angr's solver engine is called Claripy. Claripy exposes the following design:
Claripy ASTs (the subclasses of claripy.ast.Base) provide a unified way to interact with concrete and symbolic expressions
Frontend
s provide different paradigms for evaluating these expressions. For example, the FullFrontend
solves expressions using something like an SMT solver backend, while LightFrontend
handles them by using an abstract (and approximating) data domain backend.
Each Frontend
needs to, at some point, do actual operation and evaluations on an AST. ASTs don't support this on their own. Instead, Backend
s translate ASTs into backend objects (i.e., python primitives for BackendConcrete
, Z3 expressions for BackendZ3
, strided intervals for BackendVSA
, etc) and handle any appropriate statetracking objects (such as tracking the solver state in the case of BackendZ3
). Roughly speaking, frontends take ASTs as inputs and use backends to backend.convert()
those ASTs into backend objects that can be evaluated and otherwise reasoned about.
FrontendMixin
s customize the operation of Frontend
s. For example, ModelCacheMixin
caches solutions from an SMT solver.
The combination of a Frontend, a number of FrontendMixins, and a number of Backends comprise a claripy Solver
.
Internally, Claripy seamlessly mediates the cooperation of multiple disparate backends  concrete bitvectors, VSA constructs, and SAT solvers. It is pretty badass.
Most users of angr will not need to interact directly with Claripy (except for, maybe, claripy AST objects, which represent symbolic expressions)  angr handles most interactions with Claripy internally. However, for dealing with expressions, an understanding of Claripy might be useful.
Claripy ASTs abstract away the differences between mathematical constructs that Claripy supports. They define a tree of operations (i.e., (a + b) / c)
on any type of underlying data. Claripy handles the application of these operations on the underlying objects themselves by dispatching requests to the backends.
Currently, Claripy supports the following types of ASTs:
Name  Description  Supported By (Claripy Backends)  Example Code 
BV  This is a bitvector, whether symbolic (with a name) or concrete (with a value). It has a size (in bits).  BackendConcrete, BackendVSA, BackendZ3  Create a 32bit symbolic bitvector "x": `claripy.BVS('x', 32)` Create a 32bit bitvector with the value `0xc001b3475`: `claripy.BVV(0xc001b3a75, 32)` Create a 32bit "strided interval" (see VSA documentation) that can be any divisibleby10 number between 1000 and 2000: `claripy.SI(name='x', bits=32, lower_bound=1000, upper_bound=2000, stride=10)` 
FP  This is a floatingpoint number, whether symbolic (with a name) or concrete (with a value).  BackendConcrete, BackendZ3  TODO 
Bool  This is a boolean operation (True or False).  BackendConcrete, BackendVSA, BackendZ3 

All of the above creation code returns claripy.AST objects, on which operations can then be carried out.
ASTs provide several useful operations.
>>> import claripyâ€‹>>> bv = claripy.BVV(0x41424344, 32)â€‹# Size  you can get the size of an AST with .size()>>> assert bv.size() == 32â€‹# Reversing  .reversed is the reversed version of the BVV>>> assert bv.reversed is claripy.BVV(0x44434241, 32)>>> assert bv.reversed.reversed is bvâ€‹# Depth  you can get the depth of the AST>>> print(bv.depth)>>> assert bv.depth == 1>>> x = claripy.BVS('x', 32)>>> assert (x+bv).depth == 2>>> assert ((x+bv)/10).depth == 3
Applying a condition (==, !=, etc) on ASTs will return an AST that represents the condition being carried out. For example:
>>> r = bv == x>>> assert isinstance(r, claripy.ast.Bool)â€‹>>> p = bv == bv>>> assert isinstance(p, claripy.ast.Bool)>>> assert p.is_true()
You can combine these conditions in different ways.
>>> q = claripy.And(claripy.Or(bv == x, bv * 2 == x, bv * 3 == x), x == 0)>>> assert isinstance(p, claripy.ast.Bool)
The usefulness of this will become apparent when we discuss Claripy solvers.
In general, Claripy supports all of the normal python operations (+, , , ==, etc), and provides additional ones via the Claripy instance object. Here's a list of available operations from the latter.
Name  Description  Example 
LShR  Logically shifts a bit expression (BVV, BV, SI) to the right. 

SignExt  Signextends a bit expression. 

ZeroExt  Zeroextends a bit expression. 

Extract  Extracts the given bits (zeroindexed from the right, inclusive) from a bit expression.  Extract the rightmost byte of x: 
Concat  Concatenates several bit expressions together into a new bit expression. 

RotateLeft  Rotates a bit expression left. 

RotateRight  Rotates a bit expression right. 

Reverse  Reverses a bit expression. 

