Each slot probe in Table::Border was lua_rawgeti + lua_type + lua_settop,
three API calls per slot, on every transpiled `#t` and every luaL_getn
heal (table.insert / getn / remove / unpack / `...`). Resolve the Table*
once with luaA_indexAcceptable and probe with the engine's own raw
integer read, luaH_getnum, which is what lua_rawgeti resolves through.
Find now mirrors 5.1's luaH_getn: a nil in the last array slot bisects
inside the array part, otherwise unbound search from the array size.
Holey constructors now match real 5.1 (`#{1,nil,3}` is 3, was 1).
The getn heal's answers are unchanged. 100k `#t` on a 1000-entry table:
0.011s -> 0.004s.
The getn heal read a nil at the stored last slot as proof that nobody
maintained the length. For a weak-VALUED table that is false: the
collector clears those slots behind the writer, so `luaL_setn` is still
exactly what the last append left and the holes below it come and go
between two reads.
Compost-2.0 is the casualty. `secondarycache` is `__mode = "v"`, filled
with `table.insert`; `GetTable` walks it with `pairs` and calls
`table.remove(cache, i)` on a live index. Once the GC clears slot 1 and
the stored last slot, the heal bisects to border 0, `luaB_tremove` takes
its `e == 0` empty-table early-out and returns NOTHING, and the next line
indexes with the nil it got back -- "table index is nil" at
Compost-2.0.lua:81, once per recycled table, in every addon that pulls
Compost in through AceLibrary.
The border search is what exposed it (5aa382b's table.insert switch is
not involved -- Compost never appends a nil, so no mark ever applied to
its tables). The walk-down it replaced returned the HIGHEST populated
slot, which is always at or above the live index the pairs loop holds, so
remove still found something; a bisect returns some border and probes its
way down to 0 through the collected slots. That tradeoff was known and
written down -- a weak cache read through table.remove is where it bites.
So gate the heal: `__mode` mentioning "v" returns the stored length
untouched, as rule 3, ahead of the mark. It costs one getmetatable and a
raw `__mode` fetch on the already-rare unhealthy path, so the bisect keeps
its win everywhere else, and the mode is read raw -- the same fetch the
collector's own check uses.
New plumbing: Game::Lua::GetMetatable, from the offset already catalogued
in LuaCAPI.md. LuaEnv::Fenv had rolled its own copy; it now shares this
one.
Verified in game: a three-slot weak table with slots 1 and 3 collected
reports getn 3 and hands back its live slot (0 and nothing before), and
the Compost errors are gone.
The heal's fallback walked down from the stale stored length one rawgeti
per slot — O(stale gap) — and since a length read never writes back, it
did so on EVERY luaL_getn read until the next append re-synced the
length. A table cleared by `pairs` and then polled through getn / concat
/ sort / unpack paid the full walk each time.
The search is now the doubling-then-bisect that luaH_getn uses for 5.1's
`#`, in one shared header (table/Border.h) that both `__len` and the heal
call; the copy that lived in Transpile.cpp is gone. The mark gate is
unchanged in meaning: `n - b == 1` said "slot n-1 is populated", so it is
now that one probe, tested before the registry lookup. For a fully
cleared table and a table stale by one the answer is identical; a stale
table with holes gets some border rather than the highest one, which is
what 5.1 itself returns for that table.
Verified in game: the reserved slot with a hole present (5, "c"), the
stale-by-one recycle (7, 2), the full clear (9, 1), 1,000 getn reads on a
cleared 20,000-slot table with no hitch, and `#` through the shared
search.
Table::Length recorded deliberate `table.insert(t, nil)` appends (#36) with
a MinHook on the engine's `luaB_tinsert`. The hook is unnecessary: that
function has exactly one xref — the table-lib luaL_reg entry — so no engine
C code calls it, and every call arrives through one of two Lua values:
`table.insert` (lib open) and the global `tinsert`, which the engine's
embedded compat snippet binds by value (`tinsert = tab.insert`, .data
0x008722E8) from FUN_00703b80 right before FUN_LOAD_SCRIPT_FUNCTIONS
(calls at 0x0048fe97 / 0x0048fe9c; on glue 0x0046a87b / 0x0046a880).
So the marking function is now registered over `table.insert` and calls
`luaB_tinsert` directly, and `tinsert` is re-bound to the same closure by
value through a new `Game::Lua::RegisterGlobalAlias` — the snippet's own
idiom, so `tinsert == table.insert` stays true. Both Lua states get the
pair; the `luaL_getn` heal is unchanged. One MinHook fewer, on a hot Lua
function.
Verified in game: the identity triple (`tinsert == table.insert` and both
`_G.ClassicAPI` mirrors), the #36 reserved slot through both names, the
#39 heal, the positional form, and the error path through the new frame.
