The mana spark modelled a tick clock that does not exist. Spending mana
restarted the sweep with a five-second span, then it cycled at two seconds
until the next spend -- so every cast reset the phase, and the spark drew a
period the server never runs. There is one regen timer for every power,
REGEN_TIME_FULL (2s), re-armed with `+=` in Player::RegenerateAll and never
touched by casting. The five-second rule is a separate countdown that
SetLastManaUse arms on any mana-costing cast; it changes what a tick pays,
never when ticks land. Inside the window the tick still fires and still pays
mp5 plus whatever share of spirit the player keeps.
That share is common and not computable here. The amounts are readable now
(C_Spell.GetSpellEffectInfo), and the talent, racial and buff sources could be
enumerated -- but the Casting Regen item ladder is equip auras that never
appear in the buff list, and m_modManaRegenInterrupt is never sent. A floor
that is wrong whenever the player wears such an item is worse than showing
what happens: the spark starts an FSR window dim and goes solid on the first
tick that lands inside it.
The sweep is a free-running phase lock on that one clock. Every positive mana
delta used to re-anchor it, and most are not the tick -- Illumination refunds,
Judgement of Wisdom, potions, a Mana Spring totem on its own 2s clock at
another phase -- so the spark snapped back to the edge mid-sweep several times
a cycle. A gain re-anchors only if it lands within 250ms of the predicted
boundary. Two rejected gains exactly one period apart are the real clock and
move the lock, which is how a wrong first lock heals. Rollover advances by
whole periods from the anchor instead of restarting from now, so frame timing
no longer accumulates into drift the lock then has to chase.
Even a correct tick used to hitch the spark: re-anchoring to arrival time
moved it by the jitter, 20-80ms, right at the wrap. Inside an 80ms band the
tick now confirms the sweep and leaves the anchor alone; only the excess is
pulled in. With no self-drift left there is nothing systematic in that band.
The rule is armed from UNIT_SPELLCAST_SUCCEEDED filtered through
C_Spell.GetSpellPowerCost, mirroring Spell::TakePower's condition, rather than
from a mana decrease -- Mana Burn lowers mana without arming the rule.
UNIT_SPELLCAST_CHANNEL_STOP re-arms it, since Unit::Update will not expire it
while the spending spell is still channeling.
The energy period no longer names a spell. Player::RegenerateAll sums
SPELL_AURA_MOD_ENERGY_REGEN_TIME and takes amount * agility / 10 off the timer
in milliseconds; everything in that sum is readable here. The player's own
spells carrying aura 217 come from C_SpellBook.GetPlayerSpellsByAura, passives
count while known (current rank only), anything else while it is up in the
buff list, and each amount is basePoints + baseDice from the effect data. So
Blade Rush is found without GetTalentInfo(2, 16) -- an ordinal into the talent
grid that does not fail when the tree changes but reads another talent's rank
(index 16 is Vigor one tab over) -- and a retune or a second source needs no
edit. The sum is cached and dropped on SPELLS_CHANGED and
PLAYER_AURAS_CHANGED, the two events its two halves move on, so a tick costs
no spellbook walk; agility stays live, it is one call.
The CHAT_MSG_SPELL_SELF_BUFF "You gain ... Energy from" filter is gone. The
phase lock rejects Relentless Strikes and Thistle Tea the same way it rejects
a paladin's refunds, it was English-only, and its ignore-the-next-gain logic
could eat a real tick if events reordered.
The central loop calls the per-plate update for every visible plate ~100 times
a second, and the plate's own throttle sat at the bottom of the function. Above
it ran the whole classification: a GetAlpha call, a castbar IsShown, a cast
lookup, and one or two libthrottle:Get resolutions -- all to work out which
throttle applied, on plates throttled to 10fps that were going to return
anyway. Forty plates makes that four thousand times a second.
Add a cheap gate first, against the floor across all four throttle categories.
Nothing that would have updated can be turned away by it -- a plate past its
real throttle is necessarily past the minimum -- and the category-specific gate
still runs after the classification, unchanged. Event flags are read before
both and bypass both, as they did before.
libthrottle:Get is not a lookup: it walks the saved variables, a defaults
fallback and a preset table, and can build a "<category>_custom" key. Resolve
the four categories in CacheConfig instead, where config changes already land,
and read them from cfg on the hot path. That covers the two calls in the
castbar branch as well.
The castbar hide moves below the cheap gate, which is safe rather than merely
tolerable: a plate showing a castbar sets isCastingNonTarget, which selects the
castbar throttle -- one of the four the floor is derived from -- so it always
clears the first gate and reaches the hide at its normal rate.
Every unit frame registered fifteen unit events globally and then sorted them
out per event:
arg1 == this.label .. this.id or (UnitGUID and arg1 == UnitGUID(this.label .. this.id))
A string concat -- and on a miss a second concat plus a UnitGUID call -- for
every frame, for every unit event fired by anything in the world. Forty-five
frames in a raid, on every health tick of every raider, to conclude "not mine".
UpdateVisibility is the single place a frame's label and id are ever assigned,
and it already pushed the result to SetAttribute("unit", ...). It now caches
the unitstr and points the subscriptions at it, so there is one hook rather
than a set of transitions to catch, and visibilityscan re-running it five
times a second makes it self-healing. A frame that is not in use drops its
unit events entirely.
