+2010-02-23 Jakub Jelinek <jakub@redhat.com>
+
+ PR target/43107
+ * config/i386/i386.c (avx_vpermilp_parallel): Reject indexes
+ greater or equal to nelt instead of 2 * nelt.
+ (expand_vec_perm_1): When op0 and op1 are equal, mask indexes
+ with nelt - 1.
+
2010-02-23 Jason Merrill <jason@redhat.com>
PR debug/42800
if (!CONST_INT_P (er))
return 0;
ei = INTVAL (er);
- if (ei >= 2 * nelt)
+ if (ei >= nelt)
return 0;
ipar[i] = ei;
}
input where SEL+CONCAT may not. */
if (d->op0 == d->op1)
{
- if (expand_vselect (d->target, d->op0, d->perm, nelt))
+ int mask = nelt - 1;
+
+ for (i = 0; i < nelt; i++)
+ perm2[i] = d->perm[i] & mask;
+
+ if (expand_vselect (d->target, d->op0, perm2, nelt))
return true;
/* There are plenty of patterns in sse.md that are written for
every other permutation operand. */
for (i = 0; i < nelt; i += 2)
{
- perm2[i] = d->perm[i];
- perm2[i+1] = d->perm[i+1] + nelt;
+ perm2[i] = d->perm[i] & mask;
+ perm2[i + 1] = (d->perm[i + 1] & mask) + nelt;
}
if (expand_vselect_vconcat (d->target, d->op0, d->op0, perm2, nelt))
return true;
/* Recognize shufps, which means adding {0, 0, nelt, nelt}. */
if (nelt >= 4)
{
- memcpy (perm2, d->perm, nelt);
- for (i = 2; i < nelt; i += 4)
+ for (i = 0; i < nelt; i += 4)
{
- perm2[i+0] += nelt;
- perm2[i+1] += nelt;
+ perm2[i + 0] = d->perm[i + 0] & mask;
+ perm2[i + 1] = d->perm[i + 1] & mask;
+ perm2[i + 2] = (d->perm[i + 2] & mask) + nelt;
+ perm2[i + 3] = (d->perm[i + 3] & mask) + nelt;
}
if (expand_vselect_vconcat (d->target, d->op0, d->op0, perm2, nelt))
--- /dev/null
+/* PR target/43107 */
+/* { dg-do compile } */
+/* { dg-options "-O3 -mavx" } */
+
+extern void bar (float b[4][4]);
+
+void
+foo ()
+{
+ float a[4][4], b[4][4];
+ int i, j;
+ for (i = 0; i < 4; i++)
+ {
+ for (j = 0; j < 4; j++)
+ a[i][j] = 0;
+ for (j = 0; j < 4; j++)
+ b[i][j] = a[i][j];
+ }
+ bar (b);
+}