The v1.12.8 gap heuristic (keep the stored length when it overshoots the
border by exactly one) protected `table.insert(t, nil)` slots (#36) but
broke every 5.1-dialect table that is stale by exactly one — a
pairs-cleared one-element table, or a recycled table whose previous life
was one slot longer. Tablet-2.0's del()/copy() pool produces the latter on
every tooltip refresh, so FuBar/Ace2/Tablet menus and tooltips shifted by
one with a nil hole at slot 1. The two states are byte-identical, and both
consumers reach the heal through table.insert — neither the table nor the
caller can split them.
Only the writer knows which dialect the table speaks, and the write is
observable: co-hook the engine's table.insert (FUN_LUA_TABLE_INSERT =
0x7FB6B0, verified against the table-lib luaL_reg pair in .data and a
decompile matching 5.0.2 tinsert line for line) and record
{table -> stored length} in a weak-keyed registry table whenever the
two-arg form appends a literal nil. Rule 3 now heals every stale table to
the border (the 5.1 answer) and keeps the stored length only when the gap
is one AND the mark still matches. Stale marks die on their own: any later
insert/remove moves the stored length off the recorded value, and the weak
key collects with the table.
Verified in-game with a 21-test suite (directed regressions for
#29/#36/#39, byte-identical twin discrimination, metamethod isolation,
coroutine-side marking, and a seeded differential fuzzer against an
executable model of the contract): 15/21 on the broken build, 21/21 with
the fix, #36's nil-append cases among the passes both times.
New plumbing: LUA_SET_METATABLE (0x6F4020, from LuaCAPI.md) and
Game::Lua::SetMetatable for building the weak mark table from C.
Fixes#39
The stale-length heal (Length.cpp) could not tell a cleared-and-reused
table from a deliberate table.insert(t, nil): both leave the stored
length pointing at a nil slot with no explicit n field, but they want
opposite answers. Rule 3 healed both, so the next table.insert on a
single-nil-appended array wrote over the nil and shifted every later
element down by one -- silent corruption in Ace2-era addons (Waterfall's
{key1,val1,...} arg builder inserts a nil whenever an option has no
passValue, the common case), with no error anywhere.
The issue's preferred fix -- make table.insert always use the raw stored
length -- reintroduces #29: Dewdrop clears `values` with a pairs loop +
no-op table.setn, then rebuilds it via table.insert, and that append is
exactly what the heal fixed. Both consumers reach the heal through
table.insert, so the caller cannot distinguish them; only the staleness
gap can. A cleared table is stale by many slots; a deliberate nil append
is off by exactly one. So heal only when the stored length overshoots
the border by more than one slot, and keep it otherwise.
Verified in-game via CAPI_DemoTableLengthHeal (AddOnCompat.lua): nil
append preserved (getn=3, t[2]=nil, t[3]=b), Waterfall pairs intact, and
the cleared-table rebuild still heals to slot 1.
Fixes#36
The 5.1 syntax backport made the Ace2-era Lua-version probe pass:
local lua51 = loadstring("return function(...) return ... end")
and true or false
Every Ace2 library gates on it, and Dewdrop-2.0 (the menu library
behind FuBarPlugin minimap buttons -- SuperAPI, NampowerSettings) uses
it to no-op table.setn: correct on real 5.1, where lengths are computed
borders, but wrong on this VM, whose table functions still read 5.0's
out-of-band length. Cleared-and-reused menu tables kept stale lengths,
table.insert appended past the nil'd slots, and the menus silently
broke ("clicking the minimap button does nothing"). Bisect confirmed:
forcing the probe back to 5.0 (VarargExpansion off) restores them.
Since we made the client PASS 5.1 probes, 5.1 length semantics must
hold. The engine funnels every length read through luaL_getn
(LUAL_GETN = 0x6F5050 -- table.insert/getn/remove/concat/sort/foreachi,
plus our unpack), so one co-hook closes the gap for every consumer:
1. reported slot populated -> unchanged (one rawgeti of cost);
2. t[n] nil but t.n is a number -> unchanged: an explicit n field is
the 5.0 vararg-arg contract, where trailing nils are intentional
(unpack({10,nil,30,n=3}) keeps all three slots);
3. otherwise the out-of-band length is stale (keys cleared without
setn) -> return the true border, the 5.1 answer. Read-only, no
write-back: reads never mutate state, and the engine's own
table.insert re-syncs the stored length on the next append.
table.setn keeps working; the heal only changes the answer when the
stored length points past the last populated slot -- exactly the case
where 5.0 and 5.1 disagree.
