Revert "Use %gs as TLS register when dlopen() is linked"

This reverts commit d71da7f.

libc/calls/close-nt.c

@@ -28,15 +28,14 @@
 #include "libc/sysv/consts/o.h"
 #include "libc/sysv/errfuns.h"
 
-void sys_fcntl_nt_lock_cleanup(int);
-
 textwindows int sys_close_nt(int fd, int fildes) {
   if (fd + 0u >= g_fds.n) return ebadf();
   struct Fd *f = g_fds.p + fd;
   switch (f->kind) {
     case kFdEmpty:
       return ebadf();
     case kFdFile:
+      void sys_fcntl_nt_lock_cleanup(int);
       if (_weaken(sys_fcntl_nt_lock_cleanup)) {
         _weaken(sys_fcntl_nt_lock_cleanup)(fildes);
       }

libc/calls/getprogramexecutablename.greg.c

@@ -136,11 +136,10 @@ static inline void InitProgramExecutableNameImpl(void) {
   }
 
   // give up and just copy argv[0] into it
-  char *p, *e;
   if ((q = __argv[0])) {
   CopyString:
-    p = g_prog.u.buf;
-    e = p + sizeof(g_prog.u.buf);
+    char *p = g_prog.u.buf;
+    char *e = p + sizeof(g_prog.u.buf);
     while ((c = *q++)) {
       if (p + 1 < e) {
         *p++ = c;

libc/calls/park.c

@@ -35,15 +35,14 @@ static textwindows int _park_thread(uint32_t msdelay, sigset_t waitmask,
   if ((sig = __sig_get(waitmask))) goto HandleSignal;
   int expect = 0;
   atomic_int futex = 0;
-  int handler_was_called;
   struct PosixThread *pt = _pthread_self();
   pt->pt_blkmask = waitmask;
   atomic_store_explicit(&pt->pt_blocker, &futex, memory_order_release);
   bool32 ok = WaitOnAddress(&futex, &expect, sizeof(int), msdelay);
   atomic_store_explicit(&pt->pt_blocker, 0, memory_order_release);
   if (ok && (sig = __sig_get(waitmask))) {
   HandleSignal:
-    handler_was_called = __sig_relay(sig, SI_KERNEL, waitmask);
+    int handler_was_called = __sig_relay(sig, SI_KERNEL, waitmask);
     if (_check_cancel() == -1) return -1;
     if (!restartable || (handler_was_called & SIG_HANDLED_NO_RESTART)) {
       return eintr();

libc/calls/read-nt.c

@@ -723,7 +723,6 @@ static textwindows int WaitForConsole(struct Fd *f, sigset_t waitmask) {
   int sig;
   int64_t sem;
   uint32_t wi, ms = -1;
-  int handler_was_called;
   if (!__ttyconf.vmin) {
     if (!__ttyconf.vtime) {
       return 0;  // non-blocking w/o raising eagain
@@ -746,7 +745,7 @@ static textwindows int WaitForConsole(struct Fd *f, sigset_t waitmask) {
   if (wi != 1) return __winerr();      // wait failed
   if (!(sig = __sig_get(waitmask))) return eintr();
 DeliverSignal:
-  handler_was_called = __sig_relay(sig, SI_KERNEL, waitmask);
+  int handler_was_called = __sig_relay(sig, SI_KERNEL, waitmask);
   if (_check_cancel() == -1) return -1;
   if (!(handler_was_called & SIG_HANDLED_NO_RESTART)) return -2;
   return eintr();

libc/calls/readwrite-nt.c

@@ -47,9 +47,7 @@ sys_readwrite_nt(int fd, void *data, size_t size, ssize_t offset,
                  bool32 ReadOrWriteFile(int64_t, void *, uint32_t, uint32_t *,
                                         struct NtOverlapped *)) {
   int sig;
-  bool eagained;
   uint32_t exchanged;
-  int handler_was_called;
   struct Fd *f = g_fds.p + fd;
 
   // win32 i/o apis generally take 32-bit values thus we implicitly
@@ -81,7 +79,7 @@ sys_readwrite_nt(int fd, void *data, size_t size, ssize_t offset,
   }
 
