[PATCH] inline-asm: Fix ICE with bitfields in "m" operands [PR100785]

Jason Merrill jason@redhat.com
Wed Jun 2 16:14:51 GMT 2021

On 6/2/21 11:25 AM, Jakub Jelinek wrote:
> On Wed, Jun 02, 2021 at 11:09:45AM -0400, Jason Merrill wrote:
>> On 6/2/21 3:59 AM, Jakub Jelinek wrote:
>>>    	      if (!allows_reg && !cxx_mark_addressable (*op))
>>>    		operand = error_mark_node;
>>> +	      else if (!allows_reg && bitfield_p (*op))
>>> +		{
>>> +		  error_at (loc, "attempt to take address of bit-field");
>> Hmm, why aren't we already catching this in cxx_mark_addressable?
> That is certainly possible, but it goes against Eric's patch
> https://gcc.gnu.org/pipermail/gcc-patches/2015-June/421483.html

Hmm, I wonder what his rationale was?

> So, do you want to keep the
>    if (bitfield_p (arg))
>      {
>        if (complain & tf_error)
>          error_at (loc, "attempt to take address of bit-field");
>        return error_mark_node;
>      }
> in cp_build_addr_expr_1 and duplicate such check in cxx_mark_addressable
> (though, that one doesn't have complain, will it be ok to do it
> unconditionally for SFINAE)?

We would need to add complain; the existing diagnostics can't happen in 
SFINAE context, but this one can.

Alternately, cxx_mark_addressable could abort on a bitfield to require 
the caller to handle it, but that seems less useful.


> Shall the C FE do the same (i.e. diagnose in both places)?

More information about the Gcc-patches mailing list