A/B results: hash_lookup prefetch/cache both regressed (original is
near-optimal at 59 bytes), VM opcode SSE2 patches were break-even.
Only luaS_newlstr hash pre-check shows consistent 40-45% per-call win.
Drop hash_lookup hook, VM jump table patches, and resize hook. Keep
newlstr Detour with per-call A/B rdtsc and OnWorldUpdate periodic dump.
Add GC_WRITE_BARRIER.md documenting the barrier globals at 0xCEEAC0/C4
found in lua_vm_execute (useful for luagc module).
bench: add Detour overhead micro-benchmark. Simulates zhook's
JMP+trampoline mechanism on mmap'd executable pages. Result: 5 cyc/call
overhead -- negligible vs the ~100 cyc/call newlstr savings.
Rootgc sweep chunked by temporarily NULLing a next pointer and calling
the original lua_gc_remove_objects on the truncated sublist. Swept
survivors are detached to a separate list to prevent re-sweeping.
Lists reconnected after final chunk.
Birth-mark via binary patch of luaC_link immediate byte (0x6F7B37)
ensures new objects born during sweep survive.
Profiling shows sweep chunks at 0-9ms each (was 225ms avg, 5s worst).
Remaining bottleneck: atomic mark (80ms) + udata/string sweep (44-92ms).
Incremental GC: atomic mark + batched rootgc sweep. Crashes from
objects born white during sweep -- needs birth-mark barrier fix.
Link hook (luaC_link) calling convention verified but interaction
with sweep_ptr at list head needs solving. Disabled pending fix.
AllocBench: realistic addon-like data structures (combat events,
player damage breakdowns with metatables, aura trackers with
closures). Cross-references shared spell/unit caches. These
complex structures reveal GC as the true stutter source -- simple
flat objects don't trigger meaningful GC pauses.
Removed arena pre-reservation (was fragmenting 32-bit address space).
Profiled 67M Lua allocations to find hot sizes. Lua table hash nodes
are 40 bytes, so hash tables produce allocs at 80, 160, 320, 640,
1280, 2560 (2..64 nodes). Added classes at these exact sizes.
Fixed size class LUT: was 16-byte granularity (size 22 mapped to
class 32 instead of 24). Now direct 1:1 lookup table, 4KB, one
byte load per alloc with zero computation.
Other changes:
- Removed pool_ctx==0 passthrough and seg_val==0 fallback (dead code)
- Removed pool_ctx param from slabFree/slabRealloc
- Added optional PROFILE mode with size histogram dump to log file
- Added AllocBench addon (/allocbench) for in-game benchmarking
Hook 0x686940 (testAABBFrustum): SSE center/extent P-vertex method replaces
x87 sign-bit corner selection. ~160M cycles in profile, 50% rejection rate.
Hook 0x686180 (FrustumCullBoundingBox 8-corner): SSE replacement transforms
all 8 AABB corners through view-proj, horizon buffer scan with 4-wide test.
Fix behind-camera bug in both FrustumCullBoundingBox variants (0x686000,
0x686180): original returned 0 for behind-camera objects, callers interpret
0 as "visible, render it". Now returns 2 (occluded) when all geometry is
behind the camera (Z < 0 in clip space).
Replace 4-byte per-allocation header with 64KB segment table for O(1)
class lookup. VirtualAlloc guarantees 64KB-aligned pages, so ptr>>16
maps directly to the class index. Zero per-allocation overhead.
- VirtualAlloc for slab pages (guaranteed 64KB alignment)
- VirtualAlloc for large allocs (>4096) with size+magic header
- pool_ctx==0 passthrough for non-Lua callers
WoW's memory_pool_allocate (0x6FAE90) does O(classes * pages) linear
scan on every free/realloc to find which pool owns a pointer. Our slab
stores slot size in a 4-byte header for O(1) lookup.
Baseline 2000ms -> testing with allocator enabled next.
- 15 size classes (16-4096) vs WoW's 6 (16-256)
- @memcpy for cross-class realloc (WoW uses manual dword loop)
- Large allocs (>4092 usable) fall through to game heap
- No C dependencies (replaces mimalloc which couldn't cross-compile)
NPC priority pass now whitelists standard vanilla interaction flags
(0x7FFF: gossip through repair) instead of treating any non-zero
npc_flags as interactable. Fixes void zones and visual-effect NPCs
(e.g. C'Thun Portal, npc_flags=0x2000000) blocking player selection.
Cascade raycast disabled entirely in GM mode (PLAYER_FLAGS_GM) so
GMs get unfiltered targeting.
