Skip to content

Latest commit

 

History

History
517 lines (391 loc) · 17.8 KB

HACKING.rst

File metadata and controls

517 lines (391 loc) · 17.8 KB
Raw

Hacking on Rainbow Delimiters 2

Testing

A test setup must meet the following criteria:

  • Test definitions must be run by with Neovim as the Lua interpreter to get access to all Neovim APIs
  • Tests must not be affected by the user's own plugins and configuration
  • Each test which mutates editor state must run in its own Neovim process

The first two points are achieved through a small command-line interface adapter script (a shim). The shim exposes the command-line interface of a Lua interpreter, and internally it sets up environment variable to point Neovim at a prepared blank directory structure. Neovim is then called with the -l flag.

We do have to use some plugins though:

Both plugins are stored under the $XDG_DATA_HOME directory, the former as a symlink and the latter as a Git submodule.

As for process isolation, this is achieved inside the tests. We start a headless embedded Neovim instance which we control through MsgPack RPC from inside the test. We can control and probe this process only indirectly, which is awkward, but this is the best solution I could find.

Unit testing

We use busted for unit testing. A unit is a self-contained module which can be used on its own independent of the editor. Execute make unit-test to run unit tests. The busted binary must be available on the system $PATH.

End to end testing

End-to-end tests run in a separate Neovim instance which we control via RPC. These are tests which mutate the state of the editor, such as adding highlighting on changes. Execute make e2e-test to run all end to end tests.

Running tests with Neotest-busted

To run tests the g:bustedprg variable must be set to './test/busted', which is the path to the shim script. If the exrc option is set the variable will be set automatically.

Highlight testing

Highlights are tested by comparing the current highlights of a sample file with previously recorded highlights known to be correct. Of course this does nothing when defining new patterns or making changes to a sample file; in this case a human has to initially approve of the highlighting. Once that is done the current state can be recorded. Automated highlighting tests are useful when making changes to the highlighting logic itself to ensure the results remain unchanged.

Execute make highlight-test to run highlighting tests.

Definitions

Sample file
A file in the language we want to highlight. The contents have to be syntactically correct, and ideally the file should compile, but it does not have to make sense. Sample files are stored under an arbitrary name (although regular is the most common) in test/highlight/samples/<lang>.
Specification or spec
A Lua file which records all rainbow delimiter extmarks for a given combination of sample file and query. Why Lua? It could have been JSON, but generating nicely formatted Lua was simpler, that's all. Each spec is just a table, there is no logic.
Recording
The act of reading a sample file, extracting all highlighting information and writing it to a spec. You could write all the specs by hand, but there is a helper function for that instead.

Recording highlighting

First make the necessary changes to the sample file or query. Then call the record_extmarks function from the rainbow-delimiters._test.highlight module. This module is not part of the runtime plugin code, so it is undocumented. The function takes three optional arguments (all strings):

  • language: The language in question
  • sample: Name of the sample file
  • query: Name of the query

If any one of these is missing the specs for all applicable languages, samples or queries are recorded. You should at least specify the language, otherwise the function can take a lot of time.

Running highlight tests

Design decisions

Tables over strings for configuration

Strategies are given as a complex table, but a string identifier would have been much more pleasant on the eye. Which of these two is easier to read and write?

-- This?
settings = {
   strategy = {
      'global'
      html = 'local'
   }
}

-- Or this?
settings = {
   strategy = {
      require 'ts-rainbow.strategy.global'
      html = require 'ts-rainbow.strategy.local'
   }
}

Using strings might seem like the more elegant choice, but it it makes the code more complicated to maintain and less flexible for the user. With tables a user can create a new custom strategy and assign it directly without the need to "register" them first under some name.

More importantly though, we have unlimited freedom where that table is coming from. Suppose we wanted to add settings to a strategy. With string identifiers we now need much more machinery to connect a string identifier and its settings. On the other hand, we can just call a function with the settings are arguments which returns the strategy table.

settings = {
    strategy = {
        require 'ts-rainbow.strategy.global',
        -- Function call evaluates to a strategy table
        latext = my_custom_strategy {
            option_1 = true,
            option_2 = 'test'
        }
    }
}

Strategies

On container nodes

Every query has to define a container capture in addition to opening and closing captures. As humans we understand the code at an abstract level, but Tree-sitter works on a more concrete level. To a human the HTML tag <div> is one atomic object, but to Tree-sitter it is actually a container with further elements.

