/* Loop manipulation code for GNU compiler.
- Copyright (C) 2002, 2003, 2004, 2005 Free Software Foundation, Inc.
+ Copyright (C) 2002, 2003, 2004, 2005, 2007, 2008, 2009 Free Software
+ Foundation, Inc.
This file is part of GCC.
GCC is free software; you can redistribute it and/or modify it under
the terms of the GNU General Public License as published by the Free
-Software Foundation; either version 2, or (at your option) any later
+Software Foundation; either version 3, or (at your option) any later
version.
GCC is distributed in the hope that it will be useful, but WITHOUT ANY
for more details.
You should have received a copy of the GNU General Public License
-along with GCC; see the file COPYING. If not, write to the Free
-Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA
-02110-1301, USA. */
+along with GCC; see the file COPYING3. If not see
+<http://www.gnu.org/licenses/>. */
#include "config.h"
#include "system.h"
#include "cfglayout.h"
#include "cfghooks.h"
#include "output.h"
+#include "tree-flow.h"
static void duplicate_subloops (struct loop *, struct loop *);
static void copy_loops_to (struct loop **, int,
struct loop *);
static void loop_redirect_edge (edge, basic_block);
-static bool loop_delete_branch_edge (edge, int);
static void remove_bbs (basic_block *, int);
-static bool rpe_enum_p (basic_block, void *);
+static bool rpe_enum_p (const_basic_block, const void *);
static int find_path (edge, basic_block **);
-static bool alp_enum_p (basic_block, void *);
static void fix_loop_placements (struct loop *, bool *);
static bool fix_bb_placement (basic_block);
static void fix_bb_placements (basic_block, bool *);
-static void place_new_loop (struct loop *);
-static void scale_loop_frequencies (struct loop *, int, int);
-static basic_block create_preheader (struct loop *, int);
static void unloop (struct loop *, bool *);
#define RDIV(X,Y) (((X) + (Y) / 2) / (Y))
/* Checks whether basic block BB is dominated by DATA. */
static bool
-rpe_enum_p (basic_block bb, void *data)
+rpe_enum_p (const_basic_block bb, const void *data)
{
- return dominated_by_p (CDI_DOMINATORS, bb, data);
+ return dominated_by_p (CDI_DOMINATORS, bb, (const_basic_block) data);
}
/* Remove basic blocks BBS. NBBS is the number of the basic blocks. */
act = e->dest->loop_father;
if (act->header == e->dest)
- act = act->outer;
+ act = loop_outer (act);
if (flow_loop_nested_p (loop, act))
loop = act;
return true;
}
+/* Fix placement of LOOP inside loop tree, i.e. find the innermost superloop
+ of LOOP to that leads at least one exit edge of LOOP, and set it
+ as the immediate superloop of LOOP. Return true if the immediate superloop
+ of LOOP changed. */
+
+static bool
+fix_loop_placement (struct loop *loop)
+{
+ unsigned i;
+ edge e;
+ VEC (edge, heap) *exits = get_loop_exit_edges (loop);
+ struct loop *father = current_loops->tree_root, *act;
+ bool ret = false;
+
+ for (i = 0; VEC_iterate (edge, exits, i, e); i++)
+ {
+ act = find_common_loop (loop, e->dest->loop_father);
+ if (flow_loop_nested_p (father, act))
+ father = act;
+ }
+
+ if (father != loop_outer (loop))
+ {
+ for (act = loop_outer (loop); act != father; act = loop_outer (act))
+ act->num_nodes -= loop->num_nodes;
+ flow_loop_tree_node_remove (loop);
+ flow_loop_tree_node_add (father, loop);
+
+ /* The exit edges of LOOP no longer exits its original immediate
+ superloops; remove them from the appropriate exit lists. */
+ for (i = 0; VEC_iterate (edge, exits, i, e); i++)
+ rescan_loop_exit (e, false, false);
+
+ ret = true;
+ }
+
+ VEC_free (edge, heap, exits);
+ return ret;
+}
+
/* Fix placements of basic blocks inside loop hierarchy stored in loops; i.e.
enforce condition condition stated in description of fix_bb_placement. We
start from basic block FROM that had some of its successors removed, so that
remove_path (edge e)
{
edge ae;
- basic_block *rem_bbs, *bord_bbs, *dom_bbs, from, bb;
- int i, nrem, n_bord_bbs, n_dom_bbs, nreml;
+ basic_block *rem_bbs, *bord_bbs, from, bb;
+ VEC (basic_block, heap) *dom_bbs;
+ int i, nrem, n_bord_bbs, nreml;
sbitmap seen;
- bool deleted, irred_invalidated = false;
+ bool irred_invalidated = false;
struct loop **deleted_loop;
- if (!loop_delete_branch_edge (e, 0))
+ if (!can_remove_branch_p (e))
return false;
/* Keep track of whether we need to update information about irreducible
we belong to. In this case first unloop the loops, then proceed
normally. We may assume that e->dest is not a header of any loop,
as it now has exactly one predecessor. */
- while (e->src->loop_father->outer
