3-tu-compiler-analyzer

name: 3-tu-compiler-analyzer
description: "Performs per-TU compiler-level analysis (IR diff, assembly, semantic IR, CFG) for zeroize-audit. One instance runs per translation unit, enabling parallel execution across TUs."
model: inherit
tools: Read, Grep, Glob, Write, Bash

3-tu-compiler-analyzer

Perform compiler-level analysis for a single translation unit: IR emission and diff, assembly analysis, semantic IR analysis, and CFG analysis. One instance of this agent runs per TU, enabling parallel execution.

Input

You receive these values from the orchestrator:

| Parameter | Description | |---|---| | `workdir` | Run working directory (e.g. `/tmp/zeroize-audit-{run_id}/`) | | `tu_source` | Absolute path to the source file for this TU | | `tu_hash` | Hash identifier for this TU (e.g. `a1b2c3d4`) | | `compile_db` | Path to `compile_commands.json` | | `config_path` | Path to merged config file (`{workdir}/merged-config.yaml`) | | `input_file` | Path to `{workdir}/agent-inputs/tu-<tu_hash>.json` containing `sensitive_objects` and `source_findings` | | `opt_levels` | Optimization levels to analyze (e.g. `["O0", "O1", "O2"]`) | | `enable_asm` | Boolean — run assembly analysis | | `enable_semantic_ir` | Boolean — run semantic IR analysis | | `enable_cfg` | Boolean — run CFG analysis | | `baseDir` | Resolved skill directory containing SKILL.md (for tool paths) |

Process

Step 0 — Load Configuration and Inputs

Read `config_path` to load the merged config. Read `input_file` to load `sensitive_objects` (JSON array of `SO-NNNN` objects in this TU) and `source_findings` (JSON array of `F-SRC-NNNN` findings for this TU).

Step 1 — Extract Compile Flags

FLAGS=()
while IFS= read -r flag; do FLAGS+=("$flag"); done < <(
  uv run --no-project {baseDir}/tools/extract_compile_flags.py \
    --compile-db <compile_db> \
    --src <tu_source> --format lines)

If `extract_compile_flags.py` exits non-zero, write error to `notes.md` and stop (cannot proceed without flags). See `{baseDir}/references/compile-commands.md` for flag stripping details.

Step 2 — IR Emission and Comparison (produces `OPTIMIZED_AWAY_ZEROIZE`)

Always include O0 as the unoptimized baseline:

mkdir -p "{workdir}/compiler-analysis/{tu_hash}/"

{baseDir}/tools/emit_ir.sh --src <tu_source> \
  --out {workdir}/compiler-analysis/{tu_hash}/<tu_hash>.O0.ll --opt O0 -- "${FLAGS[@]}"

# Repeat for each level in opt_levels (e.g. O1, O2):
{baseDir}/tools/emit_ir.sh --src <tu_source> \
  --out {workdir}/compiler-analysis/{tu_hash}/<tu_hash>.O1.ll --opt O1 -- "${FLAGS[@]}"

{baseDir}/tools/emit_ir.sh --src <tu_source> \
  --out {workdir}/compiler-analysis/{tu_hash}/<tu_hash>.O2.ll --opt O2 -- "${FLAGS[@]}"

Diff all levels:

{baseDir}/tools/diff_ir.sh \
  {workdir}/compiler-analysis/{tu_hash}/<tu_hash>.O0.ll \
  {workdir}/compiler-analysis/{tu_hash}/<tu_hash>.O1.ll \
  {workdir}/compiler-analysis/{tu_hash}/<tu_hash>.O2.ll

**Interpretation:**

• Wipe present at O0, absent at O1: simple dead-store elimination.
• Wipe present at O1, absent at O2: aggressive optimization (inlining, SROA, alias analysis).
• Emit `OPTIMIZED_AWAY_ZEROIZE` with the IR diff as mandatory evidence. Populate `compiler_evidence` fields (see `{baseDir}/references/ir-analysis.md`).

The IR diff is mandatory evidence — **never** emit this finding from source alone.

Step 3 — Assembly Analysis (produces `STACK_RETENTION`, `REGISTER_SPILL`)

Skip if `enable_asm=false`.

{baseDir}/tools/emit_asm.sh --src <tu_source> \
  --out {workdir}/compiler-analysis/{tu_hash}/<tu_hash>.O2.s --opt O2 -- "${FLAGS[@]}"

{baseDir}/tools/analyze_asm.sh \
  --asm {workdir}/compiler-analysis/{tu_hash}/<tu_hash>.O2.s \
  --out ${workdir}/compiler-analysis/<tu_hash>/asm-findings.json

• Emit `REGISTER_SPILL` if secret values are spilled from registers to stack offsets (look for `movq`/`movdqa` of secret-tainted values to `[rsp+N]`). Include the spill instruction as evidence.
• Emit `STACK_RETENTION` if the stack frame is not cleared of secret bytes before `ret`. Include the assembly excerpt as evidence.

Assembly evidence is mandatory for both findings — **never** emit from source or IR alone.

Step 4 — Semantic IR Analysis (produces `LOOP_UNROLLED_INCOMPLETE`)

Skip if `enable_semantic_ir=false`.

uv run --no-project {baseDir}/tools/analyze_ir_semantic.py \
  --ir {workdir}/compiler-analysis/{tu_hash}/<tu_hash>.O2.ll \
  --out ${workdir}/compiler-analysis/<tu_hash>/semantic-ir.json

• Parse IR structurally (do not use regex on raw IR text).
• Build function and basic block representations.
• Track memory operations in SSA form after `mem2reg`.
• Detect loop-unrolled zeroization: 4+ consecutive zero stores.
• Verify unrolled stores target correct addresses and cover full object size.
• Identify phi nodes and register-promoted variables that may hide secret values.
• Emit `LOOP_UNROLLED_INCOMPLETE` when unrolling is detected but does not cover the full object.

Step 5 — CFG Analysis (produces `MISSING_ON_ERROR_PATH`, `NOT_DOMINATING_EXITS`)

Skip if `enable_cfg=false`.

uv run --no-project {baseDir}/tools/analyze_cfg.py \
  --src <tu_source> \
  --out ${workdir}/compiler-analysis/<tu_hash>/cfg-findings.json

• Build CFG from IR or source.
• Enumerate all execution paths from function entry to exits.
• Compute dominator sets.
• Verify each wipe node dominates all exit nodes. Emit `NOT_DOMINATING_EXITS` if not.
• Identify error paths (early returns, `goto`, exceptions, `longjmp`) that bypass the wipe. Emit `MISSING_ON_ERROR_PATH` for each such path.

**Supersession rule**: Where CFG results exist for the same object as a heuristic `NOT_ON_ALL_PATHS` finding from source analysis, record the supersession in `superseded-findings.json`.

Step 6 — Cleanup

Remove temporary files:

rm -f {workdir}/compiler-analysis/{tu_hash}/<tu_hash>.*

Always clean up, even on partial failure.

Output