Integrated RT64 (not included in repo), sign extended addresses

.gitignore

@@ -38,4 +38,7 @@ bld/
 [Ll]ogs/
 
 # Visual Studio 2015/2017 cache/options directory
-.vs/
+.vs/
+
+# RT64 (since it's not public yet)
+test/RT64

recomp.h

@@ -24,25 +24,35 @@ typedef uint64_t gpr;
     ((gpr)(int32_t)((a) - (b)))
 
 #define MEM_W(offset, reg) \
-    (*(int32_t*)(rdram + ((((reg) + (offset))) & 0x3FFFFFF)))
+    (*(int32_t*)(rdram + ((((reg) + (offset))) - 0xFFFFFFFF80000000)))
+    //(*(int32_t*)(rdram + ((((reg) + (offset))) & 0x3FFFFFF)))
 
 #define MEM_H(offset, reg) \
-    (*(int16_t*)(rdram + ((((reg) + (offset)) ^ 2) & 0x3FFFFFF)))
+    (*(int16_t*)(rdram + ((((reg) + (offset)) ^ 2) - 0xFFFFFFFF80000000)))
+    //(*(int16_t*)(rdram + ((((reg) + (offset)) ^ 2) & 0x3FFFFFF)))
 
 #define MEM_B(offset, reg) \
-    (*(int8_t*)(rdram + ((((reg) + (offset)) ^ 3) & 0x3FFFFFF)))
+    (*(int8_t*)(rdram + ((((reg) + (offset)) ^ 3) - 0xFFFFFFFF80000000)))
+    //(*(int8_t*)(rdram + ((((reg) + (offset)) ^ 3) & 0x3FFFFFF)))
 
 #define MEM_HU(offset, reg) \
-    (*(uint16_t*)(rdram + ((((reg) + (offset)) ^ 2) & 0x3FFFFFF)))
+    (*(uint16_t*)(rdram + ((((reg) + (offset)) ^ 2) - 0xFFFFFFFF80000000)))
+    //(*(uint16_t*)(rdram + ((((reg) + (offset)) ^ 2) & 0x3FFFFFF)))
 
 #define MEM_BU(offset, reg) \
-    (*(uint8_t*)(rdram + ((((reg) + (offset)) ^ 3) & 0x3FFFFFF)))
+    (*(uint8_t*)(rdram + ((((reg) + (offset)) ^ 3) - 0xFFFFFFFF80000000)))
+    //(*(uint8_t*)(rdram + ((((reg) + (offset)) ^ 3) & 0x3FFFFFF)))
 
 #define SD(val, offset, reg) { \
-    *(uint32_t*)(rdram + ((((reg) + (offset) + 4)) & 0x3FFFFFF)) = (uint32_t)((val) >> 32); \
+    *(uint32_t*)(rdram + ((((reg) + (offset) + 4)) - 0xFFFFFFFF80000000)) = (uint32_t)((val) >> 0); \
-    *(uint32_t*)(rdram + ((((reg) + (offset) + 0)) & 0x3FFFFFF)) = (uint32_t)((val) >> 0); \
+    *(uint32_t*)(rdram + ((((reg) + (offset) + 0)) - 0xFFFFFFFF80000000)) = (uint32_t)((val) >> 32); \
 }
 
+//#define SD(val, offset, reg) { \
+//    *(uint32_t*)(rdram + ((((reg) + (offset) + 4)) & 0x3FFFFFF)) = (uint32_t)((val) >> 32); \
+//    *(uint32_t*)(rdram + ((((reg) + (offset) + 0)) & 0x3FFFFFF)) = (uint32_t)((val) >> 0); \
+//}
+
 static inline uint64_t load_doubleword(uint8_t* rdram, gpr reg, gpr offset) {
     uint64_t ret = 0;
     uint64_t lo = (uint64_t)(uint32_t)MEM_W(reg, offset + 4);
@@ -56,7 +66,8 @@ static inline uint64_t load_doubleword(uint8_t* rdram, gpr reg, gpr offset) {
 