+ while (loop_outer (e->src->loop_father)
&& dominated_by_p (CDI_DOMINATORS,
e->src->loop_father->latch, e->dest))
unloop (e->src->loop_father, &irred_invalidated);
/* Remove the path. */
from = e->src;
- deleted = loop_delete_branch_edge (e, 1);
- gcc_assert (deleted);
- dom_bbs = XCNEWVEC (basic_block, n_basic_blocks);
+ remove_branch (e);
+ dom_bbs = NULL;
/* Cancel loops contained in the path. */
deleted_loop = XNEWVEC (struct loop *, nrem);
if (rem_bbs[i]->loop_father->header == rem_bbs[i])
deleted_loop[nreml++] = rem_bbs[i]->loop_father;
- remove_bbs (rem_bbs, nrem);
- free (rem_bbs);
-
for (i = 0; i < nreml; i++)
cancel_loop_tree (deleted_loop[i]);
free (deleted_loop);
+ remove_bbs (rem_bbs, nrem);
+ free (rem_bbs);
+
/* Find blocks whose dominators may be affected. */
- n_dom_bbs = 0;
sbitmap_zero (seen);
for (i = 0; i < n_bord_bbs; i++)
{
ldom;
ldom = next_dom_son (CDI_DOMINATORS, ldom))
if (!dominated_by_p (CDI_DOMINATORS, from, ldom))
- dom_bbs[n_dom_bbs++] = ldom;
+ VEC_safe_push (basic_block, heap, dom_bbs, ldom);
}
free (seen);
/* Recount dominators. */
- iterate_fix_dominators (CDI_DOMINATORS, dom_bbs, n_dom_bbs);
- free (dom_bbs);
+ iterate_fix_dominators (CDI_DOMINATORS, dom_bbs, true);
+ VEC_free (basic_block, heap, dom_bbs);
free (bord_bbs);
/* Fix placements of basic blocks inside loops and the placement of
fix_loop_placements (from->loop_father, &irred_invalidated);
if (irred_invalidated
- && (current_loops->state & LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS) != 0)
+ && loops_state_satisfies_p (LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS))
mark_irreducible_loops ();
return true;
}
-/* Predicate for enumeration in add_loop. */
-static bool
-alp_enum_p (basic_block bb, void *alp_header)
+/* Creates place for a new LOOP in loops structure. */
+
+static void
+place_new_loop (struct loop *loop)
{
- return bb != (basic_block) alp_header;
+ loop->num = number_of_loops ();
+ VEC_safe_push (loop_p, gc, current_loops->larray, loop);
}
/* Given LOOP structure with filled header and latch, find the body of the
corresponding loop and add it to loops tree. Insert the LOOP as a son of
outer. */
-static void
+void
add_loop (struct loop *loop, struct loop *outer)
{
basic_block *bbs;
int i, n;
+ struct loop *subloop;
+ edge e;
+ edge_iterator ei;
/* Add it to loop structure. */
place_new_loop (loop);
flow_loop_tree_node_add (outer, loop);
/* Find its nodes. */
- bbs = XCNEWVEC (basic_block, n_basic_blocks);
- n = dfs_enumerate_from (loop->latch, 1, alp_enum_p,
- bbs, n_basic_blocks, loop->header);
+ bbs = XNEWVEC (basic_block, n_basic_blocks);
+ n = get_loop_body_with_size (loop, bbs, n_basic_blocks);
for (i = 0; i < n; i++)
{
- remove_bb_from_loops (bbs[i]);
- add_bb_to_loop (bbs[i], loop);
+ if (bbs[i]->loop_father == outer)
+ {
+ remove_bb_from_loops (bbs[i]);
+ add_bb_to_loop (bbs[i], loop);
+ continue;
+ }
+
+ loop->num_nodes++;
+
+ /* If we find a direct subloop of OUTER, move it to LOOP. */
+ subloop = bbs[i]->loop_father;
+ if (loop_outer (subloop) == outer
+ && subloop->header == bbs[i])
+ {
+ flow_loop_tree_node_remove (subloop);
+ flow_loop_tree_node_add (loop, subloop);
+ }
+ }
+
+ /* Update the information about loop exit edges. */
+ for (i = 0; i < n; i++)
+ {
+ FOR_EACH_EDGE (e, ei, bbs[i]->succs)
+ {
+ rescan_loop_exit (e, false, false);
+ }
}
- remove_bb_from_loops (loop->header);
- add_bb_to_loop (loop->header, loop);
free (bbs);
}
/* Multiply all frequencies in LOOP by NUM/DEN. */
-static void
+void
scale_loop_frequencies (struct loop *loop, int num, int den)
{
basic_block *bbs;
free (bbs);
}
+/* Recompute dominance information for basic blocks outside LOOP. */
+
+static void
+update_dominators_in_loop (struct loop *loop)
+{
+ VEC (basic_block, heap) *dom_bbs = NULL;
+ sbitmap seen;
+ basic_block *body;
+ unsigned i;
+
+ seen = sbitmap_alloc (last_basic_block);
+ sbitmap_zero (seen);
+ body = get_loop_body (loop);
+
+ for (i = 0; i < loop->num_nodes; i++)
+ SET_BIT (seen, body[i]->index);
+
+ for (i = 0; i < loop->num_nodes; i++)
+ {
+ basic_block ldom;
+
+ for (ldom = first_dom_son (CDI_DOMINATORS, body[i]);
+ ldom;
+ ldom = next_dom_son (CDI_DOMINATORS, ldom))
+ if (!TEST_BIT (seen, ldom->index))
+ {
+ SET_BIT (seen, ldom->index);
+ VEC_safe_push (basic_block, heap, dom_bbs, ldom);
+ }
+ }
+
+ iterate_fix_dominators (CDI_DOMINATORS, dom_bbs, false);
+ free (body);
+ free (seen);
+ VEC_free (basic_block, heap, dom_bbs);
+}
+
+/* Creates an if region as shown above. CONDITION is used to create
+ the test for the if.
