Use of SHR with signed integers

**Ñuño Martínez** · 01-02-2007, 12:54 PM

I'm using SHR and SHL to convert fixed point <-> integers but I've find a problem.

Code:

program project2;

  FUNCTION integer_to_fixed &#40;v&#58; LONGINT&#41;&#58; LONGINT;
  BEGIN
    integer_to_fixed &#58;= v SHL 16;
  END;

  FUNCTION fixed_to_integer &#40;v&#58; LONGINT&#41;&#58; LONGINT;
  BEGIN
    fixed_to_integer &#58;= v SHR 16;
  END;

VAR
  v&#58; LONGINT;
begin
  v &#58;= integer_to_fixed &#40;-1&#41;; &#123; Gets -1 in 16.16 fixed point. &#125;
  WriteLn &#40;v&#41;;      &#123; Writes the value as stored in memory. &#125;
  WriteLn &#40;fixed_to_integer &#40;v&#41;&#41;; &#123; Should write '-1' but writes '65535'. &#125;
  ReadLn;
end.

As you see, SHR doesn't add the '1' bit. What can I use in the "fixed_to_integer" function?

[edit] Found a solution:

Code:

  FUNCTION fixed_to_integer &#40;v&#58; LONGINT&#41;&#58; LONGINT;
  BEGIN
    IF v < 0 THEN
      fixed_to_integer &#58;= &#40;v SHR 16&#41; OR $FFFF0000
    ELSE
      fixed_to_integer &#58;= v SHR 16;
  END;

Is this the bets one?

**Mirage** · 01-02-2007, 01:17 PM

You can use DIV 65536 of course, but it's slow.

Another way is to use assembler:

Code:

FUNCTION fixed_to_integer &#40;v&#58; LONGINT&#41;&#58; LONGINT; 
  asm 
    sar       eax, 16
  END;

LP · 01-02-2007, 04:38 PM

Originally Posted by Mirage

You can use DIV 65536 of course, but it's slow.

Actually when optimizations are enabled, this division will be translated into "sar" (at least in Delphi; is it the same in FPC?)

**dmantione** · 01-02-2007, 06:53 PM

Yes, FPC uses sar, optimized or unoptimized. However, -3 div 2=-1, while -3 sar 1=-2. To compensate FPC generates this code (register variables off):

Code:

; &#91;12&#93; fixed_to_integer &#58;= v div 65536;
                mov     edx,dword &#91;ebp-4&#93;
                mov     eax,edx
                sar     eax,31
                and     eax,65535
                add     edx,eax
                sar     edx,16
                mov     dword &#91;ebp-8&#93;,edx

**BeRo** · 03-02-2007, 09:06 PM

Code:

function sar&#40;Value,Shift&#58;integer&#41;&#58;integer; &#123;$IFDEF CPU386&#125;assembler; register;&#123;$ELSE&#125;&#123;$IFDEF FPC&#125;inline;&#123;$ELSE&#125;register;&#123;$ENDIF&#125;&#123;$ENDIF&#125;
&#123;$IFDEF CPU386&#125;
asm
 mov ecx,edx
 sar eax,cl
end;
&#123;$ELSE&#125;
begin
 result&#58;=&#40;Value shr Shift&#41; or &#40;&#40;$ffffffff+&#40;1-&#40;&#40;Value and &#40;1 shl 31&#41;&#41; shr 31&#41; and ord&#40;Shift<>0&#41;&#41;&#41; shl &#40;32-Shift&#41;&#41;;
end;
&#123;$ENDIF&#125;

But Delphi does also optimize DIV (2^n) terms to signed shift SAR opcodes.

Moderation Process Reminder

Thread: Use of SHR with signed integers

Thread Tools

Display