Removed use of std::exit and changed recomp runtime to exit normally, added helpers for getting recompiled args and returning values, added example of patch code calling native code

include/recomp_helpers.h

@@ -0,0 +1,46 @@
+#ifndef __RECOMP_HELPERS__
+#define __RECOMP_HELPERS__
+
+#include "recomp.h"
+
+template<int index, typename T>
+T _arg(uint8_t* rdram, recomp_context* ctx) {
+    static_assert(index < 4, "Only args 0 through 3 supported");
+    gpr raw_arg = (&ctx->r4)[index];
+    if constexpr (std::is_same_v<T, float>) {
+        if constexpr (index < 2) {
+            static_assert(index != 1, "Floats in arg 1 not supported");
+            return ctx->f12.fl;
+        }
+        else {
+            return std::bit_cast<T>(raw_arg);
+        }
+    }
+    else if constexpr (std::is_pointer_v<T>) {
+        static_assert (!std::is_pointer_v<std::remove_pointer_t<T>>, "Double pointers not supported");
+        return TO_PTR(std::remove_pointer_t<T>, raw_arg);
+    }
+    else if constexpr (std::is_integral_v<T>) {
+        static_assert(sizeof(T) <= 4, "64-bit args not supported");
+        return static_cast<T>(raw_arg);
+    }
+    else {
+        // static_assert in else workaround
+        [] <bool flag = false>() {
+            static_assert(flag, "Unsupported type");
+        }();
+    }
+}
+
+template <typename T>
+void _return(recomp_context* ctx, T val) {
+    static_assert(sizeof(T) <= 4 && "Only 32-bit value returns supported currently");
+    if (std::is_same_v<T, float>) {
+        ctx->f0.fl = val;
+    }
+    else if (std::is_integral_v<T> && sizeof(T) <= 4) {
+        ctx->r2 = int32_t(val);
+    }
+}
+
+#endif

include/recomp_ui.h

@@ -21,5 +21,6 @@ enum class Menu {
 };
 
 void set_current_menu(Menu menu);
+void destroy_ui();
 
 #endif

include/rt64_layer.h

@@ -76,6 +76,7 @@ DLLIMPORT void ProcessRDPList(void);
 DLLIMPORT void ProcessDList(void);
 DLLIMPORT void UpdateScreen(void);
 DLLIMPORT void ChangeWindow(void);
+DLLIMPORT void PluginShutdown(void);
 
 void set_rt64_hooks();
 

patches/Makefile

@@ -5,7 +5,7 @@ LD      := ld.lld
 OBJCOPY := llvm-objcopy
 
 CFLAGS   := -target mips -mips2 -mabi=32 -O2 -mno-odd-spreg -fomit-frame-pointer -G0 -Wall -Wextra -Wno-incompatible-library-redeclaration -Wno-unused-parameter -Wno-unknown-pragmas -Wno-unused-variable
-CPPFLAGS := -nostdinc -D_LANGUAGE_C -I ../../mm/include -I ../../mm/src -I ../../mm/build -I ../../mm/assets
+CPPFLAGS := -nostdinc -D_LANGUAGE_C -DMIPS -I ../../mm/include -I ../../mm/src -I ../../mm/build -I ../../mm/assets
 LDFLAGS  := -nostdlib -T patches.ld -T syms.ld
 BINFLAGS := -O binary
 

patches/cheats.c

@@ -1,5 +1,4 @@
-#define Audio_PlaySfx play_sound
-#include "global.h"
+#include "patches.h"
 
 // Infinite magic
 s32 Magic_Consume(PlayState* play, s16 magicToConsume, s16 type) {

patches/culling.c

@@ -1,4 +1,4 @@
-#include "global.h"
+#include "patches.h"
 
