Improve OTBM load/save and map view performance
This change improves Remere's Map Editor performance in OTBM load/save paths,
object allocation, binary serialization, tile lookup, and idle map rendering.
Measured impact:
Object pool allocation:
- Slab refill events reduced from 17,830 to 2,230, about 87.5% fewer refills.
- 48-byte class refills reduced from 4,276 to 535, about 87.5% fewer refills.
- 64-byte class refills reduced from 5 to 1, about 80.0% fewer refills.
- 128-byte class refills reduced from 8,776 to 1,097, about 87.5% fewer refills.
- 1024-byte class refills reduced from 4,773 to 597, about 87.5% fewer refills.
- rme::allocatePooledObject sampled CPU share reduced from 20.01% to 14.62%,
about 26.9% lower sampled share.
- Total pooled allocation calls stayed at 42,559,971 for the measured workload.
- Heap fallback allocations stayed at 0, confirming the hot load path remains pooled.
Latest mixed load/save profile:
- GUI::LoadMap remained the dominant sampled cost at 67.02%.
- IOMapOTBM::loadMap accounted for 60.23% total sampled CPU, with 56.34% in
the inner load body.
- GUI::SaveMap and Editor::saveMap accounted for 28.16%.
- IOMapOTBM::saveMap accounted for 27.63%.
- BaseMap::forEachTileLocation during save accounted for 23.01%.
- Tile::Tile accounted for 11.53%.
- Tile::addLoadedItem accounted for 8.15%.
- Item::Create accounted for 6.99%.
- QTreeNode::createFloor accounted for 6.71%.
- BinaryNode::advance and BinaryNode::load accounted for 5.37% and 3.81%.
Map view idle and preview rendering:
- Static map-view sampled CPU dropped from 139,767 sampled units to 126 sampled
units after overlay-only refresh reuse.
- MapCanvas::OnPaint dropped from 46,229 sampled units and 32.20% to 19 sampled
units and 15.08%.
- GLRenderer::flushCommands dropped from 24,321 sampled units and 16.94% to
4 sampled units and 3.17%.
- Performance stats refresh changed from a 16 ms scene-dirty timer to a 500 ms
overlay-only timer.
- Show Preview changed from a 16 ms scene-dirty timer to a 250 ms scene-dirty
timer.
- Position indicator keeps the 16 ms scene-dirty timer because it is expected
to animate smoothly while active.
Main changes:
- Added cached floor and tile lookup while loading OTBM map data and spawn files.
- Added direct TileLocation assignment for parser paths that already resolved
the destination location.
- Added BaseMap::forEachTileLocation for direct save traversal of existing tile
locations.
- Added a small-object slab allocator for hot Item, Tile, and Floor allocations.
- Added pool owner-thread binding and diagnostics for allocation validation.
- Increased slab sizing to reduce refill pressure in large-map loads.
- Improved binary node writing by batching raw bytes and avoiding redundant cache
renewal checks.
- Avoided rewriting XML sidecar files when serialized content only differs by
line endings.
- Fixed invalid ground serialization so placeholder ground id 0 no longer drops
the rest of the tile contents during save.
- Split map canvas refresh into scene-dirty and overlay-only paths.
- Adjusted animation timer behavior for position indicator, Show Preview, and
performance stats.
- Scaled tooltip rendering with map zoom, clamped to 55% minimum.
- Kept review and Sonar cleanups away from hot path regressions with targeted
NOSONAR annotations or FORCEINLINE where needed.
- Added AGENTS.md guidance for future Git, build, and PCH discipline.
Notes:
- The Visual Studio captures are sampling profiles, so percentages represent CPU
sample share, not direct wall-clock speedup.
- The latest profile is a mixed interaction profile, not a strict load-only or
save-only benchmark.
- The allocator counters are the strongest before/after measurement in this
change.
- No OTBM format or map semantics are intended to change.
