Layout/SIMD probe, 2026-09-21 — eight findings on the surface that returned none in September
Verdict: 8 findings, 1 fixed (e99b8484), 1 refuted (finding 2), 6 open. Probed 2026-09-21 on the macOS arm64 laptop with tychoc 0.8.5 from main. Outcome: the instrument lesson — a demo-sized program is not evidence about a surface.
probe-simd-2026-09-07.md aimed at packed, align(N), vector[N]T and swizzling and came back with zero findings. This is the same four features, two weeks later, with eight. The difference is not the model and not the surface: that probe wrote a 57-line pixel kernel, this one wrote a program with an independent checking harness and then tried to break each feature on purpose. A demo-sized program on a surface is not evidence about that surface.
Host: the macOS arm64 laptop, tychoc 0.8.5 from main.
The program
tgagrade — a per-channel gain/bias colour grader for 32-bit uncompressed TGA. packed struct TgaHeader (18 bytes with a u16 starting at byte 3, i.e. deliberately misaligned), align(8) struct ChanStat in a [4], vector[4]f32 for q := px * gain + bias, and px.(r, b) = px.(b, r) as the real BGRA↔RGBA conversion rather than a demonstration.
The checking is the part worth copying. Three independent oracles, none of them the program: Python's own struct.pack/unpack_from writes the input and re-parses the output header field by field; a float64 recomputation compares every pixel; and macOS sips decodes the output TGA to PNG, which a hand-written inflater then compares to the same expectation. Plus cc measuring the emitted structs directly. A 256×256 sweep covers every byte value on every channel.
Then it broke each feature on purpose and recorded what the checks said — deleting the swizzle fails 30 of 32 pixels, widening one u8 header field to u16 shifts size_of$ to 19 and trips the program's own validation, lying about the origin byte is caught by the third decoder and by nothing else. Control that turned out not to be one: swapping width and height in the declaration changes nothing observable, because the program reads and writes through the same struct and the error cancels. A round-trip through your own layout cannot detect a symmetric layout error.
The score
| # | finding | verdict |
|---|---|---|
| 7 | on arm64 every wide vector warns about x86-64 and prescribes --target x86-64-v3, which cc then rejects on that same host | FIXED e99b8484 — FRICTION 124 |
| 2 | "packed does not mean packed in memory" — the emitted C carries no __attribute__((packed)), sizeof is 20 not 18, width sits at 14 not 12 | REFUTED — FRICTION 127 |
| 1 | size_of$ refuses every type align(N) applies to, so align(N) is unobservable from inside the language | OPEN |
| 3 | none of the four features appears anywhere in docs/reference/, the layer docs/README.md calls "the single answer layer" | OPEN |
| 4 | three of the four have zero users in examples/ and corelib/; only packed has any (4) | OPEN |
| 6 | vector[N]T has no lane-wise min/max, so the clamp — the operation a grading kernel does as often as the multiply — drops out of the vector unit | OPEN |
| 8a | §5.5 omits u8 from the ordered types, the mixed-type diagnostic omits char/u32/u64/f32, and the compiler's real rule is in neither | OPEN |
| 8b | str() on a vector, and a vector as a struct field, both work and are undocumented | OPEN |
| 8c | bytes + and [a:b] are in the spec and not in reference/types.md's bytes section — correct docs, wrong layer | OPEN |
| 8d | two -Wunused-value warnings out of core:strings in every build | OPEN — the handle probe hit the same two |
Re-checked against main on 2026-09-22: 1, 3 and 6 reproduce exactly, including the error text 'math__clamp' instantiated with T = vector[4]f32, which does not satisfy comparable(T) and size_of$(C): only a packed struct has a stated byte size.
Finding 2 is the important one, because it is wrong
It was reported as the finding "I would most want a Tycho maintainer to see": a normative spec claim (§17.1a, "no padding between fields … each field sits at the sum of the sizes before it") contradicted by the compiler's own output, with an FFI consequence attached.
It does not reproduce. tychoc --emit-c on the probe's *own* tgagrade/main.ty, run here, ends that struct with } __attribute__((packed));, and cc measuring that definition verbatim reports:
emitted-verbatim TgaHeader: sizeof=18 alignof=1 offsetof width=12 height=14 descriptor=17
— which is §17.1a exactly, and exactly the numbers the report called "byte-exact would be". The attribute has been emitted since packed shipped (5fff6d5c).
The defect is in the probe's own apparatus. probe/layout.c, the file whose header says the definitions were "copied VERBATIM out of tychoc --emit-c", has the attribute on its align(8) struct and not on the packed one. The copy dropped one line, cc faithfully measured the mis-copy, and the measurement was then read back as a language defect. Every downstream sentence — 20 bytes, width at 14, "what packed actually buys is the side-table", the §14 FFI worry — is a description of probe/layout.c.
probe-simd-2026-09-07.md closed with the lesson that a report's own citations have to be opened rather than trusted. This is the stronger form: a report's own measurements have to be re-run. A probe that builds an independent instrument can be wrong about the instrument, and it will present that error in exactly the format that makes it most convincing — a number, from a second toolchain, against a quoted spec line.
What went right
The diagnostics again, in the probe's words "the best part of the system": every rejection named the rule, the value and usually the fix. packed and align(N) together are refused by name; vector[4]u8 is refused with the element rule stated; a math.clamp on a vector names the constraint it failed. The four features themselves all did what the spec said they do — align(8) measurably raised size from 12 to 16, the swizzle lowered as read-both-then-write-both, and 262,144 graded pixels matched an independent float64 recomputation.
corelib/raster/raster.ty's BmpHeader was singled out: "taught me the whole file-format idiom in one read." It is the only in-tree teaching example any of these four features has.
What it could not do
- Observe
align(N)from inside Tycho at all (finding 1). It left the language and rancc. - Confirm the little-endian-on-every-host guarantee — both available machines are little-endian, so the interesting half of §17.1b is still untested here.
- Keep the clamp in the vector unit (finding 6).
- Find any of the four features from the documentation's own recommended reading path (finding 3): "Without the section numbers in the brief I would have concluded they do not exist."
Filing
The program is kept out of tree, and the afternoon's other one is not: tydu came in as tools/tycho-du/ because it closed a named gap (a handle built with --shim, which nothing demonstrated). tgagrade closes no gap that way — its value was the harness, not the program.
But finding 4 is the standing argument on the other side: vector[N]T, align(N) and swizzling have no worked user anywhere in the tree, and a probe cannot close that by being thrown away. Whatever eventually does close it will need a program, and this record is what the next one should be aimed past.