The impression I get is that nowadays, applications that need truly cycle-by-cycle deterministic behavior just use an FPGA.
Yeah, or, tggzzz's beloved XMOS stuff, which, fair enough, is a pretty neat concept and product line, sure.
Some of us just give more value of avoiding locked-in single-vendor
concepts, for which nothing even similar appears on any other manufacturer, no matter how good it may be.
Then again, truly needing cycle-level determinism within the whole application is exceedingly rare; I don't remember ever thinking "oh, now this xcore thing would come handy!" And all other MCUs can easily give worst-case processing within 1% or so; or +/- a few instructions, depending on details like flash accelerator / icache / branch prediction miss, synchronization across clock barriers, or maybe DMA competing for the RAM access. It is worth noting that on
desktop computing CPUs, things like these have massive impact; going from register to L1 cache might be a few times of performance loss, then to L2 cache some other hundreds of %, then to main DRAM again an order of magnitude, not to mention disk. But on microcontrollers, it's a question of if jump takes 2 or 3 cycles, not that if it takes 2 or 200 cycles. And if your critical control piece which has 1000 cycles worth of instructions has ten branches, then the question is whether it takes 1000 or maybe 1010 cycles. And if data cache is available, it has the same latency as internal RAM anyway, so people don't use it. (It has different purpose, see below.)
You have to
seriously try to program microcontrollers such that they usually work but sometimes fail
because of icache or branch prediction miss, the instruction timing jitter caused by those mechanisms is just so small.
But sure, tggzzz is completely right that these things exist, just blowing the significance out of proportion (plus adding misleading stuff, like mentioning TLBs which do not exist on microcontrollers, and does this repeatedly like any good troll no matter how many times he is proven wrong; for him, all CPUs are microcontrollers, specifically mentioning Cortex-A application processor series). Or data cache which no one uses in timing-critical microcontroller projects because they are designed for external RAM and no one uses external RAM except for special cases like, let's say, graphics.
But yeah, this exact thread is perfect for tggzzz's argument and I'm happy for him. The question was hypothetical, and XMOS is
the answer this time.