|
Compilers will generate jump statements. I don't mind, as long as I don't have to trace them, and maintain the code at that level.
Other generators may generate source format goto statements. I don't mind, as long as I don't have to trace them, and maintain the code at that level.
If the gotos and labels you presented are created by a generator, and you never will have to trace them and maintain the code at that level, I do not consider them "your" gotos. Not any more than I consider the conditional and unconditional jumps generated from your source code to be "your" jmp instructions.
I'd be curious to also see your input to Visual FA to generate this code, as well as the table driven code generated by Visual FA!
If you have any reason at all to relate to the generated code: Trading readability and maintainability for "slightly faster code" is generally a bad move. Besides: From the snippet you presented, I am surprised that a table driven variety generated from the same input can be even "slightly" slower. I really wonder what that generator generates! (That is why I'd like to see it.)
I have never been using Visual FA, but from what I gather from a quick net search, it looks like you are not at all using SM as a development too. You are just generating code for different virtual machines, one with a state oriented execution model, one with a C/C++ oriented execution model. A comparison between them is like compiling some (arbitrary language) source code for x64 and for AArch64 and observing that the x64 is "slightly faster".
To me, the SM table is not the result delivered by a generator - it is a modeling tool for the human developer. The driver is typically a score code statements, independent of the model (a.k.a. transition table). I certainly agree that an editor tailored to transition table editing is a great thing to have. I have tried to maintain a compile time initialized C++ array for a transition table, using Np++. Even for trivial SMs, that is almost impossible (unless you just copy the table from e.g. a standard document and it will be carved in stone).
Religious freedom is the freedom to say that two plus two make five.
|
|
|
|
|
trønderen wrote: If you have any reason at all to relate to the generated code. Trading readability and maintainability for "slightly faster code" is generally a bad move. . I generally agree with you. However, as we both know there are exceptions, which is why you used the word "generally" I'm sure. This is one of those cases, as lexing is always in a critical code path, and a generalized lexer must be able to handle bulk input as efficiently as possible.
My input to visual fa is one or more regular expressions. Literally just that. Here is the full input for that generated lexer, in my .rl Rolex lexer format, but it should be easy enough to discern the grammar below without knowing the format.
Object = "{"
ObjectEnd = "}"
Array = "["
ArrayEnd = "]"
FieldSeparator = ":"
Comma = ","
Number = '-?(?:0|[1-9][0-9]*)(?:\.[0-9]+)?(?:[eE][+-]?[0-9]+)?'
Boolean = 'true|false'
Null = "null"
String = '"([^\n"\\]|\\([btrnf"\\/]|(u[0-9A-Fa-f]{4})))*"'
WhiteSpace = '[ \t\r\n]+'
The table driven code is run on a flat array of integers. It might be more efficient to unflatten it in this case - maybe? I used to run a more complicated array of structs for this, and I don't remember there being a performance difference. But anyway, there is also an array of int arrays for a feature called block ends, which simulate lazy matching on a DFA. (I have the details of all of it documented in my Visual FA series). It's also simpler in operation than it looks. I do actually use gotos in a couple of places here to restart the state machine. It was much less complicated than orchestrating a while with breaks. I should state that I didn't comment the code here because it wouldn't help me. It may help others, but I didn't really care about that. This pattern is burned into my brain after writing more than half a dozen lexers that follow the same. It honestly would just clutter it for me, as the code makes immediate sense to me despite how it looks, and I didn't write it for a team.
