Richard Biener <rguent...@suse.de> writes:
> This makes sure to lower VECTOR_BOOLEAN_TYPE_P typed VEC_COND_EXPRs
> so we don't try to use vcond to expand those. That's especially
> improtant for x86 integer mode boolean vectors but eventually
> as well for aarch64 / gcn VnBImode ones.
>
> Bootstrap and regtest running on x86_64-unknown-linux-gnu.
>
> 2020-11-25 Richard Biener <rguent...@suse.de>
>
> PR middle-end/97579
> * gimple-isel.cc (gimple_expand_vec_cond_expr): Lower
> VECTOR_BOOLEAN_TYPE_P VEC_COND_EXPRs.
>
> * gcc.dg/pr97579.c: New testcase.
> ---
> gcc/gimple-isel.cc | 21 +++++++++++++++++++--
> gcc/testsuite/gcc.dg/pr97579.c | 31 +++++++++++++++++++++++++++++++
> 2 files changed, 50 insertions(+), 2 deletions(-)
> create mode 100644 gcc/testsuite/gcc.dg/pr97579.c
>
> diff --git a/gcc/gimple-isel.cc b/gcc/gimple-isel.cc
> index b5362cc4b01..eac1f62e1ff 100644
> --- a/gcc/gimple-isel.cc
> +++ b/gcc/gimple-isel.cc
> @@ -37,6 +37,7 @@ along with GCC; see the file COPYING3. If not see
> #include "tree-ssa-dce.h"
> #include "memmodel.h"
> #include "optabs.h"
> +#include "gimple-fold.h"
>
> /* Expand all ARRAY_REF(VIEW_CONVERT_EXPR) gimple assignments into calls to
> internal function based on vector type of selected expansion.
> @@ -134,6 +135,24 @@ gimple_expand_vec_cond_expr (gimple_stmt_iterator *gsi,
> lhs = gimple_assign_lhs (stmt);
> machine_mode mode = TYPE_MODE (TREE_TYPE (lhs));
>
> + /* Lower mask typed VEC_COND_EXPRs to bitwise operations. Those can
> + end up generated by folding and at least for integer mode masks
> + we cannot expect vcond expanders to exist. We lower a ? b : c
> + to (b & a) | (c & ~a). */
I think it is reasonable to provide them for vector modes though.
E.g. for SVE we have:
(define_insn "@vcond_mask_<mode><mode>"
[(set (match_operand:PRED_ALL 0 "register_operand" "=Upa")
(ior:PRED_ALL
(and:PRED_ALL
(match_operand:PRED_ALL 3 "register_operand" "Upa")
(match_operand:PRED_ALL 1 "register_operand" "Upa"))
(and:PRED_ALL
(not (match_dup 3))
(match_operand:PRED_ALL 2 "register_operand" "Upa"))))]
"TARGET_SVE"
"sel\t%0.b, %3, %1.b, %2.b"
)
So it might better to check for an integer mode as well.
Thanks,
Richard
> + if (VECTOR_BOOLEAN_TYPE_P (TREE_TYPE (lhs)))
> + {
> + gcc_assert (types_compatible_p (TREE_TYPE (op0), TREE_TYPE (op1)));
> + gimple_seq stmts = NULL;
> + tree type = TREE_TYPE (lhs);
> + location_t loc = gimple_location (stmt);
> + tree tem0 = gimple_build (&stmts, loc, BIT_AND_EXPR, type, op1, op0);
> + tree tem1 = gimple_build (&stmts, loc, BIT_NOT_EXPR, type, op0);
> + tree tem2 = gimple_build (&stmts, loc, BIT_AND_EXPR, type, op2, tem1);
> + tree tem3 = gimple_build (&stmts, loc, BIT_IOR_EXPR, type, tem0, tem2);
> + gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT);
> + return gimple_build_assign (lhs, tem3);
> + }
> +
> gcc_assert (!COMPARISON_CLASS_P (op0));
> if (TREE_CODE (op0) == SSA_NAME)
> {
> @@ -198,7 +217,6 @@ gimple_expand_vec_cond_expr (gimple_stmt_iterator *gsi,
> cmp_op_mode = TYPE_MODE (TREE_TYPE (op0a));
> unsignedp = TYPE_UNSIGNED (TREE_TYPE (op0a));
>
> -
> gcc_assert (known_eq (GET_MODE_NUNITS (mode),
> GET_MODE_NUNITS (cmp_op_mode)));
>
> @@ -246,7 +264,6 @@ gimple_expand_vec_exprs (void)
> {
> gimple *g = gimple_expand_vec_cond_expr (&gsi,
> &vec_cond_ssa_name_uses);
> -
> if (g != NULL)
> {
> tree lhs = gimple_assign_lhs (gsi_stmt (gsi));
> diff --git a/gcc/testsuite/gcc.dg/pr97579.c b/gcc/testsuite/gcc.dg/pr97579.c
> new file mode 100644
> index 00000000000..5cd5427a528
> --- /dev/null
> +++ b/gcc/testsuite/gcc.dg/pr97579.c
> @@ -0,0 +1,31 @@
> +/* { dg-do compile } */
> +/* { dg-options "-O3 --param=max-unswitch-insns=1024" } */
> +/* { dg-additional-options "-mavx512vl" { target x86_64-*-* i?86-*-* } } */
> +
> +int bad_odd_rows_0_0, rows_bad_row1, rows_bad_group_okay, calc_rows_row2;
> +
> +int
> +rows_bad() {
> + int i, in_zeroes;
> + char block;
> + i = 0;
> + for (; i < 5; i++)
> + if (rows_bad_row1 & i)
> + in_zeroes = 0;
> + else {
> + if (!in_zeroes)
> + in_zeroes = 1;
> + if (block & 1)
> + rows_bad_group_okay = 1;
> + }
> + if (in_zeroes)
> + return rows_bad_group_okay;
> +}
> +
> +void
> +calc_rows() {
> + for (; calc_rows_row2; calc_rows_row2++) {
> + rows_bad();
> + bad_odd_rows_0_0 = rows_bad();
> + }
> +}
