Fix compiling for linux arm64 (#650)

* Enable MK64 build option for torch

* Fix HMAS type error (present on arm64)

* Align memory pool and fix empty macro

* Experimental: Add 4096 memory alignment

* Add alignment for ALIGNED8 macro

* Use memset over deprecated bzero

* Ensure OPENGLES flag can be used in CMakeLists.txt

* Diagnostic: align ship2_window_i8 to rule out odd-alignment crash

---------

Co-authored-by: MegaMech <MegaMech@users.noreply.github.com>

src/code_80005FD0.c

@@ -3468,7 +3468,7 @@ void func_8000F124(void) {
 
 // Delete track waypoints
 void clear_path_point(TrackPathPoint* arg0, size_t size) {
-    bzero((void*) arg0, size * sizeof(TrackPathPoint));
+    memset((void*) arg0, 0, size * sizeof(TrackPathPoint));
 }
 
 // Appears to allocate memory for each track.

src/code_8006E9C0.c

@@ -60,7 +60,7 @@ void reset_object_variable(void) {
     s32 j;
     func_8006EB10();
     clear_object_list();
-    bzero(playerHUD, HUD_PLAYERS_SIZE * sizeof(hud_player));
+    memset(playerHUD, 0, HUD_PLAYERS_SIZE * sizeof(hud_player));
 
     for (i = 0; i < HUD_PLAYERS_SIZE; i++) {
         playerHUD[i].lapCount = 0;
@@ -114,7 +114,7 @@ void func_8006EB10(void) {
 }
 
 void clear_object_list() {
-    bzero(gObjectList, OBJECT_LIST_SIZE * sizeof(Object));
+    memset(gObjectList, 0, OBJECT_LIST_SIZE * sizeof(Object));
     objectListSize = -1;
 }
 

src/ending/podium_ceremony_actors.c

@@ -105,7 +105,7 @@ CeremonyActor* find_available_entry(void) {
 
         // Find an inactive actor.
         if ((actor->isActive & 1) == 0) {
-            bzero(actor, sizeof(CeremonyActor));
+            memset(actor, 0, sizeof(CeremonyActor));
             actor->isActive = 1;
             actor->unk24 = 1.0f;
             return actor;
@@ -311,7 +311,7 @@ void unused_80280FA8(UNUSED CeremonyActor* actor) {
 void balloons_and_fireworks_init(void) {
     D_802874D8.actorTimer = 0;
     sPodiumActorList = (CeremonyActor*) get_next_available_memory_addr(sizeof(CeremonyActor) * 200);
-    bzero(sPodiumActorList, (sizeof(CeremonyActor) * 200));
+    memset(sPodiumActorList, 0, (sizeof(CeremonyActor) * 200));
     new_actor(&initDummy);
 }
 

src/main.c

@@ -522,7 +522,7 @@ void display_and_vsync(void) {
 
 void init_segment_ending_sequences(void) {
 #ifdef TARGET_N64
-    bzero((void*) SEG_ENDING, SEG_ENDING_SIZE);
+    memset((void*) SEG_ENDING, 0, SEG_ENDING_SIZE);
     osWritebackDCacheAll();
     dma_copy((u8*) SEG_ENDING, (u8*) SEG_ENDING_ROM_START, SEG_ENDING_ROM_SIZE);
     osInvalICache((void*) SEG_ENDING, SEG_ENDING_SIZE);
@@ -532,7 +532,7 @@ void init_segment_ending_sequences(void) {
 
 void init_segment_racing(void) {
 #ifdef TARGET_N64
-    bzero((void*) SEG_RACING, SEG_RACING_SIZE);
+    memset((void*) SEG_RACING, 0, SEG_RACING_SIZE);
     osWritebackDCacheAll();
     dma_copy((u8*) SEG_RACING, (u8*) SEG_RACING_ROM_START, SEG_RACING_ROM_SIZE);
     osInvalICache((void*) SEG_RACING, SEG_RACING_SIZE);

src/os/__osViInit.c

@@ -15,7 +15,7 @@ extern OSViMode osViModeMpalLan1;
 extern OSViMode osViModeNtscLan1;
 
 void __osViInit(void) {
-    bzero(sViContexts, sizeof(sViContexts));
+    memset(sViContexts, 0, sizeof(sViContexts));
     __osViCurr = &sViContexts[0];
     __osViNext = &sViContexts[1];
     __osViNext->retraceCount = 1;

src/os/bstring.h

@@ -28,7 +28,6 @@ extern "C" {
 
 extern void bcopy(const void*, void*, size_t);
 extern int bcmp(const void*, const void*, int);
-extern void bzero(void*, size_t);
 extern void blkclr(void*, int);
 
