dcalc: CCS ReceiverModel accessors, converged-Ceff exposure, reportGateDelay crash fix, receiver-cap Ceff solve

[saurav-fermions]

Summary

Four CCS (Composite Current Source) delay-calc changes: three purely-additive read-only accessors, one Signal-11 crash fix, and one flag-gated (default-OFF) enhancement to the ccs_ceff effective-capacitance solve. With flags off / accessors not called, timing is byte-identical to upstream.

1. Read-only ReceiverModel capacitance accessors (TableModel.{hh,cc})

Adds hasCapacitanceModel() / capacitance() to evaluate the CCS receiver-capacitance tables OpenSTA already parses into capacitance_models_ but never consumes. Used only by an additive external diagnostic; the NLDM delay path does not call them → flag-off byte-identical.

2. Expose converged CCS effective capacitance (ArcDelayCalc.hh, CcsCeffDelayCalc.*)

Additive read-only effectiveCapacitance() on ArcDelayCalc (base returns a negative sentinel), overridden by CcsCeffDelayCalc to return the region-0 effective capacitance the most recent gateDelay() converged to (negative sentinel on NLDM fallback). Purely additive — no existing delay value changes.

3. reportGateDelay crash fix (CcsCeffDelayCalc.cc)

reportGateDelay() can be called directly (e.g. by report_dcalc) without a preceding gateDelay() for the same arc. It used the stale drvr_pin_ member and an unset parasitics_ member, dereferencing bad pointers and crashing (Signal 11). The fix resolves parasitics_ from the scene and uses the drvr_pin parameter, matching gateDelay().

4. Receiver capacitance in ccs_ceff Ceff solve (CcsCeffDelayCalc.*, flag-gated, default OFF)

When the receiver model is enabled, the per-region far-cap (c1) contribution of any load pin carrying a CCS receiver_capacitance model is swapped from its constant NLDM pin cap to the library's region-dependent receiver cap (Cr1 in the Miller/transition region below the receiver threshold, Cr2 in the settled region). Pins without a receiver model keep the constant NLDM contribution (graceful fallback). Default OFF → byte-identical to prior ccs_ceff.

Testing

• Full OpenSTA ctest suite: 6083/6084 pass. The single failure (tcl.power.vcd_detailed) is a pre-existing last-digit floating-point rounding diff in the power module that fails identically on unmodified master (unrelated to this change).
• set_ccs_delay_calc regressions pass; default-OFF behavior verified byte-identical on both the global-NLDM-default and ccs_ceff-receiver-off axes.

Notes

• dcalc/DelayCalc.i is a SWIG interface; clang-format intentionally not run on src/sta/* per project convention.
• Items 1–3 are additive/bug-fix; item 4 is default-off. No behavior change for existing users.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[gemini-code-assist]

Code Review

This pull request implements support for the CCS receiver-capacitance model in the Ceff delay calculations, introducing a process-global toggle, region-dependent receiver capacitance evaluations, and exposing the converged effective capacitance. The review feedback identifies three critical safety issues: an uninitialized member variable drvr_pin_ in gateDelay(), a potential null pointer dereference when accessing parasitics_ in reportGateDelay(), and a missing null check for the rf parameter in capacitanceModel().

[gemini-code-assist]

The member variable drvr_pin_ is never initialized or assigned in gateDelay(). However, it is subsequently used on line 161 (pinPvt(drvr_pin_, scene, min_max)) and in watchWaveform() (line 617). This results in undefined behavior or crashes due to dereferencing an uninitialized or stale pointer. Assign drvr_pin_ = drvr_pin; at the beginning of gateDelay() to resolve this.

  drvr_pin_ = drvr_pin;
  parasitics_ = scene->parasitics(min_max);
  parasitic_ = parasitic;
  output_waveforms_ = nullptr;
  last_ceff_ = -1.0;  // OpenROAD-fork: ccs-delay2 -- reset; set on CCS path.

[gemini-code-assist]

If scene->parasitics(min_max) returns nullptr (e.g., if parasitics are not defined or SPEF is not loaded for the current corner), dereferencing parasitics_ on lines 796, 799, and 804 will cause a crash (Signal 11). Add a null check for parasitics_ before accessing its members.

  parasitics_ = scene->parasitics(min_max);
  if (parasitics_ && parasitic && !parasitics_->isPiElmore(parasitic)) {
    // Use the drvr_pin PARAMETER, not the stale drvr_pin_ member.
    pi_elmore =
        parasitics_->reduceToPiElmore(parasitic, drvr_pin, rf, scene, min_max);
  }
  std::string report =
      table_dcalc_->reportGateDelay(drvr_pin, arc, in_slew, load_cap, pi_elmore,
                                    load_pin_index_map, scene, min_max, digits);
  if (parasitics_) {
    parasitics_->deleteDrvrReducedParasitics(drvr_pin);  // OpenROAD-fork: ccs-delay-path
  }

[gemini-code-assist]

If the rf parameter is passed as nullptr, calling rf->index() will dereference a null pointer and crash. Adding a defensive null check at the beginning of capacitanceModel() will safely protect this method and its callers (hasCapacitanceModel() and capacitance()).

{
  if (rf == nullptr)
    return nullptr;
  const size_t idx = segment * RiseFall::index_count + rf->index();

[saurav-fermions]

Thanks:

• reportGateDelay parasitics null (CcsCeffDelayCalc.cc:~800): added — if scene->parasitics(min_max) is null (parasitics not loaded for the corner) we skip the pi/Elmore reduction and cleanup instead of dereferencing.
• ReceiverModel::capacitanceModel null RiseFall (TableModel.cc): returns nullptr on a null rf instead of calling rf->index().

On the critical note about drvr_pin_ in the POCV block (CcsCeffDelayCalc.cc:~161): that pinPvt(drvr_pin_, …) usage is pre-existing code (it's on master, unchanged by this PR — see the same line there), not introduced here. This PR's crash fix is specifically the reportGateDelay() path, which previously used the stale drvr_pin_ member and an unset parasitics_; it now uses the drvr_pin parameter and resolves parasitics_ from the scene. If the pre-existing POCV drvr_pin_ path warrants hardening, that's worth a separate dedicated fix. ccs/dcalc suite passes (5/5).

[dsengupta0628]

Please push any core dcalc changes in upstream OpenSTA (parallaxsw/OpenSTA repo). We want to ensure the core algorithms are the same in the timing engine for ease of development of the upstream OpenSTA too. Only if this is very specific to OpenROAD alone, would it be okay to have in this fork and not upstream. Could you please provide a justification for this solely to be in OpenROAD/OpenSTA? And if and when you push this upstream, please add a testcase in the sta/test directory. That is the public test area of OpenSTA we have access to.