 // TODO proper lwl/lwr/swl/swr
 #define MEM_WL(offset, reg) \
-    (*(int32_t*)(rdram + ((((reg) + (offset))) & 0x3FFFFFF)))
+    (*(int32_t*)(rdram + ((((reg) + (offset))) - 0xFFFFFFFF80000000)))
+    //(*(int32_t*)(rdram + ((((reg) + (offset))) & 0x3FFFFFF)))
 
 #define S32(val) \
     ((int32_t)(val))
@@ -103,6 +114,8 @@ static inline uint64_t load_doubleword(uint8_t* rdram, gpr reg, gpr offset) {
 #define NAN_CHECK(val) \
     assert(val == val)
 
+//#define NAN_CHECK(val)
+
 typedef union {
     double d;
     struct {

src/recompilation.cpp

@@ -104,7 +104,7 @@ bool process_instruction(const RecompPort::Context& context, const RecompPort::F
         break;
     // Arithmetic
     case InstrId::cpu_lui:
-        print_line("{}{} = {:#X} << 16", ctx_gpr_prefix(rt), rt, imm);
+        print_line("{}{} = S32({:#X} << 16)", ctx_gpr_prefix(rt), rt, imm);
         break;
     case InstrId::cpu_addu:
         {

test/RecompTest.vcxproj

@@ -76,7 +76,7 @@
       <WarningLevel>Level3</WarningLevel>
       <DebugInformationFormat>ProgramDatabase</DebugInformationFormat>
       <Optimization>Disabled</Optimization>
-      <AdditionalIncludeDirectories>$(ProjectDir)..;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
+      <AdditionalIncludeDirectories>$(ProjectDir)..;$(ProjectDir)thirdparty;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
       <LanguageStandard>stdcpp20</LanguageStandard>
     </ClCompile>
     <Link>
@@ -91,7 +91,7 @@
       <RuntimeLibrary>MultiThreadedDLL</RuntimeLibrary>
       <WarningLevel>Level3</WarningLevel>
       <DebugInformationFormat>ProgramDatabase</DebugInformationFormat>
-      <AdditionalIncludeDirectories>$(ProjectDir)..;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
+      <AdditionalIncludeDirectories>$(ProjectDir)..;$(ProjectDir)thirdparty;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
       <LanguageStandard>stdcpp20</LanguageStandard>
     </ClCompile>
     <Link>
@@ -104,21 +104,28 @@
   </ItemDefinitionGroup>
   <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">
     <ClCompile>
-      <AdditionalIncludeDirectories>$(ProjectDir)..;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
+      <AdditionalIncludeDirectories>$(ProjectDir)..;$(ProjectDir)thirdparty;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
       <LanguageStandard>stdcpp20</LanguageStandard>
     </ClCompile>
     <Link>
       <SubSystem>Console</SubSystem>
+      <AdditionalDependencies>$(ProjectDir)RT64\Debug\RT64.lib;%(AdditionalDependencies)</AdditionalDependencies>
     </Link>
+    <PostBuildEvent>
+      <Command>XCOPY "$(ProjectDir)RT64\$(Configuration)\*" "$(TargetDir)" /S /Y</Command>
+    </PostBuildEvent>
   </ItemDefinitionGroup>
   <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Release|x64'">
     <ClCompile>
-      <AdditionalIncludeDirectories>$(ProjectDir)..;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
+      <AdditionalIncludeDirectories>$(ProjectDir)..;$(ProjectDir)thirdparty;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
       <LanguageStandard>stdcpp20</LanguageStandard>
     </ClCompile>
     <Link>
       <SubSystem>Console</SubSystem>
     </Link>
+    <PostBuildEvent>
+      <Command>XCOPY "$(ProjectDir)RT64\$(Configuration)\*" "$(TargetDir)" /S /Y</Command>
+    </PostBuildEvent>
   </ItemDefinitionGroup>
   <ItemGroup>
     <ClCompile Include="funcs\adjust_texture_table.c" />
@@ -2005,6 +2012,7 @@
     <ClCompile Include="portultra\task_pthreads.cpp" />
     <ClCompile Include="portultra\task_win32.cpp" />
     <ClCompile Include="portultra\threads.cpp" />
+    <ClCompile Include="RT64\rt64_layer.cpp" />
     <ClCompile Include="src\ai.cpp" />
     <ClCompile Include="src\cont.cpp" />
     <ClCompile Include="src\dp.cpp" />
@@ -2020,7 +2028,11 @@
     <ClInclude Include="..\recomp.h" />
     <ClInclude Include="portultra\multilibultra.hpp" />
     <ClInclude Include="portultra\platform_specific.h" />
+    <ClInclude Include="RT64\rt64_layer.h" />
     <ClInclude Include="portultra\ultra64.h" />
+    <ClInclude Include="thirdparty\blockingconcurrentqueue.h" />
+    <ClInclude Include="thirdparty\concurrentqueue.h" />
+    <ClInclude Include="thirdparty\lightweightsemaphore.h" />
   </ItemGroup>
   <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
   <ImportGroup Label="ExtensionTargets">

