std::vector performance regression when enabling C++11
C++PerformanceGccC++11VectorC++ Problem Overview
I have found an interesting performance regression in a small C++ snippet, when I enable C++11:
#include <vector>
struct Item
{
int a;
int b;
};
int main()
{
const std::size_t num_items = 10000000;
std::vector<Item> container;
container.reserve(num_items);
for (std::size_t i = 0; i < num_items; ++i) {
container.push_back(Item());
}
return 0;
}
With g++ (GCC) 4.8.2 20131219 (prerelease) and C++03 I get:
milian:/tmp$ g++ -O3 main.cpp && perf stat -r 10 ./a.out
Performance counter stats for './a.out' (10 runs):
35.206824 task-clock # 0.988 CPUs utilized ( +- 1.23% )
4 context-switches # 0.116 K/sec ( +- 4.38% )
0 cpu-migrations # 0.006 K/sec ( +- 66.67% )
849 page-faults # 0.024 M/sec ( +- 6.02% )
95,693,808 cycles # 2.718 GHz ( +- 1.14% ) [49.72%]
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
95,282,359 instructions # 1.00 insns per cycle ( +- 0.65% ) [75.27%]
30,104,021 branches # 855.062 M/sec ( +- 0.87% ) [77.46%]
6,038 branch-misses # 0.02% of all branches ( +- 25.73% ) [75.53%]
0.035648729 seconds time elapsed ( +- 1.22% )
With C++11 enabled on the other hand, the performance degrades significantly:
milian:/tmp$ g++ -std=c++11 -O3 main.cpp && perf stat -r 10 ./a.out
Performance counter stats for './a.out' (10 runs):
86.485313 task-clock # 0.994 CPUs utilized ( +- 0.50% )
9 context-switches # 0.104 K/sec ( +- 1.66% )
2 cpu-migrations # 0.017 K/sec ( +- 26.76% )
798 page-faults # 0.009 M/sec ( +- 8.54% )
237,982,690 cycles # 2.752 GHz ( +- 0.41% ) [51.32%]
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
135,730,319 instructions # 0.57 insns per cycle ( +- 0.32% ) [75.77%]
30,880,156 branches # 357.057 M/sec ( +- 0.25% ) [75.76%]
4,188 branch-misses # 0.01% of all branches ( +- 7.59% ) [74.08%]
0.087016724 seconds time elapsed ( +- 0.50% )
Can someone explain this? So far my experience was that the STL gets faster by enabling C++11, esp. thanks to move semantics.
EDIT: As suggested, using container.emplace_back(); instead the performance gets on par with the C++03 version. How can the C++03 version achieve the same for push_back?
milian:/tmp$ g++ -std=c++11 -O3 main.cpp && perf stat -r 10 ./a.out
Performance counter stats for './a.out' (10 runs):
36.229348 task-clock # 0.988 CPUs utilized ( +- 0.81% )
4 context-switches # 0.116 K/sec ( +- 3.17% )
1 cpu-migrations # 0.017 K/sec ( +- 36.85% )
798 page-faults # 0.022 M/sec ( +- 8.54% )
94,488,818 cycles # 2.608 GHz ( +- 1.11% ) [50.44%]
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
94,851,411 instructions # 1.00 insns per cycle ( +- 0.98% ) [75.22%]
30,468,562 branches # 840.991 M/sec ( +- 1.07% ) [76.71%]
2,723 branch-misses # 0.01% of all branches ( +- 9.84% ) [74.81%]
0.036678068 seconds time elapsed ( +- 0.80% )
C++ Solutions
Solution 1 - C++
I can reproduce your results on my machine with those options you write in your post.
