Coding Gorilla

I’ve pulled all the Antlr parse tree editing routines into its own library yesterday. This library fills a gap between what is offered in the Antlr runtime (a psuedo XPath library, AddChild and Parent accessors for ParserRuleContext), and a full-blown transformation system like ASF+SDF.

This library contains:

• a beefed-up XPath version 2 library;
• a tree construction routine from an s-expression notation and an Antlr parser and lexer.
• Antlr parse tree node replace and delete;
• an exhanced parse tree node type that supports an observer pattern to keep data structures in sync when you modify a parse tree.

Right now it’s just in C#, but I plan at some point to translate it to Java because it is very useful.

–Ken

This is just a note to myself regarding some ideas on a command-line tool for Antlr grammars. It should be clear to anyone: grammars, especially Antlr grammars, are first-class objects that can be manipulated and changed to improve readability or performance. Antlrvsix now incorporates almost two dozen transformations, ranging from unfold, fold, sort rules, remove useless rules, reformat rules, split grammars, merge grammars, etc.

The question is how best to structure these transformations for the extension, and more importantly, in a completely automated manner whereby I can read the grammar from a web page containing a spec of a particular language like Java, C#, or what have you.

The main problem in the tool is how to identify the parts of the grammar that I want to change with a transform. Here, it looks like there are two possibilities: (1) a line/column number range; or (2) a handy XPath expression(s) to identify a point(s) or range(s) in the grammar.

So, to make changes to the grammar, I could envision something like this:

cat Grammar.g4 | trash "//ruleSpec[/RULE_SPEC = 'e'] => unfold" | trash "=> split-grammar" 1> GrammarParser.g4 2> GrammarLexer.g4

Afterwards, I can write an online version of the Antlr transformation tool.

Also to note to myself this article by Figueira et al. is the only one that I found that describes the denotational semantics of XPath in an unambiguous manner. Can fold be implemented using XPath, where one entire sub-tree (implemented I suppose as a node-set) can be compared to another sub-tree?

–Ken

This is a note of an idea I posted in two Twitter threads (here, here, and here). I think it’s important to capture the idea before it gets lost when blogs and Twitter disappear.

The problem with “semantic highlighting”, or what I would just call syntactic highlighting because there are really many levels of highlighting based on lexical, CFG, static and dynamic semantics, is that it’s nearly impossible for a programmer to augment his editor with rules to perform the type of check he wants. TextMate highlights the lexical syntax of a program. LSP “semantic highlighting” considers the static semantics of the program. But, if you would like the editor to highlight something more interesting, like the live/dead analysis of a variable, or constant propagation, you’re basically out of luck.

Parsing entire parse trees does not solve the problem of identifying parts of the program that you are interested in. You are only interested in paths through the tree. XPath is the best solution here.

With a grammar and a parse tree decorated with the results of semantic analysis, many types of highlighting are now possible using an XPath-based solution. For example, using Antlr’s notation for lexical and CFG symbols, comments could be tagged with “//COMMENT => green”, keywords tagged with “//keyword => blue”, and fields tagged with “//field_declaration//variable_declarator/identifier => pink”. To employ a new highlighting, one would simply tell the editor to re-tag the text using a new collection of rules.

The only problem with this idea is implementing the static semantics for the problem you are interested in.

–Ken

I’ve finally have the right tools to now implement transforms over an Antlr parse tree.

The first part of a transform is identifying what nodes in the tree that are going to be replaced. It turns out that the best tool to do that is an XPath engine, which I’ve rewritten in C# from Java over the last month.

The other part of a transform–and far simpler to implement–is a way to express a tree that is created and spliced into an existing parse tree. Here, the work I did in Piggy for S-expressions comes in handy. For example, the expression “( ruleAltList ( labeledAlt ( alternative { child } )))” identifies the Antlr parse tree to create, and splices in a “child” node into a created “alternative” node. Note, this is a slightly different usage for “S-expressions” that you may be used to, in which a node is an unnamed pair, but conveys the same purpose.

