Phase 3 Backend Verification
Read <project>/reports/flow_engine.json from the normal public runner. The
Router selects the task window and flow/steps.json defines its ordered rows;
there is one engine run path. Identify steps by their current table position
and name, never by a retired public step ID.
For every selected backend row, read tool_ran, native_steps,
native_metrics, metrics, and the native paths in reports. Open each
underlying tool report and compare the actual measurement, operator and limit
with the original spec, PDK or owner requirement. Check that the tool consumed
the intended design and corner. An exit code or summary PASS does not prove
that the resulting silicon satisfies its engineering requirements.
AI review each result for physical plausibility, missing coverage and defects
that the reported number cannot describe. Keep the existing per-row review
and measured review duration with that result. A missing native value, missing
original limit or unexecuted selected tool remains NOT_MEASURED; disclose it
without fabricating a number or changing a limit. Do not claim tapeout readiness
without DRC=0, WNS≥0 and all applicable native measurements meeting their
original limits. Deferred work remains open work.
Use the actual report paths saved by the engine. The paths below illustrate the corresponding report type; they do not override the run's native paths.
Captured rule — Native Verilator thread default
Compiled GLS uses Verilator's native single-thread default. Pass an explicit
simulation_threads override through to the tool. Review threading against the
complete existing suite, its random seed, simulated duration and native results.
Keep the Icarus timing arm and the bench's required VPI visibility.
The retained PWM4 IC Step 75 Candidate gate-level simulation passed with four Icarus cases and four Verilator cases in both runs. Restoring the default from four threads to one reduced row wall time from 29,907.7 s to 483.2 s. The complete professional case used seed 1794359088 and simulated 82,215,675 ns in both runs, with execution time 29,756.2 s and 402.3 s respectively. Both rows applied 37,100 SDF delays with zero annotation errors and zero GLS failures. These are single retained observations on separate worker allocations. The paired source panel has the same 120 test IDs, changing 119 PASS / 1 FAIL to 120 PASS with NEW_FAILED=[]. The explicit two-thread native case passes in both arms.
The v1.41.3 IC final-window continuation on image 0.4.77 also passes Step 96 Final SDF gate-level replay in 543.9 s. All 36 Icarus cases across nine corners and all four Verilator cases pass, with 329,168 applied delays and zero annotation or functional errors. The professional case retains seed 1794359088 and 82,215,675 ns of simulated work. The retained electrical-capacitance failure continues through signoff aggregation and adoption.
Verification checklist
Synth quality:
- Read the actual Yosys log and native elaboration/synthesis result. Resolve inferred latches, width truncation, unresolved modules and other tool diagnostics in the design or tool before rerunning synthesis.
- AI residual only: read cells/area/gate-types and compare cell count to L8 typical_gate_count if defined (judgment, no fixed threshold in spec).
Floorplan / utilization:
- Read utilization from the native physical database. Retain the advisory 50–75% planning band; fixed-die designs may legitimately differ. A value at or below zero or above 100% requires investigation, not a fabricated placement result.
- AI residual only: confirm core_area covers all macro instances + std cells, and that the IO ring + power straps are not overlapping the core (structural judgment).
STA — multi-corner:
- Read
phase3/reports/sta.rpt— ALL slacks must be ≥0 - Critical paths shouldn't hit
set_max_delayboundary; identify margin - Hold violations are NOT acceptable even if setup is clean
- For complete sign-off: SS / TT / FF corner all checked, not just TT
- Read
DRC:
- Read
phase3/reports/drc.rpt - violations=0 PASS, ≥1 FAIL or WAIVED depending on rule
- If WAIVED with Calibre deck pointer: confirm offline Calibre run is scheduled
- Density / antenna violations are common late-stage issues
- Read
LVS:
- Read
phase3/reports/lvs.rpt(or note if WAIVED) - Net mismatches indicate floorplan / connectivity bug
- Device count mismatches indicate wrong macro picked
- Read
GDS sanity:
- Open the final GDS with the native layout tool. Check that the intended geometry was loaded and compare its measured extent and content with the required design scope.
- Read the GDS hierarchy with the native layout tool and confirm that the selected top cell is the intended root, with all required child cells resolved.
- AI residual only: sanity-check that the file size is consistent with a merged macro PA-GDS (size jumps when a hard macro is merged) — judgment, no fixed threshold.
Power / EM / IR estimates:
- if available, read
phase3/reports/power.rpt - Compare to L8 power budget if defined
- if available, read
Spot-check actions
- Open the GDS in KLayout (offscreen) and visually confirm metal stack reasonable.
