Add NXDK (Xbox) support

nanors/deps/obl/oblas_lite.c

@@ -58,7 +58,7 @@ void obl_axpyb32_ref(u8 *a, u32 *b, u8 u, unsigned k)
     for (unsigned idx = 0, p = 0; idx < k; idx += 8 * sizeof(u32), p++) {
         u32 tmp = b[p];
         while (tmp > 0) {
-#ifdef _MSC_VER
+#if defined(_MSC_VER) && !defined(__clang__)
             unsigned long index = 0;
             _BitScanForward(&index, tmp);
             unsigned tz = (unsigned int)index;

src/Platform.c

@@ -20,7 +20,7 @@ static int activeMutexes = 0;
 static int activeEvents = 0;
 static int activeCondVars = 0;
 
-#if defined(LC_WINDOWS)
+#if defined(LC_WINDOWS) && !defined(NXDK)
 
 #pragma pack(push, 8)
 typedef struct tagTHREADNAME_INFO
@@ -65,7 +65,9 @@ void setThreadNameWin32(const char* name) {
     }
 #endif
 }
+#endif
 
+#if defined(LC_WINDOWS)
 DWORD WINAPI ThreadProc(LPVOID lpParameter) {
     struct thread_context* ctx = (struct thread_context*)lpParameter;
 #elif defined(__WIIU__)
@@ -76,7 +78,7 @@ void* ThreadProc(void* context) {
     struct thread_context* ctx = (struct thread_context*)context;
 #endif
 
-#if defined(LC_WINDOWS)
+#if defined(LC_WINDOWS) && !defined(NXDK)
     setThreadNameWin32(ctx->name);
 #elif defined(__linux__) || defined(__FreeBSD__)
     pthread_setname_np(pthread_self(), ctx->name);
@@ -298,7 +300,7 @@ int PltCreateThread(const char* name, ThreadEntry entry, void* context, PLT_THRE
             free(ctx);
             return err;
         }
-        
+
     }
 #endif
 
@@ -309,7 +311,7 @@ int PltCreateThread(const char* name, ThreadEntry entry, void* context, PLT_THRE
 
 int PltCreateEvent(PLT_EVENT* event) {
 #if defined(LC_WINDOWS)
-    *event = CreateEventEx(NULL, NULL, CREATE_EVENT_MANUAL_RESET, EVENT_ALL_ACCESS);
+    *event = CreateEvent(NULL, TRUE, FALSE, NULL);
     if (!*event) {
         return -1;
     }
@@ -578,7 +580,7 @@ bool PltSafeStrcpy(char* dest, size_t dest_size, const char* src) {
     memset(dest, 0xFE, dest_size);
 #endif
 
-#ifdef _MSC_VER
+#if defined(_MSC_VER) && !defined(NXDK)
     // strncpy_s() with _TRUNCATE does what we need for MSVC.
     // We use this rather than strcpy_s() because we don't want
     // the invalid parameter handler invoked upon failure.

src/Platform.h

@@ -1,6 +1,6 @@
 #pragma once
 
-#ifdef _WIN32
+#if defined(_WIN32) && !defined(NXDK)
 // Prevent bogus definitions of error codes
 // that are incompatible with Winsock errors.
 #define _CRT_NO_POSIX_ERROR_CODES
@@ -13,11 +13,23 @@
 #include <string.h>
 #include <stdint.h>
 
-#ifdef _WIN32
+#if defined(_WIN32) && !defined(NXDK)
 #define WIN32_LEAN_AND_MEAN
 #include <windows.h>
 #include <winsock2.h>
 #include <ws2tcpip.h>
+#elif defined(NXDK)
+// nxdk exposes Win32-style threading primitives, but its network stack is
+// lwIP with POSIX/BSD socket compatibility headers. Pull in the socket-facing
+// headers from the POSIX side so moonlight-common-c uses the Unix backend.
+#define WIN32_LEAN_AND_MEAN
+#include <windows.h>
+#include <unistd.h>
+#include <sys/time.h>
+#include <sys/ioctl.h>
+#include <arpa/inet.h>
+#include <netinet/in.h>
+#include <fcntl.h>
 #elif defined(__APPLE__)
 #include <mach/mach_time.h>
 #include <unistd.h>
@@ -64,15 +76,19 @@
 #endif
 
