomic32* ptr,
                                          Atomic32 increment) {
  Atomic32 temp = increment;
  __asm__ __volatile__("lock; xaddl %0,%1"
                       : "+r" (temp), "+m" (*ptr)
                       : : "memory");
  // temp now holds the old value of *ptr
  return temp + increment;
}

inline Atomic32 Barrier_AtomicIncrement(volatile Atomic32* ptr,
                                        Atomic32 increment) {
  Atomic32 temp = increment;
  __asm__ __volatile__("lock; xaddl %0,%1"
                       : "+r" (temp), "+m" (*ptr)
                       : : "memory");
  // temp now holds the old value of *ptr
  if (AtomicOps_Internalx86CPUFeatures.has_amd_lock_mb_bug) {
    __asm__ __volatile__("lfence" : : : "memory");
  }
  return temp + increment;
}

inline Atomic32 Acquire_CompareAndSwap(volatile Atomic32* ptr,
                                       Atomic32 old_value,
                                       Atomic32 new_value) {
  Atomic32 x = NoBarrier_CompareAndSwap(ptr, old_value, new_value);
  if (AtomicOps_Internalx86CPUFeatures.has_amd_lock_mb_bug)
  {
    __asm__ __volatile__("lfence" : : : "memory");
  }
  return x;
}

inline Atomic32 Release_CompareAndSwap(volatile Atomic32* ptr,
                                       Atomic32 old_value,
                                       Atomic32 new_value)
{
  return NoBarrier_CompareAndSwap(ptr, old_value, new_value);
}

inline void NoBarrier_Store(volatile Atomic32* ptr, Atomic32 value)
{
  *ptr = value;
}

#if defined(__x86_64__)

// 64-bit implementations of memory barrier can be simpler, because it
// "mfence" is guaranteed to exist.
inline void MemoryBarrier()
{
  __asm__ __volatile__("mfence" : : : "memory");
}

inline void Acquire_Store(volatile Atomic32* ptr, Atomic32 value)
{
  *ptr = value;
  MemoryBarrier();
}

#else

inline void MemoryBarrier()
{
  if (AtomicOps_Internalx86CPUFeatures.has_sse2)
  {
    __asm__ __volatile__("mfence" : : : "memory");
  } else { // mfence is faster but not present on PIII
    Atomic32 x = 0;
    NoBarrier_AtomicExchange(&x, 0);  // acts as a barrier on PIII
  }
}

inline void Acquire_Store(volatile Atomic32* ptr, Atomic32 value) {
  if (AtomicOps_Internalx86CPUFeatures.has_sse2) {
    *ptr = value;
    __asm__ __volatile__("mfence" : : : "memory");
  } else {
    NoBarrier_AtomicExchange(ptr, value);
                          // acts as a barrier on PIII
  }
}
#endif

inline void Release_Store(volatile Atomic32* ptr, Atomic32 value)
{
  ATOMICOPS_COMPILER_BARRIER();
  *ptr = value; // An x86 store acts as a release barrier.
  // See comments in Atomic64 version of Release_Store(), below.
}

inline Atomic32 NoBarrier_Load(volatile const Atomic32* ptr) {
  return *ptr;
}

inline Atomic32 Acquire_Load(volatile const Atomic32* ptr) {
  Atomic32 value = *ptr; // An x86 load acts as a acquire barrier.
  // See comments in Atomic64 version of Release_Store(), below.
  ATOMICOPS_COMPILER_BARRIER();
  return value;
}

inline Atomic32 Release_Load(volatile const Atomic32* ptr) {
  MemoryBarrier();
  return *ptr;
}

#if defined(__x86_64__)

// 64-bit low-level operations on 64-bit platform.

