diff --git a/apps/codecs/lib/codeclib.c b/apps/codecs/lib/codeclib.c
index 329b0e2136..5af4e01fff 100644
--- a/apps/codecs/lib/codeclib.c
+++ b/apps/codecs/lib/codeclib.c
@@ -163,7 +163,9 @@ const uint8_t bs_clz_tab[256] ICONST_ATTR = {
 
 #ifdef RB_PROFILE
 void __cyg_profile_func_enter(void *this_fn, void *call_site) {
-#ifdef CPU_COLDFIRE
+/* This workaround is required for coldfire gcc 3.4 but is broken for 4.4
+   and 4.5, but for those the other way works. */
+#if defined(CPU_COLDFIRE) && defined(__GNUC__) && __GNUC__ < 4
     (void)call_site;
     ci->profile_func_enter(this_fn, __builtin_return_address(1));
 #else
diff --git a/firmware/thread.c b/firmware/thread.c
index 655af1a940..91fe81be4a 100644
--- a/firmware/thread.c
+++ b/firmware/thread.c
@@ -1149,7 +1149,18 @@ void switch_thread(void)
     }
 
 #ifdef RB_PROFILE
+#ifdef CPU_COLDFIRE
+    /* Call this from asm to make sure the sp is pointing to the
+       correct place before the context is saved */
+    uint16_t id = thread->id & THREAD_ID_SLOT_MASK;
+    asm volatile ("move.l %[id], -(%%sp)\n\t"
+                  "jsr profile_thread_stopped\n\t"
+                  "addq.l #4, %%sp\n\t"
+                  :: [id] "r" (id)
+                  : "cc", "memory");
+#else
     profile_thread_stopped(thread->id & THREAD_ID_SLOT_MASK);
+#endif
 #endif
 
     /* Begin task switching by saving our current context so that we can