 #ifdef LC_WINDOWS
+# if !defined(NXDK)
 // Windows doesn't have strtok_r() but it has the same
 // function named strtok_s().
 #define strtok_r strtok_s
+# endif
 
-# if defined(WINAPI_FAMILY) && WINAPI_FAMILY==WINAPI_FAMILY_APP
+# if !defined(NXDK)
-# define LC_UWP
+#  if defined(WINAPI_FAMILY) && WINAPI_FAMILY==WINAPI_FAMILY_APP
-# else
+#   define LC_UWP
-# define LC_WINDOWS_DESKTOP
+#  else
-#endif
+#   define LC_WINDOWS_DESKTOP
+#  endif
+# endif
 
 #endif
 
@@ -83,7 +99,7 @@
     if (ListenerCallbacks.logMessage) \
         ListenerCallbacks.logMessage(s, ##__VA_ARGS__)
 
-#if defined(LC_WINDOWS)
+#if defined(LC_WINDOWS) && !defined(__clang__)
 #include <crtdbg.h>
 #ifdef LC_DEBUG
 #define LC_ASSERT(x) __analysis_assume(x); \
@@ -116,21 +132,22 @@
 #define LC_ASSERT_VT(x) LC_ASSERT(x)
 #endif
 
-#ifdef _MSC_VER
+#if defined(__has_builtin) && __has_builtin(__builtin_bswap16)
+#define LC_HAS_BUILTIN_BSWAP
+#endif
+
+#if defined(LC_HAS_BUILTIN_BSWAP) || \
+        (defined(__GNUC__) && ((__GNUC__ > 4) || (__GNUC__ == 4 && __GNUC_MINOR__ >= 8)))
+#define BSWAP16(x) __builtin_bswap16(x)
+#define BSWAP32(x) __builtin_bswap32(x)
+#define BSWAP64(x) __builtin_bswap64(x)
+#elif defined(_MSC_VER)
 #pragma intrinsic(_byteswap_ushort)
 #define BSWAP16(x) _byteswap_ushort(x)
 #pragma intrinsic(_byteswap_ulong)
 #define BSWAP32(x) _byteswap_ulong(x)
 #pragma intrinsic(_byteswap_uint64)
 #define BSWAP64(x) _byteswap_uint64(x)
-#elif (__GNUC__ > 4) || (__GNUC__ == 4 && __GNUC_MINOR__ >= 8)
-#define BSWAP16(x) __builtin_bswap16(x)
-#define BSWAP32(x) __builtin_bswap32(x)
-#define BSWAP64(x) __builtin_bswap64(x)
-#elif defined(__has_builtin) && __has_builtin(__builtin_bswap16)
-#define BSWAP16(x) __builtin_bswap16(x)
-#define BSWAP32(x) __builtin_bswap32(x)
-#define BSWAP64(x) __builtin_bswap64(x)
 #else
 #error Please define your platform byteswap macros!
 #endif

src/PlatformSockets.c

@@ -12,7 +12,7 @@
 #endif
 #define TCPv6_MSS 1220
 
-#if defined(LC_WINDOWS)
+#if defined(LC_WINDOWS) && !defined(NXDK)
 
 #ifndef SIO_UDP_CONNRESET
 #define SIO_UDP_CONNRESET _WSAIOW(IOC_VENDOR, 12)
@@ -70,7 +70,7 @@ void shutdownTcpSocket(SOCKET s) {
 }
 
 int setNonFatalRecvTimeoutMs(SOCKET s, int timeoutMs) {
-#if defined(LC_WINDOWS)
+#if defined(LC_WINDOWS) && !defined(NXDK)
     // Windows says that SO_RCVTIMEO puts the socket into an indeterminate state
     // when a timeout occurs. MSDN doesn't go into it any more than that, but it
     // seems likely that they are referring to the inability to know whether a
@@ -93,7 +93,7 @@ int setNonFatalRecvTimeoutMs(SOCKET s, int timeoutMs) {
 }
 
