Incorrect bitfield aliasing with Tree SSA

[Adam Nemet]

Daniel Berlin writes:
> On 6/15/07, Adam Nemet <anemet@caviumnetworks.com> wrote:
> > get_alias_set and internally record_component_aliases makes
> > assumptions about the IR that are only valid in RTL.
> 
> What is this assumption, exactly?

That non-addressable fields are always accessed through alias set 0.
For example:

  X: (set (mem A) ...)
  Y: (set (zero_extract (mem B) 2 2) ...)

Let's say that A and B both point to the same type and have the same
alias set 1.  The bitfield store in Y would normally have the alias
set of the field, let's say 2.  Normally 2 would have to be recorded
as an alias subset of 1.  This is not done in RTL and the comment
before record_component_aliases mentions this:

/* Record that component types of TYPE, if any, are part of that type for
   aliasing purposes.  For record types, we only record component types
   for fields that are marked addressable.

If I am not mistaken this is because the functions expanding RTL
ensure that the alias set of the mem operand in Y has alias set 0.  Or
to quote expmed.c:

  /* We may be accessing data outside the field, which means
     we can alias adjacent data.  */
  if (MEM_P (op0))
    {
      op0 = shallow_copy_rtx (op0);
      set_mem_alias_set (op0, 0);
      set_mem_expr (op0, 0);
    }

Adam