[PATCH] problems with bb-reorder.c rewrite

[Richard Earnshaw]

> Michael Hayes <m.hayes@elec.canterbury.ac.nz> writes:
> 
> |> Richard Henderson writes:
> |>  > On Sat, Mar 08, 2003 at 03:42:34PM +1300, Michael Hayes wrote:
> |>  > > 	* function.c (thread_prologue_and_epilogue_insns): Use redirect_jump 
> |>  > > 	for conditional return.
> |>  > 
> |>  > Exactly.  Thanks.
> |> 
> |> Ta.  I've committed this to 3.3 and head.
> 
> Breaks bootstrap on ia64 (both branches):
> 
> stage1/xgcc -Bstage1/ -B/usr/local/ia64-suse-linux/bin/ -c   -g -O2 -DIN_GCC   -W -Wall -Wwrite-strings -Wstrict-prototypes -Wmissing-prototypes -Wtraditional -pedantic -Wno-long-long -fno-common -Werror   -DHAVE_CONFIG_H    -I. -I. -I../../gcc -I../../gcc/. -I../../gcc/config -I../../gcc/../include ../../gcc/ifcvt.c -o ifcvt.o
> ../../gcc/ifcvt.c: In function `count_bb_insns':
> ../../gcc/ifcvt.c:134: internal compiler error: Segmentation fault
> Please submit a full bug report,
> with preprocessed source if appropriate.
> See <URL:http://www.gnu.org/software/gcc/bugs.html> for instructions.

I get an abort on arm-netbsd as well, though for a different reason (I 
think).  It seems that this change has made generation of conditional 
return instructions more aggressive, and so we trip a latent bug in 
if-conversion.

in find_if_case_1 we convert a sequence of the form

bb n
	(cond a (return))

bb p
	(insn)
	(jump x)

into

bb n
	(cond ~a (insn))
	(cond ~a (jump x))
bb p
	(return)

but in doing so we delete the original bb p and then add another one with 
the same bb number.  However, when we deleted the old p we removed the 
dominance information; but when we insert the new p we don't update the 
dominance information.  Consequently when we try to convert the return 
into a conditional one (with a second call to find_if_case_1) we abort 
when trying to delete a dominance entry that doesn't exist.

This can be fixed with the following patch.  Richard, does this look ok?  
I'm not particularly familiar with the way the dominance code works:

<date>  Richard Earnshaw  <rearnsha@arm.com>

	* ifcvt.c (find_if_case_1): If we add a new bb, update the dominance
	information.


-------------- next part --------------
Index: ifcvt.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/ifcvt.c,v
retrieving revision 1.112
diff -p -r1.112 ifcvt.c
*** ifcvt.c	24 Jan 2003 02:59:58 -0000	1.112
--- ifcvt.c	15 Mar 2003 16:58:05 -0000
*************** find_if_case_1 (test_bb, then_edge, else
*** 2741,2746 ****
--- 2741,2748 ----
      {
        new_bb->index = then_bb_index;
        BASIC_BLOCK (then_bb_index) = new_bb;
+       if (post_dominators)
+ 	add_to_dominance_info (post_dominators, new_bb);
      }
    /* We've possibly created jump to next insn, cleanup_cfg will solve that
       later.  */