private FAMatch _NextImpl(
#if FALIB_SPANS
ReadOnlySpan<char> s
#else
string s
#endif
)
{
int tlen;
int tto;
int prlen;
int pmin;
int pmax;
int i;
int j;
int state = 0;
int acc;
if (position == -1)
{
++position;
}
int len = 0;
long cursor_pos = position;
int line = this.line;
int column = this.column;
int ch = -1;
Advance(s, ref ch, ref len, true);
start_dfa:
acc = _dfa[state];
++state;
tlen = _dfa[state];
++state;
for (i = 0; i < tlen; ++i)
{
tto = _dfa[state];
++state;
prlen = _dfa[state];
++state;
for (j = 0; j < prlen; ++j)
{
pmin = _dfa[state];
++state;
pmax = _dfa[state];
++state;
if (ch < pmin)
{
state += ((prlen - (j + 1)) * 2);
j = prlen;
}
else if (ch <= pmax)
{
Advance(s, ref ch, ref len, false);
state = tto;
goto start_dfa;
}
}
}
if (acc != -1)
{
int sym = acc;
int[] be = (_blockEnds != null && _blockEnds.Length > acc) ? _blockEnds[acc] : null;
if (be != null)
{
state = 0;
start_be_dfa:
acc = be[state];
++state;
tlen = be[state];
++state;
for (i = 0; i < tlen; ++i)
{
tto = be[state];
++state;
prlen = be[state];
++state;
for (j = 0; j < prlen; ++j)
{
pmin = be[state];
++state;
pmax = be[state];
++state;
if (ch < pmin)
{
state += ((prlen - (j + 1)) * 2);
j = prlen;
}
else if (ch <= pmax)
{
Advance(s, ref ch, ref len, false);
state = tto;
goto start_be_dfa;
}
}
}
if (acc != -1)
{
return FAMatch.Create(sym,
#if FALIB_SPANS
s.Slice(unchecked((int)cursor_pos), len).ToString()
#else
s.Substring(unchecked((int)cursor_pos), len)
#endif
, cursor_pos, line, column);
}
if (ch == -1)
{
return FAMatch.Create(-1,
#if FALIB_SPANS
s.Slice(unchecked((int)cursor_pos), len).ToString()
#else
s.Substring(unchecked((int)cursor_pos), len)
#endif
, cursor_pos, line, column);
}
Advance(s, ref ch, ref len, false);
state = 0;
goto start_be_dfa;
}
return FAMatch.Create(acc,
#if FALIB_SPANS
s.Slice(unchecked((int)cursor_pos), len).ToString()
#else
s.Substring(unchecked((int)cursor_pos), len)
#endif
, cursor_pos, line, column);
}
while (ch != -1)
{
var moved = false;
state = 1;
tlen = _dfa[state];
++state;
for (i = 0; i < tlen; ++i)
{
++state;
prlen = _dfa[state];
++state;
for (j = 0; j < prlen; ++j)
{
pmin = _dfa[state];
++state;
pmax = _dfa[state];
++state;
if (ch < pmin)
{
state += ((prlen - (j + 1)) * 2);
j = prlen;
}
else if (ch <= pmax)
{
moved = true;
}
}
}
if (moved)
{
break;
}
Advance(s, ref ch, ref len, false);
}
if (len == 0)
{
return FAMatch.Create(-2, null, 0, 0, 0);
}
return FAMatch.Create(-1,
#if FALIB_SPANS
s.Slice(unchecked((int)cursor_pos), len).ToString()
#else
s.Substring(unchecked((int)cursor_pos), len)
#endif
, cursor_pos, line, column);
}
Check out my IoT graphics library here:
https://honeythecodewitch.com/gfx
And my IoT UI/User Experience library here:
https://honeythecodewitch.com/uix
|
|
|
|
|
trønderen wrote: To me, the SM table is not the result delivered by a generator - it is a modeling tool for the human developer. The driver is typically a score code statements, independent of the model (a.k.a. transition table). I certainly agree that an editor tailored to transition table editing is a great thing to have. I have tried to maintain a compile time initialized C++ array for a transition table, using Np++. Even for trivial SMs, that is almost impossible (unless you just copy the table from e.g. a standard document and it will be carved in stone).
I didn't address this part of your post. I should. I don't expect transition tables to be especially readable.
