Re: [gfortran] Avoid induction variables in DO loops.

[Canqun Yang <canqun at nudt dot edu dot cn>]

Daniel Berlin <dberlin@dberlin.org>:

> 
> This may or may not be correct, it's hard to say.
> Yes it is true that our data dependence testing is 
still in the process of 
> being written.
> We have ZIV test, SIV test, and a partial MIV test 
done.
> As written below, it may not fall into the MIV test 
portion we have 
> implemented yet.
> I honestly don't remember.
> 
> It would be good to know what your goal is.
> IE why you care whether it can do this data 
dependence analysis or not.

I lead a small group trying to tuning GCC, especially 
on IA64 alike platform. 

Some progress has been made after nearly half an year. 
For example, we can lower the execution time of SPEC 
CFP2000 program 171.swim from 817 seconds to 310 
seconds by extending the alias analysis for Fortran 95 
and modifying the induction variables recognition and 
prefetch optimization of GCC backend. But the 
performance of GCC on IA64 is still far behind of 
Intel compiler (it takes less than 100 seconds to run 
171.swim).

According to our experiments, to further improve the 
performance of 171.swim, loop nested optimization 
(loop distribution/interchange) and software pipeling 
with rotating register allocation are needed to be 
implementated. 

Obviously, data dependence analysis is the base of 
loop nested optimization, so we take much attention on 
it. Very appreciate your work.


> Are you trying to perform some specific optimization?
> 

We plan to implement scalar replacement. It is also 
based on dependence analysis.

> I mean, you are welcome to implement the rest of the 
miv test if you like 
> (see tree-data-ref.c:analyze_miv_subscript for what 
is currently 
> implemented, however, it helps avoid duplicate work 
if we know what your 
> plans are, as there may already be people working on 
what you want to do.
> 
> 
I have no plan to do it now, but I can contact with 
members in my group to know whether there is someone 
has instrest to continue this work. 


Canqun Yang
Creative Compiler Research Group.
National University of Defense Technology, China.