Rick Strahl's Weblog
Rick Strahl's FoxPro and Web Connection Weblog
White Papers | Products | Message Board | News |

Workaround for horrendously slow SUBSTR Character Parsing in FoxPro


2 comments
July 03, 2020 •

FoxPro string processing generally is reasonably fast. String processing has always been optimized with a number of optimizations that make FoxPro - despite its age - reasonably competitive when it comes to effective string processing.

But there's an Achilles heel in the string processing support: There's no decent high performance way to iterate a string character by character. If you are building system components like parsers that's a key feature and it's one that FoxPro - and there is no other way to say this - sucks at.

SUBSTR(): Slow as Heck

The FoxPro SUBSTR() function is the only native language function that you can use to iterate over string character by character:

FOR lnX = 1 to LEN(lcString)
   lcChar = SUBSTR(lcString,lnx, 1)
ENDFOR

At first glance this seems fine. If you work with small strings this is reasonably fast and there's no problem. But if the longer the string gets the slower this function becomes.

How slow? Very! Consider this 1mb string and parsing through it character by character:

*** Warning this code will lock FoxPro for a long time!
LOCAL lnX, lcString
lcString = REPLICATE("1234567890",1000000)

lnLength = LEN(lcString)
TRANSFORM(lnLength,"9,999,999")

IF .T.
lnSecs = SECONDS()

FOR lnX = 1 TO lnLength
   lcVal = SUBSTR(lcString,lnX,1)
ENDFOR

*** 35 seconds! Holy crap... 
? "SUBSTR: " +  TRANSFORM(SECONDS() - lnSecs )

ENDIF

Yikes! On my machine this takes an eternity of 35 seconds. Doubling the iterations makes this take nearly 3 minutes! In other words, the slowdown isn't linear - it gets much worse as the size increases.

Here's how this shakes out (on a reasonably fast, high end I7 Dell XPS laptop):

  • 1mb - 35 seconds
  • 2mb - 168 seconds
  • 3mb - 525 seconds

Yikes. I would argue even the 1mb use case is totally unusable let alone larger strings.

Bottom line:

SUBSTR() is slow as shit on large strings!

Does this matter?

For most types of applications it's rare that you need to iterate over a string 1 character at a time. But there are some applications where that's necessary especially if you're building system components like parsers - of which I've built a few over the years and every time I look into it I run into the SUBSTR() issue.

A few years back I built my first version of a JSON parser in FoxPro code. A text parser typically needs to read strings a character at a time to determine what to do with the next character, using a sort of state machine. I figured out right away that using SUBSTR() wasn't an option for this. It was horrendously slow once string sizes got even moderately large. I opted for some other approaches that would read to specific string bound using string extraction and while that actually provided decent performance it resulted in a number of artifacts and inconsistencies that required workarounds and ended up being a constant flow of incoming special case scenario bugs. Very unsatisfying.

The performance was so bad that I ended up throwing away the original parser and instead opted to use a .NET based parser (Newtonsoft.Json) and capturing the results as FoxPro objects. Even with all the COM interop involved this solution ran circles around the native implementation and as a side effect has been rock solid because the JSON parser is a hardened and dedicated component that is regular updated and patched.

Is there a WorkAround?

Natively FoxPro doesn't offer any good workarounds for the SUBSTR() quandry. However, if you really need this functionality there are a few ways you can get around it using creative alternatives.

Two that I found are:

  • Using wwDotnetBridge::ComArray on .ToCharArray()
  • Using a file and FREAD(lnHandle,1)

wwDotnetBridge and ToCharArray()

Since FoxPro natively doesn't have a solution, it's reasonable to hold the data in another environment and then retrieve the data. One option for this is to use .NET and wwDotnetBridge which provides for the ability to store a string in .NET and manipulate it without loading the string or a character array thereof into FoxPro.