 #ifdef __cplusplus

src/os/osCartRomInit.c

@@ -29,7 +29,7 @@ OSPiHandle* osCartRomInit(void) {
     CartRomHandle.domain = PI_DOMAIN1;
     // CartRomHandle.speed = 0;
 
-    bzero(&CartRomHandle.transferInfo, sizeof(__OSTranxInfo));
+    memset(&CartRomHandle.transferInfo, 0, sizeof(__OSTranxInfo));
 
     saveMask = __osDisableInt();
     CartRomHandle.next = __osPiTable;

src/os/osInitialize.c

@@ -57,7 +57,7 @@ void osInitialize(void) {
     }
     osClockRate = osClockRate * 3 / 4;
     if (osResetType == RESET_TYPE_COLD_RESET) {
-        bzero(osAppNmiBuffer, sizeof(osAppNmiBuffer));
+        memset(osAppNmiBuffer, 0, sizeof(osAppNmiBuffer));
     }
 
     eu_sp30 = HW_REG(PI_STATUS_REG, u32);

src/os/osLeoDiskInit.c

@@ -26,7 +26,7 @@ OSPiHandle* osLeoDiskInit(void) {
     HW_REG(PI_BSD_DOM2_PWD_REG, u32) = LeoDiskHandle.pulse;
     HW_REG(PI_BSD_DOM2_PGS_REG, u32) = LeoDiskHandle.pageSize;
     HW_REG(PI_BSD_DOM2_RLS_REG, u32) = LeoDiskHandle.relDuration;
-    bzero(&LeoDiskHandle.transferInfo, sizeof(__OSTranxInfo));
+    memset(&LeoDiskHandle.transferInfo, 0, sizeof(__OSTranxInfo));
     sp1c = __osDisableInt();
     LeoDiskHandle.next = __osPiTable;
     __osPiTable = &LeoDiskHandle;

src/port/audio/HMAS.h

@@ -58,7 +58,7 @@ struct HMAS_Effect {
 struct HMAS_ChannelInfo {
     ma_sound* sound;
 
-    uint64_t cursor;
+    ma_uint64 cursor;
     float pitch;
     float volume;
 

src/racing/memory.c

@@ -25,7 +25,7 @@
 s32 sGfxSeekPosition;
 s32 sPackedSeekPosition;
 
-static u8 sMemoryPool[0xFFFFFFF]; // Stock memory pool size: 0xAB630
+static u8 sMemoryPool[0x10000000] ALIGNED4096;
 uintptr_t sPoolEnd = sMemoryPool + sizeof(sMemoryPool);
 
 uintptr_t sPoolFreeSpace;
@@ -88,17 +88,16 @@ static uintptr_t get_texture2(size_t offset, const course_texture* textures) {
  * Default memory size, 701.984 Kilobytes.
  */
 void initialize_memory_pool() {
+    // Clear pool
+    memset(sMemoryPool, 0, sizeof(sMemoryPool));
 
-    uintptr_t poolStart = sMemoryPool;
-    // uintptr_t sPoolEnd = sMemoryPool + sizeof(sMemoryPool);
+    // Force the pointer to be exactly at the start of the aligned array
+    uintptr_t poolStart = (uintptr_t)sMemoryPool;
+    
+    // Ensure sPoolEnd is exactly at the end of the 256MB block
+    sPoolEnd = poolStart + sizeof(sMemoryPool);
 
-    bzero(sMemoryPool, sizeof(sMemoryPool));
-
-    poolStart = ALIGN16(poolStart);
-    // Truncate to a 16-byte boundary.
-    sPoolEnd &= ~0xF;
-
-    gFreeMemorySize = (sPoolEnd - poolStart) - 0x10;
+    gFreeMemorySize = sPoolEnd - poolStart;
     gNextFreeMemoryAddress = poolStart;
 
     PRINT_MEMPOOL;
@@ -309,7 +308,7 @@ u8* load_lakitu_tlut_x64(const char** textureList, size_t length) {
     }
 
     u8* textures = (u8*) gNextFreeMemoryAddress;
-    gNextFreeMemoryAddress += size;
+    gNextFreeMemoryAddress = ALIGN16(gNextFreeMemoryAddress + size);
     size_t offset = 0;
     for (size_t i = 0; i < length; i++) {
         u8* tex = (u8*) LOAD_ASSET_RAW(textureList[i]);