- Diff this design's
area.rptandsta.rptvs prior known-good run; flag regressions. - Check critical timing paths: do they make sense (clock-to-clock through expected logic) or are they bizarre artifacts?
When to escalate
- DRC violations >0 → invoke
drc-fix - Setup slack <0 → invoke
sta-reviewtheneco-plan - Hold violations → invoke
hold-fix - LVS mismatch → invoke
lvs-triage - Utilization wildly off → re-run with adjusted
--die-um/--util
Output
Append findings to <project>/reports/phase3_verify.md. For each selected row, record its current position/name, native report and value, original limit, AI finding, and any FAIL or NOT_MEASURED obligation. Keep a measured PASS distinct from deferred or unmeasured work.
Captured rule — Native routed chip replay
Simulate the routed chip with its original cell and pad Verilog models. Keep all bench cases and every selected SDF corner on the native simulator that supports those models. Do not export a padless core or replace pad models. Use Verilator where the delivered models compile natively. If its compiler rejects a PDK pad model, record that simulation arm as not run with the native error, and obtain the complete chip verdict and metrics from Icarus. Preserve the compiler failure report; an unavailable arm never earns PASS. Identify the engine for each metric and leave unmeasured comparison populations explicit. Only restrict VPI visibility when DUT access is limited to port values.
Captured rule
A physical pad Liberty master can be a simulation-only transparent wrapper:
Sky130 sky130_ef_io__gpiov2_pad has no specify block and instantiates the
pin-for-pin native sky130_fd_io__top_gpiov2 as gpiov2_base. Bind the SDF
CELLTYPE/INSTANCE to that declared child scope. Preserve the original DIVIDER,
all conditional OUT/OE_N/PAD paths, unconditional INP_DIS paths, timing checks,
and every delay tuple; top-level INTERCONNECT endpoints stay at wrapper ports.
Do not synthesize specify arcs or turn conditional paths into unconditional ones.
On the original SHA256 v1.40.54 State, cpu_queue job sha256-pad-sdf-bind-oai
replayed both GLS rows: candidate SDF errors 2464 -> 0 (407862 applied records),
final nine-corner SDF errors 22638 -> 0 (3670758 applied records). Directed pad
checks passed 77/77 in every selected corner. This exposed, rather than waived,
5 candidate / 45 final functional transaction failures; full Verilator coverage
remained unmeasured. Annotation closure alone is not GLS PASS. The same packet's
broad unit selection was 402 PASS / 6 FAIL, all six in unrelated formal source
path assertions; do not describe that broad selection as green.
Captured rule
For a physical chip GLS replay, bind the external boundary environment as
well as the testbench clock period to the native SDC. An RTL bench that changes
bus controls immediately after an external posedge does not implement an SDC
set_input_delay 5.18: the controls can race the propagated internal clock.
Do not change any cell/SDF arc, expected value, transaction, timeout or clock
period. Apply the declared uniform input arrival on the existing pad-driver
bridge (never on the clock), and observe the existing read capture at the SDC
T - output_delay - uncertainty deadline. Leave nonuniform/unsupported
constraints unguessed.
On the v1.40.54 SHA256 State, native tracing showed the initially valid one-hot
FSM become 0000 at INIT with the zero-offset RTL stimulus. Changing output
sampling alone from 1 ns to the SDC 20.47 ns deadline still failed 5 candidate /
45 final transactions. Binding the original SDC 5.18 ns external input arrival
as well closed those failures: cpu_queue sha256-gls-sdc-boundary-oai measured
candidate/final gls_failures 5/45 -> 0/0, SDF annotation errors 0/0,
with 407862/3670758 applied native delay records unchanged. Candidate:
6 Icarus + 7 Verilator PASS; final: 54 Icarus + 7 Verilator PASS, no unmeasured
functional arm. The original error limit is still zero. Parser tests: 5 PASS
locally; packet narrow suite had 109 PASS plus one newly added test's malformed
single-line fixture failure; correcting the fixture's instance line boundary
passes all 5 new tests without changing implementation behavior.
Routed timing-repair margins are per-design optimization settings, not signoff
limits or global headroom defaults. Preserve PostRouteSizing's shipped defaults.
Allow an explicitly selected VIBEIC_PRR_SETUP_MARGIN / VIBEIC_PRR_HOLD_MARGIN
to survive document/native-config filtering just like the other PPA settings.