+
+ |
+ | ------------- -------------
+ | | pred_bb | | pred_bb |
+ | ------------- -------------
+ | | |
+ | | | ENTRY_EDGE
+ | | ENTRY_EDGE V
+ | | ====> -------------
+ | | | cond_bb |
+ | | | CONDITION |
+ | | -------------
+ | V / \
+ | ------------- e_false / \ e_true
+ | | succ_bb | V V
+ | ------------- ----------- -----------
+ | | false_bb | | true_bb |
+ | ----------- -----------
+ | \ /
+ | \ /
+ | V V
+ | -------------
+ | | join_bb |
+ | -------------
+ | | exit_edge (result)
+ | V
+ | -----------
+ | | succ_bb |
+ | -----------
+ |
+ */
+
+edge
+create_empty_if_region_on_edge (edge entry_edge, tree condition)
+{
+
+ basic_block succ_bb, cond_bb, true_bb, false_bb, join_bb;
+ edge e_true, e_false, exit_edge;
+ gimple cond_stmt;
+ tree simple_cond;
+ gimple_stmt_iterator gsi;
+
+ succ_bb = entry_edge->dest;
+ cond_bb = split_edge (entry_edge);
+
+ /* Insert condition in cond_bb. */
+ gsi = gsi_last_bb (cond_bb);
+ simple_cond =
+ force_gimple_operand_gsi (&gsi, condition, true, NULL,
+ false, GSI_NEW_STMT);
+ cond_stmt = gimple_build_cond_from_tree (simple_cond, NULL_TREE, NULL_TREE);
+ gsi = gsi_last_bb (cond_bb);
+ gsi_insert_after (&gsi, cond_stmt, GSI_NEW_STMT);
+
+ join_bb = split_edge (single_succ_edge (cond_bb));
+
+ e_true = single_succ_edge (cond_bb);
+ true_bb = split_edge (e_true);
+
+ e_false = make_edge (cond_bb, join_bb, 0);
+ false_bb = split_edge (e_false);
+
+ e_true->flags &= ~EDGE_FALLTHRU;
+ e_true->flags |= EDGE_TRUE_VALUE;
+ e_false->flags &= ~EDGE_FALLTHRU;
+ e_false->flags |= EDGE_FALSE_VALUE;
+
+ set_immediate_dominator (CDI_DOMINATORS, cond_bb, entry_edge->src);
+ set_immediate_dominator (CDI_DOMINATORS, true_bb, cond_bb);
+ set_immediate_dominator (CDI_DOMINATORS, false_bb, cond_bb);
+ set_immediate_dominator (CDI_DOMINATORS, join_bb, cond_bb);
+
+ exit_edge = single_succ_edge (join_bb);
+
+ if (single_pred_p (exit_edge->dest))
+ set_immediate_dominator (CDI_DOMINATORS, exit_edge->dest, join_bb);
+
+ return exit_edge;
+}
+
+/* create_empty_loop_on_edge
+ |
+ | ------------- ------------------------
+ | | pred_bb | | pred_bb |
+ | ------------- | IV_0 = INITIAL_VALUE |
+ | | ------------------------
+ | | ______ | ENTRY_EDGE
+ | | ENTRY_EDGE / V V
+ | | ====> | -----------------------------
+ | | | | IV_BEFORE = phi (IV_0, IV) |
+ | | | | loop_header |
+ | V | | IV_BEFORE <= UPPER_BOUND |
+ | ------------- | -----------------------\-----
+ | | succ_bb | | | \
+ | ------------- | | \ exit_e
+ | | V V---------
+ | | -------------- | succ_bb |
+ | | | loop_latch | ----------
+ | | |IV = IV_BEFORE + STRIDE
+ | | --------------
+ | \ /
+ | \ ___ /
+
+ Creates an empty loop as shown above, the IV_BEFORE is the SSA_NAME
+ that is used before the increment of IV. IV_BEFORE should be used for
+ adding code to the body that uses the IV. OUTER is the outer loop in
+ which the new loop should be inserted. */
+
+struct loop *
+create_empty_loop_on_edge (edge entry_edge,
+ tree initial_value,
+ tree stride, tree upper_bound,
+ tree iv,
+ tree *iv_before,
+ struct loop *outer)
+{
+ basic_block loop_header, loop_latch, succ_bb, pred_bb;
+ struct loop *loop;
+ int freq;
+ gcov_type cnt;
+ gimple_stmt_iterator gsi;
+ bool insert_after;
+ gimple_seq stmts;
+ gimple cond_expr;
+ tree exit_test;
+ edge exit_e;
+ int prob;
+ tree upper_bound_gimplified;
+
+ gcc_assert (entry_edge && initial_value && stride && upper_bound && iv);
+
+ /* Create header, latch and wire up the loop. */
+ pred_bb = entry_edge->src;
+ loop_header = split_edge (entry_edge);
+ loop_latch = split_edge (single_succ_edge (loop_header));
+ succ_bb = single_succ (loop_latch);
+ make_edge (loop_header, succ_bb, 0);
+ redirect_edge_succ_nodup (single_succ_edge (loop_latch), loop_header);
+
+ /* Set immediate dominator information. */
+ set_immediate_dominator (CDI_DOMINATORS, loop_header, pred_bb);
+ set_immediate_dominator (CDI_DOMINATORS, loop_latch, loop_header);
+ set_immediate_dominator (CDI_DOMINATORS, succ_bb, loop_header);
+
+ /* Initialize a loop structure and put it in a loop hierarchy. */
+ loop = alloc_loop ();
+ loop->header = loop_header;
+ loop->latch = loop_latch;
+ add_loop (loop, outer);
+
+ /* TODO: Fix frequencies and counts. */
+ freq = EDGE_FREQUENCY (entry_edge);
+ cnt = entry_edge->count;
+
+ prob = REG_BR_PROB_BASE / 2;
+
+ scale_loop_frequencies (loop, REG_BR_PROB_BASE - prob, REG_BR_PROB_BASE);
+
+ /* Update dominators. */
+ update_dominators_in_loop (loop);
+
+ /* Construct IV code in loop. */
+ initial_value = force_gimple_operand (initial_value, &stmts, true, iv);
+ if (stmts)
+ {
+ gsi_insert_seq_on_edge (loop_preheader_edge (loop), stmts);
+ gsi_commit_edge_inserts ();
+ }
+
+ standard_iv_increment_position (loop, &gsi, &insert_after);
+ create_iv (initial_value, stride, iv, loop, &gsi, insert_after,
+ iv_before, NULL);
+
+ /* Modify edge flags. */
+ exit_e = single_exit (loop);
+ exit_e->flags = EDGE_LOOP_EXIT | EDGE_FALSE_VALUE;
+ single_pred_edge (loop_latch)->flags = EDGE_TRUE_VALUE;
+
+ gsi = gsi_last_bb (exit_e->src);
+
+ upper_bound_gimplified =
+ force_gimple_operand_gsi (&gsi, upper_bound, true, NULL,
+ false, GSI_NEW_STMT);
+ gsi = gsi_last_bb (exit_e->src);
+
+ cond_expr = gimple_build_cond
+ (LE_EXPR, *iv_before, upper_bound_gimplified, NULL_TREE, NULL_TREE);
+
+ exit_test = gimple_cond_lhs (cond_expr);
+ exit_test = force_gimple_operand_gsi (&gsi, exit_test, true, NULL,
+ false, GSI_NEW_STMT);
+ gimple_cond_set_lhs (cond_expr, exit_test);
+ gsi = gsi_last_bb (exit_e->src);
+ gsi_insert_after (&gsi, cond_expr, GSI_NEW_STMT);
+
+ return loop;
+}
+
/* Make area between HEADER_EDGE and LATCH_EDGE a loop by connecting
latch to header and update loop tree and dominators
accordingly. Everything between them plus LATCH_EDGE destination must
LATCH_EDGE source. HEADER_EDGE is redirected to basic block SWITCH_BB,
FALSE_EDGE of SWITCH_BB to original destination of HEADER_EDGE and
TRUE_EDGE of SWITCH_BB to original destination of LATCH_EDGE.
