Section 04: IR Parser Hardening

Status: Not Started Goal: Fix known parsing failures so all 12 journeys produce complete, accurate parse results with zero parse_errors. After this section, J8 (generics) parses all monomorphized functions.

Context: The ir_parser.py regex _FUNC_NAME_RE = re.compile(r'@([\w.$]+)\s*\(') cannot parse quoted LLVM function names like @"_ori_first$24m$24int_int" produced by monomorphized generics. J8 works around this (enough unquoted functions exist to compute scores), but it’s a known gap documented in plans/code-journeys/overview.md.

WARNING (file size): ir_parser.py is already 503 lines — at the limit. Adding quoted name support (+20 lines) and invoke handling (+30 lines) will push it to ~550+ lines. As a prerequisite step for this section, split ir_parser.py into:

ir_parser.py (~300 lines) — Module/Function/Block data classes + parse_module() entry point

ir_parser_internal.py (~250 lines) — regex patterns, line-level parsing helpers, instruction classification

This split also makes it easier to test individual parsing functions in isolation.

Depends on: Nothing (independent fix).

04.0 Prerequisite: Split ir_parser.py

File(s): .claude/skills/code-journey/ir_parser.py (split into two files)

Split ir_parser.py into ir_parser.py (data classes + public API) and ir_parser_internal.py (regex patterns + line-level helpers)
Update all imports in arc_metrics.py, attribute_metrics.py, control_flow_metrics.py, instruction_metrics.py, extract-metrics.py — these import from ir_parser, which should remain the public API module
Verify python3 -m pytest tests/test_ir_parser.py passes after split

04.1 Quoted Function Names

File(s): .claude/skills/code-journey/ir_parser.py

LLVM uses quoted names when the identifier contains characters not valid in bare identifiers. Ori’s mangling uses $ which is valid, but some backends quote names with special characters.

Update _FUNC_NAME_RE to handle both bare and quoted names:

# Before:
_FUNC_NAME_RE = re.compile(r'@([\w.$]+)\s*\(')

# After: handles @name( and @"name"(
# Note: use [\w.$]+ for bare names (not \S+? which would match a single char)
_FUNC_NAME_RE = re.compile(r'@(?:"([^"]+)"|([\w.$]+))\s*\(')

Update _parse_function_header() to extract from the correct capture group:

name_match = _FUNC_NAME_RE.search(stripped)
if not name_match:
    return None
# Group 1 = quoted name, Group 2 = bare name
func_name = name_match.group(1) or name_match.group(2)
raw_name = f'@"{func_name}"' if name_match.group(1) else f"@{func_name}"

Fix name field for quoted functions: The Function.name field (without @) must strip quotes too. With the proposed regex, group(1) already gives the unquoted name. But raw_name.lstrip('@') in _parse_function_decl produces '"name"' (with quotes) for quoted names. Use func_name (from the regex group) directly instead of stripping from raw_name. This affects both _parse_function_def and _parse_function_decl since they share _parse_function_header.
Fix property predicates for quoted names: Properties that use raw_name.startswith(...) break with quoted names (e.g., @"_ori_..." does not start with @_ori_). Fix each to use self.name (without @ or quotes):
- is_user_function: self.raw_name.startswith("@_ori_") — fails for @"_ori_...". Fix: use self.name.startswith("_ori_")
- is_runtime_decl: self.raw_name.startswith("@ori_") — fails for @"ori_...". Fix: use self.name.startswith("ori_")
- is_entry_called: self.raw_name == "@_ori_main" — OK (main is never quoted)
- is_llvm_intrinsic: self.raw_name.startswith("@llvm.") — OK (intrinsics never quoted)

Add test with quoted name IR:

def test_quoted_function_name():
    ir = 'define fastcc i64 @"_ori_first$24m$24int_int"(i64 %0) {\nentry:\n  ret i64 %0\n}\n'
    module = parse_module(ir)
    assert '@"_ori_first$24m$24int_int"' in module.functions
    func = module.functions['@"_ori_first$24m$24int_int"']
    assert func.name == '_ori_first$24m$24int_int'  # No quotes in .name
    assert func.is_user_function  # Starts with _ori_

04.2 Multi-line Instruction Handling

File(s): .claude/skills/code-journey/ir_parser.py

The multi-line switch fix (already done) should be generalized for other multi-line constructs.

Audit for other multi-line patterns in LLVM IR:
- phi with many incoming values (wraps across lines)
- landingpad with multiple catch/filter clauses
- invoke with long to/unwind labels that may wrap
If any are found in journey IR, add continuation-line joining (same pattern as switch fix)
Parse invoke as a first-class instruction: invoke has a different syntax from call and is currently not handled by the RC counting regexes (_RC_INC_RE/_RC_DEC_RE only match call, not invoke; _RC_INVOKE_RE exists but is never used for balance counting):
```
%result = invoke fastcc i64 @func(i64 %0) to label %normal unwind label %cleanup
```
The parser must:
1. Recognize invoke as an opcode (currently it does extract it as opcode, but the downstream consumers ignore it)
2. Extract the callee name from invoke instructions (same as call)
3. Extract the to label %X unwind label %Y targets for CFG construction
```
_INVOKE_RE = re.compile(
    r'invoke\b.*@(?:"([^"]+)"|(\S+?))\s*\('  # callee name
)
_INVOKE_TARGETS_RE = re.compile(
    r'to\s+label\s+%(\S+)\s+unwind\s+label\s+%(\S+)'
)
```
Update arc_metrics.py: The _RC_INVOKE_RE already exists but is NEVER used in _count_rc_ops(). Either:
- Extend _RC_INC_RE / _RC_DEC_RE to also match invoke.*@ori_rc_inc etc., OR
- Add a separate count for invoke-based RC operations
- Verify: Does Ori’s codegen ever emit invoke @ori_rc_inc? If not (which is likely — RC functions don’t unwind), this is defensive hardening only.
Update extract_branch_targets() in ir_utils.py: Currently only handles br and switch. Must also handle invoke targets for correct CFG construction in control_flow_metrics.py and the new rc_state.py (Section 02).

04.3 Completion Checklist

ir_parser.py split into ir_parser.py + ir_parser_internal.py (each <=400 lines)
_FUNC_NAME_RE handles both @name and @"name" patterns
invoke instructions parsed with callee extraction and target extraction
extract_branch_targets() handles invoke to/unwind targets
arc_metrics.py counts RC ops in both call and invoke instructions
J8 (generics): all monomorphized functions parsed (0 parse errors)
All 12 journeys: 0 parse_errors in output
Tests cover: bare names, quoted names, names with $, empty module, invoke instructions
python3 -m pytest tests/test_ir_parser.py passes

Exit Criteria: parse_module() on J8’s IR returns a Module with zero parse_errors and includes all monomorphized function definitions (verified by count matching grep -c '^define' ir.txt).