diff --git a/lib/perf_counter.h b/lib/perf_counter.h
index d99cc94..e6b4400 100644
--- a/lib/perf_counter.h
+++ b/lib/perf_counter.h
@@ -211,7 +211,7 @@
                 )
 
 #define foreach3(__type, __array, __item)                                       \
-            using(__type *_ = __array, *__item = _, _ = _, )                    \
+            using(__type *_ = __array, *__item = _, _ = _, _ = _ )              \
             for (   uint_fast32_t CONNECT2(count,__LINE__) = dimof(__array);    \
                     CONNECT2(count,__LINE__) > 0;                               \
                     _++, __item = _, CONNECT2(count,__LINE__)--                 \
@@ -238,7 +238,8 @@
 
             
 #define __cycleof__(__STR, ...)                                                 \
-            using(int64_t _ = get_system_ticks(), {                             \
+            using(int64_t _ = get_system_ticks(), __cycle_count__ = _,          \
+                _=_, {                                                          \
                 _ = get_system_ticks() - _;                                     \
                 if (__PLOOC_VA_NUM_ARGS(__VA_ARGS__) == 0) {                    \
                     printf("\r\n");                                             \
diff --git a/perf_counter.c b/perf_counter.c
index 2ecc3fa..bc4529d 100644
--- a/perf_counter.c
+++ b/perf_counter.c
@@ -178,7 +178,7 @@ volatile static int64_t s_lSystemClockCounts = 0;
            function <b>SysTick_Config</b> is not included. In this case, the file <b><i>device</i>.h</b>
            must contain a vendor-specific implementation of this function.
  */
-static __attribute__((always_inline)) uint32_t SysTick_Config(uint32_t ticks)
+__STATIC_INLINE uint32_t SysTick_Config(uint32_t ticks)
 {
     if ((ticks - 1UL) > SysTick_LOAD_RELOAD_Msk)
     {
@@ -256,7 +256,7 @@ bool start_cycle_counter(void)
  *!       hence SysTick-LOAD and (SCB->ICSR & SCB_ICSR_PENDSTSET_Msk)
  *!       won't change. 
  */
-static __attribute__((always_inline)) int32_t check_systick(void)
+__STATIC_INLINE int32_t check_systick(void)
 {
     int32_t nTemp = (int32_t)SysTick->LOAD - (int32_t)SysTick->VAL;
 
@@ -306,7 +306,11 @@ int32_t stop_cycle_counter(void)
     return nTemp - s_nOffset;
 }
 
+#if defined(__IS_COMPILER_IAR__)
+__attribute__((constructor))
+#else
 __attribute__((constructor(255)))
+#endif
 void __perf_counter_init(void)
 {
     init_cycle_counter(true);
@@ -347,7 +351,9 @@ void delay_us(int32_t nUs)
  *!           and 2) do not include system header file <time.h>
  *!
  */
+#if !defined(__IS_COMPILER_IAR__)
 __attribute__((nothrow)) 
+#endif
 int64_t clock(void)
 {
     int64_t lTemp = 0;
@@ -360,7 +366,9 @@ int64_t clock(void)
 }
 
 
+#if !defined(__IS_COMPILER_IAR__)
 __attribute__((nothrow)) 
+#endif
 int64_t get_system_ticks(void)
 {
     int64_t lTemp = 0;
diff --git a/perf_counter.h b/perf_counter.h
index d99cc94..e6b4400 100644
--- a/perf_counter.h
+++ b/perf_counter.h
@@ -211,7 +211,7 @@
                 )
 
 #define foreach3(__type, __array, __item)                                       \
-            using(__type *_ = __array, *__item = _, _ = _, )                    \
+            using(__type *_ = __array, *__item = _, _ = _, _ = _ )              \
             for (   uint_fast32_t CONNECT2(count,__LINE__) = dimof(__array);    \
                     CONNECT2(count,__LINE__) > 0;                               \
                     _++, __item = _, CONNECT2(count,__LINE__)--                 \
@@ -238,7 +238,8 @@
 
             
 #define __cycleof__(__STR, ...)                                                 \
-            using(int64_t _ = get_system_ticks(), {                             \
+            using(int64_t _ = get_system_ticks(), __cycle_count__ = _,          \
+                _=_, {                                                          \
                 _ = get_system_ticks() - _;                                     \
                 if (__PLOOC_VA_NUM_ARGS(__VA_ARGS__) == 0) {                    \
                     printf("\r\n");                                             \