[Bug rtl-optimization/55719] [4.8 Regression] ICE: Segmentation fault

["krebbel at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55719

Andreas Krebbel <krebbel at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |ASSIGNED
         AssignedTo|unassigned at gcc dot       |krebbel at gcc dot gnu.org
                   |gnu.org                     |

--- Comment #6 from Andreas Krebbel <krebbel at gcc dot gnu.org> 2013-01-10 07:41:46 UTC ---
(In reply to comment #5)
> I'm out of my depth here, but perhaps someone can give me a hint on how to
> debug this further.
> 
> As alluded to earlier, this ICE only happens with sched-pressure scheduling.
> 
> The actual ICE happens in find_reloads_address_part(), where we have are
> analyzing X=:
> 
> (const:DI (plus:DI (symbol_ref:DI ("glob_b"))
>         (const_int 3 [0x3])))
> 
> prefer_reload_class() returns NO_REGS, because legitimate_reload_constant_p()
> returns false for the X above.  So when we call force_const_mem() in
> find_reloads_address_part...
> 
>   if (CONSTANT_P (x)
>       && (!targetm.legitimate_constant_p (mode, x)
>       || targetm.preferred_reload_class (x, rclass) == NO_REGS))
>     {
>       x = force_const_mem (mode, x);   <-- HERE
> 
> force_const_mem() returns NULL because s390_cannot_force_const_mem() returns
> true for the SYMBOL_REF in -fPIC mode.
> 
> Bottom line-- force_const_mem() returns NULL for a (const (plus SYMBOL_REF
> CONST)) with -fPIC, and then we dereference this NULL.
> 
> Is the RTL correct?
> 
> Any hints?

I'm already testing a patch. We have an instruction (larl) which is able to
handle pc-relative symbol references. But it only allows even addends. symref +
odd addend is handled via secondary reload. I think it is invalid that
preferred reload class returns NO_REGS here.