https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65193
--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
I already get
> ./xg++ -B. -nostdlib -fPIC -flto -shared t.ii -O2 -g
t.ii: In function ‘float_next_imp’:
t.ii:33:7: error: address taken, but ADDRESSABLE bit not set
float_next_imp (T p1, Policy p2)
^
d
t.ii:38:4: note: in statement
frexp (p1, &d);
^
# .MEM_13 = VDEF <.MEM_11>
frexp (p1_5(D), &d);
t.ii:33:7: internal compiler error: verify_gimple failed
float_next_imp (T p1, Policy p2)
^
0xaf39de verify_gimple_in_cfg(function*)
Ok. It is the execute_update_addresses_taken () call in the TODO of the
IPA inline transform pass that re-sets TREE_ADDRESSABLE here. Somehow
we have merged 'd' from two different places it seems (the optimization
is in _GLOBAL__sub_I_t.ii).
The duplicate appears in _GLOBAL__sub_I_t.ii after inlining:
{ Scope block #21 t.ii:13 Originating from : static long doubleD.2801
get_min_shift_valueD.2829 (void);
{ Scope block #23 Originating from :#0
voidD.2802 <<< Unknown tree: imported_decl >>> = ldexpD.2839;
intD.2788 dD.2905; (nonlocalized)
...
_GLOBAL__sub_I_t.ii ()
{
...
}
...
float_next_imp (intD.2788 p1D.2902, struct AD.2830 p2D.2903)
{
struct AD.2830 D.2919;
intD.2788 D.2918;
const intD.2833 ISRA.9D.2915;
intD.2788 dD.2905;
const intD.2833 D.2904;
something is messed up. The decl is streamed via streaming struct function
local-decls from float_next_imp (as expected). No idea how it ends up in
the vars for _GLOBAL__sub_I_t.ii.
Ah. We inline
(gdb) p debug_gimple_stmt (stmt)
# .MEM = VDEF <.MEM>
get_min_shift_value ();
$91 = void
but somehow we end up with
(gdb) p debug_tree (old_var)
<var_decl 0x7ffff683c8e8 d
type <integer_type 0x7ffff66b2690 int asm_written public SI
size <integer_cst 0x7ffff66b53a0 constant 32>
unit size <integer_cst 0x7ffff66b53c0 constant 4>
align 32 symtab -159219552 alias set -1 canonical type 0x7ffff66b2690
precision 32 min <integer_cst 0x7ffff66b5340 -2147483648> max <integer_cst
0x7ffff66b5360 2147483647>
pointer_to_this <pointer_type 0x7ffff66bd738>>
addressable SI file t.ii line 36 col 8 size <integer_cst 0x7ffff66b53a0 32>
unit size <integer_cst 0x7ffff66b53c0 4>
align 32 context <function_decl 0x7ffff6821b00 float_next_imp>>
in its BLOCK tree. Which gets populated by
Old value = <tree 0x0>
New value = <block 0x7ffff6828fa0>
0x0000000000cd9194 in lto_input_ts_block_tree_pointers (ib=0x7fffffffd610,
data_in=0x1bee6a0, expr=<block 0x7ffff6828fa0>)
at /space/rguenther/src/svn/gcc-4_9-branch/gcc/tree-streamer-in.c:928
928 BLOCK_SUBBLOCKS (BLOCK_SUPERCONTEXT (expr)) = expr;
but finally the valid block tree
(gdb) p debug_tree (expr)
<block 0x7ffff6828fa0
vars <imported_decl 0x7ffff681e8a0 D.2856
type <void_type 0x7ffff66bd000 void asm_written VOID
align 8 symtab 0 alias set -1 canonical type 0x7ffff66bd000
pointer_to_this <pointer_type 0x7ffff66bd0a8>>
VOID file t.ii line 39 col 2
align 1 initial <function_decl 0x7ffff682a100 ldexp>
chain <var_decl 0x7ffff66abd10 c type <real_type 0x7ffff66bd498 long
double>
public static weak XF file t.ii line 28 col 13
size <integer_cst 0x7ffff66b50a0 constant 128>
unit size <integer_cst 0x7ffff66b50c0 constant 16>
align 128 context <function_decl 0x7ffff6821a00
get_min_shift_value>
gets invalid by
Hardware watchpoint 16: *$111
Old value = <var_decl 0x7ffff66abd10 c>
New value = <var_decl 0x7ffff683c8e8 d>
0x0000000000cd579b in streamer_read_chain (ib=0x7fffffffd990,
data_in=0x1bee740)
at /space/rguenther/src/svn/gcc-4_9-branch/gcc/tree-streamer-in.c:85
85 TREE_CHAIN (prev) = curr;
where 'prev' is that IMPORTED_DECL which is streamed globally it seems (ups).
And we miss r212992 on the branch.