- Returns newly created loop. */
+ Returns the newly created loop. Frequencies and counts in the new loop
+ are scaled by FALSE_SCALE and in the old one by TRUE_SCALE. */
struct loop *
loopify (edge latch_edge, edge header_edge,
basic_block switch_bb, edge true_edge, edge false_edge,
- bool redirect_all_edges)
+ bool redirect_all_edges, unsigned true_scale, unsigned false_scale)
{
basic_block succ_bb = latch_edge->dest;
basic_block pred_bb = header_edge->src;
- basic_block *dom_bbs, *body;
- unsigned n_dom_bbs, i;
- sbitmap seen;
- struct loop *loop = XCNEW (struct loop);
- struct loop *outer = succ_bb->loop_father->outer;
- int freq, prob, tot_prob;
+ struct loop *loop = alloc_loop ();
+ struct loop *outer = loop_outer (succ_bb->loop_father);
+ int freq;
gcov_type cnt;
edge e;
edge_iterator ei;
freq = EDGE_FREQUENCY (header_edge);
cnt = header_edge->count;
- prob = EDGE_SUCC (switch_bb, 0)->probability;
- tot_prob = prob + EDGE_SUCC (switch_bb, 1)->probability;
- if (tot_prob == 0)
- tot_prob = 1;
/* Redirect edges. */
loop_redirect_edge (latch_edge, loop->header);
add_bb_to_loop (switch_bb, outer);
/* Fix frequencies. */
- switch_bb->frequency = freq;
- switch_bb->count = cnt;
- FOR_EACH_EDGE (e, ei, switch_bb->succs)
- e->count = (switch_bb->count * e->probability) / REG_BR_PROB_BASE;
- scale_loop_frequencies (loop, prob, tot_prob);
- scale_loop_frequencies (succ_bb->loop_father, tot_prob - prob, tot_prob);
-
- /* Update dominators of blocks outside of LOOP. */
- dom_bbs = XCNEWVEC (basic_block, n_basic_blocks);
- n_dom_bbs = 0;
- seen = sbitmap_alloc (last_basic_block);
- sbitmap_zero (seen);
- body = get_loop_body (loop);
-
- for (i = 0; i < loop->num_nodes; i++)
- SET_BIT (seen, body[i]->index);
-
- for (i = 0; i < loop->num_nodes; i++)
+ if (redirect_all_edges)
{
- basic_block ldom;
-
- for (ldom = first_dom_son (CDI_DOMINATORS, body[i]);
- ldom;
- ldom = next_dom_son (CDI_DOMINATORS, ldom))
- if (!TEST_BIT (seen, ldom->index))
- {
- SET_BIT (seen, ldom->index);
- dom_bbs[n_dom_bbs++] = ldom;
- }
+ switch_bb->frequency = freq;
+ switch_bb->count = cnt;
+ FOR_EACH_EDGE (e, ei, switch_bb->succs)
+ {
+ e->count = (switch_bb->count * e->probability) / REG_BR_PROB_BASE;
+ }
}
-
- iterate_fix_dominators (CDI_DOMINATORS, dom_bbs, n_dom_bbs);
-
- free (body);
- free (seen);
- free (dom_bbs);
+ scale_loop_frequencies (loop, false_scale, REG_BR_PROB_BASE);
+ scale_loop_frequencies (succ_bb->loop_father, true_scale, REG_BR_PROB_BASE);
+ update_dominators_in_loop (loop);
return loop;
}
if (body[i]->loop_father == loop)
{
remove_bb_from_loops (body[i]);
- add_bb_to_loop (body[i], loop->outer);
+ add_bb_to_loop (body[i], loop_outer (loop));
}
free(body);
{
ploop = loop->inner;
flow_loop_tree_node_remove (ploop);
- flow_loop_tree_node_add (loop->outer, ploop);
+ flow_loop_tree_node_add (loop_outer (loop), ploop);
}
/* Remove the loop and free its data. */
fix_bb_placements (latch, &dummy);
}
-/* Fix placement of LOOP inside loop tree, i.e. find the innermost superloop
- FATHER of LOOP such that all of the edges coming out of LOOP belong to
- FATHER, and set it as outer loop of LOOP. Return true if placement of
- LOOP changed. */
-
-int
-fix_loop_placement (struct loop *loop)
-{
- basic_block *body;
- unsigned i;
- edge e;
- edge_iterator ei;
- struct loop *father = loop->pred[0], *act;
-
- body = get_loop_body (loop);
- for (i = 0; i < loop->num_nodes; i++)
- FOR_EACH_EDGE (e, ei, body[i]->succs)
- if (!flow_bb_inside_loop_p (loop, e->dest))
- {
- act = find_common_loop (loop, e->dest->loop_father);
- if (flow_loop_nested_p (father, act))
- father = act;
- }
- free (body);
-
- if (father != loop->outer)
- {
- for (act = loop->outer; act != father; act = act->outer)
- act->num_nodes -= loop->num_nodes;
- flow_loop_tree_node_remove (loop);
- flow_loop_tree_node_add (father, loop);
- return 1;
- }
- return 0;
-}
-
/* Fix placement of superloops of LOOP inside loop tree, i.e. ensure that
condition stated in description of fix_loop_placement holds for them.
