Skip to content

Latest commit

 

History

History
726 lines (564 loc) · 29.3 KB

internals.md

File metadata and controls

726 lines (564 loc) · 29.3 KB

Sorbet Internals

This doc is for people interested in learning how Sorbet works for the purpose of contributing to the codebase.

For how to use Sorbet, see https://sorbet.org/docs/overview. For how to build and test Sorbet, see the README.

Otherwise, welcome!

This document is a work in progress. Please ask questions when you encounter an unfinished or confusing section!

Table of Contents

Overview

Sorbet is broken up into a main typechecking pipeline with a handful of supporting utilities and data structures.

Briefly, this is the folder structure we're working with. (Note that this is not comprehensive, but should give you a rough idea):

sorbet
│   // 1. Main
├── main
│   ├── pipeline    → Sequences the phases, feeding one phase into the next.
│   ├── lsp         → The language server-specific code.
│   ├── options     → Option parsing
│   └── autogen     → Stripe-specific, for autogen
│
│   // 2. Phases (more info below)
├── parser
├── ast
│   └── desugar
├── dsl
├── local_vars
├── namer
├── resolver
├── cfg
│   └── builder
├── infer
│
│   // 3. Other
├── common          → Non-Sorbet specific utilities and things that
│   │                 need to be everywhere.
│   └── os          → Platform-specific code.
├── core            → Sorbet-specific data structures and utilities.
│   └── types       → Our type system is shared by many passes of the
│                     pipeline code above.
│
│   // 4. Gems
├── gems
│   ├── sorbet          → The Ruby source for `srb init`.
│   └── sorbet-runtime  → The Ruby source for the runtime type checks.
└── ···

So by now you should have a rough sense for the high level structure of Sorbet. The rest of this documentation is not meant to be read linearly. Feel free to jump around between sections.

In particular, to understand the phases, you have to understand the core abstractions, but being familiar with the phases motivates why certain core abstractions exist.

Pipeline

When learning about the phases in the next, it can be helpful to look at this high-level architecture diagram of Sorbet's core type checking pipeline:

→ docs/pipeline.md

Phases

IR stands for "internal representation". Each phase either translates from one IR to another, or modifies an existing IR. This table shows the order of the phases, what IR they operate on, and whether they translate from one IR to another or make modifications within the IR they were given.

*: Even though these passes modify the IR they're given, they have another important job which is to populate GlobalState.

**: This pass doesn't even modify the AST. It just emits errors.

Translation Pass IR Rewrite Pass
source files
1 Parser, -p parse-tree
parser::Node
2 Desugar, -p desugar-tree
3 ast::Expression Rewriter
4 ast::Expression LocalVars, -p rewrite-tree
5 ast::Expression Namer, -p name-tree (*)
6 ast::Expression Resolver, -p resolve-tree (*)
6 ast::Expression [DefinitionValidator] (**)
6 ast::Expression [ClassFlatten], -p ast
7 CFG, -p cfg --stop-after cfg
8 cfg::CFG Infer, -p cfg

When you see links to files below, you should open the file and give it a quick skim before continuing. Most of the sections below are written like a guidebook to the codebase to help you get where you're going, not a comprehensive reference.

Sorbet has two flags which are invaluable for inspecting what happens from one phase to the next:

  • -p, --print <state>
    • Prints Sorbet's internal state, including any of the IRs.
    • Only some of the printing options are shown in the table above. See the help for all the options.
    • When getting started with Sorbet, oftentimes the ***-raw variants of the printing options are more useful, until you get familiar with the codebase.
  • --stop-after <phase>
  • Stop Sorbet early

We will discuss individual phases and IRs below.

Parser

The parser we're using is based on whitequark/parser, the popular Ruby parser. The Ruby 2.4 whitequark parser was ported to yacc / C++ by Charlie Somerville for use in his TypedRuby project, and has since seen many external contributions to support later Ruby versions. You can find the sources in third_party/parser/.

We interact with the TypedRuby parser using code generation to build a C++ header. To see the C++ header, first build Sorbet, then look inside bazel at bazel-genfiles/parser/Node_gen.h.

The header itself is generated using parser/tools/generate_ast.cc.

In general, the IR the parser generates models Ruby very granularly. This fine granularity is frequently extra fine for the purpose of typechecking. We use the Desugar and Rewriter passes to simplify the IR before typechecking.

Desugar

The desugar pass translates from parser::Node into ast::Expression. The goal of the desugarer is to drastically cut back on the granularity of the parser's IR.

