 |
Code and Data Location
|
Instruction and data caching can show dramatic improvement
in the execution time of many applications in which the path of
execution is single threaded. Much research has been done to show
that for many applications, both instruction and data accesses are
frequently very localized and hence benefit immensely from caching.
However, in a multitasking environment, the thread of execution
is not so linear. Task switches require stack switching and
context saving and restoral. These switches, often in response
to interrupt activity, interfere with the seqential instruction
execution process thereby negating some of the benefits of
instruction caching.
The position of code and data in memory can also affect the
caching performance as the following example illustrates.
Assume that two code modules, A and B, are located in memory
at addresses 0x1000 and 0x2000 respectively as
shown in the diagram below. Assume that each module occupies
less than 1024 bytes, the size of the instruction cache.
Also assume that addresses map directly to the cache using
the least significant 10 bits of the address.
|
Module A |
|
Module B |
|
void Ahigh()
{
}
void Alow()
{
}
0x1000
|
|
|
|
void Bhigh()
{
}
void Blow()
{
}
0x2000
|
|
If procedure Bhigh in module B repeatedly calls procedure
Alow in module A, both procedures will end up in cache and
benefit accordingly.
However, if procedure Bhigh in module B repeatedly calls
procedure Ahigh in module A, the procedures will compete
with each other for residence in the cache and performance will suffer.
Few applications are small enough to permit the fine tuning
necessary to achieve optimal code location for maximum cache benefit.
It is more likely that your code, and the procedures which it calls,
will simply end up in memory at locations which vary with every
minor code change. Move any block of code up or down in memory
with respect to the other blocks of code with which it interacts
and you can expect execution timing differences with caching enabled.
| 