 // Disable frustum culling for actors, but leave distance culling intact
 s32 func_800BA2FC(PlayState* play, Actor* actor, Vec3f* projectedPos, f32 projectedW) {

patches/input.c

@@ -0,0 +1,24 @@
+#include "patches.h"
+#include "input.h"
+
+u32 sPlayerItemButtons[] = {
+    BTN_B,
+    BTN_CLEFT,
+    BTN_CDOWN,
+    BTN_CRIGHT,
+};
+
+// Return currently-pressed button, in order of priority B, CLEFT, CDOWN, CRIGHT.
+EquipSlot func_8082FDC4(void) {
+    EquipSlot i;
+    RecompInputs cur_inputs;
+    recomp_get_item_inputs(&cur_inputs);
+
+    for (i = 0; i < ARRAY_COUNT(sPlayerItemButtons); i++) {
+        if (CHECK_BTN_ALL(cur_inputs.buttons, sPlayerItemButtons[i])) {
+            break;
+        }
+    }
+
+    return i;
+}

patches/input.h

@@ -0,0 +1,35 @@
+#ifndef __INPUT_H__
+#define __INPUT_H__
+
+#ifdef MIPS
+#include "ultra64.h"
+#else
+#include "recomp.h"
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct RecompInputs {
+    u32 buttons;
+    float x;
+    float y;
+} RecompInputs;
+
+
+#ifdef MIPS
+#   define DECLARE_FUNC(type, name, ...) \
+        type name(__VA_ARGS__);
+#else
+#   define DECLARE_FUNC(type, name, ...) \
+        void name(uint8_t* rdram, recomp_context* ctx);
+#endif
+
+DECLARE_FUNC(void, recomp_get_item_inputs, RecompInputs* inputs);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif

patches/patches.h

@@ -0,0 +1,6 @@
+#ifndef __PATCHES_H__
+#define __PATCHES_H__
+
+#include "global.h"
+
+#endif

patches/syms.ld

@@ -1,7 +1,13 @@
 __start = 0x80000000;
+
+/* Dummy addresses that get recompiled into function calls */
+recomp_get_item_inputs = 0x81000000;
+
 /* TODO pull these symbols from the elf file directly */
 Player_PostLimbDrawGameplay = 0x80128BD0;
 Player_DrawImpl = 0x801246F4;
 gRegEditor = 0x801f3f60;
-play_sound = 0x8019f0c8;
+Audio_PlaySfx = 0x8019f0c8;
 gSaveContext = 0x801ef670;
+Interface_SetHudVisibility = 0x8010ef68;
+Player_GetItemOnButton = 0x8012364C;

portultra/events.cpp

@@ -14,6 +14,7 @@
 #include "ultra64.h"
 #include "multilibultra.hpp"
 #include "recomp.h"
+#include "recomp_ui.h"
 #include "rsp.h"
 
 struct SpTaskAction {
@@ -42,17 +43,14 @@ static struct {
         OSMesg msg = (OSMesg)0;
     } sp;
     struct {
-        std::thread thread;
         PTR(OSMesgQueue) mq = NULLPTR;
         OSMesg msg = (OSMesg)0;
     } dp;
     struct {
-        std::thread thread;
         PTR(OSMesgQueue) mq = NULLPTR;
         OSMesg msg = (OSMesg)0;
     } ai;
     struct {
-        std::thread thread;
         PTR(OSMesgQueue) mq = NULLPTR;
         OSMesg msg = (OSMesg)0;
     } si;
@@ -95,6 +93,9 @@ extern "C" void osViSetEvent(RDRAM_ARG PTR(OSMesgQueue) mq_, OSMesg msg, u32 ret
 
 uint64_t total_vis = 0;
 
+
+extern std::atomic_bool exited;
+
 void set_dummy_vi();
 
 void vi_thread_func() {
@@ -106,7 +107,7 @@ void vi_thread_func() {
     
     int remaining_retraces = events_context.vi.retrace_count;
 
-    while (true) {
+    while (!exited) {
         // Determine the next VI time (more accurate than adding 16ms each VI interrupt)
         auto next = Multilibultra::get_start() + (total_vis * 1000000us) / (60 * Multilibultra::get_speed_multiplier());
         //if (next > std::chrono::system_clock::now()) {
@@ -177,6 +178,7 @@ void RT64Init(uint8_t* rom, uint8_t* rdram, Multilibultra::WindowHandle window_h
 void RT64SendDL(uint8_t* rdram, const OSTask* task);
 void RT64UpdateScreen(uint32_t vi_origin);
 void RT64ChangeWindow();
+void RT64Shutdown();
 
 uint8_t dmem[0x1000];
 uint16_t rspReciprocals[512];
@@ -225,10 +227,14 @@ void task_thread_func(uint8_t* rdram, uint8_t* rom, std::atomic_flag* thread_rea
     thread_ready->test_and_set();
     thread_ready->notify_all();
 
