diff --git a/indra/cwdebug/debug.cc b/indra/cwdebug/debug.cc
index 18b51aad0..92ffc0d13 100644
--- a/indra/cwdebug/debug.cc
+++ b/indra/cwdebug/debug.cc
@@ -416,12 +416,23 @@ void cwdebug_backtrace(int n)
 #elif defined(DEBUG_CURLIO)
 
 #include "debug.h"
+#include "aithreadid.h"
 
 namespace debug
 {
 
 libcwd_do_type const libcw_do;
-CWD_TLS int Indent::S_indentation;
+ll_thread_local int Indent::S_indentation;
+
+Indent::Indent(int indent) : M_indent(indent)
+{
+	S_indentation += M_indent;
+}
+
+Indent::~Indent()
+{
+	S_indentation -= M_indent;
+}
 
 std::ostream& operator<<(std::ostream& os, Indent::print_nt)
 {
@@ -430,6 +441,19 @@ std::ostream& operator<<(std::ostream& os, Indent::print_nt)
   return os;
 }
 
+std::ostream& operator<<(std::ostream& os, print_thread_id_t)
+{
+	if (!AIThreadID::in_main_thread_inline())
+	{
+#ifdef LL_DARWIN
+		os << std::hex << (size_t)apr_os_thread_current() << std::dec << ' ';
+#else
+		os << std::hex << (size_t)AIThreadID::getCurrentThread_inline() << std::dec << ' ';
+#endif
+	}
+	return os;
+}
+
 std::ostream& operator<<(std::ostream& os, libcwd::buf2str const& b2s)
 {
   static char const c2s_tab[7] = { 'a', 'b', 't', 'n', 'v', 'f', 'r' };
diff --git a/indra/cwdebug/debug.h b/indra/cwdebug/debug.h
index 9e832a44c..c8f15cfb9 100644
--- a/indra/cwdebug/debug.h
+++ b/indra/cwdebug/debug.h
@@ -29,14 +29,13 @@
 
 #ifdef DEBUG_CURLIO
 
-#include "aithreadid.h"
-
 // If CWDEBUG is not defined, but DEBUG_CURLIO is, then replace
 // some of the cwd macro's with something that generates viewer
 // specific debug output. Note that this generates a LOT of
 // output and should not normally be defined.
 
 #include <string>
+#include "llpreprocessor.h"
 
 namespace debug {
 namespace libcwd {
@@ -49,6 +48,7 @@ struct buf2str {
 
 } // namespace libcwd
 
+enum print_thread_id_t { print_thread_id };
 inline void init() { }
 struct libcwd_do_type {
  void on() const { }
@@ -56,14 +56,15 @@ struct libcwd_do_type {
 extern LL_COMMON_API libcwd_do_type const libcw_do;
 struct Indent {
   int M_indent;
-  static LL_COMMON_API_TLS int S_indentation;
+  static ll_thread_local int S_indentation;
   enum LL_COMMON_API print_nt { print };
-  LL_COMMON_API Indent(int indent) : M_indent(indent) { S_indentation += M_indent; }
-  LL_COMMON_API ~Indent() { S_indentation -= M_indent; }
+  LL_COMMON_API Indent(int indent);
+  LL_COMMON_API ~Indent();
 };
 
 extern LL_COMMON_API std::ostream& operator<<(std::ostream& os, libcwd::buf2str const& b2s);
 extern LL_COMMON_API std::ostream& operator<<(std::ostream& os, Indent::print_nt);
+extern LL_COMMON_API std::ostream& operator<<(std::ostream& os, print_thread_id_t);
 
 namespace dc {
 
@@ -87,13 +88,13 @@ extern LL_COMMON_API fake_channel const notice;
 } // namespace debug
 
