variable_persistence_examples

Variable-persistence examples (version8 / Enhancement-7)

Self-contained correctness example for the Enhancement-7 fix that gives ordinary Verilog-A real/integer analog-block variables genuine persistence across evaluations, covering DC, AC, and Transient analysis. Everything here uses the version8 toolchain:

• compiler : ../OpenVAF-master-20260610/target/opt/openvaf-r
• simulator: ../ngspice-46/build/src/ngspice

See ../Enhancement-7.md for the full writeup. This is the deeper of the two Enhancement-7 findings: while fixing @(initial_step) gating (see ../initial_step_examples/), testing against a real simulation revealed that ordinary variables were being silently reset to their default value on every single evaluation, not just the first — with zero event-control involved. Root cause: two todo!("hidden state") panics sitting in the pre-existing codebase (the read side was scaffolded but never backed by real storage) and an insert_var_init pass that unconditionally replaced every read of a variable's "previous value" with its declared initializer.

The model: the minimal accumulator

module persist_demo(in, out);
    inout in, out;
    electrical in, out;
    real accum;
    analog begin
        accum = accum + 1.0;
        V(out) <+ accum;
    end
endmodule

No parameters, no event-control statements at all — this isolates the persistence fix from the @(initial_step) gating fix (covered separately in ../initial_step_examples/). Before the fix, accum read back 0.0 every evaluation regardless of what any earlier evaluation had computed, so V(out) stayed flat at exactly 1.0 forever, no matter how long the transient ran.

accum is purely a diagnostic evaluation counter: it increments once per simulator evaluation (including intermediate Newton iterations), not once per physical timepoint, so it is intentionally not a meaningful function of V(in) — see the DC/AC notes below.

Results

Analysis	Sweep	Expected	Observed
DC	V(in) −1V to 1V	not a real function of V(in) — a roughly-monotonic curve driven by Newton-iteration count per sweep point, included for honesty about what this operator does and doesn't do	matches (dc_plot.png) — same "documented negative result" spirit as last_crossing_examples/'s DC plot in Enhancement-6
AC	1Hz–100MHz	~zero small-signal gain (accum's recurrence has zero dependence on V(in))	flat ~zero (ac_plot.png)
Transient	100µs, DC input	the key demo: a clean, sustained, roughly linear ramp over the whole transient, proving accum is never silently reset	matches (tran_plot.png)

Layout

variable_persistence_examples/
  persist_demo.va    minimal accumulator, no parameters/event-control
  persist_demo.osdi  compiled with version8 openvaf-r
  dc_sim.cir / ac_sim.cir / tran_sim.cir
  dc_result.txt / ac_result.txt / tran_result.txt
  dc_plot.png / ac_plot.png / tran_plot.png

Reproduce

OPENVAF=../OpenVAF-master-20260610/target/opt/openvaf-r
NGSPICE=../ngspice-46/build/src/ngspice

$OPENVAF persist_demo.va -o persist_demo.osdi
$NGSPICE -b dc_sim.cir
$NGSPICE -b ac_sim.cir
$NGSPICE -b tran_sim.cir