Hi!
This patch adds folding of constant arguments v>> and v<<, which helps to
optimize the testcase from the PR back into constant store after vectorized
loop is unrolled.
Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
2013-04-25 Jakub Jelinek <ja...@redhat.com>
PR tree-optimization/57051
* fold-const.c (const_binop): Handle VEC_LSHIFT_EXPR
and VEC_RSHIFT_EXPR if shift count is a multiple of element
bitsize.
--- gcc/fold-const.c.jj 2013-04-12 10:16:25.000000000 +0200
+++ gcc/fold-const.c 2013-04-24 12:37:11.789122719 +0200
@@ -1380,17 +1380,42 @@ const_binop (enum tree_code code, tree a
int count = TYPE_VECTOR_SUBPARTS (type), i;
tree *elts = XALLOCAVEC (tree, count);
- for (i = 0; i < count; i++)
+ if (code == VEC_LSHIFT_EXPR
+ || code == VEC_RSHIFT_EXPR)
{
- tree elem1 = VECTOR_CST_ELT (arg1, i);
-
- elts[i] = const_binop (code, elem1, arg2);
+ if (!host_integerp (arg2, 1))
+ return NULL_TREE;
- /* It is possible that const_binop cannot handle the given
- code and return NULL_TREE */
- if (elts[i] == NULL_TREE)
+ unsigned HOST_WIDE_INT shiftc = tree_low_cst (arg2, 1);
+ unsigned HOST_WIDE_INT outerc = tree_low_cst (TYPE_SIZE (type), 1);
+ unsigned HOST_WIDE_INT innerc
+ = tree_low_cst (TYPE_SIZE (TREE_TYPE (type)), 1);
+ if (shiftc >= outerc || (shiftc % innerc) != 0)
return NULL_TREE;
+ int offset = shiftc / innerc;
+ if (code == VEC_LSHIFT_EXPR)
+ offset = -offset;
+ tree zero = build_zero_cst (TREE_TYPE (type));
+ for (i = 0; i < count; i++)
+ {
+ if (i + offset < 0 || i + offset >= count)
+ elts[i] = zero;
+ else
+ elts[i] = VECTOR_CST_ELT (arg1, i + offset);
+ }
}
+ else
+ for (i = 0; i < count; i++)
+ {
+ tree elem1 = VECTOR_CST_ELT (arg1, i);
+
+ elts[i] = const_binop (code, elem1, arg2);
+
+ /* It is possible that const_binop cannot handle the given
+ code and return NULL_TREE */
+ if (elts[i] == NULL_TREE)
+ return NULL_TREE;
+ }
return build_vector (type, elts);
}
Jakub
