[Bug c++/90383] [9/10 Regression] GCC generates invalid constexpr copy/move assignment operators for types with trailing padding. (Again)

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90383

--- Comment #5 from Jakub Jelinek  ---
Author: jakub
Date: Fri May 10 08:19:44 2019
New Revision: 271058

URL: https://gcc.gnu.org/viewcvs?rev=271058=gcc=rev
Log:
PR c++/90383
* tree-inline.h (struct copy_body_data): Add do_not_fold member.
* tree-inline.c (remap_gimple_op_r): Avoid folding expressions if
id->do_not_fold.
(copy_tree_body_r): Likewise.
(copy_fn): Set id.do_not_fold to true.

* g++.dg/cpp1y/constexpr-90383-1.C: New test.
* g++.dg/cpp1y/constexpr-90383-2.C: New test.

Added:
trunk/gcc/testsuite/g++.dg/cpp1y/constexpr-90383-1.C
trunk/gcc/testsuite/g++.dg/cpp1y/constexpr-90383-2.C
Modified:
trunk/gcc/ChangeLog
trunk/gcc/testsuite/ChangeLog
trunk/gcc/tree-inline.c
trunk/gcc/tree-inline.h

[Bug c++/90383] [9/10 Regression] GCC generates invalid constexpr copy/move assignment operators for types with trailing padding. (Again)

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90383

Jakub Jelinek  changed:

   What|Removed |Added

 Status|NEW |ASSIGNED
   Assignee|unassigned at gcc dot gnu.org  |jakub at gcc dot gnu.org

--- Comment #4 from Jakub Jelinek  ---
Created attachment 46321
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=46321=edit
gcc10-pr90383.patch

And the bug is that when copy_fn makes a copy of the function for constexpr
evaluation purposes, the inliner helper functions actually fold some
expressions in it, which confuses the constexpr inline evaluation.
This untested patch prevents that folding, which we really need just when
actually inlining or performing other middle-end optimizations.

[Bug c++/90383] [9/10 Regression] GCC generates invalid constexpr copy/move assignment operators for types with trailing padding. (Again)

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90383

--- Comment #3 from Jakub Jelinek  ---
struct alignas(8) A { constexpr A (bool x) : a(x) {} A () = delete; bool a; };
struct B { A b; };

constexpr bool
foo ()
{
  B w{A (true)};
  w.b = A (true);
  return w.b.a;
}

static_assert (foo (), "");

The problem is in that w.b = A (true); store, somehow we end up with w being a
CONSTRUCTOR with type B, containing just FIELD_DECL a rather than FIELD_DECL b
with another CONSTRUCTOR with type A, containing just FIELD_DECL a.

[Bug c++/90383] [9/10 Regression] GCC generates invalid constexpr copy/move assignment operators for types with trailing padding. (Again)

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90383

Jakub Jelinek  changed:

   What|Removed |Added

 Status|UNCONFIRMED |NEW
   Last reconfirmed||2019-05-08
 CC||jakub at gcc dot gnu.org
 Ever confirmed|0   |1

--- Comment #2 from Jakub Jelinek  ---
Started with my r269078 change (but it is only rejected with -O1 and higher, at
-O0 it is accepted).  Guess because always_true isn't evaluated in constexpr
context.

[Bug c++/90383] [9/10 Regression] GCC generates invalid constexpr copy/move assignment operators for types with trailing padding. (Again)

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90383

Richard Biener  changed:

   What|Removed |Added

   Target Milestone|--- |9.2
Summary|[9 Regression] GCC  |[9/10 Regression] GCC
   |generates invalid constexpr |generates invalid constexpr
   |copy/move assignment|copy/move assignment
   |operators for types with|operators for types with
   |trailing padding. (Again)   |trailing padding. (Again)