It is used in case when we removed some edges coming out of LOOP, which
{
struct loop *outer;
- while (loop->outer)
+ while (loop_outer (loop))
{
- outer = loop->outer;
+ outer = loop_outer (loop);
if (!fix_loop_placement (loop))
break;
}
}
-/* Creates place for a new LOOP in loops structure. */
-static void
-place_new_loop (struct loop *loop)
-{
- loop->num = number_of_loops ();
- VEC_safe_push (loop_p, heap, current_loops->larray, loop);
-}
-
/* Copies copy of LOOP as subloop of TARGET loop, placing newly
created loop into loops structure. */
struct loop *
duplicate_loop (struct loop *loop, struct loop *target)
{
struct loop *cloop;
- cloop = XCNEW (struct loop);
+ cloop = alloc_loop ();
place_new_loop (cloop);
/* Mark the new loop as copy of LOOP. */
- loop->copy = cloop;
+ set_loop_copy (loop, cloop);
/* Add it to target. */
flow_loop_tree_node_add (target, cloop);
redirect_edge_and_branch_force (e, dest);
}
-/* Deletes edge E from a branch if possible. Unless REALLY_DELETE is set,
- just test whether it is possible to remove the edge. */
-static bool
-loop_delete_branch_edge (edge e, int really_delete)
-{
- basic_block src = e->src;
- basic_block newdest;
- int irr;
- edge snd;
-
- gcc_assert (EDGE_COUNT (src->succs) > 1);
-
- /* Cannot handle more than two exit edges. */
- if (EDGE_COUNT (src->succs) > 2)
- return false;
- /* And it must be just a simple branch. */
- if (!any_condjump_p (BB_END (src)))
- return false;
-
- snd = e == EDGE_SUCC (src, 0) ? EDGE_SUCC (src, 1) : EDGE_SUCC (src, 0);
- newdest = snd->dest;
- if (newdest == EXIT_BLOCK_PTR)
- return false;
-
- /* Hopefully the above conditions should suffice. */
- if (!really_delete)
- return true;
-
- /* Redirecting behaves wrongly wrto this flag. */
- irr = snd->flags & EDGE_IRREDUCIBLE_LOOP;
-
- if (!redirect_edge_and_branch (e, newdest))
- return false;
- single_succ_edge (src)->flags &= ~EDGE_IRREDUCIBLE_LOOP;
- single_succ_edge (src)->flags |= irr;
-
- return true;
-}
-
/* Check whether LOOP's body can be duplicated. */
bool
-can_duplicate_loop_p (struct loop *loop)
+can_duplicate_loop_p (const struct loop *loop)
{
int ret;
basic_block *bbs = get_loop_body (loop);
return ret;
}
-/* The NBBS blocks in BBS will get duplicated and the copies will be placed
- to LOOP. Update the single_exit information in superloops of LOOP. */
+/* Sets probability and count of edge E to zero. The probability and count
+ is redistributed evenly to the remaining edges coming from E->src. */
static void
-update_single_exits_after_duplication (basic_block *bbs, unsigned nbbs,
- struct loop *loop)
+set_zero_probability (edge e)
{
- unsigned i;
+ basic_block bb = e->src;
+ edge_iterator ei;
+ edge ae, last = NULL;
+ unsigned n = EDGE_COUNT (bb->succs);
+ gcov_type cnt = e->count, cnt1;
+ unsigned prob = e->probability, prob1;
- for (i = 0; i < nbbs; i++)
- bbs[i]->flags |= BB_DUPLICATED;
+ gcc_assert (n > 1);
+ cnt1 = cnt / (n - 1);
+ prob1 = prob / (n - 1);
- for (; loop->outer; loop = loop->outer)
+ FOR_EACH_EDGE (ae, ei, bb->succs)
{
- if (!single_exit (loop))
+ if (ae == e)
continue;
- if (single_exit (loop)->src->flags & BB_DUPLICATED)
- set_single_exit (loop, NULL);
+ ae->probability += prob1;
+ ae->count += cnt1;
+ last = ae;
}
- for (i = 0; i < nbbs; i++)
- bbs[i]->flags &= ~BB_DUPLICATED;
-}
-
-/* Updates single exit information for the copy of LOOP. */
-
-static void
-update_single_exit_for_duplicated_loop (struct loop *loop)
-{
- struct loop *copy = loop->copy;
- basic_block src, dest;
- edge exit = single_exit (loop);
-
- if (!exit)
- return;
-
- src = get_bb_copy (exit->src);
- dest = exit->dest;
- if (dest->flags & BB_DUPLICATED)
- dest = get_bb_copy (dest);
-
- exit = find_edge (src, dest);
- gcc_assert (exit != NULL);
- set_single_exit (copy, exit);
-}
-
-/* Updates single exit information for copies of ORIG_LOOPS and their subloops.