-    while (1) {
+    while (true) {
         // Wait until an RSP task has been sent
         events_context.sp_task.wait(nullptr);
 
+        if (exited) {
+            return;
+        }
+
         // Retrieve the task pointer and clear the pending RSP task
         OSTask* task = events_context.sp_task;
         events_context.sp_task.store(nullptr);
@@ -265,7 +271,7 @@ void gfx_thread_func(uint8_t* rdram, uint8_t* rom, std::atomic_flag* thread_read
     thread_ready->test_and_set();
     thread_ready->notify_all();
 
-    while (true) {
+    while (!exited) {
         // Try to pull an action from the queue
         Action action;
         if (events_context.action_queue.wait_dequeue_timed(action, 1ms)) {
@@ -284,6 +290,9 @@ void gfx_thread_func(uint8_t* rdram, uint8_t* rom, std::atomic_flag* thread_read
             }
         }
     }
+    destroy_ui();
+    // TODO restore this call once the RT64 shutdown issue is fixed.
+    // RT64Shutdown();
 }
 
 extern unsigned int VI_STATUS_REG;
@@ -469,3 +478,21 @@ void Multilibultra::init_events(uint8_t* rdram, uint8_t* rom, Multilibultra::Win
 
     events_context.vi.thread = std::thread{ vi_thread_func };
 }
+
+void Multilibultra::join_event_threads() {
+    events_context.sp.gfx_thread.join();
+    events_context.vi.thread.join();
+
+    // Send a dummy RSP task so that the task thread is able to exit it's atomic wait and terminate.
+    OSTask dummy_task{};
+    OSTask* expected = nullptr;
+
+    // Attempt to exchange the task with the dummy task one until it was nullptr, as that indicates the
+    // task thread was ready for a new task.
+    do {
+        expected = nullptr;
+    } while (!events_context.sp_task.compare_exchange_weak(expected, &dummy_task));
+    events_context.sp_task.notify_all();
+
+    events_context.sp.task_thread.join();
+}

portultra/multilibultra.hpp

@@ -125,6 +125,8 @@ struct gfx_callbacks_t {
 void start(WindowHandle window_handle, const audio_callbacks_t& audio_callbacks, const input_callbacks_t& input_callbacks, const gfx_callbacks_t& gfx_callbacks);
 void start_game(int game);
 bool is_game_started();
+void quit();
+void join_event_threads();
 
 } // namespace Multilibultra
 

portultra/scheduler.cpp

@@ -144,6 +144,8 @@ void swap_running_thread(thread_queue_t& running_thread_queue, OSThread*& cur_ru
     }
 }
 
+extern std::atomic_bool exited;
+
 void scheduler_func() {
     thread_queue_t running_thread_queue{};
     OSThread* cur_running_thread = nullptr;
@@ -169,8 +171,13 @@ void scheduler_func() {
         // Handle threads that have changed priority
         handle_thread_reprioritization(running_thread_queue);
 
-        // Determine which thread to run, stopping the current running thread if necessary
-        swap_running_thread(running_thread_queue, cur_running_thread);
+        if (!exited) {
+            // Determine which thread to run, stopping the current running thread if necessary
+            swap_running_thread(running_thread_queue, cur_running_thread);
+        }
+        else {
+            return;
+        }
 
         std::this_thread::yield();
         if (old_running_thread != cur_running_thread && old_running_thread && cur_running_thread) {

portultra/threads.cpp

@@ -161,12 +161,6 @@ extern "C" void osStartThread(RDRAM_ARG PTR(OSThread) t_) {
     } else {
         Multilibultra::schedule_running_thread(t);
     }
-
-    // The main thread "becomes" the first thread started, so join on it and exit after it completes.
-    if (is_main_thread) {
-        t->context->host_thread.join();
-        std::exit(EXIT_SUCCESS);
-    }
 }
 
