Re: [ping] Fix PR debug/66728

[Richard Sandiford <richard.sandiford@arm.com>]

Mike Stump <mikestump@comcast.net> writes:
> On Nov 2, 2015, at 8:29 AM, Richard Sandiford <richard.sandiford@arm.com> wrote:
>>   switch (GET_CODE (rtl))
>>     {
>>     case CONST_INT:
>> -      {
>> -	HOST_WIDE_INT val = INTVAL (rtl);
>> +      if (mode != BLKmode)
>
> This changes BLKmode for CONST_INT, but I didn’t see this discussed.  I
> didn’t see a test case?  I’d like to think that BLKmode things here
> would be fine.  I think they would be use for 1024 bit things that are
> representable in 20 bits, for example.

This was:

  ... Sometimes structure decls
  have BLKmode but are assigned an integer-mode rtl (e.g. when passing
  3-byte structures by value to functions).
  [...]
  loc_descriptor refuses to use CONST_INT for BLKmode decls (which aren't
  actually integers at the source level).  That seems like the right
  behaviour, so this patch does that for add_const_value_attribute too.
  It asserts that the mode is otherwise sensible for both CONST_INT
  and CONST_WIDE_INT.  Asserting for CONST_INT isn't strictly necessary
  but means that the assumption will get much more coverage than asserting
  only for CONST_WIDE_INT does.

Integer types always have an integer mode, not BLKmode.  E.g.
layout_type has:

    case BOOLEAN_TYPE:
    case INTEGER_TYPE:
    case ENUMERAL_TYPE:
      SET_TYPE_MODE (type,
		     smallest_mode_for_size (TYPE_PRECISION (type), MODE_INT));

where the smallest_mode_for_size is guaranteed to return something
that isn't BLKmode (or VOIDmode).

The BLKmodes are for non-integer things that nevertheless get passed
as integers.  I don't think debuggers would be expecting an integer
DIE to be used for them.  (loc_descriptor already punts in that case.)

> A value that is 1 (representable
> in 20 bits) can be trivially communicated the debugger.  The existing
> add_AT_unsigned I think can represent them, no?  Similarly for wide-int
> BLKmode support.  I think the real problem is simply the precision 0
> part.  In the CONST_INT and CONST_DOUBLE there is no code that handled
> precision 0, and there is no code in the wide-int case either.  From
> wide-int.h:
>
>   The precision and length of a wide_int are always greater than 0.
>
> If is was 0, then we have failed.  When that bug is fixed, then the
> precision won’t be 0 and the existing code will work.  Where is the 0
> first generated, and from what?

It's generated from:

    case CONST_WIDE_INT:
      add_AT_wide (die, DW_AT_const_value,
                  std::make_pair (rtl, GET_MODE (rtl)));
      return true;

where the precision is evaluated as:

  inline unsigned int
  wi::int_traits <rtx_mode_t>::get_precision (const rtx_mode_t &x)
  {
    return GET_MODE_PRECISION (x.second);
  }

GET_MODE (rtl) is always VOIDmode and the precision of VOIDmode is 0,
so the precision of the wide_int passed to add_AT_wide is always 0.

Like CONST_INT, the "real" mode of a CONST_WIDE_INT is determined by
context rather than being stored in the rtx itself.  The point of the
patch is to use that mode instead of VOIDmode in the make_pair.

Thanks,
Richard