GCC/g77-3.0-to-be compiles our NWP correctly - for the first time

GCC/g77-3.0-to-be compiles our NWP correctly - for the first time

[Toon Moene]

L.S.,

Now that the problems with compiling libstdc++-v3 on
alphaev6-unknown-linux-gnu have been dealt with, I am pleased to
announce that the compiler (download of 2000061010UTC) managed to build
our Numerical Weather Forecasting system (around 1300 Fortran routines
and a dozen C routines) correctly on this target.

I wouldn't be surprised if the following fix made this finally possible:

"Fixes gcc.c-torture/execute/20000605-1.c, derived from gtk,
on alphaev6.  The problem was that we'd inherit an existing
DImode value from a floating point value into the SImode
subreg here.

(insn 80 70 71 (set (reg:DI 99)
        (sign_extend:DI (subreg:SI (reg:DI 92) 0))) 1 {*lda-3}
    (nil))

Which of course results in the wrong bit patterns being 
converted back to SFmode.


r~


        * reload1.c (choose_reload_regs): Mind CLASS_CANNOT_CHANGE_SIZE
        while looking for a register to inherit from.
"

because the part that went belly up does a lot of float->int->float
conversions to group observations in geographical (i.e.
latitude/longitude) boxes.

GCC-3.0-pre is somewhat faster than GCC-2.95.2, although it seems not to
honor the -funroll-loops flag.  I'll wait investigating that until Jan
Hubicka's ("Wat zijn die Tsjechen stil") loop patches have gone in.

Cheers,

-- 
Toon Moene - mailto:toon@moene.indiv.nluug.nl - phoneto: +31 346 214290
Saturnushof 14, 3738 XG  Maartensdijk, The Netherlands
GNU Fortran 77: http://gcc.gnu.org/onlinedocs/g77_news.html
GNU Fortran 95: http://g95.sourceforge.net/ (under construction)

Re: GCC/g77-3.0-to-be compiles our NWP correctly - for the first time

[Timothy Prince]

We have continued problems with -funroll-loops (including the long
standing ICE in LAPACK) on x86, which seem to occur in cases where the
compiler prepares to unroll and then decides (probably reasonably so)
not to do so.  One loop in LFK compiles without complaint, but appears
to have built for a fixed loop count where it needs to be variable.  I
don't have multiple architectures at my disposal to see if this is
platform dependent (still no employment, other than short term tasks).

Tim Prince
----- Original Message -----
From: "Toon Moene" <toon@moene.indiv.nluug.nl>
To: <gcc@gcc.gnu.org>
Sent: Saturday, June 10, 2000 11:47 AM
Subject: GCC/g77-3.0-to-be compiles our NWP correctly - for the first
time


> GCC-3.0-pre is somewhat faster than GCC-2.95.2, although it seems not
to
> honor the -funroll-loops flag.  I'll wait investigating that until Jan
> Hubicka's ("Wat zijn die Tsjechen stil") loop patches have gone in.
>
> Cheers,
>
> --
> Toon Moene - mailto:toon@moene.indiv.nluug.nl - phoneto: +31 346
214290
> Saturnushof 14, 3738 XG  Maartensdijk, The Netherlands
> GNU Fortran 77: http://gcc.gnu.org/onlinedocs/g77_news.html
> GNU Fortran 95: http://g95.sourceforge.net/ (under construction)

Re: GCC/g77-3.0-to-be compiles our NWP correctly - for the first time

[Toon Moene]

I wrote:

> "Wat zijn die Tsjechen stil"

Well, on second thoughts, I wouldn't want to "win" this way ....

-- 
Toon Moene - mailto:toon@moene.indiv.nluug.nl - phoneto: +31 346 214290
Saturnushof 14, 3738 XG  Maartensdijk, The Netherlands
GNU Fortran 77: http://gcc.gnu.org/onlinedocs/g77_news.html
GNU Fortran 95: http://g95.sourceforge.net/ (under construction)