The
term alignment refers to the type of address
that a data object has in memory. Data objects can have 1-byte,
2-byte, 4-byte, or 8-byte alignment, meaning that the object is
stored at an address evenly divisible by 1, 2, 4, or 8. In general,
the best performance is obtained by natural alignment,
which is the alignment that corresponds to the length of the object.
For example, the natural alignment of an HP C int
is 4-byte alignment.
For simple static variables, the compilers always naturally
align data. The alignment problem occurs in C structures, Pascal
records, and Fortran common blocks. By default, HP-UX compilers
align this data naturally. However, you can specify directives in
your program or on the command line that align data in a way that
is compatible with data alignment on other HP systems. It is also
possible to use Fortran EQUIVALENCE
statements or other programming methods to obtain nondefault data
alignment. Specifying nondefault data alignment causes the compiler
to generate extra instructions, which both substantially increase
code size and substantially degrade performance. Moreover, aligning
data on a boundary less than its natural alignment boundary (for
example, aligning a double
on a 2-byte boundary) may result in a bus error or some other kind
of run-time error.
In some situations, you can improve performance by aligning
data on greater than natural addresses. This improvement is due
to two factors, both of which concern the vector routines enabled
by the +Ovectorize
option:
A single-precision array of numbers
will sometimes allow better performance if it is 8-byte aligned,
because the vector routines can use double-precision load and store
operations to move two operands at a time.
A double-precision array of numbers will sometimes
allow better performance if it is 32-byte, or cache-line, aligned.
This is because of the way the vector routines interact with the
data caches on HP 9000 systems.
Printable version
