Code gen error on PowerPC

[Franz Sirl]

At 01:01 02.03.99 , Jeffrey A Law wrote:

>  In message < 4.2.0.25.19990301224624.03f0e420@mail.lauterbach.com >you write:
>  > At 19:03 01.03.99 , Richard Henderson wrote:
>  > >On Mon, Mar 01, 1999 at 06:52:59PM +0100, Franz Sirl wrote:
>  > >> Ugh, that's not the answer I wanted to hear :-(. This probably 
> means I h
>  > ave
>  > >> to do an extensive debugging session again... Isn't there any 
> chance tha
>  > t
>  > >> you fixed a bug without knowing? :-)
>  > >
>  > >I suppose it's possible -- a lot of code changed.  It seems more
>  > >likely that differing combine_givs heuristics merely hid the problem.
>  > >
>  > >> Any hints on what I should look at first during debugging?
>  > >
>  > >If I understood your explanation right, you've got a GIV being
>  > >recognized that isn't.  I'd start by finding out why that's so.
>  > >
>  > >You'll want to put a breakpoint in strength_reduce, right before
>  > >the call to general_induction_var.  Conditionalize it on the insn
>  > >that contains the GIV.  Likely the meat of the problem will be in
>  > >simplify_giv_expr, cases REG and default.
>  >
>  > The underlying problem seems to be that the following hunk of code (it is
>  > entered with auto_inc_opt is -1, tv->insn is insn 33, v->insn is insn 71)
>  > around line 4210 generates insn 158 and inserts it at the beginning of 
> the
>  > _inner_ loop instead of the outer loop:
>[ ... ]
>OK.  I think this reinforces Richard's theory that something is being
>recognized as a general induction variable that really is not a general
>induction variable for the inner loop.
>
>So, I think you need to be looking into the calls to general_induction_var
>like Richard suggested.
>
>Some call to that function is returning a nonzero value when it should be
>returning zero if Richard's theory is correct.

Hmm, I think this giv is setup by this hunk of code in find_mem_givs:

    case MEM:
      {
        rtx src_reg;
        rtx add_val;
        rtx mult_val;
        int benefit;

        benefit = general_induction_var (XEXP (x, 0),
                                         &src_reg, &add_val, &mult_val);

        /* Don't make a DEST_ADDR giv with mult_val == 1 && add_val == 0.
           Such a giv isn't useful.  */
        if (benefit > 0 && (mult_val != const1_rtx || add_val != const0_rtx))
          {
            /* Found one; record it.  */
            struct induction *v
              = (struct induction *) oballoc (sizeof (struct induction));

            record_giv (v, insn, src_reg, addr_placeholder, mult_val,
                        add_val, benefit, DEST_ADDR, not_every_iteration,
                        &XEXP (x, 0), loop_start, loop_end);

            v->mem_mode = GET_MODE (x);
          }
      }
      return;

But it's too late now and I'm not sure about anything :-) I'll continue 
debugging tomorrow or wednesday.

Franz.