test/RecompTest.vcxproj.filters

@@ -5700,6 +5700,9 @@
     <ClCompile Include="portultra\events.cpp">
       <Filter>Source Files</Filter>
     </ClCompile>
+    <ClCompile Include="RT64\rt64_layer.cpp">
+      <Filter>Source Files</Filter>
+    </ClCompile>
   </ItemGroup>
   <ItemGroup>
     <ClInclude Include="portultra\platform_specific.h">
@@ -5714,5 +5717,17 @@
     <ClInclude Include="..\recomp.h">
       <Filter>Header Files</Filter>
     </ClInclude>
+    <ClInclude Include="RT64\rt64_layer.h">
+      <Filter>Source Files</Filter>
+    </ClInclude>
+    <ClInclude Include="thirdparty\blockingconcurrentqueue.h">
+      <Filter>Header Files</Filter>
+    </ClInclude>
+    <ClInclude Include="thirdparty\concurrentqueue.h">
+      <Filter>Header Files</Filter>
+    </ClInclude>
+    <ClInclude Include="thirdparty\lightweightsemaphore.h">
+      <Filter>Header Files</Filter>
+    </ClInclude>
   </ItemGroup>
 </Project>

test/portultra/events.cpp

@@ -2,11 +2,25 @@
 #include <atomic>
 #include <chrono>
 #include <cinttypes>
+#include <variant>
+
+#include <Windows.h>
+#include "blockingconcurrentqueue.h"
 
 #include "ultra64.h"
 #include "multilibultra.hpp"
 #include "recomp.h"
 
+struct SpTaskAction {
+    OSTask task;
+};
+
+struct SwapBuffersAction {
+    int32_t origin;
+};
+
+using Action = std::variant<SpTaskAction, SwapBuffersAction>;
+
 static struct {
     struct {
         std::thread thread;
@@ -18,8 +32,6 @@ static struct {
         std::thread thread;
         PTR(OSMesgQueue) mq = NULLPTR;
         OSMesg msg = (OSMesg)0;
-        OSTask task;
-        std::atomic_flag task_queued;
     } sp;
     struct {
         std::thread thread;
@@ -35,12 +47,12 @@ static struct {
         std::thread thread;
         PTR(OSMesgQueue) mq = NULLPTR;
         OSMesg msg = (OSMesg)0;
-        std::atomic_flag read_queued;
     } si;
     // The same message queue may be used for multiple events, so share a mutex for all of them
     std::mutex message_mutex;
     uint8_t* rdram;
     std::chrono::system_clock::time_point start;
+    moodycamel::BlockingConcurrentQueue<Action> action_queue{};
 } events_context{};
 
 extern "C" void osSetEventMesg(RDRAM_ARG OSEvent event_id, PTR(OSMesgQueue) mq_, OSMesg msg) {
@@ -73,8 +85,10 @@ extern "C" void osViSetEvent(RDRAM_ARG PTR(OSMesgQueue) mq_, OSMesg msg, u32 ret
     events_context.vi.retrace_count = retrace_count;
 }
 
