https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99281
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Keywords| |ice-on-valid-code
Ever confirmed|0 |1
Status|UNCONFIRMED |NEW
Last reconfirmed| |2021-02-26
--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
Confirmed. We're building a temporary of type X which we shouldn't do when
expanding
get<int> (&D.2164) [return slot optimization]
and we run into
#2 0x0000000000fd165a in expand_call (exp=<call_expr 0x7ffff66ed818>,
target=0x0, ignore=0) at /home/rguenther/src/gcc3/gcc/calls.c:3878
3873 else
3874 {
3875 /* For variable-sized objects, we must be called with a
target
3876 specified. If we were to allocate space on the stack
here,
3877 we would have no way of knowing when to free it. */
3878 rtx d = assign_temp (rettype, 1, 1);
3879 structure_value_addr = XEXP (d, 0);
3880 target = 0;
because target is NULL when we come via
#7 0x0000000001197afe in expand_assignment (to=<component_ref 0x7ffff66d9390>,
from=<call_expr 0x7ffff66ed818>,
nontemporal=false) at /home/rguenther/src/gcc3/gcc/expr.c:5482
5481 else
5482 result = store_field (to_rtx, bitsize, bitpos,
5483 bitregion_start, bitregion_end,
5484 mode1, from, get_alias_set (to),
5485 nontemporal, reversep);
and the large conditional on
/* If the structure is in a register or if the component
is a bit field, we cannot use addressing to access it.
Use bit-field techniques or SUBREG to store in it. */
somehow misfires on this, specifically there is
/* And except for bitwise copying of TREE_ADDRESSABLE types,
where the FIELD_DECL has the right bitsize, but TREE_TYPE (exp)
includes some extra padding. store_expr / expand_expr will in
that case call get_inner_reference that will have the bitsize
we check here and thus the block move will not clobber the
padding that shouldn't be clobbered. In the future we could
replace the TREE_ADDRESSABLE check with a check that
get_base_address needs to live in memory. */
&& (!TREE_ADDRESSABLE (TREE_TYPE (exp))
|| TREE_CODE (exp) != COMPONENT_REF
|| !multiple_p (bitsize, BITS_PER_UNIT)
|| !multiple_p (bitpos, BITS_PER_UNIT)
|| !poly_int_tree_p (DECL_SIZE (TREE_OPERAND (exp, 1)),
&decl_bitsize)
|| maybe_ne (decl_bitsize, bitsize)))
maybe sth as simple as the following works (instead of for return-slot-opt
we can also test !TREE_ADDRESSABLE here or boht for the testcase)
diff --git a/gcc/expr.c b/gcc/expr.c
index 86dc1b6c973..54007ac0efe 100644
--- a/gcc/expr.c
+++ b/gcc/expr.c
@@ -7214,7 +7214,10 @@ store_field (rtx target, poly_int64 bitsize, poly_int64
bitpos,
|| !multiple_p (bitpos, BITS_PER_UNIT)
|| !poly_int_tree_p (DECL_SIZE (TREE_OPERAND (exp, 1)),
&decl_bitsize)
- || maybe_ne (decl_bitsize, bitsize)))
+ || maybe_ne (decl_bitsize, bitsize))
+ /* A call with return-slot-opt must not need bitfield operations. */
+ && (TREE_CODE (exp) != CALL_EXPR
+ || !CALL_EXPR_RETURN_SLOT_OPT (exp)))
/* If we are expanding a MEM_REF of a non-BLKmode non-addressable
decl we must use bitfield operations. */
|| (known_size_p (bitsize)
note clang warns
t.ii:27:5: warning: expression result unused [-Wunused-value]
pack_type<X>{get(1)};
^ ~~~~~~~~