Here's what you can do with wwDotnetBridge to parse a string character by character:

do wwDotNetBridge
loBridge = GetwwDotnetBridge()

lnSecs = SECONDS()

* Returns a COM Array object
loStringArray = loBridge.InvokeMethod(lcString,"ToCharArray")
lnLength = loStringArray.Count

FOR lnX = 0 TO lnLength -1
	lcVal = CHR(loStringArray.Item(lnX))
ENDFOR

* ~ 2.5 seconds (still not very quick really) - in .NET same iteration takes 0.03 seconds
? "wwDotnetBridge ToCharArray(): " + TRANSFORM( SECONDS() - lnSecs ) 

ENDIF

2.5 seconds - while still not blazing fast - is considerably better. Performance of this solution is also linear - doubling the string size roughly doubles the time it takes to process plus a little extra overhead, but overall, linear.

The way this works is that wwDotnetBridge.InvokeMethod() returns arrays as a ComArray structure where the actual array isn't passed to FoxPro but stays in a separately stored value in .NET. The .Item() method then retrieves an indexed value out of the array which is very fast in .NET.

While better than SUBSTR() for large strings, this approach is not as fast as it could be, because there's a lot of overhead in the COM Interop involved in retrieving the individual value from the array. Still the performance with this at least is predicatable and linear.

Using FREAD() to Iterate

Another creative solution that doesn't require an external component like wwDotnetBridge is to use low level file functions to:

  • Dump the string to a file
  • Open and read the file 1 byte at a time
  • Close the File
  • Delete the file

This seems very inefficient, but file operations on a local drive are actually blazingly fast. The actual file read operations are buffered so reading the bytes is fast. Most of the overhead of this solution is likely to come from dumping the file to disk and then deleting it when you're done with the string.

String Splitting

After I posted this article and replied to an original posting on the Universal Thread, Marco Plaza replied with another even better and more performant solution that involves breaking the large string up into small chunks and then running SUBSTR() against that.

I made a few tweaks to Marcos version by making the code more generic so it doesn't hardcode the sizes:

lnSecs = SECONDS()

lnCount = 0
lnBuffer = 3500
FOR lnX = 0 TO 9999999999
	lcSub = SUBSTR(m.lcString,(lnX*lnBuffer)+1,lnBuffer)
	lnLen = LEN(lcSub)
	IF (lnLen < 1)
	    EXIT
	ENDIF
	For lnY = 1 To lnLen
		lcVal = Substr(lcSub,lnY,1)
		* lnCount = lnCount + 1  && to make sure right iterations are used
	Endfor
ENDFOR

* 0.35 seconds 
? "String Splitting: " + TRANSFORM( SECONDS() - lnSecs )  + " " + TRANSFORM(lnCount)

A few Results to Compare

To test this out I set up a small example:

CLEAR

* 1mb string
LOCAL lnX, lcString
lcString = REPLICATE("1234567890",100000)

lnLength = LEN(lcString)
? "Iterations: " + TRANSFORM(lnLength,"9,999,999")

IF .T.
lnSecs = SECONDS()

FOR lnX = 1 TO lnLength
   lcVal = SUBSTR(lcString,lnX,1)
ENDFOR

*** 35 seconds
? "SUBSTR: " +  TRANSFORM(SECONDS() - lnSecs )

ENDIF

IF .T.
do wwDotNetBridge
LOCAL loBridge as wwDotNetBridge
loBridge = GetwwDotnetBridge()

lnSecs = SECONDS()

* Returns a COM Array object
loStringArray = loBridge.InvokeMethod(lcString,"ToCharArray")
lnLength = loStringArray.Count
FOR lnX = 0 TO lnLength -1
	lcVal = CHR(loStringArray.Item(lnX))
	*lcVal = loBridge.GetIndexedproperty(loStringArray.Instance,lnX)
ENDFOR

* ~ 2.5-2.9 seconds (still not very quick really) - in .NET same iteration takes 0.03 seconds
? "wwDotnetBridge ToCharArray(): " + TRANSFORM( SECONDS() - lnSecs ) 

