<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01//EN"
"http://www.w3.org/TR/html4/strict.dtd">
<html>
<head>
<title>Checker Developer Manual</title>
<link type="text/css" rel="stylesheet" href="menu.css">
<link type="text/css" rel="stylesheet" href="content.css">
<script type="text/javascript" src="scripts/menu.js"></script>
</head>
<body>
<div id="page">
<!--#include virtual="menu.html.incl"-->
<div id="content">
<h3 style="color:red">This Page Is Under Construction</h3>
<h1>Checker Developer Manual</h1>
<p>The static analyzer engine performs path-sensitive exploration of the program and
relies on a set of checkers to implement the logic for detecting and
constructing specific bug reports. Anyone who is interested in implementing their own
checker, should check out the Building a Checker in 24 Hours talk
(<a href="https://llvm.org/devmtg/2012-11/Zaks-Rose-Checker24Hours.pdf">slides</a>
<a href="https://youtu.be/kdxlsP5QVPw">video</a>)
and refer to this page for additional information on writing a checker. The static analyzer is a
part of the Clang project, so consult <a href="https://clang.llvm.org/hacking.html">Hacking on Clang</a>
and <a href="https://llvm.org/docs/ProgrammersManual.html">LLVM Programmer's Manual</a>
for developer guidelines and post your questions and proposals to the
<a href="https://discourse.llvm.org/c/clang/static-analyzer/"> Static Analyzer</a> subcategory at
the official <a href="https://discourse.llvm.org/"> LLVM Discourse server</a>.
</p>
<ul>
<li><a href="#start">Getting Started</a></li>
<li><a href="#analyzer">Static Analyzer Overview</a>
<ul>
<li><a href="#interaction">Interaction with Checkers</a></li>
<li><a href="#values">Representing Values</a></li>
</ul></li>
<li><a href="#idea">Idea for a Checker</a></li>
<li><a href="#registration">Checker Registration</a></li>
<li><a href="#events_callbacks">Events, Callbacks, and Checker Class Structure</a></li>
<li><a href="#extendingstates">Custom Program States</a></li>
<li><a href="#bugs">Bug Reports</a></li>
<li><a href="#ast">AST Visitors</a></li>
<li><a href="#testing">Testing</a></li>
<li><a href="#commands">Useful Commands/Debugging Hints</a>
<ul>
<li><a href="#attaching">Attaching the Debugger</a></li>
<li><a href="#narrowing">Narrowing Down the Problem</a></li>
<li><a href="#visualizing">Visualizing the Analysis</a></li>
<li><a href="#debugprints">Debug Prints and Tricks</a></li>
</ul></li>
<li><a href="#additioninformation">Additional Sources of Information</a></li>
<li><a href="#links">Useful Links</a></li>
</ul>
<h2 id=start>Getting Started</h2>
<ul>
<li>To check out the source code and build the project, follow steps 1-4 of
the <a href="https://clang.llvm.org/get_started.html">Clang Getting Started</a>
page.</li>
<li>The analyzer source code is located under the Clang source tree:
<br><tt>
$ <b>cd llvm/tools/clang</b>
</tt>
<br>See: <tt>include/clang/StaticAnalyzer</tt>, <tt>lib/StaticAnalyzer</tt>,
<tt>test/Analysis</tt>.</li>
<li>The analyzer regression tests can be executed from the Clang's build
directory:
<br><tt>
$ <b>cd ../../../; cd build/tools/clang; TESTDIRS=Analysis make test</b>
</tt></li>
<li>Analyze a file with the specified checker:
<br><tt>
$ <b>clang -cc1 -analyze -analyzer-checker=core.DivideZero test.c</b>
</tt></li>
<li>List the available checkers:
<br><tt>
$ <b>clang -cc1 -analyzer-checker-help</b>
</tt></li>
<li>See the analyzer help for different output formats, fine tuning, and
debug options:
<br><tt>
$ <b>clang -cc1 -help | grep "analyzer"</b>
</tt></li>
</ul>
<h2 id=analyzer>Static Analyzer Overview</h2>
The analyzer core performs symbolic execution of the given program. All the
input values are represented with symbolic values; further, the engine deduces
the values of all the expressions in the program based on the input symbols
and the path. The execution is path sensitive and every possible path through
the program is explored. The explored execution traces are represented with
<a href="https://clang.llvm.org/doxygen/classclang_1_1ento_1_1ExplodedGraph.html">ExplodedGraph</a> object.
Each node of the graph is
<a href="https://clang.llvm.org/doxygen/classclang_1_1ento_1_1ExplodedNode.html">ExplodedNode</a>,
which consists of a <tt>ProgramPoint</tt> and a <tt>ProgramState</tt>.
<p>
<a href="https://clang.llvm.org/doxygen/classclang_1_1ProgramPoint.html">ProgramPoint</a>
represents the corresponding location in the program (or the CFG).
<tt>ProgramPoint</tt> is also used to record additional information on
when/how the state was added. For example, <tt>PostPurgeDeadSymbolsKind</tt>
kind means that the state is the result of purging dead symbols - the
analyzer's equivalent of garbage collection.
