Perform some code cleanup

libc/thread/create.c

@@ -29,6 +29,7 @@
 #include "libc/sysv/consts/clone.h"
 #include "libc/sysv/consts/map.h"
 #include "libc/sysv/consts/prot.h"
+#include "libc/thread/internal.h"
 #include "libc/thread/thread.h"
 
 STATIC_YOINK("_main_thread_ctor");
@@ -77,6 +78,7 @@ static int cthread_start(void *arg) {
     exitcode = (void *)rc.dx;
   }
   td->exitcode = exitcode;
+  _pthread_key_destruct(td->key);
   if (atomic_load(&td->state) & cthread_detached) {
     // we're still using the stack
     // thus we can't munmap it yet

libc/thread/join.c

@@ -54,7 +54,6 @@ int cthread_join(cthread_t td, void **exitcode) {
   } else {
     if (~atomic_fetch_add(&td->state, cthread_joining) & cthread_finished) {
       while ((x = atomic_load(&td->tid))) {
-        // FUTEX_WAIT_PRIVATE makes it hang
         cthread_memory_wait32(&td->tid, x, 0);
       }
     }

libc/thread/pthread_key_destruct.c

@@ -30,7 +30,7 @@ StartOver:
     x = _pthread_key_usage[i];
     while (x) {
       j = bsrl(x);
-      if ((dtor = _pthread_key_dtor[i * 64 + j]) && (value = key[i * 64 + j])) {
+      if ((value = key[i * 64 + j]) && (dtor = _pthread_key_dtor[i * 64 + j])) {
         key[i * 64 + j] = 0;
         dtor(value);
         goto StartOver;

libc/thread/wait.c

@@ -18,53 +18,25 @@
 ╚─────────────────────────────────────────────────────────────────────────────*/
 #include "libc/bits/atomic.h"
 #include "libc/calls/calls.h"
+#include "libc/calls/struct/timespec.h"
 #include "libc/dce.h"
-#include "libc/sysv/consts/clock.h"
-#include "libc/sysv/consts/futex.h"
-#include "libc/thread/freebsd.internal.h"
+#include "libc/errno.h"
+#include "libc/intrin/futex.internal.h"
 #include "libc/thread/thread.h"
 
 int cthread_memory_wait32(int* addr, int val, const struct timespec* timeout) {
   size_t size;
-  struct _umtx_time *put, ut;
   if (IsLinux() || IsOpenbsd()) {
-    return sys_futex(addr, FUTEX_WAIT, val, timeout, 0);
-
-#if 0
-  } else if (IsFreebsd()) {
-    if (!timeout) {
-      put = 0;
-      size = 0;
-    } else {
-      ut._flags = 0;
-      ut._clockid = CLOCK_REALTIME;
-      ut._timeout = *timeout;
-      put = &ut;
-      size = sizeof(ut);
-    }
-    return _umtx_op(addr, UMTX_OP_MUTEX_WAIT, 0, &size, put);
-#endif
-
+    return _futex_wait(addr, val, timeout);
   } else {
-    unsigned tries;
-    for (tries = 1; atomic_load(addr) == val; ++tries) {
-      if (tries & 7) {
-        __builtin_ia32_pause();
-      } else {
-        sched_yield();
-      }
-    }
-    return 0;
+    return sched_yield();
   }
 }
 
 int cthread_memory_wake32(int* addr, int n) {
   if (IsLinux() || IsOpenbsd()) {
-    return sys_futex(addr, FUTEX_WAKE, n, 0, 0);
-#if 0
-  } else if (IsFreebsd()) {
-    return _umtx_op(addr, UMTX_OP_MUTEX_WAKE, n, 0, 0);
-#endif
+    return _futex_wake(addr, n);
+  } else {
+    return 0;
   }
-  return -1;
 }

libc/thread/zombie.c

@@ -41,6 +41,7 @@ void cthread_zombies_add(cthread_t td) {
 
 void cthread_zombies_reap(void) {
   struct Zombie *z;
+  // TODO(jart): Is this right? Update to not use malloc/free?
   while ((z = atomic_load(&cthread_zombies)) && !atomic_load(&z->td->tid)) {
     if (atomic_compare_exchange_weak(&cthread_zombies, &z, z->next)) {
       munmap(z->td->alloc.bottom, z->td->alloc.top - z->td->alloc.bottom);