A self-hosted compiler
The lexer, parser, AST, import resolver, semantic analysis, and LLVM IR generator are written in Prismio.
A roadmap that separates what works from what must be proved next.
Prismio is in active development. This is not a release calendar: it is the engineering direction behind the compiler, with the current implementation kept distinct from measured problems, proposed work, and unsupported surface area.
Established in the current compiler.
These are present capabilities, not roadmap aspirations.
Native code through LLVM 22
Prismio lowers to LLVM IR, then links a native executable through the platform toolchain.
Explainable allocation inference
AIF assigns storage tiers, exposes a source-oriented plan, and can explain a numbered decision or verify it at runtime.
A usable systems-language core
Types, generics, traits, closures, pattern matching, native tasks, typed channels, UMS projects, C ABI interop, JSON diagnostics, and DWARF are in the shipped surface.
The next work starts with evidence, not feature theater.
The active memory optimization tracker identifies these as high-leverage directions. Each is an engineering hypothesis with its own measurement requirement, not a promised performance number.
MEM-001
Measure allocation and lifetime behavior end-to-end
Extend the verifier and benchmark pipeline with stable allocation-site IDs, byte and lifetime metrics, peak-live data, and unified JSON. This is observability work; it makes no speedup promise.
Foundational
MEM-002
Make region and cycle state safe under native tasks
Move ambient allocator and cycle-collection state toward a design that is safe to use from concurrent native threads, then establish stress coverage before chasing throughput.
Correctness first
MEM-003
Price layout choices in their container context
A layout split that looks cheaper per object can lose badly when it breaks flat storage. The next model needs to account for construction, locality, vectorization, and release work together.
Measured regression guard
MEM-004
Preserve ownership decisions in a memory-aware IR
Moves, borrows, drops, regions, reuse, and representation choices are currently known to AIF but are not first-class operations between semantic analysis and LLVM lowering.
Architectural
MEM-005
Carry ownership and region facts across calls
Function summaries can make escape, return provenance, and memory effects visible across boundaries instead of forcing conservative decisions at each call site.
Precision
Missing is not the same as scheduled.
The benchmark suite marks 16 workloads unsupported rather than supplying stand-ins. Their missing capabilities are visible here without pretending every one has a release date.
Read the benchmark coverageLibrary surface
Linked/deque containers, ordered maps and sets, priority queues, map deletion, regex, JSON, and generic serialization are not implemented in the standard surface.
Concurrency surface
User-facing atomics, locks, work-stealing pools, and async I/O are not implemented. Prismio currently uses native OS threads and blocking typed channels.
Platform surface
Sockets, memory-mapped files, explicit SIMD types, and user-provided collection allocators are also outside the current supported catalog.
Follow the work where it happens.
Architecture notes, source changes, benchmark evidence, issues, and discussions are open. Contributions are most useful when they include a reproducible case or a clear design question.