[PATCH] Fortran/PR31593 Speed up some loops

[Tobias Schlüter]

Hi Jack,

Jack Howarth wrote:
> Does this address the same issue as in PR33113? Do you see any
> instances remaining of "data ref analysis failed" in the dump
> file generated with the -fdump-tree-vect-details flag for the
> polyhedron benchmarks? I have previously posted the distribution
> of this issue across the benchmarks...
> 
> http://gcc.gnu.org/ml/gcc/2009-08/msg00079.html
> 
> It would be great if we could suppress this issue within
> gfortran since the vectorization folks don't seem to be
> in a hurry to fix it from their end.

I don't see a change in the number, but in the distribution of these 
failures.  I haven't tried before / after my patch.  Find attached the 
list with my patch.

I reran the polyhedron suite without -funroll-loops, and now I see 
enhancements with my patch throughout.  List follows near the bottom. 
First line is always without my patch, second with the patch.  The first 
number is always the average over three runs, the second the variance. 
Except for capacita, if the change is outside the variance, it is 
towards shorter times.

Finally, Tobias approved the patch on IRC but I enhanced it a bit based 
on his commentary..

ac.f90
16.8467  var = 0.0181556
17.26  var = 0.4214
aermod.f90
41.33  var = 0.567467
41.7933  var = 0.299622
air.f90
17.8467  var = 0.0901556
18.32  var = 0.0224
capacita.f90
89.6667  var = 0.00595556
91.97  var = 0.0896
channel.f90
15.0333  var = 0.00388889
13.9967  var = 0.0124222
doduc.f90
50.3167  var = 0.575556
50.75  var = 0.346467
fatigue.f90
17.0867  var = 0.360289
16.9467  var = 0.0464889
gas_dyn.f90
17.19  var = 0.161667
16.7167  var = 0.128156
induct.f90
61.66  var = 0.2342
61.0567  var = 0.751356
linpk.f90
27.6867  var = 0.0377556
26.6067  var = 0.0333556
mdbx.f90
20.7133  var = 0.176689
20.3933  var = 0.264089
nf.f90
36.4533  var = 0.281956
34.51  var = 0.0620667
protein.f90
71.4033  var = 0.168289
70.7567  var = 0.00402222
rnflow.f90
47.5833  var = 0.393156
44.51  var = 0.0284667
test_fpu.f90
25.4367  var = 0.0938889
24.47  var = 0.0274667
tfft.f90
9.48  var = 0.1986
9.61333  var = 0.223356
-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: vect_failed.txt
URL: <http://gcc.gnu.org/pipermail/fortran/attachments/20090815/b28401dc/attachment.txt>
-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: better_loops.diff.txt
URL: <http://gcc.gnu.org/pipermail/fortran/attachments/20090815/b28401dc/attachment-0001.txt>