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, 59 Temple Place - Suite 330, Boston, MA
-02111-1307, USA. */
+Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA
+02110-1301, USA. */
#include "config.h"
#include "system.h"
#include "basic-block.h"
#include "cfgloop.h"
#include "cfglayout.h"
+#include "cfghooks.h"
#include "output.h"
static void duplicate_subloops (struct loops *, struct loop *, struct loop *);
#define RDIV(X,Y) (((X) + (Y) / 2) / (Y))
-/* Splits basic block BB after INSN, returns created edge. Updates loops
- and dominators. */
-edge
-split_loop_bb (basic_block bb, void *insn)
-{
- edge e;
-
- /* Split the block. */
- e = split_block (bb, insn);
-
- /* Add dest to loop. */
- add_bb_to_loop (e->dest, e->src->loop_father);
-
- return e;
-}
-
/* Checks whether basic block BB is dominated by DATA. */
static bool
rpe_enum_p (basic_block bb, void *data)
gcc_assert (EDGE_COUNT (e->dest->preds) <= 1);
/* Find bbs in the path. */
- *bbs = xcalloc (n_basic_blocks, sizeof (basic_block));
+ *bbs = XCNEWVEC (basic_block, n_basic_blocks);
return dfs_enumerate_from (e->dest, 0, rpe_enum_p, *bbs,
n_basic_blocks, e->dest);
}
/* Prevent us from going out of the base_loop. */
SET_BIT (in_queue, base_loop->header->index);
- queue = xmalloc ((base_loop->num_nodes + 1) * sizeof (basic_block));
+ queue = XNEWVEC (basic_block, base_loop->num_nodes + 1);
qtop = queue + base_loop->num_nodes + 1;
qbeg = queue;
qend = queue + 1;
int stack_top;
sbitmap on_stack;
edge *edges, e;
- unsigned n_edges, i;
+ unsigned num_edges, i;
if (!(from->flags & BB_IRREDUCIBLE_LOOP))
return;
on_stack = sbitmap_alloc (last_basic_block);
sbitmap_zero (on_stack);
SET_BIT (on_stack, from->index);
- stack = xmalloc (from->loop_father->num_nodes * sizeof (basic_block));
+ stack = XNEWVEC (basic_block, from->loop_father->num_nodes);
stack[0] = from;
stack_top = 1;
bb->flags &= ~BB_IRREDUCIBLE_LOOP;
if (bb->loop_father->header == bb)
- edges = get_loop_exit_edges (bb->loop_father, &n_edges);
+ edges = get_loop_exit_edges (bb->loop_father, &num_edges);
else
{
- n_edges = EDGE_COUNT (bb->succs);
- edges = xmalloc (n_edges * sizeof (edge));
+ num_edges = EDGE_COUNT (bb->succs);
+ edges = XNEWVEC (edge, num_edges);
FOR_EACH_EDGE (e, ei, bb->succs)
edges[ei.index] = e;
}
- for (i = 0; i < n_edges; i++)
+ for (i = 0; i < num_edges; i++)
{
e = edges[i];
nrem = find_path (e, &rem_bbs);
n_bord_bbs = 0;
- bord_bbs = xcalloc (n_basic_blocks, sizeof (basic_block));
+ bord_bbs = XCNEWVEC (basic_block, n_basic_blocks);
seen = sbitmap_alloc (last_basic_block);
sbitmap_zero (seen);
from = e->src;
deleted = loop_delete_branch_edge (e, 1);
gcc_assert (deleted);
- dom_bbs = xcalloc (n_basic_blocks, sizeof (basic_block));
+ dom_bbs = XCNEWVEC (basic_block, n_basic_blocks);
/* Cancel loops contained in the path. */
for (i = 0; i < nrem; i++)
loop->level = 1;
/* Find its nodes. */
- bbs = xcalloc (n_basic_blocks, sizeof (basic_block));
+ bbs = XCNEWVEC (basic_block, n_basic_blocks);
