diff --git a/src/prof/mcount.s b/src/prof/mcount.s
index 3cb207d3..620c5571 100644
--- a/src/prof/mcount.s
+++ b/src/prof/mcount.s
@@ -17,30 +17,12 @@ _mcount:
         sd   $6,  32($29)
         sd   $7,  40($29)
 
-# Make sure we're not recursively called when compiling __mcount() 
-# With -pg
-        la   $4, _busy
-        lw   $5, 0($4)
-        bnez $5, done
-        nop
-
-# Mark busy
-        li   $5, 1
-        sw   $5, 0($4)
-
 # Call internal C handler
         move $4, $1
         move $5, $31
         jal  __mcount
         nop
 
-# Unmark busy
-        la   $4, _busy
-        li   $5, 0
-        sw   $5, 0($4)
-
-        done:
-
 # Restore registers
         ld   $31, 0($29)
         ld   $1,  8($29)
@@ -52,9 +34,6 @@ _mcount:
         j    $31
         move $31, $1           # restore caller's ra
 
-_busy:
-        .space	4
-
         .end _mcount
 
 		.set reorder
diff --git a/src/prof/prof.c b/src/prof/prof.c
index f4d08410..7baf7011 100644
--- a/src/prof/prof.c
+++ b/src/prof/prof.c
@@ -78,8 +78,11 @@ extern int _ftext;
 extern int _etext;
 
 /* forward declarations */
+__attribute__((__no_instrument_function__, __no_profile_instrument_function__))
 void __gprof_cleanup(void);
+__attribute__((__no_instrument_function__, __no_profile_instrument_function__))
 void __mcount(unsigned int, unsigned int);
+__attribute__((__no_instrument_function__, __no_profile_instrument_function__))
 static SceUInt timer_handler(SceUID uid, SceKernelSysClock *c1, SceKernelSysClock *c2, void *common);
 
 /** Initializes pg library
@@ -89,6 +92,7 @@ static SceUInt timer_handler(SceUID uid, SceKernelSysClock *c1, SceKernelSysCloc
     for sampling statistics. Note that this also installs a timer that
     runs at 1000 hert.
 */
+__attribute__((__no_instrument_function__, __no_profile_instrument_function__))
 static void initialize()
 {
         initialized = 1;
@@ -156,6 +160,7 @@ static void initialize()
     Called from atexit() handler; will dump out a host:gmon.out file 
     with all collected information.
 */
+__attribute__((__no_instrument_function__, __no_profile_instrument_function__))
 void __gprof_cleanup()
 {
         FILE *fp;
@@ -205,6 +210,7 @@ void __gprof_cleanup()
     beginning of each compiled routine, which eventually brings the 
     control to here. 
 */
+__attribute__((__no_instrument_function__, __no_profile_instrument_function__))
 void __mcount(unsigned int frompc, unsigned int selfpc)
 { 
         int e;
@@ -238,6 +244,7 @@ void __mcount(unsigned int frompc, unsigned int selfpc)
 
 /** Internal timer handler
 */
+__attribute__((__no_instrument_function__, __no_profile_instrument_function__))
 static SceUInt timer_handler(SceUID uid, SceKernelSysClock *requested, SceKernelSysClock *actual, void *common)
 {
         unsigned int frompc = gp.pc;