Re: [PATCH] Fix pdftex miscompilation due to get_range_strlen (PR tree-optimization/84478)

[Jakub Jelinek <jakub at redhat dot com>]

On Tue, Feb 20, 2018 at 12:03:26PM -0700, Martin Sebor wrote:
> PR tree-optimization/84478 - [8 Regression] pdftex miscompilation on i386
> 
> gcc/ChangeLog:
> 
> 	PR tree-optimization/84478
> 	* gimple-fold.c (get_range_strlen): Set *MINLEN to zero.
> 	(get_range_strlen): Reset range on failure.
> 
> gcc/testsuite/ChangeLog:
> 
> 	PR tree-optimization/84478
> 	* gcc.c-torture/execute/pr84478.c: New test.
> 
> Index: gcc/gimple-fold.c
> ===================================================================
> --- gcc/gimple-fold.c	(revision 257796)
> +++ gcc/gimple-fold.c	(working copy)
> @@ -1369,7 +1369,10 @@ get_range_strlen (tree arg, tree length[2], bitmap
>  	      tree eltype = TREE_TYPE (type);
>  	      if (TREE_CODE (type) != ARRAY_TYPE
>  		  || !INTEGRAL_TYPE_P (eltype))
> -		return false;
> +		{
> +		  *minlen = ssize_int (0);
> +		  return false;
> +		}

This is just one of the 13 spots where we return false, so this doesn't look
safe or sufficient to me, even when you actually honor the return value in
2 argument get_range_strlen.  You'd really need to do

              {
                if (fuzzy)
-		  *maxlen = build_all_ones_cst (size_type_node);
+		  {
+		    *minlen = size_int (0);
+		    *maxlen = build_all_ones_cst (size_type_node);
+		  }
                else
                  return false;
              }

or just drop that if (fuzzy) stuff from there, but that breaks the warning
tests.  It would help if you explained why you think it is a good idea
ignoring the other phi arguments if you have one (or more) where you can
determine length.

One variation of my patch could be instead of adding type 3 change
fuzzy from bool to int, and use fuzzy == 1 for the strlen value ranges and
fuzzy == 2 for the warning code (i.e. 2 operand get_range_strlen).

Note, my patch passed regtest on both x86_64-linux and i686-linux.

	Jakub