<p>
<a href="https://clang.llvm.org/doxygen/classclang_1_1ento_1_1ProgramState.html">ProgramState</a>
represents abstract state of the program. It consists of:
<ul>
<li><tt>Environment</tt> - a mapping from source code expressions to symbolic
values
<li><tt>Store</tt> - a mapping from memory locations to symbolic values
<li><tt>GenericDataMap</tt> - constraints on symbolic values
</ul>
<h3 id=interaction>Interaction with Checkers</h3>
<p>
Checkers are not merely passive receivers of the analyzer core changes - they
actively participate in the <tt>ProgramState</tt> construction through the
<tt>GenericDataMap</tt> which can be used to store the checker-defined part
of the state. Each time the analyzer engine explores a new statement, it
notifies each checker registered to listen for that statement, giving it an
opportunity to either report a bug or modify the state. (As a rule of thumb,
the checker itself should be stateless.) The checkers are called one after another
in the predefined order; thus, calling all the checkers adds a chain to the
<tt>ExplodedGraph</tt>.
</p>
<h3 id=values>Representing Values</h3>
<p>
During symbolic execution, <a href="https://clang.llvm.org/doxygen/classclang_1_1ento_1_1SVal.html">SVal</a>
objects are used to represent the semantic evaluation of expressions.
They can represent things like concrete
integers, symbolic values, or memory locations (which are memory regions).
They are a discriminated union of "values", symbolic and otherwise.
If a value isn't symbolic, usually that means there is no symbolic
information to track. For example, if the value was an integer, such as
<tt>42</tt>, it would be a <a href="https://clang.llvm.org/doxygen/classclang_1_1ento_1_1nonloc_1_1ConcreteInt.html">ConcreteInt</a>,
and the checker doesn't usually need to track any state with the concrete
number. In some cases, <tt>SVal</tt> is not a symbol, but it really should be
a symbolic value. This happens when the analyzer cannot reason about something
(yet). An example is floating point numbers. In such cases, the
<tt>SVal</tt> will evaluate to <a href="https://clang.llvm.org/doxygen/classclang_1_1ento_1_1UnknownVal.html">UnknownVal</a>.
This represents a case that is outside the realm of the analyzer's reasoning
capabilities. <tt>SVals</tt> are value objects and their values can be viewed
using the <tt>.dump()</tt> method. Often they wrap persistent objects such as
symbols or regions.
</p>
<p>
<a href="https://clang.llvm.org/doxygen/classclang_1_1ento_1_1SymExpr.html">SymExpr</a> (symbol)
is meant to represent abstract, but named, symbolic value. Symbols represent
an actual (immutable) value. We might not know what its specific value is, but
we can associate constraints with that value as we analyze a path. For
example, we might record that the value of a symbol is greater than
<tt>0</tt>, etc.
</p>
<p>
<a href="https://clang.llvm.org/doxygen/classclang_1_1ento_1_1MemRegion.html">MemRegion</a> is similar to a symbol.
It is used to provide a lexicon of how to describe abstract memory. Regions can
layer on top of other regions, providing a layered approach to representing memory.
For example, a struct object on the stack might be represented by a <tt>VarRegion</tt>,
but a <tt>FieldRegion</tt> which is a subregion of the <tt>VarRegion</tt> could
be used to represent the memory associated with a specific field of that object.
So how do we represent symbolic memory regions? That's what
<a href="https://clang.llvm.org/doxygen/classclang_1_1ento_1_1SymbolicRegion.html">SymbolicRegion</a>
is for. It is a <tt>MemRegion</tt> that has an associated symbol. Since the
symbol is unique and has a unique name; that symbol names the region.
</p>
<p>
Let's see how the analyzer processes the expressions in the following example:
</p>
<p>
<pre class="code_example">
int foo(int x) {
int y = x * 2;
int z = x;
...
}
</pre>
</p>
<p>
Let's look at how <tt>x*2</tt> gets evaluated. When <tt>x</tt> is evaluated,
we first construct an <tt>SVal</tt> that represents the lvalue of <tt>x</tt>, in
this case it is an <tt>SVal</tt> that references the <tt>MemRegion</tt> for <tt>x</tt>.
Afterwards, when we do the lvalue-to-rvalue conversion, we get a new <tt>SVal</tt>,
which references the value <b>currently bound</b> to <tt>x</tt>. That value is
symbolic; it's whatever <tt>x</tt> was bound to at the start of the function.
Let's call that symbol <tt>$0</tt>. Similarly, we evaluate the expression for <tt>2</tt>,
and get an <tt>SVal</tt> that references the concrete number <tt>2</tt>. When
we evaluate <tt>x*2</tt>, we take the two <tt>SVals</tt> of the subexpressions,
and create a new <tt>SVal</tt> that represents their multiplication (which in
this case is a new symbolic expression, which we might call <tt>$1</tt>). When we
evaluate the assignment to <tt>y</tt>, we again compute its lvalue (a <tt>MemRegion</tt>),
and then bind the <tt>SVal</tt> for the RHS (which references the symbolic value <tt>$1</tt>)
to the <tt>MemRegion</tt> in the symbolic store.