- N is the number of the loops in the ORIG_LOOPS array. */
-
-static void
-update_single_exit_for_duplicated_loops (struct loop *orig_loops[], unsigned n)
-{
- unsigned i;
+ /* Move the rest to one of the edges. */
+ last->probability += prob % (n - 1);
+ last->count += cnt % (n - 1);
- for (i = 0; i < n; i++)
- update_single_exit_for_duplicated_loop (orig_loops[i]);
+ e->probability = 0;
+ e->count = 0;
}
/* Duplicates body of LOOP to given edge E NDUPL times. Takes care of updating
unsigned i, j, n;
int is_latch = (latch == e->src);
int scale_act = 0, *scale_step = NULL, scale_main = 0;
+ int scale_after_exit = 0;
int p, freq_in, freq_le, freq_out_orig;
int prob_pass_thru, prob_pass_wont_exit, prob_pass_main;
int add_irreducible_flag;
basic_block place_after;
+ bitmap bbs_to_scale = NULL;
+ bitmap_iterator bi;
gcc_assert (e->dest == loop->header);
gcc_assert (ndupl > 0);
prob_pass_wont_exit =
RDIV (REG_BR_PROB_BASE * (freq_le + freq_out_orig), freq_in);
+ if (orig
+ && REG_BR_PROB_BASE - orig->probability != 0)
+ {
+ /* The blocks that are dominated by a removed exit edge ORIG have
+ frequencies scaled by this. */
+ scale_after_exit = RDIV (REG_BR_PROB_BASE * REG_BR_PROB_BASE,
+ REG_BR_PROB_BASE - orig->probability);
+ bbs_to_scale = BITMAP_ALLOC (NULL);
+ for (i = 0; i < n; i++)
+ {
+ if (bbs[i] != orig->src
+ && dominated_by_p (CDI_DOMINATORS, bbs[i], orig->src))
+ bitmap_set_bit (bbs_to_scale, i);
+ }
+ }
+
scale_step = XNEWVEC (int, ndupl);
- for (i = 1; i <= ndupl; i++)
- scale_step[i - 1] = TEST_BIT (wont_exit, i)
+ for (i = 1; i <= ndupl; i++)
+ scale_step[i - 1] = TEST_BIT (wont_exit, i)
? prob_pass_wont_exit
: prob_pass_thru;
for (aloop = loop->inner, i = 0; aloop; aloop = aloop->next, i++)
orig_loops[i] = aloop;
- loop->copy = target;
+ set_loop_copy (loop, target);
first_active = XNEWVEC (basic_block, n);
if (is_latch)
first_active_latch = latch;
}
- /* Update the information about single exits. */
- if (current_loops->state & LOOPS_HAVE_MARKED_SINGLE_EXITS)
- update_single_exits_after_duplication (bbs, n, target);
-
spec_edges[SE_ORIG] = orig;
spec_edges[SE_LATCH] = latch_edge;
place_after);
place_after = new_spec_edges[SE_LATCH]->src;
- if (current_loops->state & LOOPS_HAVE_MARKED_SINGLE_EXITS)
- {
- for (i = 0; i < n; i++)
- bbs[i]->flags |= BB_DUPLICATED;
- update_single_exit_for_duplicated_loops (orig_loops, n_orig_loops);
- for (i = 0; i < n; i++)
- bbs[i]->flags &= ~BB_DUPLICATED;
- }
-
if (flags & DLTHE_RECORD_COPY_NUMBER)
for (i = 0; i < n; i++)
{
{
if (to_remove)
VEC_safe_push (edge, heap, *to_remove, new_spec_edges[SE_ORIG]);
+ set_zero_probability (new_spec_edges[SE_ORIG]);
+
+ /* Scale the frequencies of the blocks dominated by the exit. */
+ if (bbs_to_scale)
+ {
+ EXECUTE_IF_SET_IN_BITMAP (bbs_to_scale, 0, i, bi)
+ {
+ scale_bbs_frequencies_int (new_bbs + i, 1, scale_after_exit,
+ REG_BR_PROB_BASE);
+ }
+ }
}
/* Record the first copy in the control flow order if it is not
{
if (to_remove)
VEC_safe_push (edge, heap, *to_remove, orig);
+ set_zero_probability (orig);
+
+ /* Scale the frequencies of the blocks dominated by the exit. */
+ if (bbs_to_scale)
+ {
+ EXECUTE_IF_SET_IN_BITMAP (bbs_to_scale, 0, i, bi)
+ {
+ scale_bbs_frequencies_int (bbs + i, 1, scale_after_exit,
+ REG_BR_PROB_BASE);
+ }
+ }
}
/* Update the original loop. */
/* Update dominators of outer blocks if affected. */
for (i = 0; i < n; i++)
{
- basic_block dominated, dom_bb, *dom_bbs;
- int n_dom_bbs,j;
+ basic_block dominated, dom_bb;
+ VEC (basic_block, heap) *dom_bbs;
+ unsigned j;
bb = bbs[i];
bb->aux = 0;
- n_dom_bbs = get_dominated_by (CDI_DOMINATORS, bb, &dom_bbs);
- for (j = 0; j < n_dom_bbs; j++)
+ dom_bbs = get_dominated_by (CDI_DOMINATORS, bb);
+ for (j = 0; VEC_iterate (basic_block, dom_bbs, j, dominated); j++)
{
- dominated = dom_bbs[j];
if (flow_bb_inside_loop_p (loop, dominated))
continue;
dom_bb = nearest_common_dominator (
CDI_DOMINATORS, first_active[i], first_active_latch);
set_immediate_dominator (CDI_DOMINATORS, dominated, dom_bb);
}
- free (dom_bbs);
+ VEC_free (basic_block, heap, dom_bbs);
}
free (first_active);
free (bbs);
+ BITMAP_FREE (bbs_to_scale);
return true;
}
MFB_KJ_EDGE to the entry part. E is the edge for that we should decide
whether to redirect it. */
-static edge mfb_kj_edge;
-static bool
+edge mfb_kj_edge;
+bool
mfb_keep_just (edge e)
{
return e != mfb_kj_edge;
}
+/* True when a candidate preheader BLOCK has predecessors from LOOP. */
+
+static bool
+has_preds_from_loop (basic_block block, struct loop *loop)
+{
+ edge e;
+ edge_iterator ei;
+
+ FOR_EACH_EDGE (e, ei, block->preds)
+ if (e->src->loop_father == loop)
+ return true;
+ return false;
+}
+
/* Creates a pre-header for a LOOP. Returns newly created block. Unless
CP_SIMPLE_PREHEADERS is set in FLAGS, we only force LOOP to have single
entry; otherwise we also force preheader block to have only one successor.