Consider the following HTML snippet:

<div>
  Hello
</div>

The tree looks like this (showing anonymous nodes):

element [0, 0] - [2, 6]
  start_tag [0, 0] - [0, 5]
    "<" [0, 0] - [0, 1]
    tag_name [0, 1] - [0, 4]
    ">" [0, 4] - [0, 5]
  text [1, 1] - [1, 6]
  end_tag [2, 0] - [2, 6]
    "</" [2, 0] - [2, 2]
    tag_name [2, 2] - [2, 5]
    ">" [2, 5] - [2, 6]

We want to highlight the lower-level nodes like tag_name or start_tag and end_tag, but we want to base our logic on the higher-level nodes like element. The @container node will not be highlighted, we use it to determine the nesting level or the relationship to other container nodes.

Determining the level of container node

In order to correctly highlight containers we need to know the nesting level of each container relative to the other containers in the document. We can use the order in which matches are returned by the iter_matches method of a query. The iterator traverses the document tree in a depth-first manner according to the visitor patter, but matches are created whenever the match is complete. This happens upon exiting the node if the child nodes are sandwiched in-between delimiters, as is the case with delimiters like parentheses or begin/end blocks. However, if the child nodes are outside the delimiters (e.g. when using Python keywords like def or while as delimiters) the child nodes are not sandwiched between delimiters and the match will be returned upon entering the node.

Sandwiching delimiters

Let us look at a practical example. Here is a hypothetical tree:

A
├─B
│ └─C
│   └─D
└─E
  ├─F
  └─G

The nodes are returned in the following order:

  1. D
  2. C
  3. B
  4. F
  5. G
  6. E
  7. A

We can only know how deeply nodes are nested relative to one another. We need to build the entire tree structure to know the absolute nesting levels. Here is an algorithm which can build up the tree, it uses the fact that the order of nodes never skips over an ancestor.

Start with an empty stack s = []. For each match m do the following:

  1. Keep popping matches off s up until we find a match m' whose @container node is not a descendant of the container node of m. Collect the popped matches (excluding m') onto a new set s_m (order does not matter)
  2. Set s_m as the child match set of m
  3. Add m to s

Eventually s will only contain root-level matches, i.e. matches of nesting level one. To apply the highlighting we can then traverse the match tree, incrementing the highlighting level by one each time we descend a level.

The order of matches among siblings in the tree does not matter. The stack s is important for determining the relationship between nodes: since we know that no ancestors will be skipped we can be certain that we can stop checking the stack for descendants of m once we encounter the first non-descendant match. Otherwise we would have to compare each match with each other match, which would tank the performance.

Here is a step-by-step illustration of the algorithm applied to the above example. The left-hand side is the current stack (with the bottom of the stack on the left) and current node, the right-hand side is the resulting stack for that iteration. If a match has no children I have omitted the braces for brevity.

Current stack Match New stack and popped-of match
[] D [D]
[D] C [], C{D}
[C{D}]
[C{D}] B [], B{C{D}}
[B{C{D}}]
[B{C{D}}] F [B{C{D}}, F]
[B{C{D}}, F] G [B{C{D}}, F, G]
[B{C{D}}, F, G] E [B{C{D}}, F], E{G}
[B{C{D}}], E{G, F}
[B{C{D}}, E{F, G}]
[B{C{D}}, E{F, G}] A [B{C{D}}], A{E{F, G}}
[], A{B{C{D}}, E{F, G}}
[A{B{C{D}}, E{F, G}}]
[A{B{C{D}}, E{F, G}}]

Without sandwiching

In some languages like Python it makes sense to define block-level delimiters which have only one delimiter. Here is an example:

def derp():
    for (k, v) in {'a': 1, 'b': 2}:
        print(k, v)

We want to highlight the def of the function definition and the for/in of the loop. This means we have a mix of sandwiching and no sandwiching. The order of matches is:

  1. def (because it is completed first)
  2. () (the parentheses of def)
  3. (k, v) (because it is completed before for/in)
  4. for/in
  5. {...}
  6. print(k, v)

