Before this commit, the PEG.js parser always created the AST using a plain JavaSctript object, but allthough simple and effective for the job, this method sacrificies performance slightly.
From now on the parser shall call a Node creator. This should help with performance, as well as in the future move some AST helpers into the new AST functions.
Before this commit error looks like (for input `start = break:'a'`)
> Expected "!", "$", "&", "(", "*", "+", ".", "/", "/*", "//", ";", "?", character class, code block, comment, end of line, identifier, literal, or whitespace but ":" found.
After this error looks like
> Label can't be a reserved word "break".
* Optimization: do not generate unreachable calls |peg$fail| and constants for it (local to the rule).
* Optimization: do not generate unreachable calls |peg$fail| and constants for it (non-local to rule).
* Remove stack manipulations from FAIL opcode and rename FAIL to EXPECT
* Always save the expected values regardless of the match result of expression
* Remove unnecessary check for match result in the `named` node
* Collect and process expectations from predicates
Currently, an open brace without a corresponding brace will emit this confusing error message:
> Expected "!", "$", "&", "(", "*", "+", ".", "/", "/*", "//", ";", "?", character class, code block, comment, end of line, identifier, literal, or whitespace but "{" found.
This change adds an error case to the grammar to make it clear what the problem is.
Running ESLint on generated code with the configuration used on PEG.js
itself produces a lot of errors. This commit fixes some unnecessary ones
caught by these rules:
- max-len
- new-cap
- newline-before-return
- no-unused-vars
See also 5dd8e797f7.
Follow-up to #407.
The idea behind linting lib/parser.js was that it would improve quality
of code generated by PEG.js in general. However, there is a couple of
problems with it:
1. Code in lib/parser.js is ES5 while the rest of the code is ES2015.
This would mean a separate ESLint configuration and a separate set
of code style rules just for lib/parser.js once code style checks
are added.
2. Code in lib/parser.js is generated. This means that even today it
violates checks like "no-unused-var", which have to be disabled.
This would get worse once code style checks are added, again
requiring a separate ESLint configuration just for lib/parser.js.
3. Linting lib/parser.js checks only small portion of possible code
generator output. For example, code generated when optimizing for
size or when tracing is not checked at all. Thus, linting
lib/parser.js gives a false sense of security.
Because of these problems I decided not to lint lib/parser.js at all and
rely on ad-hoc linting of parser files produced by PEG.js with ignoring
false-positives. I consider this more of a pragmatic cost vs. benefits
decision than a principial one.
Part of #407.
Before this commit, continuation lines of multi-line values in variable
declaration initializers were aligned with the variable name:
let foo = {
a: 5,
b: 6
};
This was highly irregular, maintenance intensive, and made declarations
look different from assignments.
This commit changes the indentation to be more regular and similar to
assignments:
let foo = {
a: 5,
b: 6
};
The lib/parser.js file is a CommonJS module like all the other files in
lib/, so setting the environment explicitly is not needed. Besides, the
environment set by "eslint-env" was wrong (since transitioning from the
AMD format).
Use one var/let/const per variable, but only for initialized variables.
Uninitialized variables are still grouped into one var/let/const
declaration as I don't see any value in separating them. This approach
reflects the fact that initialized and uninitialized var/let/const
declarations are really two different things.
See #443.
Because arrow functions work rather differently than normal functions (a
bad design mistake if you ask me), I decided to be conservative with the
conversion.
I converted:
* event handlers
* callbacks
* arguments to Array.prototype.map & co.
* small standalone lambda functions
I didn't convert:
* functions assigned to object literal properties (the new shorthand
syntax would be better here)
* functions passed to "describe", "it", etc. in specs (because Jasmine
relies on dynamic "this")
See #442.
This is purely a mechanical change, not taking advantage of block scope
of "let" and "const". Minimizing variable scope will come in the next
commit.
In general, "var" is converted into "let" and "const" is used only for
immutable variables of permanent character (generally spelled in
ALL_CAPS). Using it for any immutable variable regardless on its
permanence would feel confusing.
Any code which is not transpiled and needs to run in ES6 environment
(examples, code in grammars embedded in specs, ...) is kept unchanged.
This is also true for code generated by PEG.js.
See #442.
This change reflects the fact that PEG.js-generated parsers really
produce two kinds of syntax errors:
Structured errors
Caused by match failures, trailing input, or calls of the "expected"
function in parser code. Their messages have fixed format ("Expected
... but ... found.").
Simple errors
Caused by calls of the "error" function in parser code. Their
messages don't have any fixed format.
Each kind of error now has a separate helper function which builds its
instances.
Until now, expectations were constructed using object literals. This
commit changes the construction to use factory functions.
This change makes generated parsers slightly smaller because property
names don't have to be repeated many times and factory function calls
are more amenable to minifying.
Some numbers based on the aggregate size of parsers generated from
examples/*.pegjs:
Optimization Minified? Size before Size after Saving
------------------------------------------------------------
speed no 719066 716063 0.42%
speed yes 188998 180202 4.65%
size no 194810 197813 1.52%
size yes 108782 99947 8.12%
(Minification was done using "uglify --mangle --compress" with
uglify-js 2.4.24.)
If the described class is case-sensitive, nothing changes.
If the described class is case-insensitive, its description doesn't
indicate that anymore. The indication was awkward and it was meaningful
only for parser users familiar with PEG.js grammar syntax (typically a
minority). For cases where case insensitivity indication is vital, named
rules can be used to customize the reporting.
Note that literal descriptions already ignore the case-sensitivity flag;
this commit only makes things consistent.
Simplify regexps that specify ranges of characters to escape with "\xXX"
and "\uXXXX" in various escaping functions. Until now, these regexps
were (mostly) mutually exclusive with more selective regexps applied
before them, but this became a maintenance headache. I decided to
abandon the exclusivity, which allowed to simplify these regexps (at the
cost of introducing an ordering dependency).
Change how found strings are escaped when building syntax error
messages:
* Do not escape non-ASCII characters (U+0100-U+FFFF). They are
typically more readable in their raw form.
* Escape DEL (U+007F). It is a control character.
* Escape NUL (U+0000) as "\0", not "\x00".
* Do not use less known shortcut escape sequences ("\b", "\f"), only the
well-known ones ("\0", "\t", "\n", "\r").
These changes mirror expectation escaping changes done in
4fe682794d.
Part of work on #428.
Before this commit, descriptions of literals used in error messages were
built by applying JavaScript string escaping to their values, making the
descriptions look like JavaScript strings. Descriptions of character
classes were built using their raw text. These approaches were mutually
inconsistent and lead to descriptions which were over-escaped and not
necessarily human-friendly (in case of literals) or coupled with details
of the grammar (in case of character classes).
This commit changes description building code in both cases and unifies
it. The intent is to generate human-friendly descriptions of matched
expressions which are clean, unambiguous, and which don't escape too
many characters, while handling special characters such as newlines
well.
Fixes#127.
I no longer think that using raw literal texts in error messages is the
right thing to do. The main reason is that it couples error messages
with details of the grammar such as use of single or double quotes in
literals. A better solution is coming in the next commit.
This reverts commit 69a0f769fc.
The wrapping functions are also generated by PEG.js, so the comment
should be above them to mark them as such. This shouldn't cause any
problems technically.
Futago-za Ryuu
Mingun
felix
David Majda