Skip to content

Latest commit

 

History

History
307 lines (266 loc) · 14.8 KB

milestones.md

File metadata and controls

307 lines (266 loc) · 14.8 KB
Raw

Milestones

Table of contents

Overview

As Carbon progresses, we want to have some common long-term milestones that we orient our work around. The annual roadmap provides a specific and immediate set of priorities for the year, but we want successive years to point in a coherent direction with meaningful end goals. Milestones should typically be long-term, spanning more than a year, and have some functional motivation.

We also assign version numbers to our initial milestones to make them easy to refer to and incorporate into various versioning schemes.

Milestone 0.1: a minimum viable product (MVP) for evaluation

The first milestone is also the most concrete -- it is the MVP for C++ users and developers to begin evaluating Carbon seriously. We want to keep this milestone as minimal as we can while still enabling a sufficient initial round of evaluation.

Goals

From the perspective of outcomes, our goals for the 0.1 language are centered around what we expect evaluations to be able to include:

  • Evaluators have a clear idea of the long-term evolution strategy of Carbon and how it addresses different use cases and requirements.
  • Language design components are documented, cohesive, and understandable by evaluators without placeholders.
    • The components and language features must include the foundational core of the language. These features must also be sufficient to translate existing C++ code (except coroutines) into obvious and unsurprising Carbon code.
    • Also in-scope are additional features that impact API design or need early feedback, but only if they are low cost to both the design and implementation.
    • Example language components include: lexical structure, expressions, statements, conditions, loops, user-defined types, and their dependencies.
    • Example library components include: integer types, floating point types, strings, arrays, ranges, pointers, optionals, variants, heap allocation, and their dependencies.
    • Where these build on top of other language or library designs, those are transitively in-scope.
  • Design for both Carbon use of C++ and C++ use of Carbon, including all major C++ language features except for coroutines, is documented, cohesive, and understandable by evaluators without placeholders.
  • Evaluators can build and run tests of most C++ interoperability, including both real-world C++ code that can be built with the latest release of Clang and test Carbon code.
    • Gaps from the design need to be ones that don't undermine evaluation confidence.
  • Evaluators can effectively stress-test build speed and scaling with Carbon.
  • Evaluators can build some key benchmarks that include C++ interoperation in the critical path and get representative performance results.
    • This can in turn be a still smaller subset of all aspects of C++ interoperability based around what impacts interesting benchmarks.

Language features

These are focused on the core necessary features for us to reach a successful 0.1 language that can address our goals. Some of these features are required directly by the above goals, others are required due to dependencies or interactions with the directly required features. However, we don't try to cover all of the features in full granularity here. There will be many minor components that are necessary for these to hold together but are not directly addressed. In general, unless something is explicitly described as partial or having exceptions, everything covered by that entry should be expected in the 0.1 language.

Another important point is that this doesn't commit Carbon to any particular design for any of these bullet points. Instead, it just means that the Carbon design must have something that addresses each of these bullet points. That might be to add the named design to Carbon, but it equally might be a clear statement that Carbon will not include this design but use some other language features to address its use cases and when rewriting C++ code using that feature into Carbon.

Code organization and structuring

  • Packages
  • Libraries
  • Implementation files
  • Importing
  • Namespaces

Type system

  • User-defined types
    • C++ interop: importing C++ types into Carbon, exporting Carbon types into C++
    • Single inheritance
      • Virtual dispatch
      • C++ interop:
        • Bi-directional inheritance between C++ and Carbon
        • Type hierarchy roots in both C++ and Carbon
        • Mappings of inheritance features: abstract, final, virtual
    • Operator overloading
      • C++ interop:
        • Synthesizing Carbon overloads for imported C++ types
        • Exporting Carbon overloads into C++
    • Sum types (discriminated unions)
    • Unions (un-discriminated)
      • C++ interop: mapping to and from C++ unions.
  • Generics
    • Both generic functions and types
    • Checked generics
      • Definition-checked variadics
    • Integrated templates
      • Including template-style structural conformance to nominal constraints, both modeling the members (like interfaces) and arbitrary predicates (like C++20 expression validity predicates)
    • C++ interop:
      • Importing C++ templates, instantiating on Carbon types
      • Exporting Carbon templates, instantiating on C++ types
      • Exporting Carbon checked generics (as templates), instantiating on C++ types
      • Mapping C++20 concepts into named predicates, and named predicates into C++20 concepts

Functions, statements, expressions, ...

  • Functions
    • Separate declaration and definition
    • Function overloading
    • C++ interop:
      • Importing C++ functions and methods and calling them from Carbon
      • Exporting Carbon functions and methods and calling them from C++
      • Importing C++ overload sets into Carbon overload sets where the model (closed overloading) fits
      • Importing C++ open-overload-sets-as-extension-points (swap, etc) into synthetic Carbon interfaces for common cases (likely based on heuristics)
  • Control flow statements
    • Conditions
    • Loops
      • Range-based loops
      • Good equivalents for a range of existing C/C++ looping constructs
    • Matching
      • Good equivalents for C/C++ uses of switch
      • Working with sum-types, especially for C++ std::variant and std::optional interop
      • Both positive (if let in Rust) and negative (let else in Rust) combined match control flow and variable declaration
  • C++ interop: support for C++'s threading and atomic primitives, memory model, and synchronization tools
  • Error handling
    • Any dedicated error handling control flow constructs
    • C++ interop:
      • Mechanisms to configure how exception handling should or shouldn't be integrated into C++ interop sufficient to address both -fno-except C++ dialects and standard C++ dialects
      • Calling C++ functions which throw exceptions from Carbon and automatically using Carbon's error handling
      • Export Carbon error handling using some reasonably ergonomic mapping into C++ -- std::expected, something roughly compatible with std::expected, C++ exceptions, etc.

