[PATCH] Fix -mcpu=power8 atomic expansion (PR target/69644)

[Jakub Jelinek]

Hi!

rs6000_expand_atomic_compare_and_swap uses oldval directly in
a comparison instruction, but oldval might be a CONST_INT not suitable
for the instruction (such as in the testcase below in SImode comparison
0x8000 constant).  We need to force those into register if they don't
satisfy the predicate.

Bootstrapped/regtested on powerpc64{,le}-linux, ok for trunk?

2016-02-03  Jakub Jelinek  <jakub@redhat.com>

	PR target/69644
	* config/rs6000/rs6000.c (rs6000_expand_atomic_compare_and_swap):
	Force oldval into register if it does not satisfy reg_or_short_operand
	predicate.  Fix up formatting.

	* gcc.dg/pr69644.c: New test.

--- gcc/config/rs6000/rs6000.c.jj	2016-02-02 20:42:01.000000000 +0100
+++ gcc/config/rs6000/rs6000.c	2016-02-03 08:45:31.345521112 +0100
@@ -22275,6 +22275,9 @@ rs6000_expand_atomic_compare_and_swap (r
   else if (reg_overlap_mentioned_p (retval, oldval))
     oldval = copy_to_reg (oldval);
 
+  if (mode != TImode && !reg_or_short_operand (oldval, mode))
+    oldval = copy_to_mode_reg (mode, oldval);
+
   mem = rs6000_pre_atomic_barrier (mem, mod_s);
 
   label1 = NULL_RTX;
@@ -22289,10 +22292,8 @@ rs6000_expand_atomic_compare_and_swap (r
 
   x = retval;
   if (mask)
-    {
-      x = expand_simple_binop (SImode, AND, retval, mask,
-			       NULL_RTX, 1, OPTAB_LIB_WIDEN);
-    }
+    x = expand_simple_binop (SImode, AND, retval, mask,
+			     NULL_RTX, 1, OPTAB_LIB_WIDEN);
 
   cond = gen_reg_rtx (CCmode);
   /* If we have TImode, synthesize a comparison.  */
--- gcc/testsuite/gcc.dg/pr69644.c.jj	2016-02-03 08:42:20.827165549 +0100
+++ gcc/testsuite/gcc.dg/pr69644.c	2016-02-03 08:41:51.000000000 +0100
@@ -0,0 +1,11 @@
+/* PR target/69644 */
+/* { dg-do compile } */
+
+int
+main ()
+{
+  unsigned short x = 0x8000;
+  if (!__sync_bool_compare_and_swap (&x, 0x8000, 0) || x)
+    __builtin_abort ();
+  return 0;
+}

	Jakub