Split screen & Overhaul Wide-Screen support #767
Merged
codeHusky merged 32 commits intosmartcmd:mainfrom Mar 8, 2026
Merged
Conversation
The native keyboard scene maintained a separate C++ buffer (m_win64TextBuffer) for physical keyboard input, which was pushed to the Flash text field via setLabel(). However, when the user typed with the on-screen controller buttons, Flash updated its text field directly through ActionScript without updating the C++ buffer. This caused a desync: switching back to the physical keyboard would overwrite any text entered via controller, since m_win64TextBuffer still held the old value before the controller edits. Fix: read the current Flash text field into m_win64TextBuffer at the start of each tick(), before consuming new physical keyboard chars. This ensures both input methods always operate on the same state.
…ction state The keyboard UI mode (on-screen virtual keyboard vs direct text input) was determined by Win64_IsControllerConnected(), which checks if any XInput controller is physically plugged in. This meant that even if the player was actively using mouse and keyboard, the virtual keyboard would still appear as long as a controller was connected. Replace the connection check with g_KBMInput.IsKBMActive(), which tracks the actual last-used input device based on per-frame input detection. Now the keyboard mode is determined by what the player is currently using, not what hardware happens to be plugged in. Affected scenes: CreateWorldMenu (world naming) and LoadOrJoinMenu (world renaming).
…rect edit The direct text editing mode introduced for KBM users had several issues with the TextInput control's caret (blinking cursor) and text manipulation: 1. Caret visible when not editing: When navigating to the world name field with keyboard/mouse, Flash's Iggy focus system would show the blinking caret even though the field wasn't active for editing yet (Enter not pressed). This was misleading since typing had no effect in that state. Fix: access the FJ_TextInput's internal m_mcCaret MovieClip and force its visibility based on editing state. This is enforced every tick because setLabel() and Flash focus transitions continuously reset the caret state. 2. No cursor movement during editing: The direct edit implementation treated the text as a simple buffer with push_back/pop_back — there was no concept of cursor position. Backspace only deleted from the end, and arrow keys did nothing. Fix: track cursor position (m_iCursorPos) in C++ and use wstring insert/erase at that position. Arrow keys (Left/Right), Home, End, and Delete now work as expected. The visual caret position is synced to Flash via the FJ_TextInput's SetCaretIndex method. 3. setLabel() resetting caret position: Every call to setLabel() (when text changes) caused Flash to reset the caret to the end of the string, making the cursor jump visually even though the C++ position was correct. Fix: enforce caret position via setCaretIndex every tick during editing, so any Flash-side resets are immediately corrected. New UIControl_TextInput API: - setCaretVisible(bool): toggles m_mcCaret.visible in Flash - setCaretIndex(int): calls FJ_TextInput.SetCaretIndex in Flash
On Windows64 with KBM, moving the mouse over empty space (outside any button) would clear the Iggy focus entirely. After that, pressing arrow keys did nothing because Flash had no starting element to navigate from. Two changes here: - Don't set focus to IGGY_FOCUS_NULL when the mouse hovers over empty space. The previous hover target stays focused, so switching back to arrows keeps working seamlessly. - When a navigation key is pressed and nothing is focused at all (e.g. mouse was already on empty space when the menu opened), grab the first focusable element instead of silently dropping the input. The keypress is consumed to avoid jumping two elements at once. This makes mixed mouse+keyboard navigation feel a lot more natural. You can point at a button, then continue with arrows, or just start pressing arrows right away without having to hover first.
…cenes This is a large rework of the Windows64 KBM (keyboard+mouse) input layer. It touches the mouse hover system, the mouse click dispatch, and the direct text editing infrastructure, then applies all of it to every scene that has text input fields or non-standard clickable elements. MOUSE HOVER REWRITE (UIController.cpp tickInput) The old hover code had two structural problems: (a) Scene lookup was group-first: it iterated UI groups and checked all layers within each group. The Tooltips layer on eUIGroup_Fullscreen (which holds non-interactive overlays like button hints) would be found before in-game menus on eUIGroup_Player1. The tooltip scene focusable objects captured mouse input and prevented hover from reaching the actual menu. Fixed by switching to layer-first lookup across all groups, and skipping eUILayer_Tooltips entirely since those are never interactive. (b) On tabbed menus (LaunchMoreOptionsMenu Game vs World tabs), all controls from all tabs are registered in Flash at the same time. There was no filtering, so controls from inactive tabs had phantom hitboxes that overlapped the active tab controls, making certain buttons unhoverable. Fixed by introducing parent panel tracking: each UIControl now has a m_pParentPanel pointer, set automatically by the UI_MAP_ELEMENT macro during mapElementsAndNames(). The hover code checks the control parent panel against the scene GetMainPanel() and skips mismatches. This is the same technique the Vita touch code used, but applied to mouse hover. The coordinate conversion was also simplified. The old code had two separate scaling paths (window dimensions for hover, display dimensions for sliders). Now there is one conversion from window pixel coords to SWF coords using the scene own render dimensions. REUSING VITA TOUCH APIs FOR MOUSE (ButtonList, UIScene) Several APIs originally gated behind __PSVITA__ are now enabled for Win64: - UIControl_ButtonList::SetTouchFocus(x,y) and CanTouchTrigger(x,y): the Flash-side ActionScript methods were already registered on all platforms in setupControl(), only the C++ wrappers were ifdef-gated. Opening the ifdefs to include _WINDOWS64 lets the mouse hover code delegate to Flash for list item highlighting, which handles internal scrolling and item layout that would be impractical to replicate in C++. - UIScene::SetFocusToElement(id): programmatic focus-by-control-ID, used as a fallback when Iggy focusable objects do not match the C++ hit test. - UIScene_LaunchMoreOptionsMenu::GetMainPanel(): returns the active tab panel control, needed by the hover code to filter inactive tab controls. MOUSE CLICK DISPATCH (UIScene.cpp handleMouseClick) Left-clicking previously relied entirely on Iggy ACTION_MENU_OK dispatch, which routes to whatever Flash considers focused. This broke for custom- drawn elements that are not Flash buttons (crafting recipe slots), and for scenes where Iggy focus did not match what the user visually clicked. Added a virtual handleMouseClick(x, y) on UIScene with a default implementation that hit-tests C++ controls. When multiple controls report overlapping bounds (common in debug scenes where TextInputs report full Flash-width), it picks the one whose left edge X is closest to the click. Returns true to consume the click and suppress the normal ACTION_MENU_A dispatch via a m_mouseClickConsumedByScene flag on UIController. The default implementation handles buttons, text inputs, and checkboxes (toggling state and calling handleCheckboxToggled directly). CRAFTING MENU MOUSE CLICK (UIScene_CraftingMenu.cpp) The crafting menu recipe slots (H slots) are rendered through Iggy custom draw callback, not as Flash buttons. They have no focusable objects, so mouse clicking did nothing. The solution caches SWF-space positions during rendering: inside customDraw, when H slot 0 and H slot 1 are drawn, the code extracts SWF coordinates from the D3D11 transform matrix via gdraw_D3D11_CalculateCustomDraw_4J. The X difference between slot 0 and slot 1 gives the uniform slot spacing. handleMouseClick then uses these cached bounds to determine which recipe slot was clicked, resets the vertical slot indices (same pattern as the constructor), updates the highlight and vertical slots display, and re-shows the old slot icon. This mirrors the existing controller LEFT/RIGHT navigation in the base class handleKeyDown. DIRECT EDIT REFACTORING (UIControl_TextInput) The direct text editing feature (type directly into text fields instead of opening the virtual keyboard) was originally implemented inline in CreateWorldMenu with all the state, character consumption, cursor tracking, caret visibility, and cooldown logic hardcoded in one scene. Moved everything into UIControl_TextInput: - beginDirectEdit(charLimit): captures current label, inits cursor at end - tickDirectEdit(): consumes chars, handles Backspace/Enter/Escape, arrow keys (Left/Right/Home/End/Delete), enforces caret visibility every tick (because setLabel and Flash focus transitions continuously reset it), returns Confirmed/Cancelled/Continue - cancelDirectEdit() / confirmDirectEdit(): programmatic control - isDirectEditing() / getDirectEditCooldown() / getEditBuffer(): state query For SWFs that lack the m_mcCaret MovieClip child (like AnvilMenu), the existence check validates by reading a property from the resolved path, since IggyValuePathMakeNameRef always succeeds even for undefined refs. When no caret exists, the control inserts a _ character at the cursor position as a visual fallback. The caret check result is cached in m_bHasCaret/m_bCaretChecked to avoid repeated Iggy calls that could corrupt internal state. SCENES UPDATED WITH DIRECT EDIT + VIRTUAL KEYBOARD Every scene with text input now supports both input modes: direct editing when KBM is active, virtual keyboard (via NavigateToScene eUIScene_Keyboard) when using a controller. The mode is chosen at press time based on g_KBMInput.IsKBMActive(). - CreateWorldMenu: refactored to use the new UIControl_TextInput API, removing ~80 lines of inline editing code. - AnvilMenu: item renaming now supports direct edit. The keyboard callback uses Win64_GetKeyboardText instead of InputManager.GetText (which reads from a different buffer on Win64). The virtual keyboard is opened with eUILayer_Fullscreen + eUIGroup_Fullscreen so it does not hide the anvil container menu underneath. Added null guards on getMovie() in setCostLabel and showCross since the AnvilMenu SWF may not fully load on Win64. - SignEntryMenu: all 4 sign lines support direct edit. Clicking a different line while editing confirms the current one. Each line cooldown timer is checked independently to prevent Enter from re-opening the edit. - LaunchMoreOptionsMenu: seed field direct edit with proper input blocking. - DebugCreateSchematic: all 7 text inputs (name + start/end XYZ coords). handleMouseClick is overridden to always consume clicks during edit to prevent Iggy re-entry on empty space. - DebugSetCamera: all 5 inputs (camera XYZ + Y rotation + elevation). Clicking a different field while editing confirms the current value and opens the new one. Float display formatting changed from %f to %.2f. All keyboard completion callbacks on Win64 now use Win64_GetKeyboardText (two params: buffer + size) instead of InputManager.GetText, which reads from the correct g_Win64KeyboardResult global when using the in-game keyboard scene. SCROLL WHEEL Mouse wheel events (ACTION_MENU_OTHER_STICK_UP/DOWN) are now centrally remapped to ACTION_MENU_UP/DOWN in UIController::handleKeyPress when KBM is active. Previously each scene would need to handle OTHER_STICK actions separately, and most did not, so scroll wheel only worked in a few places.
