Re: [PATCH] fix ICE in generic_overlap (PR 84526)

[Jakub Jelinek <jakub at redhat dot com>]

On Fri, Feb 23, 2018 at 12:57:14PM -0700, Martin Sebor wrote:
> +  /* get_inner_reference is not expected to return null.  */
> +  gcc_assert (base != NULL);
> +
>    poly_int64 bytepos = exact_div (bitpos, BITS_PER_UNIT);
>  
> -  HOST_WIDE_INT const_off;
> -  if (!base || !bytepos.is_constant (&const_off))
> -    {
> -      base = get_base_address (TREE_OPERAND (expr, 0));
> -      return;
> -    }
> -
> +  /* There is no conversion from poly_int64 to offset_int even
> +     though the latter is wider, so go through HOST_WIDE_INT.
> +     The offset is expected to always be constant.  */
> +  HOST_WIDE_INT const_off = bytepos.to_constant ();

The assert is ok, but removing the bytepos.is_constant (&const_off)
is wrong, I'm sure Richard S. can come up with some SVE testcase
where it will not be constant.  If it is not constant, you can handle
it like var_off (which as I said on IRC or in the PR also seems to be
incorrect, because if the base is not a decl the variable offset could be
negative).

>    offrange[0] += const_off;
>    offrange[1] += const_off;
>  
> @@ -923,7 +923,11 @@ builtin_access::generic_overlap ()
>        /* There's no way to distinguish an access to the same member
>  	 of a structure from one to two distinct members of the same
>  	 structure.  Give up to avoid excessive false positives.  */
> -      tree basetype = TREE_TYPE (TREE_TYPE (dstref->base));
> +      tree basetype = TREE_TYPE (dstref->base);
> +      if (POINTER_TYPE_P (basetype)
> +	  || TREE_CODE (basetype) == ARRAY_TYPE)
> +	basetype = TREE_TYPE (basetype);

This doesn't address any of my concerns that it is completely random
what {dst,src}ref->base is, apples and oranges; sometimes it is a pointer
(e.g. the argument of the function), sometimes the ADDR_EXPR operand,
sometimes the base of the reference, sometimes again address (if the
base of the reference is MEM_REF).  By the lack of consistency in what
it is, just deciding on its type whether you take TREE_TYPE or
TREE_TYPE (TREE_TYPE ()) of it also gives useless result.  You could e.g
call the memcpy etc. function with ADDR_EXPR of a VAR_DECL that has pointer
type, then if dstref->base is that VAR_DECL, POINTER_TYPE_P (basetype)
would be true.

	Jakub