Standard library components

Note: we expect to heavily leverage the C++ standard library by way of interop for the vast majority of what is needed in Carbon initially. As a consequence, this is a surprisingly more minimal area than the language features.

  • Language and syntax support library components
    • Fundamental types (bool, iN, fN)
    • Any parts of tuple or array types needed in the library
    • Pointer types
    • Interfaces powering language syntax (operators, conversions, etc.)
  • Types with important language support
    • String and related types used with string literals
    • Optional
    • Slices
  • C++ interop:
    • Transparent mapping between Carbon fundamental types and C++ equivalents
    • Transparent mapping between Carbon and C++ non-owning string-related types
    • Transparent mapping between Carbon and C++ non-owning contiguous container types
      • Includes starting from an owning container and forming the non-owning view and then transparently mapping that between languages.
    • Transparent mapping between Carbon and C++ iteration abstractions

Project features

There are a few important components of the overarching Carbon project that need to be completed as part of 0.1 beyond language features:

  • A functioning Carbon toolchain:
    • Supports drop-in usage as a Clang C++ toolchain with the most common Make- and CMake-derived build systems.
    • Implements most of the features above, including C++ interop, and any remaining gaps don't undermine the ability to evaluate the remaining features or the confidence in the overall evaluation.
    • Installs on Windows, macOS, and Linux, and builds working programs for those platforms.
  • Build system integration for CMake, and documentation for integrating with Make or similar build systems.
  • Basic documentation for evaluators from getting started to FAQs.

Milestone 0.2: feature complete product for evaluation

The second milestone already concretely in mind is reaching a level of feature-completeness. The completeness metric here is based around the features necessary to credibly address the existing needs of C++ users and developers interested in moving to Carbon, and so will be heavily driven by either features already in use in C++ or necessary features to make moving off of C++ a viable tradeoff. Ultimately, we need this milestone to be sufficiently feature complete that users can complete their evaluation of Carbon. The language will continue to evolve and grow features beyond this, but at this point there shouldn't be any feature gaps that are problematic for the initial target audience coming from C++, and we should be able to finalize the Carbon experiment with this feature set.

That said, this is about as concrete as we can get for a milestone that remains years in the future. The full scope of requirements for this milestone will be defined as we complete 0.1 and begin getting feedback on it. Currently, we just call out specific features that we are actively deferring until at least 0.2 but without being listed somewhere could cause confusion.

Features explicitly deferred until at least 0.2

  • Memory safety
  • Coroutines, async, generators, etc.
  • Comprehensive story for handling effects, and how functions can be generic across effects
  • Carbon-native threading
  • Long tail of metaprogramming features
  • Mixins
  • Properties
  • Inline assembly
  • SIMD
  • Some ability to define an API that communicates & shares data across language versions, build configurations, and other FFI or ABI boundaries.
  • Necessary parts of the standard library

Why are coroutines and async in this milestone?

Specifically, why not address them earlier in 0.1? Or if they can be deferred why not defer them further?

Coroutines and async programming are large and complex topics to introduce into the language. From watching C++, Rust, Swift, Kotlin, and many other languages working in this space, we have a strong belief that trying to add these to the 0.1 language would significantly increase the amount of work and likely delay when we reach that milestone. Also, given the recency of coroutines being added to C++, we expect evaluators to be able to reason about their absence and still accomplish most of the evaluation of Carbon without issue.

However, we also expect that as coroutines start to be widely adopted in C++, they will become an essential feature of the language that would be extremely difficult to give up when moving to Carbon. So we expect coroutines to be a necessary feature for us to effectively decide that the Carbon experiment is a success and begin planning large-scale adoption and migration.

Milestone 1.0: no longer an experiment, usable in production

Even less concrete is the milestone that marks Carbon no longer being an experiment, but if successful, a usable language. Currently this is speculatively called 1.0 but even that is highly subject to change as we approach. Again, we simply call out features here that we do expect to have in a 1.0 milestone, but want to explicitly defer beyond the 0.2 milestone.

Features explicitly deferred beyond 0.2

  • Robust language evolution strategy and plan, specifically addressing:
    • The need to make on-going changes to the language to address feedback and a changing landscape.
    • The cost on users of churn and change over time and how to manage that cost.
    • A mechanism to address users who need true, durable stability over long time horizons.
  • Package management strategy and plan, and any early groundwork needed.
  • Developer experience is high enough quality to enable initial production users.
    • Includes compiler error messages and basic developer tooling.
  • Everything we've learned we need as part of the evaluation of 0.1 and 0.2