And  Logical And (on boolean expressions) 

Or  Logical Or (on boolean expressions) 

Not  Logical Not (on a boolean expression) 

If  An Ifthenelse  Choose the maximum of two expressions: 
ULE  Unsigned less than or equal to.  Check if x is less than or equal to y: 
ULT  Unsigned less than.  Check if x is less than y: 
UGE  Unsigned greater than or equal to.  Check if x is greater than or equal to y: 
UGT  Unsigned greater than.  Check if x is greater than y: 
SLE  Signed less than or equal to.  Check if x is less than or equal to y: 
SLT  Signed less than.  Check if x is less than y: 
SGE  Signed greater than or equal to.  Check if x is greater than or equal to y: 
SGT  Signed greater than.  Check if x is greater than y: 
NOTE: The default python >
, <
, >=
, and <=
are unsigned in Claripy. This is different than their behavior in Z3, because it seems more natural in binary analysis.
The main point of interaction with Claripy are the Claripy Solvers. Solvers expose an API to interpret ASTs in different ways and return usable values. There are several different solvers.
Name  Description 
Solver  This is analogous to a 
SolverVSA  This solver uses VSA to reason about values. It is an approximating solver, but produces values without performing actual constraint solves. 
SolverReplacement  This solver acts as a passthrough to a child solver, allowing the replacement of expressions onthefly. It is used as a helper by other solvers and can be used directly to implement exotic analyses. 
SolverHybrid  This solver combines the SolverReplacement and the Solver (VSA and Z3) to allow for approximating values. You can specify whether or not you want an exact result from your evaluations, and this solver does the rest. 
SolverComposite  This solver implements optimizations that solve smaller sets of constraints to speed up constraint solving. 
Some examples of solver usage:
# create the solver and an expression>>> s = claripy.Solver()>>> x = claripy.BVS('x', 8)â€‹# now let's add a constraint on x>>> s.add(claripy.ULT(x, 5))â€‹>>> assert sorted(s.eval(x, 10)) == [0, 1, 2, 3, 4]>>> assert s.max(x) == 4>>> assert s.min(x) == 0â€‹# we can also get the values of complex expressions>>> y = claripy.BVV(65, 8)>>> z = claripy.If(x == 1, x, y)>>> assert sorted(s.eval(z, 10)) == [1, 65]â€‹# and, of course, we can add constraints on complex expressions>>> s.add(z % 5 != 0)>>> assert s.eval(z, 10) == (1,)>>> assert s.eval(x, 10) == (1,) # interestingly enough, since z can't be y, x can only be 1!
Custom solvers can be built by combining a Claripy Frontend (the class that handles the actual interaction with SMT solver or the underlying data domain) and some combination of frontend mixins (that handle things like caching, filtering out duplicate constraints, doing opportunistic simplification, and so on).
Backends are Claripy's workhorses. Claripy exposes ASTs to the world, but when actual computation has to be done, it pushes those ASTs into objects that can be handled by the backends themselves. This provides a unified interface to the outside world while allowing Claripy to support different types of computation. For example, BackendConcrete provides computation support for concrete bitvectors and booleans, BackendVSA introduces VSA constructs such as StridedIntervals (and details what happens when operations are performed on them, and BackendZ3 provides support for symbolic variables and constraint solving.
There are a set of functions that a backend is expected to implement. For all of these functions, the "public" version is expected to be able to deal with claripy's AST objects, while the "private" version should only deal with objects specific to the backend itself. This is distinguished with Python idioms: a public function will be named func() while a private function will be _func(). All functions should return objects that are usable by the backend in its private methods. If this can't be done (i.e., some functionality is being attempted that the backend can't handle), the backend should raise a BackendError. In this case, Claripy will move on to the next backend in its list.
All backends must implement a convert()
function. This function receives a claripy AST and should return an object that the backend can handle in its private methods. Backends should also implement a _convert()
method, which will receive anything that is not a claripy AST object (i.e., an integer or an object from a different backend). If convert()
or _convert()
receives something that the backend can't translate to a format that is usable internally, the backend should raise BackendError, and thus won't be used for that object. All backends must also implement any functions of the base Backend
abstract class that currently raise NotImplementedError()
.
Claripy's contract with its backends is as follows: backends should be able to handle, in their private functions, any object that they return from their private or public functions. Claripy will never pass an object to any backend private function that did not originate as a return value from a private or public function of that backend. One exception to this is convert()
and _convert()
, as Claripy can try to stuff anything it feels like into _convert() to see if the backend can handle that type of object.
To perform actual, useful computation on ASTs, Claripy uses backend objects. A BackendObject
is a result of the operation represented by the AST. Claripy expects these objects to be returned from their respective backends, and will pass such objects into that backend's other functions.