Talk:Compiler/lexical analyzer

From Rosetta Code

Clarification[edit]

I like this task, thanks for contributing it. But some more clarification needs to be added before moving it out of draft status. Off the top of my head:

  • encoding: Should we expect the input files in a specific encoding? Maybe latin-1 or utf-8?
  • encoding: Should string and char literals support Unicode, or just ASCII?
  • char literals: The stated regex is 'x', but that's not actually a regex. Shouldn't it be '\\?[^']' (a.k.a. an escape sequence or any character except ', enclosed in single quotes")?
  • char literals: How can a single quote be represented as a char, if there are no other escape sequences besides \n and \\?
  • string literals: The stated regex is ".*", but this would match e.g. "foo bar" < " due to the asterisk performing greedy matching. Shouldn't it be "[^"]*" (a.k.a. "match zero or more characters except the double quote, enclosed in double quotes")?
  • string literals: How can a double quote be represented inside a string literal, if there are no other escape sequences besides \n and \\?
  • whitespace: This needs an actual thorough description, instead of just an example. Am I right to assume that zero or more whitespace characters or comments are allowed between any two tokens, with no exceptions, and that "longest token matching" is used to resolve conflicts (e.g. in order to match <= as a single token rather than the two tokens < and =)?
  • operators: How is the lexer supposed to differentiate between Sub and Uminus? And why does the third test-case print "Sub" for both?


Sorry if some of these sound pedantic, but experience on rosettacode has shown that tasks of this complexity absolutely need to be precise and unambiguous in order to not cause problems for people who will try to add solutions... :)
--Smls (talk) 14:22, 14 August 2016 (UTC)

Another small clarification. The table of valid tokens refers to a "char literal" but the error examples reference "char constants". Are these the same token? --Thundergnat (talk) 14:44, 14 August 2016 (UTC)

Response[edit]

This first category is part of a larger example. In the works, I have syntax analysis, code generation (for a stack based virtual machine) and virtual machine emulator. The code is already complete in C and Python. But no writeups yet.

There are lots of things missing from this simple compiler, as I attempted to weed out features, in order to keep the implementations down to a manageable size. Things like else, >=, ==, data declarations, functions and so on.

The goal was to be able to run simple programs like the prime number generator in the white-space example.

1) encoding (overall):

latin-1

2) encoding - string and char literals:

ASCII

Thinking about it a bit, for a hand-written scanner, there is really nothing that I am aware of preventing string literals and comments from including utf-8. Of course this does not include character literals, where the code would have to be utf-8 aware.

Not sure how/where to update the page regarding the above two.

3) char literal regex:

The (new) definition I'm using for Flex:

\'([^'\n]|\\n|\\\\)\'

Page has been updated.

4) char literals: embedded single quote?

Not supported It is one of the features I arbitrarily removed.

Page has been updated.

5) string literals: regex:

The (new)) definition I'm using for flex:
\"[^"\n]*\"

(thanks for the new definition!)

Page has been updated.

6) string literals: embedded double quotes?

Not supported It is one of the features I arbitrarily removed.

Page has been updated.

7) Whitespace:

I have updated the description.

8) Operators: Sub vs. Uminus

Uminus cannot be recognized by the scanner. It is recognized by the syntax analyzer, i.e., the parser. The token type is there since it will turn up in the parser and the code generator.

Not sure if the page should be updated regarding this.

10) char literal vs char constants

Yes, char literal and char constants represent the same thing.

Interestingly, when I was researching this I got the following doing a Google search for: ("string literal") OR ("string constant")

https://en.wikipedia.org/wiki/String_literal
A string literal or anonymous string is the representation of a string value within the source code ..... Among other things, it must be possible to encode the character that normally terminates the string constant, plus there must be some way to ...

Not sure if the page should be updated regarding this.

--Ed Davis (talk) 16:10, 15 August 2016 (UTC)

Output Format[edit]

Thanks for the clarifications. With that out the way, two questions about apparent inconsistencies in the output format:

  1. Why are character literals fed through the evaluator and their calculated value printed, but string literals are printed "raw", exactly as they appeared in the code?
    line    18  col    26 Integer         32
    line     4  col     7 String   "Hello, World!\n"

    I suppose this is arbitrary, and done because of potential newlines in strings – if so, wouldn't it be easier to remove support for \n, and let the lexer call its evaluator on strings as well, so that it can print the actual plain string value in the output? So the second line (without the \n) would look like:

    line     4  col     7 String   Hello, World!

