2 * Copyright 2000-2009 Sun Microsystems, Inc. All Rights Reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
25 #include "incls/_precompiled.incl"
26 #include "incls/_loopTransform.cpp.incl"
28 //------------------------------is_loop_exit-----------------------------------
29 // Given an IfNode, return the loop-exiting projection or NULL if both
30 // arms remain in the loop.
31 Node *IdealLoopTree::is_loop_exit(Node *iff) const {
32 if( iff->outcnt() != 2 ) return NULL; // Ignore partially dead tests
33 PhaseIdealLoop *phase = _phase;
34 // Test is an IfNode, has 2 projections. If BOTH are in the loop
35 // we need loop unswitching instead of peeling.
36 if( !is_member(phase->get_loop( iff->raw_out(0) )) )
37 return iff->raw_out(0);
38 if( !is_member(phase->get_loop( iff->raw_out(1) )) )
39 return iff->raw_out(1);
44 //=============================================================================
47 //------------------------------record_for_igvn----------------------------
48 // Put loop body on igvn work list
49 void IdealLoopTree::record_for_igvn() {
50 for( uint i = 0; i < _body.size(); i++ ) {
51 Node *n = _body.at(i);
52 _phase->_igvn._worklist.push(n);
56 //------------------------------compute_profile_trip_cnt----------------------------
57 // Compute loop trip count from profile data as
58 // (backedge_count + loop_exit_count) / loop_exit_count
59 void IdealLoopTree::compute_profile_trip_cnt( PhaseIdealLoop *phase ) {
60 if (!_head->is_CountedLoop()) {
63 CountedLoopNode* head = _head->as_CountedLoop();
64 if (head->profile_trip_cnt() != COUNT_UNKNOWN) {
65 return; // Already computed
67 float trip_cnt = (float)max_jint; // default is big
69 Node* back = head->in(LoopNode::LoopBackControl);
70 while (back != head) {
71 if ((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) &&
73 back->in(0)->is_If() &&
74 back->in(0)->as_If()->_fcnt != COUNT_UNKNOWN &&
75 back->in(0)->as_If()->_prob != PROB_UNKNOWN) {
78 back = phase->idom(back);
81 assert((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) &&
82 back->in(0), "if-projection exists");
83 IfNode* back_if = back->in(0)->as_If();
84 float loop_back_cnt = back_if->_fcnt * back_if->_prob;
86 // Now compute a loop exit count
87 float loop_exit_cnt = 0.0f;
88 for( uint i = 0; i < _body.size(); i++ ) {
91 IfNode *iff = n->as_If();
92 if( iff->_fcnt != COUNT_UNKNOWN && iff->_prob != PROB_UNKNOWN ) {
93 Node *exit = is_loop_exit(iff);
95 float exit_prob = iff->_prob;
96 if (exit->Opcode() == Op_IfFalse) exit_prob = 1.0 - exit_prob;
97 if (exit_prob > PROB_MIN) {
98 float exit_cnt = iff->_fcnt * exit_prob;
99 loop_exit_cnt += exit_cnt;
105 if (loop_exit_cnt > 0.0f) {
106 trip_cnt = (loop_back_cnt + loop_exit_cnt) / loop_exit_cnt;
108 // No exit count so use
109 trip_cnt = loop_back_cnt;
113 if (TraceProfileTripCount) {
114 tty->print_cr("compute_profile_trip_cnt lp: %d cnt: %f\n", head->_idx, trip_cnt);
117 head->set_profile_trip_cnt(trip_cnt);
120 //---------------------is_invariant_addition-----------------------------
121 // Return nonzero index of invariant operand for an Add or Sub
122 // of (nonconstant) invariant and variant values. Helper for reassociate_invariants.
123 int IdealLoopTree::is_invariant_addition(Node* n, PhaseIdealLoop *phase) {
124 int op = n->Opcode();
125 if (op == Op_AddI || op == Op_SubI) {
126 bool in1_invar = this->is_invariant(n->in(1));
127 bool in2_invar = this->is_invariant(n->in(2));
128 if (in1_invar && !in2_invar) return 1;
129 if (!in1_invar && in2_invar) return 2;
134 //---------------------reassociate_add_sub-----------------------------
135 // Reassociate invariant add and subtract expressions:
137 // inv1 + (x + inv2) => ( inv1 + inv2) + x
138 // (x + inv2) + inv1 => ( inv1 + inv2) + x
139 // inv1 + (x - inv2) => ( inv1 - inv2) + x
140 // inv1 - (inv2 - x) => ( inv1 - inv2) + x
141 // (x + inv2) - inv1 => (-inv1 + inv2) + x
142 // (x - inv2) + inv1 => ( inv1 - inv2) + x
143 // (x - inv2) - inv1 => (-inv1 - inv2) + x
144 // inv1 + (inv2 - x) => ( inv1 + inv2) - x
145 // inv1 - (x - inv2) => ( inv1 + inv2) - x
146 // (inv2 - x) + inv1 => ( inv1 + inv2) - x
147 // (inv2 - x) - inv1 => (-inv1 + inv2) - x
148 // inv1 - (x + inv2) => ( inv1 - inv2) - x
150 Node* IdealLoopTree::reassociate_add_sub(Node* n1, PhaseIdealLoop *phase) {
151 if (!n1->is_Add() && !n1->is_Sub() || n1->outcnt() == 0) return NULL;
152 if (is_invariant(n1)) return NULL;
153 int inv1_idx = is_invariant_addition(n1, phase);
154 if (!inv1_idx) return NULL;
155 // Don't mess with add of constant (igvn moves them to expression tree root.)