Remove the core mutex that prevented secondary DLLs from installing
shared infrastructure hooks (Lua registration, file serving, engine
init/shutdown). zhook already has explicit E9-chain detection -- each
DLL's trampoline chains to the previous DLL's detour, and per-module
mutexes still prevent duplicate module hooks. DLL_PROCESS_DETACH fires
in reverse load order so detach unwinds correctly.
weirdperformance's filecache hook (File_FindInArchive) stayed active
during game teardown, returning stale archive/block pointers after
Storm freed its MPQ archives. This corrupted heap metadata, causing
ERROR #124 (SGroupPtr invalid block) on game close.
Set remove_on_shutdown=true so the hook is detached during
logoutDetour before Storm archive teardown begins.
Also wires up superweirdo module and clickthrough lateInit.
The base DLL target was compiled with SSE4.1+FMA+AVX, which meant LLVM
could emit FMA instructions (VFMADD etc.) anywhere in the main module.
This crashes on pre-Haswell CPUs that lack FMA support (e.g. i7-3930K).
Split into two targets: base target (SSE2 only) for all main code, and
sse_target (SSE4.1+FMA+AVX) only for the separately-compiled SSE object
files that replace hot game functions. This ensures FMA/AVX instructions
are confined to the performance modules.
Fixes#6.
Filter ritual (type 18) and mage portal (type 22) GOs whose creator
is a player not in the local player's party or raid. Completely
unclickable, not just deprioritized.
Also:
- Add wow.isInGroup() shared group membership check (party + raid)
- Add party/raid addresses to shared offsets.zig
- Fix dpslog party member GUID address (was 0xBC7600, correct: 0xBC6F48)
Verified from client's is_player_in_allowed_list @ 0x4e7f70
- Restore MSVC ABI (was accidentally GNU since v0.6.0, broke .CRT section)
- Replace game allocator with Windows process heap for filecache and
libdeflate malloc/free - game allocator not initialized during DllMain
when injected via CreateRemoteThread
- Defer timer calibration (Sleep 500ms) to lateInit - blocks under
loader lock during DllMain
- Remove exported malloc/free symbols from DLL
- Eliminate addObject compilation units for SSE files - direct @import
with AVX target instead
- Heap-allocate filecache (was 9.3MB static BSS)
- Strip transform44 of performance/ externs, pure profiling only
- Rename performance/ to weirdperformance/ to match module convention
- Skip default-off modules in all-variants build step
- Remove dead debug vars and stride logging from particle_sse
Removed deferred divide from both ray-tri functions -- the det-scaled
epsilon comparisons lose precision for near-parallel rays, potentially
accepting triangles the original rejects. Downstream SetupBoxFrustum
then writes to wrong globals near 0xCE6738 (SGroupPtr), causing
ERROR #124 on exit.
inflate_hook: replaced manual FS:0x24 thread ID pool with Zig native
threadlocal. Fixes potential crash on Wine where FS segment layout
may differ.
Build: switched from GNU to MSVC ABI (4-5K smaller per DLL, no
.eh_frame unwind tables). Added setjmp.h stub for libdeflate MSVC
build. Added noperf build variant (zig build noperf).
GUID lookup cache: 4096-entry direct-mapped with proper invalidation.
MoveObjectToDeletedList (0x464920) hook evicts entries AFTER the
original runs (prevents re-caching from internal FindObjectByGUID call).
DestroyObjectManager (0x467700) hook flushes on zone change/logout.
93% hit rate, 1.15x speedup on FindObjectByGUID (10K+ calls/frame).
Glyph shadow cache removed: game has internal glyph cache at
GetOrCreateCharacterGlyph (0x5CA2D0). Our hook only saw cache misses
(~30/frame), providing no benefit. The 3.65% perf profile was the
game's own hash table work, not redundant computation.
Standalone guidcache module for isolated testing.
FindObjectByGUID (0x464890): 4096-entry direct-mapped cache with
validate-on-hit. 97% hit rate at 10K+ calls/frame, reducing from
0.8% to 0.3% frame time. Validates cached pointers by checking
GUID at obj+0x30/+0x34 on every hit.
AddToSpatialGrid (0x6816F0): SSE rewrite attempted, no measurable
gain (memory-bound linked list ops dominate). Not shipped.
rayTriIntersectIndexedInt (0x7C2C40): SSE Moller-Trumbore with deferred
divide, matching original's epsilon thresholds. Parity-tested against
original for edge hits, backfaces, parallel rays, and per-triangle t/uv.
JMP-patched in weirdperformance.
Added transform44 profiling hooks for ray_tri_indexed_int (0x7C2C40)
and ProcessStaticObjectsCulling (0x683BF0).
Bench: added rayTriIndexedInt bench with exhaustive parity tests.
PerformSpatialCulling (0x6B8C60): Zig rewrite with SSE outcode
computation. 1.4x speedup, JMP-patched in weirdperformance.
performCollisionDetection (0x6B88E0): fully inlined SSE Moller-Trumbore
ray-triangle intersection, eliminating 4 SetVector3 calls and the
ray_tri function pointer call per triangle. ~22 cyc/tri in bench.