n = dfs_enumerate_from (loop->latch, 1, alp_enum_p,
bbs, n_basic_blocks, loop->header);
Returns newly created loop. */
struct loop *
-loopify (struct loops *loops, edge latch_edge, edge header_edge,
+loopify (struct loops *loops, edge latch_edge, edge header_edge,
basic_block switch_bb, edge true_edge, edge false_edge,
bool redirect_all_edges)
{
basic_block *dom_bbs, *body;
unsigned n_dom_bbs, i;
sbitmap seen;
- struct loop *loop = xcalloc (1, sizeof (struct loop));
+ struct loop *loop = XCNEW (struct loop);
struct loop *outer = succ_bb->loop_father->outer;
int freq, prob, tot_prob;
gcov_type cnt;
if (redirect_all_edges)
{
loop_redirect_edge (header_edge, switch_bb);
- loop_redirect_edge (false_edge, loop->header);
-
+ loop_redirect_edge (false_edge, loop->header);
+
/* Update dominators. */
set_immediate_dominator (CDI_DOMINATORS, switch_bb, pred_bb);
set_immediate_dominator (CDI_DOMINATORS, loop->header, switch_bb);
scale_loop_frequencies (succ_bb->loop_father, tot_prob - prob, tot_prob);
/* Update dominators of blocks outside of LOOP. */
- dom_bbs = xcalloc (n_basic_blocks, sizeof (basic_block));
+ dom_bbs = XCNEWVEC (basic_block, n_basic_blocks);
n_dom_bbs = 0;
seen = sbitmap_alloc (last_basic_block);
sbitmap_zero (seen);
unsigned i, n;
basic_block latch = loop->latch;
edge *edges;
- unsigned n_edges;
+ unsigned num_edges;
/* This is relatively straightforward. The dominators are unchanged, as
loop header dominates loop latch, so the only thing we have to care of
its work. */
body = get_loop_body (loop);
- edges = get_loop_exit_edges (loop, &n_edges);
+ edges = get_loop_exit_edges (loop, &num_edges);
n = loop->num_nodes;
for (i = 0; i < n; i++)
if (body[i]->loop_father == loop)
update the irreducible marks inside its body. While it is certainly
possible to do, it is a bit complicated and this situation should be
very rare, so we just remark all loops in this case. */
- for (i = 0; i < n_edges; i++)
+ for (i = 0; i < num_edges; i++)
if (edges[i]->flags & EDGE_IRREDUCIBLE_LOOP)
break;
- if (i != n_edges)
+ if (i != num_edges)
mark_irreducible_loops (loops);
free (edges);
}
{
outer = loop->outer;
if (!fix_loop_placement (loop))
- break;
+ break;
/* Changing the placement of a loop in the loop tree may alter the
validity of condition 2) of the description of fix_bb_placement
duplicate_loop (struct loops *loops, struct loop *loop, struct loop *target)
{
struct loop *cloop;
- cloop = xcalloc (1, sizeof (struct loop));
+ cloop = XCNEW (struct loop);
place_new_loop (loops, cloop);
/* Initialize copied loop. */
edge snd;
gcc_assert (EDGE_COUNT (src->succs) > 1);
-
+
/* Cannot handle more than two exit edges. */
if (EDGE_COUNT (src->succs) > 2)
return false;
return false;
single_succ_edge (src)->flags &= ~EDGE_IRREDUCIBLE_LOOP;
single_succ_edge (src)->flags |= irr;
-
+
return true;
}
ret = can_copy_bbs_p (bbs, loop->num_nodes);
free (bbs);
-
+
return ret;
}
unsigned i;
for (i = 0; i < nbbs; i++)
