Every region and frame carries a layout sub-object at +0x24 (4.3.4's
CLayoutFrame; OFF_REGION_ANCHOR, OFF_FRAME_LAYOUT_SUBOBJECT for a frame),
and three engine calls on it were wrapped by hand in sixteen files, each
with its own typedef and often its own AnchorOf / LayoutOf / FrameOf
helper:
- CLayoutFrame::GetRect (FUN_REGION_GET_RECT, 0x00768320), the
{bottom, left, top, right} rect, false before layout;
- IsResizePending (FUN_REGION_LAYOUT_DIRTY, 0x00768310);
- CLayoutFrame::Resize (FUN_LAYOUT_RESIZE, 0x007680E0), 1 to lay out now,
0 to mark for the layout pass.
frame/Layout.h now holds them, taking the region or frame itself, with
Of and OwnerOf for the sub-object and back, and GetResolvedRect for the
read the edge getters make (Script GetLeft 0x007A1A19: a pending resize
run now, then the rect). Frame::Subclass::FrameOfLayout is replaced by
OwnerOf.
Every RegisterFrameMethods call passed its method table and doc table
with a count beside each: sizeof(t) / sizeof(t[0]), std::size, or a
literal. The literals had to be kept in step with the tables by hand.
Game::ArrayRef<T> is a view over a file-scope array that takes the
count from the array's type, the way Doc::FieldList already did for a
function's arguments and returns. FieldList is now ArrayRef<Field>,
Doc::MethodList is ArrayRef<Method>, and RegisterFrameMethods takes
(context, methods, docs). The table and count still go to the engine's
frame-method iterator (FUN_REGISTER_FRAME_METHODS) unchanged, and to
the documentation recorder.
All 84 call sites pass only the arrays. The 16 literal counts were
checked against their tables first: each was the full length, so every
registry gets the same methods as before.
MinHook takes one hook per address. Texture::Mask hooked the
CSimpleTexture constructor (0x0076FC40) to drop its state when pool memory
is reused, and Texture::Transform hooked the corner store (0x007705B0) to
turn the quad again after each layout. Texture wrap modes and tiling need
both hooks too, so each hook now has an owner that fans out to listeners:
- Texture::Construct runs its listeners before the constructor, on the
memory it is about to build. Mask listens.
- Texture::CornerStore runs its listeners after the store. Transform
listens.
The store writes only the corners. The layer batch reads them through a
pointer on every draw, so nothing marks the layer changed.
The XML number parse through SStrToFloat (0x0064AAA0) moves into
Xml::FloatAttr, which NativeSetters and Transform's rotation attributes
now share.
MinHook allows one hook per address, so the texture, font string and
button loader hooks lived in Xml::ParentKey, which then read every other
module's attribute for it (mask, atlas, wordwrap, registerForClicks), and
CountInvisibleLetters hooked the EditBox loader on its own.
Xml::TypeLoad now owns the hooks on the engine's per-type node loaders,
the way Xml::FrameLoad owns FUN_XML_FRAME_LOAD's, and fans out to
subscribers after the engine has loaded each node:
- texture FUN_XML_TEXTURE_LOAD 0x0076FE20
- font string FUN_XML_FONTSTRING_LOAD 0x00770F40
- button FUN_XML_BUTTON_LOAD 0x007788C0 (CheckButton's loader
calls it first)
- edit box FUN_XML_EDITBOX_LOAD 0x00779FB0
- status bar FUN_XML_STATUSBAR_LOAD 0x00782EF0 (new; calls
FUN_XML_FRAME_LOAD first at 0x00782F06, ret 8)
All five share the sub-object convention __thiscall(loader, node,
status), ret 8, object = loader - OFF_XML_LOADER_SUBOBJECT. An inherited
template goes through the same loader first, so a node's subscribers run
after its templates'. The texture and font string loads keep running in
the node's mixin scope (Xml::Mixin::Begin / End), as before.
Each attribute is now read in the module that owns its setter:
Texture::Mask (mask), Texture::Atlas (atlas, useAtlasSize),
FontString::WordWrap (wordwrap), Frame::ClickMask (registerForClicks) and
EditBox::CountInvisibleLetters (countInvisibleLetters). The XML-only
entry points Atlas::SetFromXml, WordWrap::Set and ClickMask::SetFromXml
become private, and ClickMask.h and WordWrap.h go. ParentKey keeps only
its own binding work. Same hook addresses, same order around the engine
call, no behavior change.