With XPath and S-expressions, I can now rewrite all the transforms that I hardcoded in C# for Antlrvsix. The code implementing both parts is here, but I will be forking this code and placing it under in the Antlrvsix source until I see a need to place this code in a separate Nuget package.

At this moment, I’m not exactly sure what language and control structures to add on top of XPath and S-expressions. For now, these two tools should suffice along with C# to glue the pieces together in order to modify an Antlr parse tree. The other consideration is having intermediate results of an XPath expression. For example, I may want to get all ruleAltList’s but continue down the tree for a particular child. I’ve fixed a bug in the Eclipse XPath library to allow an intermediate result to be used as context for another XPath expression. But I might consider extending XPath to bind the results of an intermediate result into a C# variable.

One other note–Why am I not looking at term rewriting systems? I have. The problem is that they are not practical for two reasons: (1) integrating it with Antlr parse trees would not be easy; (2) most do not express manipulations directly on a tree. XSLT is one example. Here, the language isn’t specifying tree rewrites, but the construction of an entirely new tree. I also looked at TXL. Here, the language isn’t about trees, but term rewrites in the target language. I would need to convert the Antlr grammar into TXL grammar syntax. In all these systems, I would need to fit in Antlr parse trees into the framework. Again, all I want is to manipulate trees.

What of Piggy? Unfortunately, Piggy is not a tree editing library. While it recognizes tree nodes, the problem is that it then executes user code that performs an output on a DFS traversal of the tree. Again, all I want is to manipulate an existing tree, then do something later with that tree.

–Ken

As noted in the Bison guide:

• The associativity of an operator op determines how repeated uses of the operator nest: whether ‘x op y op z’ is parsed by grouping x with y first or by grouping y with z first. %left specifies left-associativity (grouping x with y first) and %right specifies right-associativity (grouping y with z first). %nonassoc specifies no associativity, which means that ‘x op y op z’ is considered a syntax error.%precedence gives only precedence to the symbols, and defines no associativity at all. Use this to define precedence only, and leave any potential conflict due to associativity enabled.
• The precedence of an operator determines how it nests with other operators. All the tokens declared in a single precedence declaration have equal precedence and nest together according to their associativity. When two tokens declared in different precedence declarations associate, the one declared later has the higher precedence and is grouped first.

Associativity

By default, operators in Antlr are left-associative. So, %left associative operators do not require any additional declarations in Antlr.

%left op

=>

e : e op e

An operator declared as %right associative will need to be tagged in Antlr with <assoc=right>.

%right op

=>

e : <assoc=right> e op e

Antlr does not have a %nonassoc declaration. To disallow a non-associative operator in Antlr, a semantic predicate must be used.

%nonassoc op

=>

e : e op e { Input.LA(1) != op }?

Precedence

In Antlr, to make op1 higher precedence than op2, one will need to define the alternative containing op1 before the alternative for op2. In other words, the order of the alternatives of the operators in Antlr is the same as the order of appearance in the Bison grammar.

%left op1
%left op2

=>

e : e op1 e | e op2 e

What are precedence and associativity bound to?

As noted here: “A precedence and associativity is associated with each grammar rule. It is the precedence and associativity of the final token or literal in the body of the rule. If the %prec construction is used, it overrides this default value. Some grammar rules may have no precedence and associativity associated with them.”

Even though precedence and associativity are specified with an operator, the precedence and associativity are associated with a rule as well. Note, this makes sense from an implementation reference: when you see a shift/shift or shift/reduce conflict, you are deciding what to place in the parsing table for the action based on the rule and the lookahead that you see. Crucially, the %prec, %right, %left, %nonassoc clauses can appear within a rule, which overrides any clauses specified elsewhere.

–Ken

“//parserRuleSpec[RULE_REF/text() = ruleBlock/ruleAltList/labeledAlt/alternative/[name()=’element’][1]/atom/ruleref/[1]/text()]”

This XPath expression is the first important rule I wrote for Antlr grammars, which finds all rules that have direct left recursion. Finding direct left recursion is an important step for removing indirect left recursion.

