Which is faster in c++11: max(a-b, b-a) or abs(a-b)?

In c++11, which one is faster: max(a-b, b-a) or abs(a-b)?

You can use godbolt.org

#include <algorithm>
#include <cstdlib>

int abs1(int a, int b) {
   return abs(a-b);
}

int abs2(int a, int b) {
  return std::max(a-b, b-a);
}

int abs3(int a, int b) {
 return a > b ? a - b : b - a;
}

int abs4(int a, int b) {
 return a == b ? a : (a > b ? a - b : b - a);
}

Result for clang-3.7 x86 with -O4:

abs1(int, int):                              # @abs1(int, int)
    subl    %esi, %edi
    movl    %edi, %eax
    negl    %eax
    cmovll  %edi, %eax
    retq

abs2(int, int):                              # @abs2(int, int)
    movl    %edi, %eax
    subl    %esi, %eax
    subl    %edi, %esi
    cmpl    %esi, %eax
    cmovgel %eax, %esi
    movl    %esi, %eax
    retq

abs3(int, int):                              # @abs3(int, int)
    movl    %esi, %eax
    subl    %edi, %eax
    subl    %esi, %edi
    cmovlel %eax, %edi
    movl    %edi, %eax
    retq

abs4(int, int):                              # @abs4(int, int)
    movl    %edi, %eax
    subl    %esi, %eax
    jne .LBB3_2
    movl    %edi, %eax
    retq
.LBB3_2:
    jg  .LBB3_4
    subl    %edi, %esi
    movl    %esi, %eax
.LBB3_4:
    retq

Obviously abs from cstdlib will be faster for random numbers.