I've refined the patch as below.
This patch is checking "unsigned type" adn iv.no_overflow.
While I'm thinking to use "number_of_iterations_exit (loop, e, &niter,
false, false, NULL)"
and "niter.control.no_overflow" to check overflow/wrap, which maybe
more accurate, but relative "expensive".
"nowrap_type_p and scev_probably_wraps_p" may be little cheaper,
but "number_of_iterations_exit" would be more accurate.
Is this right?
BR,
Jiufu Guo.
diff --git a/gcc/tree-ssa-loop-split.c b/gcc/tree-ssa-loop-split.c
index 3a09bbc39e5..425593ca70f 100644
--- a/gcc/tree-ssa-loop-split.c
+++ b/gcc/tree-ssa-loop-split.c
@@ -41,6 +41,7 @@ along with GCC; see the file COPYING3. If not see
#include "cfghooks.h"
#include "gimple-fold.h"
#include "gimplify-me.h"
+#include "tree-ssa-loop-ivopts.h"
/* This file implements two kinds of loop splitting.
@@ -233,7 +234,8 @@ easy_exit_values (class loop *loop)
this. The loops need to fulfill easy_exit_values(). */
static void
-connect_loop_phis (class loop *loop1, class loop *loop2, edge new_e)
+connect_loop_phis (class loop *loop1, class loop *loop2, edge new_e,
+ bool use_prev = false)
{
basic_block rest = loop_preheader_edge (loop2)->src;
gcc_assert (new_e->dest == rest);
@@ -279,7 +281,8 @@ connect_loop_phis (class loop *loop1, class loop
*loop2, edge new_e)
gphi * newphi = create_phi_node (new_init, rest);
add_phi_arg (newphi, init, skip_first, UNKNOWN_LOCATION);
- add_phi_arg (newphi, next, new_e, UNKNOWN_LOCATION);
+ add_phi_arg (newphi, use_prev ? PHI_RESULT (phi_first) : next,
new_e,
+ UNKNOWN_LOCATION);
SET_USE (op, new_init);
}
}
@@ -1593,6 +1596,229 @@ split_loop_on_cond (struct loop *loop)
return do_split;
}
+/* Check if the LOOP exit branch likes "if (idx != bound)",
+ Return the branch edge which exit loop, if overflow/wrap
+ may happen on "idx". */
+
+static edge
+get_ne_cond_branch (struct loop *loop)
+{
+ int i;
+ edge e;
+
+ auto_vec<edge> edges = get_loop_exit_edges (loop);
+ FOR_EACH_VEC_ELT (edges, i, e)
+ {
+ basic_block bb = e->src;
+
+ /* Check gcond. */
+ gimple *last = last_stmt (bb);
+ if (!last || gimple_code (last) != GIMPLE_COND)
+ continue;
+ gcond *cond = as_a<gcond *> (last);
+ enum tree_code code = gimple_cond_code (cond);
+ if (!(code == NE_EXPR
+ || (code == EQ_EXPR && (e->flags & EDGE_TRUE_VALUE))))
+ continue;
+
+ /* Check if bound is invarant. */
+ tree idx = gimple_cond_lhs (cond);
+ tree bnd = gimple_cond_rhs (cond);
+ if (expr_invariant_in_loop_p (loop, idx))
+ std::swap (idx, bnd);
+ else if (!expr_invariant_in_loop_p (loop, bnd))
+ continue;
+
+ /* By default, unsigned type conversion could cause overflow. */
+ tree type = TREE_TYPE (idx);
+ if (!INTEGRAL_TYPE_P (type) || TREE_CODE (idx) != SSA_NAME
+ || !TYPE_UNSIGNED (type)
+ || TYPE_PRECISION (type) == TYPE_PRECISION (sizetype))
+ continue;
+
+ /* Avoid to split if bound is MAX/MIN val. */
+ tree bound_type = TREE_TYPE (bnd);
+ if (TREE_CODE (bnd) == INTEGER_CST && INTEGRAL_TYPE_P
(bound_type)
+ && (bnd == TYPE_MAX_VALUE (bound_type)
+ || bnd == TYPE_MIN_VALUE (bound_type)))
+ continue;
+
+ /* Extract conversion from idx. */
+ if (TREE_CODE (idx) == SSA_NAME)
+ {
+ gimple *stmt = SSA_NAME_DEF_STMT (idx);
+ if (is_gimple_assign (stmt)
+ && CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (stmt))
+ && flow_bb_inside_loop_p (loop, gimple_bb (stmt)))
+ idx = gimple_assign_rhs1 (stmt);
+ }
+
+ /* Check if idx is simple iv with possible overflow/wrap. */
+ class loop *useloop = loop_containing_stmt (cond);
+ affine_iv iv;
+ if (!simple_iv (loop, useloop, idx, &iv, false))
+ continue;
+ if (iv.no_overflow)
+ return NULL;
+
+ /* If base is know value (esplically 0/1), other optimizations
may be
+ able to analyze "idx != bnd" as "idx < bnd" or "idx > bnd". */
+ if (TREE_CODE (iv.base) == INTEGER_CST)
+ continue;
+
+ /* Check loop is simple to split. */
+ gcc_assert (bb != loop->latch);
+
+ if (single_pred_p (loop->latch)
+ && single_pred_edge (loop->latch)->src == bb
+ && (gsi_end_p (gsi_start_nondebug_bb (loop->latch))))
+ return e;
+
+ /* Cheap header. */
+ if (bb == loop->header)
+ {
+ if (get_virtual_phi (bb))
+ continue;
+
+ /* Only one phi. */
+ gphi_iterator psi = gsi_start_phis (bb);
+ if (gsi_end_p (psi))
+ continue;
+ gsi_next (&psi);
+ if (!gsi_end_p (psi))
+ continue;
+
+ /* ++i or ++i */
+ gimple_stmt_iterator gsi = gsi_start_bb (bb);
+ if (gsi_end_p (gsi))
+ continue;
+
+ gimple *s1 = gsi_stmt (gsi);
+ if (!(is_gimple_assign (s1)
+ && (idx == gimple_assign_lhs (s1)
+ || idx == gimple_assign_rhs1 (s1))))
+ continue;
+
+ gsi_next (&gsi);
+ if (!gsi_end_p (gsi) && gsi_stmt (gsi) == cond)
+ return e;
+ }
+ }
+
+ return NULL;
+}
+
+/* Split the LOOP with NE_EXPR into two loops with GT_EXPR and LT_EXPR.