+constexpr uint32_t speed_multiplier = 10;
+
 // N64 CPU counter ticks per millisecond
-constexpr uint32_t counter_per_ms = 46'875;
+constexpr uint32_t counter_per_ms = 46'875 * speed_multiplier;
 
 uint64_t duration_to_count(std::chrono::system_clock::duration duration) {
     uint64_t delta_micros = std::chrono::duration_cast<std::chrono::microseconds>(duration).count();
@@ -107,7 +121,7 @@ void vi_thread_func() {
 
     while (true) {
         // Determine the next VI time (more accurate than adding 16ms each VI interrupt)
-        auto next = events_context.start + (total_vis * 1000000us) / 60;
+        auto next = events_context.start + (total_vis * 1000000us) / (60 * speed_multiplier);
         //if (next > std::chrono::system_clock::now()) {
         //    printf("Sleeping for %" PRIu64 " us to get from %" PRIu64 " us to %" PRIu64 " us \n",
         //        (next - std::chrono::system_clock::now()) / 1us,
@@ -118,22 +132,29 @@ void vi_thread_func() {
         //}
         std::this_thread::sleep_until(next);
         // Calculate how many VIs have passed
-        uint64_t new_total_vis = ((std::chrono::system_clock::now() - events_context.start) * 60 / 1000ms) + 1;
+        uint64_t new_total_vis = ((std::chrono::system_clock::now() - events_context.start) * (60 * speed_multiplier) / 1000ms) + 1;
         if (new_total_vis > total_vis + 1) {
-            printf("Skipped % " PRId64 " frames in VI interupt thread!\n", new_total_vis - total_vis - 1);
+            //printf("Skipped % " PRId64 " frames in VI interupt thread!\n", new_total_vis - total_vis - 1);
         }
         total_vis = new_total_vis;
 
         remaining_retraces--;
 
-        if (remaining_retraces == 0) {
+        {
             std::lock_guard lock{ events_context.message_mutex };
-            remaining_retraces = events_context.vi.retrace_count;
             uint8_t* rdram = events_context.rdram;
+            if (remaining_retraces == 0) {
+                remaining_retraces = events_context.vi.retrace_count;
 
-            if (events_context.vi.mq != NULLPTR) {
+                if (events_context.vi.mq != NULLPTR) {
-                if (osSendMesg(PASS_RDRAM events_context.vi.mq, events_context.vi.msg, OS_MESG_NOBLOCK) == -1) {
+                    if (osSendMesg(PASS_RDRAM events_context.vi.mq, events_context.vi.msg, OS_MESG_NOBLOCK) == -1) {
-                    //printf("Game skipped a VI frame!\n");
+                        //printf("Game skipped a VI frame!\n");
+                    }
+                }
+            }
+            if (events_context.ai.mq != NULLPTR) {
+                if (osSendMesg(PASS_RDRAM events_context.ai.mq, events_context.ai.msg, OS_MESG_NOBLOCK) == -1) {
+                    //printf("Game skipped a AI frame!\n");
                 }
             }
         }
@@ -152,41 +173,49 @@ void dp_complete() {
     osSendMesg(PASS_RDRAM events_context.dp.mq, events_context.dp.msg, OS_MESG_NOBLOCK);
 }
 
