https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126194
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |jakub at gcc dot gnu.org,
| |jason at gcc dot gnu.org,
| |rguenth at gcc dot gnu.org
--- Comment #7 from Richard Biener <rguenth at gcc dot gnu.org> ---
The bisected rev. does not change the underlying logic, it just exposes more
paths to DSE, so if the issue is in DSE it was latent before. I assume
we are talking about struct A D.5736
- D.5736.f_ = cb;
- D.5736.p_ = &b.q;
- D.5736.a_ = 8;
in make_widget.isra since the DSE of b.q = 0B is obvious from the IL:
- b.q = 0B;
b ={v} {CLOBBER(eob)};
b ={v} {CLOBBER(eos)};
--
<L0>:
# .MEM_17 = VDEF <.MEM_15>
# USE = nonlocal null { D.5581 D.5591 D.5794 D.5795 } (nonlocal, escaped,
escaped heap, interposable)
# CLB = nonlocal null { D.5581 D.5591 D.5794 D.5795 } (nonlocal, escaped,
escaped heap, interposable)
operator delete (_33, &D.5736, &g_r.D.5569);
does not use D.5736 and the object does not escape either, but we
definitely call this operator and the operator does read from D.5736:
__attribute__((noinline))
void operator delete (void * p, const struct A & a, struct R & r)
{
# PT = nonlocal
void * p_5(D) = p;
# PT = nonlocal
# ALIGN = 8, MISALIGN = 0
const struct A & a_4(D) = a;
# PT = nonlocal
struct R & r_6(D) = r;
void * _3;
void * * _10;
<bb 2> [local count: 1073741824]:
# PT = nonlocal null
# USE = nonlocal null
_3 = __builtin_return_address (0);
# PT = nonlocal escaped null
_10 = MEM[(void * * *)a_4(D) + 8B];
# USE = nonlocal escaped null
# CLB = nonlocal escaped null
destroy.isra (_10, p_5(D), _3, r_6(D)); [tail call]
PTA does
/* If the call is to a replaceable operator delete and results
from a delete expression as opposed to a direct call to
such operator, then the effects for PTA (in particular
the escaping of the pointer) can be ignored. */
else if (fndecl
&& DECL_IS_OPERATOR_DELETE_P (fndecl)
&& gimple_call_from_new_or_delete (t))
;
and completely ignores the delete operator.
Adding && DECL_IS_REPLACEABLE_OPERATOR (fndecl) fixes the testcase.
r11-3612-g4f4ced28826ece replaced the former use of
DECL_IS_REPLACEABLE_OPERATOR_DELETE_P with DECL_IS_OPERATOR_DELETE_P
(throughout the code base). Jason, what's this about? I suppose
DECL_IS_REPLACEABLE_OPERATOR will just never be set on delete?
