This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Patch to fix compiler ICE due to bad PRE


Compiling the attached test case with -O2 using the trunk compiler,
the compiler will ICE. The proposed patch is also attached. The test
is under going, but I like to have discussion on the right fix first.

thanks,

David

Analysis:
-------------

Here is what is happening:

BB6 : (Successor: BB8)
  MEM_19 = phi(MEM_24, MEM_25)
  t_9 = operator new (100)@MEM_19 (--> MEM_26)
  mem_ref(pp_6, -16).x@MEM_26 = t_9 (-->MEM_27)   --> (1)

BB8:
  ...
  MEM_20 = PHI(MEM_27, MEM_24)
  ....
  d.2348_15 = mem_ref(pp_6, -16).x@MEM_20


In the loop header BB3 (which dominates BB6 and BB8),

   BB3:
    ..
    pp_31 = phi (pp_6, 0);
    ...
    pp_6 = pp_31 + 16



In PRE, ANTIC_IN(BB8) includes mem_ref(pp_31,0),x@MEM_20.  After phi
translate, ANTIC_OUT(BB6) gets mem_ref(pp_31,0).x@MEM_27.  However
this expression gets propagated into ANTIC_IN(BB6) even though the
memory expression is killed by (1). The root cause is that the alias
oracle reports that mem_ref(pp_6, -16) and mem_ref(pp_31, 0) are not
aliased as their points-to set do not intersect.

As as consequence of the bad aliasing, the operator new calls gets
transformed into an gimple assignment -- this gimple assignment
becomes the defining statement for MEM_26. In a  later UD chain walk,
the memory chain stops (and triggers the assert) at this new
assignment statement because there is no associated vuse for it.


Test case
--------------

The case is synthesized -- the real world examples involve lots of inlining etc.


int g, *gp[100];
struct V {
  int* x;
  int y;
};

void foo (V **p, V* end, int i)
{
  *p = 0;
  V* pp = *p;
  int s = 100;
  for (; pp < end; )
    {
      pp++;
      (pp-1)->x = &g;
      if (g)
        {
          if (g>10)
            g++;
          int *t = (int*) operator new (100);
         (pp-1)->x = t;
        }
      else
        s--;
     gp[end-pp] = (pp-1)->x + s;
  }
}


Patch
---------


Index: tree-ssa-structalias.c
===================================================================
--- tree-ssa-structalias.c      (revision 186600)
+++ tree-ssa-structalias.c      (working copy)
@@ -6137,6 +6137,9 @@ pt_solutions_intersect_1 (struct pt_solu
   if (pt1->anything || pt2->anything)
     return true;

+  if (pt1->null && pt2->null)
+    return true;
+
   /* If either points to unknown global memory and the other points to
      any global memory they alias.  */
   if ((pt1->nonlocal


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]