The generated parser API specs are mostly extracted from
generated-parser.spec.js, which got renamed to
generated-parser-behavior.spec.js to better reflect its purpose.
Unit specs are unit tests of internal stuff. API specs are tests of the
user-visible APIs and behavior.
I think it makes sense to make this distinction because then the public
API line is more clearly visible e.g. when using the specs as
documentation.
The TEXT instruction now replaces position at the top of the stack with
the input from that position until the current position. This is simpler
and cleaner semantics than the previous one, where TEXT also popped an
additional value from the stack and kept the position there.
Implement the following bytecode instructions:
* PUSH_UNDEFINED
* PUSH_NULL
* PUSH_FAILED
* PUSH_EMPTY_ARRAY
These instructions push simple JavaSccript values to the stack directly,
without going through constants. This makes the bytecode slightly
shorter and the bytecode generator somewhat simpler.
Also note that PUSH_EMPTY_ARRAY allows us to avoid a hack which protects
the [] constant from modification.
The action/predicate code didn't have access to the parser object. This
was mostly a side effect actions/predicates being implemented as nested
functions, in which |this| is a reference to the global object (an ugly
JavaScript quirk). The initializer, being implemented differently, had
access to the parser object via |this|, but this was not documented.
Because having access to the parser object can be useful, this commits
introduces a new |parser| variable which holds a reference to it, is
visible in action/predicate/initializer code, and is properly
documented.
See also:
https://groups.google.com/forum/#!topic/pegjs/Na7YWnz6Bmg
This is mostly done for consistency with the JavaScript example grammar,
from which the |Identifier| rule is taken from. See the previous commit
for details.
Instead of matching segments between blocks character by character,
match them as a whole. Also align the style with other similar rules
(e.g. the comment ones).
Before this commit, line continuations in character classes contributed
an empty string to the list of characters and character ranges matched
by a class. While this didn't lead to a buggy behavior with the current
code generator, the AST was wrong and the problem could have caused bugs
later.
This commit fixes the problem.
Semantic predicates are kind of |PrimaryExpression|, not kind of
|PrefixedExpression|. Therefore I extracted a rule for them and
referenced it from the |PrimaryExpression|.
Initializer and rules are now separated in a similar way as JavaScript
statements -- either by a semicolon or a line terminator, possibly with
whitespace and comments mixed in.
One consequence is that the grammars like this are now illegal:
foo = "a" bar = "b"
A semicolon needs to be inserted between the rules:
foo = "a";bar = "b"
I consider this a good change as the now-illegal syntax was somewhat
confusing.
This makes the |Primary| rule a bit more tidy. Also, matching the |.|
character really belongs to the lexical part of the grammar, next to
literals and character classes.
* Rename the |Action| rule to |CodeBlock| (it better describes what
the rule matches).
* Implement the rule in a simpler way and move it after more basic
lexical elements.
This change has two side effects:
* Label names can no longer be JavaScript reserved words.
* |$| is allowed again in label names. However, because of the
preference rules, names starting with it will be usually parsed as a
text operator followed by another identifier (denoting a rule
reference or label name).
Before this commit, whitespace was handled at the lexical level by
making tokens consume any whitespace coming after them. This was
accomplished by appending |__| to every token rule.
This commit changes whitespace handling to be more explicit. Tokens no
longer consume whitespace coming after them and syntactic rules have to
cope with it. While this slightly complicates the syntactic grammar, I
think it's a cleaner way. Moreover, it is what JavaScript example
grammar does.
One small side-effect of thich change is that the grammar is now
stand-alone (it doesn't require utils.js anymore).
When rule names are capitalized, it's easier to visually distinguish
them from non-capitalized label names. Moreover, I use capitalized rule
names in all my grammars these days.
Before this commit, a line continuation (backslash followed by a line
terminator character) contributed a character to a string or a character
class it was used in. In JavaScript and many other languages, line
continuation doesn't contribute anything.
This commit aligns PEG.js line continuation behavior with JavaScript.
Before this commit, the value of the |rawText| property of "class" AST
nodes was created in a hackish way from processed input and it didn't
always exactly represent the actual input text.
This commit changes the code so that the value of the |rawText| property
is created using the |text| function. This is a clean way which also
resolves the exact representation problem.