…n, craft) The crafting screen's horizontal recipe slots and category tabs are custom-drawn via Iggy callbacks rather than regular Flash buttons, so the standard mouse hover system can't interact with them. This adds handleMouseClick to derive clickable regions from the H slot positions cached during customDraw. Tab clicking: tab hitboxes are computed relative to the H slot row since the Vita TouchPanel overlays (full-screen invisible rectangles) aren't suitable for direct hit-testing on Win64. The Y bounds were tuned empirically to match the SWF tab icon positions. Clicking a tab runs the same switch logic as LB/RB: hide old highlight, update group index, reset slot indices, recalculate recipes, and refresh the display. H slot clicking: clicking a different recipe slot selects it (updating V slots, highlight, and re-showing the previous slot). Clicking the already-selected slot crafts the item by dispatching ACTION_MENU_A through handleKeyDown, reusing the existing crafting path. Empty slots (iCount == 0) are ignored. All mouse clicks on the scene are consumed (return true) to prevent misses from falling through as ACTION_MENU_A and accidentally triggering a craft. This only suppresses mouse-originated A presses via m_mouseClickConsumedByScene; keyboard and controller A remain fully functional. Also enables GetMainPanel for Win64 (was Vita-only) so the mouse hover system can filter controls by active panel, same as other tabbed menus.
The hover code was doing a redundant second hit-test against Iggy focusable object bounds after the C++ control bounds had already identified the correct control. Iggy focusable bounds are wider than the actual visible buttons and overlap vertically, so the "pick largest x0" heuristic would match focusables belonging to earlier buttons when hovering the right side of buttons 3+. Replaced the IggyPlayerGetFocusableObjects path with a direct SetFocusToElement call using the already-correct hitControlId from the C++ hit-test, same approach the click path uses in handleMouseClick. Also switched the overlap tiebreaker from "largest x0" to smallest area, consistent with how clicks resolve overlapping controls. TextInput is excluded from hover focus to avoid showing the caret on mere mouse-over (its Iggy focus is set on click).
Same overlap fix applied to handleMouseClick: when multiple controls contain the click point, prefer the one with the smallest bounding area instead of the one with the largest left-edge X. This is more robust for any layout (vertical menus, grids, overlapping panels) and matches the hover path logic. Those changes were initially made in order to fix the teleport ui for the mouse but broke every other well working ui.
When the inventory or other UI with a hidden cursor was open, alt-tabbing out would leave the cursor locked to the game window. SetWindowFocused(false) from WM_KILLFOCUS correctly released the clip and showed the cursor, but Tick() was unconditionally calling SetCursorPos every frame to re-center it, overriding the release. Added m_windowFocused to the Tick() condition so cursor manipulation only happens while the window actually has focus.
Right clicking an item stack in Java Edition picks up half of it. Console Edition already handles this via ACTION_MENU_X (the X button on controller), which sets buttonNum=1 in handleKeyDown. This maps mouse right click to that same action so KBM players get the same behavior across all container menus (inventory, chests, furnaces, hoppers, etc).
When removeControl() removes a Flash element (e.g. the Reinstall button in Help & Options, or the Debug button when disabled), the C++ control object stays in the m_controls vector. On Vita this was handled by calling setHidden(true) and checking getHidden() in the touch hit-test, but on Windows64 none of that was happening. The result: removed buttons kept phantom bounds that the hover code would match against, stealing focus from the buttons that shifted into their visual position. In the Help & Options menu with debug enabled, the removed Reinstall button (Button6) had ghost bounds overlapping where the Debug button (Button7) moved to after the removal, making Debug un-hoverable and snapping focus to Button1. The fix has three parts: - removeControl() now calls setHidden(true) on all platforms, not just Vita. The m_bHidden member was already declared on all platforms, only the accessors were ifdef'd behind __PSVITA__. - Removed the __PSVITA__ ifdef from setHidden/getHidden in UIControl.h so they're available everywhere. - Added getHidden() checks in both the hover and click hit-test loops, matching what the Vita touch code already does. The check is a simple bool read (no Flash/Iggy call), placed before the getVisible() query which hits Flash and can return stale values for removed elements.
On controller, RB (ACTION_MENU_RIGHT_SCROLL) opens the save options dialog (rename/delete) when a save is selected. Mouse right-click maps to ACTION_MENU_X, which had no Windows64 handler in this scene. Added save options handling under ACTION_MENU_X for _WINDOWS64 so right-clicking a save opens the same dialog. Also handles the mashup world hide action for right-click consistency. Console-only options (copy save, save transfer) are excluded since they don't apply here.
Mouse hover and click in split-screen was broken: the coordinate conversion from window pixels to Flash/SWF space did not account for the viewport tile-origin offset or the smaller display dimensions of each splitscreen quadrant. Now the mouse position is mapped through three steps: window pixels to UIController screen space, subtract the viewport origin (which varies per quadrant/split type), then scale from display size to SWF authoring size. This fixes hover highlighting and click targeting in all splitscreen layouts. Mouse input was also bleeding into other splitscreen players' UI groups because the scene lookup iterated all groups. Now it only checks the fullscreen group and the primary (KBM) player's group, so controller players' menus are never affected by mouse movement. Mouse grab/release (cursor lock for gameplay) was triggering for every local player's tick, causing fights between splitscreen players over the cursor state. Now only the primary pad player controls grab state. The in-game keyboard scene in PC mode had no cursor movement: typing always appended at the end and backspace always deleted from the end. Added a cursor position tracker (m_iCursorPos) so that characters are inserted at the cursor, backspace deletes behind it, and arrow keys, Home, End, and Delete all work as expected. The Flash caret is synced to the cursor position each tick. Also stopped syncing the text buffer back from Flash in PC mode, which was resetting the cursor every tick. Arrow keys in PC mode no longer get forwarded to Flash (which would move the on-screen keyboard selector instead of the text cursor). AddLocalPlayerByUserIndex was calling NotifyPlayerJoined before the IQNet slot was actually registered, passing a pointer obtained via GetLocalPlayerByUserIndex which checks customData (not set yet at that point). Now AddLocalPlayerByUserIndex is called first, and if it succeeds, the notification uses the static m_player array directly. The stub AddLocalPlayerByUserIndex now properly initialises the slot with gamertag and remote/host flags instead of being a no-op. IsSignedIn was hardcoded to return true only for pad 0, preventing splitscreen players from joining. Now it checks IsPadConnected so any connected controller can sign in. GetXUID returned INVALID_XUID for all pads except 0, which broke splitscreen player identity. Now each pad gets a unique XUID derived from the base value plus the pad index. Pinned internal resolution to 1920x1080 and removed GetSystemMetrics auto-detection which was picking up the native monitor resolution and breaking the 16:9 assumption in the viewport math and Flash layout. DPI awareness is kept for consistent pixel coordinates.
