[Bug lto/88677] [9 Regression] Divergence in -O2 and -O2 -flto early opts

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88677

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |wrong-code
             Status|ASSIGNED                    |NEW
                 CC|                            |rguenth at gcc dot gnu.org
           Assignee|rguenth at gcc dot gnu.org         |hubicka at gcc dot gnu.org

--- Comment #8 from Richard Biener <rguenth at gcc dot gnu.org> ---
 Value numbering stmt = SR.10_27 = MEM[(struct D *)&m];
-Setting value number of SR.10_27 to SR.10_27 (changed)
+Setting value number of SR.10_27 to SR.9_23 (changed)

but then of course the code-difference

   <bb 7> :
-  # SR.10_12 = PHI <0(5), SR.10_27(6)>
   return;

is somewhat spurious since nothing uses SR.10_12, we're getting to
removal only because eliding SR.10_27 enables DCE.

  <bb 6> :
  SR.10_27 = MEM[(struct D *)&m];

  <bb 7> :
  # SR.10_12 = PHI <0(5), SR.10_27(6)>
  return;


So the code-gen difference is likely because of the following which makes
dropping TYPE_NEEDS_CONSTRUCTING a suspicious change.

/* Return true if VAR may be aliased.  A variable is considered as
   maybe aliased if it has its address taken by the local TU
   or possibly by another TU and might be modified through a pointer.  */

static inline bool
may_be_aliased (const_tree var)
{
  return (TREE_CODE (var) != CONST_DECL
          && (TREE_PUBLIC (var)
              || DECL_EXTERNAL (var)
              || TREE_ADDRESSABLE (var))
          && !((TREE_STATIC (var) || TREE_PUBLIC (var) || DECL_EXTERNAL (var))
               && ((TREE_READONLY (var)
                    && !TYPE_NEEDS_CONSTRUCTING (TREE_TYPE (var)))
                   || (TREE_CODE (var) == VAR_DECL
                       && DECL_NONALIASED (var)))));
}

the variable in question is

 <var_decl 0x7ffff7fefcf0 m
    type <record_type 0x7ffff69e6a80 D readonly needs-constructing type_1
type_5 SI
        size <integer_cst 0x7ffff689c078 constant 32>
        unit-size <integer_cst 0x7ffff689c090 constant 4>
        align:32 warn_if_not_align:0 symtab:0 alias-set -1 canonical-type
0x7ffff69e6a80
        fields <function_decl 0x7ffff69f5c00 __ct  type <method_type
0x7ffff69f3000>
            public abstract external autoinline decl_3 QI t.ii:2:29 align:16
warn_if_not_align:0 context <record_type 0x7ffff69e69d8 D>
            full-name "constexpr D<j>::D(D<j>&&) noexcept (<uninstantiated>)"
            not-really-extern chain <function_decl 0x7ffff69f5e00 __ct_base >>
context <translation_unit_decl 0x7ffff6887168 t.ii>
        full-name "const class D<j>"
        needs-constructor X(constX&) this=(X&) n_parents=1 use_template=1
interface-unknown
        pointer_to_this <pointer_type 0x7ffff69f30a8> reference_to_this
<reference_type 0x7ffff69f3150>>
    readonly addressable used public external read decl_2 VOID t.ii:17:19
    align:8 warn_if_not_align:0 context <translation_unit_decl 0x7ffff6887168
t.ii> chain <type_decl 0x7ffff699cc78 j>>

may_be_aliased tries to avoid claiming R/O data can be written to, but of
course that's not true for the global ctor.

-> You can't simply remove this flag.  You could set it to true conservatively.
Or we could stop marking globals that need constructing TREE_READONLY.