Also added few missing |hasOwnProperty| calls that JSHint didn't detect
because it only looks whether there is an |if| statement wrapping the
loop body.
In case the generated parser parsed successfully part of input and left
some input unparsed (trailing input), the error message produced was
sometimes wrong. The code worked correctly only if there were no match
failures in the successfully parsed part (highly unlikely).
This commit fixes things by explicitly triggering a match failure with the
following expectation at the end of the successfully parsed part of the
input:
peg$fail({ type: "end", description: "end of input" });
This change also made it possible to simplify the |buildMessage|
function, which can now ignore the case of no expectations.
Fixes#119.
The error check was useful when actions could have returned |null| to
trigger a match failure. This is no longer supported so the check isn't
needed anymore.
Speed impact
------------
Before: 1022.70 kB/s
After: 1035.45 kB/s
Difference: 1.24%
Size impact
-----------
Before: 975434 b
After: 931540 b
Difference: -4.50%
(Measured by /tools/impact with Node.js v0.6.18 on x86_64 GNU/Linux.)
Before this commit, the |expected| and |error| functions didn't halt the
parsing immediately, but triggered a regular match failure. After they
were called, the parser could backtrack, try another branches, and only
if no other branch succeeded, it triggered an exception with information
possibly based on parameters passed to the |expected| or |error|
function (this depended on positions where failures in other branches
have occurred).
While nice in theory, this solution didn't work well in practice. There
were at least two problems:
1. Action expression could have easily triggered a match failure later
in the input than the action itself. This resulted in the
action-triggered failure to be shadowed by the expression-triggered
one.
Consider the following example:
integer = digits:[0-9]+ {
var result = parseInt(digits.join(""), 10);
if (result % 2 === 0) {
error("The number must be an odd integer.");
return;
}
return result;
}
Given input "2", the |[0-9]+| expression would record a match
failure at position 1 (an unsuccessful attempt to parse yet another
digit after "2"). However, a failure triggered by the |error| call
would occur at position 0.
This problem could have been solved by silencing match failures in
action expressions, but that would lead to severe performance
problems (yes, I tried and measured). Other possible solutions are
hacks which I didn't want to introduce into PEG.js.
2. Triggering a match failure in action code could have lead to
unexpected backtracking.
Consider the following example:
class = "[" (charRange / char)* "]"
charRange = begin:char "-" end:char {
if (begin.data.charCodeAt(0) > end.data.charCodeAt(0)) {
error("Invalid character range: " + begin + "-" + end + ".");
}
// ...
}
char = [a-zA-Z0-9_\-]
Given input "[b-a]", the |charRange| rule would fail, but the
parser would try the |char| rule and succeed repeatedly, resulting
in "b-a" being parsed as a sequence of three |char|'s, which it is
not.
This problem could have been solved by using negative predicates,
but that would complicate the grammar and still wouldn't get rid of
unintuitive behavior.
Given these problems I decided to change the semantics of the |expected|
and |error| functions. They don't interact with regular match failure
mechanism anymore, but they cause and immediate parse failure by
throwing an exception. I think this is more intuitive behavior with less
harmful side effects.
The disadvantage of the new approach is that one can't backtrack from an
action-triggered error. I don't see this as a big deal as I think this
will be rarely needed and one can always use a semantic predicate as a
workaround.
Speed impact
------------
Before: 993.84 kB/s
After: 998.05 kB/s
Difference: 0.42%
Size impact
-----------
Before: 1019968 b
After: 975434 b
Difference: -4.37%
(Measured by /tools/impact with Node.js v0.6.18 on x86_64 GNU/Linux.)
The |error| function allows users to report custom match failures inside
actions.
If the |error| function is called, and the reported match failure turns
out to be the cause of a parse error, the error message reported by the
parser will be exactly the one specified in the |error| call.
Implements part of #198.
Speed impact
------------
Before: 999.83 kB/s
After: 1000.84 kB/s
Difference: 0.10%
Size impact
-----------
Before: 1017212 b
After: 1019968 b
Difference: 0.27%
(Measured by /tools/impact with Node.js v0.6.18 on x86_64 GNU/Linux.)
The |expected| function allows users to report regular match failures
inside actions.
