• Advertisement
Sign in to follow this  

asm/c++ div and modulo

This topic is 2392 days old which is more than the 365 day threshold we allow for new replies. Please post a new topic.

If you intended to correct an error in the post then please contact us.

Recommended Posts

Hi,

How to make a asm code in C++ program that would count div and mod in one instruction.
I don't know asm at all but i think it could make performance better.

Now I have
[code]
int AFunction(int Value)
{
int a=Value/8,
b=Value%8;
//further code...

}
[/code]


Thanks in advance for any reply :)

Share this post


Link to post
Share on other sites
Advertisement
Have you measured the performance of a Release mode c++ build first? Your bottleneck is likely elsewhere.

If you have identified this as a bottleneck, the first step would be to determine if you can reduce the number of times you need to compute this. Essentially you are looking for a high level, algorithmic optimisation.

Once you've exhausted algorithmic optimisations on an actual bottleneck you might finally consider stepping into assembler. The first thing to do is see what your compiler is generating. I compiled the following, similar function using GCC with optimisation level 3.
[code]
int AFunction(int Value)
{
int a = Value / 8;
int b = Value % 8;
return a + b;
}
[/code]
It transformed that code into left and right shifts, it avoided doing any division or mod instructions.

Share this post


Link to post
Share on other sites
Thanks! :]
However I meant that in asm it is possible to get both div and mod result with one instruction :]

Share this post


Link to post
Share on other sites
It might be a single instruction, but that doesn't make it a [b]fast instruction[/b]. Some light [url="http://www.penguin.cz/~literakl/intel/intel.html"]Googling[/url] places the idiv instructions at about 20-40 clock cycles, whereas shifts are between 3 or 4. I don't know whether this still holds with modern processors, or how relevant it is given that the processor is usually waiting for memory.

Using assembly can actually slow down your program, unless you 100% know what you are doing. Some compilers, e.g. Microsoft's, treat assembly as a "black box", which they won't optimise or inline. Lack of information about what happens inside this black box reduces the available optimisations in the surrounding code.

If you don't know assembly, I'd recommend you get pretty good at that [b]first[/b], before you try to use it in a real program. Otherwise you'll just make your code more brittle without any meaningful difference to performance, or a negative difference.

Share this post


Link to post
Share on other sites
Compiler will only optimize division and modulo with shifts and ANDs if the divisor is power of two. So in your specific example where you divide by 8 it is faster to leave it as it is, since those operations are much faster than a single division instruction.

However if you were to divide by some other non POT value, then you can use the normal "div" instruction. If I remember correctly, the remainder of the division should be in EDX register. I am not sure if compiler would do this optimization automatically.

Share this post


Link to post
Share on other sites
I played a bit more with the above program. I tried a few different NPOT constants and GCC turned them into multiplications, shifts and other non-division instructions. When I changed it to a non-constant, it only issued a single idiv instruction.

Have you tried looking at what your compiler is generating?

Share this post


Link to post
Share on other sites
[quote name='BearishSun' timestamp='1309782221' post='4830916']
Compiler will only optimize division and modulo with shifts and ANDs if the divisor is power of two. So in your specific example where you divide by 8 it is faster to leave it as it is, since those operations are much faster than a single division instruction.

However if you were to divide by some other non POT value, then you can use the normal "div" instruction. If I remember correctly, the remainder of the division should be in EDX register. I am not sure if compiler would do this optimization automatically.
[/quote]
Actually that is not completely true, the pinhole optimizers can and sometimes do deal with non-power of 2 so long as the "components" are power of 2. I.e. 48*xyz is equivalent to 32*xyz + 16*xyz and some optimizers will notice this and if the two shifts and add pipeline better it can/will do this. Sorry for pointing this out, just thought it may be interesting given the context.

Share this post


Link to post
Share on other sites
Sign in to follow this  

  • Advertisement