On Mon, Jun 10, 2019 at 10:37 PM Martin Sebor <mse...@gmail.com> wrote:
>
> On 6/10/19 1:29 PM, Jakub Jelinek wrote:
> > On Mon, Jun 10, 2019 at 01:23:28PM -0600, Martin Sebor wrote:
> >> + else if (integer_zerop (TREE_OPERAND (node, 1))
> >> + /* Dump the types of INTEGER_CSTs explicitly, for we can't
> >> + infer them and MEM_ATTR caching will share MEM_REFs
> >> + with differently-typed op0s. */
> >> + && TREE_CODE (TREE_OPERAND (node, 0)) != INTEGER_CST
> >> + /* Released SSA_NAMES have no TREE_TYPE. */
> >> + && TREE_TYPE (TREE_OPERAND (node, 0)) != NULL_TREE
> >> + /* Same pointer types, but ignoring POINTER_TYPE vs.
> >> + REFERENCE_TYPE. */
> >> + && (TREE_TYPE (TREE_TYPE (TREE_OPERAND (node, 0)))
> >> + == TREE_TYPE (TREE_TYPE (TREE_OPERAND (node, 1))))
> >> + && (TYPE_MODE (TREE_TYPE (TREE_OPERAND (node, 0)))
> >> + == TYPE_MODE (TREE_TYPE (TREE_OPERAND (node, 1))))
> >> + && (TYPE_REF_CAN_ALIAS_ALL (TREE_TYPE (TREE_OPERAND (node, 0)))
> >> + == TYPE_REF_CAN_ALIAS_ALL (TREE_TYPE (TREE_OPERAND (node, 1))))
> >> + /* Same value types ignoring qualifiers. */
> >> + && (TYPE_MAIN_VARIANT (TREE_TYPE (node))
> >> + == TYPE_MAIN_VARIANT
> >> + (TREE_TYPE (TREE_TYPE (TREE_OPERAND (node, 1)))))
> >
> > You should be using types_compatible_p rather than type equality, that is
> > all GIMPLE cares about.
>
> The code above predates my change, I just factored it out into
> a separate function; it does make the diff bigger. The code
> I introduced only adds the <...> if the size of the access
> differs from the size of the operand. I used type sizes rather
> than testing their compatibility to minimize the number of tests
> that might need updating.
>
> This is the salient bit:
>
> + pp_string (pp, "MEM");
> +
> + tree nodetype = TREE_TYPE (node);
> + tree op0 = TREE_OPERAND (node, 0);
> + tree op1 = TREE_OPERAND (node, 1);
> + tree op1type = TYPE_MAIN_VARIANT (TREE_TYPE (op1));
> +
> + if (!tree_int_cst_equal (TYPE_SIZE (nodetype),
> + TYPE_SIZE (TREE_TYPE (op1type))))
> + {
> + pp_string (pp, " <");
> + /* If the size of the type of the operand is not the same
> + as the size of the MEM_REF expression include the type
> + of the latter similar to the TDF_GIMPLE output to make
> + it clear how many bytes of memory are being accessed. */
> + dump_generic_node (pp, nodetype, spc, flags | TDF_SLIM, false);
> + pp_string (pp, "> ");
> + }
You need to guard against non-constant TYPE_SIZE here (for both
involved types) so I suggest you use operand_equal_p (..., 0). If you
do that you need to guard against NULL_TREE TYPE_SIZE
(tree_int_cst_equal handled that as unequal).
OK with such change.
Richard.
> Martin
