carbon-lang/explorer

Explorer

explorer is an implementation of Carbon whose primary purpose is to act as a clear specification of the language. As an extension of that goal, it can also be used as a platform for prototyping and validating changes to the language. Consequently, it prioritizes straightforward, readable code over performance, diagnostic quality, and other conventional implementation priorities. In other words, its intended audience is people working on the design of Carbon, and it is not intended for real-world Carbon programming on any scale. See the toolchain directory for a separate implementation that's focused on the needs of Carbon users.

Overview

explorer represents Carbon code using an abstract syntax tree (AST), which is defined in the ast directory. The syntax directory contains lexer and parser, which define how the AST is generated from Carbon code. The interpreter directory contains the remainder of the implementation.

explorer is an interpreter rather than a compiler, although it attempts to separate compile time from run time, since that separation is an important constraint on Carbon's design.

Programming conventions

The class hierarchies in explorer are built to support LLVM-style RTTI, and define a kind accessor that returns an enum identifying the concrete type. explorer typically relies less on virtual dispatch, and more on using kind as the key of a switch and then down-casting in the individual cases. As a result, adding a new derived class to a hierarchy requires updating existing code to handle it. It is generally better to avoid defining default cases for RTTI switches, so that the compiler can help ensure the code is updated when a new type is added.

explorer never uses plain pointer types directly. Instead, we use the Nonnull<T*> alias for pointers that are not nullable, or std::optional<Nonnull<T*>> for pointers that are nullable.

Many of the most commonly-used objects in explorer have lifetimes that are tied to the lifespan of the entire Carbon program. We manage the lifetimes of those objects by allocating them through an Arena object, which can allocate objects of arbitrary types, and retains ownership of them. As of this writing, all of explorer uses a single Arena object, we may introduce multiple Arenas for different lifetime groups in the future.

For simplicity, explorer generally treats all errors as fatal. Errors caused by bugs in the user-provided Carbon code should be reported with the error builders in error_builders.h. Errors caused by bugs in explorer itself should be reported with CHECK or FATAL.

Decompose functions

Many of explorer's data structures provide a Decompose method, which allows simple data types to be generically decomposed into their fields. The Decompose function for a type takes a function and calls it with the fields of that type. For example:

class MyType {
 public:
  MyType(Type1 arg1, Type2 arg2) : arg1_(arg1), arg2_(arg2) {}

  template <typename F>
  auto Decompose(F f) const { return f(arg1_, arg2_); }

 private:
  Type1 arg1_;
  Type2 arg2_;
};

Where possible, a value equivalent to the original value should be created by passing the given arguments to the constructor of the type. For example, my_value.Decompose([](auto ...args) { return MyType(args...); }) should recreate the original value.

Example Programs (Regression Tests)

The testdata/ subdirectory includes some example programs with expected output.

These tests make use of LLVM's lit and FileCheck. Tests have boilerplate at the top:

// Part of the Carbon Language project, under the Apache License v2.0 with LLVM
// Exceptions. See /LICENSE for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
// AUTOUPDATE
// RUN: %{explorer-run}
// RUN: %{explorer-run-trace}
// CHECK:result: 0

package ExplorerTest api;

To explain this boilerplate:

• The standard copyright is expected.
• The AUTOUPDATE line indicates that RUN and CHECK lines will be automatically inserted immediately below by the ./lit_autoupdate.py script.
• The RUN lines indicate two commands for lit to execute using the file: one without trace and debug output, one with.
  • RUN: will be followed by the not command when failure is expected. In particular, RUN: not explorer ....
  • The full command is in lit.cfg.py; it will run explorer and pass results to FileCheck.
• The CHECK lines indicate expected output, verified by FileCheck.
  • Where a CHECK line contains text like {{.*}}, the double curly braces indicate a contained regular expression.
• The package is required in all test files, per normal Carbon syntax rules.

Useful commands

• ./lit_autodupate.py -- Updates expected output.
  • This can be combined with git diff to see changes in output.
• bazel test ... --test_output=errors -- Runs tests and prints any errors.
• bazel run testdata/DIR/FILE.carbon.run -- Runs explorer on the file.

Updating fuzzer logic after making AST changes

Please refer to Fuzzer documentation.

Trace Program Execution

When tracing is turned on (using the --trace_file=... option), explorer prints the state of the program and each step that is performed during execution.

State of the Program

The state of the program is printed in the following format, which consists of two components: (1) a stack of actions and (2) a memory.

{
stack: action1 ## action2 ## ...
memory: 0: valueA, 1: valueB, 2: valueC, ...
}

The memory is a mapping of addresses to values. The memory is used to represent both heap-allocated objects and also mutable parts of the procedure call stack, for example, for local variables. When an address is deallocated, it stays in memory but !! is printed before its value.

The stack is list of actions separated by double pound signs (##). Each action has the format:

syntax .position. [[ results ]] { scope }

which can have up to four parts.

1. The syntax for the part of the program to be executed such as an expression or statement.
2. The position of execution (an integer) for this action (each action can take multiple steps to complete).
3. The results from subexpressions of this part.
4. The scope is the variables whose lifetimes are associated with this part of the program.

The stack always begins with a function call to Main.

In the special case of a function call, when the function call finishes, the result value appears at the end of the results.

Step of Execution

Each step of execution is printed in the following format:

--- step kind syntax .position. (file-location) --->

• The syntax is the part of the program being executed.
• The kind is the syntactic category of the part, such as exp, stmt, or decl.
• The position says how far along explorer is in executing this action.
• The file-location gives the filename and line number for the syntax.

Each step of execution can push new actions on the stack, pop actions, increment the position number of an action, and add result values to an action.