To give you some sense of the scope of the desugar pass, let's see some numbers. At the time of this writing, there are 98 subclasses of parser::Node. There are only 34 subclasses of ast::Expression.

To see the desugar pass, you'll want to look at ast/desugar/Desugar.cc. You'll notice that it's mostly one large recursive function with a typecase. (See below for more on typecase; it's basically "pattern matching as a function.")

Some examples of things we desugar in this pass:

  • case expressions become chains of if/else expressions
  • compound assignment operators (+=) become normal assignments (x = x + 1)
  • unless <cond> becomes if !<cond>

If you pass the -p desugar-tree or -p desugar-tree-raw option to sorbet, you can see what a Ruby program looks like after being desugared.

Rewriter

The Rewriter pass is sort of like a domain-specific desugar pass. It takes ast::Expressions and rewrites specific Ruby DSLs and metaprogramming into code that Sorbet can analyze. DSL in this context can have a broad meaning. Some examples of DSLs that are rewritten by this pass:

  • attr_reader and friends are rewritten to simply define the methods that would be defined were the attr_reader method to be run (did you know that attr_reader is just a normal method in Ruby, not a language keyword?)

  • Chalk::ODM's prop definitions are written similarly to attr_reader

The core Rewriter pass lives in rewriter/rewriter.cc. Each Rewriter pass lives in its own file in the rewriter/ folder.

We've envisioned the Rewriter pass as a potential extension point for some sort of plugin system. This will allow for a wider audience of Rubyists to teach Sorbet about DSLs they've written. This is why we've intentionally limited the power of Rewriter passes.

Specfically, we artificially limit what code we call from Rewriter passes. Sometimes it would be convenient to call into other phases of Sorbet (like resolver or infer), but instead we've reimplemented functionality in the Rewriter pass. This keeps the surface area of the API we'd have to present to plugins in the future small.

LocalVars

This is a fairly short pass. It converts the ast::UnresolvedIdent AST nodes that correspond to local variables to ast::Local nodes (ast::UnresolvedIdent nodes are also used for instance variables, class variables, and global variables, not just local variables, but those are handled by other phases).

For the most part doing this is very straightforwardly accomplished with a tree traversal. One trick is that local variables record which Ruby block (like do ... end) they're a part of. (Ruby blocks introduce new lexical scopes; things like if / else and begin / end expressions do not.)

Namer

Namer is in charge of creating Symbols for classes, methods, globals, and method arguments. (Counterintuitively, Namer is not in charge of creating Names. See below for the difference between Symbols and Names.)

The file you'll want to see is namer/namer.cc.

Symbols are the canonical store of information about definitions in Sorbet. Namer walks the ast::Expression tree and calls various methods to create a Symbol which is owned by GlobalState and get back a reference to what was created (for example, enterMethodSymbol and enterClassSymbol). These methods return a SymbolRef, which is conceptually a newtype wrapper around a pointer to a Symbol. See below for a discussion of Symbols vs SymbolRefs.

The key datastructure for the Namer pass is the Symbol table, which we can print out. Given this file:

Click to expand docs_example_1.rb
class A
  def method(method_arg)
    local = 1 # no name will be created

    @field = 2

    $global = 3
  end

  # singleton methods are just methods on the singleton class
  def self.singleton_method(singleton_method_arg)
    # singleton fields are just fields on the singleton class
    @singleton_field = 4
  end

  @@static_field = 5
end

We'll see this symbol-table output:

❯ sorbet --no-stdlib -p symbol-table --stop-after namer docs_example_1.rb
class ::<root> ()
  field #$global @ docs_example_1.rb:7
  class ::A < ::<todo sym> () @ docs_example_1.rb:1
    method ::A#method (method_arg) @ docs_example_1.rb:2
      argument ::A#method#method_arg<> @ docs_example_1.rb:2
  class ::<Class:A> < ::<todo sym> () @ docs_example_1.rb:1
    method ::<Class:A>#singleton_method (singleton_method_arg) @ docs_example_1.rb:11
      argument ::<Class:A>#singleton_method#singleton_method_arg<> @ docs_example_1.rb:11

Some notes:

  • The output shows you the Symbol's kind (class, method, etc.).
  • The nesting matches the internal structure. A class Symbol knows how to return its members. A method Symbol knows how to return its arguments.
  • Some of the definitions in our example aren't here yet (i.e., most of the fields) because we need to know the inheritance hierarchy first. See Resolver.
  • None of the type / inheritance information is filled in yet. This is left to Resolver.

