diff --git a/delay-stm32.c b/delay-stm32.c
index 6804723..bfb036c 100644
--- a/delay-stm32.c
+++ b/delay-stm32.c
@@ -5,46 +5,5 @@
 
 #if defined TEACUP_C_INCLUDE && defined __ARM_STM32__
 
-#include "cmsis-stm32f4xx.h"  // For __ASM() and ...
-
-
-/** Delay in microseconds.
-
-  \param delay Time to wait in microseconds.
-
-  Execution times on ARM aren't as predictable as they could be, because
-  there's a code prefetch engine which can change timings depending on the
-  position of the code in Flash. We could use the System Tick Timer for this
-  task, but this timer is probably better used for more important tasks.
-  delay_us() and delay_ms() are used only rarely and not in a way which would
-  require high precision.
-
-  Nevertheless, calibrated on the oscilloscope. Measured accuracy:
-
-             delay_us(10)    ...(100)   ...(1000)  ...(10000)  ...(65000)
-    96 MHz       10.23 us    99.3 us    0.994 ms    9.93 ms    64.5 ms
-
-  CAUTION: this currently works for a 96 MHz clock, only! As other clock rates
-           appear, there's more math neccessary, see the AVR version. Or simply
-           a second implementation.
-*/
-void delay_us(uint16_t delay) {
-
-    __ASM (".balign 16");  // No gambling with the prefetch engine.
-    while (delay) {
-      __ASM volatile (
-        "   nop               \n\t"  // One nop before the loop slows about 2%.
-        "   movs  r7, #9      \n\t"  // One more loop round slows about 20%
-        "1: nop               \n\t"
-        "   sub   r7, #1      \n\t"
-        "   cmp   r7, #0      \n\t"
-        "   bne   1b          \n\t"
-        :
-        :
-        : "r7", "cc"
-      );
-      delay--;
-    }
-}
 
 #endif /* defined TEACUP_C_INCLUDE && defined __ARM_STM32__ */
diff --git a/delay-stm32.h b/delay-stm32.h
new file mode 100644
index 0000000..6adf11a
--- /dev/null
+++ b/delay-stm32.h
@@ -0,0 +1,36 @@
+#ifndef _DELAY_STM32_H
+#define _DELAY_STM32_H
+
+#include "cmsis-stm32f4xx.h" // For __ASM() and ...
+
+/** Delay in microseconds.
+
+  \param delay Time to wait in microseconds.
+
+  Execution times on ARM aren't as predictable as they could be, because
+  there's a code prefetch engine which can change timings depending on the
+  position of the code in Flash. We could use the System Tick Timer for this
+  task, but this timer is probably better used for more important tasks.
+  delay_us() and delay_ms() are used only rarely and not in a way which would
+  require high precision.
+*/
+static void _delay(uint32_t) __attribute__ ((always_inline));
+inline void _delay(uint32_t delay)
+{
+
+    if (delay)
+    {
+        __ASM volatile(
+            "1:  subs %0, %0, #1 \n\t"
+            "    bne 1b\n\t"
+            : "+r"(delay));
+    }
+    else
+        __ASM volatile("nop \n\t");
+}
+
+// void _delay(uint32_t delay);
+#define TUNER 1
+#define delay_us(delay) (_delay((uint32_t)(((uint64_t)__SYSTEM_CLOCK) * (delay) / 3000000ULL * TUNER)))
+
+#endif /* _DELAY_STM32_H */
\ No newline at end of file
diff --git a/delay.h b/delay.h
index 81bb705..8177c56 100644
--- a/delay.h
+++ b/delay.h
@@ -3,9 +3,14 @@
 
 #include	<stdint.h>
 
+#ifndef __ARM_STM32__
 // microsecond delay, does NOT reset WDT if feature enabled
 void delay_us(uint16_t delay);
 
+#else
+#include "delay-stm32.h"
+#endif
+
 // millisecond delay, does reset WDT if feature enabled
 void delay_ms(uint32_t delay);