Hello,
I'm trying to learn high-performance real-time programming.

One of my wonderings are:
Should i use int/uint for all standard arithmetic operations or int/short/byte (depends on actual case)?
I believe this question has following subquestions:
* Arithmetic computations performance
* Memory access time

My actual compiler is DMD, but I'm interested in GDC as well.

Lastly, one more question would be:
Could someone recommend any books/resources for this kind of informations and tips that could be applied to D? I'd like to defer my own experiments with generated assembly and profiling, but i suppose people already published general rules that i could apply for my programming.

Thanks,
Mariusz Gliwiński

On 4/30/2011 10:34 AM, Mariusz Gliwiński wrote:
> Hello,
> I'm trying to learn high-performance real-time programming.
>
> One of my wonderings are:
> Should i use int/uint for all standard arithmetic operations or
> int/short/byte (depends on actual case)?
> I believe this question has following subquestions:
> * Arithmetic computations performance
> * Memory access time
>
> My actual compiler is DMD, but I'm interested in GDC as well.
>
> Lastly, one more question would be:
> Could someone recommend any books/resources for this kind of
> informations and tips that could be applied to D? I'd like to defer my
> own experiments with generated assembly and profiling, but i suppose
> people already published general rules that i could apply for my
> programming.
>
> Thanks,
> Mariusz Gliwiński

My experience with this pattern of thinking is to use the largest data type that makes sense, unless you have a profiler saying you need to do something different.  However, if you get being obsessive-compulsive about having 'the perfectly sized integer types' for the code, it is possible to fall into the trap of over-using unsigned types 'because the value can never be negative'.  Unsigned 8 and 16 bit values usually have a good reason to be unsigned, but when you start getting to 32 and 64 bit values it makes a lot less sense most of the time.

When working with non-X86 platforms other problems are usually much more severe: More expensive thread synchronization primitives, lack of efficient variable bit bit-shifting (run-time determined number of bits shifted), non-existent branch prediction, or various floating point code promoting to emulated double precision code silently on hardware that can only do single precision floating point etc.

On 30.04.2011 19:34, Mariusz Gliwiński wrote:
> Hello,
> I'm trying to learn high-performance real-time programming.
>
> One of my wonderings are:
> Should i use int/uint for all standard arithmetic operations or int/short/byte (depends on actual case)?
> I believe this question has following subquestions:
> * Arithmetic computations performance
> * Memory access time
>
> My actual compiler is DMD, but I'm interested in GDC as well.
>
> Lastly, one more question would be:
> Could someone recommend any books/resources for this kind of informations and tips that could be applied to D? I'd like to defer my own experiments with generated assembly and profiling, but i suppose people already published general rules that i could apply for my programming.
>
> Thanks,
> Mariusz Gliwiński
I find Agner Fog's guides on optimization for x86 the best source on such architecture specific matters.
http://www.agner.org/optimize/

Citing releveant part from C++ optimization guide (on Integers):
>Integers of smaller sizes (char, short int) are only slightly less
>efficient. In most cases, the compiler will convert these types to integers of the default size
>when doing calculations, and then use only the lower 8 or 16 bits of the result. You can
>assume that the type conversion takes zero or one clock cycle. In 64-bit systems, there is
>only a minimal difference between the efficiency of 32-bit integers and 64-bit integers, as
>long as you are not doing divisions.


-- 
Dmitry Olshansky

This is a good point.  Further to that, keep in mind locality of reference, ie the performance impact of data getting pushed out of the caches.  While using machine word size variables for a small number of variables that really need high performance can give a small speed-up, using them extensively can increase the programs data size, increasing the frequency of cache misses and resulting in large slow-downs.


On 01/05/11 19:28, Dmitry Olshansky wrote:
> On 30.04.2011 19:34, Mariusz Gliwiński wrote:
>> Hello,
>> I'm trying to learn high-performance real-time programming.
>>
>> One of my wonderings are:
>> Should i use int/uint for all standard arithmetic operations or
>> int/short/byte (depends on actual case)?
>> I believe this question has following subquestions:
>> * Arithmetic computations performance
>> * Memory access time
>>
>> My actual compiler is DMD, but I'm interested in GDC as well.
>>
>> Lastly, one more question would be:
>> Could someone recommend any books/resources for this kind of
>> informations and tips that could be applied to D? I'd like to defer my
>> own experiments with generated assembly and profiling, but i suppose
>> people already published general rules that i could apply for my
>> programming.
>>
>> Thanks,
>> Mariusz Gliwiński
> I find Agner Fog's guides on optimization for x86 the best source on
> such architecture specific matters.
> http://www.agner.org/optimize/
>
> Citing releveant part from C++ optimization guide (on Integers):
>  >Integers of smaller sizes (char, short int) are only slightly less
>  >efficient. In most cases, the compiler will convert these types to
> integers of the default size
>  >when doing calculations, and then use only the lower 8 or 16 bits of
> the result. You can
>  >assume that the type conversion takes zero or one clock cycle. In
> 64-bit systems, there is
>  >only a minimal difference between the efficiency of 32-bit integers
> and 64-bit integers, as
>  >long as you are not doing divisions.
>
>