The KBM pause check had a IsTutorialVisible guard that blocked Escape entirely while any tutorial popup was on screen. The controller path never had this restriction. Removed the check so Escape behaves the same as Start on controller.
When a player enters a new region, RegionFile's constructor calls createFile which adds a FileEntry with length 0 to the file table. This increases the header table size (appended at the end of the save buffer) by sizeof(FileEntrySaveData) per entry, but since no actual data is written to the file, MoveDataBeyond is never called and the committed virtual memory pages are never grown to match. On the next autosave tick, saveLevelData writes level.dat first (before chunkSource->save which would have grown the buffer). If level.dat doesn't need to grow, finalizeWrite calls WriteHeader which tries to memcpy the now-larger header table past the end of committed memory, causing an access violation. This is especially likely in splitscreen where two players exploring at the same time can create multiple new RegionFile entries within a single tick, quickly exhausting the page-alignment slack in the buffer (yes i am working at splitscreen in the meanwhile :) ) The fix was deduced by tracing the crash callstack through the save system: FileHeader, ConsoleSaveFileOriginal, the stream chain, and the RegionFile/RegionFileCache layer. The root cause turned out to be a gap between createFile (which grows the header table) and MoveDataBeyond (the only place that grows the buffer), with finalizeWrite sitting right in between unprotected. The buffer growth check added here mirrors the exact same VirtualAlloc pattern already used in MoveDataBeyond (line 484-497) and in the constructor's decompression path (line 176-190), so it integrates naturally with the existing code. Same types, same page rounding, same error handling. The fast path (no new entries, buffer already big enough) is a single DWORD comparison that doesn't get taken, so there is zero overhead in the common case. This is the right place for the fix because finalizeWrite is the sole caller of WriteHeader, meaning every code path that writes the header (closeHandle, PrepareForWrite, deleteFile, Flush) is now protected by a single check point.
…e class The fake cursor character (_) used for SWFs without m_mcCaret was leaking into saved sign and anvil text. This happened because setLabel() with instant=false only updates the C++-side cache, deferring the Flash write to the next control tick. Any getLabel() call before that tick reads the old Flash value still containing the underscore. Fixed by passing instant=true in confirmDirectEdit, cancelDirectEdit, and the Enter key path inside tickDirectEdit, so the cleaned text hits Flash immediately. Mouse hover over TextInput controls (world name, anvil name, seed field) was not showing the yellow highlight border. The hover code used IggyPlayerSetFocusRS which sets Iggy's internal dispatch focus but does not trigger Flash's ChangeState callback, so no visual feedback appeared. Buttons worked fine because Iggy draws its own focus ring on them, but TextInput relies entirely on ChangeState(0) for the yellow border. Switched to SetFocusToElement which goes through the Flash-side SetFocus path, then immediately call setCaretVisible(false) to suppress the blinking caret that comes with focus. No visual flicker since rendering happens after both tickInput and scene tick complete. While direct editing, mouse hover was able to move focus away to other TextInputs on the same scene (most noticeably on the sign editor, where hovering a different line would steal focus from the line being typed). Added an isDirectEditBlocking() check in the hover path to skip focus changes when any input on the scene is actively being edited. The Done button in SignEntryMenu was unresponsive to mouse clicks during direct editing. The root cause is execution order: handleMouseClick runs before handleInput in the frame. The base handleMouseClick found the Done button and called handlePress, but handlePress bailed out because of the isDirectEditing guard. The click was marked consumed, so handleInput never saw it. Fixed by overriding handleMouseClick in SignEntryMenu to detect the Done button hit while editing and confirm + close directly. Added click-outside-to-deselect for anvil and world name text inputs. Both scenes previously required Enter to confirm the edit, which felt wrong. Now clicking anywhere outside the text field bounds confirms the current text, matching standard UI behavior. The anvil menu now updates the item name in real time while typing, like Java edition. Previously the name was only applied on Enter, so the repair cost display was stale until confirmation. The biggest change is structural: every scene that used direct editing (AnvilMenu, CreateWorldMenu, SignEntryMenu, LaunchMoreOptionsMenu, DebugCreateSchematic, DebugSetCamera) had its own copy of the same boilerplate -- tickDirectEdit loops in tick(), click-outside hit testing in handleMouseClick(), cooldown guard checks in handleInput/handlePress, and result dispatch with switch/if chains. This was around 200 lines of near-identical code scattered across 6 files, each with its own slight variations and its own bugs waiting to happen. Pulled all of it into UIScene with two virtual methods: getDirectEditInputs() where scenes register their text inputs, and onDirectEditFinished() where they handle confirmed/cancelled results. The base class tick() drives tickDirectEdit on all registered inputs, handleMouseClick() does the click-outside-to-deselect hit test generically using panel offsets, and isDirectEditBlocking() replaces all the inline cooldown checks. Scenes now just override those two methods and get everything for free. Also removed the m_activeDirectEditControl enum tracking from the debug scenes (DebugCreateSchematic, DebugSetCamera) since the base class handles lifecycle tracking through the controls themselves.
The scroll wheel was always remapped to UP/DOWN, which is fine for vertical lists but useless on horizontal controls like sliders and the texture pack selector. Track whether the mouse is hovering a horizontal control during the hover hit-test (new bool m_bMouseHoverHorizontalList, set for eTexturePackList and eSlider). When the flag is set, handleKeyPress emits LEFT/RIGHT instead of UP/DOWN for wheel events. TexturePackList is also now part of the mouse hover system with proper hit-testing, relative-coord SetTouchFocus and GetRealHeight for accurate bounds.
tickDirectEdit calls into Iggy every tick without checking if the movie is still valid, which crashes inside iggy_w64.dll when the Flash movie gets unloaded or isn't ready yet.
The mouse scroll wheel was not working in the creative inventory at all. UIController remaps wheel input from OTHER_STICK to UP/DOWN for KBM users, but the base container menu handler consumed UP/DOWN for grid navigation before it could reach the creative menu's page scrolling logic in handleAdditionalKeyPress. Fixed by detecting scroll wheel input on UP/DOWN in the base handler and forwarding it as OTHER_STICK to handleAdditionalKeyPress instead. Also fixed the controller right stick scrolling way too fast: it was jumping TabSpec::rows (5) rows per tick at 100ms repeat rate, which blew through the entire item list almost instantly. Reduced to 1 row per tick so scrolling feels controlled on both input methods.
gluPerspective was hardcoded to use g_iAspectRatio (always 16:9) instead of the aspect parameter from getFovAndAspect, which adjusts for split-screen viewports. The 3D world was horizontally stretched in top/bottom split because the projection used 16:9 while the viewport was 32:9.