For the v1.40.54 SHA256 routed State, the original native setup/hold slacks were
-0.25775649011418533 / -0.03923617097834743 ns. Selecting 0.5 / 0.1 ns margins
in that design's work/flow/native_overrides.json, then native routed sizing,
rerouting, RCX and nine-corner STA measured +0.25528735403758707 /
+0.0747262273210573 ns (cpu_queue sha256-per-design-routed-margins-oai). The
original candidate SI checks remain >= 0 ns; both pass. Narrow tests: 14 passed,
1 skipped (the existing external-tool prerequisite). The replay harness failed
only when it tried to write into an uncreated SI report directory after STA;
repair.json and native STA reports completed, and the ordinary SI comparator
on those emitted Decimal metrics passes. This is not final electrical closure:
slew/cap/fanout and layout checks still require their own native results.
Captured rule: keep simulator roles explicit
The compiled functional arm does not annotate SDF. Generated RTL and GLS benches share their SDC input-drive and output-capture schedule; do not derive it from measured propagation delay. Icarus owns native SDF annotation, including the original pad models. A functional mismatch must still fail, and a missing native measurement remains NOT_MEASURED. Keep the original limits.
Captured rule
Select the GLS subject from the active native configuration. A hard macro may share project history with a chip run; retained preparation or wrapper inputs must not change its selected DUT. Match the prepared chip DESIGN_NAME to the selected native DESIGN_NAME before applying chip-specific port binding or simulator selection.
SHA256: preserve selected-bit predicate text in the simulator, not the SDF
When native GPIO SDF cannot bind conditional paths, reduce a real CELL record
with the unchanged installed PDK model, and inspect native -sdf-warn output.
Zero SDF ERROR lines alone do not demonstrate annotation: unsupported predicate
text may leave paths skipped. Repair the simulator's predicate representation;
never replace conditional paths with unconditional ones or edit native delays.
Icarus master 0d9012941 had lost constant bit-select rendering during source
identity integration. Its existing sdf_cond_bit_select waveform test was also
unregistered. Fork PR vibeic/iverilog#20 (40e0d54b2) restores declared-index
text only for one-bit equality operands; the existing unsupported bare-select
fallback/ifnone/reject contracts and goldens remain unchanged. It registers the
existing regression and a bitwise-AND counterpart, both checking actual edges
at 103/205/307/411 ps on ascending and nonzero-based packed ranges.
Measured cpu_queue sha256-iverilog-bitcondition-regression0476-oai, image
0.4.76: identical 112 native tests, base 2 FAIL, repair 0 FAIL,
NEW_FAILED=[]. A real single-pad CELL emitted 24 skipped conditional-path
warnings with the released tool and 0 with the rebuilt repair. The known
TIMINGCHECK capability warning remains visible. This is a simulator reproducer
repair, NOT proof that SHA256's original 1824-error candidate GLS row or either
complete chain passes. Keep their original results until native replay on a
released image measures them; only the fork gatekeeper merges tool changes.
Captured rule
For timed cocotb replay, bind the constant actually passed to the imported
cocotb.clock.Clock constructor to the native SDC period, including its literal
time unit. Do not infer support from a constant's spelling or silently accept
an unbound driver. PWM4's documented reference uses Clock(dut.clk, CLOCK_NS, unit="ns"); the previous binder recognized only custom HALF_NS drivers and
reported its timed Python arm NOT_MEASURED even at the documented 40 ns.
pwm4-cocotb-constant-before-i0476-oai measured 2 native positive regressions
FAIL (3 ns / 3 us left unbound instead of 25 ns) and 2 shadowed negatives PASS.
pwm4-cocotb-constant-after-i0476-oai passed 39 tests in 124.37 s; both unit
variants measured 25000 ps, native Icarus GLS PASS, and source benches remained
byte-identical. Shadowed constants and clockless suites remain NOT_MEASURED;
existing manual-clock, pad-path and compiled-functional negatives still pass.
Full PWM4 chain proof is separate from these measured narrow tests.
Captured rule
For a cocotb GLS shim, drive declared logical inputs through the existing
bridge_bench_inputs continuous-assignment boundary before SDF annotation.
Use native chip preparation only when its configured chip top matches the
requested netlist top; retained chip metadata must not block an IP replay.
Keep output and true bidirectional ports undriven by this input bridge. A
missing SDF interconnect is an annotation failure even if functional checks
pass. An unbound clock remains NOT_MEASURED unless its actual native period
is measured in the same replay.