Visual FA is called "Visual" because it can produce images - directed graphs of the state machines that match 1:1 with the generated tables and code. For example, q0 in the graph corresponds to the q0 : label in the goto table. It's perspicuous enough and yet concise enough that I've used the graphs as a guide to hand roll lexers when i needed the lexer to perform additional work beyond what a strict DFA could provide.
I've also used them to debug. Since the correlation is 1:1 between the graphs and the compiled code, it's easier to follow along with than the tables, but both can be followed with the graphs.
The graphs effectively become in part, documentation for the state machine, and for that they work pretty well.
Check out my IoT graphics library here:
https://honeythecodewitch.com/gfx
And my IoT UI/User Experience library here:
https://honeythecodewitch.com/uix
|
|
|
|
|
Generally you are right, usually the codewitch works on small embedded systems where performances and code footprint can be extremely stringent.
I found myself doing things I would have abhorred only a few scant years ago...
GCS/GE d--(d) s-/+ a C+++ U+++ P-- L+@ E-- W+++ N+ o+ K- w+++ O? M-- V? PS+ PE Y+ PGP t+ 5? X R+++ tv-- b+(+++) DI+++ D++ G e++ h--- r+++ y+++* Weapons extension: ma- k++ F+2 X
The shortest horror story: On Error Resume Next
|
|
|
|
|
Table driven implementations usually reduces the control code to typically a couple hundred bytes (or even less), at the expense of data space for the table.
By using data elements no larger than required in the transition table entries, each entry can be kept to a very moderate size.
One possible issue is the number of states and events. It takes some experience to control both, to keep the table size (#states * #events) under control. A common trick is to introduce 'state variables'.
Some times you can use 2+ small tables rather than a huge one, e.g. if you implement a communication protocol: One table for the connect phase, one for the data transfer phase.
Many transition tables are rather sparse anyway, but a lot of methods for space efficient storage of sparse matrices are basic data structure knowledge. E.g. non-empty entries may be factored out to a packed, linear array, and the large table contains indexes to this array. Often, several transitions are identical (typically in one state, for different events, or for one event in several different states) - then a linear table need to hold only a single copy.
Certainly, really old embedded processors (such as 8051) had very little data space; expanding code space was far easier (e.g. through banking hardware). While we would usually call the transition table 'data', it is 100% read-only, and may very well be burnt in ROM (ok, call it 'flash' nowadays) together with the driver code.
If you consider CLR for an embedded CPU (don't try that on an 8051!), then you definitely can fit a packed transition table. My guess is that the total code+data size would be significantly smaller than an equivalent implementation with switch cases and/or if/else-sequences. And faster, even if a packed table will lead to a couple more indirections.
I will maintain that table driven state machines can be a very good solution for embedded processors.
Religious freedom is the freedom to say that two plus two make five.
|
|
|
|
|
trønderen wrote: If I were given the responsibility for a state machine implementation like that, I would immediately run to my boss asking for permission to rewrite the whole thing as a table driven machine.
... or as a state machine that returns function pointers instead of using tables and state variables:
#include <stdio.h>
#include <conio.h>
typedef void (*RT)( int input );
typedef RT (*TER)( int input );
extern TER state1( int input );
extern TER state2( int input );
extern TER state3( int input );
TER state1( int input )
{
printf( "one\t" );
return input < 10 ? (TER)&state2 : (TER)NULL;
}
TER state2( int input )
{
printf( "two\t" );
return (TER)&state3;
}
TER state3( int input )
{
printf( "three\t" );
return (TER)&state1;
}
int main(int argc, char* argv[])
{
int n;
TER state = (TER)&state1;
for ( n = 0 ; state ; ++n ) {
state = (TER)( state( n ) );
}
printf( "\n\nPress any key\n" );
getch();
return 0;
}
Type casts are useful because in C it's impossible to declare function pointers that return function pointers that return function pointers that return function pointers...
Regards
|
|
|
|
|
I hate function pointer dispatch code in general.
Because at some point you'll have to debug and maintain it, and you end up with impossible to follow pointer arrays hiding the flow of your app.