Collaborator
|
Please keep your PRs in draft state if they're not ready to merge |
|
Reading how XUIDs for additional players were handled, seems like this would just entirely work with LAN/Online as well. I'll test that myself later tonight. |
…port Screen resolution is now auto-detected from the monitor at startup instead of being hardcoded to 1920x1080. This fixes rendering on ultrawide (21:9), super-ultrawide (32:9), 16:10, and any other aspect ratio -- both in singleplayer and split-screen multiplayer. The 3D world renders at native resolution so the full monitor is used. Flash UI is 16:9-fitted and centered inside each viewport, pillarboxed on wide displays and letterboxed on tall ones. Logical game dimensions (used for ortho projection and HUD layout) are computed proportionally from the real screen aspect ratio, fixing the stretched world projection and HUD that the old hardcoded 1280x720 caused on non-16:9 monitors. GameRenderer::ComputeViewportForPlayer uses the actual backbuffer size instead of the logical game size, which was causing split-screen viewports to be sized incorrectly. UIScene::render fits menus to 16:9 within each split viewport using GetViewportRect + Fit16x9, keeping inventory/crafting/options screens at their designed aspect ratio instead of stretching. Panorama and MenuBackground render at full viewport size with proper tile scaling so the background fills the entire area without gaps in vertical split and quadrant layouts. HUD tile rendering uses ComputeTileScale to uniformly scale the SWF and show the bottom portion (hotbar, hearts, hunger) in horizontal and quadrant splits. repositionHud passes visible SWF-space dimensions to ActionScript for proper element centering within each viewport. Chat and Tooltips overlays use ComputeTileScale and ComputeSplitContentOffset to anchor correctly to the bottom of each player's viewport tile. Container menus apply Fit16x9 to pointer coordinate mapping so the cursor tracks correctly in split-screen. getMouseToSWFScale moved out of the header into the .cpp. Mouse input in onMouseTick is gated to pad 0 since raw mouse deltas should only drive player 1. All shared viewport math lives in UISplitScreenHelpers.h: - GetViewportRect: origin and dimensions for any viewport type - Fit16x9: aspect-correct fitting with centering offsets - ComputeTileScale: uniform scale and Y-offset for tile rendering - ComputeSplitContentOffset: content centering for overlay components
# Conflicts: # Minecraft.Client/Common/UI/UIComponent_Panorama.cpp # Minecraft.Client/Common/UI/UIController.cpp # Minecraft.Client/Common/UI/UIScene_Keyboard.cpp # Minecraft.Client/Extrax64Stubs.cpp
Contributor
Author
|
good news,i managed to get the ui resizing right! @fayaz12g |
Contributor
Author
|
@rtm516 nice catch. @codeHusky good call. I went with your second suggestion deriving split-screen XUIDs by hashing baseXuid + iPad through the existing Mix64 function. Pad 0 still returns the base XUID unchanged (save compat), pads 1-3 get fully independent 64-bit values. More than an actual overlap risk, it's just cleaner this way, the base XUID is already generated to not collide, so + iPad would've probably been fine in practice, but hashing makes the intent explicit and removes any doubt. |
Contributor
That's amazing news! Hopefully you get the chance to commit soon, I'm excited to see how you got it to scale properly |
2 tasks
|
is this ready, cus if it is it probably shouldnt be marked as a draft. idk if devs will merge if its marked as draft. |
Contributor
Author
|
Hey, yes, gonna commit really soon, it's ready in my opinion, i've tested litterally anything, many many multiplayer fixes that pre-existed, now multiplayer is really stable and doesn't cause any crash. It's taking time because i am trying to make a comprehensive commit since there are a lot of changes |
Contributor
Author
|
Here as you can see i am quitting and re-joining world multiple time wich previously caused crashes for various reasons. 2026-03-08.14-55-49.mp42026-03-08.14-59-37.mp4 |
|
if you think its ready and its been tested properly i would open the pull request so that a developer can see it and merge it. these changes would be neat to have and i cant wait for them to be in the nightly build. |
Contributor
Author
|
yes, i will commit, then merge main into this then remove draft |
This was referenced Mar 8, 2026
Closed
|
imo the lower aspect ratios could use the 4:3 menus to make better use of screen real estate |
Collaborator
|
Does this support 4 player split screen? |
…yer stability Adds the networking layer for non-host split-screen multiplayer, implements live window resize with swap chain recreation, fixes bitmap font scaling at small window sizes, and fixes several crash-causing bugs in the multiplayer stack (compression buffer overflow, TCP stream desync, chunk visibility race, CompressedTileStorage torn reads, reconnect stability). == Non-host split-screen multiplayer == Each split-screen pad on a non-host client opens its own TCP connection to the host. From the host's perspective each connection looks like a normal remote player (gets its own smallId, Socket, PlayerConnection). WinsockNetLayer: JoinSplitScreen(), CloseSplitScreenConnection(), SplitScreenRecvThreadProc, per-pad socket/thread/smallId tracking (s_splitScreenSocket[], s_splitScreenSmallId[], s_splitScreenRecvThread[]). GetLocalSocket() returns the correct TCP socket for a given local sender's smallId. GetSplitScreenSmallId() returns the host-assigned smallId for a pad. GameNetworkManager::CreateSocket: non-host path (localPlayer && !IsHost() && IsInGameplay()) calls JoinSplitScreen, sets the IQNet slot's smallId and resolvedXuid, creates a non-hostLocal Socket + ClientConnection, sends PreLoginPacket, registers via addPendingLocalConnection. PlatformNetworkManagerStub::RemoveLocalPlayerByUserIndex: implemented the formerly-empty stub. Calls NotifyPlayerLeaving, CloseSplitScreenConnection, and clears the IQNet slot fields so the pad can rejoin cleanly. SmallId pool: s_nextSmallId starts at XUSER_MAX_COUNT (4), reserving m_player[0-3] for local pads so remote players never collide. IQNetPlayer::SendData: non-host local senders now route through GetLocalSocket(m_smallId) instead of always using SendToSmallId. IQNet::GetLocalPlayerByUserIndex: rewritten. Pad 0 on non-host uses GetLocalSmallId() for direct lookup; pads 1-3 check m_player[padIdx]. C_4JProfile::IsSignedIn: pad 0 always returns true (was checking controller connection, which is unreliable on Win64). GetGamertag/GetDisplayName: for pads 1-3 with active local players, returns the pad-specific gamertag from IQNet::m_player instead of always returning the primary username. ClientConnection: isPrimaryConnection() (true on host or for the primary pad on non-host) guards relative-delta and world-modifying handlers to prevent double-processing of shared state: - Guarded: handleMoveEntity, handleMoveEntitySmall, handleChunkTilesUpdate, handleBlockRegionUpdate, handleTileUpdate, handleTakeItemEntity, handleSignUpdate, handleTileEntityData, handleTileEvent, handleTileDestruction, handleComplexItemData, handleLevelEvent, handleSoundEvent, handleParticleEvent, handleAddGlobalEntity. - handleSetEntityMotion: secondary connections only accept motion targeting their own local player (knockback). - handleExplosion: world modification (finalizeExplosion) guarded, per-player knockback unguarded. Added null check on localplayers[]. - Entity spawn/remove/teleport/data handlers left unguarded (putEntity is idempotent, absolute value setters). handleLogin: added else clause to set level when the dimension already exists (was leaving level NULL on reconnect). handleChunkVisibilityArea/handleChunkVisibility: added null check on level. handleContainerOpen: added null check on localplayers[m_userIndex]. == Reconnect stability == PendingConnection: duplicate XUID no longer rejects with eDisconnect_Banned. Instead it force-disconnects the stale old connection via stalePlayer->connection->disconnect(), queues the old smallId for recycling via queueSmallIdForRecycle(), then calls handleAcceptedLogin for the new connection. MinecraftServer: swapped tick order so players->tick() (disconnect queue) runs before connection->tick() (new logins). The old player is removed from PlayerList before the new LoginPacket's XUID check runs. PlayerList: PushFreeSmallId and ClearSocketForSmallId moved here from DoWork, called only after PlayerConnection::disconnect() completes and the read/write threads are dead. New queueSmallIdForRecycle() method lets PendingConnection push smallIds into m_smallIdsToClose, which PlayerList::tick() processes through closePlayerConnectionBySmallId() for deferred cleanup. Prevents a race where the old write thread could resolve getPlayer() to a recycled smallId's new connection and send stale packets on it. SocketInputStreamLocal::close() and SocketOutputStreamLocal::close() now actually clear their queues (std::swap with empty queue instead of calling .empty() which is a read-only no-op). ServerConnection::stop(): pending and players vectors are snapshot-copied before iterating (prevents iterator invalidation). Remote players receive a DisconnectPacket via disconnect(eDisconnect_Quitting) instead of raw close(). tick(): added else clause so flush() only runs on live connections. WinsockNetLayer::Shutdown(): accept thread stopped first (prevents new recv threads from spawning), then all recv