代码 alignment 显着影响性能

Question

Today I have found sample code which slowed down by 50%, after adding some unrelated code. After debugging I have figured out the problem was in the loop alignment. Depending of the loop code placement there is different execution time eg:

Address	Time[us]
00007FF780A01270	980us
00007FF7750B1280	1500us
00007FF7750B1290	986us
00007FF7750B12A0	1500us

I didn't expect previously that code alignment may have such a big impact. And I thought my compiler is smart enough to align the code correctly.

What exactly cause such a big difference in execution time? (I suppose some processor architecture details).

The test program I have compiled in Release mode with Visual Studio 2019 and run it on Windows 10. I have checked the program on 2 processors: i7-8700k (the results above), and on intel i5-3570k but the problem does not exist there and the execution time is always about 1250us. I have also tried to compile the program with clang, but with clang the result is always ~1500us (on i7-8700k).

My test program:

#include <chrono>
#include <iostream>
#include <intrin.h>
using namespace std;

template<int N>
__forceinline void noops()
{
    __nop(); __nop(); __nop(); __nop(); __nop(); __nop(); __nop(); __nop(); __nop(); __nop(); __nop(); __nop(); __nop(); __nop(); __nop(); __nop();
    noops<N - 1>();
}
template<>
__forceinline void noops<0>(){}

template<int OFFSET>
__declspec(noinline) void SumHorizontalLine(const unsigned char* __restrict src, int width, int a, unsigned short* __restrict dst)
{
    unsigned short sum = 0;
    const unsigned char* srcP1 = src - a - 1;
    const unsigned char* srcP2 = src + a;

    //some dummy loop,just a few iterations
    for (int i = 0; i < a; ++i)
        dst[i] = src[i] / (double)dst[i];

    noops<OFFSET>();
    //the important loop
    for (int x = a + 1; x < width - a; x++)
    {
        unsigned char v1 = srcP1[x];
        unsigned char v2 = srcP2[x];
        sum -= v1;
        sum += v2;
        dst[x] = sum;
    }

}

template<int OFFSET>
void RunTest(unsigned char* __restrict src, int width, int a, unsigned short* __restrict dst)
{
    double minTime = 99999999;
    for(int i = 0; i < 20; ++i)
    {
        auto start = chrono::steady_clock::now();

        for (int i = 0; i < 1024; ++i)
        {
            SumHorizontalLine<OFFSET>(src, width, a, dst);
        }

        auto end = chrono::steady_clock::now();
        auto us = chrono::duration_cast<chrono::microseconds>(end - start).count();
        if (us < minTime)
        {
            minTime = us;
        }
    }

    cout << OFFSET << " : " << minTime << " us" << endl;
}

int main()
{
    const int width = 2048;
    const int x = 3;
    unsigned char* src = new unsigned char[width * 5];
    unsigned short* dst = new unsigned short[width];
    memset(src, 0, sizeof(unsigned char) * width);
    memset(dst, 0, sizeof(unsigned short) * width);

    while(true)
    RunTest<1>(src, width, x, dst);
}

To verify different alignment, just recompile the program and change RunTest<0> to RunTest<1> etc. Compiler always align the code to 16bytes. In my test code I just insert additional nops to move the code a bit more.

Assembly code generated for the loop with OFFSET=1 (for other offset only the amount of npads is different):

  0007c 90       npad    1
  0007d 90       npad    1
  0007e 49 83 c1 08  add     r9, 8
  00082 90       npad    1
  00083 90       npad    1
  00084 90       npad    1
  00085 90       npad    1
  00086 90       npad    1
  00087 90       npad    1
  00088 90       npad    1
  00089 90       npad    1
  0008a 90       npad    1
  0008b 90       npad    1
  0008c 90       npad    1
  0008d 90       npad    1
  0008e 90       npad    1
  0008f 90       npad    1
$LL15@SumHorizon:

; 25   : 
; 26   :    noops<OFFSET>();
; 27   : 
; 28   :    for (int x = a + 1; x < width - a; x++)
; 29   :    {
; 30   :        unsigned char v1 = srcP1[x];
; 31   :        unsigned char v2 = srcP2[x];
; 32   :        sum -= v1;

  00090 0f b6 42 f9  movzx   eax, BYTE PTR [rdx-7]
  00094 4d 8d 49 02  lea     r9, QWORD PTR [r9+2]

; 33   :        sum += v2;

  00098 0f b6 0a     movzx   ecx, BYTE PTR [rdx]
  0009b 48 8d 52 01  lea     rdx, QWORD PTR [rdx+1]
  0009f 66 2b c8     sub     cx, ax
  000a2 66 44 03 c1  add     r8w, cx

; 34   :        dst[x] = sum;

  000a6 66 45 89 41 fe   mov     WORD PTR [r9-2], r8w
  000ab 49 83 ea 01  sub     r10, 1
  000af 75 df        jne     SHORT $LL15@SumHorizon

; 35   :    }
; 36   : 
; 37   : }

  000b1 c3       ret     0
??$SumHorizontalLine@$00@@YAXPEIBEHHPEIAG@Z ENDP    ; SumHorizont

Answer 1

In the slow cases (ie, 00007FF7750B1280 and 00007FF7750B12A0), the jne instruction crosses a 32-byte boundary. The mitigations for the "Jump Conditional Code" (JCC) erratum ( https://www.intel.com/content/dam/support/us/en/documents/processors/mitigations-jump-conditional-code-erratum.pdf ) prevent such instructions from being cached in the DSB. The JCC erratum only applies to Skylake-based CPUs, which is why the effect does not occur on your i5-3570k CPU.

As Peter Cordes pointed out in a comment, recent compilers have options that try to mitigate this effect. Intel JCC Erratum - should JCC really be treated separately? mentions MSVC's /QIntel-jcc-erratum option; another related question is How can I mitigate the impact of the Intel jcc erratum on gcc?

Answer 2

I thought my compiler is smart enough to align the code correctly.

As you said, the compiler is always aligning things to a multiple of 16 bytes. This probably does account for the direct effects of alignment. But there are limits to the "smartness" of the compiler.

Besides alignment, code placement has indirect performance effects as well, because of cache associativity. If there is too much contention for the few cache lines that can map to this address, performance will suffer. Moving to an address with less contention makes the problem go away.

The compiler may be smart enough to handle cache contention effects as well, but only IF you turn on profile-guided optimization. The interactions are far too complex to predict in a reasonable amount of work; it is much easier to watch for cache conflicts by actually running the program and that's what PGO does.