Captured rule
An unrecognized Python clock-driver style need not block a native GLS result
when that same simulation observes every declared SDC clock. The existing
shim dumps only those scalar DUT ports, without changing time precision or
the source suite. Check every observed rising and falling period in integer
native ticks and retain the compressed trace. Use each native JUnit test's
recorded simulation start, stop, duration and unit to delimit its observation.
Carry scalar state across tests and reset period history at those boundaries.
Each interval includes its start and excludes its stop, so cancellation at
the exact stop can precede an edge. Require complete rising and falling cycles
and activity through each stop, with the native VCD's final timestamp covering
that stop. Convert only reported test boundaries to the nearest native tick;
clock periods still compare exactly. Missing or unusable test intervals and
unknown, irregular or interrupted observations remain NOT_MEASURED; do not
infer pauses within a test. Existing recognized clock settings keep their
original binding. Generic observer calls without JUnit keep the strict
whole-trace check.
Native controls measured 25900 ticks at 1 ps for an otherwise unsupported
25.9 ns driver, with functional checks and SDF records applied in both cache
replays. Incorrect, absent and stopped clocks remain unmeasured. The matched
program population passes 69 cases with NEW_FAILED=[], retaining all 49
existing cases. This measures periods at native VCD resolution; it does not
establish absence of same-slot glitches suppressed by the writer. Setup,
hold and pulse timing retain their native STA obligations.
Captured rule
A hard-macro GLS boundary needs its original SDC input arrival times just as a
pad boundary does. Recording input_arrivals without driving that delay models
a different SRAM environment: new synchronous read data can reach delayed
internal clocks in the wrong cycle. Insert the declared delay on the existing
bench driver, never on native cells/SDF and never on the clock. Retain unobserved
native outputs with undriven bench wires so Icarus can annotate their original
INTERCONNECT records. Preserve existing named/escaped connections; do not mix
named and positional syntax or remove SDF errors.
Subservient IP, image 0.4.76, original 10 ns clock / 2 ns SDC input arrivals: candidate GLS went from 5 failures (3 functional + 2 missing output intermodpaths) to 0; applied delays 6060 -> 24256 across four Icarus cases, annotation errors 2 -> 0, and all four Verilator cases passed. Native final nine-corner replay improved 45 -> 5 failures with zero annotation errors: the CPU program still fails 0/5 in min_ss_100C_1v60, while the other eight program probes pass 5/5. Existing stop-on-failure leaves the remaining final suite unmeasured; this is not final replay closure. IC control still reports its original 912 annotation errors (7743 applied records), not a repaired simulator claim.
Native boundary/workload tests: baseline 29 passed, changed 33 passed. Broader
native verification tests retained the same six pre-existing formal-path
assertion failures (418 -> 421 passed); no original tests or failures were
waived. Evidence: subservient-ip-gls-boundaries-v14067-i0476-oai,
subservient-ip-final-gls-boundaries-v14067-i0476-oai,
subservient-ic-gls-escaped-control-v14067-i0476-oai, and
subservient-gls-boundaries-final-unit-i0476-oai.
Derive child pin sets and SDC names from the native parent binding
A parameterized child must be physically prepared against its parent-bound native interface, not its standalone default width. Preserve every supplied pin pattern and its edge. Place missing bits on the edge of the numerically highest already-placed bit of their bus; for a new bus or scalar, use the last declared edge. Preserve reverse-edge directives. Supply pins remain owned by the PDN, not signal IO placement. Run the native IO placer on the derived file; do not relax missing-pin diagnostics or drop interface bits.
Where a supplied child SDC still queries a parent port name, map only static
get_ports objects using unambiguous equal-width whole-port connections from
the native elaborated parent hierarchy (the design's own L8 binding). Never
infer aliases through slices, casts, internal nets or ambiguous connections.
Retain every timing value, exception, clock, load and original limit. Stage
the same derived SDC into Phase 1 and native implementation. Child-authored
work/flow/native_overrides.json remains an independent design input; parent
configuration must not leak into the child.
Historical narrow measurements on Caravel/image 0.4.76: the L8 BITS=32 binding
and supplied 16-bit pin order yielded 48 missing pins, reduced to 0 by complete
pin placement; placement and CTS changed FAIL→PASS. The wrapper user_irq to
child irq alias reduced OpenSTA missing-port diagnostics from 6 to 0 over
three corners, with unconstrained paths/endpoints remaining 0 against limit 0.
