[oharboe]

Phase A of moving register-file generation into ORFS and OpenROAD: the generator becomes an OpenROAD command, carried here as a patch, and is proven byte-identical to the tool it replaces. Nothing calls it yet, so no flow output changes. Phase B (register files through AUTO_MEMORIES, as macros that dissolve after macro placement) builds on it.

The command

Carried patch 0010-openroad-ram-generate-regfile.patch, in OpenROAD's src/ram, next to generate_ram_netlist:

read_lef <tech.lef>
read_lef <cells.lef>
generate_regfile -spec rf.regfile [-verilog rf.v] [-def rf.def] [-lef rf.lef] [-liberty rf.lib] [-check_ports rtl.v]

• The code is tools/structured_gen's, moved: the regfile and views sources, in the ram namespace, with exceptions reported as RAM-0050 and a missing -spec as RAM-0051.
• Both build systems list the new sources: Bazel, which is built here, and CMake, which isn't.
• Not upstreamed.

Proof: byte for byte

tools/structured_gen/golden_test.sh runs each spec through the tool and the command from the same LEFs, then requires an identical Verilog netlist, DEF, abstract LEF, model liberty and pre-layout liberty.

test	specs	time	result
golden_test (default)	five tiny synthetic specs covering every spec feature: banks, bank columns, bit folds, registered read with the RTL's names, banked and interleaved reads, the zero word, netlist mode	4 s	all identical
golden_xiangshan_test (manual)	the ten specs the flow uses, at real size (RobEntryFile is about 400 k cells)	48 min	all identical

The tiny specs, in tools/structured_gen/golden/, are also a fast test case to keep for the phases that follow.

Cost

The openroad binary changes, so every ORFS stage re-runs once. Building it here was 19 local actions (the ram sources and the link) on top of the remote cache.

Next

• Phase B: AUTO_MEMORIES register files, named explicitly as STRUCTURED_MEMORIES does today (parity first), placed as macros, then dissolved into a FIRM core plus a periphery for OpenROAD. It retires ORFS patches 0071, 0078, 0082 and 0084, and the tool itself.
• Improvements come after parity: automatic detection, rotation.

🤖 Generated with Claude Code

[oharboe]

Two commits added, same edit in the tool and in patch 0010 so the golden test stays byte for byte:

• Pins where their connections land, on the track grid. In netlist mode the abstract is what the parent's macro placer sees before the file dissolves, and the placer flips a macro by its pins, so each pin now sits at the centroid of its net's cells. Off the M4 track grid, rtl_macro_placer refused the macro (MPL-0005); they are snapped as the edge pins are.
• Ports declared several to a line. The port check read one port per line and missed input [4:0] ra1, ra2, wa3, which is how riscv32i declares its register file. It now reads one comma entry at a time.
• An OpenROAD regression test in the patch. src/ram/test/generate_regfile_asap7: an 8 x 4, two-bank, 2R2W file with word 0 reading zero, checked against RTL with comma-declared ports, its four views against golden files. Run by hand (OpenROAD's test targets don't analyse from bazel-orfs), the log matches .ok and all four views match.

The ORFS side that uses this, replacing STRUCTURED_*, is stacked on this branch: see the next PR.

[oharboe]

The stacked PR is #1187.