- bbs[i]->rbi->duplicated = 1;
+ bbs[i]->flags |= BB_DUPLICATED;
for (; loop->outer; loop = loop->outer)
{
if (!loop->single_exit)
continue;
- if (loop->single_exit->src->rbi->duplicated)
+ if (loop->single_exit->src->flags & BB_DUPLICATED)
loop->single_exit = NULL;
}
for (i = 0; i < nbbs; i++)
- bbs[i]->rbi->duplicated = 0;
+ bbs[i]->flags &= ~BB_DUPLICATED;
}
/* Duplicates body of LOOP to given edge E NDUPL times. Takes care of updating
original LOOP body, the other copies are numbered in order given by control
flow through them) into TO_REMOVE array. Returns false if duplication is
impossible. */
-int
+bool
duplicate_loop_to_header_edge (struct loop *loop, edge e, struct loops *loops,
unsigned int ndupl, sbitmap wont_exit,
edge orig, edge *to_remove,
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;
gcc_assert (e->dest == loop->header);
gcc_assert (ndupl > 0);
gcc_assert (!flow_bb_inside_loop_p (loop, orig->dest));
}
- bbs = get_loop_body (loop);
+ n = loop->num_nodes;
+ bbs = get_loop_body_in_dom_order (loop);
+ gcc_assert (bbs[0] == loop->header);
+ gcc_assert (bbs[n - 1] == loop->latch);
/* Check whether duplication is possible. */
if (!can_copy_bbs_p (bbs, loop->num_nodes))
free (bbs);
return false;
}
- new_bbs = xmalloc (sizeof (basic_block) * loop->num_nodes);
+ new_bbs = XNEWVEC (basic_block, loop->num_nodes);
/* In case we are doing loop peeling and the loop is in the middle of
irreducible region, the peeled copies will be inside it too. */
prob_pass_wont_exit =
RDIV (REG_BR_PROB_BASE * (freq_le + freq_out_orig), freq_in);
- scale_step = xmalloc (ndupl * sizeof (int));
+ scale_step = XNEWVEC (int, ndupl);
for (i = 1; i <= ndupl; i++)
scale_step[i - 1] = TEST_BIT (wont_exit, i)
? prob_pass_wont_exit
: prob_pass_thru;
- if (is_latch)
+ /* Complete peeling is special as the probability of exit in last
+ copy becomes 1. */
+ if (flags & DLTHE_FLAG_COMPLETTE_PEEL)
+ {
+ int wanted_freq = EDGE_FREQUENCY (e);
+
+ if (wanted_freq > freq_in)
+ wanted_freq = freq_in;
+
+ gcc_assert (!is_latch);
+ /* First copy has frequency of incoming edge. Each subsequent
+ frequency should be reduced by prob_pass_wont_exit. Caller
+ should've managed the flags so all except for original loop
+ has won't exist set. */
+ scale_act = RDIV (wanted_freq * REG_BR_PROB_BASE, freq_in);
+ /* Now simulate the duplication adjustments and compute header
+ frequency of the last copy. */
+ for (i = 0; i < ndupl; i++)
+ wanted_freq = RDIV (wanted_freq * scale_step[i], REG_BR_PROB_BASE);
+ scale_main = RDIV (wanted_freq * REG_BR_PROB_BASE, freq_in);
+ }
+ else if (is_latch)
{
prob_pass_main = TEST_BIT (wont_exit, 0)
? prob_pass_wont_exit
n_orig_loops = 0;
for (aloop = loop->inner; aloop; aloop = aloop->next)
n_orig_loops++;
- orig_loops = xcalloc (n_orig_loops, sizeof (struct loop *));
+ orig_loops = XCNEWVEC (struct loop *, n_orig_loops);
for (aloop = loop->inner, i = 0; aloop; aloop = aloop->next, i++)
orig_loops[i] = aloop;
loop->copy = target;