*/
+
+static bool
+split_ne_loop (struct loop *loop, edge cond_e)
+{
+ initialize_original_copy_tables ();
+
+ struct loop *loop2 = loop_version (loop, boolean_true_node, NULL,
+ profile_probability::always (),
+ profile_probability::never (),
+ profile_probability::always (),
+ profile_probability::always (), true);
+
+ gcc_assert (loop2);
+ update_ssa (TODO_update_ssa);
+
+ basic_block loop2_cond_exit_bb = get_bb_copy (cond_e->src);
+ free_original_copy_tables ();
+
+ gcond *gc = as_a<gcond *> (last_stmt (cond_e->src));
+ gcond *dup_gc = as_a<gcond *> (last_stmt (loop2_cond_exit_bb));
+
+ /* Change if (i != n) to LOOP1:if (i > n) and LOOP2:if (i < n) */
+ bool inv = expr_invariant_in_loop_p (loop, gimple_cond_lhs (gc));
+ enum tree_code up_code = inv ? LT_EXPR : GT_EXPR;
+ enum tree_code down_code = inv ? GT_EXPR : LT_EXPR;
+
+ gimple_cond_set_code (gc, up_code);
+ gimple_cond_set_code (dup_gc, down_code);
+
+ /* Link the exit cond edge to new loop. */
+ gcond *break_cond = as_a<gcond *> (gimple_copy (gc));
+ edge pred_e = single_pred_edge (loop->latch);
+ bool simple_loop = pred_e && pred_e->src == cond_e->src
+ && (gsi_end_p (gsi_start_nondebug_bb (loop->latch)));
+ if (simple_loop)
+ gimple_cond_set_code (break_cond, down_code);
+ else
+ gimple_cond_make_true (break_cond);
+
+ basic_block break_bb = split_edge (cond_e);
+ gimple_stmt_iterator gsi = gsi_last_bb (break_bb);
+ gsi_insert_after (&gsi, break_cond, GSI_NEW_STMT);
+
+ edge to_exit = single_succ_edge (break_bb);
+ edge to_new_loop = make_edge (break_bb, loop_preheader_edge
(loop2)->src, 0);
+ to_new_loop->flags |= EDGE_TRUE_VALUE;
+ to_exit->flags |= EDGE_FALSE_VALUE;
+ to_exit->flags &= ~EDGE_FALLTHRU;
+ to_exit->probability = cond_e->probability;
+ to_new_loop->probability = to_exit->probability.invert ();
+
+ update_ssa (TODO_update_ssa);
+
+ connect_loop_phis (loop, loop2, to_new_loop, !simple_loop);
+
+ rewrite_into_loop_closed_ssa_1 (NULL, 0, SSA_OP_USE, loop);
+ if (dump_file && (dump_flags & TDF_DETAILS))
+ fprintf (dump_file, ";; Loop split on wrap index.\n");
+
+ return true;
+}
+
+/* Checks if LOOP contains a suitable NE_EXPR conditional block to
split.
+L_H:
+ if (i!=N)
+ S;
+ i++;
+ goto L_H;
+
+The "i!=N" is like "i>N || i<N", then it could be transform to:
+
+L_H:
+ if (i>N)
+ S;
+ i++;
+ goto L_H;
+L1_H:
+ if (i<N)
+ S;
+ i++;
+ goto L1_H;
+
+The loop with "i<N" is in favor both GIMPLE and RTL passes. */
+
+static bool
+split_loop_on_ne_cond (class loop *loop)
+{
+ int num = 0;
+ basic_block *bbs = get_loop_body (loop);
+
+ if (!can_copy_bbs_p (bbs, loop->num_nodes))
+ {
+ free (bbs);
+ return false;
+ }
+
+ for (unsigned i = 0; i < loop->num_nodes; i++)
+ num += estimate_num_insns_seq (bb_seq (bbs[i]), &eni_size_weights);
+ free (bbs);
+
+ if (num > param_max_peeled_insns)
+ return false;
+
+ edge branch_edge = get_ne_cond_branch (loop);
+ if (branch_edge && split_ne_loop (loop, branch_edge))
+ return true;
+
+ return false;
+}
+
/* Main entry point. Perform loop splitting on all suitable loops. */
static unsigned int
@@ -1622,7 +1848,8 @@ tree_ssa_split_loops (void)
if (optimize_loop_for_size_p (loop))
continue;
- if (split_loop (loop) || split_loop_on_cond (loop))
+ if (split_loop (loop) || split_loop_on_cond (loop)
+ || split_loop_on_ne_cond (loop))
{
/* Mark our containing loop as having had some split inner loops.
*/
loop_outer (loop)->aux = loop;