-void gfx_thread_func() {
+void RT64Init(uint8_t* rom, uint8_t* rdram);
+void RT64SendDL(uint8_t* rdram, const OSTask* task);
+void RT64UpdateScreen(uint32_t vi_origin);
+void RT64PumpEvents();
+
+void gfx_thread_func(uint8_t* rdram, uint8_t* rom) {
+    using namespace std::chrono_literals;
+    RT64Init(rom, rdram);
+
     while (true) {
-        // Wait for a sp task to be queued
+        // Try to pull an action from the queue
-        events_context.sp.task_queued.wait(false);
+        Action action;
-
+        if (events_context.action_queue.wait_dequeue_timed(action, 1ms)) {
-        // Grab the task and inform the game that it's free to queue up a new task
+            // Determine the action type and act on it
-        OSTask current_task = events_context.sp.task;
+            if (const auto* task_action = std::get_if<SpTaskAction>(&action)) {
-        events_context.sp.task_queued.clear();
+                if (task_action->task.t.type == M_GFXTASK) {
-        events_context.sp.task_queued.notify_all();
+                    // (TODO let RT64 do this) Tell the game that the RSP and RDP tasks are complete
-
+                    RT64SendDL(rdram, &task_action->task);
-        // Process the task
+                    sp_complete();
-        if (current_task.t.type = M_GFXTASK) {
+                    dp_complete();
-            // TODO interface with RT64 here
+                } else if (task_action->task.t.type == M_AUDTASK) {
-            
+                    sp_complete();
-            // (TODO let RT64 do this) Tell the game that the RSP and RDP tasks are complete
+                } else {
-            sp_complete();
+                    fprintf(stderr, "Unknown task type: %" PRIu32 "\n", task_action->task.t.type);
-            dp_complete();
+                    std::exit(EXIT_FAILURE);
-        } else if (current_task.t.type == M_AUDTASK) {
+                }
-            sp_complete();
+            } else if (const auto* swap_action = std::get_if<SwapBuffersAction>(&action)) {
-        } else {
+                RT64UpdateScreen(swap_action->origin);
-            fprintf(stderr, "Unknown task type: %" PRIu32 "\n", current_task.t.type);
+            }
-            std::exit(EXIT_FAILURE);
         }
+
+        // Handle events
+        RT64PumpEvents();
     }
 }
 
+extern "C" void osViSwapBuffer(RDRAM_ARG PTR(void) frameBufPtr) {
+    events_context.action_queue.enqueue(SwapBuffersAction{ frameBufPtr + 640 });
+}
+
 void Multilibultra::submit_rsp_task(RDRAM_ARG PTR(OSTask) task_) {
     OSTask* task = TO_PTR(OSTask, task_);
-    // Wait for the sp thread clear the old task
+    events_context.action_queue.enqueue(SpTaskAction{ *task });
-    events_context.sp.task_queued.wait(true);
-    // Make a full copy of the task instead of just recording a pointer to it, since that's what osSpTaskLoad does
-    events_context.sp.task = *task;
-
-    events_context.sp.task_queued.test_and_set();
-    events_context.sp.task_queued.notify_all();
 }
 
 void Multilibultra::send_si_message() {
@@ -194,8 +223,8 @@ void Multilibultra::send_si_message() {
     osSendMesg(PASS_RDRAM events_context.si.mq, events_context.si.msg, OS_MESG_NOBLOCK);
 }
 
-void Multilibultra::init_events(uint8_t* rdram) {
+void Multilibultra::init_events(uint8_t* rdram, uint8_t* rom) {
     events_context.rdram = rdram;
     events_context.vi.thread = std::thread{ vi_thread_func };
-    events_context.sp.thread = std::thread{ gfx_thread_func };
+    events_context.sp.thread = std::thread{ gfx_thread_func, rdram, rom };
-}
+}

test/portultra/init.cpp

@@ -1,9 +1,9 @@
 #include "ultra64.h"
 #include "multilibultra.hpp"
 
-void Multilibultra::preinit(uint8_t* rdram) {
+void Multilibultra::preinit(uint8_t* rdram, uint8_t* rom) {
     Multilibultra::set_main_thread();
-    Multilibultra::init_events(rdram);
+    Multilibultra::init_events(rdram, rom);
 }
 
 extern "C" void osInitialize() {

test/portultra/mesgqueue.cpp

@@ -167,7 +167,7 @@ extern "C" s32 osRecvMesg(RDRAM_ARG PTR(OSMesgQueue) mq_, PTR(OSMesg) msg_, s32
         }
     }
 