However, if I also enable link time optimization (I also pass the -flto flag to gcc 4.7.2), the results are identical:
(I am compiling your original code, with container.push_back(Item());)
$ g++ -std=c++11 -O3 -flto regr.cpp && perf stat -r 10 ./a.out
Performance counter stats for './a.out' (10 runs):
35.426793 task-clock # 0.986 CPUs utilized ( +- 1.75% )
4 context-switches # 0.116 K/sec ( +- 5.69% )
0 CPU-migrations # 0.006 K/sec ( +- 66.67% )
19,801 page-faults # 0.559 M/sec
99,028,466 cycles # 2.795 GHz ( +- 1.89% ) [77.53%]
50,721,061 stalled-cycles-frontend # 51.22% frontend cycles idle ( +- 3.74% ) [79.47%]
25,585,331 stalled-cycles-backend # 25.84% backend cycles idle ( +- 4.90% ) [73.07%]
141,947,224 instructions # 1.43 insns per cycle
# 0.36 stalled cycles per insn ( +- 0.52% ) [88.72%]
37,697,368 branches # 1064.092 M/sec ( +- 0.52% ) [88.75%]
26,700 branch-misses # 0.07% of all branches ( +- 3.91% ) [83.64%]
0.035943226 seconds time elapsed ( +- 1.79% )
$ g++ -std=c++98 -O3 -flto regr.cpp && perf stat -r 10 ./a.out
Performance counter stats for './a.out' (10 runs):
35.510495 task-clock # 0.988 CPUs utilized ( +- 2.54% )
4 context-switches # 0.101 K/sec ( +- 7.41% )
0 CPU-migrations # 0.003 K/sec ( +-100.00% )
19,801 page-faults # 0.558 M/sec ( +- 0.00% )
98,463,570 cycles # 2.773 GHz ( +- 1.09% ) [77.71%]
50,079,978 stalled-cycles-frontend # 50.86% frontend cycles idle ( +- 2.20% ) [79.41%]
26,270,699 stalled-cycles-backend # 26.68% backend cycles idle ( +- 8.91% ) [74.43%]
141,427,211 instructions # 1.44 insns per cycle
# 0.35 stalled cycles per insn ( +- 0.23% ) [87.66%]
37,366,375 branches # 1052.263 M/sec ( +- 0.48% ) [88.61%]
26,621 branch-misses # 0.07% of all branches ( +- 5.28% ) [83.26%]
0.035953916 seconds time elapsed
As for the reasons, one needs to look at the generated assembly code (g++ -std=c++11 -O3 -S regr.cpp). In C++11 mode the generated code is significantly more cluttered than for C++98 mode and inlining the function
void std::vector<Item,std::allocator<Item>>::_M_emplace_back_aux<Item>(Item&&)
fails in C++11 mode with the default inline-limit.
This failed inline has a domino effect. Not because this function is being called
(it is not even called!) but because we have to be prepared: If it is called,
the function argments (Item.a and Item.b) must already be at the right place. This leads to
a pretty messy code.
Here is the relevant part of the generated code for the case where inlining succeeds:
.L42:
testq %rbx, %rbx # container$D13376$_M_impl$_M_finish
je .L3 #,
movl $0, (%rbx) #, container$D13376$_M_impl$_M_finish_136->a
movl $0, 4(%rbx) #, container$D13376$_M_impl$_M_finish_136->b
.L3:
addq $8, %rbx #, container$D13376$_M_impl$_M_finish
subq $1, %rbp #, ivtmp.106
je .L41 #,
.L14:
cmpq %rbx, %rdx # container$D13376$_M_impl$_M_finish, container$D13376$_M_impl$_M_end_of_storage
jne .L42 #,
This is a nice and compact for loop. Now, let's compare this to that of the failed inline case:
.L49:
testq %rax, %rax # D.15772
je .L26 #,
movq 16(%rsp), %rdx # D.13379, D.13379
movq %rdx, (%rax) # D.13379, *D.15772_60
.L26:
addq $8, %rax #, tmp75
subq $1, %rbx #, ivtmp.117
movq %rax, 40(%rsp) # tmp75, container.D.13376._M_impl._M_finish
je .L48 #,
.L28:
movq 40(%rsp), %rax # container.D.13376._M_impl._M_finish, D.15772
cmpq 48(%rsp), %rax # container.D.13376._M_impl._M_end_of_storage, D.15772
movl $0, 16(%rsp) #, D.13379.a
movl $0, 20(%rsp) #, D.13379.b
jne .L49 #,
leaq 16(%rsp), %rsi #,
leaq 32(%rsp), %rdi #,
call _ZNSt6vectorI4ItemSaIS0_EE19_M_emplace_back_auxIIS0_EEEvDpOT_ #
This code is cluttered and there is a lot more going on in the loop than in the previous case.
