Re: Will tree-ssa be GCC 3.5?

[Steven Bosscher <s dot bosscher at student dot tudelft dot nl>]

On Monday 01 December 2003 16:33, Richard Kenner wrote:
>     <grin> but only after duplicating about 20% of the tree-ssa work
>     (i.e. all function lowering, inlining and part of expansion), and the
>     generated code would be so horrible that no sane Fortran user would
>     dare touch it ;-)
>
> Yes of course you might need to use some of the lowering code written
> for tree-ssa to lower into older-style trees, but I don't see that as
> significant.  And I don't understand your last comment: unlike C++,
> the important optimizations for Fortran are relatively low-level and
> will benefit less from tree-ssa than other languages.

Well like I said, "been there, done that".  gfortran with the tree-ssa 
optimizations produces significantly better code than without it. 

Your last claim is simply not true.  The important optimizations for recent 
Fortran standards are not even implemented in GCC, because they all heavily 
depend on the availability of dependency information at a very high level, 
before scalarizing array syntax for example. We lose big time here because 
GCC doesn't even support multi-D arrays, let alone code generation for array 
assignments.  But at least with tree-ssa we can get kind of reasonable code 
out of gfortran.

The optimizations that Fortran needs are not the same as those necessary for 
an optimizing C++ compiler, but I'd certainly not call them low-level.

Gr.
Steven