Re: [PATCH] Merge adjacent stores of constants (PR middle-end/22141)

["Steven Bosscher" <stevenb dot gcc at gmail dot com>]

Xf. http://gcc.gnu.org/ml/gcc-patches/2008-11/msg01413.html

> > I wonder
> > if I could trust REG_EQUAL notes containing CONST_INT that forwprop
> > generates (and in which case to trust them?  Only if the insn is
> > single_set?).
>
> Yes, you can trust them if the insn is single_set.

Yup.  In fact...

>  If it is not, they
> probably do not come from fwprop (and I never really understood when are
> they valid for non-single_set insns).

...and any place where they do come from is buggy, because this is not valid..

Any REG_EQUAL note on a non-single_set() insn is a bug. From
emit-rtl.c:set_unique_reg_note():

rtx
set_unique_reg_note (rtx insn, enum reg_note kind, rtx datum)
{
  rtx note = find_reg_note (insn, kind, NULL_RTX);

  switch (kind)
    {
    case REG_EQUAL:
    case REG_EQUIV:
      /* Don't add REG_EQUAL/REG_EQUIV notes if the insn
	 has multiple sets (some callers assume single_set
	 means the insn only has one set, when in fact it
	 means the insn only has one * useful * set).  */
      if (GET_CODE (PATTERN (insn)) == PARALLEL && multiple_sets (insn))
	{
	  gcc_assert (!note);
	  return NULL_RTX;
	}
...


Indeed, it is a bug if we set REG_EQUAL notes from anywhere without
going through set_unique_reg_note().

Gr.
Steven