This page outlines the principles we aspire to follow and the more concrete objectives and goals we aim to accomplish in the Mojo standard library.
For details about our near-term objectives, see the Roadmap doc.
The following are “North Star” principles for the Mojo standard library. These principles will inform multiple future decisions, from what features we work on to what bugs we prioritize during triage. In short, the standard library vision is the ideal we may never reach, but we collectively show up every day to work towards it.
-
Foster a vibrant community collaborating globally. The community is encouraged to engage with the standard library and language evolution. We intend to ignite enthusiasm to contribute to the expanding Mojo package ecosystem.
-
The standard library prioritizes Performance, Safety, Portability, and Debuggability.
-
Respectable performance by default. The standard library should provide respectable performance by default, but not perfect performance. We support low-level controls that enable performance tuning by systems engineers. While providing consistently strong performance over time, we do so with minimal regressions.
-
Portability by default. The use of MLIR enables portability across a variety of platforms without writing per-platform code in the standard library.
-
The standard library does not have special privileges. The standard library types are not special and do not enjoy elevated privileges over user-contributed code. This empowers Mojicians to create primitives equally as expressive as core language primitives.
-
Fully utilize available hardware. The standard library should not restrict the use of available hardware on the system. On the contrary, standard library primitives should enable users to maximize the utility of all available system hardware.
-
Standard library features prioritize AI workload optimizations. Mojo ultimately aims to be a multi-purpose programming language used to solve problems in the systems programming domain. However, we will prioritize standard library features and optimizations that improve the state of the art for AI.
We reject the following principle statements:
- Tensor operations are first-class citizens. Mojo has a prime directive to optimize AI workloads. However, certain core AI primitives are tightly integrated with the MAX engine architecture, and will remain part of the MAX engine codebase.
-
Make unsafe or risky things explicit. Software using unsafe constructs is inevitable, but it must be minimized and explicit to the reader. Safe things shouldn’t look like unsafe ones and unsafe constructs should leave artifacts to see in the code.
-
Unsafe operations support dynamic checking — where pragmatic. It’s fine to have unchecked, unsafe operations for performance, but developers need the ability to check things before calling these unsafe APIs.
-
Advanced memory management features. Provides a fleet of memory allocators out of the box for kernel developers to use. The language runtime provides a decently performing default global allocator implementation (eg. thread local caches, automatic slab-size scaling based on heuristics, virtual memory-based defragmentation, and more).
-
First-class support for parallelism and concurrency. To fully utilize available hardware, the standard library will provide a complete suite of primitives that maximize the parallelism potential of the system.
-
First-class debugging features. Integration with Mojo debugger by incorporating LLDB visualizers for the standard library types and collections to make debugging easy.
-
Consistent API and behavior across all primitives. A higher-level goal of the standard library is to be an example of well-written mojo code. For example, all collections should behave consistently with:
- Default-constructed collections do not allocate memory unless they are holding an element.
- Collections provide no thread safety unless mentioned explicitly.
- Commonly implemented collection operations use the same method names across all implementations.
- Naming of public aliases/parameters common among collections (akin to
value_type
and friends in C++).
-
Interoperability with Python code. Allows progressively migrating code to Mojo over time to not force an entire rewrite just to improve the performance of code where it matters most.
While some of the following may be common goals of other languages, the values don't outweigh the costs for us right now as we are moving fast to build the language. While we don’t actively attempt to break the following, we provide no guarantees that they work — especially over multiple releases.
- Stable ABI between language/compiler and library.
- Backward or forward compatibility guarantees in APIs or semantic behaviors.
- Integrating domain specific functionality, such as GUI toolkits, game development frameworks, networking or data science libraries to name a few.