g77/egcs-19981019 - erro for a loop with assumed length array

[Mako Sasaki]

I found that g77 of egcs-19981019 snapshot can't compile some kind of loop and says "internal compiler error" when optimization is required.

# Sorry if some othe person has reported this already.

An example follows;

*--- easy.f ----------------------------------------
      subroutine sub( r, r8, n )
      implicit real*8 ( a-h,o-z )

      real r(n)
      real*8 r8(n)

c  ... non assumed length array works
cok   real r(1000) 
cok   real*8 r8(1000)

c  ... fixed iteration count works
cok   do i=1,1000
      do i=1,n
         r8(i) = r(i)*2.0
cxx      r8(i) = 1.0d0 - r(i)*2.0
cxx      r8(i) = 1.0 - r(i)*2.0
cok      r8(i) = r(i)+2.0
cok      r8(i) = r(i)*2.0d0
cok      r8(i) = 1.0 - r(i)*2.0d0
      end do
      write(*,*) r8, r
      end
*-------------------------------------------------

% g77 -c -O2 -funroll-loops -g easy.f
easy.f: In subroutine `sub':
easy.f:22: internal error--unrecognizable insn:
(insn:QI 299 190 184 (set (reg:SF 8 %st(0))
        (mem/s:SF (mem/u:SI (plus:SI (reg:SI 6 %ebp)
                    (const_int 8)) 0) 0)) -1 (nil)
    (nil))
toplev.c:1370: Internal compiler error in function fatal_insn

*-------------------------------------------------

In this case, error disappears when "-funroll-loops" is omitted, but for other more complicated source, error ocuurs without this option.  

Conditions to cause this error are ;

(0) optimization is required.
(1) loop iteration count is not a constant.
(2) a single precision assumed-length array is referred.
     ( r(n) in the case above.)
(3) result of single presision floating multiplication is promoted to double precision.


Thanks, in advance

                            Mako Sasaki   sasaki@tyo.sci.jri.co.jp