Use the TLS register if HAVE_ARM_TLS_REGISTER is defined

this will be needed for SMP, but would improve dispatching GL calls a bit
on armv6 and above.
diff --git a/opengl/libs/Android.mk b/opengl/libs/Android.mk
index 9578452..6d20e80 100644
--- a/opengl/libs/Android.mk
+++ b/opengl/libs/Android.mk
@@ -32,6 +32,10 @@
 LOCAL_CFLAGS += -DADRENO130=1
 endif
 
+ifeq ($(ARCH_ARM_HAVE_TLS_REGISTER),true)
+  LOCAL_CFLAGS += -DHAVE_ARM_TLS_REGISTER
+endif
+
 include $(BUILD_SHARED_LIBRARY)
 installed_libEGL := $(LOCAL_INSTALLED_MODULE)
 
@@ -78,6 +82,10 @@
 LOCAL_CFLAGS += -DGL_GLEXT_PROTOTYPES -DEGL_EGLEXT_PROTOTYPES
 LOCAL_CFLAGS += -fvisibility=hidden
 
+ifeq ($(ARCH_ARM_HAVE_TLS_REGISTER),true)
+  LOCAL_CFLAGS += -DHAVE_ARM_TLS_REGISTER
+endif
+
 include $(BUILD_SHARED_LIBRARY)
 
 
@@ -107,4 +115,8 @@
 LOCAL_CFLAGS += -DGL_GLEXT_PROTOTYPES -DEGL_EGLEXT_PROTOTYPES
 LOCAL_CFLAGS += -fvisibility=hidden
 
+ifeq ($(ARCH_ARM_HAVE_TLS_REGISTER),true)
+  LOCAL_CFLAGS += -DHAVE_ARM_TLS_REGISTER
+endif
+
 include $(BUILD_SHARED_LIBRARY)
diff --git a/opengl/libs/GLES2/gl2.cpp b/opengl/libs/GLES2/gl2.cpp
index 4c0ba88..b8e3283 100644
--- a/opengl/libs/GLES2/gl2.cpp
+++ b/opengl/libs/GLES2/gl2.cpp
@@ -41,12 +41,20 @@
 
 #if USE_FAST_TLS_KEY
 
+    #ifdef HAVE_ARM_TLS_REGISTER
+        #define GET_TLS(reg) \
+            "mrc p15, 0, " #reg ", c13, c0, 3 \n"
+    #else
+        #define GET_TLS(reg) \
+            "mov   " #reg ", #0xFFFF0FFF      \n"  \
+            "ldr   " #reg ", [" #reg ", #-15] \n"
+    #endif
+
     #define API_ENTRY(_api) __attribute__((naked)) _api
 
     #define CALL_GL_API(_api, ...)                              \
          asm volatile(                                          \
-            "mov   r12, #0xFFFF0FFF   \n"                       \
-            "ldr   r12, [r12, #-15]   \n"                       \
+            GET_TLS(r12)                                        \
             "ldr   r12, [r12, %[tls]] \n"                       \
             "cmp   r12, #0            \n"                       \
             "ldrne pc,  [r12, %[api]] \n"                       \
@@ -56,7 +64,7 @@
               [api] "J"(__builtin_offsetof(gl_hooks_t, gl._api))    \
             :                                                   \
             );
-    
+
     #define CALL_GL_API_RETURN(_api, ...) \
         CALL_GL_API(_api, __VA_ARGS__) \
         return 0; // placate gcc's warnings. never reached.
diff --git a/opengl/libs/GLES_CM/gl.cpp b/opengl/libs/GLES_CM/gl.cpp
index 1e4c136..0c9352e 100644
--- a/opengl/libs/GLES_CM/gl.cpp
+++ b/opengl/libs/GLES_CM/gl.cpp
@@ -76,12 +76,20 @@
 
 #if USE_FAST_TLS_KEY && !CHECK_FOR_GL_ERRORS
 
+    #ifdef HAVE_ARM_TLS_REGISTER
+        #define GET_TLS(reg) \
+            "mrc p15, 0, " #reg ", c13, c0, 3 \n"
+    #else
+        #define GET_TLS(reg) \
+            "mov   " #reg ", #0xFFFF0FFF      \n"  \
+            "ldr   " #reg ", [" #reg ", #-15] \n"
+    #endif
+
     #define API_ENTRY(_api) __attribute__((naked)) _api
 
     #define CALL_GL_API(_api, ...)                              \
          asm volatile(                                          \
-            "mov   r12, #0xFFFF0FFF   \n"                       \
-            "ldr   r12, [r12, #-15]   \n"                       \
+            GET_TLS(r12)                                        \
             "ldr   r12, [r12, %[tls]] \n"                       \
             "cmp   r12, #0            \n"                       \
             "ldrne pc,  [r12, %[api]] \n"                       \
@@ -91,7 +99,7 @@
               [api] "J"(__builtin_offsetof(gl_hooks_t, gl._api))    \
             :                                                   \
             );
-    
+
     #define CALL_GL_API_RETURN(_api, ...) \
         CALL_GL_API(_api, __VA_ARGS__) \
         return 0; // placate gcc's warnings. never reached.