Home
Fractals
Tutorials
Books
My blog
My LinkedIn Profile

BOOKS i'm reading

Napoleon Hill Keys to Success: The 17 Principles of Personal Achievement, Napoleon Hill, ISBN: 978-0452272811
The 4-Hour Workweek: Escape 9-5, Live Anywhere, and Join the New Rich (Expanded and Updated), Timothy Ferriss, ISBN: 978-0307465351
The Fountainhead, Ayn Rand, ISBN: 0452273331
Web Hosting Canada

mailto:olivier@olivierlanglois.net

Archives for: November 2007, 25

11/25/07

Permalink 05:46:39 pm, by lano1106, 366 words, 3921 views   English (CA)
Categories: C++

Optimizing C++ code for Intel processors (and possibly many others)

Pipelined processors do not like conditional branches because when conditional branch is entering in the pipeline, the processor does not know yet what will be the condition outcome but it needs to continue to fill the pipeline with new instructions. In that situation the processor will try to predict what will be the condition outcome and start filling the pipeline from its prediction. This is very costly performance wise when the processor mispredicts the condition outcome. When that happens, the processor has to flush the pipeline content and start filling it again from the branch point.

The conditional branch prediction module has 2 components. There is the branch outcome cache that will help the processor to remember what was the outcome of branches that it has encountered recently and a static branch prediction algorithm.

The static branch algorithm is very simple:

  • If the conditional branch goes further in the code, it will not be taken
  • If the conditional branch goes back in the code, it will be taken

If you read the Intel processor optimization guide, few simple guidelines are easy to apply:

1. Eliminate branches

Some conditional branches are simply useless as whether they are taken or not, it will not make any differences on the end result. 2 examples of these are:

if( i != 0 )
  i = 0;

or

if( ptr != NULL )
{
  delete ptr;
  ptr = NULL;
}

should be changed to:

i = 0;

and

delete ptr;
ptr = NULL;

It is faster to reassign the same value to a variable than conditionally assign it. The operator delete and the function free() already contain safeguards against NULL pointers and it should be quite exceptional that they are called anyway with NULL pointers.

2. Refactor if statement conditions to be true most of the time.

Here is an example:

if(error)
{
  // exceptional error handling
}
else
{
  // normal operation
}

would be faster written like:

if(!error)
{
  // normal operation
}
else
{
  // exceptional error handling
}

To conclude, these guidelines are specifically for Intel processors but since the static branch prediction algorithm used by Intel processors is widely used by other pipelined processors (PowerPC is one of them), applying them on other platforms should be beneficial as well. Also, if you are looking for other optimization techniques, you can consult these sources.

Olivier Langlois's blog

I want you to find in this blog informations about C++ programming that I had a hard time to find in the first place on the web.

November 2007
Sun Mon Tue Wed Thu Fri Sat
 << < Current> >>
        1 2 3
4 5 6 7 8 9 10
11 12 13 14 15 16 17
18 19 20 21 22 23 24
25 26 27 28 29 30  

Search

Custom Search

Misc

XML Feeds

What is RSS?

Who's Online?

  • Guest Users: 39

powered by
b2evolution