Namer used to be a relatively simple phase. It still conceptually follows this pattern (walk definitions, create Symbols for definitions) but it has been optimized for parallelism and speed.

See Namer & Resolver Pipeline for more details.

Resolver

After Namer has run, we've created Symbols for most (but not all) things, but these Symbols haven't been woven together yet. For example, after Namer, we had a bunch of Symbols marked <todo sym> representing classes' ancestors. Another example: after Namer, we'd created Symbols for methods, but none of these Symbols carried knowledge of their arguments' types.

There are many specialized jobs of Resolver, but we're just going to call out two: resolve constants, and resolve sigs. We'll discuss each in turn.

Resolve constants: After Namer, constants literals (like A::B) in our trees manifest as UnresolvedConstantLit nodes. An ast::UnresolvedConstantLit node wraps a NameRef while an ast::ConstantLit wraps a SymbolRef. In these terms, the process of resolving constants is to convert Names into Symbols (UnresolvedConstantLits into ConstantLits).

Resolve sigs: Once constants have been resolved to proper Symbols, we can fill in type information (because signatures are mostly just hashes of constants). To fill in a sig, Sorbet parses information out of the ast::Send nodes corresponding to sig builder methods, uses this to create core::Types, and stores those types on the Symbols corresponding to the method and arguments of that method.

There are a handful of other things that the Resolver does (it computes and records a linearization of the ancestor hierarchy so derivesFrom checks are fast, it computes bounds for generic type members, it handles T.type_alias declarations, it finalizes the information needed to power T.attached_class, etc.) There are pretty good comments at the top of resolver/resolver.cc which say more.

To give you an idea, this is what our Namer example looks like after the Resolver pass:

❯ sorbet --no-stdlib -p symbol-table --stop-after resolver docs_example_1.rb
class ::<root> < ::Object ()
  field #$global @ docs_example_1.rb:7
  class ::A < ::Object () @ docs_example_1.rb:1
    method ::A#<static-init> () @ docs_example_1.rb:16
    static-field ::A#@@static_field -> T.untyped @ docs_example_1.rb:16
    field ::A#@field -> T.untyped @ docs_example_1.rb:5
    method ::A#method (method_arg) @ docs_example_1.rb:2
      argument ::A#method#method_arg<> @ docs_example_1.rb:2
  class ::<Class:A> < ::<Class:Object> () @ docs_example_1.rb:1
    field ::<Class:A>#@singleton_field -> T.untyped @ docs_example_1.rb:13
    method ::<Class:A>#singleton_method (singleton_method_arg) @ docs_example_1.rb:11
      argument ::<Class:A>#singleton_method#singleton_method_arg<> @ docs_example_1.rb:11
  • Type information is filled in (everything after the -> is new).
  • All the <todo sym>s are gone (because constants have been resolved).
  • There are more fields now (like ::A#@field).

Resolver has always been a somewhat complex phase. It has been made somewhat more complex with the introduction of parallelism, though it still conceptually follows the patterns discussed here.

See Namer & Resolver Pipeline for more details.

ClassFlatten

The class_flatten is the final pass that processes the AST. The goal here is to move around all the nodes so that the final result only has top level classes and they only contain method definitions. Code that executes at the top-level of a Ruby class is gathered up into a special self.<static-init> method inside that class. The top level statements in a file are moved to a unique <static-init> method on the synthetic <root> object.

After this pass, all code that could be type checked is in a method. This means that in the next pass, CFG, we can look only at ast::MethodDef AST nodes and ignore ast::ClassDef nodes (all the information we could ever want about a class def has already been entered into GlobalState).

Note: There is also a phase in Rewriter called rewritter::Flatten. That pass is designed to pull nested ast::MethodDefs out into the top-level of a class. (This has some nice properties, e.g., it's possible to define all Symbols without traversing into method bodies. And since it's done in Rewriter, the output is cached and invalidated at the file level.)

If you have better names for these two phases to make them more distinct, they would be very welcome!

CFG

CFG is another translation pass, this time from [ast::Expressions] into cfg::CFG. CFG here stands for "control flow graph."