My efforts for getting XPath working with Antlr are starting to finally pay off.

–Ken

It’s taken a few weeks, but the daily grind has resulted in a new grammar in Antlr for XPath version 3.1, and a translation of the Eclipse engine to search the Antlr parse trees (code). This will come in handy to partially replace the hardwired code in C# in Antlrvsix to perform grammar refactorings. There is still much work to be done:

• Polishing the XPath engine to work well.
• Create an expression language to tell the transformation system what nodes found in the parse tree (using the XPath engine) how to rewrite the tree nodes.
• Rewrite the refactorings in Antlrvsix in this new language and engine.

–Ken, June 24, 2020

I’ve made another release of Antlrvsix, version 7.3, last week and I’m almost ready to release 7.4 with a few more fixes. The changes I’ve been making since last year in August have greatly enhanced what Antlrvsix can do over version 1. But, there is more work to be done. I’m now working on updating and integrating Piggy into the extension.

Part of the problem with Piggy was that I lacked a clear vision of what I was trying to solve. I now know what I’m looking to solve because I have worked on the analysis and transformations for Antlr. And, I can see there are actually two levels to this system: (a) a high-level language, that works with snippets in the language that is going to be modified, and (b) a low-level language, that works with tree patterns and routines to manipulate those trees. This was what the authors of Coccinelle did, and it is the right idea because it’s hard writing tree patterns in a language like XPath, which is just a language that specifies a collection of tree nodes.

I will first start writing an interface for XPath because all the transforms use a basic tree node find and replace. I will first try porting Eclipse XPath written in Java to C# and see where that takes me. The port will be here.

–Ken

What is the next step in Antlrvsix to help improve a grammar? To answer that, I decided to go back and take a look at where this all began, with the comparison of the Java grammars in Antlr.

Java9 is derived from version 9 of the Java spec. Parr’s Java is derived from an unknown, earlier version of the spec, but it is much faster in parsing that the other grammar.

How can one compare these grammars? A line-by-line diff of the files could be done, but it is a terrible way to compare the grammars because they are very different: one is a partitioned grammar, the other a combined grammar; the formatting is different; the rules in each grammar are ordered differently; and, the rules differ considerably between the grammars. In order to compare the grammars, a much smarter diff is needed.

What I’m planning is to do a rule-by-rule comparison, working with rules in a DFS order from the start rule. I’ll need to know how the rules differ and why. Then, using Antlrvsix, apply a transformation to make the rules the same, and continue the diff with the next rule in the DFS ordering.

So, I first applied a DFS reorder of the grammars using the Antlrvsix extension, then a reformat (it turns out Codebuff has some bugs). Starting with the start rules, compilationUnit, the difference has already started: Parr decided to unfold ordinaryCompilation and remove the rule.

I could edit the grammar manually for the rule, but I decided to save time and implement the fold transformation right now. And, in order to make this reproducible, I want Antlrvsix to take a sequence of transformations from a file and apply them.

–Ken

I’ve made a release of Antlrvsix yesterday which includes recursion removal for Antlr grammars. There are two types of recursion that I can work with: direct (aka immediate) and indirect. The algorithm that I use for direct and indirect recursion removal are the ones that Aho et al. (2006) describe. In addition, the extension provides a transform to rewrite direct recursion (left or right) into Kleene operator form, which helps clean up the grammar considerably. These transformations help fix grammar problems that you may run across. But, to use these transformations, you have to select the rule that contains or participates in the recursion.

In addition, the extension now provides an explicit mapping for colorizing the grammar. The problem with LSP and VS2019 is that neither supports languages like Antlr. So, language features like nonterminals and terminals are mapped to client-specific data types that have display properties. This mapping can be modified in the options file for Antlrvsix. In addition, I now tag almost all of the grammar.

What’s next up for Antlrvsix? The main problem I wanted to solve is the performance of the Java9 parser, which performs terribly for certain input. The issue seems to involve left factoring, but it’s not clear. I’ll now be working on this for the next month. I was hoping to add Piggy into the extension, but that will have to wait.

–Ken