Dynamic Analysis and Profiling of Multithreaded Systems - Advanced Operating Systems and Kernel Applications - page 168

Table 4. Slow-down incurred by GNU GCC profiling

Test No profiling With profiling % Slow Down

Numeric Sort 812.32 498.2 38.67

String Sort 103.24 76.499 25.90

Bitfield 4.35E+08 1.65E+08 62.11

FP Emulation 73.76 52.96 28.20

Fourier 15366 15245 0.79

Assignment 24.292 9.77 59.78

Huffman 1412.7 1088.7 22.93

Neural Net 18.091 12.734 29.61

LU Decomp 909.76 421.48 53.67

enable profiling on selected compilation units,
thereby minimizing instrumentation costs.
This type of code instrumentation is termed
static because the code is modified before execu-
tion of the program (Figure 2, label C). COTS
compiler-based instrumentation for profiling is
generally limited to function calls and iteration
counts. Another more powerful form of static
binary instrumentation involves the use of a set
of libraries and APIs that enable users to quickly
write applications that perform binary rewriting
(Hunt & Brubacher, 1999; Larus & Schnarr,
1995; Srivastava & Eustace, 1994; Romer, et al.,
1997; Hollingsworth, Miller , & Cargille, 1994).
The following capabilities are typical of binary
rewriting libraries:

EEL code “snippets” encapsulate of pieces
of code that can be inserted into existing binary
code. They are either written directly in assembly
language (which makes the instrumentation ma-
chine dependent) or written using a higher-level
language that is compiled into assembly. To graft
snippet code into existing code, each snippet
identifies registers used in the snippet that must
be assigned to unused registers.

dynamic binary-code instrumentation

An alternative to static binary instrumentation
is dynamic instrumentation. Dynamic instru-
mentation, implemented as Just-in Time (JIT)
compilation, is performed after a program has
been loaded into memory and immediately
prior to execution (Figure 2, label D). Dynamic
instrumentation has the advantage that profiling
functionality can be selectively added or removed
from the program without the need to recompile:
Trace functionality is only present when needed.
Moreover, dynamic instrumentation can be ap-
plied reactively, for example, in response to some
event in the system, such as processor slow down.
Dynamic instrumentation is particularly useful for
facilitating conditional breakpoints in code, for
example, Buck and Hollingsworth (2000) show
that this approach is 600 times more ef ficient than
conventional trap-based debug breakpoints.

• Redirection of function calls and insertion

of trampoline functions that execute the
originally called function;

• Insertion of additional code and data; and

• Control and data-flow analysis to guide

instrumentation.

The following code illustrates the use of con-
trol-flow analysis and insertion of additional code
through the Editing Executable Library (EEL)
(Larus & Schnarr, 1995), machine-independent,
executable editing API: (Box 5)