 extern "C" void osCreateThread(RDRAM_ARG PTR(OSThread) t_, OSId id, PTR(thread_func_t) entrypoint, PTR(void) arg, PTR(void) sp, OSPri pri) {

portultra/timer.cpp

@@ -121,7 +121,8 @@ void timer_thread(RDRAM_ARG1) {
             active_timers.insert(cur_timer_);
             // Process the new action
             process_timer_action(cur_action);
-        } else {
+        }
+        else {
             // Waiting for the timer completed, so send the timer's message to its message queue
             osSendMesg(PASS_RDRAM cur_timer->mq, cur_timer->msg, OS_MESG_NOBLOCK);
             // If the timer has a specified interval then reload it with that value
@@ -135,6 +136,7 @@ void timer_thread(RDRAM_ARG1) {
 
 void Multilibultra::init_timers(RDRAM_ARG1) {
     timer_context.thread = std::thread{ timer_thread, PASS_RDRAM1 };
+    timer_context.thread.detach();
 }
 
 uint32_t Multilibultra::get_speed_multiplier() {

src/cont.cpp

@@ -1,5 +1,5 @@
 #include "../portultra/multilibultra.hpp"
-#include "recomp.h"
+#include "recomp_helpers.h"
 
 static Multilibultra::input_callbacks_t input_callbacks;
 
@@ -10,8 +10,8 @@ void set_input_callbacks(const Multilibultra::input_callbacks_t& callbacks) {
 static int max_controllers = 0;
 
 extern "C" void osContInit_recomp(uint8_t* rdram, recomp_context* ctx) {
-    gpr bitpattern = ctx->r5;
-    gpr status = ctx->r6;
+    PTR(void) bitpattern = _arg<1, PTR(void)>(rdram, ctx);
+    PTR(void) status = _arg<2, PTR(void)>(rdram, ctx);
 
     // Set bit 0 to indicate that controller 0 is present
     MEM_B(0, bitpattern) = 0x01;
@@ -29,22 +29,15 @@ extern "C" void osContInit_recomp(uint8_t* rdram, recomp_context* ctx) {
         MEM_B(4 * controller + 3, status) = 0x80 >> 4; // errno: CONT_NO_RESPONSE_ERROR >> 4
     }
 
-    ctx->r2 = 0;
+    _return<s32>(ctx, 0);
 }
 
 extern "C" void osContStartReadData_recomp(uint8_t* rdram, recomp_context* ctx) {
     Multilibultra::send_si_message();
 }
 
-struct OSContPad {
-    u16 button;
-    s8 stick_x;		/* -80 <= stick_x <= 80 */
-    s8 stick_y;		/* -80 <= stick_y <= 80 */
-    u8 errno_;
-};
-
 extern "C" void osContGetReadData_recomp(uint8_t* rdram, recomp_context* ctx) {
-    int32_t pad = (int32_t)ctx->r4;
+    PTR(void) pad = _arg<0, PTR(void)>(rdram, ctx);
 
     uint16_t buttons = 0;
     float x = 0.0f;
@@ -74,7 +67,7 @@ extern "C" void osContStartQuery_recomp(uint8_t * rdram, recomp_context * ctx) {
 }
 
 extern "C" void osContGetQuery_recomp(uint8_t * rdram, recomp_context * ctx) {
-    gpr status = ctx->r4;
+    PTR(void) status = _arg<0, PTR(void)>(rdram, ctx);
 
     // Mark controller 0 as present
     MEM_H(0, status) = 0x0005; // type: CONT_TYPE_NORMAL (from joybus)
@@ -89,8 +82,8 @@ extern "C" void osContGetQuery_recomp(uint8_t * rdram, recomp_context * ctx) {
 }
 
 extern "C" void osContSetCh_recomp(uint8_t* rdram, recomp_context* ctx) {
-    max_controllers = std::min((unsigned int)ctx->r4, 4u);
-    ctx->r2 = 0;
+    max_controllers = std::min(_arg<0, u8>(rdram, ctx), u8(4));
+    _return<s32>(ctx, 0);
 }
 