Unlike ast::Expression which is highly recursive (i.e., an ast::If is pretty much just three sub ast::Expressions), the CFG is largely flat. To drastically simplify the structure of a CFG, it looks something like this:

class LocalVariable {};
class Type {};
class Instruction {};

class Binding {
  LocalVariable local;
  Type          type;
  Instruction   instruction;
}

class BasicBlock {
  vector<Binding> bindings;
  LocalVariable   finalCond;
  Type            finalType;
  BasicBlock *    whenTrue;
  BasicBlock *    whenFalse;
}

class CFG {
  vector<BasicBlock *> blocks;
}

A CFG is pretty much a vector of basic blocks, and a basic block is a vector of instructions that compute something and assign their result to a local variable. None of the instructions in a basic block can branch (whether on a condition or unconditionally). But at the end of a basic block, we allow one branch. We distinguish the contents of one of the variables in the basic block as the branch condition, and then record which other basic block to jump to depending on the value of that variable (whenTrue when truthy, whenFalse when falsy).

Again, the structure above is highly abbreviated; look at cfg/CFG.h and cfg/Instruction.h for more specifics.

Note that while basic blocks can't have internal branches, they can still "jump" by calling into other methods.

Some things which are different about Sorbet's CFG than other CFG's you might be familiar with:

  • In Ruby, nearly every instruction can raise inside a begin ... rescue and jump to the start of the rescue block before subsequent expressions in the begin block run. But in Sorbet, we pretend that the jump to the rescue block happens instead either immediately after entering the begin, or after all expressions in the begin have executed. This simiplifying assumption (either nothing has run or everything has run) is in practice enough to model the control-flow sensitive typing context of variables used inside the rescue block.

  • Our CFG is not static single assignment (SSA), because we haven't needed the power that SSA gives rise to. Instead, we largely make do by "pinning" variables in outer scopes to a certain type, and saying that assignments to those variables in inner scopes (i.e., within a loop or if) must not change the variable's type.

Sorbet's use of a CFG for type inference is pretty cool. By doing inference on the CFG and keeping careful track of file locations (see core::Loc), Sorbet's inference algorithm can be very general. We only have to implement typechecking for ~11 kinds of instructions (+ control flow) instead of all ~98 kinds of nodes in parser::Node or all ~34 kinds of nodes in ast::Expression.

It also makes it easier to implement dead code analysis and flow-sensitive typing. And because basic blocks can't jump into basic blocks from a different method, method bodies can be type checked independently of each other.

Note: if you have graphviz installed, you can render a CFG to an image:

tools/scripts/cfg-view.sh -e 'while true; puts 42; end'

Some notes about how to read these:

  • each method gets its own box
  • each method always has one entry and one exit block (with ids 0 and 1, respectively)
  • bold arrows represent the true branch; thin arrows are the false branch
  • each basic block declares the locals that must be in scope upon entry
  • "dead" blocks are the blocks where there are no types next to the locals in a binding

Infer

Infer is the last pass. It operates directly on a cfg::CFG. In particular, when the CFG is created, each binding has nullptr its local's type. By the end of inference, reachable bindings within basic blocks will have had their types annotated with the result of inference.

Inference itself pretty much iterates over a (best-effort) topological sort of the basic blocks. ("Best effort" because there might be cycles in basic blocks). For each binding in each basic block, we

  • check whether this instruction is well-typed (now that there are so few node types, this is not so hard!)
  • use this binding to update our knowledge of the types for future bindings

This loop over instructions happens in infer/environment.cc. Look for processBinding, which is a big typecase over each cfg::Instruction type.

By far the most complicated thing about inference is checking cfg::Send instructions (method calls). Checking if a method call is well typed, and figuring out what the return type should be is implemented in core/types/calls.cc.

Sorbet visits each binding at most once to decide on its type. There is no backsolving for types or iterating until a fixed point, and there is no constraint generataion plus unification step. This single-shot style of inference is fast, because we only make one decision about typing per instruction, but it restricts the power of Sorbet's inference in user-visible ways. Also, since the CFG can have cycles, we require that within a cycle of basic blocks a variable's type cannot be widened or changed. (See http://srb.help/7001.)

The inference pass itself is largely just traversing the the CFG for each method and processing bindings. It delegates much of the implementation of the type system (like getting a method's result type, checking argument types, subtyping, generics, etc.) to logic implemented core/types/. See below for a discussion of how Sorbet's type system works.

Core

Core abstractions within Sorbet.

Refs (i.e., Symbol vs SymbolRef)

Sorbet is pretty fast. There are a couple of reasons for this, but one of them is Sorbet's use of Refs. A Ref (with a capital R) in Sorbet is a way to uniquely identify an allocated object.

For example, Sorbet has a class called Symbol and another called SymbolRef. A SymbolRef is conceptually a newtype wrapper around a pointer to a Symbol. All SymbolRefs which point to the same Symbol are equal, so comparison can be done quickly. Symbols are then stored in GlobalState, so you can always get the data for a SymbolRef to look up a Symbol's fields.