 #define Debug(x) do { using namespace debug; x; } while(0)
-#define Dout(a, b) do { using namespace debug;  if ((a).mOn) { llinfos_nf << AIThreadID::DoutPrint << (a).mLabel << ": " << Indent::print << b << llendl; } } while(0)
+#define Dout(a, b) do { using namespace debug;  if ((a).mOn) { llinfos_nf << print_thread_id << (a).mLabel << ": " << Indent::print << b << llendl; } } while(0)
 #define DoutEntering(a, b) \
   int __slviewer_debug_indentation = 2; \
   { \
 	using namespace debug; \
 	if ((a).mOn) \
-	  llinfos_nf << AIThreadID::DoutPrint << (a).mLabel << ": " << Indent::print << "Entering " << b << llendl; \
+	  llinfos_nf << print_thread_id << (a).mLabel << ": " << Indent::print << "Entering " << b << llendl; \
     else \
 	  __slviewer_debug_indentation = 0; \
   } \
diff --git a/indra/llcommon/aithreadid.cpp b/indra/llcommon/aithreadid.cpp
index 4eea7d37d..64a8dd21a 100644
--- a/indra/llcommon/aithreadid.cpp
+++ b/indra/llcommon/aithreadid.cpp
@@ -27,7 +27,6 @@
  *   - Initial version, written by Aleric Inglewood @ SL
  */
 
-#include "sys.h"
 #include <iostream>
 #include <iomanip>
 #include "aithreadid.h"
@@ -36,7 +35,7 @@ AIThreadID const AIThreadID::sNone(AIThreadID::none);
 apr_os_thread_t AIThreadID::sMainThreadID;
 apr_os_thread_t const AIThreadID::undefinedID = (apr_os_thread_t)-1;
 #ifndef LL_DARWIN
-apr_os_thread_t CWD_TLS AIThreadID::lCurrentThread;
+apr_os_thread_t ll_thread_local AIThreadID::lCurrentThread;
 #endif
 
 void AIThreadID::set_main_thread_id(void)
@@ -51,21 +50,27 @@ void AIThreadID::set_current_thread_id(void)
 #endif
 }
 
+void AIThreadID::reset(void)
+{
+	mID = lCurrentThread;
+}
+
+bool AIThreadID::equals_current_thread(void) const
+{
+	return apr_os_thread_equal(mID, lCurrentThread);
+}
+
+bool AIThreadID::in_main_thread(void)
+{
+	return apr_os_thread_equal(lCurrentThread, sMainThreadID);
+}
+
+apr_os_thread_t AIThreadID::getCurrentThread(void)
+{
+	return lCurrentThread;
+}
+
 std::ostream& operator<<(std::ostream& os, AIThreadID const& id)
 {
 	return os << id.mID;
 }
-
-std::ostream& operator<<(std::ostream& os, AIThreadID::dout_print_t)
-{
-	if (!AIThreadID::in_main_thread())
-	{
-#ifdef LL_DARWIN
-		os << std::hex << (size_t)apr_os_thread_current() << std::dec << ' ';
-#else
-		os << std::hex << (size_t)AIThreadID::lCurrentThread << std::dec << ' ';
-#endif
-	}
-	return os;
-}
-
diff --git a/indra/llcommon/aithreadid.h b/indra/llcommon/aithreadid.h
index 9f7893c66..2bb3f1dac 100644
--- a/indra/llcommon/aithreadid.h
+++ b/indra/llcommon/aithreadid.h
@@ -34,7 +34,6 @@
 #include <apr_portable.h>	// apr_os_thread_t, apr_os_thread_current(), apr_os_thread_equal().
 #include <iosfwd>			// std::ostream.
 #include "llpreprocessor.h"	// LL_COMMON_API, LL_COMMON_API_TLS
-#include "llerror.h"
 