threads are collected and waited on, then connections are closed and split-screen sockets cleaned up. Clears disconnect and free-pool vectors before deleting critical sections. WinsockNetLayer::JoinGame(): waits for old s_clientRecvThread to fully exit before creating a new TCP connection. Prevents the old recv thread from reading bytes off the new socket and desynchronizing the stream. == Compression buffer overflow == CompressLZXRLE and CompressRLE wrote RLE intermediate output into a fixed 100KB buffer with no bounds checking. Full chunk columns are ~160KB and the RLE step can expand 0xFF bytes to 2 bytes each, easily overflowing into rleDecompressBuf and heap metadata. This caused delayed crashes in unrelated code (Packet::readPacket, LevelRenderer::updateDirtyChunks) after the first autosave, since that's when full chunks get compressed. Fix: dynamic allocation when worst-case RLE output (SrcSize * 2) exceeds the static buffer. Static buffer still used for small inputs (zero overhead). CompressRLE: moved LeaveCriticalSection after dynamic buffer cleanup. DecompressLZXRLE: now checks zlib return value (was completely ignored). On failure, bails out immediately with *pDestSize = 0. Added RLE input bounds checking (pucIn >= pucEnd before reading count/data bytes) and output bounds checking (pucOut + count > pucOutEnd). Same bounds checks applied to DecompressRLE. == Stream desync (Connection write thread) == The write thread had two output paths to the same TCP socket: bufferedDos (5KB buffered stream) and direct sos->writeWithFlags(). Chunk data sent via queueSend() used the direct path with shouldDelay=true, while other packets used bufferedDos. If bufferedDos had unflushed bytes, the direct write arrived at the client first, reordering the TCP stream and producing bad packet ID crashes. Fix: flush bufferedDos immediately before every direct sos->writeWithFlags(). == Chunk visibility race (empty first chunk after 30s) == BlockRegionUpdatePacket (direct socket write via queueSend) could arrive at the client before ChunkVisibilityAreaPacket (buffered). The client called getChunk() on a chunk that didn't exist yet in the cache, got EmptyLevelChunk (whose setBlocksAndData is a no-op), and silently lost the block data. On superflat this left one invisible chunk; on normal worlds it crashed the renderer. Fix: handleBlockRegionUpdate calls dimensionLevel->setChunkVisible() for full-chunk BRUPs before writing data, making it independent of packet ordering. Added post-write verification logging. CompressedTileStorage race: get() reads indicesAndData twice without a lock. compress() can swap the pointer between reads, producing indices from the old buffer paired with data from the new buffer. Fix: snapshot indicesAndData into a local variable before deriving both pointers. Same snapshot pattern applied to getData() (non-Vita path), isRenderChunkEmpty(), getHighestNonEmptyY(), getAllocatedSize(), and write(). All methods now also guard against NULL snapshots. == Window resize == ResizeD3D() destroys the old swap chain, creates a new one at the target size, then patches InternalRenderManager members directly via memory offsets (0x20=swap chain, 0x28=RTV, 0x50=SRV, 0x98=DSV, 0x5138/0x513C= backbuffer width/height). Offset verification cross-checks known pointers (device at 0x10, swap chain at 0x20) before patching. Old RTV/SRV are intentionally leaked (orphaned with the old swap chain) to avoid fighting unknown ref holders in the precompiled RenderManager. The flow: Suspend RenderManager, ClearState+Flush, release views, gdraw_D3D11_PreReset, destroy old swap chain, create new swap chain via IDXGIFactory, patch offsets, recreate RTV/SRV/DSV, rebind render targets, update UIController (updateRenderTargets + updateScreenSize), gdraw_D3D11_PostReset + SetRendertargetSize, IggyFlushInstalledFonts, Resume, PostProcesser::Init. WM_SIZE handling defers resize during window drag (WM_ENTERSIZEMOVE/ WM_EXITSIZEMOVE). Immediate resizes (maximize, programmatic) call ResizeD3D directly. Removed the old UpdateAspectRatio() function. CleanupDevice() was leaking g_pDepthStencilView and g_pDepthStencilBuffer. InitDevice: swap chain BufferUsage now includes DXGI_USAGE_SHADER_INPUT (needed for the SRV created from the backbuffer for CaptureThumbnail). New globals: g_rScreenWidth/g_rScreenHeight (real window dimensions, updated on resize) vs g_iScreenWidth/g_iScreenHeight (fixed logical resolution, stays 1920x1080). ComputeViewportForPlayer and getFovAndAspect now use g_rScreenWidth/ g_rScreenHeight instead of the fixed startup values, so 3D perspective and split-screen viewports adapt to window size. Main loop: rendering skipped when window is minimized (IsIconic check) to avoid 100% GPU usage on a hidden swap chain. Windows64_UIController: new updateRenderTargets(rtv, dsv) method updates cached D3D pointers used by gdraw_D3D11_SetTileOrigin every frame. UIController.h: new inline updateScreenSize(w, h) sets m_fScreenWidth/ m_fScreenHeight so all downstream UI code picks up the new size. == Bitmap font scaling == At small window sizes, dynamic text (scrollable list items, HowToPlay pages) showed overlapping characters. Static SWF text was unaffected because it uses embedded vector glyphs. Root cause in UIBitmapFont.cpp GetGlyphBitmap: when display scale is smaller than the bitmap's native scale (pixel_scale < truePixelScale, glyphScale stays at 1), Iggy displayed the glyph at native 1:1 pixel size but advanced the cursor by the smaller display-scale amount. At intermediate window sizes (e.g. 1678x756, scale factor ~0.7), a second bug appeared: some SWF font sizes produced pixel_scale just above truePixelScale (13 for Mojangles_11) while others fell just below, splitting glyphs across the small-display and normal cache branches. The normal branch cached all glyphs in a single [truePixelScale, 99] range, so the first glyph cached set pixel_scale_correct for every subsequent request regardless of font size. Different font sizes then got scaled by wrong ratios (e.g. 18.9/13.3 = 1.42x with point sampling), producing visibly inconsistent letter sizes. This only happened at specific window sizes where the display scale put some fonts above and others below the truePixelScale boundary. Full 1080p and very small windows were unaffected because all fonts landed in the same branch. Fix: on _WINDOWS64, always use pixel_scale_correct = truePixelScale so every cache entry is consistent regardless of which font size creates it first. Two cache ranges: downscale (pixel_scale < truePixelScale) uses bilinear for smooth reduction, upscale uses point_sample for crisp pixel-art rendering. At most two cache entries per glyph. The console code path (fixed resolution, integer-multiple scaling) is preserved behind #else. UIScene.cpp loadMovie: always load 1080.swf on _WINDOWS64 regardless of window size. The old height-based selection could pick 480 or 720 variants which either crashed or loaded the wrong skin library (skinHD.swf vs skin.swf). Display size is now set via Fit16x9 BEFORE the init tick so Iggy's ActionScript text field creation sees the same scale that render() will use. IggyFlushInstalledFonts() called after init tick to clear stale glyph cache entries from previous scenes. Font.cpp addCharacterQuad/renderCharacter: yOff was computed with m_charWidth instead of m_charHeight, producing wrong texture coordinates for non-square glyph cells. This is the world-rendering font (chat, signs, name tags), not the Iggy UI font. == XUID generation == Split-screen pad XUIDs derived by hashing baseXuid + iPad through Mix64 (DeriveXuidForPad in Windows64_Xuid.h) instead of simple addition. Pad 0 returns the base XUID unchanged for save compatibility. Includes validity fallbacks if the hash produces an invalid XUID. (Suggested by rtm516) == Misc == Packet::readPacket: thread-local ring buffer tracks last 8 good packet IDs. On bad packet ID, dumps the history plus next 32 bytes of stream for diagnosing TCP desynchronization. PendingConnection/PlayerList: debug logging for the reconnect flow (duplicate XUID handling, force-disconnect, handleAcceptedLogin, placeNewPlayer with smallId/entityId/dimension). ClientConnection::handleBlockRegionUpdate: warning log when a full chunk arrives with ys==0 (empty full chunk, data loss indicator). == Known issues / future work == SendOnSocket global lock (WinsockNetLayer.cpp): s_sendLock is a single CriticalSection serializing ALL TCP sends across ALL connections. If one client's send() blocks (TCP window full, slow network), every other write thread stalls — no data flows to any player until the slow send completes. Each PlayerConnection has its own write thread, so with 8+ players one slow client can cause latency spikes or timeout disconnects for healthy players. Fix: replace s_sendLock with per-socket locks indexed by smallId. The lock only needs to prevent header+payload interleaving on the SAME socket; sends to different sockets are independent. Deferred to a separate PR to keep this one focused. Textures::releaseTexture: early return for id <= 0, checks TextureGetTexture(id) != NULL before calling glDeleteTextures. Prevents crashes on stale texture IDs after RenderManager reset. UIController TextureSubstitutionDestroyCallback: null guard on Minecraft::GetInstance() and mc->textures before calling releaseTexture. Prevents crash during shutdown. StringTable: removed __debugbreak() on language load failure in debug builds.
