On Mon, Oct 23, 2017 at 10:04 AM, Richard Biener <rguent...@suse.de> wrote:
> On Mon, 23 Oct 2017, Jakub Jelinek wrote:
>
>> On Mon, Oct 23, 2017 at 09:48:50AM +0200, Richard Biener wrote:
>> > > --- gcc/targhooks.c.jj 2017-10-13 19:02:08.000000000 +0200
>> > > +++ gcc/targhooks.c 2017-10-20 14:26:07.945464025 +0200
>> > > @@ -177,6 +177,14 @@ default_legitimize_address_displacement
>> > > return false;
>> > > }
>> > >
>> > > +bool
>> > > +default_const_not_ok_for_debug_p (rtx x)
>> > > +{
>> > > + if (GET_CODE (x) == UNSPEC)
>> >
>> > What about UNSPEC_VOLATILE?
>>
>> This hook is called on the argument of CONST or SYMBOL_REF.
>> UNSPEC_VOLATILE can't appear inside of CONST, it wouldn't be CONST then.
>>
>> UNSPEC appearing outside of CONST is rejected unconditionally in
>> mem_loc_descriptor:
>> ...
>> case UNSPEC:
>> ...
>> /* If delegitimize_address couldn't do anything with the UNSPEC, we
>> can't express it in the debug info. This can happen e.g. with some
>> TLS UNSPECs. */
>> break;
>> and for UNSPEC_VOLATILE we just ICE, because var-tracking shouldn't let
>> those through:
>> default:
>> if (flag_checking)
>> {
>> print_rtl (stderr, rtl);
>> gcc_unreachable ();
>> }
>> break;
>
> Ok. The patch looks fine from a middle-end point of view.
LGTM for the x86 part.
Thanks,
Uros.