Both graduated from transform44 A/B testing to production JMP patches.
entity_sse.zig: reimplementations of UpdateEntityAndChunksPositions and
updateEntitiesInBounds (A/B tested, 1.2x bench, not shipped - memory
bound with negligible real-world gain).
clickthrough: fixed CheckObjectTypePermissions hook from fastcall to
thiscall (ECX preservation), fixed ClntObjMgrObjectPtr from fastcall(4)
to fastcall(5) with correct arg count.
bench: added performCollisionDetection bench with synthetic mesh data
and patched FindOrCreateHashEntry stub.
Replace the re-raycast approach with a priority cascade: loot > GO > NPC
> normal. Each pass uses custom flag bits (0x01/02/04 in upper byte) that
our CanTargetEntity hook reads to exclude non-matching objects at the
raycast level. Terrain/WMO occlusion applies per pass.
Hook CanTargetEntity (0x480610) for per-object filtering: strips custom
bits before calling original, then checks lootable/interactable/NPC
based on which pass is active. Objects that don't match the current pass
return NULL (invisible to raycast, ray continues through them).
Fix isLootable: UNIT_DYNAMIC_FLAGS was at wrong descriptor offset
(0x96*4=0x258, should be 0x8F*4=0x23C per server UpdateFields index 143).
Also add UNIT_DYNFLAG_TAPPED and UNIT_DYNFLAG_TAPPED_BY_PLAYER offsets.
Fix isLootable base pointer: use getDescriptor (obj+0x08) not
getUnitDescriptor (obj+0x110), consistent with getNpcFlags and all
other descriptor reads in the project.
Tail vertices V0/V1 read texture U/V from wrong addresses (no offset
from base instead of +8/+16 stride). Assembly reads 0x87D734/738 for
V0 and 0x87D73C/740 for V1, not the base 0x87D72C/730. This made both
base vertices share UVs, collapsing the texture and causing light beams
to taper to a point instead of maintaining width.
Also fix transformVec4 input: must be [4]f32 with w=0.0, not [3]f32.
The function reads all 4 components including [EDX+0xC]. Reading past
a 3-element array produced garbage w values that corrupted the velocity
transform, causing tail particles to extend wildly.
Fixed in both SSE and reference implementations.
Core hooks (file serving, Lua registration, engine init/shutdown) are
shared infrastructure that other loaded DLLs depend on. Only release
them in uninstall (DLL_PROCESS_DETACH), never via the runtime API.
ModuleObjects.linkFor dispatches object/library linkage per module name.
Called from both the main weirdutils build and the variant loop, so adding
a new SSE object only requires updating one place. Fixes variant builds
for weirdperformance, transform44, silicon, and ssemaths.
When both weirdutils.dll and a standalone module DLL are loaded,
the core mutex (Local\WeirdUtils_weirdutils_<PID>) ensures only one
installs the shared hooks (file serving, Lua registration, engine
init/shutdown). Module hooks are unaffected (own mutexes). Addon
registration runs for every DLL since LoadAddonTOC is idempotent.
Move fileFindDetour from main.zig into filecache.zig with its own
install/remove. Remove file_cache import and all references from
main.zig. Rename module to weirdperformance throughout (build flag,
module_name, source file, main.zig variable).
Move filecache.zig and timer_fix.zig from standalone modules into
src/performance/. Filecache no longer has its own hooks/mutex — stats
are dumped by performance's worldupdate hook. Timer calibration runs
during performance install instead of transform44. Remove filecache
from build module list (enabled automatically with performance).
Merge DLL_README sections into single Performance entry.
Replace custom getNameFromGUID with wow.getNameByGUID, use wow.readGUID
for GUID comparisons, remove getPlayerGUID wrapper. Drop unused
PARTY_MEMBER_GUIDS, FN_NAME_CACHE_LOOKUP, NAME_CACHE_OBJ from offsets.
Root cause of crashes was thread safety: libdeflate's decompressor struct
has mutable decode tables rebuilt per-block, so sharing between main thread
and FMOD audio thread caused wild writes. Fixed with a thread-local pool
keyed by Windows thread ID (FS:[0x24]).
Switched build ABI from msvc to gnu — eliminates stub headers, libdeflate
uses Zig's bundled MinGW libc. Relaxed zlib header filter from == 0x78 to
(CMF & 0x0F) == 0x08 to catch both 32K and 4K window sizes.
Benchmarked: stock=2681ms, per-call-alloc=1304ms (2.05x), tls=1194ms (2.2x)
on 84k calls. Zero errors, zero fallbacks in production.
