| name | classes |
| description | Recover C++ class hierarchies, vtables, and RTTI from a compiled binary via r2xsql's classes and class_methods tables. Use when asked about class/inheritance structure, vtable layout, virtual method tables, or 'what C++ classes does this binary define'. |
| allowed-tools | ["Bash","Read"] |
When to use
Pick this skill when the task is understanding a binary's C++ object model,
not ordinary function-level disassembly/decompilation:
- listing the classes a binary defines (with their direct base class)
- reading a class's vtable layout (method name, address, vtable slot)
- walking a base/derived relationship (single inheritance)
For ordinary function analysis, use disassembly or decompiler instead —
this skill is specifically about the class-level structure radare2's RTTI
recovery produces.
The two "classes" concepts in radare2 — use the right one
radare2 has two unrelated things both called "classes":
icj (RBinClass) — a bin-format reader for Java .class,
Objective-C, and .NET class tables, populated at load time by the matching
bin plugin. Structurally empty for every native PE/ELF binary,
regardless of whether it has C++ classes. Never use it here.
aclj (anal-side RAnalClass) — a separate, sdb-backed class
database populated by avrr (vtable search + MSVC/Itanium RTTI parsing).
This is what classes/class_methods are built on.
There is also no acj command — the top-level ac dispatcher has no j
case; the JSON list verb is aclj (ac + l + j).
The avrr non-idempotency rule (read this before touching raw commands)
radare2's own aaa already runs avrr once, internally, at file-open
time. A second explicit avrr call is confirmed NOT idempotent — it
duplicates every already-recovered class under a _1-suffixed name
(Animal → Animal, Animal_1, …). r2xsql's classes/class_methods
tables never issue avrr (or the raw av vtable search) themselves, on any
code path, including a repeat query in the same session — they only ever
read whatever aaa already recovered.
If you drop to raw radare2 commands (.avrr, .av) while working with
these tables, do not run avrr a second time in the same session — it
will corrupt the very data you're trying to read, and r2xsql cannot protect
you from a manually-triggered repeat call.
If aaa hasn't run yet in the current session (e.g. opened with
--no-analyze), both tables read as honestly empty — that is the
correct, non-broken answer, not a sign anything needs to be triggered.
Tables
| table | source | columns |
|---|
classes | aclj | name, base_name (first direct base only, NULL for a root class), vtable_addr (NULL if none) |
class_methods | aclj (methods[] per class) | class_name (FK to classes.name), name, addr, vtable_offset (NULL for a non-virtual method) |
No predicate is required or useful — no pushdown exists, and the recovered
class set is small (tens to low hundreds of classes on a realistic binary).
Common queries
SELECT name, base_name FROM classes ORDER BY name;
SELECT name, printf('0x%x', addr) AS at, vtable_offset
FROM class_methods WHERE class_name = 'Dog' ORDER BY vtable_offset;
SELECT name, vtable_addr FROM classes WHERE base_name IS NULL;
SELECT name FROM classes WHERE base_name = 'Animal';
SELECT c.name, c.base_name, COUNT(m.name) AS n_methods
FROM classes c LEFT JOIN class_methods m ON m.class_name = c.name
GROUP BY c.name ORDER BY c.name;
SELECT name, base_name FROM classes
WHERE name NOT LIKE 'std::%' AND name ;
Caveats
-
Uncurated by design. Every class aclj reports is visible, including
CRT-internal recoveries from a statically-linked runtime (type_info,
std::exception, std::bad_alloc, std::bad_array_new_length, …). This
mirrors how the family-shared names view stays uncurated — r2xsql does
not invent a curation boundary radare2 itself never drew. Filter
client-side if you want only a target program's own classes.
-
Multiple inheritance is not fully modeled. base_name is a single
nullable column, not a bases[] array. A class with more than one direct
base loses every base past the first here — the first base is genuinely
correct, just narrower than the raw recovery. There is no
classes_bases junction table in this version; if a query needs full
multi-base fidelity, drop to raw acllj <name> for that one class.
-
The Itanium ABI recovers materially less than MSVC — measured, not
assumed. Same C++ source, same radare2 (6.1.7), compiled both ways:
| MSVC / PE | Itanium / ELF (gcc 13.3) |
|---|
| classes recovered | 3 (Animal, Dog, Bird) | 2 — the abstract base Animal is MISSING |
base_name | Dog→Animal, Bird→Animal | empty — no inheritance recovered at all |
| methods per class | 3 | 4 |
This is radare2's avrr recovery, not an r2xsql defect — r2xsql surfaces
whatever avrr produces, and both backends agree with each other on both
platforms. Practical consequence on GCC/Clang targets: do not rely on
base_name to reconstruct a hierarchy, and do not treat a missing class as
proof it does not exist. Cross-check with vtable_addr and the raw
acllj output before drawing conclusions. On MSVC targets the full ground
truth (exact method counts, exact base relationships) holds.
-
Names may be demangled or still-mangled, depending on how much of a
given RTTI record radare2's demangler covers — match with rather than exact equality if a lookup unexpectedly returns
nothing.