Commit Graph
5 Commits
Author SHA1 Message Date
Cursor Agentand西风逍遥游 d80df444aa refactor(scene): register link kinds instead of hard-coding them
A scene link keeps an external object in step with the node it hangs off. The
kinds were a closed enum with a switch in TransformSystem, so scene had to
include graphics, physics and audio to do the casts, and no other module could
ever attach to the tree without editing SceneHost.h -- particles, voxel,
spritestack, avatar and weather all had to stay outside it.

Kinds are registered now (scene/SceneLink.h). The module owning the target type
supplies pushWorld / pullWorld / alive, keeps the casts on its own side where
the type is complete, and scene never learns what a Renderable3D is. Dispatch is
a function-pointer table rather than virtual calls, since link sync runs for
every dirty node every frame.

The six typed helpers (linkRenderable2D and friends) stay as conveniences and
now resolve their kind by name, so they return false rather than crash when the
module that registers it was trimmed out. The script API already named kinds
with strings, which is exactly the registry key, so it is unchanged.

Direction is now the right way round: graphics, physics and audio depend on
scene rather than scene on them, and scene drops from L5 to L1. What remains is
scene -> graphics from pickScreenAt / collectFrustumIdsAt, which take a Camera3D
and read its projection; that is recorded in the back-edge whitelist as the next
thing to move.

Each provider lives in a SceneLinks.cpp, sharing the basename so the three come
out of the build together when scene is disabled.

Co-authored-by: 西风逍遥游 <sunxfancy@gmail.com>
2026-08-18 13:46:25 +00:00
Cursor Agentand西风逍遥游 f7aacbc4a9 refactor(event): let modules translate their own platform events
The SDL pump did the translation itself, so it had to know about keyboard,
joystick, touch, window, graphics, audio and ui: eight upward dependencies from
one of the engine's lowest modules, and the source of every remaining
dependency cycle.

Each module now registers an IPlatformEventSink for the events it owns and the
pump only dispatches. Key and text events move to keyboard/sdl, joystick and
game-controller events (including the hotplug device table) to joystick/sdl,
finger events to touch/sdl, resize and the mobile background/foreground surface
watch to window/sdl, the ImGui observer to ui, and Audio::pump to an
end-of-pump hook in audio. The pump keeps only quit and window-close, which are
about the event loop rather than any one module.

Sinks run in priority order -- observer, then surface, then input -- so ImGui
still sees events first and the window is resized before anything reads its
size. The mobile surface teardown keeps using SDL_AddEventWatch rather than the
poll path, since by the time SDL_PollEvent returns the native surface is gone;
the watch is now installed by the window module on its first pump.

Also removes exceptionIfInRenderPass, whose body has been commented out since
it was written and whose only remaining effect was a graphics include.

The engine dependency graph now has zero cycles, down from 26, and the back-edge
whitelist is down to 2 (window -> graphics, window -> image). Layering is also
right way up: event sits at L0 with the input modules above it at L1/L2.

Verified: unit tests for event/input/ui/audio/thread pass, and the example
smoke suite gives the same 19 pass / 9 fail as the pre-change baseline (those 9
fail on main too, from script API drift in the examples themselves).

Co-authored-by: 西风逍遥游 <sunxfancy@gmail.com>
2026-08-18 13:46:04 +00:00
Cursor Agentand西风逍遥游 164e8c1b4c refactor(filesystem): dispatch hot reload instead of hard-coding asset kinds
HotReload knew what lived at every watched path: it included graphics to
refresh textures, particles to reload emitter configs and map to reload tile
layers. That made the engine's lowest-level module depend on three of its
highest, and it was the only reason filesystem sat in a dependency cycle.

Each module now registers an eve::caps::IAssetReloader for the assets it owns
and HotReload only routes paths. Behaviour is unchanged: automatic dispatch
still matches by extension, an explicit bind() still targets one kind and
ignores the extension, and the texture reloader runs at kTexture priority so
the GPU upload happens before emitters and tile layers re-bind to it.

The reloadable set now follows the linked module set, so a build without
particles simply has no particle reloader rather than a dangling reference.
Reload kinds also become extensible: a module or plugin can add one without
touching filesystem.

Back-edge whitelist 14 -> 11; the dependency graph is now acyclic apart from
the event pump.

(map.render.tilesContributeDepthY fails when run after the hotreload cases both
before and after this change -- pre-existing test-order state, not from here.)

Co-authored-by: 西风逍遥游 <sunxfancy@gmail.com>
2026-08-18 13:46:04 +00:00
Cursor Agentand西风逍遥游 08679654e0 feat(common): add the capability registry, use it for GPU info and main-thread posts
Introduces eve::cap in common/Capability.h: a name-keyed registry where the
consumer owns the interface, the provider registers an implementation, and
neither includes the other. It generalises what common/RenderTrace.h already
does for the render tracer, and supports both a single provider (provide/query)
and many ordered ones (addListener/forEach), keyed by string so a provider
compiled into a plugin lands in the same slot as the host.

Two upward dependencies go away:

  system -> graphics   IGpuInfo. The Vulkan physical-device fields move into
                       graphics/vulkan/GpuInfoProvider.cpp, with a WebGPU
                       counterpart, so System.cpp loses both the graphics
                       include and its EVENGINE_WEBGPU branches.

  thread -> event      IMainThreadPost. postMain and submitPost now enqueue
                       through the capability. The interface makes explicit
                       what was previously only a comment in submitPost: the
                       submitting thread calls prepare() so the worker never
                       reaches the module registry.

Both now report an error when no provider is linked, rather than lazily
creating a module the build may not contain.

Step S3 of docs/dev/模块编排与裁剪架构.md; back-edge whitelist 16 -> 14.

Co-authored-by: 西风逍遥游 <sunxfancy@gmail.com>
2026-08-18 13:46:04 +00:00
Cursor Agentand西风逍遥游 c4fd641115 docs: add module orchestration and trimming architecture analysis
Analyse how the 53 modules currently collaborate and why none of them can be
trimmed outside the hardcoded WASM axis, then lay out a target architecture and
a six-step route to per-module build options.

Adds scripts/module_depgraph.py, which produces the dependency graph, the
7-layer topological ordering and the cycle list the document cites. Its --check
mode fails on any back-edge outside the 16 known ones, so it doubles as the CI
guardrail for step S1.

Co-authored-by: 西风逍遥游 <sunxfancy@gmail.com>
2026-08-18 13:46:04 +00:00