Hotspot in a for loop

问题

I am trying to optimize this code.

static
lvh_distance levenshtein_distance( const std::string & s1, const std::string & s2 )
{
    const size_t len1 = s1.size(), len2 = s2.size();
    std::vector<unsigned int> col( len2+1 ), prevCol( len2+1 );

    const size_t prevColSize = prevCol.size();
    for( unsigned int i = 0; i < prevColSize; i++ )
        prevCol[i] = i;

    for( unsigned int i = 0, j; i < len1; ++i )
    {
        col[0] = i+1;
        const char s1i = s1[i];
        for( j = 0; j < len2; ++j )
        {
            const auto minPrev = 1 + std::min( col[j], prevCol[1 + j] );
            col[j+1] = std::min( minPrev, prevCol[j] + (  s1i == s2[j] ? 0 : 1 ) );

        }
        col.swap( prevCol );
    }
    return prevCol[len2];
}

Intel VTune shows that roughly half of the processor time is spent in the second for instruction, not the 2 lines inside the for loop. When I unroll the assembly source, I can see that the for c++ instruction has been translated into a number of opcodes, 3 of which seem to be devouring the CPU time:

Code Location   Source Line Assembly    CPU Time
        Block 14:   [Unknown]
0x420c00    31  movq  (%r12), %rcx  19.969ms
0x420c04    30  add $0x1, %r11d [Unknown]
0x420c08    32  test %rbx, %rbx [Unknown]
0x420c0b    30  movl  %r11d, (%r8)  [Unknown]
0x420c0e    31  movzxb  (%rcx,%rdx,1), %r9d 19.964ms
0x420c13    32  jz 0x420c53 <Block 17>  [Unknown]
        Block 15:   [Unknown]
0x420c15    32  movq  (%rbp), %r10  [Unknown]
0x420c19    32  mov %r11d, %edx [Unknown]
0x420c1c    32  xor %ecx, %ecx  39.928ms
0x420c1e    32  xor %edi, %edi  [Unknown]
        Block 16:   [Unknown]
0x420c20    34  add $0x1, %edi  29.994ms
0x420c23    34  mov %edi, %esi  30.956ms
0x420c25    34  movl  (%rax,%rsi,4), %r15d  180.659ms
0x420c29    34  cmp %r15d, %edx 39.896ms
0x420c2c    34  cmovbe %edx, %r15d  19.951ms
0x420c30    35  xor %edx, %edx  460.772ms
0x420c32    34  add $0x1, %r15d 19.946ms
0x420c36    35  cmpb  (%r10,%rcx,1), %r9b   169.659ms  
0x420c3a    35  setnz %dl   49.815ms
0x420c3d    35  addl  (%rax,%rcx,4), %edx   [Unknown]
0x420c40    32  mov %rsi, %rcx               210.615ms  <------------------
0x420c43    32  cmp %edx, %r15d              29.936ms
0x420c46    32  cmovbe %r15d, %edx          29.938ms
0x420c4a    32  cmp %rsi, %rbx              558.298ms  <-------------------
0x420c4d    35  movl  %edx, (%r8,%rsi,4)    19.965ms
0x420c51    32  jnbe 0x420c20 <Block 16>    200.625ms  <-------------------

I don’t understand how a simple move and compare could be that time-consuming.

回答1:

Profiler cannot show you the exact most time-consuming instructions because all modern CPUs use out-of-order and speculative execution. It is not unusual to see the largest measured time one or two lines away from the most time-consuming instruction.

As expected, the most time consuming instructions here are cmovbe (implementing std::min). You can see the largest times near them: 460.772ms and 558.298ms. cmovbe are the most time consuming instructions because usually they have large latency and more dependencies on preceding instructions.

来源：https://stackoverflow.com/questions/16145027/hotspot-in-a-for-loop