[Bug] DELAY-family ABC strategies (retime/mfs/buffer-upsize fine-tune) make block-scale netlists routing-hostile: 4.5-5.5x more DRT violations than a plain mapping script, with LOWER Fmax

[WilliamHHL]

Title: [Bug] DELAY-family ABC strategies (retime/mfs/buffer-upsize fine-tune)
make block-scale netlists routing-hostile: 4.5-5.5x more DRT violations than a
plain mapping script, with LOWER Fmax

Self checks

• I have searched existing issues, including closed issues.
• I have read the documentation and development guide.
• This is a bug report, not a question, feature request, or usage discussion.
• I will include logs, command output, or reproducer files where they help diagnose the issue.
• I have removed secrets, private paths, tokens, and other sensitive information from the report.

ECC version

ecc 1f910c3 (frozen M1); strategy code identical on current main
(yosys_synthesis.tcl DELAY family, indexes 0-13).

Affected area

Synthesis (yosys ABC strategy: tools/yosys/scripts/yosys_synthesis.tcl)

Environment

• OS: Ubuntu 26.04 LTS, x86_64
• PDK: ICS55 (icsprout55-pdk), dual-VT std cells
• Design: ~94k-cell MCU SoC with two SRAM macros, util 0.65, MET1-5, 10ns clock

Steps to reproduce

1. Synthesize a block-scale SoC with the DELAY family (e.g. "DELAY 4").
2. Re-synthesize the SAME RTL with a plain script
   (+fx;strash;refactor;resyn2;scleanup;map;&get,-n;&st;&dch;&nf;&put;
   print_stats -m — via the ABC_RAW_SCRIPT env we added locally).
3. Route both netlists with the same backend and compare DRT.

Expected behavior

Timing-driven strategies (retime, mfs, choice, map_tight, buffer/upsize/dnsize
fine-tune) should improve or at least not harm routability for their timing
cost — and should not LOSE Fmax to a plain mapping.

Actual behavior

OpenROAD 26Q3-1305, identical flow (same frozen design, MET1-5, util 0.65):

netlist	DRT1	9-iter floor	TT ideal Fmax
DELAY0-style	253,536	106,188	~273.9 MHz
plain "safe" script	46,248	12,294	~297.7 MHz

• 5.5x DRT1, 8.6x floor improvement, +24 MHz Fmax — the aggressive script
  bought nothing and lost routability.
• keep_hierarchy ON vs OFF: DRT identical — the script is the lever, not
  hierarchy.

ECC flow (our replication, same PDK/design family):

netlist	GRT overflow	DRT1
DELAY-4 (default)	54,049-56,611	529,705-572,013
safe script + X0P5 dont_use	5,559-6,038	100,150-106,951

• 4.5-5x on ECC's own router too — the effect is tool-independent.

Interpretation

The DELAY family's transforms (retime across flops, mfs deep AOI/OAI
factorization, if -K 6 choice exploration, and the buffer/upsize/dnsize
fine-tune) create deep AOI/OAI/MUX clusters and inflated cell counts. Both
correlate with local detail-route congestion (hotspots measured in AOI/OAI
pockets). On this design the aggressive scripts provided no timing benefit
(plain mapping was FASTER) while multiplying DRT violations.

Suggested fix

1. Provide a routability-friendly default strategy (the plain script above or
   a BALANCE-family equivalent without retime/fine-tune) for block-scale
   designs, and/or make the strategy family selectable per design scale.
2. Audit why retime/mfs/upsize-buffer fine-tune raises DRT ~5x on block
   designs (structural vs cell-count mechanism) — a calibration pass against
   a routability metric would prevent the timing knobs from outrunning
   routability.
3. Minor: the bundled yosys/abc does not know the 'resyn2' alias
   ("unknown command 'resyn2'") — we expanded it to its standard definition
   locally.

Related

• ecc-tools#198 (router regional-repair gate), ecc-dreamplace#35 (EGR
  density inflation), ecc-dreamplace#31, ecc#155 — same design investigation.
• ecc#257 (local congestion support request) — the routing-side tool gap
  remains after this synthesis fix.