M2Array stores a signed 32-bit self-relative offset. On x64 a separate
heap block can lie more than 2 GB from m_data, so relocated low-priority
.anim sequence data wrapped and crashed in ProcessEventCallbacks.
Add OsAllocNear/OsFreeNear (Win64): reserve a block within 1.5 GB of an
anchor. x64 now uses it for .anim buffers (load, cancel, destructor) and
for the skin vertex overflow, which again follows the Win32 structure
(m_flags |= 8, freed in ~CM2Shared). Texture combo arrays keep growing
m_data. All new code is under WOWDECOMP_ARCH_64.
Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>
- x64: texture combo, transform combo and converted vertex arrays are
appended to a grown m_data buffer instead of separate heap blocks, so
the self-relative 32-bit M2Array offsets always reach them; they are
freed with m_data (destructor frees only on Win32)
- FinishLoading clears header bits 0x20/0x40 from the file; the client
uses them to mark arrays it allocated, and later-expansion models set
them, which made the destructor free a pointer into the file buffer
Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>
M2Model.cpp, M2Shared.cpp, M2Scene.cpp, CM2SceneRender.cpp, M2Light.cpp,
M2Cache.cpp and the new M2Data.cpp/.h (0x81C000-0x834900). Only the
std::sort internals for CModelBatch stay unported.
AnimationData lookups in ResolveSequenceFallback/SequenceFallbackById use
g_animationDataDB directly: 12340 inlines its record loads into M2Model.cpp.
Shared additions: CParticleEmitter2 vtable order and the plane/sphere/spline
emitters, CRibbonMat, Services/Camera.h, Base/SIMD, Base/ObjectAlloc.h (the
engine-side ObjectAlloc interface), CFacet, CAaRect, CiRange, the C44Matrix
16-float constructor and CMath fequal/fnotequal/FloatToInt64.
Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>