<br>
The second line is similar. When we evaluate <tt>x</tt> again, we do the same
dance, and create an <tt>SVal</tt> that references the symbol <tt>$0</tt>. Note, two <tt>SVals</tt>
might reference the same underlying values.
</p>
<p>
To summarize, MemRegions are unique names for blocks of memory. Symbols are
unique names for abstract symbolic values. Some MemRegions represents abstract
symbolic chunks of memory, and thus are also based on symbols. SVals are just
references to values, and can reference either MemRegions, Symbols, or concrete
values (e.g., the number 1).
</p>
<!--
TODO: Add a picture.
<br>
Symbols<br>
FunctionalObjects are used throughout.
-->
<h2 id=idea>Idea for a Checker</h2>
Here are several questions which you should consider when evaluating your
checker idea:
<ul>
<li>Can the check be effectively implemented without path-sensitive
analysis? See <a href="#ast">AST Visitors</a>.</li>
<li>How high the false positive rate is going to be? Looking at the occurrences
of the issue you want to write a checker for in the existing code bases might
give you some ideas. </li>
<li>How the current limitations of the analysis will effect the false alarm
rate? Currently, the analyzer only reasons about one procedure at a time (no
inter-procedural analysis). Also, it uses a simple range tracking based
solver to model symbolic execution.</li>
<li>Consult the <a
href="https://github.com/llvm/llvm-project/labels/clang%3Astatic%20analyzer">GitHub Issues</a>
to get some ideas for new checkers and consider starting with improving/fixing
bugs in the existing checkers.</li>
</ul>
<p>Once an idea for a checker has been chosen, there are two key decisions that
need to be made:
<ul>
<li> Which events the checker should be tracking. This is discussed in more
detail in the section <a href="#events_callbacks">Events, Callbacks, and
Checker Class Structure</a>.
<li> What checker-specific data needs to be stored as part of the program
state (if any). This should be minimized as much as possible. More detail about
implementing custom program state is given in section <a
href="#extendingstates">Custom Program States</a>.
</ul>
<h2 id=registration>Checker Registration</h2>
All checker implementation files are located in
<tt>clang/lib/StaticAnalyzer/Checkers</tt> folder. The steps below describe
how the checker <tt>SimpleStreamChecker</tt>, which checks for misuses of
stream APIs, was registered with the analyzer.
Similar steps should be followed for a new checker.
<ol>
<li>A new checker implementation file, <tt>SimpleStreamChecker.cpp</tt>, was
created in the directory <tt>lib/StaticAnalyzer/Checkers</tt>.
<li>The following registration code was added to the implementation file:
<pre class="code_example">
void ento::registerSimpleStreamChecker(CheckerManager &mgr) {
mgr.registerChecker<SimpleStreamChecker>();
}
</pre>
<li>A package was selected for the checker and the checker was defined in the
table of checkers at <tt>include/clang/StaticAnalyzer/Checkers/Checkers.td</tt>.
Since all checkers should first be developed as "alpha", and the SimpleStreamChecker
performs UNIX API checks, the correct package is "alpha.unix", and the following
was added to the corresponding <tt>UnixAlpha</tt> section of <tt>Checkers.td</tt>:
<pre class="code_example">
let ParentPackage = UnixAlpha in {
...
def SimpleStreamChecker : Checker<"SimpleStream">,
HelpText<"Check for misuses of stream APIs">,
DescFile<"SimpleStreamChecker.cpp">;
...
} // end "alpha.unix"
</pre>
<li>The source code file was made visible to CMake by adding it to
<tt>lib/StaticAnalyzer/Checkers/CMakeLists.txt</tt>.
</ol>
After adding a new checker to the analyzer, one can verify that the new checker
was successfully added by seeing if it appears in the list of available checkers:
<br> <tt><b>$clang -cc1 -analyzer-checker-help</b></tt>
<h2 id=events_callbacks>Events, Callbacks, and Checker Class Structure</h2>
<p> All checkers inherit from the <tt><a
href="https://clang.llvm.org/doxygen/classclang_1_1ento_1_1Checker.html">
Checker</a></tt> template class; the template parameter(s) describe the type of
events that the checker is interested in processing. The various types of events
that are available are described in the file <a
href="https://clang.llvm.org/doxygen/CheckerDocumentation_8cpp_source.html">
CheckerDocumentation.cpp</a>
<p> For each event type requested, a corresponding callback function must be
defined in the checker class (<a
href="https://clang.llvm.org/doxygen/CheckerDocumentation_8cpp_source.html">
CheckerDocumentation.cpp</a> shows the
correct function name and signature for each event type).
<p> As an example, consider <tt>SimpleStreamChecker</tt>. This checker needs to
take action at the following times:
<ul>
<li>Before making a call to a function, check if the function is <tt>fclose</tt>.
If so, check the parameter being passed.
<li>After making a function call, check if the function is <tt>fopen</tt>. If
so, process the return value.
<li>When values go out of scope, check whether they are still-open file
descriptors, and report a bug if so. In addition, remove any information about
them from the program state in order to keep the state as small as possible.
<li>When file pointers "escape" (are used in a way that the analyzer can no longer
track them), mark them as such. This prevents false positives in the cases where
the analyzer cannot be sure whether the file was closed or not.
