Infer.AI is a framework for quickly developing abstract interpretation-based checkers (intraprocedural or interprocedural). You define only:
(1) An abstract domain (type of abstract state plus
(2) Transfer functions (a transformer that takes an abstract state as input and produces an abstract state as output)
and then you have an analysis that can run on all of the languages Infer supports (C, Obj-C, C++, and Java)!
If you feel like coding instead of reading, a great way to get started with Infer.AI is to go through the lab exercise here.
By example: intraprocedural analysis
This section helps you get started ASAP if you already understand abstract interpretation (or don't, but are feeling bold).
Take a look at liveness.ml. This code is performing a compilers-101 style liveness analysis over SIL, Infer's intermediate language. Since this code is fairly small and you should already understand what it's trying to do, it's a fairly good place to look in order to understand both how to use the abstract interpretation framework and what SIL is.
There are basically three important bits here: defining the domain, defining the transfer functions, and then passing the pieces to the framework to create an an analysis. Let's break down the third bit:
ProcCfg.Backward(ProcCfg.Exceptional) part says: "I want the direction of
iteration to be backward" (since liveness is a backward analysis), and "I want
to the analysis to follow exceptional edges". For a forward analysis that
ignores exceptional edges, you would do
ProcCfg.Normal instead (and many other
combinations are possible; take a look at
for more). And finally, the
TransferFunctions part says "Use the transfer
functions I defined above".
Now you have an
Analyzer module that exposes useful functions like
(take a procedure as input and compute a postcondition) and
(take a procedure and compute an invariant map from node id's to the pre/post at
each node). The next step is to hook your checker up to the Infer CLI. For the
liveness analysis, you would do this by exposing a function for running the
checker on a single procedure:
and then adding
Liveness.checker, checkers_enabled to the list of registered
you can then run
infer run -a checkers -- <your_build_command> to run your
checker on real code. See here for more
details on the build systems supported by Infer.
Basic error reporting
Useful analyses have output. Basic printing to stderr or stderr is good for
debugging, but to report a programmer-readable error that is tied to a source
code location, you'll want to use
Reporting.log_error. Some examples of
By example: interprocedural analysis
Let's assume you have already read and understood the "intraprocedural analysis" section and have an intraprocedural checker. The abstract interpretation framework makes it easy to convert your intraprocedural analysis into a modular interprocedural analysis. Let me emphasize the modular point once more; global analyses cannot be expressed in this framework.
To make your checker interprocedural, you need to:
(1) Define the type of procedure summaries for your analysis and add some boilerplate for storing your data alongside the summaries for other analyses
(2) Add logic for (a) using summaries in your transfer functions and (b) converting your intraprocedural abstract state to a summary.
A good example to look at here is siof.ml. Step (1) is just:
along with adding the
Here, the type of the abstract state and the type of the summary are the same,
which makes things easier for us (no logic to convert an abstract state to a
Part (2a) is here:
This says: "read the summary for
callee_pname from procedure
pdesc with type
tenv". You must then add logic for applying the summary to the
current abstract state (often, this is as simple as doing a join).
Because our summary type is the same as the abstract state, part (2b) can be
done for us by making use of the convenient
(for an example of what to do when the types are different, take a look at
Interprocedural module will automatically do the work of computing and
storing the summary for us. All we need to do is change the exposed
function registered in
registerCheckers.ml to call
That's it! We now have an interprocedural analysis.
One very important note here: a current (and soon-to-be-lifted) limitation prevents us from running multiple interprocedural checkers at the same time. If you register an interprocedural checker, be sure to unregister the other other ones. Otherwise, there's a risk that the checkers will clobber each other's results.
Some pointers to useful code for building new analyses, and to the implementation of the framework for the interested:
AbstractDomain.Pair(all in AbstractDomain)
Domains and domain building blocks:
Reporting errors with interprocedural traces:
IR basics: SIL, CFG's,
All of the languages analyzed by Infer are converted into a common intermediate representation. A program is represented as a control-flow graph (CFG) whose nodes contain lists of instructions in the SIL language. SIL is a small low-level language that has some similarities with C, LLVM IR, and Boogie.
are literals, program variables (
Pvar's), temporary variables (
field offset from a struct (OO features like objects are lowered into struct's),
or an index offset from an array.
There are four interesting kinds of
Load for reading into a temporary variable,
Store for writing to a program
variable, field of a struct, or an array,
Prune e (often called
other PL formalisms) blocks execution unless the expression
e evaluates to
Call represents function calls.
Instructions and expressions have
Tstruct (think: object) type has a
and it is often useful to look up metadata about the type (what fields does it
have, what methods does it declare, what is its superclass, etc.) in the type
A procedure description or
pdesc) is an abstraction of a procedure declaration: it
stores the CFG of the procedure, its signature, its annotations, and so on.
A procedure name or
pname) is an abstraction of a called procedure name.
One procname may correspond to multiple (or zero)
procdesc's after resolution.
The abstract interpretation framework has a few additional constructs that are worth explaining.
represents the CFG of a single procedure whereas (perhaps confusingly) a
the CFG for an entire file. A
ProcCfg is really a customizable view of the
underlying procedure CFG; we can get a view the CFG with its edges backward
ProcCfg.Backward), with or without exceptional edges (
respectively), or with each node holding at most one instruction
is a container that holds all of the read-only information required to analyze a
single procedure: procedure description, and
extras are custom
read-only data that are computed before analysis begins, and can be accessed
from the transfer functions. Most often, no extras are required for analysis
ProcData.empty_extras), but it can be useful to stash information like a map
from a formal to its
or an invariant
from a prior analysis in the extras.
How it works
Intro: abstract interpretation
How do I make an analysis compositional?