[lno] tree-ssa merge

[Dorit Naishlos]

> Yes, I'm working on making the analyzer work again correctly.  The
> changes brought in code that is equivalent of the loop header copy:
> instead of having the exit condition of a loop at the top of the loop,
> ....
> ....
> Based on the fact that the loop count is computed using the exit
> expression, the nb_iterations is thus sometimes one iteration less
> than the true number of iterations.  In fact, at the moment, the
> nb_iterations is the number of times the exit condition has to be
> evaluated at runtime before exiting the loop.  I'm working on a fix.

Thanks for the detailed explanation Sebastian - I have also fixed the
loop-bound transformation in the vectorizer accordingly. I'll send a patch
soon.

dorit



                                                                                                                              
                      Pop Sebastian                                                                                           
                      <pop@gauvain.u-st        To:       Dorit Naishlos/Haifa/IBM@IBMIL                                       
                      rasbg.fr>                cc:       gcc-patches@gcc.gnu.org, Zdenek Dvorak                               
                                                <rakdver@atrey.karlin.mff.cuni.cz>                                            
                      22/01/2004 16:47         Subject:  Re: [lno] tree-ssa merge                                             
                                                                                                                              




On Thu, Jan 22, 2004 at 04:40:45PM +0200, Dorit Naishlos wrote:
>
> > +          * tree-ssa-vect-*: Use -fdump-tree-vect instead of
> > +          -fdump-tree-vect-stats.
>
> Sebastian, are you sure about this? it looks like at least for the
> vectorizer the -stats is required.

You're right, I missed the fact that -fdump-tree-vect-stats works.
This was almost an automatic change.  Sorry.

> The patch below fixes that, and also
> fixes a compilation problem in one of the tests cases. (Currently the
test

Thanks.

> cases fail due to the problem raised in
> http://gcc.gnu.org/ml/gcc-patches/2004-01/msg02273.html)
>

Yes, I'm working on making the analyzer work again correctly.  The
changes brought in code that is equivalent of the loop header copy:
instead of having the exit condition of a loop at the top of the loop,

a_0 = ...
loop
  a_1 = phi (a_0, a_2)
  if (a_1 op x) exit the loop
  ...
  a_2 = ...
endloop

the LCH copies the condition out of the loop, and moves the exit of
the loop at the end of the loops body.  The transformed loop looks as
follows:

a_0 = ...
if (a_1 op x) enter the loop
  loop
    a_1 = phi (a_0, a_2)
    ...
    a_2 = ...
    if (a_2 op x) exit the loop
  endloop
endif

Based on the fact that the loop count is computed using the exit
expression, the nb_iterations is thus sometimes one iteration less
than the true number of iterations.  In fact, at the moment, the
nb_iterations is the number of times the exit condition has to be
evaluated at runtime before exiting the loop.  I'm working on a fix.