inline Atomic64 NoBarrier_CompareAndSwap(volatile Atomic64* ptr,
                                         Atomic64 old_value,
                                         Atomic64 new_value) {
  Atomic64 prev;
  __asm__ __volatile__("lock; cmpxchgq %1,%2"
                       : "=a" (prev)
                       : "q" (new_value), "m" (*ptr), "0" (old_value)
                       : "memory");
  return prev;
}

inline Atomic64 NoBarrier_AtomicExchange(volatile Atomic64* ptr,
                                         Atomic64 new_value) {
  __asm__ __volatile__("xchgq %1,%0"  // The lock prefix is implicit for xchg.
                       : "=r" (new_value)
                       : "m" (*ptr), "0" (new_value)
                       : "memory");
  return new_value;  // Now it's the previous value.
}

inline Atomic64 NoBarrier_AtomicIncrement(volatile Atomic64* ptr,
                                          Atomic64 increment) {
  Atomic64 temp = increment;
  __asm__ __volatile__("lock; xaddq %0,%1"
                       : "+r" (temp), "+m" (*ptr)
                       : : "memory");
  // temp now contains the previous value of *ptr
  return temp + increment;
}

inline Atomic64 Barrier_AtomicIncrement(volatile Atomic64* ptr,
                                        Atomic64 increment) {
  Atomic64 temp = increment;
  __asm__ __volatile__("lock; xaddq %0,%1"
                       : "+r" (temp), "+m" (*ptr)
                       : : "memory");
  // temp now contains the previous value of *ptr
  if (AtomicOps_Internalx86CPUFeatures.has_amd_lock_mb_bug) {
    __asm__ __volatile__("lfence" : : : "memory");
  }
  return temp + increment;
}

inline void NoBarrier_Store(volatile Atomic64* ptr, Atomic64 value) {
  *ptr = value;
}

inline void Acquire_Store(volatile Atomic64* ptr, Atomic64 value) {
  *ptr = value;
  MemoryBarrier();
}

inline void Release_Store(volatile Atomic64* ptr, Atomic64 value) {
  ATOMICOPS_COMPILER_BARRIER();

  *ptr = value; // An x86 store acts as a release barrier
                // for current AMD/Intel chips as of Jan 2008.
                // See also Acquire_Load(), below.

  // When new chips come out, check:
  //  IA-32 Intel Architecture Software Developer's Manual, Volume 3:
  //  System Programming Guide, Chatper 7: Multiple-processor management,
  //  Section 7.2, Memory Ordering.
  // Last seen at:
  //   http://developer.intel.com/design/pentium4/manuals/index_new.htm
  //
  // x86 stores/loads fail to act as barriers for a few instructions (clflush
  // maskmovdqu maskmovq movntdq movnti movntpd movntps movntq) but these are
  // not generated by the compiler, and are rare.  Users of these instructions
  // need to know about cache behaviour in any case since all of these involve
  // either flushing cache lines or non-temporal cache hints.
}

inline Atomic64 NoBarrier_Load(volatile const Atomic64* ptr) {
  return *ptr;
}

inline Atomic64 Acquire_Load(volatile const Atomic64* ptr) {
  Atomic64 value = *ptr; // An x86 load acts as a acquire barrier,
                         // for current AMD/Intel chips as of Jan 2008.
                         // See also Release_Store(), above.
  ATOMICOPS_COMPILER_BARRIER();
  return value;
}

inline Atomic64 Release_Load(volatile const Atomic64* ptr) {
  MemoryBarrier();
  return *ptr;
}

inline Atomic64 Acquire_CompareAndSwap(volatile Atomic64* ptr,
                                       Atomic64 old_value,
                                       Atomic64 new_value) {
  Atomic64 x = NoBarrier_CompareAndSwap(ptr, old_value, new_value);
  if (AtomicOps_Internalx86CPUFeatures.has_amd_lock_mb_bug) {
    __asm__ __volatile__("lfence" : : : "memory");
  }
  return x;
}

inline Atomic64 Release_CompareAndSwap(volatile Atomic64* ptr,
                                       Atomic64 old_value,
                                       Atomic64 new_value) {
  return NoBarrier_CompareAndSwap(ptr, old_value, new_value);
}

#endif  // defined(__x86_64__)

} // namespace base::subtle
} // namespace base

#undef ATOMICOPS_COMPILER_BARRIER

#endif  // BASE_ATOMICOPS_INTERNALS_X86_GCC_H_