OpenROAD update sta to latest from upstream 07/10

[dsengupta0628]

Summary

Update OpenSTA to latest upstream 7/10 The-OpenROAD-Project/OpenSTA#388

Type of Change

• Bug fix
• New feature

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).

[gemini-code-assist]

Code Review

This pull request updates the subproject commit for src/sta from 3ab3337d1b8f13bd15d80786ec771e1c6388d24a to 47f191d7370389a6fde18d7334fff06ba162012a. There are no review comments, and I have no feedback to provide.

[dsengupta0628]

secure ci http://secure-ci:8080/view/My/job/SB/job/secure-or_update_sta_latest_0710/
ORFS https://jenkins.openroad.tools/job/OpenROAD-flow-scripts-Private/job/secure-or_update_sta_latest_0710/

[dsengupta0628]

Does not need metric update: will close this PR The-OpenROAD-Project/OpenROAD-flow-scripts#4358

[dsengupta0628]

Global routing aborted with a hard error on some designs-hence the CI fails in Bazel build (one regoldening was needed for Ubuntu build):

[ERROR GRT-0228] Horizontal edge usage exceeds the maximum allowed. (113, 83) usage=65535 limit=3400
usage=65535 (0xFFFF) is not real congestion — it is a uint16_t underflow. The check2DEdgesUsage sanity check reports the corrupted value; the corruption originates in Graph2D::getCostNDRAware.

Root cause

For NDR nets (edgeCost > 1), getCostNDRAware adds either an inflated cost (100 × edgeCost) or the normal cost, chosen by the edge's overflow state at the time the net is routed. On rip-up it decided inflated-vs-normal from the edge's current
ndr_overflow, which other nets could have changed in the meantime.

Sequence that underflows:

1. Net A routed on an edge with capacity available → adds normal edgeCost, ndr_overflow = 0.
2. Net B later routed on the same edge, no capacity → ndr_overflow++, adds inflated cost.
3. Net A ripped up → sees ndr_overflow > 0 → subtracts the inflated 100 × edgeCost, though A only ever added edgeCost.

The over-subtraction wraps the unsigned usage (and ndr_overflow) below zero to 65535. The NDR-disable path (GRT-0273) triggers the rip-ups that expose it.

Fix

Record, per edge, whether each NDR net contributed the inflated cost, and subtract exactly that on rip-up — independent of how the edge's overflow state changed afterward. v_ndr_nets_/h_ndr_nets_ change from std::set<FrNet*> to std::map<FrNet*, bool> (the flag). Add/remove are now symmetric, so usage/ndr_overflow cannot underflow.

[dsengupta0628]

@jfgava this is the issue and the fix I did:

Problem:
graph2d.cpp + Graph2D.h (GRT-0228 underflow): getCostNDRAware added an NDR net's edge cost as inflated (100X) or normal based on the edge's overflow state at route time, but on rip-up re-derived that choice from the edge's current ndr_overflow - which other nets could have changed. A net that added the normal cost then got the inflated cost subtracted, underflowing the uint16_t usage to 65535.

Fix:
We track per-net whether each net added the inflated cost (std::set<FrNet*> --> std::map<FrNet*, bool>) and subtract exactly that on rip-up, so add/remove stay symmetric.

This fixed the Bazel crash I mentioned in Slack. And fixes the mismatch in ubuntu/bazel results

Test change:
soft_ndr_reset_cap.tcl: added suppress_message GRT 102 so the compiler-sensitive per-iteration GRT-0102 trace lines no longer appear in the log; -verbose is kept so the resource/congestion/final-result lines still get checked.

I regoldened soft_ndr_reset_cap.ok too: regenerated to match the suppressed output - 62 lines, no GRT-0102, still the 6 GRT-0273 NDR disables plus the routing-resource table, final congestion report, and WL/routed-nets result block.

[jfgava]

@jfgava this is the issue and the fix I did:

Problem: graph2d.cpp + Graph2D.h (GRT-0228 underflow): getCostNDRAware added an NDR net's edge cost as inflated (100X) or normal based on the edge's overflow state at route time, but on rip-up re-derived that choice from the edge's current ndr_overflow - which other nets could have changed. A net that added the normal cost then got the inflated cost subtracted, underflowing the uint16_t usage to 65535.

Fix: We track per-net whether each net added the inflated cost (std::set<FrNet*> --> std::map<FrNet*, bool>) and subtract exactly that on rip-up, so add/remove stay symmetric.

This fixed the Bazel crash I mentioned in Slack. And fixes the mismatch in ubuntu/bazel results

Test change: soft_ndr_reset_cap.tcl: added suppress_message GRT 102 so the compiler-sensitive per-iteration GRT-0102 trace lines no longer appear in the log; -verbose is kept so the resource/congestion/final-result lines still get checked.

I regoldened soft_ndr_reset_cap.ok too: regenerated to match the suppressed output - 62 lines, no GRT-0102, still the 6 GRT-0273 NDR disables plus the routing-resource table, final congestion report, and WL/routed-nets result block.

Thanks Deepashree!
The GRT fix is fine. I wonder if we'll have a significant impact on the CI

[dsengupta0628]

@jfgava this is the issue and the fix I did:
Problem: graph2d.cpp + Graph2D.h (GRT-0228 underflow): getCostNDRAware added an NDR net's edge cost as inflated (100X) or normal based on the edge's overflow state at route time, but on rip-up re-derived that choice from the edge's current ndr_overflow - which other nets could have changed. A net that added the normal cost then got the inflated cost subtracted, underflowing the uint16_t usage to 65535.
Fix: We track per-net whether each net added the inflated cost (std::set<FrNet*> --> std::map<FrNet*, bool>) and subtract exactly that on rip-up, so add/remove stay symmetric.
This fixed the Bazel crash I mentioned in Slack. And fixes the mismatch in ubuntu/bazel results
Test change: soft_ndr_reset_cap.tcl: added suppress_message GRT 102 so the compiler-sensitive per-iteration GRT-0102 trace lines no longer appear in the log; -verbose is kept so the resource/congestion/final-result lines still get checked.
I regoldened soft_ndr_reset_cap.ok too: regenerated to match the suppressed output - 62 lines, no GRT-0102, still the 6 GRT-0273 NDR disables plus the routing-resource table, final congestion report, and WL/routed-nets result block.

Thanks Deepashree! The GRT fix is fine. I wonder if we'll have a significant impact on the CI

Yes rerunning the CI now- But the CI was clear even without this change : http://secure-ci:8080/view/My/job/SB/job/secure-or_update_sta_latest_0710/