problem with iv-opts and valid gimple

problem with iv-opts and valid gimple

[Andrew Pinski]

Even though I cannot find a testcase on the mainline, I could reproduce
the problem with my tree combiner.

find_interesting_uses_cond does not handle TRUTH_NOT_EXPR.

Could you look into how to fix this problem?
This might be able to reproducible on the mainline but I cannot
find a testcase right now.

Thanks,
Andrew Pinski

Re: problem with iv-opts and valid gimple

[Andrew Pinski]

On Dec 10, 2004, at 3:48 PM, Andrew Pinski wrote:

> Even though I cannot find a testcase on the mainline, I could reproduce
> the problem with my tree combiner.
>
> find_interesting_uses_cond does not handle TRUTH_NOT_EXPR.
>
> Could you look into how to fix this problem?
> This might be able to reproducible on the mainline but I cannot
> find a testcase right now.

I had forgot to mention the COND_EXPR looks like:

if (!a) goto b; else goto c;

Thanks,
Andrew Pinski

Re: problem with iv-opts and valid gimple

[Zdenek Dvorak]

Hello,

> 
> On Dec 10, 2004, at 3:48 PM, Andrew Pinski wrote:
> 
> >Even though I cannot find a testcase on the mainline, I could reproduce
> >the problem with my tree combiner.
> >
> >find_interesting_uses_cond does not handle TRUTH_NOT_EXPR.
> >
> >Could you look into how to fix this problem?
> >This might be able to reproducible on the mainline but I cannot
> >find a testcase right now.

making it kinda hard to test :-). Given that find_interesting_uses_cond
would ICE with TRUTH_NOT_EXPR, it probably cannot be reproduced easily.
Try the patch below.

Zdenek

Index: tree-ssa-loop-ivopts.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/tree-ssa-loop-ivopts.c,v
retrieving revision 2.35
diff -c -3 -p -r2.35 tree-ssa-loop-ivopts.c
*** tree-ssa-loop-ivopts.c	7 Dec 2004 21:23:04 -0000	2.35
--- tree-ssa-loop-ivopts.c	10 Dec 2004 21:09:41 -0000
*************** find_interesting_uses_cond (struct ivopt
*** 1206,1211 ****
--- 1206,1216 ----
        op0_p = cond_p;
        op1_p = &zero;
      }
+   else if (TREE_CODE (*cond_p) == TRUTH_NOT_EXPR)
+     {
+       op0_p = &TREE_OPERAND (*cond_p, 0);
+       op1_p = &zero;
+     }
    else
      {
        op0_p = &TREE_OPERAND (*cond_p, 0);
*************** determine_use_iv_cost_condition (struct 
*** 3255,3260 ****
--- 3313,3320 ----
       rid of it.  */
    if (TREE_CODE (*use->op_p) == SSA_NAME)
      record_invariant (data, *use->op_p, true);
+   else if (TREE_CODE (*use->op_p) == TRUTH_NOT_EXPR)
+     record_invariant (data, TREE_OPERAND (*use->op_p, 0), true);
    else
      {
        record_invariant (data, TREE_OPERAND (*use->op_p, 0), true);
*************** rewrite_use_compare (struct ivopts_data 
*** 4544,4553 ****
    comp = unshare_expr (get_computation (data->current_loop, use, cand));
  
    cond = *use->op_p;
!   op_p = &TREE_OPERAND (cond, 0);
!   if (TREE_CODE (*op_p) != SSA_NAME
!       || zero_p (get_iv (data, *op_p)->step))
!     op_p = &TREE_OPERAND (cond, 1);
  
    op = force_gimple_operand (comp, &stmts, true, SSA_NAME_VAR (*op_p));
    if (stmts)
--- 4625,4639 ----
    comp = unshare_expr (get_computation (data->current_loop, use, cand));
  
    cond = *use->op_p;
!   if (TREE_CODE (cond) == SSA_NAME)
!     op_p = use->op_p;
!   else
!     {
!       op_p = &TREE_OPERAND (cond, 0);
!       if (TREE_CODE (*op_p) != SSA_NAME
! 	  || zero_p (get_iv (data, *op_p)->step))
! 	op_p = &TREE_OPERAND (cond, 1);
!     }
  
    op = force_gimple_operand (comp, &stmts, true, SSA_NAME_VAR (*op_p));
    if (stmts)