[PATCH] Don't lower VEC_PERM_EXPR if it can be expanded using vec_shr optab (PR target/68483)
Richard Biener
rguenther@suse.de
Mon Nov 23 21:08:00 GMT 2015
On November 23, 2015 8:14:59 PM GMT+01:00, Jakub Jelinek <jakub@redhat.com> wrote:
>Hi!
>
>The patches that removed VEC_RSHIFT_EXPR regressed the first of these
>testcases on i?86/-msse2, because can_vec_perm_p returns false for
>that,
>and indeed as can_vec_perm_p is given only the mode and mask indices,
>there is nothing it can do about it. The former VEC_RSHIFT_EXPR
>is a special VEC_PERM_EXPR with zero (bitwise, so not -0.0) as second
>argument and we can use vec_shr in that case. The expander knows that,
>but
>veclower hasn't been taught about that, which is what this patch does.
>
>The patch also fixes up the shift_amt_for_vec_perm_mask function,
>if the first index is >= nelt, then it certainly is not a vector shift,
>but
>all zeros result (we should have folded it), plus when first is < nelt,
>then it doesn't make sense to mask the result, even for first == nelt -
>1
>first + nelt - 1 is <= 2 * nelt - 1.
>
>Bootstrapped/regtested on x86_64-linux and i686-linux, ok for
>trunk/5.3?
OK.
I wonder if we want a more "powerful" can_vec_perm that gets the actual arguments for example to decide if the result of folding a perm is still valid
To include that shift case.
Richard.
>2015-11-23 Jakub Jelinek <jakub@redhat.com>
>
> PR target/68483
> * tree-vect-generic.c (lower_vec_perm): If VEC_PERM_EXPR
> is valid vec_shr pattern, don't lower it even if can_vec_perm_p
> returns false.
> * optabs.c (shift_amt_for_vec_perm_mask): Return NULL_RTX
> whenever first is nelt or above. Don't mask expected with
> 2 * nelt - 1.
>
> * gcc.target/i386/pr68483-1.c: New test.
> * gcc.target/i386/pr68483-2.c: New test.
>
>--- gcc/tree-vect-generic.c.jj 2015-11-23 13:29:41.959236201 +0100
>+++ gcc/tree-vect-generic.c 2015-11-23 14:13:10.378094173 +0100
>@@ -1272,6 +1272,30 @@ lower_vec_perm (gimple_stmt_iterator *gs
> update_stmt (stmt);
> return;
> }
>+ /* Also detect vec_shr pattern - VEC_PERM_EXPR with zero
>+ vector as VEC1 and a right element shift MASK. */
>+ if (optab_handler (vec_shr_optab, TYPE_MODE (vect_type))
>+ != CODE_FOR_nothing
>+ && TREE_CODE (vec1) == VECTOR_CST
>+ && initializer_zerop (vec1)
>+ && sel_int[0]
>+ && sel_int[0] < elements)
>+ {
>+ for (i = 1; i < elements; ++i)
>+ {
>+ unsigned int expected = i + sel_int[0];
>+ /* Indices into the second vector are all equivalent. */
>+ if (MIN (elements, (unsigned) sel_int[i])
>+ != MIN (elements, expected))
>+ break;
>+ }
>+ if (i == elements)
>+ {
>+ gimple_assign_set_rhs3 (stmt, mask);
>+ update_stmt (stmt);
>+ return;
>+ }
>+ }
> }
> else if (can_vec_perm_p (TYPE_MODE (vect_type), true, NULL))
> return;
>--- gcc/optabs.c.jj 2015-11-23 13:29:41.706239800 +0100
>+++ gcc/optabs.c 2015-11-23 13:33:14.857205132 +0100
>@@ -5232,12 +5232,12 @@ shift_amt_for_vec_perm_mask (rtx sel)
> return NULL_RTX;
>
> first = INTVAL (CONST_VECTOR_ELT (sel, 0));
>- if (first >= 2*nelt)
>+ if (first >= nelt)
> return NULL_RTX;
> for (i = 1; i < nelt; i++)
> {
> int idx = INTVAL (CONST_VECTOR_ELT (sel, i));
>- unsigned int expected = (i + first) & (2 * nelt - 1);
>+ unsigned int expected = i + first;
> /* Indices into the second vector are all equivalent. */
> if (idx < 0 || (MIN (nelt, (unsigned) idx) != MIN (nelt, expected)))
> return NULL_RTX;
>--- gcc/testsuite/gcc.target/i386/pr68483-1.c.jj 2015-11-23
>14:27:54.213534756 +0100
>+++ gcc/testsuite/gcc.target/i386/pr68483-1.c 2015-11-23
>14:33:57.810362424 +0100
>@@ -0,0 +1,22 @@
>+/* PR target/68483 */
>+/* { dg-do compile } */
>+/* { dg-options "-O2 -ftree-vectorize -msse2 -mno-sse3" } */
>+
>+void
>+test (int *input, int *out, unsigned x1, unsigned x2)
>+{
>+ unsigned i, j;
>+ unsigned end = x1;
>+
>+ for (i = j = 0; i < 1000; i++)
>+ {
>+ int sum = 0;
>+ end += x2;
>+ for (; j < end; j++)
>+ sum += input[j];
>+ out[i] = sum;
>+ }
>+}
>+
>+/* { dg-final { scan-assembler "psrldq\[^\n\r]*(8,|, 8)" { target ia32
>} } } */
>+/* { dg-final { scan-assembler "psrldq\[^\n\r]*(4,|, 4)" { target ia32
>} } } */
>--- gcc/testsuite/gcc.target/i386/pr68483-2.c.jj 2015-11-23
>14:33:22.436865628 +0100
>+++ gcc/testsuite/gcc.target/i386/pr68483-2.c 2015-11-23
>14:34:33.716851638 +0100
>@@ -0,0 +1,15 @@
>+/* PR target/68483 */
>+/* { dg-do compile } */
>+/* { dg-options "-O2 -msse2 -mno-sse3" } */
>+
>+typedef int V __attribute__((vector_size (16)));
>+
>+void
>+foo (V *a, V *b)
>+{
>+ V c = { 0, 0, 0, 0 };
>+ V d = { 1, 2, 3, 4 };
>+ *a = __builtin_shuffle (*b, c, d);
>+}
>+
>+/* { dg-final { scan-assembler "psrldq\[^\n\r]*(4,|, 4)" } } */
>
> Jakub
More information about the Gcc-patches
mailing list