 RestartOperation:
-  eagained = false;
+  bool eagained = false;
   // check for signals and cancelation
   if (_check_cancel() == -1) return -1;  // ECANCELED
   if ((sig = __sig_get(waitmask))) goto HandleInterrupt;
@@ -134,7 +132,7 @@ sys_readwrite_nt(int fd, void *data, size_t size, ssize_t offset,
     // otherwise it must be due to a kill() via __sig_cancel()
     if ((sig = __sig_get(waitmask))) {
     HandleInterrupt:
-      handler_was_called = __sig_relay(sig, SI_KERNEL, waitmask);
+      int handler_was_called = __sig_relay(sig, SI_KERNEL, waitmask);
       if (_check_cancel() == -1) return -1;  // possible if we SIGTHR'd
       // read() is @restartable unless non-SA_RESTART hands were called
       if (!(handler_was_called & SIG_HANDLED_NO_RESTART)) {

libc/dlopen/dlopen.c

@@ -39,7 +39,6 @@
 #include "libc/nt/errors.h"
 #include "libc/nt/memory.h"
 #include "libc/nt/runtime.h"
-#include "libc/runtime/internal.h"
 #include "libc/runtime/runtime.h"
 #include "libc/runtime/syslib.internal.h"
 #include "libc/stdio/stdio.h"
@@ -76,6 +75,17 @@ __static_yoink(".dlopen.aarch64.glibc.elf");
 #define XNU_RTLD_LOCAL  4
 #define XNU_RTLD_GLOBAL 8
 
+#define BEGIN_FOREIGN                         \
+  {                                           \
+    struct CosmoTib *cosmo_tib = __get_tls(); \
+    pthread_mutex_lock(&foreign.lock);        \
+    __set_tls(foreign.tib)
+
+#define END_FOREIGN                    \
+  __set_tls(cosmo_tib);                \
+  pthread_mutex_unlock(&foreign.lock); \
+  }
+
 struct loaded {
   char *base;
   char *entry;
@@ -86,6 +96,8 @@ struct loaded {
 static struct {
   atomic_uint once;
   bool is_supported;
+  struct CosmoTib *tib;
+  pthread_mutex_t lock;
   void *(*dlopen)(const char *, int);
   void *(*dlsym)(void *, const char *);
   int (*dlclose)(void *);
@@ -300,6 +312,7 @@ static void foreign_setup(void) {
   if (!dlopen_helper) {
     return;  // this platform isn't supported yet
   }
+  struct CosmoTib *cosmo_tib = __get_tls();
   if (!setjmp(foreign.jb)) {
     elf_exec(dlopen_helper, interp, 2,
              (char *[]){
@@ -310,6 +323,8 @@ static void foreign_setup(void) {
              environ);
     return;  // if it returns then it failed
   }
+  foreign.tib = __get_tls();
+  __set_tls(cosmo_tib);
   foreign.is_supported = true;
 }
 
@@ -376,17 +391,15 @@ static char *dlsym_nt_alloc_rwx_block(void) {
 static void *dlsym_nt_alloc_rwx(size_t n) {
   void *res;
   static char *block;
-  static pthread_spinlock_t lock;
-  pthread_spin_lock(&lock);
+  pthread_mutex_lock(&foreign.lock);
   if (!block || READ32LE(block) + n > 65536) {
     if (!(block = dlsym_nt_alloc_rwx_block())) {
-      pthread_spin_unlock(&lock);
       return 0;
     }
   }
   res = block + READ32LE(block);
   WRITE32LE(block, READ32LE(block) + n);
-  pthread_spin_unlock(&lock);
+  pthread_mutex_unlock(&foreign.lock);
   return res;
 }
 
@@ -482,7 +495,9 @@ void *dlopen(const char *path, int mode) {
     last_dlerror = "dlopen() isn't supported on x86-64 MacOS";
     res = 0;
   } else if (foreign_init()) {
+    BEGIN_FOREIGN;
     res = foreign.dlopen(path, mode);
+    END_FOREIGN;
   } else {
     res = 0;
   }
@@ -509,7 +524,9 @@ void *dlsym(void *handle, const char *name) {
     last_dlerror = "dlopen() isn't supported on x86-64 MacOS";
     res = 0;
   } else if (foreign_init()) {
+    BEGIN_FOREIGN;
     res = foreign.dlsym(handle, name);
+    END_FOREIGN;
   } else {
     res = 0;
   }
@@ -533,7 +550,9 @@ int dlclose(void *handle) {
     last_dlerror = "dlopen() isn't supported on x86-64 MacOS";
     res = -1;
   } else if (foreign_init()) {
+    BEGIN_FOREIGN;
     res = foreign.dlclose(handle);
+    END_FOREIGN;
   } else {
     res = -1;
   }
@@ -551,24 +570,12 @@ char *dlerror(void) {
   } else if (IsWindows() || IsXnu()) {
     res = (char *)last_dlerror;
   } else if (foreign_init()) {
+    BEGIN_FOREIGN;
     res = foreign.dlerror();
+    END_FOREIGN;
   } else {
     res = (char *)last_dlerror;
   }
   STRACE("dlerror() → %#s", res);
   return res;
 }
-
-#ifdef __x86_64__
-static textstartup void dlopen_init() {
-  if (IsLinux() || IsFreebsd() || IsNetbsd()) {
-    // switch from %fs to %gs for tls
-    struct CosmoTib *tib = __get_tls();
-    __morph_tls();
-    __set_tls(tib);
-  }
-}
-const void *const dlopen_ctor[] initarray = {
-    dlopen_init,
-};
-#endif

libc/proc/wait4-nt.c

@@ -72,7 +72,7 @@ static textwindows int __proc_wait(int pid, int *wstatus, int options,
   for (;;) {
 
