On Wed, Jun 18, 2014 at 05:23:13PM +0200, Marek Polacek wrote:
> On Wed, Jun 18, 2014 at 12:55:01PM +0800, Thomas Preud'homme wrote:
> > @@ -1646,20 +1648,23 @@ do_shift_rotate (enum tree_code code,
> > n->n <<= count;
> > break;
> > case RSHIFT_EXPR:
> > + /* Arithmetic shift of signed type: result is dependent on the
> > value. */
> > + if (!TYPE_UNSIGNED (n->type) && (n->n & (0xff << (bitsize - 8))))
> > + return false;
>
> Looks like here an undefined behavior happens:
> tree-ssa-math-opts.c:1672:53: runtime error: shift exponent 56 is too
> large for 32-bit type 'int'
Seems there are actually two spots with this, not just one.
Completely untested fix:
2014-06-18 Jakub Jelinek <ja...@redhat.com>
* tree-ssa-math-opts.c (do_shift_rotate, find_bswap_or_nop_1): Cast
0xff to uint64_t before shifting it up.
--- gcc/tree-ssa-math-opts.c 2014-06-13 08:08:42.354136356 +0200
+++ gcc/tree-ssa-math-opts.c 2014-06-18 19:50:59.486916201 +0200
@@ -1669,7 +1669,8 @@ do_shift_rotate (enum tree_code code,
break;
case RSHIFT_EXPR:
/* Arithmetic shift of signed type: result is dependent on the value. */
- if (!TYPE_UNSIGNED (n->type) && (n->n & (0xff << (bitsize - 8))))
+ if (!TYPE_UNSIGNED (n->type)
+ && (n->n & ((uint64_t) 0xff << (bitsize - 8))))
return false;
n->n >>= count;
break;
@@ -1903,7 +1904,7 @@ find_bswap_or_nop_1 (gimple stmt, struct
old_type_size = TYPE_PRECISION (n->type);
if (!TYPE_UNSIGNED (n->type)
&& type_size > old_type_size
- && n->n & (0xff << (old_type_size - 8)))
+ && n->n & ((uint64_t) 0xff << (old_type_size - 8)))
return NULL_TREE;
if (type_size / BITS_PER_UNIT < (int)(sizeof (int64_t)))
Jakub
