On Sun, Oct 18, 2015 at 7:14 PM, Jan Hubicka <hubi...@ucw.cz> wrote:
>>
>> Adding back the mode check is fine if all types with the same TYPE_CANONICAL
>> have the same mode. Otherwise we'd regress here. I thought we do for
>>
>> Struct x { int i; };
>> Typedef y x __attribute__((packed));
>>
>> And then doing
>>
>> X x;
>> Y y;
>> X = y;
>
> Do you have any idea how to turn this into a testcase? I don't think we could
> add packed attribute to typedef. Even in
> gimple_canonical_types_compatible_p
> /* Can't be the same type if they have different mode. */
> if (TYPE_MODE (t1) != TYPE_MODE (t2))
> return false;
> (which IMO may be wrong WRT -mavx flags where modes of same types may be
> different
> in different TUs)
Ok, so the following works:
struct x { int i; };
typedef struct x y __attribute__((aligned(1)));
void foo (void)
{
struct x X;
y Y;
X = Y;
}
but we use SImode for y as well even though it's alignment is just one byte ...
Not sure what happens on strict-align targets for this and not sure how this
cannot be _not_ a problem. Consider
void bar (struct x);
and
bar (Y);
or using y *Y and X = *Y or bar (*Y).
> Therefore I would say that TYPE_CANONICAL determine mode modulo the fact that
> incoplete variant of a complete type will have VOIDmode instead of complete
> type's mode (during non-LTO). That is why I allow mode changes for casts from
> complete to incomplete.
Incomplete have VOIDmode, right?
> In longer run I think that every query to useless_type_conversion_p that
> contains incomplete types is a confused query. useless_type_conversion_p is
> about operations on the value and there are no operations for incomplete type
> (and function types). I know that ipa-icf-gimple and the following code in
> gimplify-stmt checks this frequently:
> /* The FEs may end up building ADDR_EXPRs early on a decl with
> an incomplete type. Re-build ADDR_EXPRs in canonical form
> here. */
> if (!types_compatible_p (TREE_TYPE (op0), TREE_TYPE (TREE_TYPE (expr))))
> *expr_p = build_fold_addr_expr (op0);
> Taking address of incomplete type or functions, naturally, makes sense. We
> may
> want to check something else here, like simply
> TREE_TYPE (op0) != TREE_TYPE (TREE_TYPE (expr))
> and once ipa-icf is cleanded up start sanity checking in
> usless_type_conversion
> that we use it to force equality only on types that do have values.
>
> We also can trip it when checking TYPE_METHOD_BASETYPE which may be
> incomplete.
> This is in the code checking useless_type_conversion on functions that I think
> are confused querries anyway - we need the ABI matcher, I am looking into
> that.
Ok, so given we seem to be fine in practive with TYPE_MODE (type) ==
TYPE_MODE (TYPE_CANONICAL (type))
(whether that's a but or not ...) I'm fine with re-instantiating the
mode check for
aggregate types. Please do that with
Index: gcc/gimple-expr.c
===================================================================
--- gcc/gimple-expr.c (revision 228963)
+++ gcc/gimple-expr.c (working copy)
@@ -89,8 +89,7 @@ useless_type_conversion_p (tree outer_ty
/* Changes in machine mode are never useless conversions unless we
deal with aggregate types in which case we defer to later checks. */
- if (TYPE_MODE (inner_type) != TYPE_MODE (outer_type)
- && !AGGREGATE_TYPE_P (inner_type))
+ if (TYPE_MODE (inner_type) != TYPE_MODE (outer_type))
return false;
/* If both the inner and outer types are integral types, then the
Can we asses equal sizes when modes are non-BLKmode then? Thus
@@ -270,10 +269,9 @@ useless_type_conversion_p (tree outer_ty
use the types in move operations. */
else if (AGGREGATE_TYPE_P (inner_type)
&& TREE_CODE (inner_type) == TREE_CODE (outer_type))
- return (!TYPE_SIZE (outer_type)
- || (TYPE_SIZE (inner_type)
- && operand_equal_p (TYPE_SIZE (inner_type),
- TYPE_SIZE (outer_type), 0)));
+ return (TYPE_MODE (outer_type) != BLKmode
+ || operand_equal_p (TYPE_SIZE (inner_type),
+ TYPE_SIZE (outer_type), 0));
else if (TREE_CODE (inner_type) == OFFSET_TYPE
&& TREE_CODE (outer_type) == OFFSET_TYPE)
? Hoping for VOIDmode incomplete case.
Richard.
> Honza
>>
>> Richard.
>>
>>
>> >Honza
>> >>
>> >> --
>> >> Eric Botcazou
>>
