Slice operation on memoryview causes allocation

[peterhinch]

The docs lead one to believe that slicing a memoryview should be allocation free, but this seems not to be the case.

This arose in this forum thread. The user has a timer ISR which uses SPI to read first the number of bytes to read, followed by a subsequent SPI read of the payload. I suggested a memoryview, but this does not work as the slice operation causes an allocation.

I replicated the issue with this sample which can be pasted to a Pyboard:

from machine import SPI
from pyb import Timer
import micropython
import time
micropython.alloc_emergency_exception_buf(100)

s = SPI(1)
buf = bytearray(20)
mv = memoryview(buf)
tim = Timer(1)

def cb(t):
    #s.readinto(mv)  # Works
    s.readinto(mv[:5])  # Memory error

tim.init(freq=10, callback=cb)
while True:
    print(buf)
    time.sleep(1)

[jimmo]

The docs lead one to believe...

I'm guessing it's this line that probably gives this impression.

func(mv[30:2000]) # a pointer to memory is passed

I think it's worth updating the docs to clarify that slicing a memoryview creates a new memoryview (as well as requiring a heap allocation for the slice) (ref #4244 for more details).

Not sure what the solution to the underlying issue is as it will likely require extending the memoryview API (e.g. to add a way to adjust an existing memoryview as discussed in #4244 and elsewhere) but this will break CPython compatibility... perhaps we need to raise this with python-dev.

[peterhinch]

Apologies - my memory is quite hopeless these days.

In penance I will raise a docs PR :)