Docs: API.md "Stale table lengths (5.1 healing)" section + README
behaviors row.
Fixes#29
Systematic sweep of the Lua 5.1 stdlib delta against what ClassicAPI
already backports found five gaps, all confirmed by in-game probes:
* unpack(list [, i [, j]]) -- 5.0's unpack silently ignored the range
args, so unpack(t, 2) returned the WHOLE table (wrong values, no
error; the only silent gap). Replaced with a port of 5.1's
luaB_unpack: ranges via luaL_getn + lua_rawgeti, so a vararg arg
table's n field keeps its embedded-nil count. Both states.
* string.gsub table replacement -- 5.0 accepts only string/function
and errors on a table. Shim on string.gsub AND the bare gsub global:
a table becomes an adapter closure over the engine's own gsub
(FUN_LUA_STR_GSUB = 0x007FD0E0, pinned by registration-table
position, name string, and decode). Because this engine's add_value
drops a non-string function result (match -> empty) where 5.1 keeps
the match on nil/false, the shim wraps the pattern in one outer
capture (^/$ anchors kept outside) so the closure receives the whole
match and can hand it back for missing keys -- exact 5.1 keep-match.
Patterns with a %1..%9 backref skip the wrap (renumbering hazard);
their missing-key behavior stays the engine's (documented residual).
* collectgarbage 5.1 options -- "count" -> gcinfo, "collect" ->
original collectgarbage(0), "step" -> loop-safe true, stop/restart/
setpause/setstepmul -> no-op 0, numbers forwarded untouched. The
originals are function objects stashed in the state's registry
before the overwrite, so no new engine offsets and each /reload
re-stashes the fresh engine original. Both states.
* table.maxn -- 5.1 luaB_maxn port (largest positive numeric key,
full next-walk so nil holes don't stop it). Both states.
* coroutine.running -- returns the running coroutine, nil on the main
thread. 5.0 has no lua_pushthread, so the thread TValue (tag 8 +
the running L, this build's {tt@0, value@+8} layout) is written to
the stack raw; main-thread detection via the registration-time L.
Verified in-game: ranged/holed/n-field unpack shapes, all four gsub
shapes (populated table, keep-match on missing key, whole-match key,
anchored pattern), maxn past a nil hole, collectgarbage("count"/
"collect"), and coroutine.running from main + inside a wrap.
Docs: five new API.md sections + TOC entries; README Lua list updated.
Backport of retail's table.count (added 11.2.5): returns
(numTableNodes, numArrayNodes, maxArrayIndex) -- total entries, entries
with an integer key in [1..numTableNodes], and the largest positive
integral key (0 if none). Pure lua_next walk, no Lua-internal layout
dependency; keys read via ToNumber only (ToString on a numeric key
would corrupt traversal). All five cases verified to match retail,
including negative-key handling ({[-3]=x} -> 1, 0, 0).
table.wipe nil'd every key via lua_next + rawset(nil) but left the
Lua 5.0 length counter untouched. In 5.0 a table's table.insert/getn
length lives out of band from its keys (a t.n field, else a weak
sizes[t] entry in the registry's LUA_SIZES table), so it survived the
clear. After a wipe, getn stayed at its old value and subsequent
table.insert appended past the wiped slots -- t[1] stayed nil. The
3.3.5 (Lua 5.1) twipe this was ported from has no getn legacy, which
is why the ported logic looked complete.
Fix: call luaL_setn(L, 1, 0) after the clear loop. Bound the engine's
luaL_setn (FUN_006f4ea0, verified against 5.0.2; located via
table.insert FUN_007fb6b0) as Game::Lua::SetN.
Repro before: table.insert(t,1); table.insert(t,2); table.wipe(t);
table.insert(t,3) -> getn 3, t[1] nil, t[3]=3.
After: getn 2, t[1]=3, t[2]=4. Verified in-game.
The repo was internally inconsistent: the LICENSE file is GPL v3 but every
src/ header carried the LGPL v3 notice and the README said "LGPL v3 or
later". LGPL was also the wrong fit — this is an injected mod plus a
bundled addon, not a library other code links against, so LGPL's linking
exception buys nothing.
It's also required: src/voicechat/Tts.cpp is ported from awesome_wotlk's
VoiceChat.cpp, which is GPL v3, and a derivative of GPL code cannot be
relicensed under more permissive terms. Dependencies stay compatible —
MinHook (BSD-2) and the nampower-mirrored packet shapes (BSD) are
GPL-compatible, and SAPI is covered by the system-library exception.
Swept "GNU Lesser General Public License" -> "GNU General Public License"
across all 257 src/ headers and changed the README license line to match
the GPL v3 LICENSE file. Comment-only; no build impact.