-    if (msg != nullptr) {
+    if (msg_ != NULLPTR) {
         *msg = TO_PTR(OSMesg, mq->msg)[mq->first];
     }
     

test/portultra/multilibultra.hpp

@@ -17,10 +17,10 @@ struct UltraThreadContext {
 
 namespace Multilibultra {
 
-void preinit(uint8_t* rdram);
+void preinit(uint8_t* rdram, uint8_t* rom);
 void native_init();
 void init_scheduler();
-void init_events(uint8_t* rdram);
+void init_events(uint8_t* rdram, uint8_t* rom);
 void native_thread_init(OSThread *t);
 void set_self_paused(RDRAM_ARG1);
 void wait_for_resumed(RDRAM_ARG1);

test/portultra/threads.cpp

@@ -29,7 +29,7 @@ int main(int argc, char** argv) {
 #endif
 
 #if 1
-void run_thread_function(uint8_t* rdram, uint32_t addr, uint32_t sp, uint32_t arg);
+void run_thread_function(uint8_t* rdram, uint64_t addr, uint64_t sp, uint64_t arg);
 #else
 #define run_thread_function(func, sp, arg) func(arg)
 #endif
@@ -66,7 +66,7 @@ extern "C" void osStartThread(RDRAM_ARG PTR(OSThread) t_) {
     OSThread* t = TO_PTR(OSThread, t_);
     debug_printf("[os] Start Thread %d\n", t->id);
 
-    // Wait until the thread is initialized to indicate that it's task_queued to be started.
+    // Wait until the thread is initialized to indicate that it's action_queued to be started.
     t->context->initialized.wait(false);
 
     debug_printf("[os] Thread %d is ready to be started\n", t->id);

test/portultra/ultra64.h

@@ -25,12 +25,12 @@ typedef uint16_t u16;
 typedef int8_t s8;
 typedef uint8_t u8;
 
-#define PTR(x) uint32_t
+#define PTR(x) int32_t
 #define RDRAM_ARG uint8_t *rdram, 
 #define RDRAM_ARG1 uint8_t *rdram
 #define PASS_RDRAM rdram, 
 #define PASS_RDRAM1 rdram
-#define TO_PTR(type, var) ((type*)(&rdram[var & 0x3FFFFFF]))
+#define TO_PTR(type, var) ((type*)(&rdram[(uint64_t)var - 0xFFFFFFFF80000000]))
 #ifdef __cplusplus
 #define NULLPTR (PTR(void))0
 #endif
@@ -89,7 +89,7 @@ typedef struct OSThread_t {
     OSId id;
     int32_t pad3;
     UltraThreadContext* context; // An actual pointer regardless of platform
-    uint32_t sp;
+    int32_t sp;
 } OSThread;
 
 typedef u32 OSEvent;
@@ -161,6 +161,7 @@ s32 osJamMesg(RDRAM_ARG PTR(OSMesgQueue), OSMesg, s32);
 s32 osRecvMesg(RDRAM_ARG PTR(OSMesgQueue), PTR(OSMesg), s32);
 void osSetEventMesg(RDRAM_ARG OSEvent, PTR(OSMesgQueue), OSMesg);
 void osViSetEvent(RDRAM_ARG PTR(OSMesgQueue), OSMesg, u32);
+void osViSwapBuffer(RDRAM_ARG PTR(void) frameBufPtr);
 u32 osGetCount();
 OSTime osGetTime();
 

test/src/ai.cpp

@@ -1,17 +1,20 @@
 #include "recomp.h"
 
 extern "C" void osAiSetFrequency_recomp(uint8_t* restrict rdram, recomp_context* restrict ctx) {
-    ;
+    ctx->r2 = ctx->r4;
 }
 
