Running deterministic from Flash

[Traumflug]

Just saw your awesome talk on 33C3. There you claim running from Flash isn't clock cycle deterministic, because there's a prefetcher and one would never know when this kicks in.

When implementing an interruptible delay loop for Teacup Firmware I ran into a similar problem. Depending on where the compiler placed the code, the loop would take a longer or shorter execution time. The solution I found is to align code to a 16 byte boundary. This removes the prefetcher from the picture or at least makes it deterministic. Gcc/binutils even features an instruction to enforce such alignment, .balign.

The implementation taking advantage of .balign can be found here: https://github.com/Traumflug/Teacup_Firmware/blob/master/delay-arm.c

[xobs]

Hi! Thank you for your comment.

My experiences with "non-deterministic flash" consisted of the MCU spending a variable number of cycles in the main loop -- anywhere from 24 to 28 cycles per loop. That is, I managed to eliminate four instructions, and it would work sometimes but not others.

I suspect it has to do with bit stuffing, where the codepath is not constant, and instead depends on the data streams.