【整理】__builtin_expect 解惑

转载自：https://my.oschina.net/moooofly/blog/175019

最近看 GLib 的代码遇到这个东东，网上搜索一圈，发现很多人都写过这个，自己今天才研究到，汗颜一下，扫盲一个点，留此记录为证！ 

首先看一篇最官方的讲解： 

====== 
likely() and unlikely() 

What are they ? 
      In Linux kernel code, one often find calls to likely() and unlikely(), in conditions, like : 

bvl = bvec_alloc(gfp_mask, nr_iovecs, &idx);
if (unlikely(!bvl)) {
  mempool_free(bio, bio_pool);
  bio = NULL;
  goto out;
}

      In fact, these functions are hints for the compiler that allows it to correctly optimize the branch, by knowing which is the likeliest one. The definitions of these macros, found in include/linux/compiler.h are the following :  

#define likely(x)       __builtin_expect(!!(x), 1)
#define unlikely(x)     __builtin_expect(!!(x), 0)

The GCC documentation explains the role of __builtin_expect() :  

-- Built-in Function: long __builtin_expect (long EXP, long C)
     You may use `__builtin_expect' to provide the compiler with branch
     prediction information.  In general, you should prefer to use
     actual profile feedback for this (`-fprofile-arcs'), as
     programmers are notoriously bad at predicting how their programs
     actually perform.  However, there are applications in which this
     data is hard to collect.

     The return value is the value of EXP, which should be an integral
     expression.  The value of C must be a compile-time constant.  The
     semantics of the built-in are that it is expected that EXP == C.
     For example:

          if (__builtin_expect (x, 0))
            foo ();

     would indicate that we do not expect to call `foo', since we
     expect `x' to be zero.  Since you are limited to integral
     expressions for EXP, you should use constructions such as

          if (__builtin_expect (ptr != NULL, 1))
            error ();

     when testing pointer or floating-point values.

  How does it optimize things ?  
      It optimizes things by ordering the generated assembly code correctly, to optimize the usage of the processor pipeline. To do so, they arrange the code so that the likeliest branch is executed without performing any jmp instruction (which has the bad effect of flushing the processor pipeline).  

To see how it works, let's compile the following simple C user space program with gcc -O2 :  

#define likely(x)    __builtin_expect(!!(x), 1)
#define unlikely(x)  __builtin_expect(!!(x), 0)

int main(char *argv[], int argc)
{
   int a;

   /* Get the value from somewhere GCC can't optimize */
   a = atoi (argv[1]);

   if (unlikely (a == 2))
      a++;
   else
      a--;

   printf ("%d\n", a);

   return 0;
}

Now, disassemble the resulting binary using objdump -S (comments added by me) :  

080483b0 
:
 // Prologue
 80483b0:       55                      push   %ebp
 80483b1:       89 e5                   mov    %esp,%ebp
 80483b3:       50                      push   %eax
 80483b4:       50                      push   %eax
 80483b5:       83 e4 f0                and    $0xfffffff0,%esp
 //             Call atoi()
 80483b8:       8b 45 08                mov    0x8(%ebp),%eax
 80483bb:       83 ec 1c                sub    $0x1c,%esp
 80483be:       8b 48 04                mov    0x4(%eax),%ecx
 80483c1:       51                      push   %ecx
 80483c2:       e8 1d ff ff ff          call   80482e4 
 80483c7:       83 c4 10                add    $0x10,%esp
 //             Test the value
 80483ca:       83 f8 02                cmp    $0x2,%eax
 //             --------------------------------------------------------
 //             If 'a' equal to 2 (which is unlikely), then jump,
 //             otherwise continue directly, without jump, so that it
 //             doesn't flush the pipeline.
 //             --------------------------------------------------------
 80483cd:       74 12                   je     80483e1 0x31>
 80483cf:       48                      dec    %eax
 //             Call printf
 80483d0:       52                      push   %edx
 80483d1:       52                      push   %edx
 80483d2:       50                      push   %eax
 80483d3:       68 c8 84 04 08          push   $0x80484c8
 80483d8:       e8 f7 fe ff ff          call   80482d4 
 //             Return 0 and go out.
 80483dd:       31 c0                   xor    %eax,%eax
 80483df:       c9                      leave
 80483e0:       c3                      ret

Now, in the previous program, replace the unlikely() by a likely(), recompile it, and disassemble it again (again, comments added by me) :  

080483b0 
:
 //             Prologue
 80483b0:       55                      push   %ebp
 80483b1:       89 e5                   mov    %esp,%ebp
 80483b3:       50                      push   %eax
 80483b4:       50                      push   %eax
 80483b5:       83 e4 f0                and    $0xfffffff0,%esp
 //             Call atoi()
 80483b8:       8b 45 08                mov    0x8(%ebp),%eax
 80483bb:       83 ec 1c                sub    $0x1c,%esp
 80483be:       8b 48 04                mov    0x4(%eax),%ecx
 80483c1:       51                      push   %ecx
 80483c2:       e8 1d ff ff ff          call   80482e4 
 80483c7:       83 c4 10                add    $0x10,%esp
 //             --------------------------------------------------
 //             If 'a' equal 2 (which is likely), we will continue
 //             without branching, so without flusing the pipeline. The
 //             jump only occurs when a != 2, which is unlikely.
 //             ---------------------------------------------------
 80483ca:       83 f8 02                cmp    $0x2,%eax
 80483cd:       75 13                   jne    80483e2 0x32>
 //             Here the a++ incrementation has been optimized by gcc
 80483cf:       b0 03                   mov    $0x3,%al
 //             Call printf()
 80483d1:       52                      push   %edx
 80483d2:       52                      push   %edx
 80483d3:       50                      push   %eax
 80483d4:       68 c8 84 04 08          push   $0x80484c8
 80483d9:       e8 f6 fe ff ff          call   80482d4 
 //             Return 0 and go out.
 80483de:       31 c0                   xor    %eax,%eax
 80483e0:       c9                      leave
 80483e1:       c3                      ret

How should I use it ?  
      You should use it only in cases when the likeliest branch is very very very likely, or when the unlikeliest branch is very very very unlikely.  

