> - the two nested loops in the main are more efficient as ref double,
> this is something dmd will need to fix;
A test shows that on ldc the two nested loops are a little faster without the
ref. I'd like the compiler to use a "const ref" with the foreach iterates on
array items bigger than a word.
I've done a related small test:
version (Tango)
import tango.stdc.stdio: printf;
else
import std.stdio: printf;
void main() {
auto data = new double[1_000];
double tot = 0.0;
foreach (uint i, ref el; data)
tot += el * i;
printf("%f\n", tot);
}
Just the asm of the loop.
Note how ldc keeps two iteration indexes, one in xmm0 and one in ESI, because
it's faster this way. LCPI1_0 is .quad 4607182418800017408 that is the double
value 1.0.
(Here the presence of ref doesn't change the code produced by ldc, but in the
two loops of the original code it makes a little difference).
ldc -O5 -release -inline -enable-unsafe-fp-math -output-s test.d
dmd -O -release -inline test.d
-------------------------
ldc, with ref:
.LBB1_1:
movapd %xmm0, %xmm1
mulsd (%eax,%esi,8), %xmm1
movsd 16(%esp), %xmm2
addsd %xmm1, %xmm2
movsd %xmm2, 16(%esp)
incl %esi
cmpl $1000, %esi
addsd .LCPI1_0, %xmm0
jne .LBB1_1
-------------------------
ldc, without ref:
.LBB1_1:
movapd %xmm0, %xmm1
mulsd (%eax,%esi,8), %xmm1
movsd 16(%esp), %xmm2
addsd %xmm1, %xmm2
movsd %xmm2, 16(%esp)
incl %esi
cmpl $1000, %esi
addsd .LCPI1_0, %xmm0
jne .LBB1_1
-------------------------
dmd, with ref:
L45: mov 8[ESP],EDX
xor ESI,ESI
fld qword ptr [EDX*8][ECX]
mov 0Ch[ESP],ESI
mov EBX,EDX
inc EDX
fild long64 ptr 8[ESP]
fmulp ST(1),ST
fadd qword ptr 018h[ESP]
cmp EDX,010h[ESP]
fstp qword ptr 018h[ESP]
jb L45
-------------------------
dmd, without ref:
L45: mov 8[ESP],EDX
xor ESI,ESI
mov EBX,EDX
mov 0Ch[ESP],ESI
fild long64 ptr 8[ESP]
fmul qword ptr [EDX*8][ECX]
inc EDX
cmp EDX,010h[ESP]
fadd qword ptr 018h[ESP]
fstp qword ptr 018h[ESP]
jb L45
Bye,
bearophile
