+2019-05-16 Jakub Jelinek <jakub@redhat.com>
+
+ * omp-low.c (lower_rec_input_clauses): If OMP_CLAUSE_IF
+ has non-constant expression, force sctx.lane and use two
+ argument IFN_GOMP_SIMD_LANE instead of single argument.
+ * tree-ssa-dce.c (eliminate_unnecessary_stmts): Don't DCE
+ two argument IFN_GOMP_SIMD_LANE without lhs.
+ * tree-vectorizer.h (struct _loop_vec_info): Add simd_if_cond
+ member.
+ (LOOP_VINFO_SIMD_IF_COND, LOOP_REQUIRES_VERSIONING_FOR_SIMD_IF_COND):
+ Define.
+ (LOOP_REQUIRES_VERSIONING): Or in
+ LOOP_REQUIRES_VERSIONING_FOR_SIMD_IF_COND.
+ * tree-vect-loop.c (_loop_vec_info::_loop_vec_info): Initialize
+ simd_if_cond.
+ (vect_analyze_loop_2): Punt if LOOP_VINFO_SIMD_IF_COND is constant 0.
+ * tree-vect-loop-manip.c (vect_loop_versioning): Add runtime check
+ from simd if clause if needed.
+
2019-05-16 Richard Biener <rguenther@suse.de>
* tree-affine.c (expr_to_aff_combination): New function split
tree simt_lane = NULL_TREE, simtrec = NULL_TREE;
tree ivar = NULL_TREE, lvar = NULL_TREE, uid = NULL_TREE;
gimple_seq llist[3] = { };
+ tree nonconst_simd_if = NULL_TREE;
copyin_seq = NULL;
sctx.is_simt = is_simd && omp_find_clause (clauses, OMP_CLAUSE__SIMT_);
case OMP_CLAUSE_IF:
if (integer_zerop (OMP_CLAUSE_IF_EXPR (c)))
sctx.max_vf = 1;
+ else if (TREE_CODE (OMP_CLAUSE_IF_EXPR (c)) != INTEGER_CST)
+ nonconst_simd_if = OMP_CLAUSE_IF_EXPR (c);
break;
case OMP_CLAUSE_SIMDLEN:
if (integer_onep (OMP_CLAUSE_SIMDLEN_EXPR (c)))
if (known_eq (sctx.max_vf, 1U))
sctx.is_simt = false;
+ if (nonconst_simd_if)
+ {
+ if (sctx.lane == NULL_TREE)
+ {
+ sctx.idx = create_tmp_var (unsigned_type_node);
+ sctx.lane = create_tmp_var (unsigned_type_node);
+ }
+ /* FIXME: For now. */
+ sctx.is_simt = false;
+ }
+
if (sctx.lane || sctx.is_simt)
{
uid = create_tmp_var (ptr_type_node, "simduid");
}
if (sctx.lane)
{
- gimple *g
- = gimple_build_call_internal (IFN_GOMP_SIMD_LANE, 1, uid);
+ gimple *g = gimple_build_call_internal (IFN_GOMP_SIMD_LANE,
+ 1 + (nonconst_simd_if != NULL),
+ uid, nonconst_simd_if);
gimple_call_set_lhs (g, sctx.lane);
gimple_stmt_iterator gsi = gsi_start_1 (gimple_omp_body_ptr (ctx->stmt));
gsi_insert_before_without_update (&gsi, g, GSI_SAME_STMT);
+2019-05-16 Jakub Jelinek <jakub@redhat.com>
+
+ * gcc.dg/vect/vect-simd-1.c: New test.
+ * gcc.dg/vect/vect-simd-2.c: New test.
+ * gcc.dg/vect/vect-simd-3.c: New test.
+ * gcc.dg/vect/vect-simd-4.c: New test.
