emitTLSLoad doesn't work on OSX with emulated TLS

- Provide a TLS solution that works on OSX with __emutls_*

hphp/runtime/vm/jit/x64-util.h

@@ -17,6 +17,11 @@
 #define X64_UTIL_H_
 
 #include "hphp/util/asm-x64.h"
+#ifndef USE_GCC_FAST_TLS
+#include <pthread.h>
+#include "hphp/runtime/vm/jit/abi-x64.h"
+#include "hphp/runtime/vm/jit/physreg.h"
+#endif
 #include "hphp/runtime/vm/jit/translator-inline.h"
 
 namespace HPHP {
@@ -29,6 +34,8 @@ translator_not_reached(X64Assembler &a) {
   }
 }
 
+#ifdef USE_GCC_FAST_TLS
+
 /*
  * TLS access: XXX we currently only support static-style TLS directly
  * linked off of FS.
@@ -46,23 +53,32 @@ translator_not_reached(X64Assembler &a) {
  *
  * The virtual addresses of TLS data are not exposed to C/C++. To figure it
  * out, we take a datum's linear address, and subtract it from the linear
- * address where TLS starts. If you use the void* variant here, it's up to
- * the programmer to ensure that it really is a TLS address.
+ * address where TLS starts.
  */
-static inline void
-emitTLSLoad(X64Assembler& a, const void* datum, RegNumber reg) {
-  uintptr_t virtualAddress = uintptr_t(datum) - tlsBase();
-  a.    fs();
-  a.    load_disp32_reg64(virtualAddress, reg);
-}
-
-//TODO(2525714) Will not work without USE_GCC_FAST_TLS being defined
 template<typename T>
 static inline void
 emitTLSLoad(X64Assembler& a, const ThreadLocalNoCheck<T>& datum,
             RegNumber reg) {
-  emitTLSLoad(a, &datum.m_node.m_p, reg);
+  uintptr_t virtualAddress = uintptr_t(&datum.m_node.m_p) - tlsBase();
+  a.    fs();
+  a.    load_disp32_reg64(virtualAddress, reg);
 }
 
+#else // USE_GCC_FAST_TLS
+
+template<typename T>
+static inline void
+emitTLSLoad(X64Assembler& a, const ThreadLocalNoCheck<T>& datum,
+            RegNumber reg) {
+  PhysRegSaver(a, kGPCallerSaved); // we don't know for sure what's alive
+  a.    emitImmReg(&datum.m_key, argNumToRegName[0]);
+  a.    call((TCA)pthread_getspecific);
+  if (reg != reg::rax) {
+    a.    mov_reg64_reg64(reg::rax, reg);
+  }
+}
+
+#endif // USE_GCC_FAST_TLS
+
 } }
 #endif

hphp/util/thread_local.h

@@ -31,12 +31,22 @@ namespace HPHP {
 // Clang seems to have added this feature, or at the very least it is ignoring
 // __thread keyword and compiling anyway
 //
+// On OSX, gcc does emulate TLS but in a manner that invalidates assumptions
+// we have made about __thread and makes accessing thread-local variables in a
+// JIT-friendly fashion difficult (as the compiler is doing a lot of magic that
+// is not contractual or documented that we would need to duplicate in emitted
+// code) so for now we're not going to use it. One possibility if we really
+// want to do this is to generate functions that access variables of interest
+// in ThreadLocal* (all of them are NoCheck right now) and use the bytes of
+// gcc's compiled functions to find the values we would need to pass to
+// __emutls_get_address.
+//
 // icc 13.0.0 appears to support it as well but we end up with
 // assembler warnings of unknown importance about incorrect section
 // types
 
-#if !defined(NO_TLS) &&                                         \
-  ((__llvm__ && __clang__) ||                                  \
+#if !defined(NO_TLS) && !defined(__APPLE__) &&                  \
+  ((__llvm__ && __clang__) ||                                   \
    __GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ > 3) ||     \
    __INTEL_COMPILER)
 #define USE_GCC_FAST_TLS
@@ -65,7 +75,7 @@ inline void ThreadLocalCreateKey(pthread_key_t *key, void (*del)(void*)) {
 
 /**
  * A thread-local object is a "global" object within a thread. This is useful
- * for writing apartment-threaded code, where nothing is actullay shared
+ * for writing apartment-threaded code, where nothing is actually shared
  * between different threads (hence no locking) but those variables are not
  * on stack in local scope. To use it, just do something like this,
  *
@@ -426,7 +436,7 @@ public:
     return *getNoCheck();
   }
 
-private:
+public:
   pthread_key_t m_key;
 };