The probe stands in for Broadway with StarOS's own bench loader code, so
that StarOS's H-1A bench (docs/architecture/h-1a.md, D-H1A-7) runs whole:
its Starlet image on the ARM core, its Broadway side from StarOS's
boot/wii/loader hw0c.c, hw0d.c and h1a.c, vendored unchanged in
Tools/StarletBootstrapProbe/StarOSLoader (StarOS 480119d, stored -text so
their bytes stay StarOS's) with their contract headers, and built over a
host HAL in BroadwayHAL/. CMake computes the nine files' SHA-256s at
configure; the probe prints them when the mode starts.
As StarOS's loader around the jump: hw0c_setup() before the Starlet's
first instruction, one hw0c_step() between each two slices, the timebase
243 MHz / 4 from the Starlet's cycles, until it says 1; once the image's
trace is terminal and the bench is done, the trace, then the report
between main.c's status and END lines on stdout. If the budget runs out,
the report under SBL:TIMEOUT, and the run fails. The client reaches PPC
memory only in the bench window 0x13500000..0x13510000, by whole words:
the first access outside fails the run, and is not made. Dolphin's memory
is coherent: the run rehearses the protocol, the capture and the judge,
not the console's caches.
--cycles may reach 2^34 in this mode only (MAX_CLIENT_CYCLES), 2^30 as
before otherwise. --broadway-h1a takes neither --ipc-script nor
--stage-spk, whose behaviour is unchanged.
StarletProbeBroadwayTest: the window's bounds, the clock, Gecko's number
formats, the budgets, the built-in sources; and, each in a death test's
own process (the loader's state is global), the real loader code against
a scripted Starlet: 3000 rounds and a COHERENT record, no READY, no
Starlet END, a budget too short; a second client refused.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Yaya48andClaude Opus 5.5
f391ae6c85