From: nakata AT ccm DOT cl DOT nec DOT co DOT jp (Toshiyuki Nakata)
Subject: Re: Support for malign-double?
26 Aug 1998 17:32:55 -0700
Message-ID: <199808241600.BAA21520.cygnus.gnu-win32@csls3.csl.cl.nec.co.jp>
References: <199808190618 DOT PAA01030 AT csls3 DOT csl DOT cl DOT nec DOT co DOT jp>
Reply-To: Toshiyuki Nakata <nakata AT ccm DOT cl DOT nec DOT co DOT jp>
To: N8TM AT aol DOT com
Cc: gnu-win32 AT cygnus DOT com

; I sent the folowing mail to Mr. Khan, but I suspect some other people
; might be interested so I am sending the mail here as well.(With
; modifications to make the mail abit more understandable)
; Best Regards
; Toshiyuki
; nakata AT ccm DOT cl DOT nec DOT co DOT jp
; ------- Forwarded Message
; 
; I have one question on performance related issue.
; 
; Do you support -malign-double in mingwin-32 and gnuwin-32 (egcs-1.0.2)?
; 

; The reason is that I have been comparing the performance of a user's
; application on linux and Winnt and the interesting thing to note is that
; 
; They are comparable for integer performance but  linux performance is
; about 20% better for floating point intensive application.

Sorry, it seems that this has been fixed for the egcs1.0.2 for gnuwin32 b19
but not for egcs1.0.2 on  ming32.

In a nutshell, older versions (as well as the current mingw32 version)
did not seem to have aligned the common blocks on an 8 word boundary.

if you compile the following 2 files by 
gcc -c tm.c
g77 -O -malign-double a.f tm.o

and then do 
nm a.exe | egrep xx

then on egcs ver 1.0.2 on  mingwin32 I get,

0040a000 b _xx.4
0040f28c B _xx1_
004111d4 B _xx2_
0040d344 B _xx3_
0040b3fc B _xx4_


while on cygwin-32 b19 + egcs 1.0.2, I get

0040a000 b _xx.4
0040f330 B _xx1_
004102d8 B _xx2_
0040d3e8 B _xx3_
0040b4a0 B _xx4_

And the timing values in millisecs 
differ. ( I would like to have the no. 1 and no. 3 faster than no.2 and no.4)

-----a.f
      program main
      implicit double precision (a-h,o-z)
      common /xx1/ list1(2002)
      common /xx2/ list2(2002)
      common /xx3/ dlist1(1001)
      common /xx4/ dlist2(1001)
      dimension list3(3)
      equivalence (dlist2(1), list3(1))
      call milltimer(is)
      call sub(list1(1),dble(0.2))
      call milltimer(ie)
      call sub(list2(2),dble(0.2))
      call milltimer(if)
      call sub(dlist1,dble(0.2))
      call milltimer(ix)
      call sub(list3(2),dble(0.2))
      call milltimer(iy)
      write(*,*), ie-is,if-ie,ix-if,iy-ix
      stop
      end
      subroutine sub(d,ds)
      implicit double precision (a-h,o-z)
      dimension d(1000)
      do i=1,100000
         do j=1,1000
         d(j)= ds
         d(j+1)=d(j)*ds
         end do
      end do
      return
      end
-----------tm.c
#include <sys/timeb.h>



milltimer_(itime)
int *itime;
{
    struct timeb mytime;
    int ires;
/*    gettimeofday( &mytime, 0);*/
    ftime(&mytime);
    ires= 1000* (mytime.time & 0xffff) + (mytime.millitm);
    *itime=ires;
}

Best Regards
Toshiyuki Nakata
-
For help on using this list (especially unsubscribing), send a message to
"gnu-win32-request AT cygnus DOT com" with one line of text: "help".