[Bug ipa/68220] [5/6 Regression] Devirtualization ICE in record_target_from_binfo, at ipa-devirt.c:2389

[hubicka at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68220

--- Comment #4 from Jan Hubicka <hubicka at gcc dot gnu.org> ---
OK, we have:

(gdb) p context
$4 = {offset = 0, speculative_offset = 96, outer_type = 0x7ffff6c77c78,
speculative_outer_type = 0x7ffff6cd5348, maybe_in_construction = 1,
maybe_derived_type = 1, 
  speculative_maybe_derived_type = 1, invalid = 0, dynamic = 1}

and the code gets lost in analyzing speculative_outer_type, because there is
simply no binfo on the address:

 <record_type 0x7ffff6cd5348 Fragment addressable tree_2 needs-constructing
type_4 type_5 type_6 BLK
    size <integer_cst 0x7ffff6ad7e58 type <integer_type 0x7ffff6adb2a0
bitsizetype> constant 64>
    unit size <integer_cst 0x7ffff6ad7e70 type <integer_type 0x7ffff6adb1f8
sizetype> constant 8>
    align 64 symtab 0 alias set 3 canonical type 0x7ffff6cd5348
    fields <field_decl 0x7ffff6ccbed8 _vptr.Fragment
        type <pointer_type 0x7ffff6c34540 type <pointer_type 0x7ffff6c343f0
__vtbl_ptr_type>
            public unsigned DI size <integer_cst 0x7ffff6ad7e58 64> unit size
<integer_cst 0x7ffff6ad7e70 8>
            align 64 symtab 0 alias set 4 canonical type 0x7ffff6c34540>
        unsigned virtual DI file /aux/hubicka/t.C line 225 col 8 size
<integer_cst 0x7ffff6ad7e58 64> unit size <integer_cst 0x7ffff6ad7e70 8>
        align 64 offset_align 128
        offset <integer_cst 0x7ffff6ad7e88 constant 0>
        bit offset <integer_cst 0x7ffff6ad7ed0 constant 0> context <record_type
0x7ffff6cd5348 Fragment>
        chain <type_decl 0x7ffff6ccbe40 Fragment type <record_type
0x7ffff6cd53f0 Fragment>
            used nonlocal decl_4 VOID file /aux/hubicka/t.C line 225 col 17
            align 1 context <record_type 0x7ffff6cd5348 Fragment> result
<record_type 0x7ffff6cd5348 Fragment>
           >> context <translation_unit_decl 0x7ffff7ff81e0 D.1>
    full-name "struct Fragment"
    needs-constructor needs-destructor X() X(constX&) this=(X&) n_parents=0
use_template=0 interface-only
    pointer_to_this <pointer_type 0x7ffff6cd5930> reference_to_this
<reference_type 0x7ffff6cd5d20> chain <type_decl 0x7ffff6ccbda8 Fragment>>

Well, the type is simply too small. The assert is simply too overactive, but
restrict_to_inner_type could probably drop the info when it knows the offset is
off the range.