These measurements motivated the general preparation code; hand-run per-design
recipes are not a valid fresh-chain fix and have been removed. The general
code is exercised with a neutral WIDTH=4 tile, including native dir:: glob
preprocessing, preserved original edge placements and SDC values, and rejected
ambiguous/sliced aliases. Fresh from-documents packet
caravel-ip-fromdocs-v14063-childinputs-glob-i0476-oai on v1.40.63 plus this
fix ran without manual preparation: native missing-pin diagnostics 48→0,
SDC invalid diagnostics 6→0 (limit 0), placement/clock planning/CTS and
pre-layout timing PASS. All 16 targeted tests passed. Parent result was
38 PASS / 7 FAIL / 0 NOT_MEASURED, NEW_FAILED=[] relative to the provided
chain. The macro row still fails on later child setup timing (-0.438153 ns),
SDF annotation (99 candidate errors), and final electrical analysis;
child functional simulation/activity remain unmeasured. Do not describe this
input-preparation repair as whole-macro or parent signoff closure.
Captured rule
A sequential UDP's asynchronous level rows and edge rows update the same
state. If a simulator lowers the former to combinational blocking assignments
and the latter to NBA writes, repair that lowering in the simulator; do not
suppress BLKANDNBLK or replace the PDK models. PWM4 native compiled GLS on
image 0.4.76 produced four BLKANDNBLK errors and no JUnit verdict in
pwm4-ic-compiled-core-native-i0476-oai. The repair is submitted for fork
review at https://github.com/vibeic/verilator/pull/2 (branch
fix/udp-async-nba-output, head 0acd73e7a, base 806bddcf3); the author
has not merged it. pwm4-verilator-udp-configured-tests-i0476-oai built with
--enable-ccwarn and passed all 22 native UDP regressions, 0 failed/skipped.
pwm4-ic-udp-functional-runtime-i0476-oai used that built compiler on the
retained routed core: both the retained bench and document-derived expert
reference passed, 2/2 JUnit tests, GLS failures 0. The split workload's separate
Icarus SDF arm and full stock-image chain still require their own proof after
the fork gatekeeper integrates the fix. No native limits or diagnostic gates
were changed.
Captured rule
Bind generated CPU-case startup reset launch to the native SDC reference edge; do not add a rising-edge input arrival to an unrelated falling-edge release. The case producer owns initialization stimulus, not the checker. Launch the CPU program's startup release nonblocking after its already-counted rising edge. For the generated retention case, retain both original measured windows and in-flight reset assertion exactly, hold reset for one additional unmeasured setup edge, then launch only initial deassertion nonblocking. The existing native input driver supplies the original arrival delay. Do not initialize RF state, change memory contents, extend program/retention budgets or suppress Xs.
Measured Subservient IP, image 0.4.76, original 10 ns clock / 2 ns input arrival: min_ss startup previously advanced bsel at 107.073 ns but pending at 117.101 ns, skipping the byte-zero tag. RF rdata0[1] first went known-to-X at 127.052 ns; regzero staying X was a consequence, not permission to initialize it. The paired nom_ss control aligned bsel/pending at 117.102/117.132 ns. With the owned startup launch, the pin reset release is 97 ns rather than 102 ns, bsel/pending align at 107.073/107.101 ns, the first byte is captured at 127.091 ns, and regzero resolves normally. RAM preload and expected program words are unchanged.
Native final GLS: 5 functional failures -> 0, all 36 Icarus cases across nine
corners and all four Verilator cases PASS; 218304 delays applied, 0 annotation
errors. The min_ss four-case control also PASSes (24256 applied delays). Native
boundary/workload tests: 33 -> 38 passed. A blanket clock-resampling experiment
was rejected because it added an extra active-operation latency and produced
9 reset-retention failures; neither that source nor a moved checker window is
used. Evidence: subservient-ip-final-gls-producer-launch-i0476-oai,
subservient-ip-minss-producer-launch-trace-i0476-oai, and
subservient-producer-reset-launch-unit-i0476-oai. Automatic producer helpers
and their native-SDC call sites are the program-level capture.
Captured rule
SDF routes through continuous output assignments must retain bit identity
When a native INTERCONNECT from a scalar cell output to a vector output bit
reports Could not find/insert intermodpath, inspect the compiled route before
changing the SDF or cell model. An explicit continuous assignment can retain a
Z-preserving BUFZ before the vector concatenation. That assignment is part of
the valid route, not a logic gate or an absent endpoint. Icarus must traverse it
while retaining the node and its strength conversion; do not remove the record,
mask its error, replace the tie model, or count it as an applied delay.