 extern "C" void __osMotorAccess_recomp(uint8_t* rdram, recomp_context* ctx) {
@@ -108,3 +101,16 @@ extern "C" void osMotorStart_recomp(uint8_t* rdram, recomp_context* ctx) {
 extern "C" void osMotorStop_recomp(uint8_t* rdram, recomp_context* ctx) {
     ;
 }
+
+#include "../patches/input.h"
+
+extern "C" void recomp_get_item_inputs(uint8_t* rdram, recomp_context* ctx) {
+    RecompInputs* inputs = _arg<0, RecompInputs*>(rdram, ctx);
+
+    if (input_callbacks.get_input) {
+        u16 buttons;
+        input_callbacks.get_input(&buttons, &inputs->x, &inputs->y);
+        // TODO remap the inputs for items here
+        inputs->buttons = buttons;
+    }
+}

src/main/main.cpp

@@ -99,7 +99,7 @@ std::vector<GameControllerButtonMapping> controller_button_map{
 
 std::vector<SDL_JoystickID> controllers{};
 
-int sdl_event_filter(void* userdata, SDL_Event* event) {
+bool sdl_event_filter(void* userdata, SDL_Event* event) {
     switch (event->type) {
     //case SDL_EventType::SDL_KEYUP:
     //case SDL_EventType::SDL_KEYDOWN:
@@ -138,13 +138,13 @@ int sdl_event_filter(void* userdata, SDL_Event* event) {
         }
         break;
     case SDL_EventType::SDL_QUIT:
-        std::quick_exit(EXIT_SUCCESS);
-        break;
+        Multilibultra::quit();
+        return true;
     default:
         queue_event(*event);
         break;
     }
-    return 1;
+    return false;
 }
 
 Multilibultra::gfx_callbacks_t::gfx_data_t create_gfx() {
@@ -185,8 +185,9 @@ void update_gfx(void*) {
     constexpr int max_events_per_frame = 16;
     SDL_Event cur_event;
     int i = 0;
-    while (i++ < max_events_per_frame && SDL_PollEvent(&cur_event)) {
-        sdl_event_filter(nullptr, &cur_event);
+    static bool exited = false;
+    while (i++ < max_events_per_frame && SDL_PollEvent(&cur_event) && !exited) {
+        exited = sdl_event_filter(nullptr, &cur_event);
     }
 }
 

src/recomp.cpp

@@ -122,6 +122,28 @@ extern "C" void unload_overlays(int32_t ram_addr, uint32_t size);
 std::unique_ptr<uint8_t[]> rdram_buffer;
 recomp_context context{};
 
+void read_patch_data(uint8_t* rdram, gpr patch_data_address) {
+    const char patches_data_file_path[] = "patches/patches.bin";
+    std::ifstream patches_data_file{ patches_data_file_path, std::ios::binary };
+
+    if (!patches_data_file) {
+        fprintf(stderr, "Failed to open patches data file: %s\n", patches_data_file_path);
+        exit(EXIT_FAILURE);
+    }
+
+    patches_data_file.seekg(0, std::ios::end);
+    size_t patches_data_size = patches_data_file.tellg();
+    patches_data_file.seekg(0, std::ios::beg);
+
+    std::unique_ptr<uint8_t[]> patches_data = std::make_unique<uint8_t[]>(patches_data_size);
+
+    patches_data_file.read(reinterpret_cast<char*>(patches_data.get()), patches_data_size);
+
+    for (size_t i = 0; i < patches_data_size; i++) {
+        MEM_B(i, patch_data_address) = patches_data[i];
+    }
+}
+
 EXPORT extern "C" void init() {
     {
         std::ifstream rom_file{ get_rom_name(), std::ios::binary };
@@ -160,6 +182,9 @@ EXPORT extern "C" void init() {
     // Initial 1MB DMA (rom address 0x1000 = physical address 0x10001000)
     do_rom_read(rdram_buffer.get(), entrypoint, 0x10001000, 0x100000);
 
+    // Read in any extra data from patches
+    read_patch_data(rdram_buffer.get(), (gpr)(s32)0x80800100);
+
     // Set up stack pointer
     context.r29 = 0xFFFFFFFF803FFFF0u;
 
@@ -197,6 +222,15 @@ bool Multilibultra::is_game_started() {
 void set_audio_callbacks(const Multilibultra::audio_callbacks_t& callbacks);
 void set_input_callbacks(const Multilibultra::input_callbacks_t& callback);
 
