Re: patch for flow.c: flow_delete_insn

[graham <grahams at rcp dot co dot uk>]

Denis

Richard Henderson wrote:
> 
> On Tue, May 09, 2000 at 08:35:30PM +0400, Denis Chertykov wrote:
> >       * flow.c (flow_delete_insn): Use INSN_DELETED_P for marking insn
> >       as deleted.
> 
> Ok.
> 
> r~

Have you bootstrapped with this changed?

The reason I ask is that with this patch applied
a bootstrap on i686-linux fails due to ICE.

The ICE occurs in add_insn_after () because it is 
asked to add a insn after a insn which has been 
marked as deleted.

The problem is flow uses flow_delete_insn ()
to remove a NOTE insn and then later add it 
back.

Here's the relevant code from flow which is using
flow_delete_insn.

	    /* Look for basic block notes with which to keep the 
	       basic_block_info pointers stable.  Unthread the note now;
	       we'll put it back at the right place in create_basic_block.
	       Or not at all if we've already found a note in this block.  */
	    else if (kind == NOTE_INSN_BASIC_BLOCK)
	      {
		if (bb_note == NULL_RTX)
		   bb_note = insn;

	        next = flow_delete_insn (insn);
	      }



I have a work around which is to reset the DELETED
flag after calling flow_delte_insn () as follows.

	    /* Look for basic block notes with which to keep the 
	       basic_block_info pointers stable.  Unthread the note now;
	       we'll put it back at the right place in create_basic_block.
	       Or not at all if we've already found a note in this block.  */
	    else if (kind == NOTE_INSN_BASIC_BLOCK)
	      {
	        next = flow_delete_insn (insn);

		if (bb_note == NULL_RTX)
		  {
		    bb_note = insn;
		    INSN_DELETED_P (bb_note) = 0;
		  }
	      }


I don't like this it's functional but it's a hack.

Richard do you have a better solution ?

Graham