How could I implement logical implication with bitwise or other efficient code in C?

前端 未结 5 1523
没有蜡笔的小新
没有蜡笔的小新 2021-02-20 17:52

I want to implement a logical operation that works as efficient as possible. I need this truth table:

p    q    p → q
T    T      T
T    F      F
F    T      T
F         


        
5条回答
  •  既然无缘
    2021-02-20 18:48

    Another solution for C booleans (a bit dirty, but works):

    ((unsigned int)(p) <= (unsigned int)(q))

    It works since by the C standard, 0 represents false, and any other value true (1 is returned for true by boolean operators, int type).

    The "dirtiness" is that I use booleans (p and q) as integers, which contradicts some strong typing policies (such as MISRA), well, this is an optimization question. You may always #define it as a macro to hide the dirty stuff.

    For proper boolean p and q (having either 0 or 1 binary representations) it works. Otherwise T->T might fail to produce T if p and q have arbitrary nonzero values for representing true.

    If you need to store the result only, since the Pentium II, there is the cmovcc (Conditional Move) instruction (as shown in Derobert's answer). For booleans, however even the 386 had a branchless option, the setcc instruction, which produces 0 or 1 in a result byte location (byte register or memory). You can also see that in Derobert's answer, and this solution also compiles to a result involving a setcc (setbe: Set if below or equal).

    Derobert and Chris Dolan's ~p | q variant should be the fastest for processing large quantities of data since it can process the implication on all bits of p and q individually.

    Notice that not even the !p || q solution compiles to branching code on the x86: it uses setcc instructions. That's the best solution if p or q may contain arbitrary nonzero values representing true. If you use the _Bool type, it will generate very few instructions.

    I got the following figures when compiling for the x86:

    __attribute__((fastcall)) int imp1(int a, int b)
    {
     return ((unsigned int)(a) <= (unsigned int)(b));
    }
    
    __attribute__((fastcall)) int imp2(int a, int b)
    {
     return (!a || b);
    }
    
    __attribute__((fastcall)) _Bool imp3(_Bool a, _Bool b)
    {
     return (!a || b);
    }
    
    __attribute__((fastcall)) int imp4(int a, int b)
    {
     return (~a | b);
    }
    

    Assembly result:

    00000000 :
       0:   31 c0                   xor    %eax,%eax
       2:   39 d1                   cmp    %edx,%ecx
       4:   0f 96 c0                setbe  %al
       7:   c3                      ret    
    
    00000010 :
      10:   85 d2                   test   %edx,%edx
      12:   0f 95 c0                setne  %al
      15:   85 c9                   test   %ecx,%ecx
      17:   0f 94 c2                sete   %dl
      1a:   09 d0                   or     %edx,%eax
      1c:   0f b6 c0                movzbl %al,%eax
      1f:   c3                      ret    
    
    00000020 :
      20:   89 c8                   mov    %ecx,%eax
      22:   83 f0 01                xor    $0x1,%eax
      25:   09 d0                   or     %edx,%eax
      27:   c3                      ret    
    
    00000030 :
      30:   89 d0                   mov    %edx,%eax
      32:   f7 d1                   not    %ecx
      34:   09 c8                   or     %ecx,%eax
      36:   c3                      ret    
    

    When using the _Bool type, the compiler clearly exploits that it only has two possible values (0 for false and 1 for true), producing a very similar result to the ~a | b solution (the only difference being that the latter performs a complement on all bits instead of just the lowest bit).

    Compiling for 64 bits gives just about the same results.

    Anyway, it is clear, the method doesn't really matter from the point of avoiding producing conditionals.

提交回复
热议问题