Booth

Booth — Changelog
=================

## Booth 0.6.0

### Runtime

- `kath run`, `kath build` and `kath doctor`, so one command builds a source
  and runs it on whatever device is there (Zane Hambly, 2026-09-14)

### Backends

- `--nvidia-cubin` writes a cubin the card will load, with no NVCC anywhere in
  the chain. All 67 of ggml-cuda's files now reach the IR
  (Zane Hambly, 2026-09-09)

### Frontend

- `constexpr` and `const` objects fold at every use, and anything the folder
  cannot evaluate refuses with E128 (Zane Hambly, 2026-09-03)

- class templates, specialisations, default template arguments and
  `enum class` parse, so 47 of ggml-cuda's 67 files reach the lowerer
  (Zane Hambly, 2026-09-04)

- the lowerer's tables no longer run out of room on a large translation unit,
  taking ggml-cuda's lowering errors from 774 to 210 (Zane Hambly, 2026-09-04)


## Booth 0.5.3

### Runtime

- #169: the runtime is split by where it runs, the `BC_ERR_*` codes no
  longer collide, and the examples and NVIDIA harness are built
  (Zane Hambly, 2026-08-23)


### Frontend

- variadic template parameter packs, several `.cu` files as separate
  translation units, `mma.sync` and `mfma` lowering, and an i1 that no
  longer strides by zero (Zane Hambly, 2026-09-03)

- `(a) + (b)` adds again; the parser treated any parenthesised identifier as a
  type name without asking whether it named one, so the left operand vanished
  into a cast with no diagnostic (Zane Hambly, 2026-09-03)

- the cast test is now the type name registry, so the registry has to be
  complete. Template type parameters, `using X = T` aliases and the type names
  sema resolves without a typedef (`size_t`, `uint32_t`, `float4` and the rest)
  all reach it. A compound literal through a typedef, `(pair){1, 2}`, parses
  for the first time, and `sizeof(name)` where the name is a type reads as a
  type rather than an expression (Zane Hambly, 2026-09-03)

- llama.cpp's ggml-cuda preprocesses, all 67 files; `#pragma once` is
  honoured, variadic and multi-line macro invocations expand, and an
  expansion too big for the output buffer is E053 rather than an
  unterminated buffer the lexer reads past (Zane Hambly, 2026-09-03)

- `kath --mlir` reads MLIR text, no LLVM in the path. Čertík's pure-C
  reader vendored under `src/mlir/vendor` (mlir 826b69c9, corec a160199d),
  reached only through `src/mlir/mlir_fe.c` (Zane Hambly, 2026-08-11)

- `src/mlir/lower.c` walks the parsed module into BIR: `func.func`, `return`,
  `arith.constant` and every arith binop, compare and conversion the reader
  classifies. From there it is the pipeline CUDA and Triton already use, and
  MLIR reaches all four backends. `--mlir --pp` reprints instead
  (Zane Hambly, 2026-08-11)

- an op outside the subset stops the lowering and names itself. Skipping it
  would leave a function that compiles and computes something else
  (Zane Hambly, 2026-08-11)

- five fixes to the vendored reader, all worth upstreaming, and four of them