How to calculate the worst and the mean execution time of an assembly instruction in the STM32F103?

[flammmable]

I have write two simple code examples.
First:

  LDR R0, =0x00000000
  LDR R1, =0x0000FFFF
  LDR R2, =GPIOA_ODR

  turnON:
  STR R1, [R2]
  turnOFF:
  STR R0, [R2]
  delayDone:
  B turnON 

 Second:

  LDR R0, =0x00000000
  LDR R1, =0x0000FFFF
  LDR R2, =GPIOA_ODR

  turnON:
  STR R1, [R2]
  B turnOFF
  turnOFF:
  STR R0, [R2]
  delayDone:
  B turnON 

The first example cause next output:
The red plot is the bitbang output, the blue plot is the SYSCLK from MSO

 The second example cause next output:
The red plot is the bitbang output, the blue plot is the SYSCLK from MSO

The first example gives a 2-2-2-4 clock bitbang sequence. Respectively we can say that
the best loop execution time is 4 cycles,
the mean loop execution time is 5 cycles,
the worst loop execution time is 6 cycles.
And the second example gives a 3-3-3-3 clock bitbang sequence. Respectively we can say that
the best loop execution time is 6 cycles,
the mean loop execution time is 6 cycles,
the worst loop execution time is 6 cycles.
I understand that this behaviour cause is the branch prediction.
Is it possible to calculate/simulate for some code (such as a macro or a procedure) writed with assembly instructions the worst execution time?

[Tesla DeLorean]

You should be able to use DWT CYCCNT to count cycles. You could perhaps post-process the .LST file and have it estimate instruction timings.

The F1 doesn't cache the FLASH. Trying to remember if the Write Buffers on the CM3 are one or two words deep, the AHB write is of the order of four cycles as I recall.

Could try aligning the branch targets, and unwinding the loops a little.