</ul>
<p>These events that will be used for each of these actions are, respectively, <a
href="https://clang.llvm.org/doxygen/classclang_1_1ento_1_1check_1_1PreCall.html">PreCall</a>,
<a
href="https://clang.llvm.org/doxygen/classclang_1_1ento_1_1check_1_1PostCall.html">PostCall</a>,
<a
href="https://clang.llvm.org/doxygen/classclang_1_1ento_1_1check_1_1DeadSymbols.html">DeadSymbols</a>,
and <a
href="https://clang.llvm.org/doxygen/classclang_1_1ento_1_1check_1_1PointerEscape.html">PointerEscape</a>.
The high-level structure of the checker's class is thus:
<pre class="code_example">
class SimpleStreamChecker : public Checker<check::PreCall,
check::PostCall,
check::DeadSymbols,
check::PointerEscape> {
public:
void checkPreCall(const CallEvent &Call, CheckerContext &C) const;
void checkPostCall(const CallEvent &Call, CheckerContext &C) const;
void checkDeadSymbols(SymbolReaper &SR, CheckerContext &C) const;
ProgramStateRef checkPointerEscape(ProgramStateRef State,
const InvalidatedSymbols &Escaped,
const CallEvent *Call,
PointerEscapeKind Kind) const;
};
</pre>
<h2 id=extendingstates>Custom Program States</h2>
<p> Checkers often need to keep track of information specific to the checks they
perform. However, since checkers have no guarantee about the order in which the
program will be explored, or even that all possible paths will be explored, this
state information cannot be kept within individual checkers. Therefore, if
checkers need to store custom information, they need to add new categories of
data to the <tt>ProgramState</tt>. The preferred way to do so is to use one of
several macros designed for this purpose. They are:
<ul>
<li><a
href="https://clang.llvm.org/doxygen/ProgramStateTrait_8h.html#ae4cddb54383cd702a045d7c61b009147">REGISTER_TRAIT_WITH_PROGRAMSTATE</a>:
Used when the state information is a single value. The methods available for
state types declared with this macro are <tt>get</tt>, <tt>set</tt>, and
<tt>remove</tt>.
<li><a
href="https://clang.llvm.org/doxygen/ProgramStateTrait_8h.html#aa27656fa0ce65b0d9ba12eb3c02e8be9">REGISTER_LIST_WITH_PROGRAMSTATE</a>:
Used when the state information is a list of values. The methods available for
state types declared with this macro are <tt>add</tt>, <tt>get</tt>,
<tt>remove</tt>, and <tt>contains</tt>.
<li><a
href="https://clang.llvm.org/doxygen/ProgramStateTrait_8h.html#ad90f9387b94b344eaaf499afec05f4d1">REGISTER_SET_WITH_PROGRAMSTATE</a>:
Used when the state information is a set of values. The methods available for
state types declared with this macro are <tt>add</tt>, <tt>get</tt>,
<tt>remove</tt>, and <tt>contains</tt>.
<li><a
href="https://clang.llvm.org/doxygen/ProgramStateTrait_8h.html#a6d1893bb8c18543337b6c363c1319fcf">REGISTER_MAP_WITH_PROGRAMSTATE</a>:
Used when the state information is a map from a key to a value. The methods
available for state types declared with this macro are <tt>add</tt>,
<tt>set</tt>, <tt>get</tt>, <tt>remove</tt>, and <tt>contains</tt>.
</ul>
<p>All of these macros take as parameters the name to be used for the custom
category of state information and the data type(s) to be used for storage. The
data type(s) specified will become the parameter type and/or return type of the
methods that manipulate the new category of state information. Each of these
methods are templated with the name of the custom data type.
<p>For example, a common case is the need to track data associated with a
symbolic expression; a map type is the most logical way to implement this. The
key for this map will be a pointer to a symbolic expression
(<tt>SymbolRef</tt>). If the data type to be associated with the symbolic
expression is an integer, then the custom category of state information would be
declared as
<pre class="code_example">
REGISTER_MAP_WITH_PROGRAMSTATE(ExampleDataType, SymbolRef, int)
</pre>
The data would be accessed with the function
<pre class="code_example">
ProgramStateRef state;
SymbolRef Sym;
...
int currentlValue = state->get<ExampleDataType>(Sym);
</pre>
and set with the function
<pre class="code_example">
ProgramStateRef state;
SymbolRef Sym;
int newValue;
...
ProgramStateRef newState = state->set<ExampleDataType>(Sym, newValue);
</pre>
<p>In addition, the macros define a data type used for storing the data of the
new data category; the name of this type is the name of the data category with
"Ty" appended. For <tt>REGISTER_TRAIT_WITH_PROGRAMSTATE</tt>, this will simply
be passed data type; for the other three macros, this will be a specialized
version of the <a
href="https://llvm.org/doxygen/classllvm_1_1ImmutableList.html">llvm::ImmutableList</a>,
<a
href="https://llvm.org/doxygen/classllvm_1_1ImmutableSet.html">llvm::ImmutableSet</a>,
or <a
href="https://llvm.org/doxygen/classllvm_1_1ImmutableMap.html">llvm::ImmutableMap</a>
templated class. For the <tt>ExampleDataType</tt> example above, the type
created would be equivalent to writing the declaration:
<pre class="code_example">
using ExampleDataTypeTy = llvm::ImmutableMap<SymbolRef, int>;
</pre>
<p>These macros will cover a majority of use cases; however, they still have a
few limitations. They cannot be used inside namespaces (since they expand to
contain top-level namespace references), and the data types that they define
cannot be referenced from more than one file.
