• bearophile <bearophileHUGS lycos.com> Nov 11 2008
• Kagamin <spam here.lot> Nov 12 2008
• Kagamin <spam here.lot> Nov 12 2008
• bearophile <bearophileHUGS lycos.com> Nov 12 2008
• Don <nospam nospam.com> Nov 12 2008
• bearophile <bearophileHUGS lycos.com> Nov 12 2008

bearophile <bearophileHUGS lycos.com> writes:

I have found an interesting small article about optimization, so I've tried the
code in C and D, and I have found strange results (the D code shows timings
opposite of the article).
This is the article, look at the "Branch Prediction" section:
http://www.ddj.com/184405848

The C code:
http://codepad.org/QSGIije4
And its asm (MinGW 4.2.1):
http://codepad.org/c7ZRiXGI

The similar D code:
http://codepad.org/slhcSJEA
Its asm (DMD 1.036):
http://codepad.org/AjlraEs9

There is also about 2X performance difference.

Bye,
bearophile

Kagamin <spam here.lot> writes:

            if (i % 4 == 1) {
                if (i % 4 == 0) {
                    counter1++;
                } else {
                    counter2++;
                }
            } else {
                if (i % 4 == 2) {
                    counter3++;
                } else {
                    counter4++;
                }
            }
this is incorrect

Kagamin <spam here.lot> writes:

You didn't run your code, lol.

bearophile <bearophileHUGS lycos.com> writes:

Kagamin:
 this is incorrect


Thank you for spotting the silly bug, I'll fix it now. (It seems it's more easy
to leave bugs in such kind of code because it does nothing useful).

But note that in both programs:
#define FIRST
static if (1) {
So the first part only is run in both D and C code, not the wrong one...

So the code is like this:

#include "stdio.h"

int main() {
    int counter0 = 0, counter1 = 0, counter2 = 0, counter3 = 0;
    int i = 300000000;
    while (i--) {
        // 0.63 s
        if (i % 4 == 0) {
            counter0++;
        } else if (i % 4 == 1) {
            counter1++;
        } else if (i % 4 == 2) {
            counter2++;
        } else {
            counter3++;
        }
    }

    printf("%d %d %d %d\n", counter0, counter1, counter2, counter3);
    return 0;
}

So the problem and timings of the first part I have shown are correct still :-)

Bye,
bearophile

Don <nospam nospam.com> writes:

bearophile wrote:
 I have found an interesting small article about optimization, so I've tried
the code in C and D, and I have found strange results (the D code shows timings
opposite of the article).
 This is the article, look at the "Branch Prediction" section:
 http://www.ddj.com/184405848
 
 The C code:
 http://codepad.org/QSGIije4
 And its asm (MinGW 4.2.1):
 http://codepad.org/c7ZRiXGI
 
 The similar D code:
 http://codepad.org/slhcSJEA
 Its asm (DMD 1.036):
 http://codepad.org/AjlraEs9
 
 There is also about 2X performance difference.
 
 Bye,
 bearophile


Are you running it on a Pentium 4? Pentium 4 has *horrific* branch 
misprediction (minimum 24 cycles, 45 uops). No other processor is nearly 
as bad, eg it's 15 cycles on Core2; it was just 4 cycles on PMMX.

bearophile <bearophileHUGS lycos.com> writes:

Don:
 Are you running it on a Pentium 4? Pentium 4 has *horrific* branch 
 misprediction (minimum 24 cycles, 45 uops). No other processor is nearly 
 as bad, eg it's 15 cycles on Core2; it was just 4 cycles on PMMX.


Sorry, I am using a Core2   2GHz.
The fixed C code with timings:

#include "stdio.h"

//#define FIRST

int main() {
    int counter0 = 0, counter1 = 0, counter2 = 0, counter3 = 0;
    int i = 300000000;
    while (i--) {
        #ifdef FIRST
            // 0.63 s
            if (i % 4 == 0) {
                counter0++;
            } else if (i % 4 == 1) {
                counter1++;
            } else if (i % 4 == 2) {
                counter2++;
            } else {
                counter3++;
            }
        #else
            // 0.66 s
            if (i & 2) {
                if (i & 1) {
                    counter3++;
                } else {
                    counter2++;
                }
            } else {
                if (i & 1) {
                    counter1++;
                } else {
                    counter0++;
                }
            }
        #endif
    }

    printf("%d %d %d %d\n", counter0, counter1, counter2, counter3);
    return 0;
}


Fixed D code with timings:

void main() {
    int counter0, counter1, counter2, counter3;

    int i = 300000000;
    while (i--)
        static if (0) { // 1.24 s
            if (i % 4 == 0) {
                counter0++;
            } else if (i % 4 == 1) {
                counter1++;
            } else if (i % 4 == 2) {
                counter2++;
            } else {
                counter3++;
            }
        } else { // 1.01 s
            if (i & 2) {
                if (i & 1) {
                    counter3++;
                } else {
                    counter2++;
                }
            } else {
                if (i & 1) {
                    counter1++;
                } else {
                    counter0++;
                }
            }
        }

    printf("%d %d %d %d\n", counter0, counter1, counter2, counter3);
}

As you can see the C version (GCC 4.2.1-dw2) is twice faster than the D one,
and it shows the scan as faster than the binary search, as says the article I
have linked.

Bye,
bearophile