IExpand: replace the SYB traversal in removeInlinedPositions

[nanavati]

Under -keep-method-conds, removeInlinedPositions strips the inlined-position bookkeeping from every ICon head once generateMethodPreds has harvested it into the COND defs. It did this with a Data.Generics "everywhere" traversal, which pays a generic-dispatch cost at every node of the module; on a large module the pass alone took about a third of elaboration time (and ~1 GB of rebuilt heap).

Replace it with a direct traversal over just the fields that carry expressions (local defs, rules, interface, state-instance args, clock/reset/inout wires, and the expression-bearing IConInfo constructors). The traversal is deliberately kept as lazy as the SYB one was: the output must stay byte-identical, and downstream declaration order depends on the order in which the generated names are first forced (FString interning feeds Ord Id, which feeds tsort tie-breaks and the Set-sorted lists in the Verilog backend). An earlier version that indexed the defs eagerly — a Map from def id to stripped body, to share ICValue rebuilds — reordered the generated Verilog; recursing into each ICValue payload directly avoids that and reproduces the SYB output exactly.

This is a performance change with no output difference. Correctness is covered by the existing bsc.misc/method_conditions tests, which check both the COND signals in the generated .v and the bluetcl position readout from the .ba (134 pass). Measured on a module elaborating to ~98k defs under -keep-method-conds, the pass dropped from ~27s to under a second.

🤖 Generated with Claude Code

https://claude.ai/code/session_01VVN2QMZtLUwkag1vCYLaao

---

Generated by Claude Code