Unfortunately, for POPCNT assembler, SSE 4.2 is required. For older
CPUs, after some tests, I'd finally recommend
function CShift(ANum : QWord) : Cardinal;
var
n1, n2 : QWord;
begin
n1 := ANum - (ANum shr 1 and $5555555555555555);
n2 := (n1 and $3333333333333333) + (n1 shr 2 and $3333333333333333);
Result :=
(((n2 + (n2 shr 4)) and $0f0f0f0f0f0f0f0f) * $0101010101010101) shr 56;
end;
What do you do, when it comes to fast pop. counts on vintage CPUs?
The "official" alternative, provided with FPC 2.7.1, seems ca. 10 x
slower. Did I grab the wrong version?
Is there a even better solution?
_______________________________________________
fpc-pascal maillist - fpc-pascal@lists.freepascal.org
http://lists.freepascal.org/mailman/listinfo/fpc-pascal