ENDIF


IF .T.

lnSecs = SECONDS()

STRTOFILE(lcString,"c:\temp\test.txt")
lnHandle = FOPEN("c:\temp\test.txt",0)

DO WHILE !FEOF(lnHandle)
  lcVal = FREAD(lnHandle,1)
ENDDO
FCLOSE(lnHandle)
ERASE("c:\temp\test.txt")

* 3.6 seconds 
? "FREAD : " + TRANSFORM( SECONDS() - lnSecs ) 

ENDIF

Running this with:

1,000,000

The divergence here is so large that SUBSTR() is basically unusable at this size:

  • SUBSTR: 35s
  • ToCharArray: 2.6
  • FREAD: 3.7
  • Split: 0.319

500,000

Here SUBSTR() is borderline but already very slow to process the string.

  • SUBSTR: 8.62s
  • ToCharArray: 1.32
  • FREAD: 1.85
  • Split: 0.022

100,000

Here SUBSTR() is roughly on par with the .NET solution. This is roughly the tipping point where SUBSTR() becomes significantly less efficient for anything larger.

  • SUBSTR: 0.26
  • ToCharArray: 2.66
  • FREAD: 0.39
  • Split: 0.319

10,000

At smaller sizes SUBSTR() performs considerably better than the other solutions. here you can see that SUBSTR() is 10x faster than the other solutions.

  • SUBSTR: 0.003
  • ToCharArray: 0.025
  • FREAD: 0.041
  • Split: 0.002

Summary

From the results above it would appear that SUBSTR() is reasonably fast up to about 100kb string, but anything larger than that and its performance starts dropping of extremely quickly and if you need to parse character by character you really need to look at other approaches.

Once that happens you need to really look at one of the alternatives. Marco's string splitting approach is by far the fastest solution, but it'll require a bit of logic to make sure you're iterating the string correctly (inner and outer loops) - if you need to grab the individual character in several places and it's not easily isolated into a single location the double loop approach can be tricky to keep consistent. The wwDotnetBridge approach also works reasonably fast and other than the initial string load, retains basically the same behavior as `SUBSTR(). But then again - if you're running into perf issues because your strings are too big you'll probably want to go with the fastest possible solution which is the string splitting.

It's not too often that you'll need this sort of code, but it's good to know there are workarounds to make this work if necessary.

this post created and published with the Markdown Monster Editor
Posted in: FoxPro

Feedback for this Weblog Entry


Re: Workaround for horrendously slow SUBSTR Character Parsing in FoxPro



Christof Wollenhaupt
July 05, 2020

It'll be faster if you copy the string into memory and then use SYS(2600) to read byte by byte.

LOCAL lnX, lcString
lcString = REPLICATE("1234567890",100000)

	Declare Integer HeapAlloc in Win32Api Integer, Integer, Integer
	Declare Integer HeapFree in Win32APi Integer, Integer, Integer
	Declare Long GetProcessHeap in Win32API
      
lnLength = LEN(lcString)
TRANSFORM(lnLength,"9,999,999")

IF .T.
lnSecs = SECONDS()

lnBase = HeapAlloc( GetProcessHeap(), 0, m.lnLength )
Sys(2600, m.lnBase, m.lnLength, m.lcString)
FOR lnX = 1 TO lnLength
	lcVal = Sys(2600,m.lnBase-1+m.lnX,1)
ENDFOR
HeapFree( GetProcessHeap(), 0, m.lnBase )

? "SYS(2600): " +  TRANSFORM(SECONDS() - lnSecs )

ENDIF

That's 0.33 seconds on my machine vs 37 seconds with SUBSTR().

Re: Workaround for horrendously slow SUBSTR Character Parsing in FoxPro



Rick Strahl
July 06, 2020

@Christof - that's great! Another great solution - looks like roughly on par with Marco's improvement.

 
© Rick Strahl, West Wind Technologies, 2003 - 2024