[Bug tree-optimization/68756] [6 Regression] ICE w/ -O1 -floop-nest-optimize and isl 0.15: isl-0.15/isl_id.c:213: unable to find id

["asolokha at gmx dot com" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68756

--- Comment #3 from Arseny Solokha <asolokha at gmx dot com> ---
The following observations were made w/ gcc-6.0.0-alpha20151220 snapshot.

isl fails w/

> isl-0.15/isl_ctx.c:227: Assertion "ctx->ref > 0" failed

when compiling the following snippet w/ the same options as above:

int s9;
int id[2];

void
aj(void)
{
  int l6, p0, dg, it, gs = 0;
  for (it = 0; it < 2; ++it) {
    --gs;
    for (s9 = 0; s9 < 2; ++s9) {
      if (gs != 0)
        break;
      for (dg = 0; dg < 2; ++dg)
        for (p0 = 0; p0 < 2; ++p0)
          for (l6 = 0; l6 < 2; ++l6)
            id[l6] = 0;
    }
  }
}

After moving s9 declarations into aj() cc1 starts to segfault w/ a new
backtrace:

--- qtcykola.c  2015-12-22 17:00:04.827862358 +0700
+++ eu1fk1op.c  2015-12-22 17:00:39.119531288 +0700
@@ -1,9 +1,9 @@
-int s9;
 int id[2];

 void
 aj(void)
 {
+  int s9;
   int l6, p0, dg, it, gs = 0;
   for (it = 0; it < 2; ++it) {
     --gs;

#0  0x000003fff7eefeca in isl_map_dim () from /usr/lib64/libisl.so.15
#1  0x00000000011c3a0a in
translate_isl_ast_to_gimple::get_max_schedule_dimensions(scop*) ()
#2  0x00000000011c3ad9 in
translate_isl_ast_to_gimple::generate_isl_schedule(scop*) ()
#3  0x00000000011c5924 in translate_isl_ast_to_gimple::scop_to_isl_ast(scop*,
std::map<isl_id*, tree_node*, std::less<isl_id*>,
std::allocator<std::pair<isl_id* const, tree_node*> > >&) ()
#4  0x00000000011c9012 in graphite_regenerate_ast_isl(scop*) ()
#5  0x00000000011c0366 in graphite_transform_loops() ()
[trimmed]

Further minimization of these two new testcases yield the same failure as in
#c0.

Another question is, why segfaults in libisl.so aren't propagated to cc1 and,
eventually, to a driver?