Before the function call (last line shown), the arguments must be placed appropriately:
leaq 16(%rsp), %rsi #,
leaq 32(%rsp), %rdi #,
call _ZNSt6vectorI4ItemSaIS0_EE19_M_emplace_back_auxIIS0_EEEvDpOT_ #
Even though this is never actually executed, the loop arranges the things before:
movl $0, 16(%rsp) #, D.13379.a
movl $0, 20(%rsp) #, D.13379.b
This leads to the messy code. If there is no function call because inlining succeeds,
we have only 2 move instructions in the loop and there is no messing going with the %rsp (stack pointer). However, if the inlining fails, we get 6 moves and we mess a lot with the %rsp.
Just to substantiate my theory (note the -finline-limit), both in C++11 mode:
$ g++ -std=c++11 -O3 -finline-limit=105 regr.cpp && perf stat -r 10 ./a.out
Performance counter stats for './a.out' (10 runs):
84.739057 task-clock # 0.993 CPUs utilized ( +- 1.34% )
8 context-switches # 0.096 K/sec ( +- 2.22% )
1 CPU-migrations # 0.009 K/sec ( +- 64.01% )
19,801 page-faults # 0.234 M/sec
266,809,312 cycles # 3.149 GHz ( +- 0.58% ) [81.20%]
206,804,948 stalled-cycles-frontend # 77.51% frontend cycles idle ( +- 0.91% ) [81.25%]
129,078,683 stalled-cycles-backend # 48.38% backend cycles idle ( +- 1.37% ) [69.49%]
183,130,306 instructions # 0.69 insns per cycle
# 1.13 stalled cycles per insn ( +- 0.85% ) [85.35%]
38,759,720 branches # 457.401 M/sec ( +- 0.29% ) [85.43%]
24,527 branch-misses # 0.06% of all branches ( +- 2.66% ) [83.52%]
0.085359326 seconds time elapsed ( +- 1.31% )
$ g++ -std=c++11 -O3 -finline-limit=106 regr.cpp && perf stat -r 10 ./a.out
Performance counter stats for './a.out' (10 runs):
37.790325 task-clock # 0.990 CPUs utilized ( +- 2.06% )
4 context-switches # 0.098 K/sec ( +- 5.77% )
0 CPU-migrations # 0.011 K/sec ( +- 55.28% )
19,801 page-faults # 0.524 M/sec
104,699,973 cycles # 2.771 GHz ( +- 2.04% ) [78.91%]
58,023,151 stalled-cycles-frontend # 55.42% frontend cycles idle ( +- 4.03% ) [78.88%]
30,572,036 stalled-cycles-backend # 29.20% backend cycles idle ( +- 5.31% ) [71.40%]
140,669,773 instructions # 1.34 insns per cycle
# 0.41 stalled cycles per insn ( +- 1.40% ) [88.14%]
38,117,067 branches # 1008.646 M/sec ( +- 0.65% ) [89.38%]
27,519 branch-misses # 0.07% of all branches ( +- 4.01% ) [86.16%]
0.038187580 seconds time elapsed ( +- 2.05% )
Indeed, if we ask the compiler to try just a little bit harder to inline that function, the difference in performance goes away.
So what is the take away from this story? That failed inlines can cost you a lot and you should make full use of the compiler capabilities: I can only recommend link time optimization. It gave a significant performance boost to my programs (up to 2.5x) and all I needed to do is to pass the -flto flag. That's a pretty good deal! ;)
However, I do not recommend trashing your code with the inline keyword; let the compiler decide what to do. (The optimizer is allowed to treat the inline keyword as white space anyway.)
Great question, +1!