Program capture is Icarus fork PR https://github.com/vibeic/iverilog/pull/21,
commit d32f49f, not a plugin waiver. The native regression checks independent
2/3/4 ns delays on 0/1/X/Z transitions plus constant outputs. Before it failed
with six rejected records and an undelayed edge; after it passed. The complete
native VVP suite passes 1103/1103, with optional waveform backends enabled.
Evidence: caravel-iverilog-output-alias-artifact-i0476-oai (exit 0).
Paired actual Caravel wrapper replay, unchanged routed netlists/SDF and original
limits, base image 0.4.76, stock tool before and fork-built tool after:
caravel-parent-iverilog-output-alias-native-i0476-oai:
- Candidate: annotation errors 111→0, applied delays 7432→7543,
gls_failures111→0. Icarus 3/3 and Verilator 3/3 tests PASS. - Final: annotation errors 999→0, applied delays 66888→67887,
final_gls_function_timing_errors999→0. All nine Icarus corners 3/3 tests each and the Verilator 3/3 suite PASS. - Original requirements remain annotation/function errors 0 and applied delays >0. Both rows FAIL→PASS, NEW_FAILED=[] in this two-row comparison.
Scalar destination links retain their value filter
A scalar output alias can terminate directly after a one-bit BUFZ strength
conversion. Icarus must match that destination and insert the SDF delay with
the destination signal's value filter. Preserve the conversion and source
observations. Fork regression interconnect_scalar_output_alias measures
independent 2 ns and 5 ns delays, shared wiring, unannotated fanout and assignment
strength. The full 1,136-ID VVP panel changes from 1,117 PASS / 1 FAIL / 18 Not
Implemented to 1,118 PASS / 18 Not Implemented, with NEW_FAILED=[].
PWM4 IP v1.41.3 with author-tested Icarus head
a0786952db4e82eb50950f693303a809ca79ebdf inside image 0.4.77 passes complete
Steps 75 and 96. Candidate coverage is three Icarus and four Verilator cases,
with 25,401 applied delays. Final coverage is 27 Icarus cases across nine corners
and four Verilator cases, with 224,490 applied delays. Annotation errors fall
from 1 / 9 to 0 / 0. Before repair the timing arm stopped after its first case
at each corner. The after runs complete the original directed and professional
suites with zero GLS failures. Electrical capacitance remains open. The native
repair and regression are in Icarus PR 23.
The exact increase in applied records is evidence that the missing paths were annotated, not discarded. These are native narrow proofs using the fixed fork; they do not claim that an unmodified old image or the complete design chain passes. Tool deployment remains with the fork gatekeeper; the author does not merge the PR. Child functional/activity and final electrical DRVs still need independent closure.
Routed sizing needs headroom for legalization and rerouting
Do not use a zero-slack optimization target when the native routed sizing step
is followed by legalization, global routing, timing repair and detailed routing.
Keep every signoff scene, input/output delay, derate and acceptance limit. For
an unspecified routed sizing target, reuse the existing design derating budget
(CLOCK_PERIOD * TIME_DERATING_CONSTRAINT / 100) or the native GRT setup
margin, whichever is larger; retain the native GRT hold margin. Project-native
and document-authored routed targets, including explicit zero, still override
these defaults. This is implementation headroom, not a relaxed timing check.
Measured Caravel child, image 0.4.76, base main v1.40.79: the native sizing log
reached +0.431 ns setup slack before subsequent routing, but signoff on the
resulting extracted route had setup -0.630002734 ns and hold
-0.067879038 ns. Native routed optimization targets were 0/0 ns. With the
existing 25 ns period and 5% derating budget, the default setup target becomes
1.25 ns and the existing GRT hold target 0.2 ns. Fresh native child
implementation through candidate SI then measured setup +0.643368044 ns
and hold +0.157165840 ns, both against unchanged 0 ns limits:
candidate_signal_integrity FAIL→PASS, 129 tests passed, NEW_FAILED=[].
Evidence: caravel-routed-headroom-native-v14079-i0476-oai versus
caravel-child-timing-native-v14077-i0476-oai (same relevant native flow).
This is automatic program capture in flow_config.native_defaults.
Do not infer whole signoff closure: max-slew violations remained 579 (before 586), max-cap violations 13 (before 14), and fanout violations 7. Native GLS and missing child functional activity still require separate work. The earlier CTS-scene experiment was rejected: selecting the measured slow scene changed setup only to -0.621924 ns and hold to -0.045779 ns; it did not pass and no scene-selection change is included in this fix.