Contributor
Author
|
@TorutheRedFox nice idea, will look forward to this but it's not necessary right now. @rtm516 yes |
Contributor
Author
|
Took some time but it's now perfect |
# Conflicts: # Minecraft.Client/Common/UI/UIScene_Keyboard.cpp
Collaborator
|
Thanks for this, awesome PR! fixes a ton of bugs |
|
This PR breaks multiplayer functionality. |
+1, this pr probably needs to be reverted and tested properly |
Collaborator
|
we're not reverting this whole PR based on cross-version multiplayer not working. if there's an issue with it, it needs to be clearly reported properly so it can be fixed. |
This comment was marked as spam.
This comment was marked as spam.
Repository owner
locked as off-topic and limited conversation to collaborators
Mar 9, 2026
Collaborator
|
I will no longer be merging multi topic PRs due to the issues that have arisen after merging this PR. Going forward PRs should be limited in scope to a single purpose to avoid issues. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to subscribe to this conversation on GitHub.
Already have an account?
Sign in.
9 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Description
Split-screen multiplayer for Windows64, plus window resize support, bitmap font fixes, and several crash-causing multiplayer stability fixes. Split-screen was completely non-functional on this platform, window resize wasn't supported, and the networking stack had multiple bugs that surfaced under real multiplayer load (compression buffer overflows, TCP stream desync, chunk visibility races, CompressedTileStorage torn reads, reconnect crashes).
Changes
1. Split-screen (local multiplayer)
Previous Behavior
Split-screen did not work at all on Windows64. Secondary controllers couldn't sign in (
IsSignedInwas hardcoded to pad 0),AddLocalPlayerByUserIndexwas a no-op stub, andGetXUIDreturned invalid for all pads except 0. Even if a second player somehow joined, the 3D world was horizontally stretched because the projection matrix ignored per-viewport aspect ratios. Mouse input only hit the top-left quadrant with no concept of viewport offsets or per-quadrant scaling. The cursor fought between players every tick. On ultrawide monitors the game auto-detected native resolution viaGetSystemMetrics, breaking the 16:9 assumption in viewport math and Flash layout entirely.Root Cause
The Win64 stubs in
Extrax64Stubs.cppwere never implemented for multi-pad support.IsSignedInonly returned true for pad 0,GetXUIDonly returned a valid ID for pad 0, andAddLocalPlayerByUserIndexdid nothing.gluPerspectiveinglWrapper.cppignored its aspect parameter and always used the globalg_iAspectRatio(full-screen 16:9), so split-screen viewports (which are half-width or half-height) got the wrong projection. Mouse coordinate conversion had no concept of viewport offsets or per-quadrant scaling. Screen resolution was pulled from the monitor instead of being pinned to the game's internal 1920x1080.New Behavior
Split-screen works on Windows64. Up to 4 players can join with controllers, each getting a correctly rendered viewport with proper aspect ratio. Mouse hover and click work in all quadrants for the primary (KBM) player. Internal resolution is pinned to 1920x1080, keeping viewport math and Flash layout correct on any monitor aspect ratio.
Fix Implementation
IsSignedInnow checksIsPadConnectedinstead of hardcoded pad 0.GetXUIDreturns a unique value per pad viaDeriveXuidForPad(Mix64 hash inWindows64_Xuid.h, pad 0 returns base XUID unchanged for save compatibility, includes validity fallbacks if the hash produces an invalid XUID; suggested by rtm516).AddLocalPlayerByUserIndexproperly initializes the slot with gamertag and remote/host flags.NotifyPlayerJoineduses the staticm_playerarray directly sinceGetLocalPlayerByUserIndexcan't find the player beforecustomDatais set.gluPerspectivepasses the caller's aspect parameter toRenderManager.MatrixPerspectiveinstead of the globalg_iAspectRatio. Split-screen viewports already compute the correct per-viewport aspect ingetFovAndAspect.iPad == ProfileManager.GetPrimaryPad()so split-screen players don't fight over cursor state. RemovedIsTutorialVisibleguard from Escape key pause check.GetSystemMetricsauto-detection, pinnedg_iScreenWidth/g_iScreenHeightto 1920x1080.2. Non-host split-screen networking
Previous Behavior
Only the primary pad on a non-host client had a network connection to the host. Secondary split-screen pads on a non-host machine had no way to send or receive packets, making non-host split-screen multiplayer impossible.
Root Cause
The networking layer assumed one TCP connection per machine. There was no mechanism for additional local pads to open their own connections, no per-pad socket tracking, and no routing logic to send packets through the correct socket based on which pad originated them.
New Behavior
Each split-screen pad on a non-host client opens its own TCP connection to the host. From the host's perspective, each connection looks like a normal remote player (gets its own smallId, Socket, PlayerConnection). Secondary pads can join and leave independently.
Fix Implementation
JoinSplitScreen(),CloseSplitScreenConnection(),SplitScreenRecvThreadProcfor per-pad socket/thread/smallId tracking (s_splitScreenSocket[],s_splitScreenSmallId[],s_splitScreenRecvThread[]).GetLocalSocket()returns the correct TCP socket for a given local sender's smallId.GetSplitScreenSmallId()returns the host-assigned smallId for a pad.localPlayer && !IsHost() && IsInGameplay()) callsJoinSplitScreen, sets the IQNet slot's smallId and resolvedXuid, creates a non-hostLocal Socket + ClientConnection, sends PreLoginPacket, registers viaaddPendingLocalConnection.NotifyPlayerLeaving,CloseSplitScreenConnection, and clears the IQNet slot fields so the pad can rejoin cleanly.s_nextSmallIdstarts atXUSER_MAX_COUNT(4), reservingm_player[0-3]for local pads so remote players never collide.GetLocalSocket(m_smallId)instead of always usingSendToSmallId.GetLocalSmallId()for direct lookup; pads 1-3 checkm_player[padIdx].IQNet::m_playerinstead of always returning the primary username.isPrimaryConnection()(true on host or for the primary pad on non-host) guards relative-delta and world-modifying handlers to prevent double-processing of shared state:handleMoveEntity,handleMoveEntitySmall,handleChunkTilesUpdate,handleBlockRegionUpdate,handleTileUpdate,handleTakeItemEntity,handleSignUpdate,handleTileEntityData,handleTileEvent,handleTileDestruction,handleComplexItemData,handleLevelEvent,handleSoundEvent,handleParticleEvent,handleAddGlobalEntity.handleSetEntityMotion: secondary connections only accept motion targeting their own local player (knockback).handleExplosion: world modification (finalizeExplosion) guarded, per-player knockback unguarded. Added null check onlocalplayers[].putEntityis idempotent, absolute value setters).localplayers[m_userIndex].3. Reconnect stability
Previous Behavior
Reconnecting to a server after a disconnect would crash or produce undefined behavior. Duplicate XUIDs were rejected with
eDisconnect_Banned. The old player's write thread could race with the new connection's smallId assignment, sending stale packets on the wrong connection.SocketInputStream/OutputStream::close()didn't actually clear their queues.ServerConnection::stop()had iterator invalidation bugs and didn't send proper disconnect packets to remote players.Root Cause
Multiple race conditions in the disconnect/reconnect flow. The old player wasn't fully cleaned up before the new login was processed. SmallId recycling happened too early (before write threads were dead). Stream close methods called
.empty()(a read-only check) instead of actually clearing the queue. Server shutdown iterated vectors while modifying them.New Behavior
Reconnecting works cleanly. The stale connection is force-disconnected, its smallId is queued for deferred recycling, and the new login proceeds only after the old player is fully removed.
