[oharboe]

The flow decides which parasitics to time with in four places: open.tcl, generate_abstract.tcl, write_endpoint_slack and final_outputs.tcl. Each decides from the stage number, grt::have_routes and whether a .spef file exists. Those rules are easy to get wrong. #4688 fixes one copy: an abstract written from global route was timed with placement parasitics, 6.9 ns instead of 1.8 ns on one arc.

Gemini Code Assist's review of #4688 is a concrete example. It suggested reading $stem.spef from stage 5 on whenever the file exists (comment). It reads as a sensible improvement, and it's wrong:

• Only the final stage writes a SPEF, so before final that branch can only match a leftover file.
• Re-running a flow from place or global route leaves the previous final's SPEF in RESULTS_DIR. The re-run stage's timing would silently use parasitics from a different netlist, and nothing in the log would show it.

The rules are subtle enough that an AI reviewer, given the code, proposed a change that turns a loud failure into a silent wrong number. A reviewer reading only the diff, human or bot, couldn't have caught it either: the decision is spread over four scripts and depends on which files a previous run left behind. So the rules belong somewhere they can't be gotten wrong.

Whether a file is on disk is not a valid test, and a stage number is a second copy of a fact OpenROAD already has. So:

• OpenROAD decides what: a flagless estimate_parasitics uses global routes if they exist, otherwise placement. It fails loudly on an unplaced design, zero wire RC, or a congested route that wasn't accepted, and does nothing once read_spef has run.
• The flow decides when: log_cmd estimate_parasitics wherever it wants to pay for an estimate (the log shows the time), and an unconditional read_spef exactly where its configuration says a SPEF was made. final_extracts_parasitics says so from RCX_RULES and SKIP_DETAILED_ROUTE. A missing SPEF fails loudly, and a stale one is never read.

On gcd (asap7 and nangate45), the flow with this change and the OpenROAD one is byte-identical to today's at every stage: the .odb files from place through final, the abstract's .lib and .lef, and 6_final.spef apart from its *DATE line. What changes is what happens to a SPEF that is missing or doesn't belong to the design:

generate_abstract on gcd's results	today	with this change
6_final, SPEF missing	succeeds without a word, estimates	fails: cannot read file .../6_final.spef
6_final, detailed route skipped, leftover SPEF	reads the leftover	estimates from the global routes
6_final, extracted SPEF	reads it	reads it

Changes in behaviour:

• With RCX configured, a missing 6_final.spef fails open and generate_abstract instead of falling back to an estimate.
• A global route that is congested and wasn't accepted fails estimate_parasitics instead of being timed without parasitics. To look at one, open it with GUI_TIMING=0. A failed route written with GENERATE_ARTIFACTS_ON_FAILURE gets no endpoint-slack snapshot.
• generate_abstract from global route or later is timed with global-route parasitics instead of placement parasitics. This supersedes generate_abstract: time a global-route abstract with the route's parasitics #4688's change.
• The endpoint-slack snapshot records where OpenROAD's parasitics came from: 6_final says spef instead of rcx, and 3_1_place_gp_skip_io with unplaced pins says none instead of placement.

SKIP_DETAILED_ROUTE now reaches generate_abstract, which needs it to know whether 6_final has a SPEF.

Integration test: The-OpenROAD-Project/bazel-orfs#1284, which runs ORFS's generate_abstract.tcl on copies of gcd's results (missing, leftover and extracted SPEF, and an abstract from global route) on asap7 and nangate45.

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request simplifies and standardizes parasitics estimation across various flow scripts by replacing explicit estimation flags (such as -placement or -global_routing) with a bare estimate_parasitics call, allowing OpenROAD to automatically determine the correct source based on the design's state. It also introduces the final_extracts_parasitics helper function, updates documentation, and extends the SKIP_DETAILED_ROUTE variable to the generate_abstract stage. There are no review comments, so I have no feedback to provide.

[openroad-ci]

🔍 QoR check

Metrics reflect the PR merge build — i.e. what will land on the target branch.