SetAlphaGradient(index, gradient), ClearAlphaGradient and HasAlphaGradient,
what ScrollBox's SetEdgeFadeLength calls. As probed on retail: index 0 fades
in from the left and top edges, index 1 from the right and bottom, and the
fade shows only while the frame clips its children, its own regions
included.
Each index is one texture stage over the draw: a 2x2 ramp texture, filtered
and clamped, whose alpha multiplies in (preset 1). Coordinates come from the
camera-space position through a projected texture matrix that maps clip
space to the screen and then to the ramps, reloaded on every projection and
viewport set while a gradient draws. Textures, text (drawn in pixels under
its own projection) and clipped draws fade on the same screen spot.
ClipChildren now batches regions by nearest clipping frame and nearest
frame with a gradient; Texture::Mask takes its stages above the gradient's.
D3D only; a desaturated or masked texture does not fade.
DesaturateHierarchy(desaturation [, excludeRoot]) sets the desaturation of
every texture of the frame and of the frames under it, through
Texture::Desaturation.
A Cooldown was vanilla's indicator Model with CooldownFrameTemplate loaded
into it, so SetReverse and SetDrawEdge stored flags that changed nothing,
and every cooldown paid a Model's per-frame cost.
Cooldown is now a CSimpleFrame subclass built the way CGCooldown::Create
builds it: one SMemAlloc block, the engine's frame constructor, and copies
of CSimpleFrame's two vtables with 4.3.4's overrides in their 1.12 slots
(update, layer emit, alpha, size change, LoadXML, identity, destructor).
It runs 4.3.4's sweep / flash / reverse state machine and queues the star
fan, flash and edge as layer-batch strips animated in place.
- The five methods live in a class method hash of their own, so only a
Cooldown has them; it is released after CGGameUI's unregister pass.
- GetObjectType / GetFrameType answer through the engine's virtuals; the
Lua re-registrations are gone.
- star4.blp ships with 1.12; edge.blp does not, so its 16x16 image is
created in memory through TextureCreateWithCallback.
- A Cooldown inheriting the stock Model CooldownFrameTemplate applies
4.3.4's template (setAllPoints, hidden); reverse / drawEdge XML
attributes are read.
- CooldownFrame_SetTimer calls SetCooldown on a Cooldown and keeps 1.12's
body for the Models stock buttons use.
- Mask.cpp's path loader moves to texture/Load; frame/RenderBatch queues
quads into a layer batch.
Verified in-game: sweep, flash and hide; reverse with the edge line;
stock action-button cooldowns unchanged; repeated /reload.
- <Texture mask="path"> applies texture:SetMask(path).
- <MaskTexture> in a <Layer> builds a MaskTexture, and its
<MaskedTextures> entries (childKey or target) attach it to textures.
xml/Layers.cpp replaces the engine's layer loader with the same loop
plus a MaskTexture branch, so the engine no longer warns on the tag.
- SetMask closes the mask texture's handle on replace, clear, region
reuse and reload, as SetTexture does with the texture it held.
- SetMask keeps its previous mask when the file does not load, using
SetTexture's own status test, instead of masking with the opaque
fallback texture.
- API.md documents the XML forms and that a mask with an opaque edge
does not clip past its rectangle (the device has no border address
mode).
0x007680E0 is 4.3.4's CLayoutFrame::Resize(force): force 0 marks the
layout dirty, force 1 resolves it now, as Script GetLeft calls it. It was
named FUN_FONTSTRING_LAYOUT_INVALIDATE, which hid the second use; rename
it FUN_LAYOUT_RESIZE.
Move the walk of a frame's intrusive region and child lists out of
Frame::Alpha into Game::ForEachFrameListObject, so the bounds walk can
use the same code.
Every engine frame method validates its `self` in three steps before
touching it: lua_type == table, resolve to the CFrameScriptObject, then
the IsA(typeId) vmethod at vtable+0x10. We did the first two and skipped
the third, so ResolveObject handed back a pointer for ANY frame and the
method cast it to whatever type it expected.
That was reachable from plain Lua, not just through a mirror or a hook.
The per-type method dispatcher is an __index handler that rawgeti's the
method closure out of the Lua registry, so frame methods are ordinary
first-class values, and
GameTooltip.SetSpellByID(ShapeshiftButton1, 133)
reached the tooltip builder with a Button, writing tooltip fields through
Button memory. The engine refuses that same call with "Wrong object type
for member function".
