grokfield miscompiled with -O9

[Richard Henderson]

On Sat, Jan 08, 2000 at 05:52:22PM -0200, Alexandre Oliva wrote:
> It turned out that expand_compound_operation would disregard the
> zero_extend because reg_nonzero_bits[28] said only its two
> least-significant bits could be zero...

This makes the compiler produce correct results. 

But I don't understand the intent of the code I've killed, particularly
in context of the other places this function is called.  It may be that
this code is correct, but that set_nonzero_bits_and_sign_copies shouldn't
be using expand_field_assignment.

Can someone more familiar with this code lend an eye?  This is real
antique stuff -- mid 1991 according to Cygnus' cvs annotate...



r~



Index: combine.c
===================================================================
RCS file: /cvs/gcc/egcs/gcc/combine.c,v
retrieving revision 1.107
diff -c -p -d -r1.107 combine.c
*** combine.c	2000/01/04 08:01:42	1.107
--- combine.c	2000/01/08 20:18:57
*************** expand_field_assignment (x)
*** 5653,5660 ****
--- 5653,5665 ----
  	    }
  	}
  
+ #if 0
        /* A SUBREG between two modes that occupy the same numbers of words
  	 can be done by moving the SUBREG to the source.  */
+       /* ??? No it can't.  Or at least not when called from 
+ 	 set_nonzero_bits_and_sign_copies, since we lose the fact that
+ 	 the upper bits of an assignment into a non-paradoxical subreg
+ 	 are now undefined.  */
        else if (GET_CODE (SET_DEST (x)) == SUBREG
  	       && (((GET_MODE_SIZE (GET_MODE (SET_DEST (x)))
  		     + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD)
*************** expand_field_assignment (x)
*** 5667,5672 ****
--- 5672,5678 ----
  			    SET_SRC (x)));
  	  continue;
  	}
+ #endif
        else
  	break;