156 if (n1->is_Add() && n1->in(2)->is_Con()) return NULL;
157 Node* inv1 = n1->in(inv1_idx);
158 Node* n2 = n1->in(3 - inv1_idx);
159 int inv2_idx = is_invariant_addition(n2, phase);
160 if (!inv2_idx) return NULL;
161 Node* x = n2->in(3 - inv2_idx);
162 Node* inv2 = n2->in(inv2_idx);
164 bool neg_x = n2->is_Sub() && inv2_idx == 1;
165 bool neg_inv2 = n2->is_Sub() && inv2_idx == 2;
166 bool neg_inv1 = n1->is_Sub() && inv1_idx == 2;
167 if (n1->is_Sub() && inv1_idx == 1) {
169 neg_inv2 = !neg_inv2;
171 Node* inv1_c = phase->get_ctrl(inv1);
172 Node* inv2_c = phase->get_ctrl(inv2);
175 Node *zero = phase->_igvn.intcon(0);
176 phase->set_ctrl(zero, phase->C->root());
177 n_inv1 = new (phase->C, 3) SubINode(zero, inv1);
178 phase->register_new_node(n_inv1, inv1_c);
184 inv = new (phase->C, 3) SubINode(n_inv1, inv2);
186 inv = new (phase->C, 3) AddINode(n_inv1, inv2);
188 phase->register_new_node(inv, phase->get_early_ctrl(inv));
192 addx = new (phase->C, 3) SubINode(inv, x);
194 addx = new (phase->C, 3) AddINode(x, inv);
196 phase->register_new_node(addx, phase->get_ctrl(x));
197 phase->_igvn.hash_delete(n1);
198 phase->_igvn.subsume_node(n1, addx);
202 //---------------------reassociate_invariants-----------------------------
203 // Reassociate invariant expressions:
204 void IdealLoopTree::reassociate_invariants(PhaseIdealLoop *phase) {
205 for (int i = _body.size() - 1; i >= 0; i--) {
206 Node *n = _body.at(i);
207 for (int j = 0; j < 5; j++) {
208 Node* nn = reassociate_add_sub(n, phase);
209 if (nn == NULL) break;
215 //------------------------------policy_peeling---------------------------------
216 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
217 // make some loop-invariant test (usually a null-check) happen before the loop.
218 bool IdealLoopTree::policy_peeling( PhaseIdealLoop *phase ) const {
219 Node *test = ((IdealLoopTree*)this)->tail();
220 int body_size = ((IdealLoopTree*)this)->_body.size();
221 int uniq = phase->C->unique();
222 // Peeling does loop cloning which can result in O(N^2) node construction
223 if( body_size > 255 /* Prevent overflow for large body_size */
224 || (body_size * body_size + uniq > MaxNodeLimit) ) {
225 return false; // too large to safely clone
227 while( test != _head ) { // Scan till run off top of loop
228 if( test->is_If() ) { // Test?
229 Node *ctrl = phase->get_ctrl(test->in(1));
231 return false; // Found dead test on live IF? No peeling!
232 // Standard IF only has one input value to check for loop invariance
233 assert( test->Opcode() == Op_If || test->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added");
234 // Condition is not a member of this loop?
235 if( !is_member(phase->get_loop(ctrl)) &&
237 return true; // Found reason to peel!
239 // Walk up dominators to loop _head looking for test which is
240 // executed on every path thru loop.
241 test = phase->idom(test);
246 //------------------------------peeled_dom_test_elim---------------------------
247 // If we got the effect of peeling, either by actually peeling or by making
248 // a pre-loop which must execute at least once, we can remove all
249 // loop-invariant dominated tests in the main body.
250 void PhaseIdealLoop::peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new ) {
251 bool progress = true;
253 progress = false; // Reset for next iteration
254 Node *prev = loop->_head->in(LoopNode::LoopBackControl);//loop->tail();
255 Node *test = prev->in(0);
256 while( test != loop->_head ) { // Scan till run off top of loop
258 int p_op = prev->Opcode();
259 if( (p_op == Op_IfFalse || p_op == Op_IfTrue) &&
260 test->is_If() && // Test?
261 !test->in(1)->is_Con() && // And not already obvious?
262 // Condition is not a member of this loop?
263 !loop->is_member(get_loop(get_ctrl(test->in(1))))){
264 // Walk loop body looking for instances of this test
265 for( uint i = 0; i < loop->_body.size(); i++ ) {
266 Node *n = loop->_body.at(i);
267 if( n->is_If() && n->in(1) == test->in(1) /*&& n != loop->tail()->in(0)*/ ) {
268 // IfNode was dominated by version in peeled loop body
270 dominated_by( old_new[prev->_idx], n );
276 } // End of scan tests in loop
278 } // End of while( progress )
281 //------------------------------do_peeling-------------------------------------
282 // Peel the first iteration of the given loop.
283 // Step 1: Clone the loop body. The clone becomes the peeled iteration.
284 // The pre-loop illegally has 2 control users (old & new loops).
285 // Step 2: Make the old-loop fall-in edges point to the peeled iteration.
286 // Do this by making the old-loop fall-in edges act as if they came
287 // around the loopback from the prior iteration (follow the old-loop
288 // backedges) and then map to the new peeled iteration. This leaves
289 // the pre-loop with only 1 user (the new peeled iteration), but the
290 // peeled-loop backedge has 2 users.
291 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the
292 // extra backedge user.
293 void PhaseIdealLoop::do_peeling( IdealLoopTree *loop, Node_List &old_new ) {
295 C->set_major_progress();
296 // Peeling a 'main' loop in a pre/main/post situation obfuscates the
297 // 'pre' loop from the main and the 'pre' can no longer have it's
298 // iterations adjusted. Therefore, we need to declare this loop as
299 // no longer a 'main' loop; it will need new pre and post loops before
300 // we can do further RCE.
301 Node *h = loop->_head;
302 if( h->is_CountedLoop() ) {
303 CountedLoopNode *cl = h->as_CountedLoop();
304 assert(cl->trip_count() > 0, "peeling a fully unrolled loop");
305 cl->set_trip_count(cl->trip_count() - 1);
306 if( cl->is_main_loop() ) {
307 cl->set_normal_loop();
309 if( PrintOpto && VerifyLoopOptimizations ) {
310 tty->print("Peeling a 'main' loop; resetting to 'normal' ");
317 // Step 1: Clone the loop body. The clone becomes the peeled iteration.
318 // The pre-loop illegally has 2 control users (old & new loops).
