[Bug middle-end/67133] [6 Regression] ICE at -Os and above on x86_64-linux-gnu in gimple_op, at gimple.h:2274

[manu at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67133

Manuel López-Ibáñez <manu at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |law at gcc dot gnu.org

--- Comment #7 from Manuel López-Ibáñez <manu at gcc dot gnu.org> ---
(In reply to Zhendong Su from comment #0)
> The following code causes an ICE when compiled with the current gcc trunk at
> -Os and above on x86_64-linux-gnu in both 32-bit and 64-bit modes.
> 
> It is a regression from 5.1.x. 

This was already latent in r226635, just compile with options:

-Os -fisolate-erroneous-paths-dereference -fisolate-erroneous-paths-attribute

The problem is that infer_nonnull_range gets:

(gdb) p debug_gimple_stmt(stmt)
__builtin_unreachable ();

which is decorated with attributes:

(gdb) p debug_tree(attrs)
 <tree_list 0x7ffff64b3bb8
    purpose <identifier_node 0x7ffff644b3c0 nonnull>
    value <tree_list 0x7ffff645edc0
        value <integer_cst 0x7ffff6444060 constant 1>>
    chain <tree_list 0x7ffff64b3b90
        purpose <identifier_node 0x7ffff644b2d0 format>
        value <tree_list 0x7ffff6460730
            value <identifier_node 0x7ffff644b550 printf>
            chain <tree_list 0x7ffff645ef00 value <integer_cst 0x7ffff6444060
1>
                chain <tree_list 0x7ffff645ede8
                    value <integer_cst 0x7ffff6444318 constant 2>>>>>>

That is, this __builtin_unreachable has __attribute__((nonnull(1))), which
doesn't make any sense and breaks infer_nonnull_range when it tries to access
its first argument.

I have no idea whatsoever how __builtin_unreachable gets to have attribute
nonnull, but that is a pre-existing bug. Surprising it doesn't break more
stuff.