+
2019-05-16 Martin Liska <mliska@suse.cz>
PR lto/90500
--- /dev/null
+/* { dg-additional-options "-fopenmp-simd" } */
+/* { dg-additional-options "-mavx" { target avx_runtime } } */
+
+#include "tree-vect.h"
+
+#define N 1024
+int a[N];
+int x;
+
+__attribute__((noipa)) int
+bar (void)
+{
+ return x;
+}
+
+__attribute__((noipa)) void
+foo (void)
+{
+ #pragma omp simd if (bar ())
+ for (int i = 0; i < N; ++i)
+ a[i] = a[i] + 1;
+}
+
+__attribute__((noipa)) void
+baz (void)
+{
+ int c = 0;
+ #pragma omp simd if (c)
+ for (int i = 0; i < N; ++i)
+ a[i] = a[i] + 1;
+}
+
+__attribute__((noipa)) void
+qux (void)
+{
+ int c = 1;
+ #pragma omp simd if (c)
+ for (int i = 0; i < N; ++i)
+ a[i] = a[i] + 1;
+}
+
+int
+main ()
+{
+ check_vect ();
+ foo ();
+ for (int i = 0; i < N; ++i)
+ if (a[i] != 1)
+ abort ();
+ x = 1;
+ foo ();
+ for (int i = 0; i < N; ++i)
+ if (a[i] != 2)
+ abort ();
+ baz ();
+ for (int i = 0; i < N; ++i)
+ if (a[i] != 3)
+ abort ();
+ qux ();
+ for (int i = 0; i < N; ++i)
+ if (a[i] != 4)
+ abort ();
+ return 0;
+}
--- /dev/null
+/* { dg-do compile } */
+/* { dg-require-effective-target vect_int } */
+/* { dg-additional-options "-fopenmp-simd" } */
+
+#define N 1024
+int a[N];
+int bar (void);
+
+void
+foo (void)
+{
+ #pragma omp simd if (bar ())
+ for (int i = 0; i < N; ++i)
+ a[i] = a[i] + 1;
+}
+
+/* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" } } */
+/* { dg-final { scan-tree-dump-times "created versioning for simd if condition check" 1 "vect" } } */
--- /dev/null
+/* { dg-do compile } */
+/* { dg-require-effective-target vect_int } */
+/* { dg-additional-options "-fopenmp-simd" } */
+
+#define N 1024
+int a[N];
+
+void
+foo (void)
+{
+ int c = 0;
+ #pragma omp simd if (c)
+ for (int i = 0; i < N; ++i)
+ a[i] = a[i] + 1;
+}
+
+/* { dg-final { scan-tree-dump-times "vectorized 0 loops" 1 "vect" } } */
--- /dev/null
+/* { dg-do compile } */
+/* { dg-require-effective-target vect_int } */
+/* { dg-additional-options "-fopenmp-simd" } */
+
+#define N 1024
+int a[N];
+
+void
+foo (void)
+{
+ int c = 1;
+ #pragma omp simd if (c)
+ for (int i = 0; i < N; ++i)
+ a[i] = a[i] + 1;
+}
+
+/* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" } } */
+/* { dg-final { scan-tree-dump-not "created versioning for simd if condition check" "vect" } } */
update_stmt (stmt);
release_ssa_name (name);
- /* GOMP_SIMD_LANE or ASAN_POISON without lhs is not
- needed. */
+ /* GOMP_SIMD_LANE (unless two argument) or ASAN_POISON
+ without lhs is not needed. */
if (gimple_call_internal_p (stmt))
switch (gimple_call_internal_fn (stmt))
{
case IFN_GOMP_SIMD_LANE:
+ if (gimple_call_num_args (stmt) >= 2
+ && !integer_nonzerop (gimple_call_arg (stmt, 1)))
+ break;
+ /* FALLTHRU */
case IFN_ASAN_POISON:
remove_dead_stmt (&gsi, bb, to_remove_edges);
break;
bool version_align = LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (loop_vinfo);
bool version_alias = LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo);
bool version_niter = LOOP_REQUIRES_VERSIONING_FOR_NITERS (loop_vinfo);
+ tree version_simd_if_cond
+ = LOOP_REQUIRES_VERSIONING_FOR_SIMD_IF_COND (loop_vinfo);
if (check_profitability)
cond_expr = fold_build2 (GE_EXPR, boolean_type_node, scalar_loop_iters,
vect_create_cond_for_alias_checks (loop_vinfo, &cond_expr);
}
+ if (version_simd_if_cond)
+ {
+ gcc_assert (dom_info_available_p (CDI_DOMINATORS));
+ if (flag_checking)
+ if (basic_block bb
+ = gimple_bb (SSA_NAME_DEF_STMT (version_simd_if_cond)))
+ gcc_assert (bb != loop->header
+ && dominated_by_p (CDI_DOMINATORS, loop->header, bb)
