est: delete the parasitic network when its dbNet is destroyed

[dsengupta0628]

Summary

Fixes #11192. Fixes the crashes reported in #11178.

Problem

sta::Net* is the dbNet* (dbNetwork::dbToSta is a reinterpret_cast), and ConcreteParasitics keys parasitic_network_map_ by it. Destroying a dbNet only disconnected the parasitic network's pin nodes — the network itself stayed annotated on the dead net, because EstimateParasitics::eraseParasitics() only erased the net from parasitics_invalid_.

odb recycles dbNet slots, so the next net created lands on the same sta::Net* and inherits a network that has wire nodes but no node for its driver pin.

flowchart TD
    A["repair_timing edits netlist<br/>dbNet::destroy()"] --> B["inDbNetDestroy() fires"]
    B --> C["eraseParasitics() clears only<br/>the parasitics_invalid_ marker"]
    C --> D["Network stays in parasitic_network_map_,<br/>pin nodes already disconnected"]
    D --> E["odb recycles the dbNet slot"]
    E --> F["New net = same sta::Net*<br/>inherits a driver-less network"]
    F --> G{"Delay calculator"}
    G -->|table| H["Reduces it to pi-Elmore.<br/>Silently wrong, no crash."]
    G -->|prima| I["No driver node<br/>node_count_ == 0"]
    I --> J["Zero-node circuit:<br/>empty threshold_times_,<br/>singular matrix → SIGSEGV"]

Loading

Only CCS/prima runs crash. The table calculator reduces the stale network and never inspects individual nodes, so it just uses wrong numbers.

Both stacks in #11178 reach this through the same middle section — the incremental router re-entering the timer — differing only in the resizer entry point:

rsz::Resizer::repairSetup -> SetupLegacyPolicy::repairEndpoint  \
rsz::Resizer::repairHold  -> RepairHold::makeHoldDelay          /
  -> est::EstimateParasitics::updateParasitics
    -> grt::GlobalRouter::updateDirtyRoutes -> FastRouteCore::layerAssignment
      -> updateSlacks -> sta::Sta::slack -> ... -> PrimaDelayCalc::simulate1

Why this fix is in OpenROAD, not OpenSTA

• The premise is OpenROAD's. sta::Net*-is-dbNet* is a dbSta invention and slot recycling is odb's. Upstream OpenSTA has no notion of one net pointer later meaning a different net; teaching it that would export an OpenROAD-specific representation upstream.
• est owns these annotations. They are produced by estimate_parasitics, not read from SPEF. est already had a net-destroy hook, eraseParasitics(), meant for exactly this cleanup — it just did too little. sta::Parasitics is a passive store.

Fix

eraseParasitics() now deletes the annotation across all scenes, min and max.

It calls deleteParasiticNetwork() and not deleteParasitics(): the latter resolves drivers(net), and dbNet::destroy() has already disconnected every iterm by the time inDbNetDestroy() fires, so that lookup finds no driver and re-populates dbNetwork's driver cache with an empty entry keyed on a dbNet that is about to be freed.

Test

src/est/test/prima_net_recycle - 6 instances, one net destroyed and rebuilt with an identical pin set. Crashes before the fix in PrimaDelayCalc::measureThresholds-->simulate1; passes after this fix.

Two conditions needed to be met, and hence commented in the test:

• prima must be selected before estimate_parasitics, or the default calculator reduces each network to pi-Elmore and deletes it, leaving nothing to inherit.
• the driven net needs >= 1.44 fF (the CCS tables' lowest total_output_net_capacitance), or PrimaDelayCalc::checkArgs() falls back to the table calculator and the test is vacuous. Hence the extra inverter loads.

Adds test/asap7/asap7_small_ccs.lib.gz with public filegroups //test:asap7_ccs_data and //test:asap7_data. OpenROAD had no asap7 CCS liberty outside the src/sta submodule, and the shared asap7 libs are NLDM only, so prima could not be exercised from a module test before this.

Type of Change

• Bug fix

Impact

[How does this change the tool's behavior?]

Verification

• I have verified that the local build succeeds (./etc/Build.sh).
• I have run the relevant tests and they pass.
• My code follows the repository's formatting guidelines.
• I have included tests to prevent regressions.
• I have signed my commits (DCO).

Related Issues

#11178
#11192

[gemini-code-assist]

Code Review

This pull request ensures that destroying a dbNet properly deletes its associated parasitic network annotation. Previously, dbNet::destroy() only disconnected pin nodes, leaving a driver-less network on the dead net, which could then be inherited by a new net when the dbNet slot was recycled. The fix explicitly deletes the parasitic network in EstimateParasitics::eraseParasitics. A new regression test prima_net_recycle has been added to verify this behavior, along with its corresponding test files and build targets. There are no review comments, and the implementation is clean and well-tested.

[eder-matheus]

@dsengupta0628 Does it need a secure-ci run? You mentioned wrong values being calculated on the NLDM path. Perhaps this could affect some designs.

[dsengupta0628]

@dsengupta0628 Does it need a secure-ci run? You mentioned wrong values being calculated on the NLDM path. Perhaps this could affect some designs.

Hi Eder. Yes I plan to run the full ORFS. I didn’t do it yet as it was very late last night by the time I filed this PR :)

[dsengupta0628]

@dsengupta0628 Does it need a secure-ci run? You mentioned wrong values being calculated on the NLDM path. Perhaps this could affect some designs.

The-OpenROAD-Project/OpenROAD-flow-scripts#4472 shows all clean
secure ci http://secure-ci:8080/job/SB/job/secure-est_fix/ looks passed