     // check for signals and cancelation
-    int rc, sig, handler_was_called;
+    int sig, handler_was_called;
     if (_check_cancel() == -1) {
       return -1;
     }
@@ -89,7 +89,7 @@ static textwindows int __proc_wait(int pid, int *wstatus, int options,
     // check for zombie to harvest
     __proc_lock();
   CheckForZombies:
-    rc = __proc_check(pid, wstatus, rusage);
+    int rc = __proc_check(pid, wstatus, rusage);
     if (rc || (options & WNOHANG)) {
       __proc_unlock();
       return rc;

libc/runtime/morph_tls.c

@@ -55,13 +55,11 @@ privileged void __morph_tls(void) {
     // address 0x30 was promised to us, according to Go team
     // https://github.com/golang/go/issues/23617
     dis = 0x30;
-  } else if (IsWindows()) {
+  } else {
     // MSVC __declspec(thread) generates binary code for this
     // %gs:0x1480 abi. So long as TlsAlloc() isn't called >64
     // times we should be good.
     dis = 0x1480 + __tls_index * 8;
-  } else {
-    dis = 0;
   }
 
   // iterate over modifiable code looking for 9 byte instruction
@@ -114,7 +112,6 @@ privileged void __morph_tls(void) {
     }
   }
 
-  __tls_morphed = true;
   __morph_end();
 #endif
 }

libc/runtime/set_tls.c

@@ -26,9 +26,6 @@
 #include "libc/thread/tls.h"
 #include "libc/thread/tls2.internal.h"
 
-#define AMD64_SET_FSBASE 129
-#define AMD64_SET_GSBASE 131
-
 int sys_set_tls();
 
 textstartup void __set_tls(struct CosmoTib *tib) {
@@ -39,9 +36,9 @@ textstartup void __set_tls(struct CosmoTib *tib) {
     npassert(0 <= __tls_index && __tls_index < 64);
     asm("mov\t%1,%%gs:%0" : "=m"(*((long *)0x1480 + __tls_index)) : "r"(tib));
   } else if (IsFreebsd()) {
-    sys_set_tls(__tls_morphed ? AMD64_SET_GSBASE : AMD64_SET_FSBASE, tib);
+    sys_set_tls(129 /*AMD64_SET_FSBASE*/, tib);
   } else if (IsLinux()) {
-    sys_set_tls(__tls_morphed ? ARCH_SET_GS : ARCH_SET_FS, tib);
+    sys_set_tls(ARCH_SET_FS, tib);
   } else if (IsNetbsd()) {
     // netbsd has sysarch(X86_SET_FSBASE) but we can't use that because
     // signal handlers will cause it to be reset due to not setting the

libc/stdio/fmt.c

@@ -705,7 +705,7 @@ static int __fmt_bround(struct FPBits *b, int prec, int prec1) {
   return prec;
 }
 
-static int __fmt_noop(const char *x, void *y, size_t z) {
+static int __fmt_noop(const char *, void *, size_t) {
   return 0;
 }
 

libc/thread/tls.h

@@ -40,7 +40,6 @@ struct CosmoTib {
 };
 
 extern int __threaded;
-extern bool __tls_morphed;
 extern unsigned __tls_index;
 
 char *_mktls(struct CosmoTib **);

libc/thread/tls2.internal.h

@@ -15,11 +15,7 @@ COSMOPOLITAN_C_START_
 __funline struct CosmoTib *__get_tls_privileged(void) {
   char *tib, *lin = (char *)0x30;
   if (IsLinux() || IsFreebsd() || IsNetbsd() || IsOpenbsd() || IsMetal()) {
-    if (!__tls_morphed) {
-      asm("mov\t%%fs:(%1),%0" : "=a"(tib) : "r"(lin) : "memory");
-    } else {
-      asm("mov\t%%gs:(%1),%0" : "=a"(tib) : "r"(lin) : "memory");
-    }
+    asm("mov\t%%fs:(%1),%0" : "=a"(tib) : "r"(lin) : "memory");
   } else {
     asm("mov\t%%gs:(%1),%0" : "=a"(tib) : "r"(lin) : "memory");
     if (IsWindows()) {