Repository navigation
C extractor: functions with #ifdef-split braces are dropped from the graph (no Function node) #961
Description
Summary
A C function whose body braces are split across #ifdef/#else branches is silently dropped from the code knowledge graph — no Function node is emitted for it, even though its callers remain and reference it as an unresolved name.
Root cause
The extractor parses raw source with tree-sitter without running the preprocessor, so both #ifdef branches are present at once. When a construct like this opens a brace in each branch but shares one close brace:
#ifdef _WIN32
if (a() && b()) { // open brace (branch 1)
...
#else
if (c() && d()) { // a different open brace (branch 2)
#endif
...shared body...
} // single shared close bracetree-sitter sees two if { opens and one } close → unbalanced braces → an ERROR node spanning the function body → no Function node is created. Recovery is localized (neighbouring functions extract fine), so the symptom is one function silently absent.
Reproduction (reproduce-first)
Found via a graph-vs-grep audit: cbm_path_within_root in src/mcp/mcp.c was absent from the graph while its 6 callers were present. Refactoring that specific function to a single unconditional control flow (#960) made it index. The general pattern remains unhandled.
Suggested RED reproduction for the bug-repro suite: a fixture C function with the #ifdef-splits-brace shape, asserting a Function node is produced for it. It should fail on the current extractor and pass once the extractor handles the pattern.
Impact
Low frequency but real: any function using this (uncommon) preprocessor pattern is invisible to search_graph/trace_path/callers, and dead-code / impact analysis will misjudge it. Also a latent correctness gap for graph-based audits.
Possible directions (extractor-side, harder than the per-function workaround)
- Detect the
ERROR-node-spanning-a-definition case and attempt a light preprocessor pass (pick one#ifdefbranch) before re-parsing that region. - Or a targeted brace-rebalancing heuristic for
#ifdef/#else/#endifregions inside a definition.
Not urgent; filing so the limitation is tracked rather than rediscovered.
Activity
DeusData commentedon Jul 9, 2026
The coverage signal (#963, shipped in #968 / 97c6750) now detects this class automatically. A fresh full self-index flags 44 files as parse_partial; drilling into the two largest flagged regions:
src/mcp/mcp.c[4106-6406]: 49 functions in the source range, 48 recovered into the graph, exactly one missing —cbm_path_within_root(the known instance; its body contains an#ifdef _WIN32split).src/ui/http_server.c[574-1868]: 31 functions in range, one missing —handle_process_kill(new instance, same in-function#ifdef _WIN32platform split).Smaller recurring classes from the same sweep (containing functions all recovered, only inner declarations lost):
static CBM_TLS ...declarations across ~20 pipeline files, and the#ifdef __cplusplus }extern-C close braces in the vendored tree-sitter headers.Current confirmed loss list for this issue:
cbm_path_within_root,handle_process_kill. The full flagged set is queryable any time viaindex_statusorquery_graph(graph="missed").DeusData commentedon Jul 12, 2026
Fixed on main by 65a176e (merged in db330bf, PR #1050), shipping in v0.9.1.
The general pattern is now handled, exactly along the lines the issue suggested: the preprocessed second pass (simplecpp, which already existed for macro-hidden CALLS) evaluates the conditionals so one branch survives and the tree parses clean — and since same-file token lines stay aligned, definitions recovered from it carry their original line numbers. Recovery is deliberately narrow: a def is adopted only when it intersects a raw-tree ERROR region, its name is visible on the corresponding original source line, and its qualified name wasn't already extracted — so clean files never gain phantom or duplicate defs.
Both confirmed instances verified with the fixed binary:
handle_process_kill(http_server.c, 580–647) andcbm_path_within_root(mcp.c, 5070–5085) now index with correct ranges — the second no longer only thanks to the #960 refactor. The reproduce-first fixture from this issue (#ifdef-splits-brace shape) is in the suite as the permanent guard, plus an inverse guard that clean#ifdeffiles gain nothing.The parse_partial coverage signal (#963) stays the detector for any residual instances of this class — a fresh self-index after this fix is the easiest way to confirm the flagged-region count drops.
fix(extract): recover C/C++ defs lost to #ifdef-split brace ERROR reg…
fix(extract): recover C/C++ defs lost to #ifdef-split brace ERROR reg…
fix(extract): one definition per site when the raw and preprocessed p…
fix(cpp): one spelling per template specialization in the preprocesse…