Check out my IoT graphics library here:
https://honeythecodewitch.com/gfx
And my IoT UI/User Experience library here:
https://honeythecodewitch.com/uix
|
|
|
|
|
honey the codewitch wrote: with impossible to follow pointer arrays
There are no pointer arrays in my code.
|
|
|
|
|
Sorry, I was speaking generally about dispatch function pointers. Your statement just remind me of it. Sorry I wasn't clear. I just woke up when I wrote that.
Check out my IoT graphics library here:
https://honeythecodewitch.com/gfx
And my IoT UI/User Experience library here:
https://honeythecodewitch.com/uix
|
|
|
|
|
No problem. I'm not a native English speaker so I always fear being misunderstood.
|
|
|
|
|
I'm not sure why it wouldn't be pretty straightforward to [TestCase()] for each of the branching?
I don't think this code is very cyclomatically complex?
But yeah when you say table driven state machine I'm pretty sure that's where my head is too if you're basically talking a direct map of the case statements to data.
|
|
|
|
|
There is one issue with that.
The compiled ones can be augmented in a way that the table driven ones cannot.
For example, I wrote an embedded JSON pull parser in C++. I used compiled DFA code, and then I parsed floats, ints, and bools out of the stream *as* I was lexing, making the API both easier to use and marginally more performant because you didn't have to get the string back and then reexamine it in order to call atoi() or whatever. It was a simple little surgery on the generated code, with excellent results.
I admit this isn't the most common case out there, but I have used this technique several times.
Edited to add: It's also easier in practice to debug and step through a generated lexer than it is a table driven lexer. And with my Visual FA project, it produces images of directed graphs that map one to one to the labels/jump points in the code. q0: maps to the state q0 in the graph. It makes it really easy to see what it's doing, in terms of documenting it.
Check out my IoT graphics library here:
https://honeythecodewitch.com/gfx
And my IoT UI/User Experience library here:
https://honeythecodewitch.com/uix
|
|
|
|
|
Wow, too much with the GOTO's already - they put it in the language for a reason. It will always be in certain languages for the same reasons. We are just debating normal human failings that have nothing to do with GOTO.
We are engineers, we should know that ALL humans a fallible and can make a mess of anything.
Careless use of GOTO helps us make a mess faster, careful use of GOTO makes our code run faster.
|
|
|
|
|
Fair enough. Implement a DFA state machine without gotos, achieving comparable performance. I'll wait.
Check out my IoT graphics library here:
https://honeythecodewitch.com/gfx
And my IoT UI/User Experience library here:
https://honeythecodewitch.com/uix
|
|
|
|
|
I was agreeing with you - my point was, nobody should care if you are using GOTO's, they should only care if you are making a mess with them.
I have never seen you produce a mess, quite the opposite in fact.
|
|
|
|
|
I clearly misunderstood you. Sorry.
And thanks!
Check out my IoT graphics library here:
https://honeythecodewitch.com/gfx
And my IoT UI/User Experience library here:
https://honeythecodewitch.com/uix
|
|
|
|
|
honey the codewitch wrote: I clearly misunderstood you. Sorry.
Np, irony is easily missed in short messaging!
|
|
|
|
|
My CP sig used to be something like, "If you think GOTO's are bad try coding in Assembler without using JMP."
A programmer I once worked with had the opinion that subroutines (that's what methods were called back then) should only have one exit point. Since you couldn't RETURN from where the routine might need to exit and you couldn't use GOTO his longer methods tended to have dozens of embedded IF blocks. Yuck.
There are no solutions, only trade-offs. - Thomas Sowell
A day can really slip by when you're deliberately avoiding what you're supposed to do. - Calvin (Bill Watterson, Calvin & Hobbes)
|
|
|
|
|
When all you have is a hammer, every problem looks like a nail.
Goto's have their place, the problem is when they're being abused because all the developer sees is nails and can't come up with a better solution.
