IRA fix for 47992

IRA fix for 47992

[Bernd Schmidt]

[-- Attachment #1: Type: text/plain, Size: 789 bytes --]

This is a report of a crash in IRA. If you debug it with a sufficiently 
old compiler, you'll find that we manage to delete some basic blocks 
from within IRA. Later on, reload calls alter_reg for all unallocated 
pseudos, including one that only occurs in the deleted blocks. reload 
does not notice it is unused, because REG_N_REFS is still 3. We crash in 
an IRA callback because the pseudo has no allocno.

Fixed as below. A similar patch cures the problem in gcc-4.6. Adding a 
testcase seems pointless - the crashing code was derived from an 
existing Fortran testcase with exotic options, and the whole thing only 
ever triggered in one gcc version AFAICT. Current gcc can't trigger it 
because it's not using reload on x86_64.

Bootstrapped and tested on x86_64-linux, ok?


Bernd

[-- Attachment #2: 47992.diff --]
[-- Type: text/x-patch, Size: 1228 bytes --]

	PR rtl-optimization/47992
	* ira.c (ira): Update regstat data if we deleted insns.

Index: gcc/ira.c
===================================================================
--- gcc/ira.c	(revision 232339)
+++ gcc/ira.c	(working copy)
@@ -5185,19 +5185,27 @@ ira (FILE *f)
   setup_reg_equiv ();
   setup_reg_equiv_init ();
 
+  bool update_regstat = false;
+
   if (optimize && rebuild_p)
     {
       timevar_push (TV_JUMP);
       rebuild_jump_labels (get_insns ());
       if (purge_all_dead_edges ())
-	delete_unreachable_blocks ();
+	{
+	  delete_unreachable_blocks ();
+	  update_regstat = true;
+	}
       timevar_pop (TV_JUMP);
     }
 
   allocated_reg_info_size = max_reg_num ();
 
   if (delete_trivially_dead_insns (get_insns (), max_reg_num ()))
-    df_analyze ();
+    {
+      df_analyze ();
+      update_regstat = true;
+    }
 
   /* It is not worth to do such improvement when we use a simple
      allocation because of -O0 usage or because the function is too
@@ -5308,7 +5316,7 @@ ira (FILE *f)
     check_allocation ();
 #endif
 
-  if (max_regno != max_regno_before_ira)
+  if (update_regstat || max_regno != max_regno_before_ira)
     {
       regstat_free_n_sets_and_refs ();
       regstat_free_ri ();

Re: IRA fix for 47992

[Jeff Law]

On 01/15/2016 06:42 AM, Bernd Schmidt wrote:
> This is a report of a crash in IRA. If you debug it with a sufficiently
> old compiler, you'll find that we manage to delete some basic blocks
> from within IRA. Later on, reload calls alter_reg for all unallocated
> pseudos, including one that only occurs in the deleted blocks. reload
> does not notice it is unused, because REG_N_REFS is still 3. We crash in
> an IRA callback because the pseudo has no allocno.
>
> Fixed as below. A similar patch cures the problem in gcc-4.6. Adding a
> testcase seems pointless - the crashing code was derived from an
> existing Fortran testcase with exotic options, and the whole thing only
> ever triggered in one gcc version AFAICT. Current gcc can't trigger it
> because it's not using reload on x86_64.
>
> Bootstrapped and tested on x86_64-linux, ok?
OK.
jeff