319 clone_loop( loop, old_new, dom_depth(loop->_head) );
322 // Step 2: Make the old-loop fall-in edges point to the peeled iteration.
323 // Do this by making the old-loop fall-in edges act as if they came
324 // around the loopback from the prior iteration (follow the old-loop
325 // backedges) and then map to the new peeled iteration. This leaves
326 // the pre-loop with only 1 user (the new peeled iteration), but the
327 // peeled-loop backedge has 2 users.
328 for (DUIterator_Fast jmax, j = loop->_head->fast_outs(jmax); j < jmax; j++) {
329 Node* old = loop->_head->fast_out(j);
330 if( old->in(0) == loop->_head && old->req() == 3 &&
331 (old->is_Loop() || old->is_Phi()) ) {
332 Node *new_exit_value = old_new[old->in(LoopNode::LoopBackControl)->_idx];
333 if( !new_exit_value ) // Backedge value is ALSO loop invariant?
334 // Then loop body backedge value remains the same.
335 new_exit_value = old->in(LoopNode::LoopBackControl);
336 _igvn.hash_delete(old);
337 old->set_req(LoopNode::EntryControl, new_exit_value);
342 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the
343 // extra backedge user.
344 Node *nnn = old_new[loop->_head->_idx];
345 _igvn.hash_delete(nnn);
346 nnn->set_req(LoopNode::LoopBackControl, C->top());
347 for (DUIterator_Fast j2max, j2 = nnn->fast_outs(j2max); j2 < j2max; j2++) {
348 Node* use = nnn->fast_out(j2);
349 if( use->in(0) == nnn && use->req() == 3 && use->is_Phi() ) {
350 _igvn.hash_delete(use);
351 use->set_req(LoopNode::LoopBackControl, C->top());
356 // Step 4: Correct dom-depth info. Set to loop-head depth.
357 int dd = dom_depth(loop->_head);
358 set_idom(loop->_head, loop->_head->in(1), dd);
359 for (uint j3 = 0; j3 < loop->_body.size(); j3++) {
360 Node *old = loop->_body.at(j3);
361 Node *nnn = old_new[old->_idx];
363 set_idom(nnn, idom(nnn), dd-1);
364 // While we're at it, remove any SafePoints from the peeled code
365 if( old->Opcode() == Op_SafePoint ) {
366 Node *nnn = old_new[old->_idx];
367 lazy_replace(nnn,nnn->in(TypeFunc::Control));
371 // Now force out all loop-invariant dominating tests. The optimizer
372 // finds some, but we _know_ they are all useless.
373 peeled_dom_test_elim(loop,old_new);
375 loop->record_for_igvn();
378 //------------------------------policy_maximally_unroll------------------------
379 // Return exact loop trip count, or 0 if not maximally unrolling
380 bool IdealLoopTree::policy_maximally_unroll( PhaseIdealLoop *phase ) const {
381 CountedLoopNode *cl = _head->as_CountedLoop();
382 assert( cl->is_normal_loop(), "" );
384 Node *init_n = cl->init_trip();
385 Node *limit_n = cl->limit();
387 // Non-constant bounds
388 if( init_n == NULL || !init_n->is_Con() ||
389 limit_n == NULL || !limit_n->is_Con() ||
390 // protect against stride not being a constant
391 !cl->stride_is_con() ) {
394 int init = init_n->get_int();
395 int limit = limit_n->get_int();
396 int span = limit - init;
397 int stride = cl->stride_con();
399 if (init >= limit || stride > span) {
400 // return a false (no maximally unroll) and the regular unroll/peel
401 // route will make a small mess which CCP will fold away.
404 uint trip_count = span/stride; // trip_count can be greater than 2 Gig.
405 assert( (int)trip_count*stride == span, "must divide evenly" );
407 // Real policy: if we maximally unroll, does it get too big?
408 // Allow the unrolled mess to get larger than standard loop
409 // size. After all, it will no longer be a loop.
410 uint body_size = _body.size();
411 uint unroll_limit = (uint)LoopUnrollLimit * 4;
412 assert( (intx)unroll_limit == LoopUnrollLimit * 4, "LoopUnrollLimit must fit in 32bits");
413 cl->set_trip_count(trip_count);
414 if( trip_count <= unroll_limit && body_size <= unroll_limit ) {
415 uint new_body_size = body_size * trip_count;
416 if (new_body_size <= unroll_limit &&
417 body_size == new_body_size / trip_count &&
418 // Unrolling can result in a large amount of node construction
419 new_body_size < MaxNodeLimit - phase->C->unique()) {
420 return true; // maximally unroll
424 return false; // Do not maximally unroll
428 //------------------------------policy_unroll----------------------------------
429 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if
430 // the loop is a CountedLoop and the body is small enough.
431 bool IdealLoopTree::policy_unroll( PhaseIdealLoop *phase ) const {
433 CountedLoopNode *cl = _head->as_CountedLoop();
434 assert( cl->is_normal_loop() || cl->is_main_loop(), "" );
436 // protect against stride not being a constant
437 if( !cl->stride_is_con() ) return false;
439 // protect against over-unrolling
440 if( cl->trip_count() <= 1 ) return false;
442 int future_unroll_ct = cl->unrolled_count() * 2;
444 // Don't unroll if the next round of unrolling would push us
445 // over the expected trip count of the loop. One is subtracted
446 // from the expected trip count because the pre-loop normally
447 // executes 1 iteration.
448 if (UnrollLimitForProfileCheck > 0 &&
449 cl->profile_trip_cnt() != COUNT_UNKNOWN &&
450 future_unroll_ct > UnrollLimitForProfileCheck &&
451 (float)future_unroll_ct > cl->profile_trip_cnt() - 1.0) {
455 // When unroll count is greater than LoopUnrollMin, don't unroll if:
456 // the residual iterations are more than 10% of the trip count
457 // and rounds of "unroll,optimize" are not making significant progress
458 // Progress defined as current size less than 20% larger than previous size.
