Call speed (penalty) viper vs. asm_thumb

[GitHubsSilverBullet]

It seems that a viper function has significantly less call-overhead than an asm_thumb function.
What could be the reason for this behavior? Is there any good explanation for this?
Do I have a mistake massaged into the time taking code?
A difference of 1.4µs translates to an additional 280 CPU cycles and that's a lot!

#!/micropython
# -*- coding: UTF-8 -*-
# vim: fileencoding=utf-8: ts=4: sw=4: expandtab:

from array import array
from time import ticks_us, ticks_diff
from machine import freq as cpufreq

@micropython.viper
def vip_cb():           # empty viper function takes ~1.7 µs @ 200 MHz
    return

@micropython.asm_thumb
def asm_cb():           # empty asm_thumb function takes ~3.1 µs @ 200 MHz
    mov     (r0, r0)    # takes only 5ns

@micropython.native
def timetest():
    N = 1_000
    timings = array('I', range(N))
    OVERHEAD = .600        # measured with both functions commented out
    µticks = ticks_us
    for index in range(N):
        t0 = µticks()
        # vip_cb()
        asm_cb()
        t1 = µticks()
        timings[index] = ticks_diff(t1,t0)

    print(f'average: {(sum(timings)-N*OVERHEAD) / N:.3f} µs', end='   ')
    print(f'fastest: {min(timings)} µs (w/o compensation)', end='   ')
    print(f'slowest: {max(timings)} µs (w/o compensation)')

cpufreq(200_000_000)
for i in range(10):
    timetest()

[robert-hh]

Repeating your test on a RP2040, I get ~1.75µs for ASM and ~1.62 for Viper.

[GitHubsSilverBullet]

Hi Robert,
that's interesting, thanks for the test. Seeing your results, I remembered that I was still using 1.24.
Now that I flashed to 1.25.0, the numbers have changed (similar to yours, here 1.66µs vs 1.92µs)
Still about a 260ns (about 54 CPU cycles) difference. It could be that Micropython saves and restores registers as if there's no tomorrow and that there is at least an addition call layer somewhere hidden.

At least my original routine is now able to answer ~70000 IRQs / second. Some finetuning and a moderate increase in CPU clock will suffice.