EnableEvents no longer registers them: a registration keeps its kind, so
plain-registering first would make every later RegisterUnitEvent a no-op.
UNIT_PET and UNIT_HAPPINESS stay plain deliberately. Their branches key on the
frame's label, and UNIT_PET's arg1 is the pet's OWNER -- "player" for a "pet"
frame -- so filtering them by the frame's own unit would drop them.
The GUID alternative is gone. These events fire once per token that resolves
to the unit AND once with the raw GUID, so the token form always arrives and
the GUID clause was only ever a duplicate wake. It dates from b7a0912e, before
ClassicAPI's TokenObserver made synthetic tokens first-class event sources.
Two teardown paths needed help. UnregisterAllEvents drops the filters with the
registrations, so the cached unit has to be cleared or the next
UpdateVisibility believes they are still in place. And PLAYER_LOGOUT silences
the frame while visibilityscan keeps ticking, which would re-register
everything it just dropped -- the crash 132 that branch exists to prevent -- so
the frame now takes itself off the scan list.
The central handler registered UNIT_AURA, UNIT_FLAGS and the four
UNIT_SPELLCAST_* globally, so every unit event in the world woke it for a
"^nameplate" prefix test, and each of the four branches that survived the test
then resolved the plate back out of arg1 -- UnitGUID + plateByGuid, or
GetNamePlateForUnit.
Register them per plate instead, on the plate's own frame against its own
token. The token is the subscription, so the event only reaches the plate it
concerns and `this` is already that plate: the aura branch is now one
assignment where it used to be a prefix test, an API call and a table lookup.
This is also what makes the approach viable at all. A central listener would
have had to name every token it might care about, and nameplate slots have no
cap -- Events.cpp keeps them in a vector that grows by push_back, with only a
"<80 even in AV-scale scenes" comment for sizing. Any nameplate1..N list would
have been a guess that fails silently in exactly the crowded scenes where
plates matter.
Lifecycle: NAME_PLATE_UNIT_ADDED points the registrations at the new token,
replacing the previous unit rather than stacking, since RegisterUnitEvent over
an already-filtered registration swaps the units. NAME_PLATE_UNIT_REMOVED
unregisters, which is required -- freed slots are reused, and a stale token
would feed the frame another unit's events.
PLAYER_LOGOUT now tears down the plates too. It silenced only the central
frame before; leaving forty plate frames dispatching through logout is the
crash 132 that branch exists to prevent.
UNIT_HEALTH/UNIT_MAXHEALTH register through RegisterUnitEvent for mark1
through mark8, so the scanner stops waking for every unit in the world. Eight
tokens is past the four other clients accept; ClassicAPI takes as many as it
is given, so this file is fork-only.
That makes arg1 always one of eight known strings, so the handler looks the
row index up in a table built alongside markerTokens, instead of running
string.match + tonumber -- which allocated a capture and parsed it on every
health tick of every marked unit.
The fallback poll was the larger cost and the filtering did not touch it: a
full eight-row rebuild once a second for the whole session, whether or not a
marker existed anywhere. Raid markers are a group feature, so the scanner now
watches PARTY_MEMBERS_CHANGED / RAID_ROSTER_UPDATE and creates or cancels the
ticker with group membership -- outside a group no timer is queued at all,
rather than one waking every second to return early. It is deliberately not
keyed on a mark being visible: a marker set on a unit that is out of range
shows no row, and that is precisely the case the poll exists to catch.
The mark-token capability probe at the top goes too. ClassicAPI is a hard
dependency, so the tokens are always there to resolve.
ClassicAPI's RegisterUnitEvent registers for an event but only delivers it
when arg1 is one of the given units, so a handler for one unit stops waking
for every other one in the world. Convert the 26 registrations whose unit set
is fixed and known at registration time.
The rule throughout is register the superset and keep the handler's own
check: the filter narrows what arrives, it does not decide what to act on.
castbar is the case that matters -- the player's own casts only ever fire
arg1=="player", never "target"/"focus", so the target and focus bars must
register "player" too, and their UnitIsUnit test is what still rejects a
"player" event while you are targeting a mob.
Two that are not one-line swaps:
actionbar's two unit events are keys in tables that also drive dispatch, so
they cannot leave the tables -- and they cannot be plain-registered first,
because a registration keeps its kind and RegisterUnitEvent over a plain
one stays plain. They route through a small event_units map instead.
swingtimer's UNIT_DIED carries a GUID, not a token. The filter is a plain
case-insensitive string compare with no GUID resolution, so the player's
own GUID -- fixed for the session -- filters it exactly.
Guards that are now unreachable stay put: the filter applies only when arg1
is a string, so an event that ever fires with a number or no argument is
delivered as if plainly registered, and the handler's own test is what still
rejects it.
Left alone: api/unitframes.lua (the frame's unit changes at runtime and the
handler also matches on GUID), nameplates' six nameplateN events, libdebuff
and libpredict's UNIT_HEALTH (genuinely any unit), and raid.lua's UNIT_PET
(any raid member can own the pet).
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
Brues