- n = loop->num_nodes;
-
- first_active = xmalloc (n * sizeof (basic_block));
+ first_active = XNEWVEC (basic_block, n);
if (is_latch)
{
memcpy (first_active, bbs, n * sizeof (basic_block));
spec_edges[SE_ORIG] = orig;
spec_edges[SE_LATCH] = latch_edge;
+ place_after = e->src;
for (j = 0; j < ndupl; j++)
{
/* Copy loops. */
copy_loops_to (loops, orig_loops, n_orig_loops, target);
/* Copy bbs. */
- copy_bbs (bbs, n, new_bbs, spec_edges, 2, new_spec_edges, loop);
+ copy_bbs (bbs, n, new_bbs, spec_edges, 2, new_spec_edges, loop,
+ place_after);
+ place_after = new_spec_edges[SE_LATCH]->src;
- for (i = 0; i < n; i++)
- new_bbs[i]->rbi->copy_number = j + 1;
+ if (flags & DLTHE_RECORD_COPY_NUMBER)
+ for (i = 0; i < n; i++)
+ {
+ gcc_assert (!new_bbs[i]->aux);
+ new_bbs[i]->aux = (void *)(size_t)(j + 1);
+ }
/* Note whether the blocks and edges belong to an irreducible loop. */
if (add_irreducible_flag)
{
for (i = 0; i < n; i++)
- new_bbs[i]->rbi->duplicated = 1;
+ new_bbs[i]->flags |= BB_DUPLICATED;
for (i = 0; i < n; i++)
{
edge_iterator ei;
new_bb->flags |= BB_IRREDUCIBLE_LOOP;
FOR_EACH_EDGE (ae, ei, new_bb->succs)
- if (ae->dest->rbi->duplicated
+ if ((ae->dest->flags & BB_DUPLICATED)
&& (ae->src->loop_father == target
|| ae->dest->loop_father == target))
ae->flags |= EDGE_IRREDUCIBLE_LOOP;
}
for (i = 0; i < n; i++)
- new_bbs[i]->rbi->duplicated = 0;
+ new_bbs[i]->flags &= ~BB_DUPLICATED;
}
/* Redirect the special edges. */
redirect_edge_and_branch_force (new_spec_edges[SE_LATCH],
loop->header);
set_immediate_dominator (CDI_DOMINATORS, new_bbs[0], latch);
- latch = loop->latch = new_bbs[1];
+ latch = loop->latch = new_bbs[n - 1];
e = latch_edge = new_spec_edges[SE_LATCH];
}
else
if (!first_active_latch)
{
memcpy (first_active, new_bbs, n * sizeof (basic_block));
- first_active_latch = new_bbs[1];
+ first_active_latch = new_bbs[n - 1];
}
/* Set counts and frequencies. */
}
free (new_bbs);
free (orig_loops);
-
+
/* Update the original loop. */
if (!is_latch)
set_immediate_dominator (CDI_DOMINATORS, e->dest, e->src);
int n_dom_bbs,j;
bb = bbs[i];
- bb->rbi->copy_number = 0;
+ bb->aux = 0;
n_dom_bbs = get_dominated_by (CDI_DOMINATORS, bb, &dom_bbs);
for (j = 0; j < n_dom_bbs; j++)
continue;
dom_bb = nearest_common_dominator (
CDI_DOMINATORS, first_active[i], first_active_latch);
- set_immediate_dominator (CDI_DOMINATORS, dominated, dom_bb);
+ set_immediate_dominator (CDI_DOMINATORS, dominated, dom_bb);
}
free (dom_bbs);
}
return new_bb;
}
-/* Uses the natural loop discovery to recreate loop notes. */
-void
-create_loop_notes (void)
+/* This function is called from loop_version. It splits the entry edge
+ of the loop we want to version, adds the versioning condition, and
+ adjust the edges to the two versions of the loop appropriately.
+ e is an incoming edge. Returns the basic block containing the
+ condition.
+
+ --- edge e ---- > [second_head]
+
+ Split it and insert new conditional expression and adjust edges.