Fix Implementation
eDisconnect_Banned. Instead it force-disconnects the stale old connection viastalePlayer->connection->disconnect(), queues the old smallId for recycling viaqueueSmallIdForRecycle(), then callshandleAcceptedLoginfor the new connection.players->tick()(disconnect queue) runs beforeconnection->tick()(new logins). The old player is removed from PlayerList before the new LoginPacket's XUID check runs.PushFreeSmallIdandClearSocketForSmallIdmoved here from DoWork, called only afterPlayerConnection::disconnect()completes and the read/write threads are dead. NewqueueSmallIdForRecycle()method lets PendingConnection push smallIds intom_smallIdsToClose, whichPlayerList::tick()processes throughclosePlayerConnectionBySmallId()for deferred cleanup. Prevents a race where the old write thread could resolvegetPlayer()to a recycled smallId's new connection and send stale packets on it.std::swapwith empty queue instead of calling.empty()which is a read-only no-op).disconnect(eDisconnect_Quitting)instead of rawclose().tick(): added else clause soflush()only runs on live connections.s_clientRecvThreadto fully exit before creating a new TCP connection. Prevents the old recv thread from reading bytes off the new socket and desynchronizing the stream.4. Compression buffer overflow
Previous Behavior
Delayed crashes in unrelated code (
Packet::readPacket,LevelRenderer::updateDirtyChunks) after the first autosave (~30 seconds into a multiplayer game). The crashes appeared random and were extremely difficult to reproduce consistently.Root Cause
CompressLZXRLEandCompressRLEwrote RLE intermediate output into a fixed 100KB buffer with no bounds checking. Full chunk columns are ~160KB and the RLE step can expand 0xFF bytes to 2 bytes each, easily overflowing intorleDecompressBufand heap metadata. The corruption was silent until unrelated code touched the corrupted memory, which is why crashes appeared inreadPacketorupdateDirtyChunksrather than in the compression code itself.New Behavior
Compression handles arbitrarily large inputs without overflow. Small inputs still use the static buffer (zero allocation overhead). Decompression validates all return values and bounds.
Fix Implementation
SrcSize * 2) exceeds the static buffer. Static buffer still used for small inputs.CompressRLE: movedLeaveCriticalSectionafter dynamic buffer cleanup.*pDestSize = 0. Added RLE input bounds checking (pucIn >= pucEndbefore reading count/data bytes) and output bounds checking (pucOut + count > pucOutEnd). Same bounds checks applied toDecompressRLE.5. TCP stream desync
Previous Behavior
Occasional bad packet ID crashes during multiplayer, especially under load. The crash happened on the client side with an impossible packet ID value.
Root Cause
The write thread had two output paths to the same TCP socket:
bufferedDos(5KB buffered stream) and directsos->writeWithFlags(). Chunk data sent viaqueueSend()used the direct path withshouldDelay=true, while other packets usedbufferedDos. IfbufferedDoshad unflushed bytes, the direct write arrived at the client first, reordering the TCP stream and producing bad packet ID crashes.New Behavior
All writes to the TCP socket are properly ordered. Direct writes always flush the buffered stream first.
Fix Implementation
Flush
bufferedDosimmediately before every directsos->writeWithFlags()call.6. Chunk visibility race
Previous Behavior
On superflat worlds, one chunk would occasionally be invisible (empty air). On normal worlds with terrain, the same bug crashed the renderer. This happened roughly 30 seconds into a multiplayer session and was tied to the autosave interval.
Root Cause
BlockRegionUpdatePacket(sent via direct socket write throughqueueSend) could arrive at the client beforeChunkVisibilityAreaPacket(sent through the buffered stream). The client calledgetChunk()on a chunk that didn't exist yet in the cache, gotEmptyLevelChunk(whosesetBlocksAndDatais a no-op), and silently lost the block data.A separate race existed in
CompressedTileStorage:get()readsindicesAndDatatwice without a lock.compress()can swap the pointer between reads, producing indices from the old buffer paired with data from the new buffer (torn read).New Behavior
Block data is never silently dropped. Chunk visibility is set before processing block data, making the handler independent of packet ordering.
Fix Implementation
dimensionLevel->setChunkVisible()for full-chunk BRUPs before writing data, making it independent of packet ordering. Added post-write verification logging. Warning log when a full chunk arrives withys==0(empty full chunk, data loss indicator).indicesAndDatainto a local variable before deriving both pointers. Same snapshot pattern applied togetData()(non-Vita path),isRenderChunkEmpty(),getHighestNonEmptyY(),getAllocatedSize(), andwrite(). All methods now also guard against NULL snapshots.7. Window resize
Previous Behavior
The game window could not be resized at all on Windows64. The swap chain was created once at startup and never updated. Resizing the window or switching between windowed/fullscreen either did nothing or crashed.
Root Cause
Two obstacles prevented standard D3D11 resize:
ResizeBuffersfails. The precompiledRenderManager(4J lib) holds internal references to the backbuffer thatSuspend()does not release.Suspend()/Resume()are just boolean flag setters (decompiled:m_bSuspended = true/false) with no resource release, no thread synchronization, and no memory barriers. Even after releasing every known ref (g_pRenderTargetView, the Renderer's own RTV at offset 0x28, its SRV at 0x50, depth stencil at 0x98), callingClearState() + Flush()on the immediate context,gdraw_D3D11_PreReset(), andPostProcesser::Cleanup(),ResizeBuffersstill returnsDXGI_ERROR_INVALID_CALL(0x887A0001). The remaining refs are held by unknown internal code, likely the Renderer's command buffer system, deferred contexts, or TLS-cached render contexts that can't be ClearState'd from outside the lib.Initialise()wipes textures.RenderManager.Initialise()creates all D3D state from scratch (backbuffer RTV, SRV texture, mip-level textures, shaders, index buffers), then doesmemset(m_textures, 0, sizeof(m_textures))— wiping all 512 texture slots (each 40 bytes:{allocated, ID3D11Texture2D*, ID3D11ShaderResourceView*, flags, mipLevels, format, samplerParams}) without callingRelease()on existing textures. The game crashes when scene destruction callbacks fireTextureFree()on stale IDs.The key insight was that
C4JRender RenderManageris a stateless wrapper with zero member variables — all methods delegate toRenderer InternalRenderManager. Patching&RenderManagerdirectly hit garbage memory. The actual D3D state lives at&InternalRenderManager, and its memory layout was mapped by decompiling the 4J lib and cross-checking known pointers at runtime.Another critical discovery:
StartFrame()bindsrenderTargetViewanddepthStencilViewviaOMSetRenderTargetsand sets the viewport frombackBufferWidth/backBufferHeightevery single frame. If these members aren't patched to the new dimensions, every frame resets the viewport to the old size, making it look like the resize didn't work even though the swap chain was correctly recreated.Approaches that were tried and failed:
ResizeBuffersafter Suspend (outstanding backbuffer ref),ResizeBufferswith RTV patching (unknown refs remain), direct patching with&RenderManagerbase (no members — offsets hit garbage), force-releasing backbuffer viaRelease()loop (D3D11 debug layer crashes on over-release), swap chain recreation +Initialise()(texture wipe), callingInitialise()while suspended or beforegdraw_D3D11_PostReset()(both crash), keeping old swap chain alive alongside new one (crash).New Behavior
The game window can be freely resized by dragging borders, maximizing, or switching to fullscreen. The 3D world renders at the new native resolution, Flash UI adapts via Fit16x9 pillarboxing. Rendering is skipped when the window is minimized (avoids 100% GPU usage on a hidden swap chain).
