 |
» |
|
|
|
|  | |
Profile-based optimization (PBO) is a set of code-improving
transformations that are based on feedback concerning the run-time
characteristics of an application. Run-time profile data is collected
during program execution. This information is fed back to the optimizer,
which performs a variety of optimizations based upon how frequently
certain code is executed and how frequently calls are made between
different code segments. In general, each higher level of optimization
takes increasing advantage of the PBO-generated information. One of the goals of PBO is to improve the efficiency of memory
access by increasing the hit rates for the instruction cache, memory
pages, and the Translation Lookaside Buffer (TLB). PBO can also
improve the compiler's inlining decisions. One basis for the decision
is call frequency, which, without PBO, the compiler can only estimate.
With PBO, however, the compiler uses actual frequencies as the basis
for its decision. For complete details on PBO, see HP-UX Linker and
Libraries Online User Guide. | | | |  | NOTE: HP Pascal/HP-UX is limited to procedure repositioning,
not basic block repositioning. | | | | |
Invoking PBO | |
You must perform the following steps to invoke PBO: Compile and link with the +I
option to produce the instrumented program. Collect execution profile statistics by running
the instrumented program. Optimize the program with the +P
option.
Instrumenting the Program To instrument the program, use the +I
compile-line option when compiling and linking the object files
of an application. Instrumenting the program inserts code into the
program to collect execution profile statistics. Execution profile
statistics include a count of the calls between procedures. Also,
note that when you use the +I
compile-line option to compile source files, instrumentation can
be added within the code for each subroutine in that file. In the following example, the source file sample.p
is compiled, instrumented, and linked into sample.inst: pc -o sample.inst -O +I sample.p |
Collecting Execution Profile Statistics To collect execution profile statistics, run your program
using reasonably representative data. The profile database file,
flow.data, is
created the first time the program is run, and is updated for each
subsequent execution of the program. The following example collects execution profile statistics
by running the sample.inst
program with representative data from two input files: sample.inst < input.file1
sample.inst < input.file2 |
This step, by default, logs the profile statistics in a file
called flow.data.
See “Maintaining Multiple Profile Data
Files ” to change
this default. To perform PBO, re-link the program with the +P
compile-line option to specify that you wish to use the collected
profile data: pc -o sample.opt -O +P +pgm sample.inst sample.o |
The +pgm
compile-line option allows you to specify an executable name that
is different from the current output file name. In the preceding
command line, the +pgm
option indicates that the name of the instrumented executable (sample.inst)
differs from the name of the optimized executable (sample.opt)
that is specified with the -o
option. Source files compiled with the +I
option do not need to be recompiled after collecting the profile
data. Simply relink the application with the same options that you
used in the first step to instrument the program, but replace +I
with +P. For
more information about how the compiler and linker work together
to perform profile based-optimizations, refer to HP-UX
Linker and Libraries Online User Guide. Maintaining Multiple Profile Data
Files | |
By default, PBO logs the profile statistics in a file called
flow.data. You
can specify another name with the +df
compile-line option. The name of the executable file used during profiling is the
name under which the profile data is stored in the database file.
If you specify a different executable output file name during the
optimization phase, you need to use the +pgm
option to specify the program name used during profiling. The following
steps and example demonstrate the use of +df
and +pgm. Rename flow.data
after performing the data collection step described previously: Perform PBO on this application by re-linking the
program as follows: pc -o sample.opt -O +P +pgm sample.inst +df sample.data sample.o |
The default value for the +df
compile-line option is flow.data.
The +df option
is used because the profile data file for the program has been moved
from flow.data
to sample.data. The +pgm
option is used because the instrumented program file is sample.inst,
and the optimized program file is sample.opt. You can also use the FLOW_DATA
environment variable to specify a different path name for the profile
database file. Note, however, that the +df
compile-line option takes precedence over the FLOW_DATA
environment variable. For more information about the profile database
and the FLOW_DATA
environment variable, see HP-UX Linker and Libraries
Online User Guide.
|
Printable version