+static uint32_t ai_length = 0;
+
 extern "C" void osAiSetNextBuffer_recomp(uint8_t* restrict rdram, recomp_context* restrict ctx) {
-    ;
+    ai_length = (uint32_t)ctx->r5;
+    ctx->r2 = 0;
 }
 
 extern "C" void osAiGetLength_recomp(uint8_t* restrict rdram, recomp_context* restrict ctx) {
-    ;
+    ctx->r2 = ai_length;
 }
 
 extern "C" void osAiGetStatus_recomp(uint8_t* restrict rdram, recomp_context* restrict ctx) {
-    ;
+    ctx->r2 = 0x80000000;
-}
+}

test/src/cont.cpp

@@ -9,8 +9,24 @@ extern "C" void osContStartReadData_recomp(uint8_t* restrict rdram, recomp_conte
     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* restrict rdram, recomp_context* restrict ctx) {
-    ;
+    int32_t pad = (uint32_t)ctx->r4;
+
+    // button
+    MEM_H(0, pad) = 0;
+    // stick_x
+    MEM_B(2, pad) = 0;
+    // stick_y
+    MEM_B(3, pad) = 0;
+    // errno
+    MEM_B(4, pad) = 0;
 }
 
 extern "C" void osMotorInit_recomp(uint8_t* restrict rdram, recomp_context* restrict ctx) {

test/src/pi.cpp

@@ -16,7 +16,7 @@ extern "C" void osCreatePiManager_recomp(uint8_t* restrict rdram, recomp_context
 
 constexpr uint32_t rom_base = 0xB0000000;
 
-void do_rom_read(uint8_t* rdram, uint32_t ram_address, uint32_t dev_address, size_t num_bytes) {
+void do_rom_read(uint8_t* rdram, int32_t ram_address, uint32_t dev_address, size_t num_bytes) {
     // TODO use word copies when possible
     uint8_t* rom_addr = rom.get() + (dev_address | rom_base) - rom_base;
     for (size_t i = 0; i < num_bytes; i++) {
@@ -30,7 +30,7 @@ extern "C" void osPiStartDma_recomp(uint8_t* restrict rdram, recomp_context* res
     uint32_t pri = ctx->r5;
     uint32_t direction = ctx->r6;
     uint32_t devAddr = ctx->r7;
-    uint32_t dramAddr = MEM_W(0x10, ctx->r29);
+    int32_t dramAddr = MEM_W(0x10, ctx->r29);
     uint32_t size = MEM_W(0x14, ctx->r29);
     uint32_t mq_ = MEM_W(0x18, ctx->r29);
     OSMesgQueue* mq = TO_PTR(OSMesgQueue, mq_);

test/src/recomp.cpp

@@ -35,8 +35,8 @@ extern "C" recomp_func_t* get_function(uint32_t addr) {
 }
 
 extern "C" void bzero(uint8_t* restrict rdram, recomp_context* restrict ctx) {
-    uint32_t start_addr = ctx->r4;
+    gpr start_addr = ctx->r4;
-    uint32_t size = ctx->r5;
+    gpr size = ctx->r5;
 
     for (uint32_t i = 0; i < size; i++) {
         MEM_B(start_addr, i) = 0;
@@ -53,7 +53,7 @@ extern "C" void do_break(uint32_t vram) {
     exit(EXIT_FAILURE);
 }
 
-void run_thread_function(uint8_t* rdram, uint32_t addr, uint32_t sp, uint32_t arg) {
+void run_thread_function(uint8_t* rdram, uint64_t addr, uint64_t sp, uint64_t arg) {
     recomp_context ctx{};
     ctx.r29 = sp;
     ctx.r4 = arg;
@@ -62,7 +62,7 @@ void run_thread_function(uint8_t* rdram, uint32_t addr, uint32_t sp, uint32_t ar
 }
 
 extern "C" void game_init(uint8_t* restrict rdram, recomp_context* restrict ctx);
-void do_rom_read(uint8_t* rdram, uint32_t ram_address, uint32_t dev_address, size_t num_bytes);
+void do_rom_read(uint8_t* rdram, int32_t ram_address, uint32_t dev_address, size_t num_bytes);
 
 std::unique_ptr<uint8_t[]> rom;
 size_t rom_size;
@@ -105,7 +105,7 @@ int main(int argc, char **argv) {
 