459 if (UseSuperWord && cl->node_count_before_unroll() > 0 &&
460 future_unroll_ct > LoopUnrollMin &&
461 (future_unroll_ct - 1) * 10.0 > cl->profile_trip_cnt() &&
462 1.2 * cl->node_count_before_unroll() < (double)_body.size()) {
466 Node *init_n = cl->init_trip();
467 Node *limit_n = cl->limit();
468 // Non-constant bounds.
469 // Protect against over-unrolling when init or/and limit are not constant
470 // (so that trip_count's init value is maxint) but iv range is known.
471 if( init_n == NULL || !init_n->is_Con() ||
472 limit_n == NULL || !limit_n->is_Con() ) {
473 Node* phi = cl->phi();
475 assert(phi->is_Phi() && phi->in(0) == _head, "Counted loop should have iv phi.");
476 const TypeInt* iv_type = phase->_igvn.type(phi)->is_int();
477 int next_stride = cl->stride_con() * 2; // stride after this unroll
478 if( next_stride > 0 ) {
479 if( iv_type->_lo + next_stride <= iv_type->_lo || // overflow
480 iv_type->_lo + next_stride > iv_type->_hi ) {
481 return false; // over-unrolling
483 } else if( next_stride < 0 ) {
484 if( iv_type->_hi + next_stride >= iv_type->_hi || // overflow
485 iv_type->_hi + next_stride < iv_type->_lo ) {
486 return false; // over-unrolling
492 // Adjust body_size to determine if we unroll or not
493 uint body_size = _body.size();
494 // Key test to unroll CaffeineMark's Logic test
495 int xors_in_loop = 0;
496 // Also count ModL, DivL and MulL which expand mightly
497 for( uint k = 0; k < _body.size(); k++ ) {
498 switch( _body.at(k)->Opcode() ) {
499 case Op_XorI: xors_in_loop++; break; // CaffeineMark's Logic test
500 case Op_ModL: body_size += 30; break;
501 case Op_DivL: body_size += 30; break;
502 case Op_MulL: body_size += 10; break;
506 // Check for being too big
507 if( body_size > (uint)LoopUnrollLimit ) {
508 if( xors_in_loop >= 4 && body_size < (uint)LoopUnrollLimit*4) return true;
509 // Normal case: loop too big
513 // Check for stride being a small enough constant
514 if( abs(cl->stride_con()) > (1<<3) ) return false;
516 // Unroll once! (Each trip will soon do double iterations)
520 //------------------------------policy_align-----------------------------------
521 // Return TRUE or FALSE if the loop should be cache-line aligned. Gather the
522 // expression that does the alignment. Note that only one array base can be
523 // aligned in a loop (unless the VM guarantees mutual alignment). Note that
524 // if we vectorize short memory ops into longer memory ops, we may want to
525 // increase alignment.
526 bool IdealLoopTree::policy_align( PhaseIdealLoop *phase ) const {
530 //------------------------------policy_range_check-----------------------------
531 // Return TRUE or FALSE if the loop should be range-check-eliminated.
532 // Actually we do iteration-splitting, a more powerful form of RCE.
533 bool IdealLoopTree::policy_range_check( PhaseIdealLoop *phase ) const {
534 if( !RangeCheckElimination ) return false;
536 CountedLoopNode *cl = _head->as_CountedLoop();
537 // If we unrolled with no intention of doing RCE and we later
538 // changed our minds, we got no pre-loop. Either we need to
539 // make a new pre-loop, or we gotta disallow RCE.
540 if( cl->is_main_no_pre_loop() ) return false; // Disallowed for now.
541 Node *trip_counter = cl->phi();
543 // Check loop body for tests of trip-counter plus loop-invariant vs
545 for( uint i = 0; i < _body.size(); i++ ) {
546 Node *iff = _body[i];
547 if( iff->Opcode() == Op_If ) { // Test?
549 // Comparing trip+off vs limit
550 Node *bol = iff->in(1);
551 if( bol->req() != 2 ) continue; // dead constant test
552 if (!bol->is_Bool()) {
553 assert(UseLoopPredicate && bol->Opcode() == Op_Conv2B, "predicate check only");
556 Node *cmp = bol->in(1);
558 Node *rc_exp = cmp->in(1);
559 Node *limit = cmp->in(2);
561 Node *limit_c = phase->get_ctrl(limit);
562 if( limit_c == phase->C->top() )
563 return false; // Found dead test on live IF? No RCE!
564 if( is_member(phase->get_loop(limit_c) ) ) {
565 // Compare might have operands swapped; commute them
568 limit_c = phase->get_ctrl(limit);
569 if( is_member(phase->get_loop(limit_c) ) )
570 continue; // Both inputs are loop varying; cannot RCE
573 if (!phase->is_scaled_iv_plus_offset(rc_exp, trip_counter, NULL, NULL)) {
576 // Yeah! Found a test like 'trip+off vs limit'
577 // Test is an IfNode, has 2 projections. If BOTH are in the loop
578 // we need loop unswitching instead of iteration splitting.
579 if( is_loop_exit(iff) )
580 return true; // Found reason to split iterations
587 //------------------------------policy_peel_only-------------------------------
588 // Return TRUE or FALSE if the loop should NEVER be RCE'd or aligned. Useful
589 // for unrolling loops with NO array accesses.
590 bool IdealLoopTree::policy_peel_only( PhaseIdealLoop *phase ) const {
592 for( uint i = 0; i < _body.size(); i++ )
593 if( _body[i]->is_Mem() )
596 // No memory accesses at all!
600 //------------------------------clone_up_backedge_goo--------------------------
601 // If Node n lives in the back_ctrl block and cannot float, we clone a private
602 // version of n in preheader_ctrl block and return that, otherwise return n.
603 Node *PhaseIdealLoop::clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n ) {
604 if( get_ctrl(n) != back_ctrl ) return n;
606 Node *x = NULL; // If required, a clone of 'n'
607 // Check for 'n' being pinned in the backedge.
608 if( n->in(0) && n->in(0) == back_ctrl ) {
609 x = n->clone(); // Clone a copy of 'n' to preheader
610 x->set_req( 0, preheader_ctrl ); // Fix x's control input to preheader
613 // Recursive fixup any other input edges into x.
