TYPE_PTRMEMFUNC_P

[Zdenek Dvorak <rakdver at atrey dot karlin dot mff dot cuni dot cz>]

Hello,

I am toying with lowering address arithmetics in late phases of tree-ssa
on lno-branch, and I have encountered the following problem:

I have this testcase:

struct A { };

int func(A *oa, int (A::*pmf)(int))
{ return (oa->*pmf)(2); }

Which translates to (irrelevant parts stripped):

int func(A*, int (A::*)(int)) (oa, pmf)
{
  A:: * pmf$__pfn;

  pmf$__pfn_25 = pmf.__pfn;

  ...
}

Address arithmetics lowering alters this to:

  A:: * pmf$__pfn;
  A:: * * ruatmp.18;
  A:: * * fgotmp.17;

  fgotmp.17_30 = (A:: * *)&pmf;
  ruatmp.18_31 = fgotmp.17_30;
  pmf$__pfn_25 = *ruatmp.18_31;

Which is later (wrongly) simplified to

  pmf$__pfn_25 = pmf;

Which later crashes, since type of pmf is

 struct
  {
    A:: * __pfn;
    long int __delta;
  }

The problem seems to be that cp/typeck.c:comptypes considers
A:: * (whatever this type means) and the above type equivalent,
since TYPE_PTRMEMFUNC_P is true for the above type and it therefore
compares TYPE_PTRMEMFUNC_FN_TYPE of it instead, which happens to be
A:: *.

Since I do not quite understand what TYPE_PTRMEMFUNC_P and related
stuff means, I don't know where to fix the problem.  Could someone help,
please?

Zdenek