+
+ --- edge e ---> [cond expr] ---> [first_head]
+ |
+ +---------> [second_head]
+*/
+
+static basic_block
+lv_adjust_loop_entry_edge (basic_block first_head,
+ basic_block second_head,
+ edge e,
+ void *cond_expr)
{
- rtx insn, head, end;
- struct loops loops;
- struct loop *loop;
- basic_block *first, *last, bb, pbb;
- struct loop **stack, **top;
-
-#ifdef ENABLE_CHECKING
- /* Verify that there really are no loop notes. */
- for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
- gcc_assert (!NOTE_P (insn) ||
- NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_BEG);
-#endif
-
- flow_loops_find (&loops);
- free_dominance_info (CDI_DOMINATORS);
- if (loops.num > 1)
+ basic_block new_head = NULL;
+ edge e1;
+
+ gcc_assert (e->dest == second_head);
+
+ /* Split edge 'e'. This will create a new basic block, where we can
+ 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. */
+ e1 = make_edge (new_head, first_head, ir_type () ? EDGE_TRUE_VALUE : 0);
+ set_immediate_dominator (CDI_DOMINATORS, first_head, new_head);
+ set_immediate_dominator (CDI_DOMINATORS, second_head, new_head);
+
+ /* Adjust loop header phi nodes. */
+ lv_adjust_loop_header_phi (first_head, second_head, new_head, e1);
+
+ return new_head;
+}
+
+/* Main entry point for Loop Versioning transformation.
+
+ This transformation given a condition and a loop, creates
+ -if (condition) { loop_copy1 } else { loop_copy2 },
+ where loop_copy1 is the loop transformed in one way, and loop_copy2
+ is the loop transformed in another way (or unchanged). 'condition'
+ may be a run time test for things that were not resolved by static
+ analysis (overlapping ranges (anti-aliasing), alignment, etc.).
+
+ 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 loops *loops, struct loop * loop,
+ void *cond_expr, basic_block *condition_bb,
+ bool place_after)
+{
+ basic_block first_head, second_head;
+ edge entry, latch_edge, exit, 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;
+ entry->flags &= ~EDGE_IRREDUCIBLE_LOOP;
+
+ /* Note down head of loop as first_head. */
+ first_head = entry->dest;
+
+ /* Duplicate loop. */
+ if (!cfg_hook_duplicate_loop_to_header_edge (loop, entry, loops, 1,
+ NULL, NULL, NULL, NULL, 0))
+ return NULL;
+
+ /* After duplication entry edge now points to new loop head block.
+ Note down new head as second_head. */
+ second_head = entry->dest;
+
+ /* 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);
+ if (condition_bb)
+ *condition_bb = cond_bb;
+
+ if (!cond_bb)
{
- last = xcalloc (loops.num, sizeof (basic_block));
+ entry->flags |= irred_flag;
+ return NULL;
+ }
- FOR_EACH_BB (bb)
- {
- for (loop = bb->loop_father; loop->outer; loop = loop->outer)
- last[loop->num] = bb;
- }
+ latch_edge = single_succ_edge (get_bb_copy (loop->latch));
+
+ extract_cond_bb_edges (cond_bb, &true_edge, &false_edge);
+ nloop = loopify (loops,
+ latch_edge,
+ single_pred_edge (get_bb_copy (loop->header)),
+ cond_bb, true_edge, false_edge,
+ false /* Do not redirect all edges. */);
- first = xcalloc (loops.num, sizeof (basic_block));
- stack = xcalloc (loops.num, sizeof (struct loop *));
- top = stack;
+ exit = loop->single_exit;
+ if (exit)
