combine: Don't record for UNDO_MODE pointers into regno_reg_rtx array [PR104985]

The testcase in the PR fails under valgrind on mips64 (but only Martin
can reproduce, I couldn't).
But the problem reported there is that SUBST_MODE remembers addresses
into the regno_reg_rtx array, then some splitter needs a new pseudo
and calls gen_reg_rtx, which reallocates the regno_reg_rtx array
and finally undo operation is done and dereferences the old regno_reg_rtx
entry.
The rtx values stored in regno_reg_rtx array seems to be created
by gen_reg_rtx only and since then aren't modified, all we do for it
is adjusting its fields (e.g. adjust_reg_mode that SUBST_MODE does).

So, I think it is useless to use where.r for UNDO_MODE and store
&regno_reg_rtx[regno] in struct undo, we can store just
regno_reg_rtx[regno] (i.e. pointer to the REG itself instead of
pointer to pointer to REG) or could also store just the regno.

The following patch does the latter, and because SUBST_MODE no longer
needs to be a macro, changes all SUBST_MODE uses to subst_mode.

2022-04-06  Jakub Jelinek  <jakub@redhat.com>

	PR rtl-optimization/104985
	* combine.cc (struct undo): Add where.regno member.
	(do_SUBST_MODE): Rename to ...
	(subst_mode): ... this.  Change first argument from rtx * into int,
	operate on regno_reg_rtx[regno] and save regno into where.regno.
	(SUBST_MODE): Remove.
	(try_combine): Use subst_mode instead of SUBST_MODE, change first
	argument from regno_reg_rtx[whatever] to whatever.  For UNDO_MODE, use
	regno_reg_rtx[undo->where.regno] instead of *undo->where.r.
	(undo_to_marker): For UNDO_MODE, use regno_reg_rtx[undo->where.regno]
	instead of *undo->where.r.
	(simplify_set): Use subst_mode instead of SUBST_MODE, change first
	argument from regno_reg_rtx[whatever] to whatever.