That being said, I'm looking glancing at your code and clearly I'm in no position to criticize.
|
|
|
|
|
Try writing Assembly with out them (the fabled JMP!). They are a tool that get misused (kinda like the powered screw driver).
|
|
|
|
|
![Go to Parent](https://www.codeproject.com/App_Themes/CodeProject/Img/arrow-up24.png) Code runs in LinqPad.
Code runs in LinqPad.
This should be significantly faster than your original code because it speeds up the conditionals by using pattern matching instead of overloadable operators. Also, the local functions can be in-lined, meaning they will be executed in place, which is even more efficient than the `Goto` statements. And now it's not pure spaghetti.
string json = """
{
"test": 0,
"data": "value"
}
""";
JsonStringRunner runner = new();
List<FAMatch> matches = new();
FAMatch current = default;
Stopwatch sw = new();
sw.Start();
do{
current = runner.GetMatch(json);
matches.Add(current);
} while(!runner.isDone);
sw.Stop();
matches.Dump();
sw.Dump();
internal record struct FAMatch(int token, string match, int position, int length, int column)
{
internal static FAMatch Create(int token, string match, int position, int length, int column)
=> new(token, match, position, length, column);
}
internal abstract class FAStringRunner
{
protected int position = -1, line = 0, column = 0;
internal bool isDone = false;
}
internal sealed partial class JsonStringRunner : FAStringRunner
{
private void Advance(string s, ref int ch, ref int len, bool flag)
{
ch = s[position];
position++;
len++;
isDone = !(position < s.Length);
}
private FAMatch NextMatchImpl(string s)
{
int ch;
int len;
int l;
int c;
ch = -1;
len = 0;
if ((this.position is -1))
{
this.position = 0;
}
int p = this.position;
l = this.line;
c = this.column;
this.Advance(s, ref ch, ref len, true);
switch (ch)
{
case 9 or 10 or 13 or 32:
if(ch is 10 or 13){
l = line++;
}
return q1();
case 34:
return q2();
case 44:
return q9();
case 45:
return q10();
case 48:
return q11();
case >= 49 and <= 57:
return q17();
case 58:
return q18();
case 91:
return q19();
case 93:
return q20();
case 102:
return q21();
case 110:
return q26();
case 116:
return q30();
case 123:
return q32();
case 125:
return q33();
}
return errorout();
[MethodImpl(MethodImplOptions.AggressiveInlining)] FAMatch q1()
{
while (ch is 9 or 10 or 13 or 32)
{
this.Advance(s, ref ch, ref len, false);
}
return FAMatch.Create(10, s.Substring(p, len), p, l, c);
}
[MethodImpl(MethodImplOptions.AggressiveInlining)] FAMatch q2()
{
while (ch is (((((>= 0)
and (<= 9))
or ((>= 11)
and (<= 33)))
or ((>= 35)
and (<= 91)))
or ((>= 93)
and (<= 1114111))))
{
this.Advance(s, ref ch, ref len, false);
}
if ((ch is 34))
{
this.Advance(s, ref ch, ref len, false);
return q3();
}
if ((ch is 92))
{
this.Advance(s, ref ch, ref len, false);
return q4();
}
return default;
}
[MethodImpl(MethodImplOptions.AggressiveInlining)] FAMatch q3() => FAMatch.Create(9, s.Substring(p, len), p, l, c);
[MethodImpl(MethodImplOptions.AggressiveInlining)] FAMatch q4()
{
if (ch is 34 or 47
or 92
or 98
or 102
or 110
or 114
or 116)
{
this.Advance(s, ref ch, ref len, false);
return q2();
}
if (ch is 117)
{
this.Advance(s, ref ch, ref len, false);
return q5();
}
return errorout();
}
[MethodImpl(MethodImplOptions.AggressiveInlining)] FAMatch q5()
{
if (ch is (>= 48 and <= 57) or (>= 65 and <= 70)
or (>= 97 and <= 102))
{
this.Advance(s, ref ch, ref len, false);
return q6();
