[debug] Inline enum variants into debug var/subfield intrinsics

[fkhaidari]

Summary

Drop the separate circt_debug_enumdef intrinsic. A ChiselEnum-typed circt_debug_var/circt_debug_subfield now carries its variants (JSON) and bit width inline alongside enumTypeName/enumFqn, so every enum use site is self-contained and no per-circuit dedup is needed. This removes an outlier intrinsic and matches the CIRCT side, which resolves enum metadata from the var/subfield itself rather than a staged enumdef table.

Type of Improvement

Interal

Desired Merge Strategy

Rebase: You will rebase the PR onto master and it will be merged with a merge commit.

Release Notes

Drop the separate circt_debug_enumdef intrinsic. A ChiselEnum-typed circt_debug_var/circt_debug_subfield now carries its variants (JSON) and bit width inline alongside enumTypeName/enumFqn

[fkhaidari]

Hi @jackkoenig!

I’ve reworked the debug intrinsic design to cut down on the translation complexity on the CIRCT side.

The separate circt_debug_enumdef intrinsic is gone. A ChiselEnum-typed circt_debug_var/circt_debug_subfield now carries its variants (as JSON) and bit width inline, next to enumTypeName/enumFqn. Every enum use site is self-contained, so there’s no per-circuit dedup pass and nothing for CIRCT to resolve against a staged enumdef table; the enum metadata travels with the value.

The matching CIRCT change reads the variants straight off the var/subfield and materializes the dbg.enum at the use site. I’ll push that side once this lands so the two stay in sync.

Could you take a look?

[fkhaidari]

While I’m here, I’d like your take on one thing

For a bidirectional or aggregate signal, circt_debug_var is emitted with no SSA operand, so CIRCT has to recover the root signal by matching the var’s name against the wires/ports/registers in the module (resolveRootSignal in llvm/circt#10647). That name-based lookup is the fragile part: it has to disambiguate by name and special-case memories, and it breaks down whenever PrettifyVerilogNames or lowering renames the underlying signal.

The reason it can’t just take the signal as an operand is that firrtl.int.generic operands must be passive (FIRRTLIntrinsics.td), so a bidirectional value can’t be passed through. Is there a way to attach the signal as a proper operand (or otherwise give CIRCT a stable handle) that I’m missing, or is the name-based binding the intended contract here?

[fkhaidari]

@jackkoenig, could you take a look?

[jackkoenig]

Without being an expert on the CIRCT representation, this makes sense to me and sounds a lot better.