Dynamic Analysis and Profiling of Multithreaded Systems - Advanced Operating Systems and Kernel Applications - page 171

the same ISA as the underlying host. As shown
in Figure 5, the Pin system consists of:

idiosyncrasies of the underlying instruction set
and allows passing of context information, such
as register contents, to the injected code as param-
eters. Pin programs typically consist of analysis
and instrumentation elements. The basic building
blocks for defining instrumentation points are
machine instructions, basic blocks, procedures,
images, and applications. For example, the C++
code below shows the use of the Pin API to in-
strument the target code with trace functions each
time sleep() is invoked.
First, a replacement function is defined with
the same signature as the function that is being
replaced (in this example sleep() ). (Box 7)
A callback function ImageLoad() is used to
intercept binary-image loads that are executed by
the target application. The Pin API can then be used
to obtain the function that will be replaced with
the new tracing/trampoline function. (Box 8)
The instrumentation function is “hooked” onto
image loads through IMG _ AddInstrument-
Function() as follows:

• Instrumentation API, which is used to write

programs that perform dynamic interception
and replacement;

• JIT Compiler, which compiles and instru-

ments the bytecode immediately before
execution;

• Code cache, which caches translations be-

tween the target binary and the JIT compiled
code;

• Emulation unit, which interprets instructions

that cannot be executed directly; and

• Dispatcher, which retrieves and executes

code from the code cache.

To analyze a program's behavior (multithread-
ed or not), Pin users must write a program that
performs dynamic instrumentation. Pin provides
an API based on the ATOM API (Srivastava &
Eustace, 1994) that shields programmers from

Figure 5. The Pin software architecture