Fe trace attribution and EVM PC mapping

[cburgdorf]

Summary

Add the minimal Sonatina substrate required by Fe's ethdebug-distill work:

• carry opaque Fe origin keys through IR optimization,
• expose deterministic views over optimized IR,
• stamp post-optimization instruction provenance,
• and join that provenance to EVM bytecode through byte-exact PC ranges.

This is intentionally scoped to Fe's trace pipeline rather than being a general-purpose debug-info framework.

Design

Each instruction can carry two independent optional attribution channels:

• Frontend origin: attached by Fe during MIR-to-Sonatina lowering and preserved through optimization.
• Post-optimization provenance: stamped by Fe after optimization and carried through EVM lowering into the final PC map.

Attribution strings use shared Arc<str> storage. Inlining and machine lowering therefore preserve identity without duplicating long frontend keys per instruction.

Function::propagate_inst_attribution copies both channels when a pass replaces or expands an instruction. This contract is used by the relevant optimizer passes and both inliners.

Trace views

The PR exposes deterministic, panic-free views over:

• functions,
• blocks,
• instructions,
• CFG edges,
• and instruction kinds.

It also exposes the optimized module before backend preparation so Fe can emit post-optimization trace facts and stamp instruction provenance.

EVM observability

EVM lowering distinguishes between:

• exact 1:1 lowering, which carries both attribution channels, and
• synthesized glue instructions, which may inherit frontend context but must not impersonate a stamped post-optimization instruction.

The object layer emits byte-reconciled PC ranges with either:

• frontend provenance for mapped ranges, or
• an explicit unmapped reason.

Synthetic observability units are numbered after all real module functions, avoiding identity collisions.

Attribution correctness

The PR also fixes incorrect attribution in branch canonicalization: replacement compares now inherit attribution from the compare they replace, rather than borrowing it from the enclosing branch.

Missing attribution remains missing. Instructions created without a valid attribution anchor are deliberately left unmapped instead of borrowing nearby provenance.

Known uncovered creation sites remain in parts of aggregate/ABI transformation and GVN/LSR/CFG-edit phi construction.

Verification

Sonatina:

• cargo test -p sonatina-ir
• cargo test -p sonatina-codegen
• all 819 Sonatina codegen unit tests and integration tests pass

Fe, using a temporary local Cargo patch:

• cargo test -p fe-codegen
• cargo test -p fe
• Fe codegen, CLI, trace, ethdebug, E2E, and formatter-roundtrip suites pass

[cburgdorf]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Nice work!

Reviewed commit: 9d9325e32d

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[cburgdorf]

@micahscopes @sbillig These changes are needed for tracing to work post-optimization. It's based on @micahscopes 's initial work.