Fast Pos & StringReplace for 64 bit

[Tom de Neef 12]

I've seen several queries about 64 bit equivalents for the fast string routines in Win32. And I needed them myself as well.

I have published (assembly based) functions that mimic system.Pos and sysUtils.StringReplace for 32 & 64 bit compiling (both String and AnsiString). Extra feature: searching can be case sensitive or case insensitive.

 

Timing for string search (1k substrings in 1M string, microsec, see documentation; NextStrPos is new):

mode	32 bit			64 bit
search	System.pos	NextStrPos		System.pos	NextStrPos
String	550	500		730	400
ANSIstring	530	460		3.500	360

 

Timing for String replace (1k substrings in 1M string, microsec, see documentation; StrReplace is new):

mode	32 bit			64 bit
replace	SysUtils. StringReplace	StrReplace		SysUtils. StringReplace	StrReplace
String	6.500	1.050		6.000	1.050
ANSIstring	7.400	900		7.000	850

 

Files and documentation on https://sourceforge.net/projects/delphi-fast-pos-stringreplace/

[Stefan Glienke 2066]

Certainly better than the RTL - however if you want it really fast and justify implementing in asm then SSE should be used.

For reference: https://www.strchr.com/strcmp_and_strlen_using_sse_4.2

 

Also, people might not want to use your code because of GPLv3 unless they put it into an extra DLL.

Edited June 1, 2021 by Stefan Glienke

[Tom de Neef 12]

Pls explain your comment about the license. I have no wish to limit use in any way. What license condition should I use then ?

[Stefan Glienke 2066]

https://choosealicense.com/

[dummzeuch 1579]

28 minutes ago, Tom de Neef said:

Pls explain your comment about the license. I have no wish to limit use in any way. What license condition should I use then ?

Apache or MPL are the most commonly used when you want few restrictions.

Edited June 2, 2021 by dummzeuch

[Lars Fosdal 1826]

MIT license is no limit in use.

[Leif Uneus 43]

I still find the fastcoders purepascal Pos() faster for 64 bit in this modification:

https://stackoverflow.com/a/20947429/576719

 

 

[Arnaud Bouchez 407]

On 6/1/2021 at 10:54 PM, Stefan Glienke said:

Certainly better than the RTL - however if you want it really fast and justify implementing in asm then SSE should be used.

For reference: https://www.strchr.com/strcmp_and_strlen_using_sse_4.2

In practice, SSE 4.2 is not faster than regular SSE 2 code for strcmp and strlen. More complex process may benefit of SSE 4.2 - but the fastest JSON parser I have seen doesn't use it, and focuses on micro-parallelized branchless process with regular SIMD instructions - see https://github.com/simdjson/simdjson
Memory access is the bottleneck. This is what Agner measured.

 

About any asm, it is mandatory to refer to https://agner.org/optimize

There are reference code and reference documentation about how modern asm should be written.

 

The PosEx_Sha_Pas_2 version is one of the fastest, and probably faster than your version, even if it is written in pure pascal. For instance, reading a register then shr + cmp is not the fastest pattern today.
Pascal version will also work on Mac and  Linux, whereas your asm version would need additional code to support the POSIX ABI.
We included it (with minimal tweaks like using NativeInt instead of integer, and using an inline stub for register usage) in https://github.com/synopse/mORMot2/blob/master/src/core/mormot.core.base.pas#L7974

 

First thing is to benchmark and compare your code with proper timing, and regular tests.
Try with some complex process, not in naive loops which tends to be biased because in naive tests the data remains in the CPU L1 cache, so numbers are not realistic.

 

Edited June 3, 2021 by Arnaud Bouchez

[Arnaud Bouchez 407]

Here is for instance how FPC compile our PosExPas() function for Linux x86_64.

https://gist.github.com/synopse/1e30b30a77f6b0288310115085401c1e

 

You can see the resulting asm is very efficient.
Thanks to the goto used in the pascal code, and proper use of pointers/registers, to be fair. 😉

 

You may find some inspiring string process in our https://github.com/synopse/mORMot2/blob/master/src/core/mormot.core.unicode.pas#L1404 unit.

This link give you some efficient case-insensitive process over not only latin A-Z chars, but on the whole Unicode 10.0 case folding tables (i.e. work with greek, cyrilic, or other folds).

This code is UTF-8 focused, because we use it instead of UTF-16 in our framework for faster processing with minimal memory allocation and usage.
But you would find some pascal code which is as fast as manual asm with no SIMD support. For better performance, branchless SIMD is the key, but it is much more complex and error prone.
The main trick about case insensitive search is that a branchless version using a lookup table for A-Z chars is faster than cascaded cmp/ja/jb branches, on modern CPUs. We just enhanced this idea to Unicode case folding.

Edited June 3, 2021 by Arnaud Bouchez

[Tom de Neef 12]

I have taken up two suggestions: change the license (to MIT) and use SSE4.2 (both 32 and 64 bit Windows platforms). Timing over 10M trials are as follows:

SEARCH:

mode	32 bit			64 bit
search	System.pos	NextStrPos		System.pos	NextStrPos
String	550	250		730	400
ANSIstring	530	460		3.500	360

 

REPLACE:

mode	32 bit			64 bit
replace	SysUtils. StringReplace	StrReplace		SysUtils. StringReplace	StrReplace
String	6.500	1.050		6.000	1.050
ANSIstring	7.400	900		7.000	850

 

[Stefan Glienke 2066]

64bit StrPos raises AV when either string is empty - also the overall design of your API is a bit weird to me (or I misunderstand something) - why do I need different functions when doing a one time search vs repeated searches. For example when using StrPos in a loop it gets terribly slow because of all the initialization happen in that function.

Edited June 12, 2021 by Stefan Glienke

[Tom de Neef 12]

Thanks for the report - I've corrected.

 

The functions are meant for repeated searches:

var t : TstrPosRec; k : integer;

begin

  t:=SetupStrPos(target,source,ignoreCase);

  repeat

     k:=NextStrPos(t);

     if k=0 then break;

     // user code, based on k

  until false;

 

Thus the initialization overhead is outside the loop.

 

When you code like this:

  k:=1-target.length;

  repeat

     k:=StrPos(target,source,k+target.length);

     if k=0 then break;

     // user code, based on k

  until false;

you bring all the initialization within the loop. The StrPos function is provided only as a replacement for system.pos (and does not offer ignoreCase for that reason).

[Tom de Neef 12]

And I correct the timings. As someone wrote, they vary a lot with circumstances. But the relativities are a reasonable indication.

 

SEARCH:

mode	32 bit			64 bit
search	System.pos	NextStrPos		System.pos	NextStrPos
String	550	250		650	135
ANSIstring	530	460		3.500	65

 

REPLACE:

mode	32 bit			64 bit
replace	SysUtils. StringReplace	StrReplace		SysUtils. StringReplace	StrReplace
String	6.500	1.050		6.000	800
ANSIstring	7.400	900		7.000	650