Skip to content

Apple's open-source XNU kernel on Android phones: its own loader and drivers, Metal on the Adreno. Runs from RAM; a reboot gives the phone back to Android.

C

docs/air: the texture-as-parameter attempt, reverted, with what it cost to learn

vcvkk6cbb6e7a205a
diff --git a/docs/air.md b/docs/air.md
index 180d802..8da9866 100644
--- a/docs/air.md
+++ b/docs/air.md
@@ -309,6 +309,31 @@ functions and not entry points at all.
they are evidence about is binding, descriptors and ir3's texture instructions — exactly what
they were built to be — and silence about everything Metal's sampler state decides. Two modules
of the corpus declare a `constexpr sampler`; everywhere else the API supplies it.
+- **Four modules call a function that takes a texture, and that is attempted and not done.**
+ `air.convert`-style arguments pass through a `nir_call` happily; a texture does not, because
+ this frontend resolves a texture at translation time into a binding and `tex_handle` wants an
+ `AIR_VAL_PTR`, not an SSA def. The attempt, kept here so it is not repeated from the start:
+
+ - binding the pointer-class arguments from the caller's own resolution and passing only the
+ rest as NIR parameters **works** — "argument 0 is not a texture" goes away, and one of the
+ four, `FilmGrainChromaKernel_Uint`, gets past the call to a different limit (a shuffle mask
+ that is not constant);
+ - the module's own values have to be carried into the callee's fresh value table rather than
+ re-registered, because `translate_globals` is not idempotent: it appends to
+ `info->resources` and adds to `shared_size`, so a second run invents an inline sampler and
+ doubles the threadgroup memory. Copying the `module_values` set is the right carry;
+ - and then three of the four **crash**, in `nir_lower_goto_ifs` → `nir_repair_ssa` →
+ `nir_block_is_unreachable`, on an assertion that dominance metadata is valid. Marking the
+ hand-built callee impl's metadata invalid (`nir_metadata_preserve(impl, nir_metadata_none)`)
+ does not fix it, and the obvious hypothesis — that the callee has unreachable blocks, which
+ is what that predicate is about — is **wrong**: these callees have no `unreachable`
+ terminator at all.
+
+ So the work is reverted rather than left half-built, and the state above is the committed one.
+ What is missing is a diagnosis, not another fix: two speculative ones have already failed, and
+ a third would be guessing. Each of the four has exactly one call site, so specialising per site
+ costs nothing in duplication when it does work.
+
- **Of Apple's own 47 kernels, 6 are dispatched, all on software.** `tools/air/kernelcheck.py`
prints where each one stops, as a table whose rows are checked against the total before it is
printed: