Undefined behavior


In computer programming, undefined behavior is the result of executing a program whose behavior is prescribed to be unpredictable, in the language specification to which the computer code adheres. This is different from unspecified behavior, for which the language specification does not prescribe a result and implementation defined behavior that defers to the documentation of another component of the platform.
In the C community, undefined behavior may be humorously referred to as "nasal demons", after a comp.std.c post that explained undefined behavior as allowing the compiler to do anything it chooses, even "to make demons fly out of your nose".

Overview

Some programming languages allow a program to operate differently or even have a different control flow than the source code, as long as it exhibits the same user-visible side effects, if undefined behavior never happens during program execution. Undefined behavior is the name of a list of conditions that the program must not meet.
In the early versions of C, undefined behavior's primary advantage was the production of performant compilers for a wide variety of machines: a specific construct could be mapped to a machine-specific feature, and the compiler did not have to generate additional code for the runtime to adapt the side effects to match semantics imposed by the language. The program source code was written with prior knowledge of the specific compiler and of the platforms that it would support.
However, progressive standardization of the platforms has made this less of an advantage, especially in newer versions of C. Now, the cases for undefined behavior typically represent unambiguous bugs in the code, for example indexing an array outside of its bounds. By definition, the runtime can assume that undefined behavior never happens; therefore, some invalid conditions do not need to be checked against. For a compiler, this also means that various program transformations become valid, or their proofs of correctness are simplified; this allows for various kinds of premature optimization and micro-optimization, which lead to incorrect behavior if the program state meets any of such conditions. The compiler can also remove explicit checks that may have been in the source code, without notifying the programmer; for example, detecting undefined behavior by testing whether it happened is not guaranteed to work, by definition. This makes it hard or impossible to program a portable fail-safe option.
Current compiler development usually evaluates and compares compiler performance with benchmarks designed around micro-optimizations, even on platforms that are mostly used on the general-purpose desktop and laptop market. Therefore, undefined behavior provides ample room for compiler performance improvement, as the source code for a specific source code statement is allowed to be mapped to anything at runtime.
For C and C++, the compiler is allowed to give a compile-time diagnostic in these cases, but is not required to: the implementation will be considered correct whatever it does in such cases, analogous to don't-care terms in digital logic. It is the responsibility of the programmer to write code that never invokes undefined behavior, although compiler implementations are allowed to issue diagnostics when this happens. Compilers nowadays have flags that enable such diagnostics, for example, -fsanitize enables the "undefined behavior sanitizer" in gcc 4.9 and in clang. However, this flag is not the default and enabling it is a choice of who builds the code.
Under some circumstances there can be specific restrictions on undefined behavior. For example, the instruction set specifications of a CPU might leave the behavior of some forms of an instruction undefined, but if the CPU supports memory protection then the specification will probably include a blanket rule stating that no user-accessible instruction may cause a hole in the operating system's security; so an actual CPU would be permitted to corrupt user registers in response to such an instruction, but would not be allowed to, for example, switch into supervisor mode.
The runtime platform can also provide some restrictions or guarantees on undefined behavior, if the toolchain or the runtime explicitly document that specific constructs found in the source code are mapped to specific well-defined mechanisms available at runtime. For example, an interpreter may document a certain behavior for some operations that are undefined in the language specification, while other interpreters or compilers for the same language may not. A compiler produces executable code for a specific ABI, filling the semantic gap in ways that depend on the compiler version: the documentation for that compiler version and the ABI specification can provide restrictions on undefined behavior. Relying on these implementation details makes the software non-portable, however portability may not be a concern if the software is not supposed to be used outside of a specific runtime.
Undefined behavior can result in a program crash or even in failures that are harder to detect and make the program look like it is working normally, such as silent loss of data and production of incorrect results.

Benefits

Documenting an operation as undefined behavior allows compilers to assume that this operation will never happen in a conforming program. This gives the compiler more information about the code and this information can lead to more optimization opportunities.
An example for the C language:

int foo

The value of x cannot be negative and, given that signed integer overflow is undefined behavior in C, the compiler can assume that value < 2147483600 will always be false. Thus the if statement, including the call to the function bar, can be ignored by the compiler since the test expression in the if has no side effects and its condition will never be satisfied. The code is therefore semantically equivalent to:

int foo

Had the compiler been forced to assume that signed integer overflow has wraparound behavior, then the transformation above would not have been legal.
Such optimizations become hard to spot by humans when the code is more complex and other optimizations, like inlining, take place. For example, another function may call the above function:

void run_tasks

The compiler is free to optimize away the while-loop here by applying value range analysis: by inspecting foo, it knows that the initial value pointed to by ptrx cannot possibly exceed 47, therefore the initial check of *ptrx > 60 will always be false in a conforming program. Going further, since the result z is now never used and foo has no side effects, the compiler can optimize run_tasks to be an empty function that returns immediately. The disappearance of the while-loop may be especially surprising if foo is defined in a separately compiled object file.
Another benefit from allowing signed integer overflow to be undefined is that it makes it possible to store and manipulate a variable's value in a processor register that is larger than the size of the variable in the source code. For example, if the type of a variable as specified in the source code is narrower than the native register width, then the compiler can safely use a signed 64-bit integer for the variable in the machine code it produces, without changing the defined behavior of the code. If a program depended on the behavior of a 32-bit integer overflow, then a compiler would have to insert additional logic when compiling for a 64-bit machine, because the overflow behavior of most machine instructions depends on the register width.
Undefined behavior also allows more compile-time checks by both compilers and static program analysis.

Risks

C and C++ standards have several forms of undefined behavior throughout, which offer increased liberty in compiler implementations and compile-time checks at the expense of undefined run-time behavior if present. In particular, the ISO standard for C has an appendix listing common sources of undefined behavior. Moreover, compilers are not required to diagnose code that relies on undefined behavior. Hence, it is common for programmers, even experienced ones, to rely on undefined behavior either by mistake, or simply because they are not well-versed in the rules of the language that can span hundreds of pages. This can result in bugs that are exposed when a different compiler, or different settings, are used. Testing or fuzzing with dynamic undefined behavior checks enabled, e.g., the Clang sanitizers, can help to catch undefined behavior not diagnosed by the compiler or static analyzers.
Undefined behavior can lead to security vulnerabilities in software. For example, buffer overflows and other security vulnerabilities in the major web browsers are due to undefined behavior. The Year 2038 problem is another example due to signed integer overflow. When GCC's developers changed their compiler in 2008 such that it omitted certain overflow checks that relied on undefined behavior, CERT issued a warning against the newer versions of the compiler. Linux Weekly News pointed out that the same behavior was observed in PathScale C, Microsoft Visual C++ 2005 and several other compilers; the warning was later amended to warn about various compilers.

Examples in C and C++

The major forms of undefined behavior in C can be broadly classified as: spatial memory safety violations, temporal memory safety violations, integer overflow, strict aliasing violations, alignment violations, unsequenced modifications, data races, and loops that neither perform I/O nor terminate.
In C the use of any automatic variable before it has been initialized yields undefined behavior, as does integer division by zero, signed integer overflow, indexing an array outside of its defined bounds, or null pointer dereferencing. In general, any instance of undefined behavior leaves the abstract execution machine in an unknown state, and causes the behavior of the entire program to be undefined.
Attempting to modify a string literal causes undefined behavior:ISO/IEC. ISO/IEC 14882:2003: Programming Languages - C++ §2.13.4 String literals para. 2

char *p = "wikipedia"; // valid C, deprecated in C++98/C++03, ill-formed as of C++11
p = 'W'; // undefined behavior

Integer division by zero results in undefined behavior:ISO/IEC. ISO/IEC 14882:2003: Programming Languages - C++ §5.6 Multiplicative operators para. 4

int x = 1;
return x / 0; // undefined behavior

Certain pointer operations may result in undefined behavior:ISO/IEC. ISO/IEC 14882:2003: Programming Languages - C++ §5.7 Additive operators para. 5

int arr = ;
int *p = arr + 5; // undefined behavior for indexing out of bounds
p = 0;
int a = *p; // undefined behavior for dereferencing a null pointer

In C and C++, the relational comparison of pointers to objects is only strictly defined if the pointers point to members of the same object, or elements of the same array.ISO/IEC. ISO/IEC 14882:2003: Programming Languages - C++ §5.9 Relational operators para. 2 Example:

int main

Reaching the end of a value-returning function without a return statement results in undefined behavior if the value of the function call is used by the caller:ISO/IEC. ISO/IEC 9899:2007: Programming Languages - C §6.9 External definitions para. 1

int f
/* undefined behavior if the value of the function call is used*/

Modifying an object between two sequence points more than once produces undefined behavior. There are considerable changes in what causes undefined behavior in relation to sequence points as of C++11. The following example will however cause undefined behavior in both C++ and C.

i = i++ + 1; // undefined behavior

When modifying an object between two sequence points, reading the value of the object for any other purpose than determining the value to be stored is also undefined behavior.ISO/IEC. ISO/IEC 9899:1999: Programming Languages - C §6.5 Expressions para. 2

a = i++; // undefined behavior
printf; // also undefined behavior

In C/C++ bitwise shifting a value by a number of bits which is either a negative number or is greater than or equal to the total number of bits in this value results in undefined behavior. The safest way is to always keep the number of bits to shift within the range: <0, sizeof*CHAR_BIT - 1>.

int num = -1;
unsigned int val = 1 << num; //shifting by a negative number - undefined behavior
num = 32; //or whatever number greater than 31
val = 1 << num; //the literal '1' is typed as a 32-bit integer - in this case shifting by more than 31 bits is undefined behavior
num = 64; //or whatever number greater than 63
unsigned long long val2 = 1ULL << num; //the literal '1ULL' is typed as a 64-bit integer - in this case shifting by more than 63 bits is undefined behavior