<p>Note that <tt>ProgramStates</tt> are immutable; instead of modifying an existing
one, functions that modify the state will return a copy of the previous state
with the change applied. This updated state must be then provided to the
analyzer core by calling the <tt>CheckerContext::addTransition</tt> function.
<h2 id=bugs>Bug Reports</h2>
<p> When a checker detects a mistake in the analyzed code, it needs a way to
report it to the analyzer core so that it can be displayed. The two classes used
to construct this report are <tt><a
href="https://clang.llvm.org/doxygen/classclang_1_1ento_1_1BugType.html">BugType</a></tt>
and <tt><a
href="https://clang.llvm.org/doxygen/classclang_1_1ento_1_1BugReport.html">
BugReport</a></tt>.
<p>
<tt>BugType</tt>, as the name would suggest, represents a type of bug. The
constructor for <tt>BugType</tt> takes two parameters: The name of the bug
type, and the name of the category of the bug. These are used (e.g.) in the
summary page generated by the scan-build tool.
<P>
The <tt>BugReport</tt> class represents a specific occurrence of a bug. In
the most common case, three parameters are used to form a <tt>BugReport</tt>:
<ol>
<li>The type of bug, specified as an instance of the <tt>BugType</tt> class.
<li>A short descriptive string. This is placed at the location of the bug in
the detailed line-by-line output generated by scan-build.
<li>The context in which the bug occurred. This includes both the location of
the bug in the program and the program's state when the location is reached. These are
both encapsulated in an <tt>ExplodedNode</tt>.
</ol>
<p>In order to obtain the correct <tt>ExplodedNode</tt>, a decision must be made
as to whether or not analysis can continue along the current path. This decision
is based on whether the detected bug is one that would prevent the program under
analysis from continuing. For example, leaking of a resource should not stop
analysis, as the program can continue to run after the leak. Dereferencing a
null pointer, on the other hand, should stop analysis, as there is no way for
the program to meaningfully continue after such an error.
<p>If analysis can continue, then the most recent <tt>ExplodedNode</tt>
generated by the checker can be passed to the <tt>BugReport</tt> constructor
without additional modification. This <tt>ExplodedNode</tt> will be the one
returned by the most recent call to <a
href="https://clang.llvm.org/doxygen/classclang_1_1ento_1_1CheckerContext.html#a9717efea3fbc71523984160ae7ae9d41">CheckerContext::addTransition</a>.
If no transition has been performed during the current callback, the checker should call <a
href="https://clang.llvm.org/doxygen/classclang_1_1ento_1_1CheckerContext.html#a9717efea3fbc71523984160ae7ae9d41">CheckerContext::addTransition()</a>
and use the returned node for bug reporting.
<p>If analysis can not continue, then the current state should be transitioned
into a so-called <i>sink node</i>, a node from which no further analysis will be
performed. This is done by calling the <a
href="https://clang.llvm.org/doxygen/classclang_1_1ento_1_1CheckerContext.html#a249245cdf2384738921f134c8d7d909a">
CheckerContext::generateSink</a> function; this function is the same as the
<tt>addTransition</tt> function, but marks the state as a sink node. Like
<tt>addTransition</tt>, this returns an <tt>ExplodedNode</tt> with the updated
state, which can then be passed to the <tt>BugReport</tt> constructor.
<p>
After a <tt>BugReport</tt> is created, it should be passed to the analyzer core
by calling <a href="https://clang.llvm.org/doxygen/classclang_1_1ento_1_1CheckerContext.html#af50a9f46f6ea787a2a8e4ad7f86576e7">CheckerContext::emitReport</a>.
<h2 id=ast>AST Visitors</h2>
Some checks might not require path-sensitivity to be effective. Simple AST walk
might be sufficient. If that is the case, consider implementing a Clang
compiler warning. On the other hand, a check might not be acceptable as a compiler
warning; for example, because of a relatively high false positive rate. In this
situation, AST callbacks <tt><b>checkASTDecl</b></tt> and
<tt><b>checkASTCodeBody</b></tt> are your best friends.
