Patch: fix crash with -fprofile-arcs

[Dale Johannesen <dalej at apple dot com>]

/* { dg-do compile } */
/* { dg-options "-fprofile-arcs" } */
#include <setjmp.h>
int foo (jmp_buf buf) { longjmp(buf, 1); }

> On Monday, December 9, 2002, at 11:25  AM, Richard Henderson wrote:
> > It seems the block that calls longjmp has no successors, and the code
> > in profile.c is not expecting that.  Could somebody clarify which code
> > is right?
> Before the invocation of the profiling routines, no successors is 
> correct.
> During the invocation of the profiling routines, however, we add fake
> edges from calls to the exit block.  I would not expect this to change
> for a noreturn function like longjmp.
>
> So I guess the quesion is "where did you get this null, exactly?"
The block has no successors, but the longjmp call is not marked 
NORETURN,
so need_fake_edge_p thinks we need to execute the special case at the
beginning of flow_call_edges_add.  This doesn't allow for bb->succ==0.
I don't understand why the longjmp call doesn't get marked NORETURN,
but expand_call goes to considerable lengths to handle longjmp 
separately,
and I assume there's a reason for that... The patch here just checks
for the failing case in flow_call_edges_add (the necessary fake edge is
added later by add_noreturn_fake_exit_edges, which looks for bb->succ==0
rather than NORETURN). Bootstrapped and tested on Darwin.

2002-12-11  Dale Johannesen  <dalej@apple.com>

        * cfganal.c (flow_call_edges_add):  Handle longjmp in last 
block.

Index: cfganal.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/cfganal.c,v
retrieving revision 1.30
diff -u -d -b -w -c -3 -p -r1.30 cfganal.c
cvs server: conflicting specifications of output style
*** cfganal.c   11 Oct 2002 21:09:59 -0000      1.30
--- cfganal.c   11 Dec 2002 18:38:59 -0000
*************** flow_call_edges_add (blocks)
*** 309,314 ****
--- 309,319 ----
        basic_block bb = EXIT_BLOCK_PTR->prev_bb;
        rtx insn = bb->end;

+       /* When this block ends with a longjmp call, bb->succ will be 
null,
+        yet the call is not marked NORETURN.  This case will be handled
+        later; avoid a crash for now. */
+       if (bb->succ)
+         {
            /* Back up past insns that must be kept in the same block 
as a call.  */
            while (insn != bb->head
                 && keep_with_call_p (insn))
*************** flow_call_edges_add (blocks)
*** 324,329 ****
--- 329,335 ----

                insert_insn_on_edge (gen_rtx_USE (VOIDmode, 
const0_rtx), e);
                commit_edge_insertions ();
+             }
        }
      }