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] Fix SLP miscompilation with constant shifts (PR tree-optimization/52760)


Hi!

The shift count operand of shifts/rotates uses int type which is different
from lhs/rhs1 type.  If shift/rotate counts are constant, vectorizable_shift
converts it to the right type, but then the types in
vect_get_constant_vectors aren't compatible and we end up folding a ushort
VCE of int, which on big-endian results in 0 shift count.

Fixed thusly, acked by Richard on IRC, bootstrapped/regtested on
x86_64-linux and i686-linux, committed to trunk.

2012-03-29  Jakub Jelinek  <jakub@redhat.com>

	PR tree-optimization/52760
	* tree-vect-slp.c (vect_get_constant_vectors): Convert constant_p
	shift count for {L,R}{SHIFT,ROTATE}_EXPR to TREE_TYPE (vector_type).

	* gcc.c-torture/execute/pr52760.c: New test.

--- gcc/tree-vect-slp.c.jj	2012-03-26 11:53:20.000000000 +0200
+++ gcc/tree-vect-slp.c	2012-03-29 13:32:26.153408588 +0200
@@ -2337,8 +2337,23 @@ vect_get_constant_vectors (tree op, slp_
 		    op = gimple_call_arg (stmt, op_num);
 		    break;
 
+		  case LSHIFT_EXPR:
+		  case RSHIFT_EXPR:
+		  case LROTATE_EXPR:
+		  case RROTATE_EXPR:
+		    op = gimple_op (stmt, op_num + 1);
+		    /* Unlike the other binary operators, shifts/rotates have
+		       the shift count being int, instead of the same type as
+		       the lhs, so make sure the scalar is the right type if
+		       we are dealing with vectors of
+		       long long/long/short/char.  */
+		    if (op_num == 1 && constant_p)
+		      op = fold_convert (TREE_TYPE (vector_type), op);
+		    break;
+
 		  default:
 		    op = gimple_op (stmt, op_num + 1);
+		    break;
 		}
 	    }
 
--- gcc/testsuite/gcc.c-torture/execute/pr52760.c.jj	2012-03-29 11:29:08.185685328 +0200
+++ gcc/testsuite/gcc.c-torture/execute/pr52760.c	2012-03-29 11:28:40.000000000 +0200
@@ -0,0 +1,27 @@
+/* PR tree-optimization/52760 */
+
+struct T { unsigned short a, b, c, d; };
+
+__attribute__((noinline, noclone)) void
+foo (int x, struct T *y)
+{
+  int i;
+
+  for (i = 0; i < x; i++)
+    {
+      y[i].a = ((0x00ff & y[i].a >> 8) | (0xff00 & y[i].a << 8));
+      y[i].b = ((0x00ff & y[i].b >> 8) | (0xff00 & y[i].b << 8));
+      y[i].c = ((0x00ff & y[i].c >> 8) | (0xff00 & y[i].c << 8));
+      y[i].d = ((0x00ff & y[i].d >> 8) | (0xff00 & y[i].d << 8));
+    }
+}
+
+int
+main ()
+{
+  struct T t = { 0x0001, 0x0203, 0x0405, 0x0607 };
+  foo (1, &t);
+  if (t.a != 0x0100 || t.b != 0x0302 || t.c != 0x0504 || t.d != 0x0706)
+    __builtin_abort ();
+  return 0;
+}

	Jakub


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