<h2 id=testing>Testing</h2>
Every patch should be well tested with Clang regression tests. The checker tests
live in <tt>clang/test/Analysis</tt> folder. To run all of the analyzer tests,
execute the following from the <tt>clang</tt> build directory:
<pre class="code">
$ <b>bin/llvm-lit -sv ../llvm/tools/clang/test/Analysis</b>
</pre>
<h2 id=commands>Useful Commands/Debugging Hints</h2>
<h3 id=attaching>Attaching the Debugger</h3>
<p>When your command contains the <tt><b>-cc1</b></tt> flag, you can attach the
debugger to it directly:</p>
<pre class="code">
$ <b>gdb --args clang -cc1 -analyze -analyzer-checker=core test.c</b>
$ <b>lldb -- clang -cc1 -analyze -analyzer-checker=core test.c</b>
</pre>
<p>
Otherwise, if your command line contains <tt><b>--analyze</b></tt>,
the actual clang instance would be run in a separate process. In
order to debug it, use the <tt><b>-###</b></tt> flag for obtaining
the command line of the child process:
</p>
<pre class="code">
$ <b>clang --analyze test.c -\#\#\#</b>
</pre>
<p>
Below we describe a few useful command line arguments, all of which assume that
you are running <tt><b>clang -cc1</b></tt>.
</p>
<h3 id=narrowing>Narrowing Down the Problem</h3>
<p>While investigating a checker-related issue, instruct the analyzer to only
execute a single checker:
</p>
<pre class="code">
$ <b>clang -cc1 -analyze -analyzer-checker=osx.KeychainAPI test.c</b>
</pre>
<p>If you are experiencing a crash, to see which function is failing while
processing a large file use the <tt><b>-analyzer-display-progress</b></tt>
option.</p>
<p>To selectively analyze only the given function, use the
<tt><b>-analyze-function</b></tt> option:</p>
<pre class="code">
$ <b>clang -cc1 -analyze -analyzer-checker=core test.c -analyzer-display-progress</b>
ANALYZE (Syntax): test.c foo
ANALYZE (Syntax): test.c bar
ANALYZE (Path, Inline_Regular): test.c bar
ANALYZE (Path, Inline_Regular): test.c foo
$ <b>clang -cc1 -analyze -analyzer-checker=core test.c -analyzer-display-progress -analyze-function=foo</b>
ANALYZE (Syntax): test.c foo
ANALYZE (Path, Inline_Regular): test.c foo
</pre>
<b>Note: </b> a fully qualified function name has to be used when selecting
C++ functions and methods, Objective-C methods and blocks, e.g.:
<pre class="code">
$ <b>clang -cc1 -analyze -analyzer-checker=core test.cc -analyze-function='foo(int)'</b>
</pre>
The fully qualified name can be found from the
<tt><b>-analyzer-display-progress</b></tt> output.
<p>The bug reporter mechanism removes path diagnostics inside intermediate
function calls that have returned by the time the bug was found and contain
no interesting pieces. Usually it is up to the checkers to produce more
interesting pieces by adding custom <tt>BugReporterVisitor</tt> objects.
However, you can disable path pruning while debugging with the
<tt><b>-analyzer-config prune-paths=false</b></tt> option.
<h3 id=visualizing>Visualizing the Analysis</h3>
<p>To dump the AST, which often helps understanding how the program should
behave:</p>
<pre class="code">
$ <b>clang -cc1 -ast-dump test.c</b>
</pre>
<p>To view/dump CFG use <tt>debug.ViewCFG</tt> or <tt>debug.DumpCFG</tt>
checkers:</p>
<pre class="code">
$ <b>clang -cc1 -analyze -analyzer-checker=debug.ViewCFG test.c</b>
</pre>
<p><tt>ExplodedGraph</tt> (the state graph explored by the analyzer) can be
visualized with another debug checker:</p>
<pre class="code">
$ <b>clang -cc1 -analyze -analyzer-checker=debug.ViewExplodedGraph test.c</b>
</pre>
<p>Or, equivalently, with <tt><b>-analyzer-viz-egraph-graphviz</b></tt>
option, which does the same thing - dumps the exploded graph in graphviz
<tt><b>.dot</b></tt> format.</p>
<p>You can convert <tt><b>.dot</b></tt> files into other formats - in
particular, converting to <tt><b>.svg</b></tt> and viewing in your web
browser might be more comfortable than using a <tt><b>.dot</b></tt> viewer:</p>
<pre class="code">
$ <b>dot -Tsvg ExprEngine-501e2e.dot -o ExprEngine-501e2e.svg</b>
</pre>
<p>The <tt><b>-trim-egraph</b></tt> option removes all paths except those
leading to bug reports from the exploded graph dump. This is useful
because exploded graphs are often huge and hard to navigate.</p>
<p>Viewing <tt>ExplodedGraph</tt> is your most powerful tool for understanding
the analyzer's false positives, because it gives comprehensive information
on every decision made by the analyzer across all analysis paths.</p>
<p>There are more debug checkers available. To see all available debug checkers:
</p>
<pre class="code">
$ <b>clang -cc1 -analyzer-checker-help | grep "debug"</b>
</pre>
<h3 id=debugprints>Debug Prints and Tricks</h3>
<p>To view "half-baked" <tt>ExplodedGraph</tt> while debugging, jump to a frame
that has <tt>clang::ento::ExprEngine</tt> object and execute:</p>
<pre class="code">
(gdb) <b>p ViewGraph(0)</b>
</pre>
<p>To see the <tt>ProgramState</tt> while debugging use the following command.
