re PR rtl-optimization/8599 (loop unroll bug with -march=k6-3)

PR optimization/8599
	* doloop.c (doloop_modify_runtime): Revert 2002-11-22 change.
	* loop.c (loop_invariant_p): Likewise.

From-SVN: r60392

gcc/ChangeLog

@@ -1,3 +1,9 @@
+2002-12-21  Eric Botcazou  <ebotcazou@libertysurf.fr>
+
+	PR optimization/8599
+	* doloop.c (doloop_modify_runtime): Revert 2002-11-22 change.
+	* loop.c (loop_invariant_p): Likewise.
+
 2002-12-19  Eric Botcazou  <ebotcazou@libertysurf.fr>
 
 	PR optimization/8988

gcc/doloop.c

@@ -598,19 +598,16 @@ doloop_modify_runtime (loop, iterations_max,
 
      If the loop has been unrolled, the full calculation is
 
-       t1 = abs_inc * unroll_number;		        increment per loop
-       n = (abs (final - initial) + abs_inc - 1) / t1;    full loops
-       n += (abs (final - initial) + abs_inc - 1) % t1) >= abs_inc;
-                                                          partial loop
-     which works out to be equivalent to
+       t1 = abs_inc * unroll_number;		increment per loop
+       n = abs (final - initial) / t1;		full loops
+       n += (abs (final - initial) % t1) != 0;	partial loop
 
-       n = (abs (final - initial) + t1 - 1) / t1;
-
-     In the case where the loop was preconditioned, a few iterations
-     may have been executed earlier; but 'initial' was adjusted as they
-     were executed, so we don't need anything special for that case here.
-     As above, when t1 is a power of two we don't need to worry about
-     overflow.
+     However, in certain cases the unrolled loop will be preconditioned
+     by emitting copies of the loop body with conditional branches,
+     so that the unrolled loop is always a full loop and thus needs
+     no exit tests.  In this case we don't want to add the partial
+     loop count.  As above, when t1 is a power of two we don't need to
+     worry about overflow.
 
      The division and modulo operations can be avoided by requiring
      that the increment is a power of 2 (precondition_loop_p enforces
@@ -685,10 +682,10 @@ doloop_modify_runtime (loop, iterations_max,
       if (shift_count < 0)
 	abort ();
 
-      /* (abs (final - initial) + abs_inc * unroll_number - 1) */
-      diff = expand_simple_binop (GET_MODE (diff), PLUS,
-				  diff, GEN_INT (abs_loop_inc - 1),
-				  diff, 1, OPTAB_LIB_WIDEN);
+      if (!loop_info->preconditioned)
+	diff = expand_simple_binop (GET_MODE (diff), PLUS,
+				    diff, GEN_INT (abs_loop_inc - 1),
+				    diff, 1, OPTAB_LIB_WIDEN);
 
       /* (abs (final - initial) + abs_inc * unroll_number - 1)
 	 / (abs_inc * unroll_number)  */

gcc/loop.c

@@ -3278,13 +3278,6 @@ loop_invariant_p (loop, x)
 	  && REGNO (x) < FIRST_PSEUDO_REGISTER && call_used_regs[REGNO (x)])
 	return 0;
 
-      /* Out-of-range regs can occur when we are called from unrolling.
-	 These have always been created by the unroller and are set in
-	 the loop, hence are never invariant. */
-
-      if (REGNO (x) >= regs->num)
-	return 0;
-
       if (regs->array[REGNO (x)].set_in_loop < 0)
 	return 2;
 

gcc/testsuite/ChangeLog

@@ -1,3 +1,7 @@
+2002-12-21  Eric Botcazou  <ebotcazou@libertysurf.fr>
+
+	* gcc.dg/unroll-1.c: New test.
+
 2002-12-19  Eric Botcazou  <ebotcazou@libertysurf.fr>
 
 	* gcc.c-torture/execute/20021219-1.c: New test.

gcc/testsuite/gcc.dg/unroll-1.c

@@ -0,0 +1,26 @@
+/* PR optimization/8599 */
+/* { dg-do run } */
+/* { dg-options "-O2 -funroll-loops" } */
+/* { dg-options "-mcpu=k6 -O2 -funroll-loops" { target i?86-*-* } } */
+
+extern void abort (void);
+
+int array[6] = { 1,2,3,4,5,6 };
+
+void foo()
+{
+  int i;
+
+  for (i = 0; i < 5; i++)
+    array[i] = 0;
+}
+
+int main()
+{
+  foo();
+  if (array[0] || array [1] || array[2] || array[3] || array[4])
+    abort ();
+  if (array[5] != 6)
+    abort ();
+  return 0;
+}