How to speed up this conversion function? - performance

Question

I have this function called IntToStrLen which converts an integer to a string of a padded size. For example, an integer value of 42 with a size of 4 would result in a string __42 (padded with a given character, space by default).
The problem is when this function is used in bulk (for example 1,000,000 times in a loop), it adds extra weight onto the loop. The loop I'm using this in, without this function, takes about 20 seconds, but with this function, I'm still waiting as of right now for this function to complete, after about 5 minutes.
How can I speed up the following function?
function IntToStrLen(const Value: Integer; const Len: Integer;
const Fill: String = ' '): String;
var
T: String;
begin
Result:= IntToStr(Value); //convert result
if Length(Result) > Len then
Result:= Copy(Result, 1, Len) //forcefully truncate
else if Length(Result) < Len then begin
T:= '';
while Length(T) < (Len - Length(Result)) do //fill space with character
T:= T + Fill;
Result:= T + Result; //return combination
end;
end;

Answer 1

The absolute number one change that you can make here is to avoid heap allocations. Your code presently has multiple heap allocations. Your goal here is to write a function with zero heap allocations.
That means you need the caller to allocate and provide the buffer. Typically they will do so with a stack allocated buffer, which hence costs nothing to allocate (or deallocate). If the caller really needs a string, which is always allocated on the heap, then the caller can wrap your function with a call to SetString.
The function prototype might look like this:
procedure IntToStrLen(const Value, Len: Integer; var Buffer: array of Char;
const Fill: Char = ' ');
The first point to stress here is that Fill must be a Char. Your use of string is inefficient and allows the caller to call a fill "character" with length not equal to one. Doing so would of course break your function because it would return a value with length not equal to Len.
Do note also that the implementation must not call IntToStr because that involves heap allocation. So you need to write your own heap allocation free integer to decimal text conversion code because, astonishingly, the RTL does not offer such functionality. When I do this I use code like so:
procedure DivMod(Dividend, Divisor: Cardinal;
out Quotient, Remainder: Cardinal);
{$IFDEF CPUX86}
asm
PUSH EBX
MOV EBX,EDX
XOR EDX,EDX
DIV EBX
MOV [ECX],EAX
MOV EBX,Remainder
MOV [EBX],EDX
POP EBX
end;
{$ELSE IF Defined(CPUX64)}
asm
.NOFRAME
MOV EAX,ECX
MOV ECX,EDX
XOR EDX,EDX
DIV ECX
MOV [R8],EAX
MOV [R9],EDX
end;
{$ELSE}
{$Message Error 'Unrecognised platform.'}
{$ENDIF}
function CopyIntegerToCharBuffer(const Value: Integer;
var Buffer: array of Char): Integer;
var
i, j: Integer;
val, remainder: Cardinal;
negative: Boolean;
tmp: array [0..15] of Char;
begin
negative := Value<0;
val := abs(Value);
Result := 0;
repeat
DivMod(val, 10, val, remainder);
tmp[Result] := Chr(remainder + ord('0'));
inc(Result);
until val=0;
if negative then begin
tmp[Result] := '-';
inc(Result);
end;
Assert(Result<=Length(Buffer));
i := 0;
j := Result-1;
while i<Result do begin
Buffer[i] := tmp[j];
inc(i);
dec(j);
end;
end;
Now you can make decimal text representations of integers without touching the heap. From there it's a short way to your function.
procedure IntToStrLen(const Value, Len: Integer; var Buffer: array of Char;
const Fill: Char = ' ');
var
tmp: array [0..15] of Char;
i, N: Integer;
begin
Assert(Length(Buffer)>=Len);
N := CopyIntegerToCharBuffer(Value, tmp);
if N>=Len then begin
Move(tmp, Buffer, SizeOf(Char)*Len);
end else begin
for i := 0 to Len-N-1 do begin
Buffer[i] := Fill;
end;
Move(tmp, Buffer[Len-N], SizeOf(Char)*N);
end;
end;
At this point you'll have gained the bulk of the available performance benefits. From here on in you will be into diminishing returns. You could micro-optimise CopyIntegerToCharBuffer as is done in SysUtils._IntToStr32 for instance. Beyond that I'm sure the implementation of IntToStrLen could be optimised with judicious use of assembler. But such optimisations will yield nothing like the benefit you have gained so far from avoiding the heap.
Of course, all this assumes that you have correctly identified your performance bottleneck. It's all too easy to assume that you know where the performance bottleneck is by statically analysing the code. Unless you've actually profiled it carefully expect to find that your intuition is a poor judge of where to invest optimisation effort.

Answer 2

Try this variant
(and you don't treat negative numbers)
function IntToStrLen(const Value: Integer; const Len: Integer;
const Fill: Char = ' '): String;
var
T: String;
begin
Result:= IntToStr(Value);
if Length(Result) > Len then
SetLength(Result, Len) //forcefully truncate
else if Length(Result) < Len then
Result := StringOfChar(Fill, Len - Length(Result)) + Result;
end;
P.S. So strange truncation (2014=>20) - is what you really want?