}
return errorout();
}
[MethodImpl(MethodImplOptions.AggressiveInlining)] FAMatch q6()
{
if (ch is (>= 48 and <= 57) or (>= 65 and <= 70)
or (>= 97 and <= 102))
{
this.Advance(s, ref ch, ref len, false);
return q7();
}
return errorout();
}
[MethodImpl(MethodImplOptions.AggressiveInlining)] FAMatch q7() => FAMatch.Create(7, s.Substring(p, len), p, l, c);
[MethodImpl(MethodImplOptions.AggressiveInlining)] FAMatch q9() => FAMatch.Create(9, s.Substring(p, len), p, l, c);
[MethodImpl(MethodImplOptions.AggressiveInlining)] FAMatch q10() => FAMatch.Create(10, s.Substring(p, len), p, l, c);
[MethodImpl(MethodImplOptions.AggressiveInlining)] FAMatch q11() => FAMatch.Create(11, s.Substring(p, len), p, l, c);
[MethodImpl(MethodImplOptions.AggressiveInlining)] FAMatch q17() => FAMatch.Create(17, s.Substring(p, len), p, l, c);
[MethodImpl(MethodImplOptions.AggressiveInlining)] FAMatch q18() => FAMatch.Create(18, s.Substring(p, len), p, l, c);
[MethodImpl(MethodImplOptions.AggressiveInlining)] FAMatch q19() => FAMatch.Create(19, s.Substring(p, len), p, l, c);
[MethodImpl(MethodImplOptions.AggressiveInlining)] FAMatch q20() => FAMatch.Create(20, s.Substring(p, len), p, l, c);
[MethodImpl(MethodImplOptions.AggressiveInlining)] FAMatch q21() => FAMatch.Create(21, s.Substring(p, len), p, l, c);
[MethodImpl(MethodImplOptions.AggressiveInlining)] FAMatch q26() => FAMatch.Create(26, s.Substring(p, len), p, l, c);
[MethodImpl(MethodImplOptions.AggressiveInlining)] FAMatch q30() => FAMatch.Create(110, s.Substring(p, len), p, l, c);
[MethodImpl(MethodImplOptions.AggressiveInlining)] FAMatch q32() => FAMatch.Create(123, s.Substring(p, len), p, l, c);
[MethodImpl(MethodImplOptions.AggressiveInlining)] FAMatch q33() => FAMatch.Create(125, s.Substring(p, len), p, l, c);
[MethodImpl(MethodImplOptions.AggressiveInlining)] FAMatch errorout() => FAMatch.Create(0, s.Substring(p, len), p, l, c);
}
public FAMatch GetMatch(string s) => NextMatchImpl(s);
}
|
|
|
|
|
I'll have to try a variation of this, but what you produced won't function due to the returns. How are you going to loop?
Check out my IoT graphics library here:
https://honeythecodewitch.com/gfx
And my IoT UI/User Experience library here:
https://honeythecodewitch.com/uix
|
|
|
|
|
Without the full code I didn't know what the logic inside of the various labelled location did, so I simply returned the current substring as a FAMatch. Your method dumps out as an FAMatch so I defaulted to that behavior. The point is that inlined local methods are going to be just as fast as gotos and the pattern matching is much more efficient.
|
|
|
|
|
Sure, I understand. I did say it was a DFA state machine implementation but unless you're a total FA nerd like I am that probably doesn't mean anything.
I'm very curious about the inlined local method and pattern matching approach, particularly the IL it generates, because I don't understand how it would be faster than the IL my code produces - particularly my direct compiler which can short circuit the if tests because the comparisons are in sorted order.
Check out my IoT graphics library here:
https://honeythecodewitch.com/gfx
And my IoT UI/User Experience library here:
https://honeythecodewitch.com/uix
|
|
|
|
|
We're running a heavy process that's causing some of our servers to creak a little. Sorry about any slowness you might experience today. It'll be over soon, I hope.
cheers
Chris Maunder
|
|
|
|
|