https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102540

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |aldyh at gcc dot gnu.org,
                   |                            |amacleod at redhat dot com
             Blocks|                            |85316
            Version|unknown                     |12.0

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
FRE1 has the following difference, simplifying the (unsigned int) truncation.

   <bb 2> :
   a.0_1 = a;
   _2 = (unsigned int) a.0_1;
   b = _2;
-  c_10 = (long int) _2;
+  _6 = a.0_1 & 4294967295;
+  c_10 = _6;
   if (c_10 != 0)
     goto <bb 3>; [INV]
   else

where the EVRP which now uses ranger retains (diff from GCC 11 to trunk):

   <bb 2> :
   a.0_1 = a;
   _2 = (unsigned int) a.0_1;
   b = _2;
-  c_10 = (long int) _2;
+  _6 = a.0_1 & 4294967295;
+  c_10 = _6;
   if (c_10 != 0)
     goto <bb 3>; [INV]
   else
-    goto <bb 4>; [INV]
+    goto <bb 6>; [INV]

   <bb 3> :
   _4 = c_10 + 1;
   iftmp.2_12 = 2 / _4;
+  if (iftmp.2_12 != 0)
+    goto <bb 4>; [INV]
+  else
+    goto <bb 6>; [INV]

   <bb 4> :
+  if (_2 == 0)
+    goto <bb 5>; [INV]
+  else
+    goto <bb 6>; [INV]
+
+  <bb 5> :
+  foo ();
+
+  <bb 6> :
   a = 0;
   return 0;


after EVRP we have

+  # RANGE [0, 4294967295] NONZERO 4294967295
   c_10 = _6;
...
+  # RANGE [2, 4294967296] NONZERO 8589934591
   _4 = c_10 + 1;
+  # RANGE [0, 1] NONZERO 1
   iftmp.2_12 = 2 / _4;
   if (iftmp.2_12 != 0)

what we did in GCC 11 is simplified the following check

  <bb 4> :
  if (_2 == 0)
    goto <bb 5>; [INV]
  else
    goto <bb 6>; [INV]

based on iftmp.2_12 == [1, 1] via ranger and

evrp visiting BB4
Visiting controlling predicate if (iftmp.2_12 != 0)
Adding assert for iftmp.2_12 from iftmp.2_12 != 0
Intersecting
  long int ~[0, 0]  EQUIVALENCES: { iftmp.2_12 } (1 elements)
and
  long int [0, 1]
to
  long int [1, 1]  EQUIVALENCES: { iftmp.2_12 } (1 elements)
Intersecting
  long int [0, 1]
and
  long int [1, 1]
to
  long int [1, 1]
pushing new range for iftmp.2_12: long int [1, 1]  EQUIVALENCES: { iftmp.2_12 }
(1 elements)
evrp visiting stmt if (_2 == 0)
Folding statement: if (_2 == 0)

Visiting conditional with predicate: if (_2 == 0)

With known ranges
        _2: unsigned int [1, +INF]  EQUIVALENCES: { _2 } (1 elements)

Predicate evaluates to: 0
Folded into: if (0 != 0)

that's now missing, somehow due to the folded IL if the bisect is correct.


Referenced Bugs:

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=85316
[Bug 85316] [meta-bug] VRP range propagation missed cases

Reply via email to