Before voting on this answer, please test (and verify) this on your machine and comment/add the results. Note that I used a vector size of 1000*1000*1000 for my tests. Currently, this answer has 19 upvotes but only one posted results, and these results did not show the effect described below (though obtained with a different test code, see comments).
There seems to be an optimizer bug/artifact. Compare the times of:
template<typename _ForwardIterator, typename _Compare>
_ForwardIterator
my_max_element_orig(_ForwardIterator __first, _ForwardIterator __last,
_Compare __comp)
{
if (__first == __last) return __first;
_ForwardIterator __result = __first;
while(++__first != __last)
if (__comp(__result, __first))
__result = __first;
return __result;
}
template<typename _ForwardIterator, typename _Compare>
_ForwardIterator
my_max_element_changed(_ForwardIterator __first, _ForwardIterator __last,
_Compare __comp)
{
if (__first == __last) return __first;
_ForwardIterator __result = __first;
++__first;
for(; __first != __last; ++__first)
if (__comp(__result, __first))
__result = __first;
return __result;
}
The first is the original libstdc++ implementation, the second one should be a transformation without any changes in behaviour or requirements. Clang++ produces very similar run times for those two functions, whereas g++4.8.2 is four times faster with the second version.
Following Maxim's proposal, changing the vector from int
to int64_t
, the changed version is not 4, but only 1.7 times faster than the original version (g++4.8.2).
The difference is in predictive commoning of *result
, that is, storing the value of the current max element so that it does not have to be reloaded from memory each time. This gives a far cleaner cache access pattern:
w/o commoning with commoning
* *
** *
** *
** *
* * *
* * *
* * *
Here's the asm for comparison (rdi
/rsi
contain the first/last iterators respectively):
With the while loop (2.88743 ms; gist):
movq %rdi, %rax
jmp .L49
.L51:
movl (%rdi), %edx
cmpl %edx, (%rax)
cmovl %rdi, %rax
.L49:
addq $4, %rdi
cmpq %rsi, %rdi
jne .L51
With the for loop (1235.55 μs):
leaq 4(%rdi), %rdx
movq %rdi, %rax
cmpq %rsi, %rdx
je .L53
movl (%rdi), %ecx
.L54:
movl (%rdx), %r8d
cmpl %r8d, %ecx
cmovl %rdx, %rax
cmovl %r8d, %ecx
addq $4, %rdx
cmpq %rdx, %rsi
jne .L54
.L53:
If I force commoning by explicitly storing *result
into a variable prev
at the start and whenever result
is updated, and using prev
instead of *result
in the comparison, I get an even faster loop (377.601 μs):
movl (%rdi), %ecx
movq %rdi, %rax
.L57:
addq $4, %rdi
cmpq %rsi, %rdi
je .L60
.L59:
movl (%rdi), %edx
cmpl %edx, %ecx
jge .L57
movq %rdi, %rax
addq $4, %rdi
movl %edx, %ecx
cmpq %rsi, %rdi
jne .L59
.L60:
The reason this is faster than the for
loop is that the conditional moves (cmovl
) in the above are a pessimisation as they are executed so rarely (Linus says that cmov is only a good idea if the branch is unpredictable). Note that for randomly distributed data the branch is expected to be taken Hn times, which is a negligible proportion (Hn grows logarithmically, so Hn/n rapidly approaches 0). The conditional-move code will only be better on pathological data e.g. [1, 0, 3, 2, 5, 4, ...].