+ && (scalar_loop == NULL
+ || (bb != scalar_loop->header
+ && dominated_by_p (CDI_DOMINATORS,
+ scalar_loop->header, bb))));
+ tree zero = build_zero_cst (TREE_TYPE (version_simd_if_cond));
+ tree c = fold_build2 (NE_EXPR, boolean_type_node,
+ version_simd_if_cond, zero);
+ if (cond_expr)
+ cond_expr = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
+ c, cond_expr);
+ else
+ cond_expr = c;
+ if (dump_enabled_p ())
+ dump_printf_loc (MSG_NOTE, vect_location,
+ "created versioning for simd if condition check.\n");
+ }
+
cond_expr = force_gimple_operand_1 (unshare_expr (cond_expr),
&gimplify_stmt_list,
is_gimple_condexpr, NULL_TREE);
max_vectorization_factor (0),
mask_skip_niters (NULL_TREE),
mask_compare_type (NULL_TREE),
+ simd_if_cond (NULL_TREE),
unaligned_dr (NULL),
peeling_for_alignment (0),
ptr_mask (0),
gimple *stmt = gsi_stmt (si);
gimple_set_uid (stmt, 0);
add_stmt (stmt);
+ /* If .GOMP_SIMD_LANE call for the current loop has 2 arguments, the
+ second argument is the #pragma omp simd if (x) condition, when 0,
+ loop shouldn't be vectorized, when non-zero constant, it should
+ be vectorized normally, otherwise versioned with vectorized loop
+ done if the condition is non-zero at runtime. */
+ if (loop_in->simduid
+ && is_gimple_call (stmt)
+ && gimple_call_internal_p (stmt)
+ && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE
+ && gimple_call_num_args (stmt) >= 2
+ && TREE_CODE (gimple_call_arg (stmt, 0)) == SSA_NAME
+ && (loop_in->simduid
+ == SSA_NAME_VAR (gimple_call_arg (stmt, 0))))
+ {
+ tree arg = gimple_call_arg (stmt, 1);
+ if (integer_zerop (arg) || TREE_CODE (arg) == SSA_NAME)
+ simd_if_cond = arg;
+ else
+ gcc_assert (integer_nonzerop (arg));
+ }
}
}
}
/* The first group of checks is independent of the vector size. */
fatal = true;
+ if (LOOP_VINFO_SIMD_IF_COND (loop_vinfo)
+ && integer_zerop (LOOP_VINFO_SIMD_IF_COND (loop_vinfo)))
+ return opt_result::failure_at (vect_location,
+ "not vectorized: simd if(0)\n");
+
/* Find all data references in the loop (which correspond to vdefs/vuses)
and analyze their evolution in the loop. */
loops. */
tree mask_compare_type;
+ /* For #pragma omp simd if (x) loops the x expression. If constant 0,
+ the loop should not be vectorized, if constant non-zero, simd_if_cond
+ shouldn't be set and loop vectorized normally, if SSA_NAME, the loop
+ should be versioned on that condition, using scalar loop if the condition
+ is false and vectorized loop otherwise. */
+ tree simd_if_cond;
+
/* Unknown DRs according to which loop was peeled. */
struct dr_vec_info *unaligned_dr;
#define LOOP_VINFO_SCALAR_ITERATION_COST(L) (L)->scalar_cost_vec
#define LOOP_VINFO_SINGLE_SCALAR_ITERATION_COST(L) (L)->single_scalar_iteration_cost
#define LOOP_VINFO_ORIG_LOOP_INFO(L) (L)->orig_loop_info
+#define LOOP_VINFO_SIMD_IF_COND(L) (L)->simd_if_cond
#define LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT(L) \
((L)->may_misalign_stmts.length () > 0)
|| (L)->lower_bounds.length () > 0)
#define LOOP_REQUIRES_VERSIONING_FOR_NITERS(L) \
(LOOP_VINFO_NITERS_ASSUMPTIONS (L))
+#define LOOP_REQUIRES_VERSIONING_FOR_SIMD_IF_COND(L) \
+ (LOOP_VINFO_SIMD_IF_COND (L))
#define LOOP_REQUIRES_VERSIONING(L) \
(LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (L) \
|| LOOP_REQUIRES_VERSIONING_FOR_ALIAS (L) \
- || LOOP_REQUIRES_VERSIONING_FOR_NITERS (L))
+ || LOOP_REQUIRES_VERSIONING_FOR_NITERS (L) \
+ || LOOP_REQUIRES_VERSIONING_FOR_SIMD_IF_COND (L))
#define LOOP_VINFO_NITERS_KNOWN_P(L) \
(tree_fits_shwi_p ((L)->num_iters) && tree_to_shwi ((L)->num_iters) > 0)