re PR tree-optimization/72817 (wrong code at -O3 on x86_64-linux-gnu (in both 32...

	PR tree-optimization/72817
	PR tree-optimization/73450
	* tree-ssa-loop-niter.c (number_of_iterations_ne): Check
	multiple_of_p for adjusted IV.base.

	gcc/testsuite
	* gcc.dg/tree-ssa/pr72817.c: New test.
	* gcc.dg/tree-ssa/pr73450.c: New test.

From-SVN: r239494

diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index c380045cafef2dc6c1b1df85540d9911dd438737..99fb2d2c82cd0ec2d4d838725e50d47698b2d318 100644 (file)
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1,3 +1,10 @@
+2016-08-16  Bin Cheng  <bin.cheng@arm.com>
+
+       PR tree-optimization/72817
+       PR tree-optimization/73450
+       * tree-ssa-loop-niter.c (number_of_iterations_ne): Check
+       multiple_of_p for adjusted IV.base.
+
 2016-08-15  Uros Bizjak  <ubizjak@gmail.com>
 
        PR target/72867

diff --git a/gcc/testsuite/ChangeLog b/gcc/testsuite/ChangeLog
index 810aca1a4a8fa8ff569d9256da998a2f4d48cd36..26266dfae670958df22dfca6f67bc734d9a525ea 100644 (file)
--- a/gcc/testsuite/ChangeLog
+++ b/gcc/testsuite/ChangeLog
@@ -1,3 +1,10 @@
+2016-08-16  Bin Cheng  <bin.cheng@arm.com>
+
+       PR tree-optimization/72817
+       PR tree-optimization/73450
+       * gcc.dg/tree-ssa/pr72817.c: New test.
+       * gcc.dg/tree-ssa/pr73450.c: New test.
+
 2016-08-15  Fritz Reese  <fritzoreese@gmail.com>
 
        * gfortran.dg/init_flag_13.f90: New testcase.

diff --git a/gcc/testsuite/gcc.dg/tree-ssa/pr72817.c b/gcc/testsuite/gcc.dg/tree-ssa/pr72817.c
new file mode 100644 (file)
index 0000000..6ee0b08
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/tree-ssa/pr72817.c
@@ -0,0 +1,13 @@
+/* { dg-do run } */
+/* { dg-options "-O3" } */
+
+char a;
+short b;
+
+int main ()
+{
+  for (a = 3; a != -1; a -= 5)
+    while (b)
+      ;
+  return 0;
+}

diff --git a/gcc/testsuite/gcc.dg/tree-ssa/pr73450.c b/gcc/testsuite/gcc.dg/tree-ssa/pr73450.c
new file mode 100644 (file)
index 0000000..7dd44db
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/tree-ssa/pr73450.c
@@ -0,0 +1,14 @@
+/* { dg-do run } */
+/* { dg-options "-O3" } */
+
+int a;
+char b;
+int main() {
+  char c = 0;
+  for (; c != 3; c = c + 7) {
+    a = b & a;
+    if (a)
+      break;
+  }
+  return 0;
+}

diff --git a/gcc/tree-ssa-loop-niter.c b/gcc/tree-ssa-loop-niter.c
index a50d2b4b43ee88eececa1fb59ea7758545904351..5041c83169b4947b2301c7e1dde9646178b577dc 100644 (file)
--- a/gcc/tree-ssa-loop-niter.c
+++ b/gcc/tree-ssa-loop-niter.c
@@ -999,33 +999,36 @@ number_of_iterations_ne (struct loop *loop, tree type, affine_iv *iv,
   mpz_clear (max);
 
   /* Compute no-overflow information for the control iv.  This can be
-     proven when below two conditions hold.
-
-       1) |FINAL - base| is an exact multiple of step.
-       2) IV evaluates toward FINAL at beginning, i.e:
+     proven when below two conditions are satisfied:
 
+       1) IV evaluates toward FINAL at beginning, i.e:
            base <= FINAL ; step > 0
            base >= FINAL ; step < 0
 
-         Note the first condition holds, the second can be then relaxed
-         to below condition.
+       2) |FINAL - base| is an exact multiple of step.
+
+     Unfortunately, it's hard to prove above conditions after pass loop-ch
+     because loop with exit condition (IV != FINAL) usually will be guarded
+     by initial-condition (IV.base - IV.step != FINAL).  In this case, we
+     can alternatively try to prove below conditions:
+
+       1') IV evaluates toward FINAL at beginning, i.e:
+           new_base = base - step < FINAL ; step > 0
+                                            && base - step doesn't underflow
+           new_base = base - step > FINAL ; step < 0
+                                            && base - step doesn't overflow
 
-           base - step < FINAL ; step > 0
-                                 && base - step doesn't underflow
-           base - step > FINAL ; step < 0
-                                 && base - step doesn't overflow
+       2') |FINAL - new_base| is an exact multiple of step.
 
-         The relaxation is important because after pass loop-ch, loop
-         with exit condition (IV != FINAL) will usually be guarded by
-         pre-condition (IV.base - IV.step != FINAL).  Please refer to
-         PR34114 as an example.
+     Please refer to PR34114 as an example of loop-ch's impact, also refer
+     to PR72817 as an example why condition 2') is necessary.
 
-     Also note, for NE_EXPR, base equals to FINAL is a special case, in
+     Note, for NE_EXPR, base equals to FINAL is a special case, in
      which the loop exits immediately, and the iv does not overflow.  */
   if (!niter->control.no_overflow
       && (integer_onep (s) || multiple_of_p (type, c, s)))
     {
-      tree t, cond, relaxed_cond = boolean_false_node;
+      tree t, cond, new_c, relaxed_cond = boolean_false_node;
 
       if (tree_int_cst_sign_bit (iv->step))
        {
@@ -1039,8 +1042,12 @@ number_of_iterations_ne (struct loop *loop, tree type, affine_iv *iv,
              if (integer_nonzerop (t))
                {
                  t = fold_build2 (MINUS_EXPR, type, iv->base, iv->step);
-                 relaxed_cond = fold_build2 (GT_EXPR, boolean_type_node,
-                                             t, final);
+                 new_c = fold_build2 (MINUS_EXPR, niter_type,
+                                      fold_convert (niter_type, t),
+                                      fold_convert (niter_type, final));
+                 if (multiple_of_p (type, new_c, s))
+                   relaxed_cond = fold_build2 (GT_EXPR, boolean_type_node,
+                                               t, final);
                }
            }
        }
@@ -1056,8 +1063,12 @@ number_of_iterations_ne (struct loop *loop, tree type, affine_iv *iv,
              if (integer_nonzerop (t))
                {
                  t = fold_build2 (MINUS_EXPR, type, iv->base, iv->step);
-                 relaxed_cond = fold_build2 (LT_EXPR, boolean_type_node,
-                                             t, final);
+                 new_c = fold_build2 (MINUS_EXPR, niter_type,
+                                      fold_convert (niter_type, final),
+                                      fold_convert (niter_type, t));
+                 if (multiple_of_p (type, new_c, s))
+                   relaxed_cond = fold_build2 (LT_EXPR, boolean_type_node,
+                                               t, final);
                }
            }
        }