x64 SwapEndian(<16bit>) via ROL is faster?
lazarus / 64-FPC development

----------------------------------------------------------------------
AlexTP · 2023-09-08 13:58

This page suggests to do SwapEndian via ROL
https://stackoverflow.com/questions/3065335/how-to-convert-big-endian-
numbers-to-native-numbers-delphi
Code: Pascal  [Select][+][-]window.onload = function(){var x1 =
document.getElementById("main_content_section"); if (x1) { var x =
document.getElementsByClassName("geshi");for (var i = 0; i function
SwapEndian16(Value: smallint): smallint; register;asm  rol   ax,
8end; 
and we now have this in  rtl/x86_64/x86_64.inc
Code: Pascal  [Select][+][-]window.onload = function(){var x1 =
document.getElementById("main_content_section"); if (x1) { var x =
document.getElementsByClassName("geshi");for (var i = 0; i {
SwapEndian(<16 Bit>) being inlined is faster than using assembler
}function SwapEndian(const AValue: SmallInt): SmallInt;{$ifdef
SYSTEMINLINE}inline;{$endif}  begin    { the extra Word type cast is
necessary because the "AValue shr 8" }    { is turned into
"longint(AValue) shr 8", so if AValue < 0 then    }    { the sign bits
from the upper 16 bits are shifted in rather than  }    { zeroes.    
                                                    }    Result :=
SmallInt(((Word(AValue) shr 8) or (Word(AValue) shl 8)) and $ffff); 
end;  function SwapEndian(const AValue: Word): Word;{$ifdef
SYSTEMINLINE}inline;{$endif}  begin    Result := ((AValue shr 8) or
(AValue shl 8)) and $ffff;  end; 
Which is faster?

----------------------------------------------------------------------
birin · 2023-09-08 14:35

If SwapEndian16 are not inlined, then she needs CALL and RET.
Instruction flow interruption is time consuming.

I think the second version of SwapEndian doesn't need the and $ffff
because all values are Word, and Word are unsigned.