Game::Lua::ResolveTypedObject runs the engine's prologue and raises the
engine's own three error strings, read from .data rather than reinvented.
FrameScriptTypeId mirrors the lazy id assignment, so a check is valid
before any stock method of the type has run. Seven per-type wrappers stop
a call site pairing a method with the wrong class's id, and
ResolveTypedObjectAny covers a method that more than one type reaches.
45 call sites across 19 files now gate self. model/DisplayInfo.cpp had
already worked this out for Model alone and loses its private copy.
Type ids are per CLASS, one shared by all of that class's methods. Each
was read from the prologue of a method that can only belong to its class:
a base-class method registered into several tables carries the BASE id, so
sampling GetObjectType or GetDrawLayer would have produced a gate that
accepts unrelated types. Frame comes from Script_RegisterEvent and Texture
from Script_SetDesaturated, both cross-checked against the first entry of
their table.
Deliberately left ungated: the binding `owner` in bindings/Api.cpp and the
three co-hook sites in texture/Atlas.cpp. Each only uses the pointer as an
identity key and never casts it, so a gate would tighten the contract for
no safety gain -- and raising inside a co-hook would fire before the
original ran.
Verified in-game: the Button-as-tooltip call now raises the engine's
error; a Button still reaches a Frame-registry method
(button:SetAttribute round-trips), so IsA does accept derived types; and
fontstring:GetRotation() still answers -- that reader is registered on
both the Texture and FontString registries and needs the id pair rather
than the Texture id alone.
The engine's mip-chain byte total FUN_GX_TEXTURE_MIP_BYTES (0x00594CE0)
walks a texture's levels by halving both dimensions until each reaches
exactly 1, and gates every halving on JBE. A dimension of 0 therefore
never shifts and never reaches 1, so the loop at 0x00594D47 spins
forever. It does no allocation and no I/O, which is why it presents as a
hard freeze at 100% of one core rather than a crash, and why the client
never recovers.
Caught live at 0x00594D78 by a SIGSEGV-forced backtrace taken during
login under Wine. The frame's four printed stack slots are this
function's own reused parameter slots, and read back as 4 bytes per block
(uncompressed 32-bit) with both dimensions 0. The call chain checks out
edge for edge from the region-layer draw down through the state flush and
the texture-stage apply, so it is a real stack rather than a heuristic
scan. Note that the state flush applies a dirty list, so the UI draw at
the bottom is where the flush happened, not necessarily what bound the
texture.
One branch above the loop the engine already handles the analogous case
correctly: a texture with no mip chain returns its base-level bytes and
stops. The guard gives a degenerate dimension the same answer. It
resolves the caller's arguments the way the original does and, when
either dimension would shift to zero, raises that one to the smallest
value the accounting can express before calling through. The loop then
terminates on its first pass. No mip math, format table or reduction
level is reimplemented. Clamping at the entry rather than at a call site
also covers the second caller, FUN_00594C80, which accounts for a lock by
passing a dirty rect's width and height and hangs identically on an empty
but still flagged rect.
A dimension can also reach zero without the texture being 0x0, via the
reduction level the accounting shifts by -- a caps field (caps + 0xE0,
the block being device + 0x23C per FUN_005923D0) filled from the graphics
backend. Measured 0 on a native D3D9 driver, unmeasured under wined3d.
So the guard tests the shifted value and clamps to 1 << reduce rather
than assuming either case.
Who feeds it a zero is a separate question, and the guard does not bury
it. ClassicAPI constructs no textures, and DimensionGate cannot admit a
zero it would otherwise reject, since the validator's power-of-two test
already yields "is a power of two" for 0. So the first occurrence is
logged with everything readable off the texture, plus
Texture::Mask::AnyMaskActive() to rule the mask path in or out -- passive
evidence that costs whoever hits the freeze nothing.
SetMask rode the minimap BLIP recipe (explicit second UV stream, uv1 = uv0),
so the mask sampled at the base's texcoords - a SetTexCoord'd sprite made
unit 1 sample that same sub-rect of the MASK image, showing a corner chunk
of the shape instead of the shape.
Retire that path: the path mask is now one extra unit in the texgen draw
(MaskedRegionDrawImpl), its matrix built from the BASE region's own rect
(SetTexCoordModifiesRect crop gate included) plus the base's SetRotation so
a mask keeps following a rotated texture. The angle-0 matrix is exactly the
old recipe's formula re-expressed through positions (see BuildMaskMatrix),
so default 0..1-texcoord textures render identically; sprites become the
fixed case, matching retail (SetMask ignores SetTexCoord).