    Maybe require a single TAB as separator between the output fields to avoid whitespace ambiguity, and maybe replace the print built-in with println so strings can be printed on their own line.

  2. The output consists of four data fields (line number, column number, token type, token value) - how come only the first two fields have a label included in the output? Shouldn't either none or all fields include the label? i.e.:
    4	7	String	Hello, World!
    line	4	col	7	token	String	value	Hello, World!

--Smls (talk) 16:35, 15 August 2016 (UTC)

1) Yes, it is arbitrary. There isn't a right way to do this, as far as I'm aware. You could do it either way I suppose. However, since character literals are really just small integer constants, that is what the parser needs, and so they have to be converted somewhere. Since the scanner readily has that information, it is easy to do the conversion there. Again, there isn't a right a wrong way. This is just the way I've always done it. For the integration of the scanner with the parser, this seemed to be the easier way. I personally like the way it is done now :-) Again, just my opinion. I hope I have answered the question(s)!

2) I would rather remove the "line" and "col" labels than add additional ones. Consider it done. It will actually make some of the code a tad shorter :-)

--Ed Davis (talk) 17:32, 15 August 2016 (UTC)

Token names[edit]

Some more bikeshedding. I hope you don't mind it - it's just that once the task goes out of draft status and gets many solutions, it will be very annoying to make changes, so now is the time to get it in as good a shape as possible.

What caught my eye regarding the token names is, that the extensive use of uncommon abbreviations makes them less "user-friendly" to deal with than they could be. For this task, itself, it doesn't matter too much, but when it gets to syntax analysis, I image people who will try to write solutions will find themselves saying like "Huh, 'Lss Ident Semi', what tokens were those again?" and having to look it up. How would you feel about renaming them to the following slightly more verbose but friendlier names?

Old New
Mul OP_MULTIPLY
Div OP_DIVIDE
Add OP_ADD
Sub OP_SUBTRACT
Uminus OP_NEGATE
Lss OP_LESSEQUAL
Leq OP_LESS
Gtr OP_GREATER
Neq OP_NOTEQUAL
Assign OP_ASSIGN
And OP_AND
Old New
If KEYWORD_IF
While KEYWORD_WHILE
Print KEYWORD_PRINT
Putc KEYWORD_PUTC
Old New
Lparen LEFTPAREN
Rparen RIGHTPAREN
Lbrace LEFTBRACE
Rbrace RIGHTBRACE
Semi SEMICOLON
Comma COMMA
Old New
Ident IDENTIFIER
Integer INTEGER
String STRING

--Smls (talk) 17:04, 15 August 2016 (UTC)

These names came from textbooks and other compiler source code.

Interestingly, I remember when I had my compilers course back in 1978, I could not figure out what "Lss" stood for :-) I guess I've been using those names for so long that now they are second nature.

I can change them.

--Ed Davis (talk) 17:46, 15 August 2016 (UTC)

I have updated the token names, as well as remove the "line" and "col" labels. --Ed Davis (talk) 21:04, 15 August 2016 (UTC)

What is next?[edit]

What is left to do, to make this a non-draft task? Thanks!

--Ed Davis (talk) 13:39, 16 August 2016 (UTC)

It's customary to wait until a few people other than the task author have added a solution, but I don't think that's a hard rule. I'm writing a Perl solution right now, btw. --Smls (talk) 12:36, 17 August 2016 (UTC)

I was hoping someone would do a Perl solution. I've only done a very little with Perl - basically maintain a few scripts over the years - but I was always impressed with it. Cool! :-)

--Ed Davis (talk) 12:53, 17 August 2016 (UTC)

Wow! I really like your Perl solution. I don't understand much of it, but what I do understand is really cool. --Ed Davis (talk) 10:28, 18 August 2016 (UTC)

Thanks! :) I tried to be clever and dynamically construct a single regex (with one branch per token) to act as the scanner, since it's safe to assume that the Perl regex engine is more bug-free and better optimized than a substr-based scanner that I could have written by hand. But then I realized that there's no easy way to get the line and column number of a regex match, so I had to scan and accumulate those separately, which introduced overhead again. I wonder if the approach was still worth it, performance-wise. Not that a solution in an interpreted language like Perl could ever compete with the C solution, but it might be interesting to benchmark it against the Python solution for large input files... --Smls (talk) 17:06, 18 August 2016 (UTC)

It's easy to get line and column numbers out of a regex. See the Alternate. --Tybalt89 (talk) 14:10, 24 May 2018 (UTC)