+ When CP_FALLTHRU_PREHEADERS is set in FLAGS, we force the preheader block
+ to be a fallthru predecessor to the loop header and to have only
+ predecessors from outside of the loop.
The function also updates dominators. */
-static basic_block
+basic_block
create_preheader (struct loop *loop, int flags)
{
edge e, fallthru;
int nentry = 0;
bool irred = false;
bool latch_edge_was_fallthru;
- edge one_succ_pred = 0;
+ edge one_succ_pred = NULL, single_entry = NULL;
edge_iterator ei;
FOR_EACH_EDGE (e, ei, loop->header->preds)
continue;
irred |= (e->flags & EDGE_IRREDUCIBLE_LOOP) != 0;
nentry++;
+ single_entry = e;
if (single_succ_p (e->src))
one_succ_pred = e;
}
gcc_assert (nentry);
if (nentry == 1)
{
- /* Get an edge that is different from the one from loop->latch
- to loop->header. */
- e = EDGE_PRED (loop->header,
- EDGE_PRED (loop->header, 0)->src == loop->latch);
-
- if (!(flags & CP_SIMPLE_PREHEADERS) || single_succ_p (e->src))
+ bool need_forwarder_block = false;
+
+ /* We do not allow entry block to be the loop preheader, since we
+ cannot emit code there. */
+ if (single_entry->src == ENTRY_BLOCK_PTR)
+ need_forwarder_block = true;
+ else
+ {
+ /* If we want simple preheaders, also force the preheader to have
+ just a single successor. */
+ if ((flags & CP_SIMPLE_PREHEADERS)
+ && !single_succ_p (single_entry->src))
+ need_forwarder_block = true;
+ /* If we want fallthru preheaders, also create forwarder block when
+ preheader ends with a jump or has predecessors from loop. */
+ else if ((flags & CP_FALLTHRU_PREHEADERS)
+ && (JUMP_P (BB_END (single_entry->src))
+ || has_preds_from_loop (single_entry->src, loop)))
+ need_forwarder_block = true;
+ }
+ if (! need_forwarder_block)
return NULL;
}
if (dump_file)
fprintf (dump_file, "Created preheader block for loop %i\n",
loop->num);
+
+ if (flags & CP_FALLTHRU_PREHEADERS)
+ gcc_assert ((single_succ_edge (dummy)->flags & EDGE_FALLTHRU)
+ && !JUMP_P (BB_END (dummy)));
return dummy;
}
loop_iterator li;
struct loop *loop;
+ if (!current_loops)
+ return;
+
FOR_EACH_LOOP (li, loop, 0)
create_preheader (loop, flags);
- current_loops->state |= LOOPS_HAVE_PREHEADERS;
+ loops_state_set (LOOPS_HAVE_PREHEADERS);
}
/* Forces all loop latches to have only single successor. */
split_edge (e);
}
- current_loops->state |= LOOPS_HAVE_SIMPLE_LATCHES;
+ loops_state_set (LOOPS_HAVE_SIMPLE_LATCHES);
}
/* This function is called from loop_version. It splits the entry edge
--- edge e ---> [cond expr] ---> [first_head]
|
+---------> [second_head]
-*/
+
+ THEN_PROB is the probability of then branch of the condition. */
static basic_block
-lv_adjust_loop_entry_edge (basic_block first_head,
- basic_block second_head,
- edge e,
- void *cond_expr)
+lv_adjust_loop_entry_edge (basic_block first_head, basic_block second_head,
+ edge e, void *cond_expr, unsigned then_prob)
{
basic_block new_head = NULL;
edge e1;
insert conditional expr. */
new_head = split_edge (e);
-
lv_add_condition_to_bb (first_head, second_head, new_head,
cond_expr);
/* Don't set EDGE_TRUE_VALUE in RTL mode, as it's invalid there. */
+ e = single_succ_edge (new_head);
e1 = make_edge (new_head, first_head,
current_ir_type () == IR_GIMPLE ? EDGE_TRUE_VALUE : 0);
+ e1->probability = then_prob;
+ e->probability = REG_BR_PROB_BASE - then_prob;
+ e1->count = RDIV (e->count * e1->probability, REG_BR_PROB_BASE);
+ e->count = RDIV (e->count * e->probability, REG_BR_PROB_BASE);
+
set_immediate_dominator (CDI_DOMINATORS, first_head, new_head);
set_immediate_dominator (CDI_DOMINATORS, second_head, new_head);
may be a run time test for things that were not resolved by static
analysis (overlapping ranges (anti-aliasing), alignment, etc.).
+ THEN_PROB is the probability of the then edge of the if. THEN_SCALE
+ is the ratio by that the frequencies in the original loop should
+ be scaled. ELSE_SCALE is the ratio by that the frequencies in the
+ new loop should be scaled.