 // Lightweight wrapper around apr_os_thread_t.
 // This class introduces no extra assembly code after optimization; it's only intend is to provide type-safety.
@@ -45,12 +44,11 @@ private:
 	static LL_COMMON_API apr_os_thread_t sMainThreadID;
 	static LL_COMMON_API apr_os_thread_t const undefinedID;
 #ifndef LL_DARWIN
-	static LL_COMMON_API_TLS apr_os_thread_t lCurrentThread;
+	static ll_thread_local apr_os_thread_t lCurrentThread;
 #endif
 public:
 	static LL_COMMON_API AIThreadID const sNone;
 	enum undefined_thread_t { none };
-	enum dout_print_t { DoutPrint };
 
 public:
 	AIThreadID(void) : mID(apr_os_thread_current()) { }
@@ -62,7 +60,6 @@ public:
 	friend LL_COMMON_API bool operator==(AIThreadID const& id1, AIThreadID const& id2) { return apr_os_thread_equal(id1.mID, id2.mID); }
 	friend LL_COMMON_API bool operator!=(AIThreadID const& id1, AIThreadID const& id2) { return !apr_os_thread_equal(id1.mID, id2.mID); }
 	friend LL_COMMON_API std::ostream& operator<<(std::ostream& os, AIThreadID const& id);
-	friend LL_COMMON_API std::ostream& operator<<(std::ostream& os, dout_print_t);
 	static void set_main_thread_id(void);					// Called once to set sMainThreadID.
 	static void set_current_thread_id(void);				// Called once for every thread to set lCurrentThread.
 #ifdef LL_DARWIN
@@ -70,9 +67,15 @@ public:
 	bool equals_current_thread(void) const { return apr_os_thread_equal(mID, apr_os_thread_current()); }
 	static bool in_main_thread(void) { return apr_os_thread_equal(apr_os_thread_current(), sMainThreadID); }
 #else
-	void reset(void) { mID = lCurrentThread; }
-	bool equals_current_thread(void) const { return apr_os_thread_equal(mID, lCurrentThread); }
-	static bool in_main_thread(void) { return apr_os_thread_equal(lCurrentThread, sMainThreadID); }
+	LL_COMMON_API void reset(void);
+	LL_COMMON_API bool equals_current_thread(void) const;
+	LL_COMMON_API static bool in_main_thread(void);
+	LL_COMMON_API static apr_os_thread_t getCurrentThread(void);
+	// The *_inline variants cannot be exported because they access a thread-local member.
+	void reset_inline(void) { mID = lCurrentThread; }
+	bool equals_current_thread_inline(void) const { return apr_os_thread_equal(mID, lCurrentThread); }
+	static bool in_main_thread_inline(void) { return apr_os_thread_equal(lCurrentThread, sMainThreadID); }
+	static apr_os_thread_t getCurrentThread_inline(void) { return lCurrentThread; }
 #endif
 };
 
diff --git a/indra/llcommon/llaprpool.cpp b/indra/llcommon/llaprpool.cpp
index de5761157..3dffa8300 100644
--- a/indra/llcommon/llaprpool.cpp
+++ b/indra/llcommon/llaprpool.cpp
@@ -60,10 +60,10 @@ void LLAPRPool::create(LLAPRPool& parent)
 	//
 	// In other words, it's safe for any thread to create a (sub)pool, independent of who
 	// owns the parent pool.
-	mOwner.reset();
+	mOwner.reset_inline();
 #else
 	mOwner = mParent->mOwner;
-	llassert(mOwner.equals_current_thread());
+	llassert(mOwner.equals_current_thread_inline());
 #endif
 	apr_status_t const apr_pool_create_status = apr_pool_create(&mPool, mParent->mPool);
 	llassert_always(apr_pool_create_status == APR_SUCCESS);
@@ -83,7 +83,7 @@ void LLAPRPool::destroy(void)
 		// of course. Otherwise, if we are a subpool, only the thread that owns
 		// the parent may destruct us, since that is the pool that is still alive,
 		// possibly being used by others and being altered here.
-		llassert(!mParent || mParent->mOwner.equals_current_thread());
+		llassert(!mParent || mParent->mOwner.equals_current_thread_inline());
 #endif
 		apr_pool_t* pool = mPool;
 		mPool = NULL;				// Mark that we are BEING destructed.
diff --git a/indra/llcommon/llpreprocessor.h b/indra/llcommon/llpreprocessor.h
index be0ec94a3..a1f598d4a 100644
--- a/indra/llcommon/llpreprocessor.h
+++ b/indra/llcommon/llpreprocessor.h
@@ -198,12 +198,9 @@
 // Darwin does not support thread-local data.
 #ifndef LL_DARWIN
 #if LL_WINDOWS
-// On windows, thread-local data is automatically exported.
-#define LL_COMMON_API_TLS __declspec(thread)
-#define CWD_TLS __declspec(thread)
+#define ll_thread_local __declspec(thread)
 #else // Linux
-#define LL_COMMON_API_TLS LL_COMMON_API __thread
-#define CWD_TLS __thread
+#define ll_thread_local __thread
 #endif
 #endif
 
