An unusual Performance approach using Synthetic registers

[Andy Walker]

Thank you for your kind comments.  Also, thank you very much for the 
reference to the Lal George paper.

On Friday 27 December 2002 03:03 am, Chris Lattner wrote:

<snip>

> You may win a suprising amount strictly because of the register
> renaming hardware found on modern x86 processors, which would eliminate
> the L1 reference. 

I had not considered this.  

<snip>

> On the other hand, I think this approach is not the right one to take if
> improving optimizer effectiveness is the goal.  

My intuitin says this can be a fertile area for investigation, for these 
reasons:  Algorithms for register use optimization are very good.  My 
experience with numerical analysis and evaluation of algorithms, in graduate 
school, leads me to believe that color-graphing leads to a result that is 
either optimal or near-optimal.  

Algorithms for memory use optimization are not nearly so well developed.  

In one sense, this approach co-opts the optimization algorithms for registers 
and uses them on small amounts of memory.  

Color-graphing works better, the more registers there are available for use.  
Six registers is sub-optimal.  Sixteen registers is a good start.  Thirty-two 
plus five gives thirty-seven, enough for the algorithm to excel.

<snip>

> You will probably find that the number of "virtual" registers varies
> widely across different incarnations of the architecture, implying that
> the -march options should change the number of virtual registers.  Maybe
> keeping this in mind when you design the code will help down the line.  :)

I expect the benefit of Synthetic registers to vary with the complexity of a 
function.  For itty-bitty subroutines, they will be a detriment, not a 
benefit.  For longer, complicated, routines, there will be more benefit.  I 
do not hold out my coding skills as an example of the best, but my keyboard 
does not really get warmed up until after I have added a couple of hundred 
lines to a routine.  

On machines like the intel 80386 that do not use cache, I expect Synthetic 
registers to have very limited benefits.

> This will certainly be an interesting project, please keep the list
> informed with what you find.  :)

Certainly.

> -Chris