<pre class="code">
(gdb) <b>p State->dump()</b>
</pre>
<p>To see <tt>clang::Expr</tt> while debugging use the following command. If you
pass in a <tt>SourceManager</tt> object, it will also dump the corresponding line in the
source code.</p>
<pre class="code">
(gdb) <b>p E->dump()</b>
</pre>
<p>To dump AST of a method that the current <tt>ExplodedNode</tt> belongs
to:</p>
<pre class="code">
(gdb) <b>p C.getPredecessor()->getCodeDecl().getBody()->dump()</b>
</pre>
<h2 id=links>Making Your Checker Better</h2>
<ul>
<li>User facing documentation is important for adoption! Make sure the <a href="/available_checks.html">checker list </a>is updated
at the homepage of the analyzer. Also ensure the description is clear to
non-analyzer-developers in <tt>Checkers.td</tt>.</li>
<li>Warning and note messages should be clear and easy to understand, even if a bit long.</li>
<ul>
<li>Messages should start with a capital letter (unlike Clang warnings!) and should not
end with <tt>.</tt>.</li>
<li>Articles are usually omitted, eg. <tt>Dereference of a null pointer</tt> ->
<tt>Dereference of null pointer</tt>.</li>
<li>Introduce <tt>BugReporterVisitor</tt>s to emit additional notes that explain the warning
to the user better. There are some existing visitors that might be useful for your check,
e.g. <tt>trackNullOrUndefValue</tt>. For example, SimpleStreamChecker should highlight
the event of opening the file when reporting a file descriptor leak.</li>
</ul>
<li>If the check tracks anything in the program state, it needs to implement the
<tt>checkDeadSymbols</tt>callback to clean the state up.</li>
<li>The check should conservatively assume that the program is correct when a tracked symbol
is passed to a function that is unknown to the analyzer.
<tt>checkPointerEscape</tt> callback could help you handle that case.</li>
<li>Use safe and convenient APIs!</li>
<ul>
<li>Always use <tt>CheckerContext::generateErrorNode</tt> and
<tt>CheckerContext::generateNonFatalErrorNode</tt> for emitting bug reports.
Most importantly, never emit report against <tt>CheckerContext::getPredecessor</tt>.</li>
<li>Prefer <tt>checkPreCall</tt> and <tt>checkPostCall</tt> to
<tt>checkPreStmt<CallExpr></tt> and <tt>checkPostStmt<CallExpr></tt>.</li>
<li>Use <tt>CallDescription</tt> to detect hardcoded API calls in the program.</li>
<li>Simplify <tt>C.getState()->getSVal(E, C.getLocationContext())</tt> to <tt>C.getSVal(E)</tt>.</li>
</ul>
<li>Common sources of crashes:</li>
<ul>
<li><tt>CallEvent::getOriginExpr</tt> is nullable - for example, it returns null for an
automatic destructor of a variable. The same applies to some values generated while the
call was modeled, eg. <tt>SymbolConjured::getStmt</tt> is nullable.</li>
<li><tt>CallEvent::getDecl</tt> is nullable - for example, it returns null for a
call of symbolic function pointer.</li>
<li><tt>addTransition</tt>, <tt>generateSink</tt>, <tt>generateNonFatalErrorNode</tt>,
<tt>generateErrorNode</tt> are nullable because you can transition to a node that you have already visited.</li>
<li>Methods of <tt>CallExpr</tt>/<tt>FunctionDecl</tt>/<tt>CallEvent</tt> that
return arguments crash when the argument is out-of-bounds. If you checked the function name,
it doesn't mean that the function has the expected number of arguments!
Which is why you should use <tt>CallDescription</tt>.</li>
<li>Nullability of different entities within different kinds of symbols and regions is usually
documented via assertions in their constructors.</li>
<li><tt>NamedDecl::getName</tt> will fail if the name of the declaration is not a single token,
e.g. for destructors. You could use <tt>NamedDecl::getNameAsString</tt> for those cases.
Note that this method is much slower and should be used sparringly, e.g. only when generating reports
but not during analysis.</li>
<li>Is <tt>-analyzer-checker=core</tt> included in all test <tt>RUN:</tt> lines? It was never supported
to run the analyzer with the core checks disabled. It might cause unexpected behavior and
crashes. You should do all your testing with the core checks enabled.</li>
</ul>
</ul>
<li>Patterns that you should most likely avoid even if they're not technically wrong:</li>
<ul>
<li><tt>BugReporterVisitor</tt> should most likely not match the AST of the current program point
to decide when to emit a note. It is much easier to determine that by observing changes in
the program state.</li>
<li>In <tt>State->getSVal(Region)</tt>, if <tt>Region</tt> is not known to be a <tt>TypedValueRegion</tt>
and the optional type argument is not specified, the checker may accidentally try to dereference a
void pointer.</li>
<li>Checker logic should not depend on whether a certain value is a <tt>Loc</tt> or <tt>NonLoc</tt>.
It should be immediately obvious whether the <tt>SVal</tt> is a <tt>Loc</tt> or a
<tt>NonLoc</tt> depending on the AST that is being checked. Checking whether a value
is <tt>Loc</tt> or <tt>Unknown</tt>/<tt>Undefined</tt> or whether the value is
<tt>NonLoc</tt> or <tt>Unknown</tt>/<tt>Undefined</tt> is totally fine.</li>
<li>New symbols should not be constructed in the checker via direct calls to <tt>SymbolManager</tt>,
unless they are of <tt>SymbolMetadata</tt> class tagged by the checker,
or they represent newly created values such as the return value in <tt>evalCall</tt>.
