🚀 go-pugleaf

Jim Robinson writes:
   
   GCC-2.1 was configured as m88k-sysv3 on a Motorola SVR3 88k (R32V3 M88100).
   The standard procedure was followed and everything  worked up until the
   comparison of stage 2 object files with stage 3 object files. A number of
   files differed:
   	cse.o emit-rtl.o integrate.o regclass.o reload1.o sched.o tree.o
   
It appears this is a bug in the native C compiler (Green Hills 1.8.5).
A few folks have run into this problem and each has come to this conclusion.
I suspect that the stage 3 compiler is actually good.  If this is true,
you should be able to make the stage 4 compiler and compare those objects
with the ones in stage3.

          make stage3
          make CC=stage3/gcc CFLAGS="-g -O -Bstage3/"

          for file in *.o; do
          tail +10c $file > foo1
          tail +10c stage3/$file > foo2
          cmp foo1 foo2 || echo $file
          done

Should you find that the miscompares go away, this indicates a bug in the
Green Hills compiler.

A second approach is to use GNU C to bootstrap (in a sense, that's what
the above is all about).  However, there is a 88open BCS compliant binary
release of GNU C version DG-1.37.29 available via anonymous ftp to
dg-rtp.dg.com:

-r--r--r--   1 1040753 Mar  6  1991 gcc-dg.1.37.29-88k.usr.tar.Z
-r--r--r--   1 2907221 Mar  6  1991 gcc-dg.1.37.29.tar.Z

The *.usr.tar.Z is the binary release; the *.tar.Z is the source release.

I have also been told that the miscompares did not happen with GCC 2.0.
A third alternative would be to use that to bootstrap GCC 2.1.