+
If PLACE_AFTER is true, we place the new loop after LOOP in the
instruction stream, otherwise it is placed before LOOP. */
struct loop *
loop_version (struct loop *loop,
void *cond_expr, basic_block *condition_bb,
+ unsigned then_prob, unsigned then_scale, unsigned else_scale,
bool place_after)
{
basic_block first_head, second_head;
- edge entry, latch_edge, exit, true_edge, false_edge;
+ edge entry, latch_edge, true_edge, false_edge;
int irred_flag;
struct loop *nloop;
basic_block cond_bb;
- /* CHECKME: Loop versioning does not handle nested loop at this point. */
- if (loop->inner)
- return NULL;
-
/* Record entry and latch edges for the loop */
entry = loop_preheader_edge (loop);
irred_flag = entry->flags & EDGE_IRREDUCIBLE_LOOP;
/* Split loop entry edge and insert new block with cond expr. */
cond_bb = lv_adjust_loop_entry_edge (first_head, second_head,
- entry, cond_expr);
+ entry, cond_expr, then_prob);
if (condition_bb)
*condition_bb = cond_bb;
nloop = loopify (latch_edge,
single_pred_edge (get_bb_copy (loop->header)),
cond_bb, true_edge, false_edge,
- false /* Do not redirect all edges. */);
-
- exit = single_exit (loop);
- if (exit)
- set_single_exit (nloop, find_edge (get_bb_copy (exit->src), exit->dest));
+ false /* Do not redirect all edges. */,
+ then_scale, else_scale);
/* loopify redirected latch_edge. Update its PENDING_STMTS. */
lv_flush_pending_stmts (latch_edge);
free (bbs);
}
- /* At this point condition_bb is loop predheader with two successors,
- first_head and second_head. Make sure that loop predheader has only
+ /* At this point condition_bb is loop preheader with two successors,
+ first_head and second_head. Make sure that loop preheader has only
one successor. */
split_edge (loop_preheader_edge (loop));
split_edge (loop_preheader_edge (nloop));
removed (thus the loop nesting may be wrong), and some blocks and edges
were changed (so the information about bb --> loop mapping does not have
to be correct). But still for the remaining loops the header dominates
- the latch, and loops did not get new subloobs (new loops might possibly
+ the latch, and loops did not get new subloops (new loops might possibly
get created, but we are not interested in them). Fix up the mess.
If CHANGED_BBS is not NULL, basic blocks whose loop has changed are
basic_block bb;
struct loop *loop, *ploop;
loop_iterator li;
+ bool record_exits = false;
+ struct loop **superloop = XNEWVEC (struct loop *, number_of_loops ());
- /* Remove the old bb -> loop mapping. */
+ /* Remove the old bb -> loop mapping. Remember the depth of the blocks in
+ the loop hierarchy, so that we can recognize blocks whose loop nesting
+ relationship has changed. */
FOR_EACH_BB (bb)
{
- bb->aux = (void *) (size_t) bb->loop_father->depth;
+ if (changed_bbs)
+ bb->aux = (void *) (size_t) loop_depth (bb->loop_father);
bb->loop_father = current_loops->tree_root;
}
- /* Remove the dead loops from structures. */
- current_loops->tree_root->num_nodes = n_basic_blocks;
- FOR_EACH_LOOP (li, loop, 0)
+ if (loops_state_satisfies_p (LOOPS_HAVE_RECORDED_EXITS))
+ {
+ release_recorded_exits ();
+ record_exits = true;
+ }
+
+ /* Remove the dead loops from structures. We start from the innermost
+ loops, so that when we remove the loops, we know that the loops inside
+ are preserved, and do not waste time relinking loops that will be
+ removed later. */
+ FOR_EACH_LOOP (li, loop, LI_FROM_INNERMOST)
{
- loop->num_nodes = 0;
if (loop->header)
continue;
{
ploop = loop->inner;
flow_loop_tree_node_remove (ploop);
- flow_loop_tree_node_add (loop->outer, ploop);
+ flow_loop_tree_node_add (loop_outer (loop), ploop);
}
/* Remove the loop and free its data. */
delete_loop (loop);
}
- /* Rescan the bodies of loops, starting from the outermost. */
+ /* Rescan the bodies of loops, starting from the outermost ones. We assume
+ that no optimization interchanges the order of the loops, i.e., it cannot
+ happen that L1 was superloop of L2 before and it is subloop of L2 now
+ (without explicitly updating loop information). At the same time, we also
+ determine the new loop structure. */
+ current_loops->tree_root->num_nodes = n_basic_blocks;
FOR_EACH_LOOP (li, loop, 0)
{
+ superloop[loop->num] = loop->header->loop_father;
loop->num_nodes = flow_loop_nodes_find (loop->header, loop);
}
/* Now fix the loop nesting. */
FOR_EACH_LOOP (li, loop, 0)
{
- bb = loop_preheader_edge (loop)->src;
- if (bb->loop_father != loop->outer)
+ ploop = superloop[loop->num];
+ if (ploop != loop_outer (loop))
{
flow_loop_tree_node_remove (loop);
- flow_loop_tree_node_add (bb->loop_father, loop);
+ flow_loop_tree_node_add (ploop, loop);
}
}
+ free (superloop);
/* Mark the blocks whose loop has changed. */
- FOR_EACH_BB (bb)
+ if (changed_bbs)
{
- if (changed_bbs
- && (void *) (size_t) bb->loop_father->depth != bb->aux)
- bitmap_set_bit (changed_bbs, bb->index);
+ FOR_EACH_BB (bb)
+ {
+ if ((void *) (size_t) loop_depth (bb->loop_father) != bb->aux)
+ bitmap_set_bit (changed_bbs, bb->index);
- bb->aux = NULL;
+ bb->aux = NULL;
+ }
}
- if (current_loops->state & LOOPS_HAVE_MARKED_SINGLE_EXITS)
- mark_single_exit_loops ();
- if (current_loops->state & LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS)
+ if (loops_state_satisfies_p (LOOPS_HAVE_PREHEADERS))
+ create_preheaders (CP_SIMPLE_PREHEADERS);
+
+ if (loops_state_satisfies_p (LOOPS_HAVE_SIMPLE_LATCHES))
+ force_single_succ_latches ();
+
+ if (loops_state_satisfies_p (LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS))
mark_irreducible_loops ();
+
+ if (record_exits)
+ record_loop_exits ();
+
+#ifdef ENABLE_CHECKING
+ verify_loop_structure ();
+#endif
}
OSDN Git Service