     // Get entrypoint from ROM
     // TODO fix this for other IPL3 versions
-    uint32_t entrypoint = byteswap(*reinterpret_cast<uint32_t*>(rom.get() + 0x8));
+    int32_t entrypoint = byteswap(*reinterpret_cast<uint32_t*>(rom.get() + 0x8));
 
     // Allocate rdram_buffer
     std::unique_ptr<uint8_t[]> rdram_buffer = std::make_unique<uint8_t[]>(8 * 1024 * 1024);
@@ -134,17 +134,17 @@ int main(int argc, char **argv) {
 #endif
 
     // Set up stack pointer
-    context.r29 = 0x803FFFF0u;
+    context.r29 = 0xFFFFFFFF803FFFF0u;
 
     // Initialize variables normally set by IPL3
-    constexpr uint32_t osTvType = 0x80000300;
+    constexpr int32_t osTvType = 0x80000300;
-    constexpr uint32_t osRomType = 0x80000304;
+    constexpr int32_t osRomType = 0x80000304;
-    constexpr uint32_t osRomBase = 0x80000308;
+    constexpr int32_t osRomBase = 0x80000308;
-    constexpr uint32_t osResetType = 0x8000030c;
+    constexpr int32_t osResetType = 0x8000030c;
-    constexpr uint32_t osCicId = 0x80000310;
+    constexpr int32_t osCicId = 0x80000310;
-    constexpr uint32_t osVersion = 0x80000314;
+    constexpr int32_t osVersion = 0x80000314;
-    constexpr uint32_t osMemSize = 0x80000318;
+    constexpr int32_t osMemSize = 0x80000318;
-    constexpr uint32_t osAppNMIBuffer = 0x8000031c;
+    constexpr int32_t osAppNMIBuffer = 0x8000031c;
     uint8_t *rdram = rdram_buffer.get();
     MEM_W(osTvType, 0) = 1; // NTSC
     MEM_W(osRomBase, 0) = 0xB0000000u; // standard rom base
@@ -157,7 +157,7 @@ int main(int argc, char **argv) {
 
     debug_printf("[Recomp] Starting\n");
 
-    Multilibultra::preinit(rdram_buffer.get());
+    Multilibultra::preinit(rdram_buffer.get(), rom.get());
 
     game_init(rdram_buffer.get(), &context);
 

test/src/sp.cpp

@@ -10,7 +10,7 @@ extern "C" void osSpTaskStartGo_recomp(uint8_t* restrict rdram, recomp_context*
     //printf("[sp] osSpTaskStartGo(0x%08X)\n", (uint32_t)ctx->r4);
     OSTask* task = TO_PTR(OSTask, ctx->r4);
     if (task->t.type == M_GFXTASK) {
-        printf("[sp] Gfx task: %08X\n", (uint32_t)ctx->r4);
+        //printf("[sp] Gfx task: %08X\n", (uint32_t)ctx->r4);
     } else if (task->t.type == M_AUDTASK) {
         printf("[sp] Audio task: %08X\n", (uint32_t)ctx->r4);
     }

test/src/vi.cpp

@@ -1,3 +1,4 @@
+#include "../portultra/multilibultra.hpp"
 #include "recomp.h"
 
 extern "C" void osCreateViManager_recomp(uint8_t* restrict rdram, recomp_context* restrict ctx) {
@@ -21,7 +22,7 @@ extern "C" void osViGetNextFramebuffer_recomp(uint8_t* restrict rdram, recomp_co
 }
 
 extern "C" void osViSwapBuffer_recomp(uint8_t* restrict rdram, recomp_context* restrict ctx) {
-    ;
+    osViSwapBuffer(rdram, ctx->r4);
 }
 
 extern "C" void osViSetMode_recomp(uint8_t* restrict rdram, recomp_context* restrict ctx) {