Captured rule — the SDF arm needs directed cases the documents declare
A gate-level row that runs for hours first needs its actual SDF workload
read. gls_workload.run uses existing directed HDL cases first. If none exist,
it calls the existing stated-vector producer when the declared L9 logical
surface matches that producer. The producer uses the provided L4 register map
and L10 literal message and expected answer; it does not compute an answer or
invent register addresses. Generated cases go into the run's directed_cases/
folder, with ordinary producer refusals retained. Repeat payloads stay in the
complete compiled suite. If no case can be produced, the existing whole-suite
SDF fallback remains explicit. Keep every selected corner, signal pad path,
original assertion, limit and complete compiled functional suite.
The selection is the documents' and only theirs (2026-11-12): a bench is replayed with SDF when its file name is the name of a test case the verification plan declares (L10). The reader no longer picks benches by their language or length. When the documents declare no case that the suite delivers, the SDF arm replays the whole suite. Both gate-level rows replay every corner in native State, and the row names the benches replayed with SDF with the document line that declares each, the benches that ran at zero delay only, and the corners.
Fix it in the design documents: declare short directed cases, one per
timing-relevant transaction (a write to every register observed at an output,
an output edge on every channel, reset release, each asynchronous input), as
- **Name**: stimulus; expect ... bullets in the verification-plan document,
and deliver each case's test program as input/sim/tb/<case name>.v
(module <case name>;, a `timescale, a free-running
always #<half> clk = ~clk;, the logical top instantiated with named
connections, [TB ...] FAIL lines and ORACLE_TB_DONE pass=n/m). The
stimulus text must come before the first "expect/check/assert" word or Phase 1
refuses the case. The full contract test stays in the suite for the compiled
zero-delay arm.
PWM4 IC, gf180mcuD, image 0.4.76, from documents: three directed cases
(reset release; nine register writes with a PWM edge on every channel;
asynchronous reset while running). SDF arm before: tb_chip_top.py
(82,215,675 ns, 2.06 M clock cycles, 9171 s per corner in Icarus), row wall
9192 s candidate / 24644 s final. After: 3 directed cases + pad paths, PASS at
all nine corners, sdf_annotation_errors 0, failures 0, 3-24 s per case; row
wall 27 s / 66 s; chain 11 h -> 26 min. Both rows stay FAIL on the unchanged
compiled arm: Verilator rejects the gf180 UDP primitives.v (BLKANDNBLK), so
the contract test still has no zero-delay gate-level run on this image; it
passes in functional_simulation_and_coverage.
SHA256 IC, sky130A, stock image 0.4.78, native Step 75 on the original retained routed checkpoint: five cases produced from provided L4/L9/L10 documents plus the 77-path signal-pad probe pass at the selected TT corner. Each of the six Icarus runs applies 67,123 SDF delays, totaling 402,738, with zero annotation errors and zero failures. The complete compiled suite passes 6/6 with no skips; its native log includes the one-million-byte message and 76,451,050.20 ns of simulation. Icarus does not implement SDF timing checks; setup/hold stay with STA. The producer change's 116 matched native program cases improve 114 PASS/2 FAIL to 116 PASS/0 FAIL, with no new failures. Five nonfailing controls retain the compiled bench bytes, seed samples and native simulated intervals. The separate wrong-answer control fails timing and leaves the compiled arm unexecuted. This is the candidate GLS row's result; final GLS and whole-chain acceptance still need their own native results.
Captured rule
When a verification step has several engines, each engine declares its role in one data table, and tests exercise every input kind on every declared engine. Do not infer timing ownership from the compiler or bench kind. Measured GLS example: SHA256's plain bench passed, but SPM's generated scenario and serial_parallel benches sent Verilator's routed-core functional arm into the chip/STA observation path (Steps 75/96 NOT_MEASURED in v1.40.57). Icarus owns chip SDF/STA timing; Verilator owns zero-delay functional coverage. Both must pass, with the same checks, limits and transaction populations.
