Facet: Neuro-Symbolic Indirect-Call Analysis under Opaque Pointers

Facet resolves indirect calls in C programs compiled to LLVM bitcode, and keeps the evidence behind every call target it adds or removes.

Open the example report Read the paper on arXiv

The example is hand-authored to show the interface. Its numbers are not measured results.

The report

PartWhat it shows
Call locationsEvery indirect call that has a debug location, with its number of targets before and after the analysis.
Call graphThe listed call locations and the functions they can reach, drawn as one graph. Each edge is marked as kept, added or removed.
TargetsFor the selected call, each function that was added, removed or kept.
EvidenceFor each target, the check that supports it and where the analyzer recorded it.
SourceThe lines around the call, read from the source directory you supply.
ProfilesThe sound, checked and precise results of one run, chosen from a menu.
Pending decisionsChanges the analyzer proposed and did not adopt, each with its reason.

Method

Type-based indirect-call analyses resolve a call through a structure field by matching the field's type against the functions assigned to fields of that type. Since LLVM 17, pointers in the IR are opaque. The IR no longer says which field a call loads its function pointer from, so that matching key is gone.

Facet reconstructs the relation itself. For each indirect call it recovers the dispatched field from debug information. For each field it recovers the set of functions the program assigns to it, counting initializers, stores and aggregate copies. The two sides are joined by field identity.

Every change to the starting call graph is one fact: one edge added or one edge pruned. A fact is accepted only when a rule verifies it from program evidence, and it keeps a record of the rule, the evidence it used and the premise it holds under. A language model may be asked about the facts that remain undecided, among candidates the symbolic stages have already bounded, and its answers pass through the same rules.

One run yields two call graphs: one that preserves recall and one that is refined.

Results across language models

The language model answers only the questions the symbolic stages leave open. To see how much the result depends on which model answers them, the same pipeline ran on 14 open-source C programs three times: with Gemini 3.7 Flash, with Claude Opus 5.5 at high reasoning effort, and with GLM-5.3, a model from a third vendor, at maximum reasoning effort. The recall-preserving graph accepts no model decision, so it is the same in all three runs.

Mean F1 over the 14 programs

Recall-preserving graph, no model decision0.448
Refined graph, Gemini 3.7 Flash0.709
Refined graph, Claude Opus 5.5, high reasoning effort0.717
Refined graph, GLM-5.3, maximum reasoning effort0.714

F1 for each program

  • Recall-preserving graph, no model decision
  • Refined graph, Gemini 3.7 Flash
  • Refined graph, Claude Opus 5.5, high reasoning effort
  • Refined graph, GLM-5.3, maximum reasoning effort
Values as a table
ProgramRecall-preserving graphRefined graph, Gemini 3.7 FlashRefined graph, Claude Opus 5.5Refined graph, GLM-5.3
Mean of 140.4480.7090.7170.714
nginx0.2550.4570.5100.500
sqlite0.1060.5090.5150.515
dovecot0.3580.4810.4820.471
libarchive0.2670.7100.7180.740
tmux0.3370.6550.6550.655
redis0.8060.9040.9300.912
libjpeg-turbo0.1010.8050.8110.788
lua0.8280.9130.9140.914
gdbm0.5830.6930.6930.693
oniguruma0.5780.9890.9890.989
libxml20.1990.4740.4850.477
libdwarf0.5930.6310.6310.631
md4c0.7440.7440.7440.744
sudo0.5110.9620.9620.962

F1 compares the targets a graph gives each indirect call with the targets observed at that call while the program ran a workload, averaged over the calls. Observation is a lower bound: a target that was never observed counts against precision even when the call can reach it.

Earlier run, reported in the paper

The paper repeats the analysis under Gemini 2.5 Flash, Gemini 3.7 Flash and GLM-5.3. These are the mean F1 values it reports for the refined graph. Its Gemini 3.7 Flash value comes from an earlier run and differs slightly from the one above.

Refined graph, Gemini 2.5 Flash0.689
Refined graph, Gemini 3.7 Flash0.708
Refined graph, GLM-5.30.714

Requirements

Input
Whole-program LLVM bitcode, preferably built with debug information.
Where it runs
On your own machine. The browser workspace listens only on the loopback address.
API key
None for the symbolic analysis. The language model is optional.
Output
A self-contained HTML report that opens offline.

Availability

The source code will come soon.