g77/686 -O2 in 19991122 breaks Kahan sum

[N8TM@aol.com]

While ss 19991122 has fixed certain bugs with -O2 on i686, it seems now that 
-O2 is violating parentheses.  Kahan' s sum formula works accurately at -Os 
but not at -O2, as it does in gcc-2.95.2.  i.e.

      s= 0
      y= 0
C
C Kahan Summation: Knuth, D.E. 1981 "The Art of Programmming"
C Addison-Wesley, Reading MA USA vol II 2nd ed (proof p. 572)
      do k= 1,n
      y= y+(k)*v(k)
      t= s+y
      y= y-(t-s)
      s= t
    enddo

no longer produces accurate results.  It makes no difference whether 
-ffast-math is in effect.  If such "optimizations" are to be available, they 
should not be a default part of a normal optimization option.  It would be 
bad enough to have this intentional violation of language standards thrown 
together with the fast comparison in -ffast-math.  BTW, -ffast-math is still 
totally breaking some cases of max() and min().

Tim
tprince@computer.org