[patch, libgfortran] Re-factored write_float, by kind (RFC)

[Steve Kargl]

On Tue, Aug 07, 2007 at 10:44:47PM -0700, Jerry DeLisle wrote:
> 
> Well it has taken me a while to get to this point.  The attached patch 
> provides some small performance improvement in formatted output of reals.  
> This gets rid of the translation of all types into GFC_REAL_LARGEST before 
> doing the output. It also avoids using some instances of memcpy.
> 
(snip)
> 
> Discussion welcome,
> 

Ran the Polyhedron benchmark with and without your patch.  The
average of 4 runs is shown

          without   with
ac         19.17   19.19
aermod     46.20   46.04
air        19.61   19.36
capacita  118.42  117.19
channel    15.16   15.17
doduc      54.82   53.90
fatigue    21.26   20.39
gas_dyn    22.21   22.61
induct     60.20   60.26
linpk      31.19   31.67
mdbx       23.62   23.70
nf        131.04  127.21
protein    67.51   67.94
rnflow     60.21   60.63
test_fpu   29.22   29.42
tfft        9.72    9.72

There are small speed ups in some case with nf showing a 3%
improvement.  I still need to inspect output from some of
my codes to see if there are any inconsistencies.

-- 
Steve