If the |expected| function is called, and the reported match failure
turns out to be the cause of a parse error, the error message reported
by the parser will be in the usual "Expected ... but found ..." format
with the description specified in the |expected| call used as part of
the message.
Implements part of #198.
Speed impact
------------
Before: 1146.82 kB/s
After: 1031.25 kB/s
Difference: -10.08%
Size impact
-----------
Before: 950817 b
After: 973269 b
Difference: 2.36%
(Measured by /tools/impact with Node.js v0.6.18 on x86_64 GNU/Linux.)
After making the |?| operator return |null| instead of an empty string
in the previous commit, empty strings were still returned from
predicates. This didn't make much sense.
Return value of a predicate is unimportant (if you have one in hand, you
already know the predicate succeeded) and one could even argue that
predicates shouldn't return any value at all. The closest thing to
"return no value" in JavaScript is returning |undefined|, so I decided
to make predicates return exactly that.
Implements part of #198.
Before this commit, the |?| operator returned an empty string upon
unsuccessful match. This commit changes the returned value to |null|. It
also updates the PEG.js grammar and the example grammars, which used the
value returned by |?| quite often.
Returning |null| is possible because it no longer indicates a match
failure.
I expect that this change will simplify many real-world grammars, as an
empty string is almost never desirable as a return value (except some
lexer-level rules) and it is often translated into |null| or some other
value in action code.
Implements part of #198.
Using a special value to indicate match failure instead of |null| allows
actions to return |null| as a regular value. This simplifies e.g. the
JSON parser.
Note the special value is internal and intentionally undocumented. This
means that there is currently no official way how to trigger a match
failure from an action. This is a temporary state which will be fixed
soon.
The negative performance impact (see below) is probably caused by
changing lot of comparisons against |null| (which likely check the value
against a fixed constant representing |null| in the interpreter) to
comparisons against the special value (which likely check the value
against another value in the interpreter).
Implements part of #198.
Speed impact
------------
Before: 1146.82 kB/s
After: 1031.25 kB/s
Difference: -10.08%
Size impact
-----------
Before: 950817 b
After: 973269 b
Difference: 2.36%
(Measured by /tools/impact with Node.js v0.6.18 on x86_64 GNU/Linux.)
Before this commit, the |expected| property of an exception object
thrown when a generated parser encountered an error contained
expectations as strings. These strings were in a human-readable format
suitable for displaying in the UI but not suitable for machine
processing. For example, expected string literals included quotes and a
string "any character" was used when any character was expected.
This commit makes expectations structured objects. This makes the
machine processing easier, while still allowing to generate a
human-readable representation if needed.
Implements part of #198.
Speed impact
------------
Before: 1180.41 kB/s
After: 1165.31 kB/s
Difference: -1.28%
Size impact
-----------
Before: 863523 b
After: 950817 b
Difference: 10.10%
(Measured by /tools/impact with Node.js v0.6.18 on x86_64 GNU/Linux.)
The deduplication skipped over an expected string right after the one
that was removed because the index variable was incorrectly incremented
in that case.
Based on a patch by @fresheneesz:
https://github.com/dmajda/pegjs/pull/146
The compiler passes are now split into three stages:
* check -- passes that check for various error conditions
* transform -- passes that transform the AST (e.g. to perform
optimizations)
* generate -- passes that are related to code generation
Splitting the passes into stages is important for plugins. For example,
if a plugin wants to add a new optimization pass, it can add it at the
end of the "transform" stage without any knowledge about other passes it
contains. Similarly, if it wants to generate something else than the
default code generator does from the AST, it can just replace all passes
in the "generate" stage by its own one(s).
More generally, the stages make it possible to write plugins that do not
depend on names and actions of specific passes (which I consider
internal and subject of change), just on the definition of stages (which
I consider a public API with to which semver rules apply).
Implements part of GH-106.
Fixes the following JSHint errors (which I think are JSHint bugs):
spec/parser.spec.js: line 142, col 20, Bad for in variable 'key'.
spec/generated-parser.spec.js: line 119, col 20, Bad for in variable 'key'.
This is a complete rewrite of the PEG.js code generator. Its goals are:
1. Allow optimizing the generated parser code for code size as well as
for parsing speed.
