Zero-initialize struct-typed function-local variables (fixes Area Lights, OpenPBR Analytic Lights)

[bkaradzic-microsoft]

Summary

Babylon.js v6+ shaders for Area Lights (computeAreaLighting) and a number of other PBR/OpenPBR helpers declare a struct local without an initializer, then immediately do read-modify-write on its fields:

lightingInfo result;
result.specular += value;
result.diffuse  += value;
return result;

This is undefined behavior under the GLSL spec (locals are not zero-initialized), but every WebGL driver in practice tolerates it. SPIRV-Cross faithfully reproduces the pattern in its HLSL output, and D3DCompile correctly rejects it with error X4000: variable used without having been completely initialized. Result: Area Lights - Standard Material and any other test that exercises this code path fails to compile its fragment shader on D3D11 / D3D12 today.

Fix

Add a new AST traverser, ZeroInitializeStructLocals, that walks every function body and prepends EOpAssign(symbol, EOpConstructStruct(0, 0, ...)) for each unique struct-typed EvqTemporary symbol it references. The SPIR-V emitter then sees a normal store; SPIRV-Cross emits HLSL with the struct local definitely initialized; D3DCompile accepts it. Any subsequent real assignment in the body is dead-code eliminated by DXC's optimizer, so functional behavior is unchanged.

Wired into the DXBC, DXIL, Metal, and Vulkan compile pipelines (same as the existing sampler-handling and dFdy traversers). Arrays of structs are intentionally skipped — they are uncommon in BJS shaders and would require constructing a synthetic array constructor.

Verification

End-to-end on build/win32-default (D3D11, Chakra, RelWithDebInfo) and build/win32-gl (OpenGL/ANGLE, V8, Release):

Test	D3D11	OpenGL
idx 191 Area Lights - Standard Material	PASS (1027 px diff)	PASS (1131 px diff)
idx 194 Area Lights - PBR (regression check for #1718)	PASS	PASS
idx 558 OpenPBR Fuzz Weight vs Fuzz Roughness - Analytic Lights	PASS	PASS
idx 559 OpenPBR Fuzz Weight vs Coat Weight w/ Normal Maps - Analytic Lights	PASS	PASS
idx 574 OpenPBR Transmission Dispersion VS IOR - Analytic Lights	PASS	PASS
idx 0/39/47/165 (baseline regression checks)	PASS	PASS

Wide regression sweep:

• D3D11: 135 indices (1-30, 100-115, 160-180, 285-305, 380-405, 540-560) → 134 PASS / 1 FAIL. The single failure is idx 22 Outline with Error: Unsupported alpha mode: -1 — pre-existing, unrelated to shader pipeline.
• OpenGL: 113 indices (1-50, 100-120, 250-270, 400-420) → 52 PASS / 60 SKIP / 1 FAIL. Same idx 22 failure.

No new regressions on either backend.

Why a BN traverser (not a SPIRV-Cross fix)

Both the SPIRV-Cross fork-extension route and the BN traverser route would solve this. We chose the traverser because:

1. It avoids extending the BabylonJS/SPIRV-Cross fork further.
2. The fix lives next to the existing sampler/dFdy traversers, which use the exact same pattern.
3. It applies uniformly to all four backends (DXBC, DXIL, Metal, Vulkan) with one change.

Out of scope

• idx 192/193 (Textured - Area Lights - *) still fail with Unsupported texture format or type: format 5, type 3553. This is a NativeEngine::PrepareImage whitelist gap and orthogonal to this PR.
• A separate cluster of OpenPBR-IBL tests (idx 557/561/575) fails with a JS-side TypeError: 'FLOAT' undefined before any shader compile. Will be filed as a separate issue.

[bkaradzic-microsoft]

Closed and superseded by #1721 (branch was renamed for housekeeping).