 int pollSockets(struct pollfd* pollFds, int pollFdsCount, int timeoutMs) {
-#if defined(LC_WINDOWS)
+#if defined(LC_WINDOWS) && !defined(NXDK)
     // We could have used WSAPoll() but it has some nasty bugs
     // https://daniel.haxx.se/blog/2012/10/10/wsapoll-is-broken/
     //
@@ -206,7 +206,7 @@ int recvUdpSocket(SOCKET s, char* buffer, int size, bool useSelect) {
                     (LastSocketError() == EWOULDBLOCK ||
                      LastSocketError() == EINTR ||
                      LastSocketError() == EAGAIN ||
-         #if defined(LC_WINDOWS)
+         #if defined(LC_WINDOWS) && !defined(NXDK)
                      // This error is specific to overlapped I/O which isn't even
                      // possible to perform with recvfrom(). It seems to randomly
                      // be returned instead of WSAETIMEDOUT on certain systems.
@@ -221,7 +221,7 @@ int recvUdpSocket(SOCKET s, char* buffer, int size, bool useSelect) {
     // We may receive an error due to a previous ICMP Port Unreachable error received
     // by this socket. We want to ignore those and continue reading. If the remote party
     // is really dead, ENet or TCP connection failures will trigger connection teardown.
-#if defined(LC_WINDOWS)
+#if defined(LC_WINDOWS) && !defined(NXDK)
     } while (err < 0 && LastSocketError() == WSAECONNRESET);
 #else
     } while (err < 0 && LastSocketError() == ECONNREFUSED);
@@ -231,7 +231,7 @@ int recvUdpSocket(SOCKET s, char* buffer, int size, bool useSelect) {
 }
 
 void closeSocket(SOCKET s) {
-#if defined(LC_WINDOWS)
+#if defined(LC_WINDOWS) && !defined(NXDK)
     closesocket(s);
 #else
     close(s);
@@ -335,7 +335,7 @@ SOCKET bindUdpSocket(int addressFamily, struct sockaddr_storage* localAddr, SOCK
         int val = 1;
         setsockopt(s, SOL_SOCKET, SO_NOSIGPIPE, (char*)&val, sizeof(val));
     }
-#elif defined(LC_WINDOWS)
+#elif defined(LC_WINDOWS) && !defined(NXDK)
     {
         // Disable WSAECONNRESET for UDP sockets on Windows
         BOOL val = FALSE;
@@ -409,10 +409,10 @@ int setSocketNonBlocking(SOCKET s, bool enabled) {
 #if defined(__vita__) || defined(__HAIKU__)
     int val = enabled ? 1 : 0;
     return setsockopt(s, SOL_SOCKET, SO_NONBLOCK, (char*)&val, sizeof(val));
-#elif defined(O_NONBLOCK)
+#elif defined(O_NONBLOCK) && !defined(NXDK)
     return fcntl(s, F_SETFL, (enabled ? O_NONBLOCK : 0) | (fcntl(s, F_GETFL) & ~O_NONBLOCK));
 #elif defined(FIONBIO)
-#ifdef LC_WINDOWS
+#if defined(LC_WINDOWS) && !defined(NXDK)
     u_long val = enabled ? 1 : 0;
 #else
     int val = enabled ? 1 : 0;
@@ -452,7 +452,6 @@ SOCKET connectTcpSocket(struct sockaddr_storage* dstaddr, SOCKADDR_LEN addrlen,
     LC_SOCKADDR addr;
     struct pollfd pfd;
     int err;
-    int val;
 
