Re: [patch, ice, selective scheduler] Patch for PR 41697

[Andrey Belevantsev <abel at ispras dot ru>]

Hello,

Andrey Belevantsev wrote:
> Andrey Belevantsev wrote:
> > Hi Steve,
> >
> > Steve Ellcey wrote:
> > > This patch fixes an ICE on IA64 with the testcase
> > > gcc.c-torture/compile/20090917-1.c.  That test will generate an ICE in
> > > the selective scheduler because we try to access the successor to a
> > > basic block before verifying that the basic block has any successors.
> > > This patch adds a check and fixes the ICE.

I have missed the second instance of this issue that can be seen when 
using selective scheduler on the above test with -O2.  The problem is 
that a block ending on a conditional jump has a single successor, thus 
using FALLTHRU_EDGE on it like in fallthru_bb_of_jump crashes.  The 
following simple patch avoids this.  Is it ok for trunk?  The patch is 
safe, and the bug technically is probably a regression, though I don't 
have a 4.4.x build on ia64 to back up this.

Bootstrapped and tested on ia64 with sel-sched enabled at -O2.

Andrey

2009-11-12  Andrey Belevantsev  <abel@ispras.ru>

	PR rtl-optimization/41697
	* sel-sched-ir.c (fallthru_bb_of_jump): Bail out when a block with a 
conditional jump has a single successor.

Index: gcc/sel-sched-ir.c
===================================================================
*** gcc/sel-sched-ir.c  (revision 154088)
--- gcc/sel-sched-ir.c  (working copy)
*************** fallthru_bb_of_jump (rtx jump)
*** 4290,4295 ****
--- 4290,4300 ----
    if (!any_condjump_p (jump))
      return NULL;

+   /* A basic block that ends with a conditional jump may still have 
one successor
+      (and be followed by a barrier), we are not interested.  */
+   if (single_succ_p (BLOCK_FOR_INSN (jump)))
+     return NULL;
+
    return FALLTHRU_EDGE (BLOCK_FOR_INSN (jump))->dest;
  }