Related Work: Attempts at an Improved C

PFE in action: before designing Elseon, survey the attempts that came before, take the ideas, and leave the rest.

This document surveys existing attempts at an improved modern version of C. It is a living knowledge document: it records what was tried, what converged, and what Elseon can borrow.

Exclusions (why some well-known candidates are not models)

Rust. The ideas within Rust may hold merit (ownership, move semantics, bounds safety), but the choice of an entirely different surface representation is pure ego, not reason. Elseon’s syntax family is C and JavaScript; a language whose surface rejects C’s form fails the minimal-cognitive-friction principle. Rust is an idea source, not a model (see below).
D. The work of that author is valued, but garbage collection makes sense for application programming and little or no sense for kernel programming. D fails the Kernel Test (no GC). D is an idea source, not a model (see below).
Go falls under the same standard as D: GC’d and application-oriented, so it is not a kernel candidate. It is at most an idea source (tooling and simplicity).

Camp 1: Fresh C-like designs

Clean-slate languages that keep C’s shape but fix its pain points. These are the closest to Elseon’s design space: no GC, manual memory, LLVM backend.

Language	Take on “improved C”	Notes
C3	Builds on C’s syntax, evolving it into a better language; zero-cost abstractions, C-compatible FFI, LLVM backend	Probably the closest philosophical sibling to Elseon
C2	“Cleaner C”: removes the preprocessor in favor of modules and imports; LLVM-based	Smaller community
Zig	Modern low-level alternative to C: no GC, explicit allocators, usable as a drop-in C compiler (`zig cc`), LLVM backend	The most successful “better C” positioning to date
Jai	C-like, no GC, compile-time metaprogramming, LLVM backend; designed for game engines	Long-awaited, still not publicly released
Odin	Data-oriented, C-like, no GC, explicit allocators, LLVM backend	Active
Hare	Simple, C-like, no memory management required; minimal and static; for UNIX-like OSes	Used to build a real OS
Val	C-like syntax with mutable value semantics, to kill aliasing bugs	Newer

Honorable mentions with caveats: Nim compiles to C (GC optional, so kernel candidacy depends on the configuration chosen).

Camp 2: Supersets and safe dialects of C

These stay compatible with C and patch its wounds in place:

C++ — the original attempt (1985), but it became a second language bolted onto C’s semantics rather than a repair.
Cake — “C23 and beyond”: a superset of C with modern features that can desugar back to C.
Cyclone (Cornell, 2002) — a safe dialect of C: region-based memory, checked pointers, no GC. The classic academic answer to “safe C.”
Checked C (Microsoft, 2016) — bounds-checked pointer types layered onto C.
CHERI / Morello — not a new language but a hardware repair: capability-based pointers make C memory errors architecturally impossible.

Idea sources (PFE: take the idea, leave the surface or runtime)

From Rust: ownership and move semantics as a concept — the insight that a value has exactly one owner and transfers ownership explicitly. Also bounds safety as a goal. Rejected: the surface representation.
From D: contracts (in/out/invariant), slices with length, immutable and pure attributes. Rejected: the GC runtime. (Each idea is admitted only if it passes the Kernel Test — e.g. bounds checks are explicit safety, not hidden control flow.)
From Go: simplicity of tooling and formatting. Rejected: the GC runtime.

What this means for Elseon

Two failure modes. Supersets (C++, Cake, Checked C) stay compatible but inherit C’s baggage; clean-slate designs (C2, C3, Zig, Jai) are clean but must rebuild the ecosystem. Elseon’s PFE position — ingest C headers as first-class semantic modules — is a third position: compatibility without superset baggage, and no need to re-declare the world.
The Kernel Test is the differentiator. Every serious candidate converges on the same constraints Elseon chose: no GC, no hidden allocations, no hidden control flow, deterministic layout, LLVM backend, C-compatible ABI.
Concrete borrowable features:
- C3/C2: preprocessing replaced by modules — the same instinct as Elseon’s semantic header import.
- Zig: explicit allocators as a first-class concept (Elseon’s “no hidden allocations”).
- Jai/Odin: compile-time metaprogramming without a macro preprocessor.
- Cyclone/Checked C/CHERI: the menu of C-safety mechanisms (checked pointers, regions, bounds annotations) — Elseon picks from the menu rather than reinventing it.
- Rust (ideas): ownership/move semantics; bounds safety as a goal.
- D (ideas): contracts, slices, immutable/pure attributes.

Status

Surveyed and recorded in the Elseon project (phase 1 — research).
The borrowable-features list feeds the open questions in TODO.md and the design document documents/07-elseon-language.md.
Episode 1-A (the C-comparison study series) is complete: five study notes compare C3, Odin, Hare, Val, and Cake with C as the baseline on the common criteria — documents/13-c3-vs-c.md, documents/14-odin-vs-c.md, documents/15-hare-vs-c.md, documents/16-val-vs-c.md, and documents/17-cake-vs-c.md.
All five candidates classify as idea sources for Elseon; none classifies as a model under the kernel test.
The studies deepen and refine this survey’s rows: C3 is a documented fork of C2 (one lineage, not two independent rows); Odin is C-shaped but uses type-last declarations and an explicit allocator mechanism with an implicit per-scope default; Hare is spec-first with a QBE backend rather than LLVM and builds real kernels (Helios, Hermes, Ares) with the stock toolchain; Val is Swift-shaped, renamed Hylo in 2023, and remains pre-1.0 and officially unusable; Cake is a Camp 2 front end that transpiles C23 to C89, with compile-time-only ownership checks.
The row corrections are recorded in the synthesis; the survey tables above keep the pre-study wording until the human decides to fold the corrections in.
The synthesis documents/18-c-comparison-synthesis.md pools the five studies: a cross-candidate comparison on the criteria, what was tried before and what worked and did not work across all five, the Elseon use and avoid lists with reasoning and source studies, candidate-drop recommendations for the human to decide, and proposed revisions to prompts/common/04-c-comparison-criteria.md (proposals only — the criteria file is unchanged).
The synthesis is the input that informs phase 2 (language specification).