Measured stack-heavy work against the memory holding the stack

[fdesbiens]

#635 found BTCM worth ~7.4% on a context switch with no determinism advantage, and explained why: a switch saves sixteen registers, roughly one cache line, so the stack's cache state has almost nothing to contribute. It named the interesting case — work with a large stack working set — said it had not been measured, and said it should not be assumed.

This measures it, and the answer inverts the earlier one.

Method

deep_touch recurses 24 frames, writing a frame on the way down and reading it on the way up, so the working set is the whole descent rather than one frame. The cache is cleaned and invalidated before each sample, so every descent starts cold.

Two threads, one stack in BTCM and one in DRAM0. No partner threads and no relinquish — the timed region is entirely within one thread, which removes the confound that made #635's numbers hard to read, where the timed region included the partner's execution.

Result

Reproducible across runs:

stack in	min	mean	max	spread
BTCM	42,868	43,031	44,890	2,020
DRAM0	42,982	43,291	49,842	6,860

The mean is the same to within 0.6%. The worst case is 10% lower for BTCM and the spread is 3.4× tighter.

No sample in either configuration exceeded twice the minimum (interrupted 0 of 64), so these maxima are the workload rather than a timer tick — the mistake that produced a false 16× jitter claim in #635, and the reason the interrupted count is printed.

Two measurements, opposite conclusions, both true

	footprint	throughput	determinism
context switch (#635)	~1 cache line, every switch	+7.4%	none
stack-heavy work (this)	24 frames, once	+0.6%	10% worst case, 3.4× spread

The reason is that this workload is compute-bound at the optimisation level this BSP builds at — 43,000 cycles for 24 frames is dominated by call and loop overhead, so line fills are a few percent of the total and barely move the mean. What they do is vary, and that variance is what a bank with no cache in the path removes.

For anyone citing this

The determinism argument for TCM does hold here, but it is worth 10% of worst case and a threefold narrowing of spread — not an order of magnitude. Cite those numbers, not a larger claim.

Verification

All three targets build clean with no warnings; the demo passes and the boot image still passes six of six probes. BTCM usage is 10,240 of 16,384 bytes.