mirror of
https://github.com/libretro/boom3.git
synced 2026-11-07 10:15:59 -10:00
libretroadmin
895b8ccaf8
R_FrameAlloc rounds every request to 16 but handed out memory from a block whose payload began at offset 20 on 64-bit: the header's pad was sized for 32-bit pointers, and the block itself came from Mem_Alloc, which promises 8. Every view, draw surface, render command and vertex array of a frame was therefore 4 bytes off its alignment, which UBSan reported on nearly a thousand accesses per run and which costs unaligned loads on every one of them. The header is now padded to a multiple of 16 on either pointer width and the blocks come from Mem_Alloc16, and the struct moves to renderer/frame_block.h so the layout can be checked on its own. The collision model polygons sit end to end in one block at their exact byte size, and edges[] is 4 bytes an entry, so an odd edge count left the next polygon's pointer and bounds on a 4 byte boundary. One measure, CM_PolygonSize, rounds the stride to 8 for the allocator, the free accounting and the block size the .cm writer records; a .cm sized by the old measure runs its block out a few polygons early and those take the allocator's existing fallback. tools/alignment checks the frame block layout at the 64-bit and, where the toolchain has -m32, the 32-bit layout, and the polygon stride; both fail on the old code, and the Linux workflow runs them.