     // Create a non-blocking TCP socket
     s = createSocket(dstaddr->ss_family, SOCK_STREAM, IPPROTO_TCP, true);
@@ -471,7 +470,9 @@ SOCKET connectTcpSocket(struct sockaddr_storage* dstaddr, SOCKADDR_LEN addrlen,
     // Note: This only changes the max packet size we can *receive* from the host PC.
     // We still must split our own sends into smaller chunks with TCP_NODELAY enabled to
     // avoid MTU issues on the way out to to the target.
-#if defined(LC_WINDOWS)
+#if defined(LC_WINDOWS) && !defined(NXDK)
+    int val;
+
     // Windows doesn't support setting TCP_MAXSEG but IP_PMTUDISC_DONT forces the MSS to the protocol
     // minimum which is what we want here. Linux doesn't do this (disabling PMTUD just avoids setting DF).
     if (dstaddr->ss_family == AF_INET) {
@@ -491,12 +492,13 @@ SOCKET connectTcpSocket(struct sockaddr_storage* dstaddr, SOCKADDR_LEN addrlen,
     // restrict MSS to the minimum. It strips all options out of the SYN packet which
     // forces the remote party to fall back to the minimum MSS. TCP_MAXSEG doesn't seem
     // to work correctly for outbound connections on macOS/iOS.
-    val = 1;
+    int val = 1;
     if (setsockopt(s, IPPROTO_TCP, TCP_NOOPT, (char*)&val, sizeof(val)) < 0) {
         Limelog("setsockopt(TCP_NOOPT, %d) failed: %d\n", val, (int)LastSocketError());
     }
 #elif defined(TCP_MAXSEG)
-    val = dstaddr->ss_family == AF_INET ? TCPv4_MSS : TCPv6_MSS;
+    int val = dstaddr->ss_family == AF_INET ? TCPv4_MSS : TCPv6_MSS;
+
     if (setsockopt(s, IPPROTO_TCP, TCP_MAXSEG, (char*)&val, sizeof(val)) < 0) {
         Limelog("setsockopt(TCP_MAXSEG, %d) failed: %d\n", val, (int)LastSocketError());
     }
@@ -1004,7 +1006,7 @@ void exitLowLatencyMode(void) {
 }
 
 int initializePlatformSockets(void) {
-#if defined(LC_WINDOWS)
+#if defined(LC_WINDOWS) && !defined(NXDK)
     WSADATA data;
     return WSAStartup(MAKEWORD(2, 0), &data);
 #elif defined(__vita__) || defined(__WIIU__) || defined(__3DS__)
@@ -1027,7 +1029,7 @@ int initializePlatformSockets(void) {
 }
 
 void cleanupPlatformSockets(void) {
-#if defined(LC_WINDOWS)
+#if defined(LC_WINDOWS) && !defined(NXDK)
     WSACleanup();
 #else
 #endif

src/PlatformSockets.h

@@ -18,7 +18,7 @@ extern in_port_t n3ds_udp_port;
 #endif
 #endif
 
-#ifdef _WIN32
+#if defined(_WIN32) && !defined(NXDK)
 #define WIN32_LEAN_AND_MEAN
 #include <windows.h>
 #include <wlanapi.h>
@@ -83,7 +83,9 @@ typedef int SOCKADDR_LEN;
 #include <signal.h>
 #include <poll.h>
 
+#ifndef ioctlsocket
 #define ioctlsocket ioctl
+#endif
 #define LastSocketError() errno
 #define SetLastSocketError(x) errno = x
 #define INVALID_SOCKET -1

src/rswrapper.c

@@ -117,7 +117,7 @@ reed_solomon_release_t reed_solomon_release_fn;
 reed_solomon_encode_t reed_solomon_encode_fn;
 reed_solomon_decode_t reed_solomon_decode_fn;
 
-#if defined(_MSC_VER) && (defined(_M_IX86) || defined(_M_AMD64))
+#if defined(_MSC_VER) && !defined(NXDK) && (defined(_M_IX86) || defined(_M_AMD64))
 
   #if defined(_M_AMD64)
   // For some reason this is needed to avoid a "C1189 No target architecture" error from winnt.h
@@ -135,7 +135,7 @@ reed_solomon_decode_t reed_solomon_decode_fn;
  * @details The streaming code will directly invoke these function pointers during encoding.
  */
 void reed_solomon_init(void) {
-#if defined(_MSC_VER) && (defined(_M_IX86) || defined(_M_AMD64))
+#if defined(_MSC_VER) && !defined(NXDK) && (defined(_M_IX86) || defined(_M_AMD64))
   // Visual Studio
   if (_msc_supports_avx512f()) {
     reed_solomon_new_fn = reed_solomon_new_avx512;