dasSpirv: fix vec*=scalar (OpVectorTimesScalar) + SPIR-V 1.4 entry-point interface

[borisbat]

Two dasSpirv emitter bugs surfaced by the first mesh-shader-with-a-UBO (a GPU-tessellated teapot tutorial in dasVulkan), both proven via spirv-val and guarded by new fixtures.

1. vec *= scalar emitted an illegal OpFMul(vec, scalar)

A compound-assignment multiply with a vector LHS and a scalar RHS lowered to a component-wise OpFMul — illegal, since OpFMul requires both operands to equal the result type — which crashed the driver. The * binop already routed mixed shapes through emit_mul (OpVectorTimesScalar / OpMatrixTimes*); the *= path bypassed it and emitted bo.code directly.

Fix: route *= multiply through emit_mul too; same-shape (vec*=vec, scalar*=scalar) still falls back to the component-wise op. (A runtime-scalar vec*float binop was already fine; the *= form was wrong for both constant and runtime scalars.)

2. SPIR-V ≥ 1.4 entry-point interface omitted UBOs

SPIR-V 1.4 requires the OpEntryPoint interface to list every global the entry point references — not just Input/Output (the ≤ 1.3 rule). The emitter only pushed Input/Output vars, so a mesh/task shader reading a UBO produced a blob whose Uniform variable was missing from the interface (invalid; some drivers tolerated it, others would not — and a strict validator/lavapipe rejects it).

Fix: at entry-point emission, for version ≥ 1.4, append the remaining used globals (Uniform / StorageBuffer / PushConstant / UniformConstant / Workgroup), ordered by var-id (= emission order) so the blob stays byte-reproducible.

Tests

• vts — vec *= 0.5 (const) and vec *= x (runtime) → two OpVectorTimesScalar, spirv-val clean.
• mesh_ubo — a mesh shader reading a @uniform → the Uniform variable is listed in the OpEntryPoint interface (count 4: 3 outputs + the UBO), spirv-val clean.

Both fail on the unfixed emitter. Full dasSpirv suite green (154/154: golden + census + spirv-val); no existing golden changed (nothing used vec*=scalar, and the existing mesh fixtures have no descriptor globals to add).

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Fixes two SPIR-V emission correctness issues in dasSpirv that were surfaced by mesh-shader + UBO usage and validated via spirv-val, with new fixtures guarding regressions.

Changes:

• Fix compound-assignment multiply (*=) so mixed-shape cases (e.g., vec *= scalar) route through emit_mul and emit legal OpVectorTimesScalar / OpMatrixTimes* ops instead of an illegal shape-mismatched Op*Mul.
• For SPIR-V version ≥ 1.4, expand OpEntryPoint interface to include all referenced globals (beyond Input/Output), appending any missing global variable IDs deterministically (sorted by var id).
• Add new SPIR-V test fixtures (vts, mesh_ubo) and wire them into the opcode census to ensure ongoing coverage.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
modules/dasSpirv/spirv/spirv_emit.das	Fix *= lowering for mixed shapes via emit_mul; for SPIR-V ≥ 1.4, include all referenced globals in OpEntryPoint interface.
tests/spirv/_spirv_common.das	Add new fixtures vts_spv (vec*=scalar) and mesh_ubo_spv (mesh shader reads UBO) and expose *_words() helpers.
tests/spirv/test_vecarith.das	Add test asserting vec *= scalar lowers to exactly two OpVectorTimesScalar and validates cleanly.
tests/spirv/test_mesh.das	Add test asserting mesh shader UBO is included in OpEntryPoint interface (SPIR-V 1.4) and validates cleanly.
tests/spirv/test_census.das	Include the new fixtures in the opcode census union.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.