2. Prepare ground for future optimizations and big features (like
incremental parsing).
2. Replace the old template-based code-generation system with
something more lightweight and flexible.
4. General code cleanup (structure, style, variable names, ...).
New Architecture
----------------
The new code generator consists of two steps:
* Bytecode generator -- produces bytecode for an abstract virtual
machine
* JavaScript generator -- produces JavaScript code based on the
bytecode
The abstract virtual machine is stack-based. Originally I wanted to make
it register-based, but it turned out that all the code related to it
would be more complex and the bytecode itself would be longer (because
of explicit register specifications in instructions). The only downsides
of the stack-based approach seem to be few small inefficiencies (see
e.g. the |NIP| instruction), which seem to be insignificant.
The new generator allows optimizing for parsing speed or code size (you
can choose using the |optimize| option of the |PEG.buildParser| method
or the --optimize/-o option on the command-line).
When optimizing for size, the JavaScript generator emits the bytecode
together with its constant table and a generic bytecode interpreter.
Because the interpreter is small and the bytecode and constant table
grow only slowly with size of the grammar, the resulting parser is also
small.
When optimizing for speed, the JavaScript generator just compiles the
bytecode into JavaScript. The generated code is relatively efficient, so
the resulting parser is fast.
Internal Identifiers
--------------------
As a small bonus, all internal identifiers visible to user code in the
initializer, actions and predicates are prefixed by |peg$|. This lowers
the chance that identifiers in user code will conflict with the ones
from PEG.js. It also makes using any internals in user code ugly, which
is a good thing. This solves GH-92.
Performance
-----------
The new code generator improved parsing speed and parser code size
significantly. The generated parsers are now:
* 39% faster when optimizing for speed
* 69% smaller when optimizing for size (without minification)
* 31% smaller when optimizing for size (with minification)
(Parsing speed was measured using the |benchmark/run| script. Code size
was measured by generating parsers for examples in the |examples|
directory and adding up the file sizes. Minification was done by |uglify
--ascii| in version 1.3.4.)
Final Note
----------
This is just a beginning! The new code generator lays a foundation upon
which many optimizations and improvements can (and will) be made.
Stay tuned :-)
When called inside an action, the |text| function returns the text
matched by action's expression. It can be also called inside an
initializer or a predicate where it returns an empty string.
The |text| function will be useful mainly in cases where one needs a
structured representation of the input and simultaneously the raw text.
Until now, the only way to get the raw text in these cases was to
painfully build it from the structured representation.
Fixes GH-131.
Implement a new syntax to extract matched strings from expressions. For
example, instead of:
identifier = first:[a-zA-Z_] rest:[a-zA-Z0-9_]* { return first + rest.join(""); }
you can now just write:
identifier = $([a-zA-Z_] [a-zA-Z0-9_]*)
This is useful mostly for "lexical" rules at the bottom of many
grammars.
Note that structured match results are still built for the expressions
prefixed by "$", they are just ignored. I plan to optimize this later
(sometime after the code generator rewrite).
Getting rid of the |trackLineAndColumn| simplifies the code generator
(by unifying two paths in the code).
The |line| and |column| functions currently always compute all the
position info from scratch, which is horribly ineffective. This will be
improved in later commit(s).
This will allow to compute position data lazily and get rid of the
|trackLineAndColumn| option without affecting performance of generated
parsers that don't use position data.
Before this commit, |PEG.buildParser| always returned a parser object.
The only way to get its source code was to call the |toSource| method on
it. While this method worked for parsers produced by |PEG.buildParser|
directly, it didn't work for parsers instantiated by executing their
source code. In other words, it was unreliable.
This commit remvoes the |toSource| method on generated parsers and
introduces a new |output| option to |PEG.buildParser|. It allows callers
to specify whether they want to get back the parser object
(|options.output === "parser"|) or its source code (|options.output ===
"source"|). This is much better and more reliable API.
Includes:
* Moving the source code from /src to /lib.
* Adding an explicit file list to package.json
* Updating the Makefile.
* Updating the spec and benchmark suites and their READMEs.
Part of a fix for GH-32.
Before this commit, generated parser were able to start parsing from any
rule. This was nice, but it made rule code inlining impossible.
Since this commit, the list of allowed start rules has to be specified
explicitly using the |allowedStartRules| option of the |PEG.buildParser|
method (or the --allowed-start-rule option on the command-line). These
rules will be excluded from inlining when it's implemented.
This commit replaces the |startRule| parameter of the |parse| method in
generated parsers with more generic |options| -- an options object. This
options object can be used to pass custom options to the parser because
it is visible as the |options| variable inside parser code.
The start rule can now be specified as the |startRule| option. This
means you have to replace all calls like:
parser.parse("input", "myStartRule");
with
parser.parse("input", { startRule: "myStartRule" });
Closes GH-37.
The purpose of this change is to avoid the need to index register
variables storing match results of sequences whose elements are labeled.
The indexing happened when match results of labeled elements were passed
to action/predicate functions.
In order to avoid indexing, the register allocator needs to ensure that
registers storing match results of any labeled sequence elements are
still "alive" after finishing parsing of the sequence. They should not
be used to store anything else at least until code of all actions and
predicates that can see the labels is executed. This requires that the
|allocateRegisters| pass has the knowledge of scoping. Because that
knowledge was already implicitly embedded in the |coputeParams| pass,
the logical step to prevent duplication was to merge it with the
|allocateRegisters| pass. This is what this commit does.
As a part of the merge the tests of both passes were largely refactored.
This is both to accomodate the merge and to make the tests in sync with
the code again (the tests became a bit out-of-sync during the last few
commits -- they tested more than was needed).
The speed/size impact is slightly positive:
Speed impact
------------
Before: 849.86 kB/s
After: 858.16 kB/s
Difference: 0.97%
Size impact
-----------
Before: 876618 b
After: 875602 b
Difference: -0.12%
(Measured by /tools/impact with Node.js v0.6.18 on x86_64 GNU/Linux.)
This commit changes the model underlying parser variables used to store
match results and parse positions. Until now they were treated as a
stack, now they are thought of as registers. The actual behavior does
not change (yet), only the terminology.
More specifically, this commit:
* Changes parser variable names from |result0|, |result1|, etc. to
|r0|, |r1|, etc.
* Changes various internal names and comments to match the new model.
* Renames the |computeVarIndices| pass to |allocateRegisters|.
One stack is conceptually simpler, requires less code and will make a
transition to a register-based machine easier.
Note that the stack variables are now named a bit incorrectly
(|result0|, |result1|, etc. even when they store also parse positions).
I didn't bother with renaming because a transition to a register-based
machine will follow soon and the names will change anyway.
The speed/size impact is insignificant.
Speed impact
------------
Before: 839.05 kB/s
After: 839.67 kB/s
Difference: 0.07%
Size impact
-----------
Before: 949783 b
After: 961578 b
Difference: 1.24%
(Measured by /tools/impact with Node.js v0.6.18 on x86_64 GNU/Linux.)
This commit replaces all variable name computations in |computeVarNames|
and |computeParams| passes by computations of indices. The actual names
are computed later in the |generateCode| pass.
This change makes the code generator the only place that deals with the
actual variable names, making them easier to change for example.
The code generator code seems bit more complicated after the change, but
this complexity will pay off (and mostly disappear) later.
Places all code that does something with "action" AST nodes under code
handling "choice" nodes.
This ordering is logical because now all the node handling code matches
the sequence in which various node types usually appear when descending
through the AST tree.
Changes all code that does something with "literal", "class" or "any"
AST nodes so that the code deals with these in the follwing order:
1. literal
2. class
3. any
Previously the code used this ordering:
1. literal
2. any
3. class
The new ordering is more logical because the nodes are handled from the
most specific to the most generic.
PEG.js grammar rules are represented by |rule| nodes in the AST. Until
now, all such nodes had a |displayName| property which was either |null|
or stored rule's human-readable name. This commit gets rid of the
|displayName| property and starts representing rules with a
human-readable name using a new |named| node (a child of the |rule|
node).
This change simplifies code generation code a bit as tests for
|displayName| can be removed (see changes in generate-code.js). It also
separates different concerns from each other nicely.
This change makes code using |oneRuleGrammar| less verbose + prepares
for passing of the initializer (will be used by code added in the next
few commits).
David Majda
David Majda
Andrei Neculau