[ds54e]

Fixes #991

always_comb and always @* missed changes on a whole read of an unpacked struct that is a continuous copy of another and is folded into it by the identity port-net collapse (assign q = p; always_comb c = q;). The testcase is in the issue.

Cause. A whole-struct read depends on the struct's member signals. The map from a struct name to those signals (unpacked_struct_sens) was built in Simulator::new before its identity port-net collapse, and compile can run the collapse again. The collapse points the member names of a continuous copy at its source, so the procedures waited on the copy's original member signals, which nothing writes after the collapse.

Change (code and tests in separate commits).

• The map is built right after the last collapse call in compile, and no longer in Simulator::new. Its only reader, build_comb_entries, runs after that point.
• The map is still built once per compile, so nothing is added to event processing apart from the missing wake-ups.
• PREPARED_COMB_MAGIC goes from XZCMB030 to XZCMB031, as for An always_comb that writes one packed element and reads another never wakes (§9.2.2.2.1) #844, because caches made before this change hold the stale edges.

Testing

• New tests/scheduling/unpacked_struct_copy_chain.rs:

  • copy chains of one to three steps, read by always_comb and always @*;
  • nested and real members changed one at a time;
  • copies inside a child instance and through an output port;
  • clocked and initial-block readers as controls (IEEE 1800-2023 §9.2.2.2, §9.4.2.2 and §7.2).

  With the src/ change reverted, three of the four tests fail; the clocked/initial control passes either way.

• All nine generated test programs pass slang 12.1 with no diagnostics.

• Full suites on this branch:

  • xezim: 6,336 passed, 0 failed, 1 ignored (the ignored test is also ignored on main);
  • xezim-core (unchanged): 345 passed.

• Same output on a struct-nettype charge-pump regulator model. Three alternating runs on a shared machine give medians of 100.5 s on main and 101.2 s with this change, which is within noise. WSL has no hardware counters, so there is no perf stat count. The change is outside the event loop.

🤖 Generated with Claude Code