Thomas Matthews wrote:[color=blue]
> Victor Bazarov wrote:
>[color=green]
>> "Alex" <alex.fake@gmail.com> wrote...
>>[color=darkred]
>>> I apoligise in advance if this is an os or platform based question, I
>>> don't know.[/color]
>>
>>
>>
>> It's not platform-based. It is, however, implementation-defined.
>>
>>[color=darkred]
>>> I was wondering how register integers (and other types of register
>>> variables) are managed by c++. For example, on a pentium 4, there are
>>> 8 register integers in the cpu. If you define more than 8, or if there
>>> are other programs using this space, how are the variables allocated.
>>> This is for a simulation program that needs to be very fast.
>>> For example:
>>>
>>> int foo()
>>> {
>>> register int a;
>>> register int b;
>>> [...][/color]
>>
>>
>>
>> 'register' just like 'inline' is not a directive, but rather a suggestion
>> for your compiler. It may take a hint and try placing the object in a
>> CPU
>> register, or it may decide against it and do whatever it sees fit. Just
>> like you noticed, it's not always possible, besides, it's not always the
>> desired course of action.
>>
>> One thing you should probably learn at this point: compilers are
>> better at
>> deciding what needs to go where than programmers. So, spend your time on
>> _algorithms_ and getting things _right_, before ever attempting to get
>> them
>> "very fast". Let the compiler decide what registers to use and when. Do
>> _not_ concern yourself with performance of your code on any particular
>> CPU
>> type or model; instead concern yourself with portability and
>> maintainability
>> of your code. IOW, forget the "register" keyword exists or what it's
>> for,
>> except that you're not allowed to use it anywhere else. Trust me,
>> your life
>> will be simpler that way.
>>
>> Victor
>>[/color]
>
> However, one can always assist a compiler by writing code that
> makes the decisions easier for the compiler.
>
> For example, if a processor has a string search instruction, a
> person may want to set up the code so that the compiler recognizes
> the "pattern" and implements the string search function.
>
> I've done this on the ARM processor using registers and its
> special transfer instructions.[/color]
Oh, there is no doubt about it. The point[s] of my post was/were that
(a) 'register' is not the way to do that, (b) before attempting doing that
one should understand the inner workings of the compiler/CPU combination
very well and (c) portability and maintainability are often more important
than minute improvements in performance: productivity is not measured by
the performance of a single program fragment but by the overall output of
the [sometimes quite large] group of people that involves using and
advancing that program (and I am not sure if I read it somewhere or that
I just made it up :-).
Victor