614 // If there are no changes we can just return 'n', otherwise
615 // we need to clone a private copy and change it.
616 for( uint i = 1; i < n->req(); i++ ) {
617 Node *g = clone_up_backedge_goo( back_ctrl, preheader_ctrl, n->in(i) );
618 if( g != n->in(i) ) {
624 if( x ) { // x can legally float to pre-header location
625 register_new_node( x, preheader_ctrl );
627 } else { // raise n to cover LCA of uses
628 set_ctrl( n, find_non_split_ctrl(back_ctrl->in(0)) );
633 //------------------------------insert_pre_post_loops--------------------------
634 // Insert pre and post loops. If peel_only is set, the pre-loop can not have
635 // more iterations added. It acts as a 'peel' only, no lower-bound RCE, no
636 // alignment. Useful to unroll loops that do no array accesses.
637 void PhaseIdealLoop::insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ) {
639 C->set_major_progress();
641 // Find common pieces of the loop being guarded with pre & post loops
642 CountedLoopNode *main_head = loop->_head->as_CountedLoop();
643 assert( main_head->is_normal_loop(), "" );
644 CountedLoopEndNode *main_end = main_head->loopexit();
645 assert( main_end->outcnt() == 2, "1 true, 1 false path only" );
646 uint dd_main_head = dom_depth(main_head);
647 uint max = main_head->outcnt();
649 Node *pre_header= main_head->in(LoopNode::EntryControl);
650 Node *init = main_head->init_trip();
651 Node *incr = main_end ->incr();
652 Node *limit = main_end ->limit();
653 Node *stride = main_end ->stride();
654 Node *cmp = main_end ->cmp_node();
655 BoolTest::mask b_test = main_end->test_trip();
657 // Need only 1 user of 'bol' because I will be hacking the loop bounds.
658 Node *bol = main_end->in(CountedLoopEndNode::TestValue);
659 if( bol->outcnt() != 1 ) {
661 register_new_node(bol,main_end->in(CountedLoopEndNode::TestControl));
662 _igvn.hash_delete(main_end);
663 main_end->set_req(CountedLoopEndNode::TestValue, bol);
665 // Need only 1 user of 'cmp' because I will be hacking the loop bounds.
666 if( cmp->outcnt() != 1 ) {
668 register_new_node(cmp,main_end->in(CountedLoopEndNode::TestControl));
669 _igvn.hash_delete(bol);
670 bol->set_req(1, cmp);
673 //------------------------------
674 // Step A: Create Post-Loop.
675 Node* main_exit = main_end->proj_out(false);
676 assert( main_exit->Opcode() == Op_IfFalse, "" );
677 int dd_main_exit = dom_depth(main_exit);
679 // Step A1: Clone the loop body. The clone becomes the post-loop. The main
680 // loop pre-header illegally has 2 control users (old & new loops).
681 clone_loop( loop, old_new, dd_main_exit );
682 assert( old_new[main_end ->_idx]->Opcode() == Op_CountedLoopEnd, "" );
683 CountedLoopNode *post_head = old_new[main_head->_idx]->as_CountedLoop();
684 post_head->set_post_loop(main_head);
686 // Reduce the post-loop trip count.
687 CountedLoopEndNode* post_end = old_new[main_end ->_idx]->as_CountedLoopEnd();
688 post_end->_prob = PROB_FAIR;
690 // Build the main-loop normal exit.
691 IfFalseNode *new_main_exit = new (C, 1) IfFalseNode(main_end);
692 _igvn.register_new_node_with_optimizer( new_main_exit );
693 set_idom(new_main_exit, main_end, dd_main_exit );
694 set_loop(new_main_exit, loop->_parent);
696 // Step A2: Build a zero-trip guard for the post-loop. After leaving the
697 // main-loop, the post-loop may not execute at all. We 'opaque' the incr
698 // (the main-loop trip-counter exit value) because we will be changing
699 // the exit value (via unrolling) so we cannot constant-fold away the zero
700 // trip guard until all unrolling is done.
701 Node *zer_opaq = new (C, 2) Opaque1Node(C, incr);
702 Node *zer_cmp = new (C, 3) CmpINode( zer_opaq, limit );
703 Node *zer_bol = new (C, 2) BoolNode( zer_cmp, b_test );
704 register_new_node( zer_opaq, new_main_exit );
705 register_new_node( zer_cmp , new_main_exit );
706 register_new_node( zer_bol , new_main_exit );
709 IfNode *zer_iff = new (C, 2) IfNode( new_main_exit, zer_bol, PROB_FAIR, COUNT_UNKNOWN );
710 _igvn.register_new_node_with_optimizer( zer_iff );
711 set_idom(zer_iff, new_main_exit, dd_main_exit);
712 set_loop(zer_iff, loop->_parent);
714 // Plug in the false-path, taken if we need to skip post-loop
715 _igvn.hash_delete( main_exit );
716 main_exit->set_req(0, zer_iff);
717 _igvn._worklist.push(main_exit);
718 set_idom(main_exit, zer_iff, dd_main_exit);
719 set_idom(main_exit->unique_out(), zer_iff, dd_main_exit);
720 // Make the true-path, must enter the post loop
721 Node *zer_taken = new (C, 1) IfTrueNode( zer_iff );
722 _igvn.register_new_node_with_optimizer( zer_taken );
723 set_idom(zer_taken, zer_iff, dd_main_exit);
724 set_loop(zer_taken, loop->_parent);
725 // Plug in the true path
726 _igvn.hash_delete( post_head );
727 post_head->set_req(LoopNode::EntryControl, zer_taken);
728 set_idom(post_head, zer_taken, dd_main_exit);
730 // Step A3: Make the fall-in values to the post-loop come from the
731 // fall-out values of the main-loop.
