Performance improvement strategies for VM / interpreter?

问题

I have written a simple VM in C, using a simple switch of instructions, without any instruction decoding whatsoever, but performance is terrible.

For simple aritmetic operations the VM is about 4000 times slower than native C code for the same operations. I tested with a group of arrays of length 10 million, the first consisting of the program instructions, random + - * / operations, 2 arrays holding random integers and the third array being the operation target storage.

I was expecting to see 3-4 times drop in arithmetic performance, so that `4000x really blew me away. Even the slowest interpreted languages seem to offer higher performance. So where I am going wrong with my approach and how can I improve performance without resorting to JIT compilation to machine code?

The implementation is... basically the simplest I could come up with:

begin:
    {
        switch (*(op+(c++)))
        {
        case 0:
            add(in1+c, in2+c, out+c); goto begin;

        case 1:
            sub(in1+c, in2+c, out+c); goto begin;

        case 2:
            mul(in1+c, in2+c, out+c); goto begin;

        case 3:
            div(in1+c, in2+c, out+c); goto begin;

        case 4:
            cout << "end of program" << endl;
            goto end;

        default:
            cout << "ERROR!!!" << endl;

        }
    }

end:

UPDATE: I was toying with the length of the program when I noticed the QElapsedTimer I was using to profile was actually broken. Now I am using the clock() function from and according to it the computed goto is actually running on par with the native code, maybe a tad lower. Is that result legit??? Here is the full source (it is ugly I know, it's just for testing after all):

#include <QtGlobal>
#include <iostream>
#include <stdio.h>
#include <ctime>

using namespace std;

#define LENGTH 70000000

void add(int & a, int & b, int & r) {r = a * b;}
void sub(int & a, int & b, int & r) {r = a - b;}
void mul(int & a, int & b, int & r) {r = a * b;}
void div(int & a, int & b, int & r) {r = a / b;}

int main()
{
    char * op = new char[LENGTH];
    int * in1 = new int[LENGTH];
    int * in2 = new int[LENGTH];
    int * out = new int[LENGTH];

    for (int i = 0; i < LENGTH; ++i)
    {
        *(op+i) = i % 4;
        *(in1+i) = qrand();
        *(in2+i) = qrand()+1;
    }

    *(op+LENGTH-1) = 4; // end of program


    long long  sClock, fClock;


    unsigned int c = 0;
    sClock = clock();

    cout << "Program begins" << endl;

    static void* table[] = {
        &&do_add,
        &&do_sub,
        &&do_mul,
        &&do_div,
        &&do_end,
        &&do_err,
        &&do_fin};

#define jump() goto *table[op[c++]]

    jump();
do_add:
    add(in1[c], in2[c], out[c]); jump();
do_sub:
    sub(in1[c], in2[c], out[c]); jump();
do_mul:
    mul(in1[c], in2[c], out[c]); jump();
do_div:
    div(in1[c], in2[c], out[c]); jump();
do_end:
    cout << "end of program" << endl; goto *table[6];
do_err:
    cout << "ERROR!!!" << endl; goto *table[6];
do_fin:

    fClock = clock();
    cout << fClock - sClock << endl;

    delete [] op;
    delete [] in1;
    delete [] in2;
    delete [] out;

    in1 = new int[LENGTH];
    in2 = new int[LENGTH];
    out = new int[LENGTH];

    for (int i = 0; i < LENGTH; ++i)
    {
        *(in1+i) = qrand();
        *(in2+i) = qrand()+1;
    }

    cout << "Native begins" << endl;

    sClock = clock();

    for (int i = 0; i < LENGTH; i += 4)
    {

        *(out+i) = *(in1+i) + *(in2+i);
        *(out+i+1) = *(in1+i+1) - *(in2+i+1);
        *(out+i+2) = *(in1+i+2) * *(in2+i+2);
        *(out+i+3) = *(in1+i+3) / *(in2+i+3);
    }

    fClock = clock();
    cout << fClock - sClock << endl;

    delete [] in1;
    delete [] in2;
    delete [] out;

    return 0;
}

回答1:

Darek Mihocka has a good and in-depth writeup on creating fast interpreters in portable C: http://www.emulators.com/docs/nx25_nostradamus.htm

来源：https://stackoverflow.com/questions/11720357/performance-improvement-strategies-for-vm-interpreter