Side effect: SetMask and AddMaskTexture now compose (alpha intersection)
instead of object masks winning; the path mask occupies one of the 7 units.
Modules that keep reload-fragile file-static state (Lua registry refs,
frame/object-pointer-keyed maps, cached event-slot indices) each expose a
PrepareForReload() that DllMain called from the FrameScript_Initialize hook.
That list was hand-maintained, so a new module could silently forget to wire
itself in -- the failure mode behind the Frame::Attributes stale-ref bug and
Tooltip::SetEvents issue #33.
Add Game::ReloadAutoRegister + Game::RunReloadCleanups(), mirroring the
existing ModuleAutoRegister / HookAutoRegister / WorldTick::AutoSubscribe
patterns. Each module self-registers with one file-scope line after its
PrepareForReload:
static const Game::ReloadAutoRegister _reloadReg{&PrepareForReload};
DllMain's 11-call manual list collapses to a single RunReloadCleanups(); a
new module can no longer be forgotten. Player::NameCache::Flush stays an
explicit call -- it's a save-before-teardown, not a stale-state clear, and
also runs on the glue-return path.
A sweep of all file-static state found one uncovered module: Time::Timer
pins C_Timer.After/NewTimer/NewTicker callbacks in the Lua registry but never
released them on /reload -- a surviving/stale ref could fire a dead closure.
Add Time::Timer::PrepareForReload (release each pending timer's registry
callback, clear the queue) and self-register it.
- Game.{h,cpp}: ReloadAutoRegister + RunReloadCleanups.
- DllMain.cpp: collapse the manual list to RunReloadCleanups(); drop 6
now-dead includes.
- 11 existing reload modules + Time::Timer: add the self-register line.
Behavior-preserving (same cleanups, before the Lua teardown); auto-register
order is unspecified across TUs but every callback only clears its own state.
Backport frame:CreateMaskTexture plus texture:AddMaskTexture /
RemoveMaskTexture / GetNumMaskTextures / GetMaskTexture -- positioned,
rect-based masks, the object form later clients use (SetMask(path) stays
the simple single-shape form).
Positioned masks sample via texgen (object-space vertex position through
a per-unit texture matrix), the mechanism of the engine's own minimap
disc mask, so no vertex UV stream is consumed per mask. That lifts both
v1 limits at once:
- Up to 7 masks apply at once (device stage cap); a texture shows only
the intersection of all of them.
- mask:SetRotation rotates the clip shape (folded into the matrix), and a
mask can move/resize each frame, so the clip animates.
State is restored in a co-hook on the indexed submit (FUN_GX_PRIM_INDEXED)
the engine issues after the stream call, not inside the stream hook --
restoring earlier reverts the state the queued draw is flushed with.
Two incidental fixes: a SetTexCoord-cropped base now masks against the
same rect the engine's corner store uses, and a recycled mask region is
detached from every base still holding it.
Offsets.h: add the texgen / coord-transform / texture-matrix / state-journal
/ indexed-submit constants (decoded from the D3D9 backend state flush
FUN_005a2f00), and fix the FUN_GX_TEXUNIT_ENABLE/_DISABLE mislabel (they
are texture-matrix push/pop, not stage enable/disable).
Backports the modern Texture:SetMask, clipping a texture to a mask's alpha
shape - round portraits, round minimap buttons, and other shaped art. Vanilla
1.12 has no general masking (only the special-cased minimap disc).
Mirrors the engine's own masked-quad draw (the minimap blip mask FUN_004eae10):
a co-hook on the vertex-stream primitive FUN_0058a2a0 recovers the owning
CSimpleTexture from the batch positions pointer (region = positions -
OFF_SIMPLETEXTURE_CORNERS), and for a region with a mask set binds the mask on
texture unit 1 with a MODULATE combiner, then re-runs the engine's own submit
with uv1 = uv0 so unit 1 samples the mask at the texture's own coordinates. The
unit is enabled/disabled around the draw so the mask never leaks onto other
textures. A cold co-hook on the region ctor clears a region's mask on pooled
reuse; a kill switch and SEH latch guard the draw path.
The recipe was pinned by an in-game probe on the live GL backend and matched
against FUN_004eae10; the full RE is in the "Texture masking" block in Offsets.h
and the Mask.cpp header. Masks must be BLP or uncompressed 32-bit TGA with the
shape in the alpha channel; two round masks ship in the client, other shapes are
addon-supplied content.