732 for (DUIterator_Fast imax, i = main_head->fast_outs(imax); i < imax; i++) {
733 Node* main_phi = main_head->fast_out(i);
734 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() >0 ) {
735 Node *post_phi = old_new[main_phi->_idx];
736 Node *fallmain = clone_up_backedge_goo(main_head->back_control(),
737 post_head->init_control(),
738 main_phi->in(LoopNode::LoopBackControl));
739 _igvn.hash_delete(post_phi);
740 post_phi->set_req( LoopNode::EntryControl, fallmain );
744 // Update local caches for next stanza
745 main_exit = new_main_exit;
748 //------------------------------
749 // Step B: Create Pre-Loop.
751 // Step B1: Clone the loop body. The clone becomes the pre-loop. The main
752 // loop pre-header illegally has 2 control users (old & new loops).
753 clone_loop( loop, old_new, dd_main_head );
754 CountedLoopNode* pre_head = old_new[main_head->_idx]->as_CountedLoop();
755 CountedLoopEndNode* pre_end = old_new[main_end ->_idx]->as_CountedLoopEnd();
756 pre_head->set_pre_loop(main_head);
757 Node *pre_incr = old_new[incr->_idx];
759 // Reduce the pre-loop trip count.
760 pre_end->_prob = PROB_FAIR;
762 // Find the pre-loop normal exit.
763 Node* pre_exit = pre_end->proj_out(false);
764 assert( pre_exit->Opcode() == Op_IfFalse, "" );
765 IfFalseNode *new_pre_exit = new (C, 1) IfFalseNode(pre_end);
766 _igvn.register_new_node_with_optimizer( new_pre_exit );
767 set_idom(new_pre_exit, pre_end, dd_main_head);
768 set_loop(new_pre_exit, loop->_parent);
770 // Step B2: Build a zero-trip guard for the main-loop. After leaving the
771 // pre-loop, the main-loop may not execute at all. Later in life this
772 // zero-trip guard will become the minimum-trip guard when we unroll
774 Node *min_opaq = new (C, 2) Opaque1Node(C, limit);
775 Node *min_cmp = new (C, 3) CmpINode( pre_incr, min_opaq );
776 Node *min_bol = new (C, 2) BoolNode( min_cmp, b_test );
777 register_new_node( min_opaq, new_pre_exit );
778 register_new_node( min_cmp , new_pre_exit );
779 register_new_node( min_bol , new_pre_exit );
781 // Build the IfNode (assume the main-loop is executed always).
782 IfNode *min_iff = new (C, 2) IfNode( new_pre_exit, min_bol, PROB_ALWAYS, COUNT_UNKNOWN );
783 _igvn.register_new_node_with_optimizer( min_iff );
784 set_idom(min_iff, new_pre_exit, dd_main_head);
785 set_loop(min_iff, loop->_parent);
787 // Plug in the false-path, taken if we need to skip main-loop
788 _igvn.hash_delete( pre_exit );
789 pre_exit->set_req(0, min_iff);
790 set_idom(pre_exit, min_iff, dd_main_head);
791 set_idom(pre_exit->unique_out(), min_iff, dd_main_head);
792 // Make the true-path, must enter the main loop
793 Node *min_taken = new (C, 1) IfTrueNode( min_iff );
794 _igvn.register_new_node_with_optimizer( min_taken );
795 set_idom(min_taken, min_iff, dd_main_head);
796 set_loop(min_taken, loop->_parent);
797 // Plug in the true path
798 _igvn.hash_delete( main_head );
799 main_head->set_req(LoopNode::EntryControl, min_taken);
800 set_idom(main_head, min_taken, dd_main_head);
802 // Step B3: Make the fall-in values to the main-loop come from the
803 // fall-out values of the pre-loop.
804 for (DUIterator_Fast i2max, i2 = main_head->fast_outs(i2max); i2 < i2max; i2++) {
805 Node* main_phi = main_head->fast_out(i2);
806 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() > 0 ) {
807 Node *pre_phi = old_new[main_phi->_idx];
808 Node *fallpre = clone_up_backedge_goo(pre_head->back_control(),
809 main_head->init_control(),
810 pre_phi->in(LoopNode::LoopBackControl));
811 _igvn.hash_delete(main_phi);
812 main_phi->set_req( LoopNode::EntryControl, fallpre );
816 // Step B4: Shorten the pre-loop to run only 1 iteration (for now).
817 // RCE and alignment may change this later.
818 Node *cmp_end = pre_end->cmp_node();
819 assert( cmp_end->in(2) == limit, "" );
820 Node *pre_limit = new (C, 3) AddINode( init, stride );
822 // Save the original loop limit in this Opaque1 node for
823 // use by range check elimination.
824 Node *pre_opaq = new (C, 3) Opaque1Node(C, pre_limit, limit);
826 register_new_node( pre_limit, pre_head->in(0) );
827 register_new_node( pre_opaq , pre_head->in(0) );
829 // Since no other users of pre-loop compare, I can hack limit directly
830 assert( cmp_end->outcnt() == 1, "no other users" );
831 _igvn.hash_delete(cmp_end);
832 cmp_end->set_req(2, peel_only ? pre_limit : pre_opaq);
834 // Special case for not-equal loop bounds:
835 // Change pre loop test, main loop test, and the
836 // main loop guard test to use lt or gt depending on stride
838 // positive stride use <
839 // negative stride use >
841 if (pre_end->in(CountedLoopEndNode::TestValue)->as_Bool()->_test._test == BoolTest::ne) {
843 BoolTest::mask new_test = (main_end->stride_con() > 0) ? BoolTest::lt : BoolTest::gt;
844 // Modify pre loop end condition
845 Node* pre_bol = pre_end->in(CountedLoopEndNode::TestValue)->as_Bool();
846 BoolNode* new_bol0 = new (C, 2) BoolNode(pre_bol->in(1), new_test);
847 register_new_node( new_bol0, pre_head->in(0) );
848 _igvn.hash_delete(pre_end);
849 pre_end->set_req(CountedLoopEndNode::TestValue, new_bol0);
850 // Modify main loop guard condition
851 assert(min_iff->in(CountedLoopEndNode::TestValue) == min_bol, "guard okay");
852 BoolNode* new_bol1 = new (C, 2) BoolNode(min_bol->in(1), new_test);
853 register_new_node( new_bol1, new_pre_exit );
854 _igvn.hash_delete(min_iff);
855 min_iff->set_req(CountedLoopEndNode::TestValue, new_bol1);
856 // Modify main loop end condition
857 BoolNode* main_bol = main_end->in(CountedLoopEndNode::TestValue)->as_Bool();
858 BoolNode* new_bol2 = new (C, 2) BoolNode(main_bol->in(1), new_test);
859 register_new_node( new_bol2, main_end->in(CountedLoopEndNode::TestControl) );
860 _igvn.hash_delete(main_end);
861 main_end->set_req(CountedLoopEndNode::TestValue, new_bol2);
865 main_head->set_main_loop();
866 if( peel_only ) main_head->set_main_no_pre_loop();
868 // It's difficult to be precise about the trip-counts
869 // for the pre/post loops. They are usually very short,
870 // so guess that 4 trips is a reasonable value.