The intended match tree should look like this according to the syntax tree:

def
├ ()
└ for/in
  ├ (k, v)
  ├ {...}
  └ print(k, v)

Eyeballing the code however suggest a match tree like this:

├def
└ ()
  ├ for/in
  │ ├ (k, v)
  │ └ print(k, v)
  └ {...}

The idea is that matches which logicaly appear together (such as the head of a for-loop) should be cousins. This raises the question of what belongs together. I will probably need to add a new capture like @body which matches the delimited content. In the sandwich case the body was implicitly that which is between both delimiters, but here we would need to be explicit about it. Example:

(for_statement
  "for" @delimiter
  "in" @delimiter
  body: _ @body) @container

(list
  "[" @delimiter
  _ @body
  "]" @delimiter) @container

Then a match is a child of a parent if and only if the @container of the child is contained inside the @body of the parent.

Not only can the parent-child order be reversed, we can also skip over generations. In the above example (k, v) is a grandchild of def, but it comes directly after it. We need to revise the algorithm to account for this case. All in all we have the following cases:

  • The new node and the top of the stack are cousins
  • The new node is an ancestor of the top node
  • The new node is a descendant of the top node

Here the term “cousin” is cross-generational, i.e. if A is the parent of B and C, and D the child of C, then B and D are considered cousins. They have a common ancestor, but share no lineage from one to the other. Siblings are also considered cousins.

The local highlight strategy

Consider the following bit of contrived HTML code:

<div id="Alpha">
  <div id="Bravo">
     <div id="Charlie">
     </div>
  </div>
  <div id="Delta">
  </div>
</div>

Supposed the cursor was inside the angle brackets of Bravo, which tags should we highlight? From eyeballing the obvious answer is Alpha, Bravo and Charlie. Obviously Alpha and Bravo both contain the cursor within the range, but how do we know that we need to highlight Charlie? Charlie is contained inside Bravo, which contains the cursor, but on the other hand Delta is contained inside Alpha, which also contains the cursor. We cannot simply check whether the parent contains the cursor.

When working with the Tree-sitter API and iterating through matches and captures we have no way of knowing that any of the captures within Charlie are contained within Bravo. However, due to the order of traversal we do know that Bravo is the lowest node to still contain the cursor.

Therefore we that the first match which contains the cursor is the lowest one. If a match does not contain the cursor we can check whether it is a descendant of the cursor container match.

The problem with nested languages

The language tree of a buffer is a tree of parsers. Some languages like Markdown can contain other languages, which complicates things.

Foreign extmarks

Extmarks move along with the text they belong to. This is generally a good thing, but it can become a problem if we move text from one language to another. Consider the following Markdown code:

Hello world

```lua
print {{{{}}}}
print {{{{}}}}
```

We can move the cursor to line 4 and move that line out of the Lua block by executing :move 1 to move it to the second line. However, this will preserve the extmarks and we will end up with Lua delimiter highlighting inside Markdown.

My solution is on every change to delete all rainbow delimiter extmarks which do not belong to the current language.

Overwritten extmarks

Take the following Markdown code:

Hello world

```c
puts("This is an injected language")
{
    {
        {
            {
                {
                    return ((((((2)))))) + ((((3))))
                }
            }
        }
    }
}
```

If we put the cursor on the line with the puts statement and move it up one line (:move -2) we get the following changes:

  • Markdown - { 2, 0, 3, 0 }

This means lines 3 and 4 of the Markdown tree have changed; we have changed the contents of the fifth line and added one more line. This is all as expected. However, let us now move the line back down by executing :move +1. We get the following changes:

  • Markdown - { 3, 0, 15, 0 }
  • C - { 3, 0, 4, 0 }

The changes to the C tree are what we expect. However, the changes to the Markdown tree span the code block as well. This is a problem when we start deleting foreign extmarks (see above). If we work from the outside we wipe out all non-Markdown extmarks in the range, which includes the C extmarks. Then we apply the C extmarks inside the C block, but the C change does not span the entire C tree. Thus we will only apply highlighting to the changed C line, but not the remainder of the C block.

The solution at the moment is to overwrite the changes of nested languages. If the changes belong to a language tree with parent language we replace all the changes with a range that spans the entire tree for that language.