Re: [Gomp-discuss] Start of an OpenMP Implementation Spec

[Ross Towle]

On Thu, 26 Aug 2004 16:45:25 +0200, Biagio Lucini <address@hidden> wrote:

> Going into the specific, your note looks fine with me, although it is not
> clear to me whether you want to do a rewriting of the code and then a
> gimplification or if the operations you describe apply to the middle end.
> 

I used the term "middle end" in the same way as the GCC Fortran95
project uses it.  It is
the part of the compiler after the front end and before the gimplification.

I would have prefered to do the code rewriting later.  I would llike
to perform many (but not all)
scalar type of optimizations such as code motion, common subexpression
elimination.
Remember, the OpenMP directives do not change the semantics of the
language; so these
optimizations can be performed on the parallel region prior to
"outlining".  However, once the
region has been outlined these optimizations become harder to perform.
 It requires some
high quality interprocedural optimization.  Even with outlining, it is
still possible to optimize
easily within the outlined region.  Strength reduction is going to be
the optimization that
is essential for performance.

As I stated in my GOMP proposal, because of GIMPLE, the outlining needs to be 
performed before gimplification.  By adding the calls to the OpenMP
heleper routines and
the outlining before gimplification, the intermediate representation
given to the back end
for programs that had OpenMP directives, look no different than
non-OpenMP programs.
No changes will be needed for the back end. The back end will have no
idea that it is
optimizing and generating code for an OpenMP  program.  

-Ross A Towle