I recently came across a strange deoptimization (or rather missed optimization opportunity).
Consider this function for efficient unpacking of arrays of 3-bit integers to 8-bit integers. It unpacks 16 ints in each loop iteration:
void unpack3bit(uint8_t* target, char* source, int size) {
while(size > 0){
uint64_t t = *reinterpret_cast<uint64_t*>(source);
target[0] = t & 0x7;
target[1] = (t >> 3) & 0x7;
target[2] = (t >> 6) & 0x7;
target[3] = (t >> 9) & 0x7;
target[4] = (t >> 12) & 0x7;
target[5] = (t >> 15) & 0x7;
target[6] = (t >> 18) & 0x7;
target[7] = (t >> 21) & 0x7;
target[8] = (t >> 24) & 0x7;
target[9] = (t >> 27) & 0x7;
target[10] = (t >> 30) & 0x7;
target[11] = (t >> 33) & 0x7;
target[12] = (t >> 36) & 0x7;
target[13] = (t >> 39) & 0x7;
target[14] = (t >> 42) & 0x7;
target[15] = (t >> 45) & 0x7;
source+=6;
size-=6;
target+=16;
}
}
Here is the generated assembly for parts of the code:
...
367: 48 89 c1 mov rcx,rax
36a: 48 c1 e9 09 shr rcx,0x9
36e: 83 e1 07 and ecx,0x7
371: 48 89 4f 18 mov QWORD PTR [rdi+0x18],rcx
375: 48 89 c1 mov rcx,rax
378: 48 c1 e9 0c shr rcx,0xc
37c: 83 e1 07 and ecx,0x7
37f: 48 89 4f 20 mov QWORD PTR [rdi+0x20],rcx
383: 48 89 c1 mov rcx,rax
386: 48 c1 e9 0f shr rcx,0xf
38a: 83 e1 07 and ecx,0x7
38d: 48 89 4f 28 mov QWORD PTR [rdi+0x28],rcx
391: 48 89 c1 mov rcx,rax
394: 48 c1 e9 12 shr rcx,0x12
398: 83 e1 07 and ecx,0x7
39b: 48 89 4f 30 mov QWORD PTR [rdi+0x30],rcx
...
It looks quite efficent. Simply a shift right
followed by an and
, and then a store
to the target
buffer. But now, look what happens when I change the function to a method in a struct:
struct T{
uint8_t* target;
char* source;
void unpack3bit( int size);
};
void T::unpack3bit(int size) {
while(size > 0){
uint64_t t = *reinterpret_cast<uint64_t*>(source);
target[0] = t & 0x7;
target[1] = (t >> 3) & 0x7;
target[2] = (t >> 6) & 0x7;
target[3] = (t >> 9) & 0x7;
target[4] = (t >> 12) & 0x7;
target[5] = (t >> 15) & 0x7;
target[6] = (t >> 18) & 0x7;
target[7] = (t >> 21) & 0x7;
target[8] = (t >> 24) & 0x7;
target[9] = (t >> 27) & 0x7;
target[10] = (t >> 30) & 0x7;
target[11] = (t >> 33) & 0x7;
target[12] = (t >> 36) & 0x7;
target[13] = (t >> 39) & 0x7;
target[14] = (t >> 42) & 0x7;
target[15] = (t >> 45) & 0x7;
source+=6;
size-=6;
target+=16;
}
}
I thought the generated assembly should be quite the same, but it isn't. Here is a part of it:
...
2b3: 48 c1 e9 15 shr rcx,0x15
2b7: 83 e1 07 and ecx,0x7
2ba: 88 4a 07 mov BYTE PTR [rdx+0x7],cl
2bd: 48 89 c1 mov rcx,rax
2c0: 48 8b 17 mov rdx,QWORD PTR [rdi] // Load, BAD!
2c3: 48 c1 e9 18 shr rcx,0x18
2c7: 83 e1 07 and ecx,0x7
2ca: 88 4a 08 mov BYTE PTR [rdx+0x8],cl
2cd: 48 89 c1 mov rcx,rax
2d0: 48 8b 17 mov rdx,QWORD PTR [rdi] // Load, BAD!
2d3: 48 c1 e9 1b shr rcx,0x1b
2d7: 83 e1 07 and ecx,0x7
2da: 88 4a 09 mov BYTE PTR [rdx+0x9],cl
2dd: 48 89 c1 mov rcx,rax
2e0: 48 8b 17 mov rdx,QWORD PTR [rdi] // Load, BAD!
2e3: 48 c1 e9 1e shr rcx,0x1e
2e7: 83 e1 07 and ecx,0x7
2ea: 88 4a 0a mov BYTE PTR [rdx+0xa],cl
2ed: 48 89 c1 mov rcx,rax
2f0: 48 8b 17 mov rdx,QWORD PTR [rdi] // Load, BAD!
...
As you see, we introduced an additional redundant load
from memory before each shift (mov rdx,QWORD PTR [rdi]
). It seems like the target
pointer (which is now a member instead of a local variable) has to be always reloaded before storing into it. This slows down the code considerably (around 15% in my measurements).
First I thought maybe the C++ memory model enforces that a member pointer may not be stored in a register but has to be reloaded, but this seemed like an awkward choice, as it would make a lot of viable optimizations impossible. So I was very surprised that the compiler did not store target
in a register here.
I tried caching the member pointer myself into a local variable:
void T::unpack3bit(int size) {
while(size > 0){
uint64_t t = *reinterpret_cast<uint64_t*>(source);
uint8_t* target = this->target; // << ptr cached in local variable
target[0] = t & 0x7;
target[1] = (t >> 3) & 0x7;
target[2] = (t >> 6) & 0x7;
target[3] = (t >> 9) & 0x7;
target[4] = (t >> 12) & 0x7;
target[5] = (t >> 15) & 0x7;
target[6] = (t >> 18) & 0x7;
target[7] = (t >> 21) & 0x7;
target[8] = (t >> 24) & 0x7;
target[9] = (t >> 27) & 0x7;
target[10] = (t >> 30) & 0x7;
target[11] = (t >> 33) & 0x7;
target[12] = (t >> 36) & 0x7;
target[13] = (t >> 39) & 0x7;
target[14] = (t >> 42) & 0x7;
target[15] = (t >> 45) & 0x7;
source+=6;
size-=6;
this->target+=16;
}
}
This code also yields the "good" assembler without additional stores. So my guess is: The compiler is not allowed to hoist the load of a member pointer of a struct, so such a "hot pointer" should always be stored in a local variable.
- So, why is the compiler unable to optimize out these loads?
- Is it the C++ memory model that forbids this? Or is it simply a shortcoming of my compiler?
- Is my guess correct or what is the exact reason why the optimization can't be performed?
The compiler in use was g++ 4.8.2-19ubuntu1
with -O3
optimization. I also tried clang++ 3.4-1ubuntu3
with similar results: Clang is even able to vectorize the method with the local target
pointer. However, using the this->target
pointer yields the same result: An extra load of the pointer before each store.
I checked the assembler of some similar methods and the result is the same: It seems that a member of this
always has to be reloaded before a store, even if such a load could simply be hoisted outside the loop. I will have to rewrite a lot of code to get rid of these additional stores, mainly by caching the pointer myself into a local variable that is declared above the hot code. But I always thought fiddling with such details as caching a pointer in a local variable would surely qualify for premature optimization in these days where compilers have gotten so clever. But it seems I am wrong here. Caching a member pointer in a hot loop seems to be a necessary manual optimization technique.