======  

看完最权威的，下面看下“民间”的说法：

======  
likely,unlikely宏与GCC内建函数__builtin_expect()  

在 GCC 手册中对 __builtin_expect() 的描述是这样的：  

      由于大部分程序员在分支预测方面做得很糟糕，所以 GCC 提供了这个内建函数来帮助程序员处理分支预测，优化程序。其第一个参数 exp 为一个整型表达式，这个内建函数的返回值也是这个 exp ，而 c 为一个编译期常量。这个函数的语义是：你期望 exp 表达式的值等于常量 c ，从而 GCC 为你优化程序，将符合这个条件的分支放在合适的地方。一般情况下，你也许会更喜欢使用 gcc 的一个参数 '-fprofile-arcs' 来收集程序运行的关于执行流程和分支走向的实际反馈信息。  
      因为这个程序只提供了整型表达式，所以如果你要优化其他类型的表达式，可以采用指针的形式。  

likely 和 unlikely 是 gcc 扩展的跟处理器相关的宏：   

#define  likely(x)        __builtin_expect(!!(x), 1) 
#define  unlikely(x)      __builtin_expect(!!(x), 0)

        现在处理器都是流水线的，有些里面有多个逻辑运算单元，系统可以提前取多条指令进行并行处理，但遇到跳转时，则需要重新取指令，这相对于不用重新去指令就降低了速度。   
      所以就引入了 likely 和 unlikely ，目的是增加条件分支预测的准确性，cpu 会提前装载后面的指令，遇到条件转移指令时会提前预测并装载某个分 支的指令。unlikely 表示你可以确认该条件是极少发生的，相反 likely 表示该条件多数情况下会发生。编译器会产生相应的代码来优化 cpu 执行效率。  

因此程序员在编写代码时可以根据判断条件发生的概率来优化处理器的取指操作。   
例如：  

int x, y; 
if(unlikely(x > 0)) 
    y = 1; 
else 
    y = -1;

      上面的代码中 gcc 编译的指令会预先读取 y = -1 这条指令，这适合 x 的值大于 0 的概率比较小的情况。    如果 x 的值在大部分情况下是大于 0 的，就应该用 likely(x > 0)，这样编译出的指令是预先读取 y = 1 这条指令了。这样系统在运行时就会减少重新取指了。  


======  
内核中的 likely() 与 unlikely()  

首先要明确：  

• if(likely(value)) 等价于 if(value)
• if(unlikely(value)) 也等价于 if(value)

        __builtin_expect() 是 GCC (version >= 2.96）提供给程序员使用的，目的是将“分支转移”的信息提供给编译器，这样编译器可以对代码进行优化，以减少指令跳转带来的性能下降。  

__builtin_expect((x),1) 表示 x 的值为真的可能性更大；  
__builtin_expect((x),0) 表示 x 的值为假的可能性更大。  

      也就是说，使用 likely() ，执行 if 后面的语句 的机会更大，使用 unlikely()，执行 else 后面的语句的机会更大。通过这种方式，编译器在编译过程中，会将可能性更大的代码紧跟着起面的代码，从而减少指令跳转带来的性能上的下降。  

======  

看完一圈别   人写的东西，自己也要输出点干货，列举 GLib-2.35.4 中    gmacros.h 代码如下：  

/*
 * The G_LIKELY and G_UNLIKELY macros let the programmer give hints to 
 * the compiler about the expected result of an expression. Some compilers
 * can use this information for optimizations.
 *
 * The _G_BOOLEAN_EXPR macro is intended to trigger a gcc warning when
 * putting assignments in g_return_if_fail ().  
 */
#if defined(__GNUC__) && (__GNUC__ > 2) && defined(__OPTIMIZE__)
#define _G_BOOLEAN_EXPR(expr)                   \
 G_GNUC_EXTENSION ({                            \
   int _g_boolean_var_;                         \
   if (expr)                                    \
      _g_boolean_var_ = 1;                      \
   else                                         \
      _g_boolean_var_ = 0;                      \
   _g_boolean_var_;                             \
})
// 为条件判断提供程序员期望的结果-- 用于编译器优化
#define G_LIKELY(expr) (__builtin_expect (_G_BOOLEAN_EXPR(expr), 1))
#define G_UNLIKELY(expr) (__builtin_expect (_G_BOOLEAN_EXPR(expr), 0))
#else
#define G_LIKELY(expr) (expr)
#define G_UNLIKELY(expr) (expr)
#endif

由上可以看出，   GLib 中使用    _G_BOOLEAN_EXPR(expr) 代替了 !!(expr)   。但功能上是一样的。