• bearophile <bearophileHUGS lycos.com> Oct 15 2010

bearophile <bearophileHUGS lycos.com> writes:

Walter Bright:

 bearophile wrote:
 With D sometimes array-based code is faster than pointer-based. With LDC they
are usually equally efficient.


 ??? This makes no sense.
 
 The (ptr+i*T.sizeof) is an addressing mode on the x86, and comes at ZERO cost.


A little test program:


import std.c.stdio: printf;

size_t add1(size_t[] arr) {
    size_t sum = 0;
    for (size_t i = 0; i < arr.length; i++)
        sum += arr[i];
    return sum;
}

size_t add2(size_t[] arr) {
    size_t sum = 0;
    size_t* p = arr.ptr;
    for (size_t i = 0; i < arr.length; i++)
        sum += p[i];
    return sum;
}

size_t add3(size_t[] arr) {
    size_t sum = 0;
    size_t* p = arr.ptr;
    for (size_t i = 0; i < arr.length; i++)
        sum += *p++;
    return sum;
}

void main() {
    auto arr = new size_t[10_000_000];
    foreach (size_t i, ref el; arr)
        el = i;
    printf("%u\n", add1(arr));
    printf("%u\n", add2(arr));
    printf("%u\n", add3(arr));
}

---------------------

dmd 2.049
dmd -O -release -inline

_D4test4add1FAkZk   comdat
        push EAX
        xor ECX,ECX
        xor EDX,EDX
        push EBX
        cmp 0Ch[ESP],ECX
        je  L28
        mov 4[ESP],EDX
        mov EDX,010h[ESP]
        mov EBX,EDX
        mov EAX,0Ch[ESP]
        mov EDX,4[ESP]
L1E:    add ECX,[EDX*4][EBX]
        inc EDX
        cmp EDX,0Ch[ESP]
        jb  L1E
L28:    pop EBX
        mov EAX,ECX
        pop ECX
        ret 8

_D4test4add2FAkZk   comdat
        xor ECX,ECX
        xor EDX,EDX
        cmp 4[ESP],ECX
        je  L18
LA:     mov EAX,8[ESP]
        add ECX,[EDX*4][EAX]
        inc EDX
        cmp EDX,4[ESP]
        jb  LA
L18:    mov EAX,ECX
        ret 8

_D4test4add3FAkZk   comdat
        push EBX
        xor EDX,EDX
        xor ECX,ECX
        cmp 8[ESP],ECX
        mov EBX,0Ch[ESP]
        je  L1D
LF:     mov EAX,EBX
        add EBX,4
        inc ECX
        add EDX,[EAX]
        cmp ECX,8[ESP]
        jb  LF
L1D:    pop EBX
        mov EAX,EDX
        ret 8


This has an influence on running time too.

Bye,
bearophile