Re: target/3920: wrong register number for CTR in stabs

["Mattias Engdegård" <mattias@virtutech.se>]

The following reply was made to PR target/3920; it has been noted by GNATS.

From: "=?ISO-8859-1?Q?Mattias Engdeg=E5rd?=" <mattias@virtutech.se>
To: geoffk@redhat.com
Cc: rth@gcc.gnu.org, gcc-bugs@gcc.gnu.org, gcc-prs@gcc.gnu.org,
   geoffk@gcc.gnu.org, gcc-gnats@gcc.gnu.org
Subject: Re: target/3920: wrong register number for CTR in stabs
Date: Mon, 24 Jun 2002 12:02:53 +0200

 >Why do you think this is incorrect?  The assembler code generated by
 >GCC 3.1 with -O2 -g -mregnames is:
 [...]
 >.L9:
 >        add %r0,%r0,%r3
 >        addi %r3,%r3,-1
 >        bdnz .L9
 
 You are right of course. Gcc 2.95 didn't use an integer register that
 worked in parallel with CTR and would instead do a mfctr in the loop
 (which is less efficient of course), so I was confused in my reporting.
 Sorry about that.
 
 A better example would be
 
 int sum(int n)
 {
         int i, j = 0, s = 0;
         for(i = n; i != 0; i--) {
 		s += j;
 		j *= 2;
 	}
         return s;
 }
 
 the loop of which gcc 3.0.2 (I didn't have 3.1 handy) turns into the nice code
 
 .L8:
         add %r3,%r3,%r0
         slwi %r0,%r0,1
         bdnz .L8
 
 but where the stabs claim 'i' lives in register 66, which does not make sense
 for CTR. The DWARF2 output (compiling with -gdwarf-2) seems to use the same
 incorrect numbering.
 
 I will check 3.1 as soon as I have built it.