Commit 28e4347 · Jenkins build #13 · Baseline: build · View build on dashboard

62 design(s) checked — 0 with regression(s), 1 without a comparable baseline.

[oharboe]

@eder-matheus 🥺 ?

tools/OpenROAD must be bumped, which I will let someone else do.

If you approve + automerge, it will go in when pr-merge goes green.

[oharboe]

@maliberty @eder-matheus pr-merge is red, but that's just a QoR problem that will fix itself when any ORFS pr is merged. So... this PR is as good as any.

Merge?

[maliberty]

why not an elseif - is there a case where you want to both read_spef and estimate parasitics?

[oharboe]

No case wants both. It is an elseif now, in orfs_load_parasitics in util.tcl (92bda58); the two ifs only worked because estimate_parasitics does nothing after read_spef.

[maliberty]

same elseif question

[oharboe]

No case wants both. It is an elseif now, in orfs_load_parasitics in util.tcl (92bda58); the two ifs only worked because estimate_parasitics does nothing after read_spef.

[maliberty]

Codex (trimmed)

The other findings:

• Undeclared input: generate_abstract.tcl now always reads 6_final.spef when RCX applies, but the generate_abstract target in the Makefile doesn't declare that file. In bazel-orfs, which only provides declared inputs, the step would fail.
• Copied logic: the same "read SPEF, otherwise estimate if placed" block appears in open.tcl, generate_abstract.tcl and util.tcl. TclContract.md says logic needed in two places becomes a proc in util.tcl.
• Speed:
  • The final stage on platforms without RCX estimates twice in a row with identical results.
  • Each guarded estimate walks the full netlist about four times.
• Wrong label in reports: util.tcl:379 reports the parasitics source from the last estimate run, which can be out of date when the estimate is skipped.
• Doc mismatch: in EndpointSlackTracking.md, the 6_final row leaves out the SKIP_DETAILED_ROUTE=1 fallback.

[maliberty]

needs a merge of master to pick up a later OR version with: That commit includes OR #11788 (merge commit 99bd6268d8, "est-parasitics-auto").

[oharboe]

Merged master, which brings OpenROAD 98c6045c, which includes #11788. New commits, one per finding:

• Undeclared input: e8398b5. generate_abstract lists $(RESULTS_DIR)/6_final.spef, a result of 6_report like 6_final.def. Without RCX the file is not written and make goes on without it, so the graph is the same for every design. (bazel-orfs already declares 6_final.spef as a final-stage result; the abstract tests in bazel-orfs ran with it.)

• Copied logic: 92bda58. orfs_load_parasitics <step> in util.tcl is the one copy, as an if/elseif; open.tcl, generate_abstract.tcl and write_endpoint_slack call it.

• Speed, final estimates twice: 92bda58. final_outputs.tcl loads its parasitics itself and calls write_endpoint_slack 6_final 0.

• Speed, netlist walks: one of the four is the flow's (grt::design_is_placed), now once per call; the other three are inside estimate_parasitics (designParasiticsSrc, haveRoutes, haveCongestedRoutes). Measured at global route on asap7, three runs each:

  design	instances	grt::design_is_placed	estimate_parasitics
  aes_cipher_top	12,446	2.0–2.5 ms	77–89 ms
  jpeg_encoder	58,595	10–12 ms	299–339 ms

• Wrong label: 92bda58. The snapshot records what orfs_load_parasitics returns: none when the design is not placed, else est::parasitics_source right after the read_spef or estimate it ran.

• Doc mismatch: 3da44b6.

gcd asap7 with this change against without it: 2_floorplan, 3_place, 4_cts, 5_1_grt, 5_route and 6_final .odb, the abstract's .lib and .lef byte-identical, 6_final.spef apart from *DATE. The asap7 cases of the bazel-orfs integration test pass (extracted SPEF read, missing SPEF fails, leftover SPEF unread, abstract from global route). The nangate45 cases were not rerun: gcd on nangate45 fails at global route in bazel-orfs (DPL-0033 after repair at 98.2% utilization) with and without this change.