Fix Implementation
ResizeBuffers(no outstanding ref issues) andInitialise()(no texture wipe).Windows64_Minecraft.cpp: Destroys the old swap chain, creates a new one at the target size viaIDXGIFactory::CreateSwapChain, then patchesInternalRenderManagermembers directly via memory offsets:0x20=m_pSwapChain(soPresent()targets the new swap chain)0x28=renderTargetView(soStartFrame()binds the correct render target)0x50=renderTargetShaderResourceView(soCaptureThumbnail()reads from the right SRV)0x98=depthStencilView(soStartFrame()binds the correct depth buffer)0x5138/0x513C=backBufferWidth/backBufferHeight(soStartFrame()sets the correct viewport)m_pDevice(0x10) andm_pSwapChain(0x20) against known globals.g_bResizeReady), (2) verify offsets at runtime, (3)Suspend()+ wait forSuspended(), (4)PostProcesser::Cleanup(), (5)ClearState() + Flush(), (6) releaseg_pRenderTargetView,g_pDepthStencilView,g_pDepthStencilBuffer, (7)gdraw_D3D11_PreReset(), (8)g_pSwapChain->Release(), (9)IDXGIFactory::CreateSwapChainat new dimensions, (10) patch swap chain pointer at 0x20, (11)GetBuffer(0)+CreateRenderTargetView+ patch RTV at 0x28, (12) create backbuffer-sized texture (RENDER_TARGET|SHADER_RESOURCE) +CreateShaderResourceView+ patch SRV at 0x50, (13) create depth texture +CreateDepthStencilView+ patch DSV at 0x98, (14) patchbackBufferWidth/backBufferHeightat 0x5138/0x513C, (15)OMSetRenderTargets+RSSetViewports, (16)updateRenderTargets()+updateScreenSize()on UIController, (17)gdraw_D3D11_PostReset()+SetRendertargetSize(), (18)IggyFlushInstalledFonts(), (19)Resume(), (20)PostProcesser::Init().Suspend()before any resource release (stopsStartFrame()/Present()calls).ClearState() + Flush()before swap chain destruction (unbinds all pipeline state).gdraw_D3D11_PreReset()before swap chain destruction (GDraw releases its internal render targets). Swap chain pointer patch immediately after creating the new swap chain (Present()must target the new chain whenResume()fires).gdraw_D3D11_PostReset()beforeResume()(GDraw must recreate state objects before the next frame).PostProcesser::Init()after new swap chain exists (it callsGetBuffer(0)to read backbuffer dimensions).WM_ENTERSIZEMOVE/WM_EXITSIZEMOVE). Immediate resizes (maximize, programmatic) callResizeD3Ddirectly.g_rScreenWidth/g_rScreenHeight(real window dimensions, updated on resize) vsg_iScreenWidth/g_iScreenHeight(fixed logical resolution, stays 1920x1080).g_rScreenWidth/g_rScreenHeightinstead of the fixed startup values, so 3D perspective and split-screen viewports adapt to window size.IsIconiccheck).ConsoleUIControllercachesm_pRenderTargetViewandm_pDepthStencilViewas raw pointer copies (NOT AddRef'd). These are passed togdraw_D3D11_SetTileOrigin()every frame duringrender()andsetTileOrigin()(per-viewport tile origin for split-screen). NewupdateRenderTargets(rtv, dsv)method updates these cached pointers after new views are created during resize.updateScreenSize(w, h)setsm_fScreenWidth/m_fScreenHeight. These are used bysetupRenderPosition()(Fit16x9 pillarbox offsets),getRenderDimensions()(Flash movie display sizes), andIggyPlayerSetDisplaySize()calls during scene rendering. Must only be called when the resize actually succeeds — calling it with new dimensions while the backbuffer is still old-sized causes ghost/double rendering.gdraw_D3D11_PreReset()before swap chain destruction (releases GDraw's internal render targets),gdraw_D3D11_PostReset()after new swap chain creation (recreates GDraw state objects, must happen beforeResume()),gdraw_D3D11_SetRendertargetSize(w, h)after PostReset (tells GDraw the new backbuffer dimensions).IggyFlushInstalledFonts()called after PostReset to flush cached bitmap font textures that were sized for the old resolution.Cleanup()releases all resources before swap chain destruction,Init()called after new swap chain exists — it callsGetBuffer(0)to read backbuffer dimensions, then creates new offscreen resources at the new size.BufferUsagenow includesDXGI_USAGE_SHADER_INPUT(needed for the SRV created from the backbuffer forCaptureThumbnail).g_pDepthStencilViewandg_pDepthStencilBuffer.UpdateAspectRatio()function.8. Bitmap font scaling
Previous Behavior
At small window sizes, dynamic text (scrollable list items, HowToPlay pages) showed overlapping/inconsistent character sizes. Static SWF text (pre-authored buttons) was unaffected. At intermediate window sizes (e.g. 1678x756), some characters appeared at different scales than others within the same text block.
Root Cause
Two bugs in
UIBitmapFont.cpp GetGlyphBitmap:Small window overlap: When display scale < bitmap native scale (
pixel_scale < truePixelScale),glyphScalestays at 1, so Iggy displayed the glyph at native 1:1 pixel size but advanced the cursor by the smaller display-scale amount. Result: glyphs wider than their advance, massive overlap.Intermediate window inconsistency: Some SWF font sizes produced
pixel_scalejust abovetruePixelScale(13 for Mojangles_11) while others fell just below, splitting glyphs across the small-display and normal cache branches. The normal branch cached all glyphs in a single[truePixelScale, 99]range, so the first glyph cached setpixel_scale_correctfor every subsequent request regardless of font size. Different font sizes then got scaled by wrong ratios (e.g.18.9/13.3 = 1.42xwith point sampling), producing visibly inconsistent letter sizes.A third unrelated bug in
UIScene.cpp loadMovie: the height-based SWF selection could pick 480 or 720 variants on Windows64, which either crashed or loaded the wrong skin library (skinHD.swfvsskin.swf).New Behavior
Dynamic text renders correctly at all window sizes. Characters are consistent regardless of which font size gets cached first. 1080.swf is always loaded on Windows64.
Fix Implementation
_WINDOWS64, always usepixel_scale_correct = truePixelScaleso every cache entry is consistent regardless of which font size creates it first. Two cache ranges: downscale (pixel_scale < truePixelScale) uses bilinear for smooth reduction, upscale usespoint_samplefor crisp pixel-art rendering. At most two cache entries per glyph. The console code path (fixed resolution, integer-multiple scaling) is preserved behind#else.1080.swfon_WINDOWS64regardless of window size. Display size is set viaFit16x9BEFORE the init tick so Iggy's ActionScript text field creation sees the same scale thatrender()will use.IggyFlushInstalledFonts()called after init tick to clear stale glyph cache entries from previous scenes.yOffwas computed withm_charWidthinstead ofm_charHeight, producing wrong texture coordinates for non-square glyph cells. This is the world-rendering font (chat, signs, name tags), not the Iggy UI font.9. In-game keyboard improvements
Previous Behavior
The in-game keyboard scene had no cursor movement support in PC mode. Text could only be appended or deleted from the end.
New Behavior
Full cursor navigation in PC mode: arrow keys move the cursor, Home/End jump to start/end, Delete removes the character after the cursor. Flash caret syncs with cursor position.
Fix Implementation
m_iCursorPostracking for PC mode with full arrow key, Home, End, Delete support and Flash caret sync viaSetCaretIndex.10. Misc fixes and diagnostics
id <= 0, checksTextureGetTexture(id) != NULLbefore callingglDeleteTextures. Prevents crashes on stale texture IDs after RenderManager reset.Minecraft::GetInstance()andmc->texturesbefore callingreleaseTexture. Prevents crash during shutdown.__debugbreak()on language load failure in debug builds.handleAcceptedLogin,placeNewPlayerwith smallId/entityId/dimension).Known issues / future work
SendOnSocket global lock (
WinsockNetLayer.cpp):s_sendLockis a single CriticalSection serializing ALL TCP sends across ALL connections. If one client'ssend()blocks (TCP window full, slow network), every other write thread stalls. Each PlayerConnection has its own write thread, so with 8+ players one slow client can cause latency spikes or timeout disconnects for healthy players. Fix: replaces_sendLockwith per-socket locks indexed by smallId. Deferred to a separate PR.AI Use Disclosure
Claude 4.6 Opus was responsibly used as a senior dev assistant throughout this work:
ResizeBuffersandInitialise()couldn't work, suggesting which offsets to investigate and why.All code changes were written by me. The few repetitive and tedious pieces of code that were written with AI assistance were human-reviewed and tested locally. All architecture decisions, debugging strategies, and implementation choices are mine.