I\'ve been digging through some parts of the Linux kernel, and found calls like this:
if (unlikely(fd < 0))
{
/* Do something */
}
o
Let's decompile to see what GCC 4.8 does with it
Without __builtin_expect
#include "stdio.h"
#include "time.h"
int main() {
/* Use time to prevent it from being optimized away. */
int i = !time(NULL);
if (i)
printf("%d\n", i);
puts("a");
return 0;
}
Compile and decompile with GCC 4.8.2 x86_64 Linux:
gcc -c -O3 -std=gnu11 main.c
objdump -dr main.o
Output:
0000000000000000 <main>:
0: 48 83 ec 08 sub $0x8,%rsp
4: 31 ff xor %edi,%edi
6: e8 00 00 00 00 callq b <main+0xb>
7: R_X86_64_PC32 time-0x4
b: 48 85 c0 test %rax,%rax
e: 75 14 jne 24 <main+0x24>
10: ba 01 00 00 00 mov $0x1,%edx
15: be 00 00 00 00 mov $0x0,%esi
16: R_X86_64_32 .rodata.str1.1
1a: bf 01 00 00 00 mov $0x1,%edi
1f: e8 00 00 00 00 callq 24 <main+0x24>
20: R_X86_64_PC32 __printf_chk-0x4
24: bf 00 00 00 00 mov $0x0,%edi
25: R_X86_64_32 .rodata.str1.1+0x4
29: e8 00 00 00 00 callq 2e <main+0x2e>
2a: R_X86_64_PC32 puts-0x4
2e: 31 c0 xor %eax,%eax
30: 48 83 c4 08 add $0x8,%rsp
34: c3 retq
The instruction order in memory was unchanged: first the printf
and then puts
and the retq
return.
With __builtin_expect
Now replace if (i)
with:
if (__builtin_expect(i, 0))
and we get:
0000000000000000 <main>:
0: 48 83 ec 08 sub $0x8,%rsp
4: 31 ff xor %edi,%edi
6: e8 00 00 00 00 callq b <main+0xb>
7: R_X86_64_PC32 time-0x4
b: 48 85 c0 test %rax,%rax
e: 74 11 je 21 <main+0x21>
10: bf 00 00 00 00 mov $0x0,%edi
11: R_X86_64_32 .rodata.str1.1+0x4
15: e8 00 00 00 00 callq 1a <main+0x1a>
16: R_X86_64_PC32 puts-0x4
1a: 31 c0 xor %eax,%eax
1c: 48 83 c4 08 add $0x8,%rsp
20: c3 retq
21: ba 01 00 00 00 mov $0x1,%edx
26: be 00 00 00 00 mov $0x0,%esi
27: R_X86_64_32 .rodata.str1.1
2b: bf 01 00 00 00 mov $0x1,%edi
30: e8 00 00 00 00 callq 35 <main+0x35>
31: R_X86_64_PC32 __printf_chk-0x4
35: eb d9 jmp 10 <main+0x10>
The printf
(compiled to __printf_chk
) was moved to the very end of the function, after puts
and the return to improve branch prediction as mentioned by other answers.
So it is basically the same as:
int main() {
int i = !time(NULL);
if (i)
goto printf;
puts:
puts("a");
return 0;
printf:
printf("%d\n", i);
goto puts;
}
This optimization was not done with -O0
.
But good luck on writing an example that runs faster with __builtin_expect
than without, CPUs are really smart these days. My naive attempts are here.
C++20 [[likely]]
and [[unlikely]]
C++20 has standardized those C++ built-ins: How to use C++20's likely/unlikely attribute in if-else statement They will likely (a pun!) do the same thing.
(general comment - other answers cover the details)
There's no reason that you should lose portability by using them.
You always have the option of creating a simple nil-effect "inline" or macro that will allow you to compile on other platforms with other compilers.
You just won't get the benefit of the optimization if you're on other platforms.
They're hints to the compiler to generate the hint prefixes on branches. On x86/x64, they take up one byte, so you'll get at most a one-byte increase for each branch. As for performance, it entirely depends on the application -- in most cases, the branch predictor on the processor will ignore them, these days.
Edit: Forgot about one place they can actually really help with. It can allow the compiler to reorder the control-flow graph to reduce the number of branches taken for the 'likely' path. This can have a marked improvement in loops where you're checking multiple exit cases.
In many linux release, you can find complier.h in /usr/linux/ , you can include it for use simply. And another opinion, unlikely() is more useful rather than likely(), because
if ( likely( ... ) ) {
doSomething();
}
it can be optimized as well in many compiler.
And by the way, if you want to observe the detail behavior of the code, you can do simply as follow:
gcc -c test.c objdump -d test.o > obj.s
Then, open obj.s, you can find the answer.
They are hint to the compiler to emit instructions that will cause branch prediction to favour the "likely" side of a jump instruction. This can be a big win, if the prediction is correct it means that the jump instruction is basically free and will take zero cycles. On the other hand if the prediction is wrong, then it means the processor pipeline needs to be flushed and it can cost several cycles. So long as the prediction is correct most of the time, this will tend to be good for performance.
Like all such performance optimisations you should only do it after extensive profiling to ensure the code really is in a bottleneck, and probably given the micro nature, that it is being run in a tight loop. Generally the Linux developers are pretty experienced so I would imagine they would have done that. They don't really care too much about portability as they only target gcc, and they have a very close idea of the assembly they want it to generate.
These are macros that give hints to the compiler about which way a branch may go. The macros expand to GCC specific extensions, if they're available.
GCC uses these to to optimize for branch prediction. For example, if you have something like the following
if (unlikely(x)) {
dosomething();
}
return x;
Then it can restructure this code to be something more like:
if (!x) {
return x;
}
dosomething();
return x;
The benefit of this is that when the processor takes a branch the first time, there is significant overhead, because it may have been speculatively loading and executing code further ahead. When it determines it will take the branch, then it has to invalidate that, and start at the branch target.
Most modern processors now have some sort of branch prediction, but that only assists when you've been through the branch before, and the branch is still in the branch prediction cache.
There are a number of other strategies that the compiler and processor can use in these scenarios. You can find more details on how branch predictors work at Wikipedia: http://en.wikipedia.org/wiki/Branch_predictor