Introduce FreeBSD ARM64 support

It's 100% passing test fleet. Solid as a rock.

libc/intrin/exit.c

@@ -59,12 +59,12 @@ wontreturn void _Exit(int exitcode) {
                    : "rcx", "r11", "memory");
 #elif defined(__aarch64__)
       register long x0 asm("x0") = exitcode;
-      asm volatile("mov\tx8,%0\n\t"
-                   "mov\tx16,%1\n\t"
+      register long x8 asm("x8") = IsLinux() ? 94 : 1;
+      asm volatile("mov\tx16,%1\n\t"
                    "svc\t0"
                    : /* no outputs */
-                   : "i"(94), "i"(1), "r"(x0)
-                   : "x8", "x16", "memory");
+                   : "r"(x8), "i"(1), "r"(x0)
+                   : "x16", "memory");
 #else
 #error "unsupported architecture"
 #endif
@@ -77,12 +77,12 @@ wontreturn void _Exit(int exitcode) {
                  : "rcx", "r11", "memory");
 #else
     register long r0 asm("x0") = exitcode;
-    asm volatile("mov\tx8,%0\n\t"
-                 "mov\tx16,%1\n\t"
+    register long x8 asm("x8") = IsLinux() ? 93 : 431;
+    asm volatile("mov\tx16,%1\n\t"
                  "svc\t0"
                  : /* no outputs */
-                 : "i"(93), "i"(0x169), "r"(r0)
-                 : "x8", "memory");
+                 : "r"(x8), "i"(0x169), "r"(r0)
+                 : "memory");
 #endif
   } else if (IsWindows()) {
     uint32_t waitstatus;

libc/intrin/exit1.greg.c

@@ -74,13 +74,27 @@ wontreturn void _Exit1(int rc) {
   }
   notpossible;
 #elif defined(__aarch64__)
-  if (IsLinux()) {
-    register long r0 asm("x0") = rc;
-    asm volatile("mov\tx8,%0\n\t"
-                 "svc\t0"
-                 : /* no outputs */
-                 : "i"(93), "r"(r0)
-                 : "x8", "memory");
+  if (IsLinux() || IsFreebsd()) {
+    register int x0 asm("x0") = rc;
+    register int x8 asm("x8");
+    if (IsLinux()) {
+      x8 = 93;  // exit
+    } else if (IsFreebsd()) {
+      x8 = 431;  // thr_exit
+    } else {
+      __builtin_unreachable();
+    }
+    asm volatile("svc\t0" : "+r"(x0) : "r"(x8) : "memory");
+    if (SupportsFreebsd()) {
+      // On FreeBSD, thr_exit() fails if the current thread is orphaned.
+      // In that case we're really better off just calling plain _exit()
+      x0 = rc;
+      asm volatile("mov\tx8,#1\n\t"
+                   "svc\t0"
+                   : /* no outputs */
+                   : "r"(x0)
+                   : "memory");
+    }
   } else if (IsXnu()) {
     __syslib->__pthread_exit(0);
   }

libc/intrin/futex.S

@@ -32,10 +32,19 @@ _futex:
 	neg	%eax
 1:	pop	%rbp
 #elif defined(__aarch64__)
-	mov	x8,#0x062
-	svc	#0
+	ldr	x7,=__hostos
+	ldr	w7,[x7]
+	tst	x7,1			// IsLinux()
+	mov	x7,98			// futex (Linux)
+	mov	x8,454			// _umtx_op (FreeBSD)
+	csel	x8,x7,x8,ne		// choose syscall magnum
+	mov	x7,0			// clear carry (for Linux)
+	adds	x7,x7,0			// clear carry
+	svc	#0			// call kernel
+	bcc	1f			// jump if not carry
+	neg	x0,x0			// linux style errno
 #else
 #error "unsupported architecture"
 #endif /* __x86_64__ */
-	ret
+1:	ret
 	.endfn	_futex,globl,hidden

libc/intrin/g_fds.c

@@ -79,7 +79,7 @@ static textwindows void SetupWinStd(struct Fds *fds, int i, uint32_t x) {
 
 textstartup void __init_fds(int argc, char **argv, char **envp) {
   struct Fds *fds;
-  fds = __veil("r", &g_fds);
+  fds = &g_fds;
   fds->n = 4;
   atomic_store_explicit(&fds->f, 3, memory_order_relaxed);
   if (_weaken(_extend)) {

libc/intrin/sys_gettid.greg.c

@@ -18,6 +18,7 @@
 ╚─────────────────────────────────────────────────────────────────────────────*/
 #include "libc/calls/syscall-sysv.internal.h"
 #include "libc/dce.h"
+#include "libc/intrin/kprintf.h"
 #include "libc/nt/thread.h"
 #include "libc/nt/thunk/msabi.h"
 #include "libc/runtime/internal.h"
@@ -25,9 +26,9 @@
 __msabi extern typeof(GetCurrentThreadId) *const __imp_GetCurrentThreadId;
 
 int sys_gettid(void) {
+  int64_t wut;
 #ifdef __x86_64__
   int tid;
-  int64_t wut;
   if (IsWindows()) {
     tid = __imp_GetCurrentThreadId();
   } else if (IsLinux()) {
@@ -65,11 +66,23 @@ int sys_gettid(void) {
 #elif defined(__aarch64__)
   // this can't be used on xnu
   register long res asm("x0");
-  asm volatile("mov\tx8,%1\n\t"
-               "svc\t0"
-               : "=r"(res)
-               : "i"(178)
-               : "x8", "memory");
+  if (IsLinux()) {
+    asm volatile("mov\tx8,%1\n\t"
+                 "svc\t0"
+                 : "=r"(res)
+                 : "i"(178)
+                 : "x8", "memory");
+  } else if (IsFreebsd()) {
+    res = (long)&wut;
+    asm volatile("mov\tx8,%2\n\t"
+                 "svc\t0"
+                 : "+r"(res), "=m"(wut)
+                 : "i"(432)  // thr_self()
+                 : "x8", "memory");
+    res = wut;
+  } else {
+    res = __pid;
+  }
   return res;
 #else
 #error "arch unsupported"

libc/intrin/sys_umtx_timedwait_uint.c

@@ -19,7 +19,6 @@
 #include "libc/atomic.h"
 #include "libc/sysv/consts/clock.h"
 #include "libc/thread/freebsd.internal.h"
-#ifdef __x86_64__
 
 int sys_umtx_timedwait_uint_cp(atomic_int *, int, int, size_t,
                                struct _umtx_time *) asm("sys_futex_cp");
@@ -46,5 +45,3 @@ int sys_umtx_timedwait_uint(atomic_int *p, int expect, bool pshare,
   }
   return sys_umtx_timedwait_uint_cp(p, op, expect, size, tm_p);
 }
-
-#endif