There are a handful of such paired data structures: Symbol/SymbolRef, and Name/NameRef, and File/FileRef are the most widespread.

Why not just use pointers? Refs are usually smaller than an 8-byte pointer. It's also nice to have the distinction reinforced in the type system. Also the "dereference" operation on these types is written foo.data*() instead of *foo or foo->. Sorbet subscribes to the philosophy that slow operations should be longer to type.

We use various .enterFoo methods on GlobalState to create new objects owned directly by GlobalState. These methods return FooRefs. These objects can't be constructed any other way.

Also, the SymbolRef, NameRef, and FileRef types are nilable by default. Any such Ref might not exist (you can check with the .exists() method).

Symbols

Symbols are canonical stores of semantic information about definitions. They carry types, parents, the kind of definition, locations where the definitions was defined, etc. Most of the job of the Namer and Resolver passes are to populate GlobalState with accurate Symbols representing every definition in a Ruby program.

Symbols are then consumed by later passes to do interesting things, chiefly to report type checking errors.

Symbols are somewhat polymorphic. There is one Symbol class, but a Symbol might represent a class, method, field, argument, etc. There is some bit packing going on to represent all of these things with little overhead, so be sure to use the public Symbol methods which ENFORCE that the operations you're trying to do on a Symbol make sense for this kind of Symbol.

See core/Symbols.h and core/SymbolRef.h for more information.


Work In Progress

These are just extra things I don't have a solid idea of where to put yet.

Names

  • Names: essentially "strings" but as numbers instead
  • (not quite "just strings" because two Names with identical printed name but different numbers are not equal.)
  • (which is why we need substitutions to renumber merged trees)
  • generate_names is changed frequently

ast::Expression (aka Trees)

  • GlobalState + treemap instead of explicit recursion
    • Avoids having to construct and deconstruct trees
    • Can instead just index into one tree
    • Link to chapter 4 of dmitry's thesis
  • gotcha: cast_tree will just return nullptr if you pass in nullptr. if you expect that the thing you're trying to cast is not null, ENFORCE it!
  • tagged pointers to avoid virtual dispatch

typecase

  • Basically pattern matching using C++ template hackery.

core::Loc

  • pronounced 'lohk' not 'lock'
  • fast / bit hacks
  • philosophy around useful error messages
  • there is a canonical loc, because symbols can have multiple locs, and most code only wants to have to care about the "best" loc.

beginError and strictness levels

  • We use beginError which will first check whether we'll report this error
  • Returns a builder, so that we can avoid the expensive parts of constructing the error message if we're not going to even report this error at the current file's strictness level.

The type system

  • document the type system of core/Types.h
  • T.self_type is return type of Object.dup, conceptually
  • lub → 'or'
  • glb → 'and'
  • intrinsic: compute result type of method as a function in C++, rather than through statically declaring its type with sigs.
  • ground types vs proxy types (dmitry thesis 2.4)
  • dependent object types (thesis)

Parallelism in the pipeline, and substitutions

  • Find the picture from dmitry's Sorbet Internals video

LSP

  • Idk how LSP works. Depending on scope, might want to have whole new section / doc to describe just LSP.
  • LSP might change significantly soon. Might not want to document it yet.
  • How to debug LSP in VSCode / locally?

RBI files and the payload

  • There are two builds of sorbet
    • one which reads stdlib rbi files from disk
    • one which stores all stdlib rbi files inside our binary and reads them directly into global state on startup

General tips for learning about sorbet

  • no but actually, click through all the links here and read the source
  • set up jump-to-def in your editor (it is possible)
  • use sorbet -p and https://sorbet.run liberally
    • look at the before and after of each pass
  • use lldb to break on specific functions and step through the logic on small examples
  • Finish this doc on things I learned about C++

Other topics to document

  • sanityChecks and ENFORCEs
    • both: only in debug builds
    • sanityCheck: internal consistency check, run at pre-determined times (like "after desugar")
    • ENFORCE: in-line assertion (pre- / post-condition)
  • show vs toString
    • most data structures have both methods
    • show is "something that could be shown to the user" (like Rust Display trait)
    • toString is "internal representation" (like Rust Debug trait)
  • gems/sorbet/ (srb init)
  • gems/sorbet-runtime/
  • bazel build //foo --copt=-ftime-trace --spawn_strategy=local
    • generates chrome://tracing profiles for clang