Captured rule
A seal ring and a pad may not share ground. The PDK seal generator draws its ring inside the frame it is given; pads placed at the die edge are then crossed by the ring's diffusion, which the finishing XOR reports as added geometry touching design geometry and the foundry-style DRC as diffusion spacing. When the design declares no pad edge spacing, measure the ring on an empty probe before pad placement and set the pads back past its drawn depth on a straight edge, counting only the layer pairs the pad and cell libraries draw on themselves: marker-only layers and the corner pieces reach far further in and are not what a pad can collide with. A declared spacing is left as declared. Measured on a SKY130 pad-limited die: drawn ring diffusion 1.45 to 3.55 um from the frame (a marker layer reaches 101.6 um), pads set back 5 um, die 2398x908 -> 2408x918 um; finishing XOR 2824 -> 0 and KLayout DRC 1152 -> 53 at the original limits. A die that grows also moves the density windows: read the density row of the same run before accepting the change.
Captured rule
One bench clock generator can drive several DUT clock ports. Bind every port
to its declared SDC period, but replace the generator's delay once. Retain
each port in the binding report. If shared ports require different periods,
the bench cannot supply both and the binding stays unmeasured. An undeclared
clock port remains undeclared when it shares a generator with a declared one.
programs/sdf_gate_sim.py keeps one replacement per original generator span
and refuses conflicting periods before changing the bench text.
Measured on an authored AES bench with two DUT clock ports connected to one
generator: applying both edits to the original #5 span produced #255, even
though each port's binding metadata claimed #25. Native Icarus measured
510 ns against the declared 50 ns. Editing that generator once produced
#25, and native Icarus measured 50 ns against 50 ns. The matched generic
clock checks improved from 15 of 19 to 19 of 19 with no new failures.
This verifies clock preparation using the generator from the actual bench. The complete gate-level simulation still requires its native compilation, SDF and functional results.
Captured rule
An authored gate-level bench can depend on packages and helper modules that the implemented DUT does not use. Resolve that bench's dependencies through the existing simulation source closure, keeping the native gate netlist and PDK models as the selected providers. A helper outside the synthesized RTL still belongs in the simulation input when the bench instantiates it.
Verilator consumes the package-ordered sources. For Icarus, use native sv2v
on the bench and its support closure, then compile the converted result with
the original gate netlist and PDK timing models. Keep the source bench, SDF
annotation, selected corners and original functional checks unchanged.
Measured on an AES chip bench with stock image 0.4.81: the original Icarus
compile exited 14 at a package-qualified declaration. Adding the source
closure alone still aborted at unsupported SystemVerilog inside
expressions. Native sv2v conversion of only the bench and support exited
zero, and Icarus then compiled against the unchanged gate and PDK models with
exit zero. This establishes compiler preparation, not a full GLS verdict.
Carry the bench's declared timescale across this conversion. sv2v removes
that directive while retaining numeric delays, so a compiled #25 alone
does not prove a 25 ns delay. Restore the exact declaration already present
on the injected bench, without inferring units or changing the delay. This
lowering scope contains one timed bench and its event-driven support; the
gate and PDK models keep their original time declarations. Support with its
own differently timed behavior needs those units preserved independently.
Review a native path-delay or clock-period observation after lowering;
eventual functional success alone does not establish the intended timing.
Native Icarus measured the converted AES generator at 50 seconds before the directive was restored and 50 ns afterward, against the declared 50 ns. A timing-sensitive support fixture changed from FAIL to PASS while checking the native 1.5 ns path delay. The actual nominal-TT FIPS-197 C.1 case then compiled and applied 127436 SDF delays with zero annotation errors, but its bus read still timed out after 20001 cycles. That case remains FAIL; compiler preparation and correct clock units do not establish functional acceptance.
Captured rule
Apply a uniform, nonnegative SDC input-arrival offset to its declared input port even when the design declares several clocks. That offset is a boundary driver delay in the SDC time unit; it does not require choosing one clock period. Keep clocks undelayed and leave nonuniform bus or min/max offsets unbound. A delayed native clock can otherwise accept an undelayed request before the external bench expects the handshake, so the bench can miss a response that the netlist produced. Trace external and core valid/ready signals before attributing such a timeout to reset or peripheral readiness.
Captured rule
Select the library's functional model view for compiled functional GLS.
Removing its unit-delay macro does not select that view. A library timing
model can feed sequential primitives from delayed clock, data and reset nets
that only timing checks drive. A simulator that ignores those checks cannot
use that model to establish functional behavior. The compiled functional arm
defines FUNCTIONAL; the Icarus SDF arm keeps the timing view and its existing
annotation checks. Keep the authored functional checks and source stimulus.
Captured rule — a parent with a hardened child: what each side reads
A hierarchical IP (a wrapper that instantiates a child hardened first) fails i
…(truncated)