Patchwork Fix dead_debug_insert_before ICE (PR debug/49522)

login
register
mail settings
Submitter Jakub Jelinek
Date July 5, 2011, 8:06 p.m.
Message ID <20110705200651.GG2687@tyan-ft48-01.lab.bos.redhat.com>
Download mbox | patch
Permalink /patch/103381/
State New
Headers show

Comments

Jakub Jelinek - July 5, 2011, 8:06 p.m.
On Tue, Jul 05, 2011 at 10:35:11AM +0200, Eric Botcazou wrote:
> > There are two kinds of changes we do on the debug insns without immediate
> > rescanning:
> > 1) reset the debug insn
> > 2) replace a reg use with DEBUG_EXPR of the same mode or
> >    subreg of a larger DEBUG_EXPR with the same outer mode as the reg
> >
> > In the attached testcase on arm a debug insn is reset, because a multi-reg
> > register has been used there and as the debug insn location was that
> > multi-reg register before, it is now VOIDmode after the reset - (clobber
> > (const_int 0)).
> 
> That can happen only in this case, right?  Otherwise, for a single register, 
> the debug insn would have been removed from debug->head already.  If so, how 
> simpler would it be to remove the other uses in dead_debug_reset instead?

So you prefer something like this (untested) instead?
Without the second loop I have no idea how to make it work in
dead_debug_reset, the other dead_debug_use referencing the same insn might
be earlier or later in the chain.



	Jakub
Eric Botcazou - July 6, 2011, 8:31 p.m.
> So you prefer something like this (untested) instead?

I think that, ideally, we should avoid leaving the dead_debug chain in the 
semi-broken state that we currently have.

> Without the second loop I have no idea how to make it work in
> dead_debug_reset, the other dead_debug_use referencing the same insn might
> be earlier or later in the chain.

I guess I was somehow hoping that you could use one of the numerous DF links to 
get to the other uses; probably not, in the end, indeed.  But you can set a 
flag in the first loop in order to decide whether to run the second loop.

But I don't really have a strong opinon so, if you think that the original 
patch is good enough, fine with me.  Maybe use gcc_checking_assert though.

Patch

--- gcc/df-problems.c	2011-07-04 19:17:50.757435754 +0200
+++ gcc/df-problems.c	2011-07-05 22:04:19.817464710 +0200
@@ -3117,6 +3117,25 @@  dead_debug_reset (struct dead_debug *deb
       else
 	tailp = &(*tailp)->next;
     }
+
+  /* If any other dead_debug_use structs refer to the debug insns
+     that have been reset above, remove them too.  */
+  if (debug->to_rescan != NULL)
+    {
+      tailp = &debug->head;
+      while ((cur = *tailp))
+	{
+	  insn = DF_REF_INSN (cur->use);
+	  if (bitmap_bit_p (debug->to_rescan, INSN_UID (insn))
+	      && VAR_LOC_UNKNOWN_P (INSN_VAR_LOCATION_LOC (insn)))
+	    {
+	      *tailp = cur->next;
+	      XDELETE (cur);
+	    }
+	  else
+	    tailp = &(*tailp)->next;
+	}
+    }
 }
 
 /* Add USE to DEBUG.  It must be a dead reference to UREGNO in a debug