+std::atomic_bool exited = false;
+
+void Multilibultra::quit() {
+    exited.store(true);
+    int desired = -1;
+    game_started.compare_exchange_strong(desired, -2);
+    game_started.notify_all();
+}
+
 void Multilibultra::start(WindowHandle window_handle, const audio_callbacks_t& audio_callbacks, const input_callbacks_t& input_callbacks, const gfx_callbacks_t& gfx_callbacks_) {
     set_audio_callbacks(audio_callbacks);
     set_input_callbacks(input_callbacks);
@@ -231,16 +265,20 @@ void Multilibultra::start(WindowHandle window_handle, const audio_callbacks_t& a
             case 0:
                 recomp_entrypoint(rdram_buffer.get(), &context);
                 break;
+            case -2:
+                break;
         }
         
         debug_printf("[Recomp] Quitting\n");
     }, window_handle};
 
-    while (true) {
+    while (!exited) {
         using namespace std::chrono_literals;
         std::this_thread::sleep_for(1ms);
         if (gfx_callbacks.update_gfx != nullptr) {
             gfx_callbacks.update_gfx(gfx_data);
         }
     }
+    game_thread.join();
+    Multilibultra::join_event_threads();
 }

src/rt64_layer.cpp

@@ -130,3 +130,7 @@ void RT64UpdateScreen(uint32_t vi_origin) {
 void RT64ChangeWindow() {
     ChangeWindow();
 }
+
+void RT64Shutdown() {
+    PluginShutdown();
+}

src/ui/ui_renderer.cpp

@@ -587,6 +587,10 @@ struct {
         Rml::Context* context;
         std::unique_ptr<Rml::EventListenerInstancer> event_listener_instancer;
 
+        void unload() {
+            render_interface.reset();
+        }
+
         void swap_document(Menu menu) {
             if (current_document != nullptr) {
                 current_document->Hide();
@@ -752,3 +756,8 @@ void set_rt64_hooks() {
 void set_current_menu(Menu menu) {
     open_menu.store(menu);
 }
+
+void destroy_ui() {
+    Rml::Shutdown();
+    UIContext.rml.unload();
+}

us.rev1.toml

@@ -10,10 +10,15 @@ relocatable_sections_path = "overlays.us.rev1.txt"
 [patches]
 stubs = [
     # Stub out unused functions that directly manipulate RCP status.
-    "func_80084940",
-    "func_80084968",
+    "RcpUtils_PrintRegisterStatus",
+    "RcpUtils_Reset",
     # Stub out an unnecessary function that accesses kseg1 addresses.
-    "func_800818F4"
+    "CIC6105_Init"
+]
+
+ignored = [
+    # Not actually a function
+    "D_80186028"
 ]
 
 # Hooks
@@ -25,23 +30,14 @@ args = ["u32", "u32", "u32"]
 
 # Function hooks for overlay loading.
 [[patches.hook]]
-func = "Idle_InitCodeAndMemory"
-calls = "load_overlays"
-args = ["a2", "a1", "a3"]
+func = "Main_Init"
+text = "    load_overlays((uint32_t)ctx->r6, (uint32_t)ctx->r5, (uint32_t)ctx->r7);"
 after_vram = 0x800802A4
 
 [[patches.hook]]
-func = "Load2_LoadOverlay"
-calls = "load_overlays"
-# args = [
-#     "a0", # $a0 contains rom start
-#     {operation = "load", type = "u32", base = "sp", offset = 0x10}, # sp + 10 contains the ram address
-#     {operation = "subtract", arguments = ["a1", "a0"]} # Calculate size from rom end - rom start
-# ]
-args = ["a1", "a0", "a2"]
-# This vram address is an instruction in a delay slot. In that case, the recompiler will emit the
-# hook call after this instruction is run and before the function is called.
-after_vram = 0x80085048
+func = "Overlay_Load"
+text = "    load_overlays((uint32_t)ctx->r4, MEM_W(0x10, ctx->r29), (uint32_t)(ctx->r5 - ctx->r4));"
+# No after_vram means this will be placed at the start of the function
 
 # Single-instruction patches