871 post_head->set_profile_trip_cnt(4.0);
872 pre_head->set_profile_trip_cnt(4.0);
874 // Now force out all loop-invariant dominating tests. The optimizer
875 // finds some, but we _know_ they are all useless.
876 peeled_dom_test_elim(loop,old_new);
879 //------------------------------is_invariant-----------------------------
880 // Return true if n is invariant
881 bool IdealLoopTree::is_invariant(Node* n) const {
882 Node *n_c = _phase->has_ctrl(n) ? _phase->get_ctrl(n) : n;
883 if (n_c->is_top()) return false;
884 return !is_member(_phase->get_loop(n_c));
888 //------------------------------do_unroll--------------------------------------
889 // Unroll the loop body one step - make each trip do 2 iterations.
890 void PhaseIdealLoop::do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ) {
891 assert( LoopUnrollLimit, "" );
893 if( PrintOpto && VerifyLoopOptimizations ) {
894 tty->print("Unrolling ");
898 CountedLoopNode *loop_head = loop->_head->as_CountedLoop();
899 CountedLoopEndNode *loop_end = loop_head->loopexit();
900 assert( loop_end, "" );
902 // Remember loop node count before unrolling to detect
903 // if rounds of unroll,optimize are making progress
904 loop_head->set_node_count_before_unroll(loop->_body.size());
906 Node *ctrl = loop_head->in(LoopNode::EntryControl);
907 Node *limit = loop_head->limit();
908 Node *init = loop_head->init_trip();
909 Node *strid = loop_head->stride();
912 if( adjust_min_trip ) { // If not maximally unrolling, need adjustment
913 assert( loop_head->is_main_loop(), "" );
914 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );
915 Node *iff = ctrl->in(0);
916 assert( iff->Opcode() == Op_If, "" );
917 Node *bol = iff->in(1);
918 assert( bol->Opcode() == Op_Bool, "" );
919 Node *cmp = bol->in(1);
920 assert( cmp->Opcode() == Op_CmpI, "" );
922 // Occasionally it's possible for a pre-loop Opaque1 node to be
923 // optimized away and then another round of loop opts attempted.
924 // We can not optimize this particular loop in that case.
925 if( opaq->Opcode() != Op_Opaque1 )
926 return; // Cannot find pre-loop! Bail out!
929 C->set_major_progress();
931 // Adjust max trip count. The trip count is intentionally rounded
932 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll,
933 // the main, unrolled, part of the loop will never execute as it is protected
934 // by the min-trip test. See bug 4834191 for a case where we over-unrolled
935 // and later determined that part of the unrolled loop was dead.
936 loop_head->set_trip_count(loop_head->trip_count() / 2);
938 // Double the count of original iterations in the unrolled loop body.
939 loop_head->double_unrolled_count();
942 // Step 2: Cut back the trip counter for an unroll amount of 2.
943 // Loop will normally trip (limit - init)/stride_con. Since it's a
944 // CountedLoop this is exact (stride divides limit-init exactly).
945 // We are going to double the loop body, so we want to knock off any
946 // odd iteration: (trip_cnt & ~1). Then back compute a new limit.
947 Node *span = new (C, 3) SubINode( limit, init );
948 register_new_node( span, ctrl );
949 Node *trip = new (C, 3) DivINode( 0, span, strid );
950 register_new_node( trip, ctrl );
951 Node *mtwo = _igvn.intcon(-2);
952 set_ctrl(mtwo, C->root());
953 Node *rond = new (C, 3) AndINode( trip, mtwo );
954 register_new_node( rond, ctrl );
955 Node *spn2 = new (C, 3) MulINode( rond, strid );
956 register_new_node( spn2, ctrl );
957 Node *lim2 = new (C, 3) AddINode( spn2, init );
958 register_new_node( lim2, ctrl );
960 // Hammer in the new limit
961 Node *ctrl2 = loop_end->in(0);
962 Node *cmp2 = new (C, 3) CmpINode( loop_head->incr(), lim2 );
963 register_new_node( cmp2, ctrl2 );
964 Node *bol2 = new (C, 2) BoolNode( cmp2, loop_end->test_trip() );
965 register_new_node( bol2, ctrl2 );
966 _igvn.hash_delete(loop_end);
967 loop_end->set_req(CountedLoopEndNode::TestValue, bol2);
969 // Step 3: Find the min-trip test guaranteed before a 'main' loop.
970 // Make it a 1-trip test (means at least 2 trips).
971 if( adjust_min_trip ) {
972 // Guard test uses an 'opaque' node which is not shared. Hence I
973 // can edit it's inputs directly. Hammer in the new limit for the
974 // minimum-trip guard.
975 assert( opaq->outcnt() == 1, "" );
976 _igvn.hash_delete(opaq);
977 opaq->set_req(1, lim2);
981 // Step 4: Clone the loop body. Move it inside the loop. This loop body
982 // represents the odd iterations; since the loop trips an even number of
983 // times its backedge is never taken. Kill the backedge.
984 uint dd = dom_depth(loop_head);
985 clone_loop( loop, old_new, dd );
987 // Make backedges of the clone equal to backedges of the original.
988 // Make the fall-in from the original come from the fall-out of the clone.
