Re: Performance; was: 68k problem.

Re: Performance; was: 68k problem.

[Toon Moene]

Jeff wrote:

>  I'm not including -frerun-loop yet; I want to think about
>  it a little -- my experiences with trying to rerun passes
>  hasn't been good.  At various times over the years I've
>  done this with cse and flow, and in both cases doing so
>  has destabilized things pretty badly.

I must say we were also pretty surprised that it Just Worked.  It  
has been in g77 since g77-0.5.18, i.e. April 1996.  Note, however,  
that we had these flags turned off for all non-Fortran compiles by  
default.

Toon.

Re: Performance; was: 68k problem.

[Jeffrey A Law]

  In message <9708272100.AA01051@moene.indiv.nluug.nl>you write:
  > i.e. not as bad as I thought [38000+ seconds], but certainly worse  
  > than the g77-0.5.21-970811 + gcc-2.7.2.2 result (~ 10 %).
Just a note; I've got -freduce-all-givs and -fmove-all-movables
in my source tree, and barring any natual disasters they should
get checked in and appear in the next snapshot (tonight/tomorrow).

I'm not including -frerun-loop yet; I want to think about it a
little -- my experiences with trying to rerun passes hasn't been
good.  At various times over the years I've done this with cse
and flow, and in both cases doing so has destabilized things
pretty badly.

Jeff

Re: Performance; was: 68k problem.

[Toon Moene]

I wrote, in a previous message on this subject:

QUOTE
Quoting the log files of both runs (all times in seconds;  
optimisation -O2 -fforce-addr):

For g77-0.5.21-970811 + gcc-2.7.2.2:

HL_EXP12230704: FORECAST TOOK      32427.2949 SECONDS
UNQUOTE

Here are some more results:

For egcs-ss-970821:

HL_EXP26193432: FORECAST TOOK      35981.2305 SECONDS

i.e. not as bad as I thought [38000+ seconds], but certainly worse  
than the g77-0.5.21-970811 + gcc-2.7.2.2 result (~ 10 %).

Cheers,
Toon.

BTW, how do you want your egcs today: Fried, scrambled, cooked or  
continental ? [ Hi, Kate ]

And I'll refrain from explaining what "De Muyter" means in Dutch  
;-)

Performance; was: 68k problem.

[Toon Moene]

Yesterday, I wrote:

>  Subsequently, I succesfully built HIRLAM (see URL:
>  http://www.knmi.nl/hirlam ) and it is running now.  If
>  egcs performs as well as g77-0.5.21-970811 + gcc-2.7.2.2,
>  this will finish within 24 hours.

Unfortunately, it doesn't.  This is not really unexpected, as the  
egcs snapshot doesn't work around the heuristics in loop.c meant to  
prevent "too much register pressure".

Quoting the log files of both runs (all times in seconds;  
optimisation -O2 -fforce-addr):

For g77-0.5.21-970811 + gcc-2.7.2.2:

HL_EXP12230704:0SUPOBS TOOK :       20.02850341797
HL_EXP12230704:0DATACH TOOK :      9476.3925781250
HL_EXP12230704:0ANAEVA TOOK :     11479.2958984375
HL_EXP12230704:0GRPEVA TOOK :     22881.7519531250
HL_EXP12230704:0HUMSUP TOOK :        1.60937500000
HL_EXP12230704:0DATACH TOOK :       579.2109375000
HL_EXP12230704:0HUMEVA TOOK :       292.9531250000
HL_EXP12230704:0GRPEVA TOOK :       384.8281250000
HL_EXP12230704: PREPARATIONS TOOK         49.1833 SECONDS
HL_EXP12230704: FORECAST TOOK        286.4648 SECONDS
HL_EXP12230704: PREPARATIONS TOOK         55.1645 SECONDS
HL_EXP12230704: FORECAST TOOK      32427.2949 SECONDS

For egcs-ss-970821:

HL_EXP26193432:0SUPOBS TOOK :       21.24188232422
HL_EXP26193432:0DATACH TOOK :     13005.2167968750
HL_EXP26193432:0ANAEVA TOOK :     12855.5527343750
HL_EXP26193432:0GRPEVA TOOK :     24094.5722656250
HL_EXP26193432:0HUMSUP TOOK :        1.87500000000
HL_EXP26193432:0DATACH TOOK :       715.5117187500
HL_EXP26193432:0HUMEVA TOOK :       318.5078125000
HL_EXP26193432:0GRPEVA TOOK :       402.7421875000
HL_EXP26193432: PREPARATIONS TOOK         52.4014 SECONDS
HL_EXP26193432: FORECAST TOOK        296.5845 SECONDS
HL_EXP26193432: PREPARATIONS TOOK         57.6717 SECONDS

[ The forecast is not finished yet, but I expect it to clock in at
  38000+ seconds ]

In other words:  There are Lies, Damn Lies, and Heuristics

Fortunately, the EGCS project to the rescue: "If the compiler lies  
to you, you can get your revenge" *).

Cheers,
Toon.

*) That is, without sneaking in secret backend patches for g77.