Add back missing TlsAlloc() call

Cosmopolitan Libc once called this important function although somewhere
along the way, possibly in a refactoring, it got removed and __tls_alloc
has always been zero ever since.

Author: jart

libc/proc/fork-nt.c

@@ -65,6 +65,7 @@
 #include "libc/thread/itimer.internal.h"
 #include "libc/thread/posixthread.internal.h"
 #include "libc/thread/tls.h"
+#include "libc/thread/tls2.internal.h"
 #ifdef __x86_64__
 
 extern long __klog_handle;
@@ -112,7 +113,8 @@ static dontinline textwindows ssize_t ForkIo2(
   ssize_t rc = ForkIo(h, buf, n, fn);
   if (ischild) {
     // prevent crashes
-    __tls_enabled_set(false);
+    __threaded = false;
+    __tls_enabled = false;
     __pid = __imp_GetCurrentProcessId();
     __klog_handle = 0;
     __maps.maps = 0;
@@ -265,9 +267,8 @@ textwindows void WinMainForked(void) {
   ReadOrDie(reader, __data_start, __data_end - __data_start);
   ReadOrDie(reader, __bss_start, __bss_end - __bss_start);
   kStartTsc = savetsc;
+  __tls_enabled = false;
   __threaded = false;
-  __tls_index = 0;
-  __tls_enabled_set(false);
 
   // fixup memory manager
   __maps.free = 0;
@@ -456,9 +457,10 @@ textwindows int sys_fork_nt(uint32_t dwCreationFlags) {
   } else {
     rc = 0;
     // re-apply code morphing for thread-local storage
-    __set_tls(tib);
+    __tls_index = TlsAlloc();
+    __set_tls_win32(tib);
     __morph_tls();
-    __tls_enabled_set(true);
+    __tls_enabled = true;
     // the child's pending signals is initially empty
     atomic_store_explicit(&__sig.pending, 0, memory_order_relaxed);
     atomic_store_explicit(&tib->tib_sigpending, 0, memory_order_relaxed);

libc/runtime/cosmo.S

@@ -45,7 +45,7 @@ cosmo:	push	%rbp
 #endif /* SYSDEBUG */
 
 #ifndef NOX87
-//	Windows always initializes FPU to douuble precision.
+//	Windows always initializes FPU to double precision.
 //	WSL breaks Linux ABI by initializing FPU to double precision.
 //	This code makes long double long again.
 //

libc/runtime/enable_tls.c

@@ -251,6 +251,8 @@ textstartup void __enable_tls(void) {
   atomic_store_explicit(&_pthread_static.ptid, tid, memory_order_release);
 
   // ask the operating system to change the x86 segment register
+  if (IsWindows())
+    __tls_index = TlsAlloc();
   __set_tls(tib);
 
 #ifdef __x86_64__

libc/runtime/set_tls.c

@@ -38,7 +38,7 @@ dontinstrument textstartup void __set_tls(struct CosmoTib *tib) {
 #ifdef __x86_64__
   // ask the operating system to change the x86 segment register
   if (IsWindows()) {
-    asm("mov\t%1,%%gs:%0" : "=m"(*((long *)0x1480 + __tls_index)) : "r"(tib));
+    __set_tls_win32(tib);
   } else if (IsLinux()) {
     sys_set_tls(ARCH_SET_GS, tib);
   } else if (IsFreebsd()) {

libc/thread/tls2.internal.h

@@ -17,9 +17,8 @@ forceinline struct CosmoTib *__get_tls_privileged(void) {
     __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()) {
+    if (IsWindows())
       tib = *(char **)(tib + 0x1480 + __tls_index * 8);
-    }
   }
   return (struct CosmoTib *)tib;
 }