For modeling arithmetic/bitwise/comparison operations, <tt>SValBuilder</tt> should be used.</li>
<li>Custom <tt>ProgramPointTag</tt>s should not be created within the checker. There is usually
no good reason for a checker to chain multiple nodes together, because checkers aren't worklists.</li>
</ul>
<li>Checkers are encouraged to actively participate in the analysis by sharing
their knowledge about the program state with the rest of the analyzer,
but they should not be disrupting the analysis unnecessarily:</li>
<ul>
<li>If a checker splits program state, this must be based on knowledge that
the newly appearing branches are definitely possible and worth exploring
from the user's perspective. Otherwise the state split should be delayed
until there's an indication that one of the paths is taken, or one of the
paths needs to be dropped entirely. For example, it is fine to eagerly split
paths while modeling <tt>isalpha(x)</tt> as long as <tt>x</tt> is constrained accordingly on
each path. At the same time, it is not a good idea to split paths over the
return value of <tt>printf()</tt> while modeling the call because nobody ever checks
for errors in <tt>printf</tt>; at best, it'd just double the remaining analysis time.
</li>
<li>Caution is advised when using <tt>CheckerContext::generateNonFatalErrorNode</tt>
because it generates an independent transition, much like <tt>addTransition</tt>.
It is easy to accidentally split paths while using it. Ideally, try to
structure the code so that it was obvious that every <tt>addTransition</tt> or
<tt>generateNonFatalErrorNode</tt> (or sequence of such if the split is intended) is
immediately followed by return from the checker callback.</li>
<li>Multiple implementations of <tt>evalCall</tt> in different checkers should not conflict.</li>
<li>When implementing <tt>evalAssume</tt>, the checker should always return a non-null state
for either the true assumption or the false assumption (or both).</li>
<li>Checkers shall not mutate values of expressions, i.e. use the <tt>ProgramState::BindExpr</tt> API,
unless they are fully responsible for computing the value.
Under no circumstances should they change non-<tt>Unknown</tt> values of expressions.
Currently the only valid use case for this API in checkers is to model the return value in the <tt>evalCall</tt> callback.
If expression values are incorrect, <tt>ExprEngine</tt> needs to be fixed instead.</li>
</ul>
<h2 id=additioninformation>Additional Sources of Information</h2>
Here are some additional resources that are useful when working on the Clang
Static Analyzer:
<ul>
<li><a href="https://lcs.ios.ac.cn/~xzx/memmodel.pdf">Xu, Zhongxing &
Kremenek, Ted & Zhang, Jian. (2010). A Memory Model for Static Analysis of C
Programs.</a></li>
<li><a href="https://github.com/llvm/llvm-project/blob/main/clang/lib/StaticAnalyzer/README.txt">
The Clang Static Analyzer README</a></li>
<li><a href="https://github.com/llvm/llvm-project/blob/main/clang/docs/analyzer/developer-docs/RegionStore.rst">
Documentation for how the Store works</a></li>
<li><a href="https://github.com/llvm/llvm-project/blob/main/clang/docs/analyzer/developer-docs/IPA.rst">
Documentation about inlining</a></li>
<li> The "Building a Checker in 24 hours" presentation given at the <a
href="https://llvm.org/devmtg/2012-11">November 2012 LLVM Developer's
meeting</a>. Describes the construction of SimpleStreamChecker. <a
href="https://llvm.org/devmtg/2012-11/Zaks-Rose-Checker24Hours.pdf">Slides</a>
and <a
href="https://youtu.be/kdxlsP5QVPw">video</a>
are available.</li>
<li>
<a href="https://github.com/haoNoQ/clang-analyzer-guide/releases/download/v0.1/clang-analyzer-guide-v0.1.pdf">
Artem Degrachev: Clang Static Analyzer: A Checker Developer's Guide
</a> (reading the previous items first might be a good idea)</li>
<li>The list of <a href="implicit_checks.html">Implicit Checkers</a></li>
<li> <a href="https://clang.llvm.org/doxygen">Clang doxygen</a>. Contains
up-to-date documentation about the APIs available in Clang. Relevant entries
have been linked throughout this page. Also of use is the
<a href="https://llvm.org/doxygen">LLVM doxygen</a>, when dealing with classes
from LLVM.</li>
<li>
The <a href="https://discourse.llvm.org/c/clang/"> Clang Frontend Discourse site</a>.
This is the primary forum discussing ideas and posting questions about Clang development.
For posting Clang Static Analyzer specific questions, please visit the
<a href="https://discourse.llvm.org/c/clang/static-analyzer/"> Static Analyzer subcategory</a>
of the same site. In the past, Static Analyzer discussions took place at the
<a href="https://lists.llvm.org/pipermail/cfe-dev/"> cfe-dev</a> mailing list, which is now
archived and superseeded by the mentioned Discourse site.
</li>
</ul>
</div>
</div>
</body>
</html>
Codebase Browser
llvm