Re: cfg merge part 19 - gcse tweek

[Jan Hubicka <jh at suse dot cz>]

> > 
> > The problem is that gcse_emit_move_after, called from pre_delete creates an
> > instruction:
> > 
> > 
> > (insn 858 467 468 25 (nil) (set (reg/v:SF 210)
> >         (reg:SF 258)) -1 (nil)
> >     (nil))
> > 
> > If you are trying to generate a move insn that 'just works', I suggest you look
> > at
> > gen_move_insn.
> 
> Hmm, I sitll keep it reading that way.  want_to_gcse does:
> 
> 
>   /* Otherwise, check if we can make a valid insn from it.  First initialize
>      our test insn if we haven't already.  */
>   if (test_insn == 0)
>     {
>       test_insn
> 	= make_insn_raw (gen_rtx_SET (VOIDmode,
> 				      gen_rtx_REG (word_mode,
> 						   FIRST_PSEUDO_REGISTER * 2),
> 				      const0_rtx));
>       NEXT_INSN (test_insn) = PREV_INSN (test_insn) = 0;
>     }
> 
>   /* Now make an insn like the one we would make when GCSE'ing and see if
>      valid.  */
>   PUT_MODE (SET_DEST (PATTERN (test_insn)), GET_MODE (x));
>   SET_SRC (PATTERN (test_insn)) = x;
>   return ((icode = recog (PATTERN (test_insn), test_insn, &num_clobbers)) >= 0
> 	  && (num_clobbers == 0 || ! added_clobbers_hard_reg_p (icode)));
> 
> So basically it test that (set (destination) (register)) is a valid
> instruction and that is exactly what I emit.  What did I missed?
> What I am affraid of is the fact that the dest don't have to be
> gneeral_operand, so passing it into gen_move is de-facto invalid.

I see it now.  The code is allowing pseudo register clobbers to be
added.  I guess proper fix is the to teach emit_gcse_move to add the
clobbers as well, like apply_change_group does? Makes this sense?

Honza
> 
> Honza