DecompressData_WithOptions modifies the input buffer during overlap
handling. Our hook must save the input before calling the original,
then pass the saved copy to libdeflate. Also documents all compression
types, function signatures, and bugs found during integration.
libdeflate sanity test passes (hello → hello, cpu_features=0x8000001F).
All 93K decompression calls during loading are pure zlib (type 0x02).
Header bytes 78 9C confirm standard zlib-wrapped deflate. No PKWare,
bzip2, or ADPCM observed.
Volume: 170MB compressed → 380MB decompressed in 2.9s during loading,
~5K calls/period during gameplay.
Format: [0x02] [zlib_stream...] — straightforward for libdeflate.
Previous crashes were from reading out_size AFTER original modified it
and from writing to out_buf unsafely.
Hooks DecompressData_WithOptions (0x661A80) to run both the original
zlib inflate and libdeflate on the same data. Times both, logs the
comparison every ~450 frames, counts mismatches.
malloc/free provided via game's Storm memory manager (SMemAlloc/SMemFree)
since we link without libc on the msvc target.
libdeflate compiled with -g0 to avoid .debug_frame COFF section name
warning. AVX-512 disabled (32-bit), SSE2/AVX2 paths active.
Vendored libdeflate decompress-only C sources into src/performance/.
Compiled as static lib (x86-windows-gnu for libc headers), linked into
the DLL. AVX-512 disabled (not available on 32-bit x86), SSE2/AVX2
paths active.
Target: replace WoW's embedded zlib inflate (~3% CPU) with libdeflate's
~2.3x faster implementation. Hook integration next.
- performance module default=true, transform44 and silicon default=false
- Silicon JMP patches (22 functions) moved into performance.zig directly
- Single flag: 'zig build' gives all optimizations, zero profiling
- Dev mode: 'zig build -Dtransform44=true' for A/B testing
Clean module at src/performance/ with all verified permanent
optimizations and zero profiling overhead. No rdtsc, no A/B testing,
no probe counters, no dumpStats.
Hooks: transformMatrix4x4 bone SSE, RenderParticleSprites SSE,
GetOrCreateCharacterGlyph cache, OnWorldUpdate cache reset,
teardown guard.
SSE compilation units (bone_sse, particle_sse, clip_sse, silicon_sse)
moved to src/performance/ as canonical location.
Build: zig build -Dperformance=true -Dsilicon=true
(transform44 module remains available for profiling/development)
Color channel packing missing & 0xFF after >> 14 extraction — upper
bits bled into adjacent channels causing broken particle fading.
Same class of bug as the alpha output fix earlier.
Added particle_sse_reference.zig (faithful recreation from commit
574f96f) as a separate compilation unit for correctness comparison.
Faithful recreation of RenderSpriteQuads (0x5A0F50) with hoisted
invariant division and inlined DisplayMode_CalculateOffset. No
measurable improvement — cost is dominated by getAdapterInfo (7
sub-calls to D3D device per invocation × 3203 calls/frame) and
DrawPrimitive/DrawIndexedPrimitive virtual dispatch.
Added decompilation and assembly dumps for future reference.
Faithful recreation attempt of SetupParticleRendering (0x7B3D20).
All game function CCs verified from assembly. Vertex buffer setup
works (8 verts produced), but D3D draw submission doesn't produce
visible output. Disabled pending investigation of GfxDeviceMethod
param struct layout. The function remains as timing-only pass-through.
Fixes found during work:
- max_particle_sprites global: 0xCF58F4 → 0xCF5B60
- billboard_matrix global: 0xCF5898 → 0xCF5888
- index_buffer_6/12: 0xCF58D0/D4 → 0xCF5BAC/0xCF5AF4
- BuildIndexBuffer takes renders_count not field_28
- Identity matrices must be mutable (game writes to them)
10% of frame time, ~275 calls/frame, ~5800 cyc/call. Builds identity
matrices (32 stores of dead work), does 1-5 matmul calls, copies to
g_worldMatrix global. Translation matrix is identity+offset — matmul
chain can be simplified to direct computation.
- V4 store (vmovups) writes xyz+color as one 16-byte op instead of 4
scalar stores. Unaligned but still 1 μop on modern CPUs.
- inline for unrolls 4-vertex loops, letting LLVM schedule stores
across vertices and fill pipeline bubbles.
- Hoisted world_pos to locals to prevent array re-reads.
- A/B: BASELINE ~433ms → CUSTOM ~299ms (~31% per-period reduction).
Total from original: 520ms → 299ms = 42% reduction.
- Detected interleaved 24-byte vertex layout: xyz(12)+color(4)+uv(8).
Fast path writes 6 sequential u32s instead of scattered stores.
- Normal stride=0 (shared global) — write once in writeback, not 4x.
- Added stride_info export + logging for vertex layout analysis.
- A/B: ~30% peak reduction, baseline also improved due to less overhead.
MarcelineVQ