The lexer is the code generated by the flexer that is actually responsible for lexing Enso source code. It chunks the character stream into a (structured) token stream in order to make later processing faster, and to identify blocks

Lexer Architecture

The structure of the flexer’s code generation forces the lexer to be split into two parts: the definition, and the generation. As the latter is the point from which the lexer will be used, the second subproject is the one that is graced with the name lexer.

Libraries in the Lexer Definition

The lexer generation subproject needs to be able to make the assumption that all imports will be in the same place (relative to the crate root). To this end, the definition subproject exports public modules library and prelude. These are re-imported and used in the generation subproject to ensure that all components are found at the same paths relative to the crate root.

This does mean, however, that all imports from within the current crate in the definition subproject must be imported from the library module, not from their paths directly from the crate root.

Lexer Functionality

The lexer provides the following functionality as part of the parser.

  • It consumes the source lazily, character by character, and produces a structured token stream consisting of the lexer ast.
  • It succeeds any input, even if there are invalid constructs in the token stream, represented by Invalid tokens.

The Lexer AST

In contrast to the full parser ast, the lexer operates on a simplified AST that we call a ‘structured token stream’. While most lexers output a linear token stream, it is very important in Enso that we encode the nature of blocks into the token stream, hence giving it structure.

This encoding of blocks is crucial to the functionality of Enso as it ensures that no later stages of the parser can ignore blocks, and hence maintains them for use by the GUI.

It contains the following constructs:

  • Referent: Referrent identifiers (e.g. Some_Ref_Ident).
  • Variable: Variable identifiers (e.g. some_var_ident).
  • External: External identifiers (e.g. someJavaName).
  • Blank: The blank name _.
  • Operator: Operator identifiers (e.g. -->>).
  • Modifier: Modifier operators (e.g. +=).
  • Annotation: An annotation (e.g. @Tail_Call).
  • Number: Numbers (16_FFFF).
  • DanglingBase: An explicit base without an associated number (e.g. 16_).
  • TextLine: A single-line text literal.
  • TextInlineBlock: An inline block text literal.
  • TextBlock: A text block literal.
  • InvalidQuote: An invalid set of quotes for a text literal.
  • TextSegmentRaw: A raw text segment in which the contents should be interpreted literally.
  • TextSegmentEscape: A text segment containing an escape sequence.
  • TextSegmentInterpolate: A text segment containing an arbitrary interpolated expression.
  • TextSegmentUnclosedInterpolate: An unclosed interpolation text segment.
  • Line: A line in a block that contains tokens.
  • BlankLine: A line in a block that contains only whitespace.
  • Block: Syntactic blocks in the language.
  • InvalidSuffix: Invalid tokens when in a given state that would otherwise be valid.
  • Unrecognized: Tokens that the lexer doesn’t recognise.
  • DisableComment: A standard comment that disables interpretation of the commented code (i.e. #).
  • DocComment: A documentation comment (e.g. ##). Documentation syntax is not lexed by this lexer.

The distinction is made here between the various kinds of identifiers in order to keep lexing fast, but also in order to allow macros to switch on the kinds of identifiers.

The actionables for this section are:

  • Determine if we want to have separate ASTs for the lexer and the parser, or not.

Benchmarking the Lexer

As the lexer is the first port of call when getting an Enso program to run it needs to be quick. To that end, we insist on comprehensive benchmarks for any change made to the lexer. The lexer benchmarks are written using criterion.rs, and include both examples of whole program definitions and more specific benchmark examples.

Baseline Commit: e5695e6f5d44cba4094380545036a3a5cbbf6973

The benchmarking process for the lexer is as follows:

  1. Check out the current baseline commit, listed above.
  2. In lexer_bench_sources.rs change the line that reads .retain_baseline to instead read .save_baseline. This will save the current baseline (taken on your machine).
  3. Run the benchmarks using cargo bench. Please note that running these benchmarks takes approximately two hours, so sufficient time should be allotted.
  4. Once the baseline run has completed, change the above-mentioned line back to .retain_baseline. This will disable overwriting the saved baseline, and will perform its regression reporting against it.
  5. Make your changes.
  6. Run the benchmark suite again. It will report any performance regressions in the benchmark report, measured against your saved baseline.

Unfortunately, the use of time-based benchmarks means that we can’t commit the baseline to the repository. There is far too much variance between machines for this to be useful.

Running a Subset of the Benchmarks

The benchmarks are very comprehensive, running a wide range of program text through the lexer while replicating it out to various sizes (see lexer_bench_sources.rs for the full list). However, in order to decrease iteration time it can be useful to run a subset of these.

There are two main tuning points for this:

  1. The sizes of inputs being executed on.
  2. The benchmarks being executed.

The sizes can be tuned by editing the SIZES array in the lexer_bench_sources.rs file. The benchmarks themselves are best tuned by changing the macro definitions in lexer_time_bench.rs to exclude certain benchmarks or groups of benchmarks.

While it is possible to tune the benchmarking config (bench_config in lexer_bench_sources.rs) to decrease benchmarking time, this is not recommended. The current settings are tuned to provide reliable results.

Changing the Lexer

When changing the lexer the full benchmark suite must be run against the current baseline before the changes can be merged. This suite run must use the provided settings for the benchmarking library, and should be performed using the process described above.