diff --git a/indra/llcommon/llrefcount.cpp b/indra/llcommon/llrefcount.cpp
index d92735aac..0956e1628 100644
--- a/indra/llcommon/llrefcount.cpp
+++ b/indra/llcommon/llrefcount.cpp
@@ -75,7 +75,7 @@ void LLRefCount::ref() const
 		}
 
 		mMutex.lock() ;
-		mLockedThreadID.reset();
+		mLockedThreadID.reset_inline();
 
 		mRef++; 
 
@@ -96,7 +96,7 @@ S32 LLRefCount::unref() const
 		}
 
 		mMutex.lock() ;
-		mLockedThreadID.reset();
+		mLockedThreadID.reset_inline();
 		
 		llassert(mRef >= 1);
 		if (0 == --mRef) 
diff --git a/indra/llcommon/llthread.cpp b/indra/llcommon/llthread.cpp
index c8d0be2f4..5c1a29f24 100644
--- a/indra/llcommon/llthread.cpp
+++ b/indra/llcommon/llthread.cpp
@@ -72,7 +72,7 @@ LLAtomicS32	LLThread::sRunning = 0;
 
 LL_COMMON_API void assert_main_thread()
 {
-	if (!AIThreadID::in_main_thread())
+	if (!AIThreadID::in_main_thread_inline())
 	{
 		llerrs << "Illegal execution outside main thread." << llendl;
 	}
@@ -407,12 +407,12 @@ LLMutexBase::LLMutexBase() :
 
 bool LLMutexBase::isSelfLocked() const
 {
-	return mLockingThread.equals_current_thread();
+	return mLockingThread.equals_current_thread_inline();
 }
 
 void LLMutexBase::lock() 
 { 
-	if (mLockingThread.equals_current_thread())
+	if (mLockingThread.equals_current_thread_inline())
 	{ //redundant lock
 		mCount++;
 		return;
@@ -420,12 +420,12 @@ void LLMutexBase::lock()
 
 	apr_thread_mutex_lock(mAPRMutexp);
 	
-	mLockingThread.reset();
+	mLockingThread.reset_inline();
 }
 
 bool LLMutexBase::tryLock()
 {
-	if (mLockingThread.equals_current_thread())
+	if (mLockingThread.equals_current_thread_inline())
 	{ //redundant lock
 		mCount++;
 		return true;
@@ -433,7 +433,7 @@ bool LLMutexBase::tryLock()
 	bool success = !APR_STATUS_IS_EBUSY(apr_thread_mutex_trylock(mAPRMutexp));
 	if (success)
 	{
-		mLockingThread.reset();
+		mLockingThread.reset_inline();
 	}
 	return success;
 }
@@ -441,7 +441,7 @@ bool LLMutexBase::tryLock()
 // non-blocking, but does do a lock/unlock so not free
 bool LLMutexBase::isLocked() const
 {
-	if (mLockingThread.equals_current_thread())
+	if (mLockingThread.equals_current_thread_inline())
 	  return false;		// A call to lock() won't block.
 	if (APR_STATUS_IS_EBUSY(apr_thread_mutex_trylock(mAPRMutexp)))
 	  return true;