Walter Bright:
Sorry for the late reply.
You've got array bounds checking turned on for
algorithm.reverse, and you avoided that in your code by using
unchecked pointers.
Right, sorry. This is compiled with the latest DMD 2.061alpha
with -O -release -inline -noboundscheck. I have used the same
program source code as above.
Now the calls inside std.algorithm.reverse are fully inlined,
below there are no "call" istrunctions.
But the number of instructions performed in the loop is rather
different.
dpaste is currently down and I don't have gdc2/dc2 installed
here, so I can't show the asm from those other compilers.
I have also done benchmarks, and I have seen performance
difference in real code between std.algorithm.reverse and a
reverseArr function.
reverseArr:
push EBX
mov ECX, 0Ch[ESP]
mov EBX, 0Ch[ESP]
push ESI
mov ESI, 0Ch[ESP]
lea ESI, -4[ESI*4][ECX]
cmp ECX, ESI
jae L2B
L19: mov EAX, [EBX]
mov EDX, [ESI]
mov [EBX], EDX
add EBX, 4
mov [ESI], EAX
add ESI, 0FFFFFFFCh
cmp EBX, ESI
jb L19
L2B: pop ESI
pop EBX
ret 8
std.algorithm.reverse:
sub ESP, 014h
push EBX
push ESI
cmp dword ptr 020h[ESP], 0
je L76
LC: mov EAX, 020h[ESP]
mov EDX, 024h[ESP]
mov EBX, 020h[ESP]
mov 0Ch[ESP], EDX
mov ECX, 0Ch[ESP]
mov ESI, [ECX]
mov 014h[ESP], EDX
mov EDX, 014h[ESP]
lea EBX, -4[EBX*4][EDX]
mov 8[ESP], EAX
mov EDX, [EBX]
mov [ECX], EDX
mov ECX, 020h[ESP]
mov EDX, 024h[ESP]
mov 010h[ESP], EAX
lea EAX, -1[ECX]
add EDX, 4
mov 020h[ESP], EAX
mov [EBX], ESI
mov 024h[ESP], EDX
cmp dword ptr 020h[ESP], 0
je L76
mov EBX, 020h[ESP]
lea ESI, -1[EBX]
mov ECX, 024h[ESP]
mov 020h[ESP], ESI
mov 024h[ESP], ECX
cmp dword ptr 020h[ESP], 0
jne LC
L76: pop ESI
pop EBX
add ESP, 014h
ret 8
Bye,
bearophile
