AUTO_MEMORIES_REGFILES: register files as macros of placed standard cells

[oharboe]

What a user sees

• New: AUTO_MEMORIES_REGFILES lists register-file spec files, one per RTL module. During canonicalization OpenROAD's generate_regfile builds each file from placed standard cells: a netlist, an abstract LEF and a timing model, which sit beside the FakeRAM views. In mode netlist the file is a macro through synthesis and macro placement, and then dissolves into its cells: the core stays FIRM where the macro was placed, and the address decode is left to placement.
• New: AUTO_MEMORIES_MACRO_PLACE lists the register files that are placed as macros. Any other listed file is inlined: synthesis reads its generated netlist in place of the module.
• New OpenROAD command: generate_regfile, in the ram module.
• Works with a keep threshold: each register file's line in blackboxes.txt carries the area its .lib states, so SYNTH_MINIMUM_KEEP_SIZE can cost it (the format AUTO_MEMORIES: a macro's blackbox costs its view's area #1205 introduced).
• Breaking: STRUCTURED_MEMORIES and STRUCTURED_PLACEMENT, which were bazel-orfs's own carried ORFS patches 0071/0078/0082/0084, are retired. A flow that still sets either one fails, and the error names the replacement. In arguments/sources it fails at load time; in user_arguments, or on a stage rule directly, it fails at analysis. ORFS ignores variables it does not know, so a silent removal would have built without the register files the flow asked for.

Carried patches

patch	upstream	content
OpenROAD 0010	The-OpenROAD-Project/OpenROAD#11669	the PR's net diff at its head (691d8e15), all eight commits in one patch
ORFS 0089	The-OpenROAD-Project/OpenROAD-flow-scripts#4651, commit 1	AUTO_MEMORIES_REGFILES
ORFS 0091	#4651, commit 2	AUTO_MEMORIES_MACRO_PLACE
ORFS 0099	#4651, commit 3	inline a register file in place of a parameterized module

The ORFS patches are #4651's commits as rebuilt on #4650, with the example designs left out (flow/designs/asap7/regfile, flow/designs/asap7/riscv32i-regfile and flow/designs/src/regfile). Here the flow is tested on a fixture of our own, test/regfile, which is the 16×8 2R2W example as a bazel-orfs orfs_flow. They apply after 0090, #4650, which #1205 merged. With 0071 retired, 0090 loses the hunks it carried for 0071 only (the DEF-die area of STRUCTURED_MEMORIES files); the rest of it is unchanged. Each patch retires at a bump onto an upstream that carries it.

OpenROAD's own src/ram/test targets: bazel-orfs does not build or test them: CI runs bazelisk test ..., this repository's own targets. Run here anyway, the seven generate_regfile_*_asap7 tests fail to build, because a patch applied by archive_override cannot create the src/ram/test/asap7 symlink that #11669 adds, so their LEF and Liberty inputs are missing. The openroad binary is unaffected, and the four older make_* ram tests pass. The generate_regfile tests run in OpenROAD's own CI on #11669.

Migrated

test/structured_netlist, test/planned_parent, the XiangShan flows and tools/structured_gen/test move to the new variables. tools/macro_select/plan_floorplan.py now places the plan's register files by instance in place_macros.tcl, so it no longer writes netlists.txt. The FIRM stand-aside dance around rtl_macro_placer is gone.

Tests

bazelisk test over //test/regfile/..., //tools/structured_gen/..., //tools/macro_select/..., //test/planned_parent/..., //test/structured_netlist/..., //test/auto_memories/..., the XiangShan package and //test:all, plus @openroad//src/ram/test/...: 172 pass. The 7 failures are the OpenROAD generate_regfile tests described above.

• //test/regfile: the dissolve, every register clocked (macro and inlined), global route on both, simulation of both write styles against riscv32i/regfile.v, and synthesis with SYNTH_MINIMUM_KEEP_SIZE=1, including a RegFile <area> line in blackboxes.txt.
• //test:retired_structured_*_refused_test: the two refusals.
• //:fix_lint, //:public_surface and //:host_tools pass.
• The manual //test/planned_parent:mini_top_grt stops on DRT-0073 (no access point for u_block*/din[0]). It stops the same way on main at 7519e7c, so this PR doesn't cause it.

This is split out of #1187, which stays the integration and scaffolding branch: the register-file campaign, the ladder designs and the study variants.

🤖 Generated with Claude Code