[ASDAlexander77]

A struct's tuple now spells every field C++ lays out, so structs holding library types are tuples with that library's layout instead of Opaque.

What changes

• Specializations held by value are declared, named by their arguments: std::string → basic_string_char, nvrhi::static_vector<IBindingSet *, 8> → static_vector_IBindingSet_8, RefCountPtr<IShader> → RefCountPtr_IShader = [ptr_: IShader | null]. A pointer to a specialization alone stays Opaque /* ... */.
• Bases: a derived struct has its bases' fields first (an empty base has none). MSVC's tail padding after a base becomes a _pad field; the Itanium ABI's reuse of it is kept.
• Arrays: T x[N] → x: [_0: T, ..., _{N-1}: T] (up to 1024 elements; tslang has no fixed-size array type).
• Unions: the first member when it alone has the union's size and alignment, raw words otherwise (union { char buf[16]; char *ptr; } → [_0: u64, _1: u64]). Anonymous unions are named after their first member; anonymous structs are flattened.
• Bitfields: a run is raw words over its bytes, named after its first bitfield.
• [[no_unique_address]] empty members take no storage (libc++'s string allocator).
• One layout pass (spellRecord) both decides and spells, for recordSkipReason and mapFields.
• Fix: a struct aligned beyond its fields (alignas, DECLSPEC_ALIGN(16)) is now skipped — a tuple holding it would have been placed at the wrong offset.
• Fix: an operator with C linkage (guiddef.h's operator== for GUID) is no longer emitted as a nameless declare function (...).

Member names inside library types are that library's own and differ between MSVC's STL and libc++ (_Elems vs __elems_); tuple literals are positional, so only name-based reads are per-ABI.

Results

NVRHI: every desc is now a tuple. BufferDesc is 80 bytes under MSVC and 72 under libc++, matching C++ (nvrhi.layoutMatches()).

Tests

• New LayoutTest.cpp (arrays, bitfields, unions, held specializations, MSVC- and libc++-shaped std::string, base tail padding, inherited name clashes — under both windowsTarget and linuxTarget), Naming.OperatorWithCLinkageIsLeftOut; flipped expectations in SkipTest/NamespacedStructTest. 76/76.
• test-bindgen-cpp: the fixture passes a static_vector-shaped container (private array base + count), a buffer/pointer union and a bitfield run to real compiled C++ (exe + JIT). ctest -R "test-bindgen|cppvtable|cpp_slot": 46/46.
• Generated and compiled with tslang: NVRHI (MSVC, Android x86_64/aarch64), the MSVC STL and libc++ (--filter 'std::*').
• native_ts_graphics with every NVRHI struct generated (hand tuples removed): nvrhi_direct + headless pass on Windows (dx12/vk, AOT + JIT, -debug) and the Android emulator (x86_64, arm64).

Found, not fixed

• A type used before its declaration (including through Reference<X>) fails in tslang: windows.h + d3d12.h give 284 such errors with both this and the previous tsbindgen. Sorting the output by dependency is a follow-up.
• @cpp class fields (addClassFields) don't use the layout pass yet: a union or bitfield in a class still drops its fields. NVRHI's interfaces have none.
• On Android, plain char arrays are s8 on x86_64 and u8 on aarch64 (faithful; same layout).

Downstream

native_ts_graphics' branch nvrhi-generated-descs uses this (generated descs, hand tuples removed); it will be opened after a tslang with this change is deployed. Native main keeps building in the meantime.

🤖 Generated with Claude Code