Simple benchmark[edit]

I ran some simple benchmarks, using a source file consisting of the following two programs, repeated over and over, until I got to 248,880 lines.

 
count = 1;
n = 1;
limit = 100;
while (n < limit) {
k=3;
p=1;
n=n+2;
while ((k*k<=n) && (p)) {
p=n/k*k!=n;
k=k+2;
}
if (p) {
print(n, " ");
count = count + 1;
}
}
print(count, "\n");
 
{
/*
This is an integer ascii Mandelbrot generator
*/

left_edge = -420;
right_edge = 300;
top_edge = 300;
bottom_edge = -300;
x_step = 7;
y_step = 15;
 
max_iter = 200;
 
y0 = top_edge;
while (y0 > bottom_edge) {
x0 = left_edge;
while (x0 < right_edge) {
y = 0;
x = 0;
the_char = ' ';
i = 0;
while (i < max_iter) {
x_x = (x * x) / 200;
y_y = (y * y) / 200;
if (x_x + y_y > 800 ) {
the_char = '0' + i;
if (i > 9) {
the_char = '@';
}
i = max_iter;
}
y = x * y / 100 + y0;
x = x_x - y_y + x0;
i = i + 1;
}
putc(the_char);
x0 = x0 + x_step;
}
putc('\n');
y0 = y0 - y_step;
}
}
 

I ran them as follows:

timer python lex.py big.t >foo.bar

So the startup time for Python, Perl, and Euphoria is also included in the timings.

All the output files were 1,101,601 lines in length.

I ran each test 3 times, and took the shortest run.

Here are the specs for my machine:

Windows 7, Service Pack 1, 64-bit

Intel Core i7-3720QM CPU @2.60GHz

16.0 GB (15.9 usable)


Processor Time
C(1) 1.08
Flex 1.13
C 1.34
Euphoria 4.15
Perl 8.36
Python 9.24

(1) I swapped out getc(fp) with _fgetc_nolock(fp), and added setvbuf(fp, NULL, _IOFBF, 1024*1024).

To me, the Euphoria, Perl and Python times are very impressive.

Indeed, I would have expected a larger difference between C and the interpreted languages. --Smls (talk) 13:23, 19 August 2016 (UTC)

Future[edit]

My goal is to add the following related tasks:

Syntax Analysis
this is basically a parser that outputs a textural parse tree
Code Generation
code generation for a simple stack based virtual machine - outputs stack vm assembly code
Virtual Machine
virtual machine code interpreter - interprets the vm assembly code

I have already implemented all 3 of these in C and Python. I can do something like:

Given the following program:

count = 1;
while (count < 10) {
    print("count is: ", count, "\n");
    count = count + 1;
}

Running:

lex count.t | parse

Will output a parse tree in textural format:

Sequence
Sequence
;
Assign
Identifier     count
Integer        1
While
Less
Identifier     count
Integer        10
Sequence
Sequence
;
Sequence
Sequence
Sequence
;
Prts
String         "count is: "
;
Prti
Identifier     count
;
Prts
String         "\n"
;
Assign
Identifier     count
Add
Identifier     count
Integer        1

Running:

lex count.t | parse | gen

Will output the following virtual assembly code:

Datasize: 1 Strings: 2
"count is: "
"\n"
    0 push  1
    5 store [0]
   10 fetch [0]
   15 push  10
   20 lt
   21 jz     (43) 65
   26 push  0
   31 prts
   32 fetch [0]
   37 prti
   38 push  1
   43 prts
   44 fetch [0]
   49 push  1
   54 add
   55 store [0]
   60 jmp    (-51) 10
   65 halt

Running:

lex count.t | parse | gen | vm

And it will output the following.

count is: 1
count is: 2
count is: 3
count is: 4
count is: 5
count is: 6
count is: 7
count is: 8
count is: 9

And, I can mix and match - I can use the C lexer, the Python parser, the C code generator, and the Python vm, if that makes sense.

I've already started the write-ups for these. I'll make them draft tasks. Question: What is the protocol, e.g., is it acceptable to post very-rough draft work, in order to solicit feedback? Or should I wait until I have it specified more clearly?

--Ed Davis (talk) 15:01, 13 September 2016 (UTC)

Status[edit]

I vetted this on 2 programming forums/mailing lists, and 2 compiler specific forums. Got feedback, and got two additional solutions! I believe it is ready to go! --Ed Davis (talk) 10:55, 22 October 2016 (UTC)