+ nloop->single_exit = find_edge (get_bb_copy (exit->src), exit->dest);
- FOR_EACH_BB (bb)
+ /* loopify redirected latch_edge. Update its PENDING_STMTS. */
+ lv_flush_pending_stmts (latch_edge);
+
+ /* loopify redirected condition_bb's succ edge. Update its PENDING_STMTS. */
+ extract_cond_bb_edges (cond_bb, &true_edge, &false_edge);
+ lv_flush_pending_stmts (false_edge);
+ /* Adjust irreducible flag. */
+ if (irred_flag)
+ {
+ cond_bb->flags |= BB_IRREDUCIBLE_LOOP;
+ loop_preheader_edge (loop)->flags |= EDGE_IRREDUCIBLE_LOOP;
+ loop_preheader_edge (nloop)->flags |= EDGE_IRREDUCIBLE_LOOP;
+ single_pred_edge (cond_bb)->flags |= EDGE_IRREDUCIBLE_LOOP;
+ }
+
+ if (place_after)
+ {
+ basic_block *bbs = get_loop_body_in_dom_order (nloop), after;
+ unsigned i;
+
+ after = loop->latch;
+
+ for (i = 0; i < nloop->num_nodes; i++)
{
- for (loop = bb->loop_father; loop->outer; loop = loop->outer)
- {
- if (!first[loop->num])
- {
- *top++ = loop;
- first[loop->num] = bb;
- }
-
- if (bb == last[loop->num])
- {
- /* Prevent loops from overlapping. */
- while (*--top != loop)
- last[(*top)->num] = EXIT_BLOCK_PTR;
-
- /* If loop starts with jump into it, place the note in
- front of the jump. */
- insn = PREV_INSN (BB_HEAD (first[loop->num]));
- if (insn
- && BARRIER_P (insn))
- insn = PREV_INSN (insn);
-
- if (insn
- && JUMP_P (insn)
- && any_uncondjump_p (insn)
- && onlyjump_p (insn))
- {
- pbb = BLOCK_FOR_INSN (insn);
- gcc_assert (pbb && single_succ_p (pbb));
-
- if (!flow_bb_inside_loop_p (loop, single_succ (pbb)))
- insn = BB_HEAD (first[loop->num]);
- }
- else
- insn = BB_HEAD (first[loop->num]);
-
- head = BB_HEAD (first[loop->num]);
- emit_note_before (NOTE_INSN_LOOP_BEG, insn);
- BB_HEAD (first[loop->num]) = head;
-
- /* Position the note correctly wrto barrier. */
- insn = BB_END (last[loop->num]);
- if (NEXT_INSN (insn)
- && BARRIER_P (NEXT_INSN (insn)))
- insn = NEXT_INSN (insn);
-
- end = BB_END (last[loop->num]);
- emit_note_after (NOTE_INSN_LOOP_END, insn);
- BB_END (last[loop->num]) = end;
- }
- }
+ move_block_after (bbs[i], after);
+ after = bbs[i];
}
-
- free (first);
- free (last);
- free (stack);
+ free (bbs);
}
- flow_loops_free (&loops);
+
+ /* At this point condition_bb is loop predheader with two successors,
+ first_head and second_head. Make sure that loop predheader has only
+ one successor. */
+ loop_split_edge_with (loop_preheader_edge (loop), NULL);
+ loop_split_edge_with (loop_preheader_edge (nloop), NULL);
+
+ return nloop;
}
/* The structure of LOOPS might have changed. Some loops might get removed
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
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
marked in it. */
}
/* Remove the dead loops from structures. */
- loops->tree_root->num_nodes = n_basic_blocks + 2;
+ loops->tree_root->num_nodes = n_basic_blocks;
for (i = 1; i < loops->num; i++)
{
loop = loops->parray[i];
bb->aux = NULL;
}
- mark_single_exit_loops (loops);
- mark_irreducible_loops (loops);
+ if (loops->state & LOOPS_HAVE_MARKED_SINGLE_EXITS)
+ mark_single_exit_loops (loops);
+ if (loops->state & LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS)
+ mark_irreducible_loops (loops);
}