> Please submit a full bug report,
> with preprocessed source if appropriate.
> See <file:///usr/share/doc/gcc-4.6/README.Bugs> for instructions.
> 
> Please let me know if I have forgotten anything you need.

Preprocessed source and the version of Boost would help, thanks

Created attachment 26019 [details]
preprocessed source (gzipped because too large otherwise)

(In reply to comment #1)
> > Please submit a full bug report,
> > with preprocessed source if appropriate.
> > See <file:///usr/share/doc/gcc-4.6/README.Bugs> for instructions.
> > 
> > Please let me know if I have forgotten anything you need.
> 
> Preprocessed source and the version of Boost would help, thanks

I've attached the preprocessed source as a gzipped plain text file. It's over 2 mb and is too large to attach it as plain text.

The boost version is boost 1.37.0

Confirmed with -O1.

Program received signal SIGSEGV, Segmentation fault.
0x0000000000bb9511 in VEC_edge_base_index (vec_=0x7ffff52e2cf8, ix_=0, file_=Cannot access memory at address 0x7fffff7feff8
)
    at /space/rguenther/src/svn/gcc-4_6-branch/gcc/basic-block.h:64
64      DEF_VEC_P(edge);
(gdb) bt
#1  0x0000000000bb9500 in ei_edge (i=...)
    at /space/rguenther/src/svn/gcc-4_6-branch/gcc/basic-block.h:671
#2  0x0000000000bb942d in ei_cond (ei=..., p=0x7fffff7ff0e8)
    at /space/rguenther/src/svn/gcc-4_6-branch/gcc/basic-block.h:692
#3  0x0000000000bbdd75 in predict_paths_for_bb (cur=0x7ffff19a3888, 
    bb=0x7ffff1b2f270, pred=PRED_CONTINUE, taken=NOT_TAKEN)
    at /space/rguenther/src/svn/gcc-4_6-branch/gcc/predict.c:1801
#4  0x0000000000bbdda6 in predict_paths_for_bb (cur=0x7ffff1b2f270, 
    bb=0x7ffff1b2f270, pred=PRED_CONTINUE, taken=NOT_TAKEN)
    at /space/rguenther/src/svn/gcc-4_6-branch/gcc/predict.c:1836
...

we endlessly recurse in predict_paths_for_bb walking

  for (son = first_dom_son (CDI_POST_DOMINATORS, cur);
       son;
       son = next_dom_son (CDI_POST_DOMINATORS, son))
    predict_paths_for_bb (son, bb, pred, taken);

(we never do anything else in the loop above it).  Sequence of recursion
is (cur, bb): (7, 9) (3, 3) (9, 9) (7, 9) ...

;; basic block 2, loop depth 0, count 0
;; prev block 0, next block 9
;; pred:       ENTRY (fallthru,exec)
;; succ:       9 [100.0%]  (fallthru)
<bb 2>:

;; basic block 9, loop depth 1, count 0
;; prev block 2, next block 3
;; pred:       2 [100.0%]  (fallthru) 7 (fallthru,exec)
;; succ:       3 [100.0%]  (fallthru,exec) 8 (eh,exec)
<bb 9>:
f_1 = operator new (80);
D.102841_9 = &f_1->bar;
boost::mutex::mutex (D.102841_9);

;; basic block 3, loop depth 1, count 0
;; prev block 9, next block 4
;; pred:       9 [100.0%]  (fallthru,exec)
;; succ:       7 [100.0%]  (fallthru,exec) 4 (eh,exec)
<bb 3>:
D.102840_10 = &f_1->baz;
boost::mutex::mutex (D.102840_10);
goto <bb 7>;

;; basic block 7, loop depth 1, count 0
;; prev block 6, next block 8
;; pred:       3 [100.0%]  (fallthru,exec)
;; succ:       9 (fallthru,exec)
<bb 7>:
// predicted unlikely by continue predictor.
goto <bb 9>;

... so it seems post-dominators do not honor EH-edges, or at least
it is inconsistent with the way we add fake exit edges for infinite
loops.

Reduced testcase that fails with both 4.7 and 4.6:

class mx {
public:
    mx();
};

int main()
{
  while (true) {
      mx *bar = new mx;
      mx *baz = new mx;
      continue;
  }
  return 0;
}

4.5 happens to work.

For the reduced testcase we recurse endlessly through

      /* If there is non-abnormal path leaving e->src, predict edge
         using predictor.  Otherwise we need to look for paths
         leading to e->src.  */
      if (found)
        predict_edge_def (e, pred, taken);
      else
        predict_paths_for_bb (e->src, e->src, pred, taken);

instead.  That recursion is new in 4.6.  I think it wants to recurse
not to e->src but get_immediate_dominator (CDI_DOMINATORS, cur) (outside of the loop?)

Caused by rev. 161691, thus Honza.

I think this is a dup of bug 48600 which also deals with predict_paths_for_bb .

indeed.

*** This bug has been marked as a duplicate of bug 48600 ***