Files

123 lines
5.5 KiB
TeX

%% SCRAP: architecture/03-architecture/hal/migration-plan
%% SOURCE: docs/working/architecture/03-architecture/hal/migration-plan.md
%% STATUS: WORKING
%% FITS: dev-guide/ch-platform
%% EDITORIAL: lifted — prose rewritten to press voice
\section{HAL Migration Plan}
This section sets out the staged refactor that moves the existing StarForth
code base onto the Hardware Abstraction Layer. The migration is incremental by
design: one subsystem at a time, with a passing build and a passing test suite
required after every step, and platform parity maintained throughout. The
governing constraint is zero functional regression --- all regression tests must
continue to pass and the runtime's zero algorithmic variance must be preserved.
\subsection{Phase 1: Define the Interfaces}
The first phase writes the HAL headers and nothing else. The header directory is
created, the six interface headers are authored against the published contract,
the build is taught to find them, and a headers-only compile confirms the tree
is sound. No interpreter code changes, so the risk is low.
\subsection{Phase 2: Implement the Linux HAL}
The reference platform is implemented next. An audit locates existing POSIX
calls --- timing, allocation, and console I/O scattered through the tree --- and
each is folded behind a HAL function. Timing wraps \texttt{clock\_gettime};
allocation wraps \texttt{malloc} and \texttt{free} with mandatory
zero-initialization; the console wraps standard I/O; interrupts are emulated with
signals; panic writes to standard error and aborts. The VM still builds against
its old paths at the end of this phase --- the HAL exists but is not yet called.
\subsection{Phase 3: Migrate the VM Core}
The interpreter core, the external API, the dictionary allocator, and the VM
header now adopt the HAL. Direct allocation and timing calls are replaced with
their HAL equivalents, HAL initialization is added to startup in dependency
order, and platform conditionals are deleted in favor of single portable calls.
\begin{lstlisting}[language=C]
/* before */
struct timespec ts;
clock_gettime(CLOCK_MONOTONIC, &ts);
uint64_t now = ts.tv_sec * 1000000000ULL + ts.tv_nsec;
/* after */
uint64_t now = hal_time_now_ns();
\end{lstlisting}
This is the highest-risk phase. Compilation, link, and initialization-order
errors are expected and are fixed incrementally. The phase is complete when the
core builds against the HAL, the full test suite passes, and the interpreter
contains no platform-specific code.
\subsection{Phase 4: Migrate the Physics Subsystems}
The physics subsystems --- heat tracking, the rolling window, the hot-words
cache, the pipelining metrics, the inference engine, and the heartbeat --- adopt
the HAL for all timing. The most consequential change is moving the heartbeat
from a thread-and-sleep loop onto a HAL periodic timer whose callback runs in
interrupt context. That callback must be ISR-safe: no allocation, no blocking
I/O, only lock-free ring-buffer updates. Determinism is the acceptance gate here.
A design-of-experiments run before and after the change must produce identical
output, confirming zero algorithmic variance survives the migration.
\subsection{Phase 5: Migrate the REPL and I/O Words}
The read-eval-print loop and the I/O words (\texttt{EMIT}, \texttt{KEY}, and
their relatives) move from standard I/O onto the console HAL. This phase is low
risk; the success test is that the REPL behaves exactly as before.
\subsection{Phase 6: Implement the L4Re HAL (Optional)}
Implementing a second hosted platform is optional but strongly recommended,
because it proves the abstraction actually abstracts. The L4Re HAL binds time to
the kernel clock, memory to dataspaces, console to the L4Re console service, and
interrupts to IRQ objects. Success means the VM runs on L4Re without source
changes and passes its tests there.
\subsection{Phase 7: Validate Determinism}
The final phase is validation. The design-of-experiments harness is run
repeatedly on Linux; every run must be byte-identical, and identical to the
pre-migration baseline, demonstrating the HAL added no algorithmic overhead. The
full test suite must pass and the benchmark must show under five percent
performance delta from baseline.
\subsection{Rollback and Cleanup}
The migration proceeds on a dedicated branch with one commit per passing phase,
so any failed phase reverts cleanly. Where a clean cut-over is risky, a feature
flag can gate the HAL path against the legacy path. After validation, superseded
platform code is removed, documentation is updated, and build artifacts for all
platforms are ignored.
\subsection{Timeline}
\begin{table}[h]
\centering
\begin{tabular}{lll}
\toprule
Phase & Duration & Cumulative \\
\midrule
1. Define interfaces & 1--2 days & 1--2 days \\
2. Linux HAL & 3--5 days & 4--7 days \\
3. VM core & 2--3 days & 6--10 days \\
4. Physics subsystems & 2--3 days & 8--13 days \\
5. REPL and I/O words & 1--2 days & 9--15 days \\
6. L4Re HAL (optional) & 3--5 days & 12--20 days \\
7. Validation & 1 day & 13--21 days \\
\bottomrule
\end{tabular}
\caption{Estimated migration schedule, two to four weeks depending on whether
the optional L4Re phase is included.}
\end{table}
After a successful migration the next work is the StarKernel platform itself: the
UEFI boot loader, the freestanding HAL, and the path to an \texttt{ok} prompt on
QEMU with OVMF.
%% TODO(bob): the source lists migration as a forward plan; confirm which phases
%% are now complete so STATUS can be sharpened to HISTORICAL for finished work.