3 people f57f53361e Replace the engine-specific body retargets with project-supplied assembly overrides (#67) ...

* codegen: keep fused-tree operands ahead of their consumers after the f16 gather rewrite

rewriteF16GatherGroup appends the collapsed v128_load16x4_u at the end
of the node list and points the chain's tail (now f16x4_cvt) at it, a
forward reference. The asm splicers walk the tree recursively and never
noticed, but the pure twin body emits nodes in list order, so any tree
that reached it through this rewrite declared the load after its use:
"undefined: n6 / declared and not used: n6" from the gcasm capture
build. It surfaced once ggml's wasm f16 vector loads started producing
the gather idiom everywhere (vec_dot_f16, vec_mad_f16, flash attention)
instead of only in the repack kernels.

compactDeadNodes, which every rewrite already runs, now places a node's
operands before the node while otherwise keeping list order — the
collapsed load lands where the chain's last lane load used to be.

* helpers: detect FEAT_I8MM on linux/arm64 (HWCAP2 bit 13)

The kernel-override dispatch and the SMMLA splices gate on it; darwin
already reported it through hw.optional.arm.FEAT_I8MM.

Co-Authored-By: Claude Fable 5.1 <[email protected]>

* Replace the engine-specific body retargets with project-supplied kernel overrides

wasm2go's job is to read a wasm module and emit Go and, where it can do
better than gc, assembly derived from that wasm. Two features had grown
past that line: the repack GEMM retarget swapped the body of an export
named dbg_gemm_q8_0_4x4 for a hand-written kernel that knows one
engine's quantized block layout, and the vec_dot row/column pairing
recognized one engine's trait-table layout and handed the paired
companion to another hand-written kernel of the same kind. Neither body
came from the wasm, and both tied the transpiler to that engine. On the
current engine build the pairing also no longer measures (prompt and
decode throughput within noise with it off), because the repack path
carries the hot matmuls.

Both are removed: the retarget and its -no-repack-gemm switch, the
pairing and its -vec-dot-pair-entry / -vec-dot-rows switches, the
kernels and the trait-table recognizer. The passive-segment placement
scan the f16-table detection shares moves to its own file.

In their place, a generic mechanism: -kernel-overrides <manifest>. The
project that produced the wasm may supply the assembly body of an
exported leaf function, per architecture and CPU feature level.
wasm2go validates the manifest against the module (export present,
signature and pointer width as declared), wraps each body in a fixed
ABI — the linear-memory base and size in named registers, arguments by
the ABI0 layout rule, a kov_oob label that reaches the module's
out-of-bounds trap — emits a feature-dispatch stub (i8mm > dotprod >
neon; avx512vnni > avx2 > sse4) with the lowered body as the portable
fallback, and pins the exports out of line. It never inspects what a
body computes; the body must stay inside the wasm execution model
(linear memory through the given base and size, bounds-checked, no
calls), and the manifest loader rejects TEXT, calls, system
instructions and foreign symbol references. docs/kernel-overrides.md
states the contract and the verification the project owes.

Tests: manifest validation, ABI text and dispatch shape, and an
end-to-end transpile of a module with and without an override body on
the host architecture — same sums, and the override traps past the end
of memory.

Co-Authored-By: Claude Fable 5.1 <[email protected]>

* helpers: detect FEAT_I8MM on darwin/arm64 too

The linux and fallback twins gained CPUI8MM; a darwin build of a bundle
whose kernel overrides declare an i8mm body needs the same symbol in
the base package.

Co-Authored-By: Claude Fable 5.1 <[email protected]>

* Rename kernel overrides to assembly overrides

"Kernel" is numeric-computing vocabulary; the mechanism itself knows
nothing about what the function computes — it replaces the body of an
exported function with project-supplied assembly. The name now says
that: -asm-overrides / Options.AsmOverrides, docs/asm-overrides.md,
the manifest's "functions" array, the ovr_oob label and ovr_ data
prefix, and Fn<idx>ovr<feature> body symbols.

Co-Authored-By: Claude Fable 5.1 <[email protected]>

---------

Co-authored-by: Claude Fable 5.1 <[email protected]>

2026-09-04 00:10:00 +09:00

..

asm-overrides.md

Replace the engine-specific body retargets with project-supplied assembly overrides (#67)

2026-09-04 00:10:00 +09:00