HowtheTokenizerWorks
Understanding the Tokenizer is not for the faint-hearted. It is by far the most complex and twisty piece
of perl I've ever written that is actually still built properly and isn't a terrible spaghetti-like mess.
In fact, you probably want to skip this section.
But if you really want to understand, well then here goes.
SourceInputandCleanUp
The Tokenizer starts by taking source in a variety of forms, sucking it all in and merging into one big
string, and doing our own internal line split, using a "universal line separator" which allows the
Tokenizer to take source for any platform (and even supports a few known types of broken newlines caused
by mixed mac/pc/*nix editor screw ups).
The resulting array of lines is used to feed the tokenizer, and is also accessed directly by the heredoc-
logic to do the line-oriented part of here-doc support.
DoingThingstheOldFashionedWay
Due to the complexity of perl, and after 2 previously aborted parser attempts, in the end the tokenizer
was fashioned around a line-buffered character-by-character method.
That is, the Tokenizer pulls and holds a line at a time into a line buffer, and then iterates a cursor
along it. At each cursor position, a method is called in whatever token class we are currently in, which
will examine the character at the current position, and handle it.
As the handler methods in the various token classes are called, they build up an output token array for
the source code.
Various parts of the Tokenizer use look-ahead, arbitrary-distance look-behind (although currently the
maximum is three significant tokens), or both, and various other heuristic guesses.
I've been told it is officially termed a "backtrackingparserwithinfinitelookaheads".
StateVariables
Aside from the current line and the character cursor, the Tokenizer maintains a number of different state
variables.
Current Class
The Tokenizer maintains the current token class at all times. Much of the time is just going to be
the "Whitespace" class, which is what the base of a document is. As the tokenizer executes the
various character handlers, the class changes a lot as it moves a long. In fact, in some instances,
the character handler may not handle the character directly itself, but rather change the "current
class" and then hand off to the character handler for the new class.
Because of this, and some other things I'll deal with later, the number of times the character
handlers are called does not in fact have a direct relationship to the number of actual characters in
the document.
Current Zone
Rather than create a class stack to allow for infinitely nested layers of classes, the Tokenizer
recognises just a single layer.
To put it a different way, in various parts of the file, the Tokenizer will recognise different
"base" or "substrate" classes. When a Token such as a comment or a number is finalised by the
tokenizer, it "falls back" to the base state.
This allows proper tokenization of special areas such as __DATA__ and __END__ blocks, which also
contain things like comments and POD, without allowing the creation of any significant Tokens inside
these areas.
For the main part of a document we use PPI::Token::Whitespace for this, with the idea being that code
is "floating in a sea of whitespace".
Current Token
The final main state variable is the "current token". This is the Token that is currently being built
by the Tokenizer. For certain types, it can be manipulated and morphed and change class quite a bit
while being assembled, as the Tokenizer's understanding of the token content changes.
When the Tokenizer is confident that it has seen the end of the Token, it will be "finalized", which
adds it to the output token array and resets the current class to that of the zone that we are
currently in.
I should also note at this point that the "current token" variable is optional. The Tokenizer is
capable of knowing what class it is currently set to, without actually having accumulated any
characters in the Token.
MakingItFaster
As I'm sure you can imagine, calling several different methods for each character and running regexes and
other complex heuristics made the first fully working version of the tokenizer extremely slow.
During testing, I created a metric to measure parsing speed called LPGC, or "lines per gigacycle" . A
gigacycle is simple a billion CPU cycles on a typical single-core CPU, and so a Tokenizer running at
"1000 lines per gigacycle" should generate around 1200 lines of tokenized code when running on a 1200 MHz
processor.
The first working version of the tokenizer ran at only 350 LPGC, so to tokenize a typical large module
such as ExtUtils::MakeMaker took 10-15 seconds. This sluggishness made it unpractical for many uses.
So in the current parser, there are multiple layers of optimisation very carefully built in to the basic.
This has brought the tokenizer up to a more reasonable 1000 LPGC, at the expense of making the code quite
a bit twistier.
MakingItFaster-WholeLineClassification
The first step in the optimisation process was to add a hew handler to enable several of the more basic
classes (whitespace, comments) to be able to be parsed a line at a time. At the start of each line, a
special optional handler (only supported by a few classes) is called to check and see if the entire line
can be parsed in one go.
This is used mainly to handle things like POD, comments, empty lines, and a few other minor special
cases.
MakingItFaster-Inlining
The second stage of the optimisation involved inlining a small number of critical methods that were
repeated an extremely high number of times. Profiling suggested that there were about 1,000,000
individual method calls per gigacycle, and by cutting these by two thirds a significant speed improvement
was gained, in the order of about 50%.
You may notice that many methods in the "PPI::Tokenizer" code look very nested and long hand. This is
primarily due to this inlining.
At around this time, some statistics code that existed in the early versions of the parser was also
removed, as it was determined that it was consuming around 15% of the CPU for the entire parser, while
making the core more complicated.
A judgment call was made that with the difficulties likely to be encountered with future planned
enhancements, and given the relatively high cost involved, the statistics features would be removed from
the Tokenizer.
MakingItFaster-QuoteEngine
Once inlining had reached diminishing returns, it became obvious from the profiling results that a huge
amount of time was being spent stepping a char at a time though long, simple and "syntactically boring"
code such as comments and strings.
The existing regex engine was expanded to also encompass quotes and other quote-like things, and a
special abstract base class was added that provided a number of specialised parsing methods that would
"scan ahead", looking out ahead to find the end of a string, and updating the cursor to leave it in a
valid position for the next call.
This is also the point at which the number of character handler calls began to greatly differ from the
number of characters. But it has been done in a way that allows the parser to retain the power of the
original version at the critical points, while skipping through the "boring bits" as needed for
additional speed.
The addition of this feature allowed the tokenizer to exceed 1000 LPGC for the first time.
MakingItFaster-The"Complete"Mechanism
As it became evident that great speed increases were available by using this "skipping ahead" mechanism,
a new handler method was added that explicitly handles the parsing of an entire token, where the
structure of the token is relatively simple. Tokens such as symbols fit this case, as once we are passed
the initial sigil and word char, we know that we can skip ahead and "complete" the rest of the token much
more easily.
A number of these have been added for most or possibly all of the common cases, with most of these
"complete" handlers implemented using regular expressions.
In fact, so many have been added that at this point, you could arguably reclassify the tokenizer as a
"hybrid regex, char-by=char heuristic tokenizer". More tokens are now consumed in "complete" methods in a
typical program than are handled by the normal char-by-char methods.
Many of the these complete-handlers were implemented during the writing of the Lexer, and this has
allowed the full parser to maintain around 1000 LPGC despite the increasing weight of the Lexer.
MakingItFaster-PortingToC(InProgress)
While it would be extraordinarily difficult to port all of the Tokenizer to C, work has started on a
PPI::XS "accelerator" package which acts as a separate and automatically-detected add-on to the main PPI
package.
PPI::XS implements faster versions of a variety of functions scattered over the entire PPI codebase, from
the Tokenizer Core, Quote Engine, and various other places, and implements them identically in XS/C.
In particular, the skip-ahead methods from the Quote Engine would appear to be extremely amenable to
being done in C, and a number of other functions could be cherry-picked one at a time and implemented in
C.
Each method is heavily tested to ensure that the functionality is identical, and a versioning mechanism
is included to ensure that if a function gets out of sync, PPI::XS will degrade gracefully and just not
replace that single method.
Install Now
PNG Icons