989 for (DUIterator_Fast jmax, j = loop_head->fast_outs(jmax); j < jmax; j++) {
990 Node* phi = loop_head->fast_out(j);
991 if( phi->is_Phi() && phi->in(0) == loop_head && phi->outcnt() > 0 ) {
992 Node *newphi = old_new[phi->_idx];
993 _igvn.hash_delete( phi );
994 _igvn.hash_delete( newphi );
996 phi ->set_req(LoopNode:: EntryControl, newphi->in(LoopNode::LoopBackControl));
997 newphi->set_req(LoopNode::LoopBackControl, phi ->in(LoopNode::LoopBackControl));
998 phi ->set_req(LoopNode::LoopBackControl, C->top());
1001 Node *clone_head = old_new[loop_head->_idx];
1002 _igvn.hash_delete( clone_head );
1003 loop_head ->set_req(LoopNode:: EntryControl, clone_head->in(LoopNode::LoopBackControl));
1004 clone_head->set_req(LoopNode::LoopBackControl, loop_head ->in(LoopNode::LoopBackControl));
1005 loop_head ->set_req(LoopNode::LoopBackControl, C->top());
1006 loop->_head = clone_head; // New loop header
1008 set_idom(loop_head, loop_head ->in(LoopNode::EntryControl), dd);
1009 set_idom(clone_head, clone_head->in(LoopNode::EntryControl), dd);
1011 // Kill the clone's backedge
1012 Node *newcle = old_new[loop_end->_idx];
1013 _igvn.hash_delete( newcle );
1014 Node *one = _igvn.intcon(1);
1015 set_ctrl(one, C->root());
1016 newcle->set_req(1, one);
1017 // Force clone into same loop body
1018 uint max = loop->_body.size();
1019 for( uint k = 0; k < max; k++ ) {
1020 Node *old = loop->_body.at(k);
1021 Node *nnn = old_new[old->_idx];
1022 loop->_body.push(nnn);
1024 set_loop(nnn, loop);
1027 loop->record_for_igvn();
1030 //------------------------------do_maximally_unroll----------------------------
1032 void PhaseIdealLoop::do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ) {
1033 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1034 assert( cl->trip_count() > 0, "");
1036 // If loop is tripping an odd number of times, peel odd iteration
1037 if( (cl->trip_count() & 1) == 1 ) {
1038 do_peeling( loop, old_new );
1041 // Now its tripping an even number of times remaining. Double loop body.
1042 // Do not adjust pre-guards; they are not needed and do not exist.
1043 if( cl->trip_count() > 0 ) {
1044 do_unroll( loop, old_new, false );
1048 //------------------------------dominates_backedge---------------------------------
1049 // Returns true if ctrl is executed on every complete iteration
1050 bool IdealLoopTree::dominates_backedge(Node* ctrl) {
1051 assert(ctrl->is_CFG(), "must be control");
1052 Node* backedge = _head->as_Loop()->in(LoopNode::LoopBackControl);
1053 return _phase->dom_lca_internal(ctrl, backedge) == ctrl;
1056 //------------------------------add_constraint---------------------------------
1057 // Constrain the main loop iterations so the condition:
1058 // scale_con * I + offset < limit
1059 // always holds true. That is, either increase the number of iterations in
1060 // the pre-loop or the post-loop until the condition holds true in the main
1061 // loop. Stride, scale, offset and limit are all loop invariant. Further,
1062 // stride and scale are constants (offset and limit often are).
1063 void PhaseIdealLoop::add_constraint( int stride_con, int scale_con, Node *offset, Node *limit, Node *pre_ctrl, Node **pre_limit, Node **main_limit ) {
1065 // Compute "I :: (limit-offset)/scale_con"
1066 Node *con = new (C, 3) SubINode( limit, offset );
1067 register_new_node( con, pre_ctrl );
1068 Node *scale = _igvn.intcon(scale_con);
1069 set_ctrl(scale, C->root());
1070 Node *X = new (C, 3) DivINode( 0, con, scale );
1071 register_new_node( X, pre_ctrl );
1073 // For positive stride, the pre-loop limit always uses a MAX function
1074 // and the main loop a MIN function. For negative stride these are
1077 // Also for positive stride*scale the affine function is increasing, so the
1078 // pre-loop must check for underflow and the post-loop for overflow.
1079 // Negative stride*scale reverses this; pre-loop checks for overflow and
1080 // post-loop for underflow.
1081 if( stride_con*scale_con > 0 ) {
1082 // Compute I < (limit-offset)/scale_con
1083 // Adjust main-loop last iteration to be MIN/MAX(main_loop,X)
1084 *main_limit = (stride_con > 0)
1085 ? (Node*)(new (C, 3) MinINode( *main_limit, X ))
1086 : (Node*)(new (C, 3) MaxINode( *main_limit, X ));
1087 register_new_node( *main_limit, pre_ctrl );
1090 // Compute (limit-offset)/scale_con + SGN(-scale_con) <= I
1091 // Add the negation of the main-loop constraint to the pre-loop.
1092 // See footnote [++] below for a derivation of the limit expression.
1093 Node *incr = _igvn.intcon(scale_con > 0 ? -1 : 1);
1094 set_ctrl(incr, C->root());
1095 Node *adj = new (C, 3) AddINode( X, incr );
1096 register_new_node( adj, pre_ctrl );
1097 *pre_limit = (scale_con > 0)
1098 ? (Node*)new (C, 3) MinINode( *pre_limit, adj )
1099 : (Node*)new (C, 3) MaxINode( *pre_limit, adj );
1100 register_new_node( *pre_limit, pre_ctrl );
1102 // [++] Here's the algebra that justifies the pre-loop limit expression:
1104 // NOT( scale_con * I + offset < limit )
1106 // scale_con * I + offset >= limit
1108 // SGN(scale_con) * I >= (limit-offset)/|scale_con|
1110 // (limit-offset)/|scale_con| <= I * SGN(scale_con)
1112 // (limit-offset)/|scale_con|-1 < I * SGN(scale_con)
1114 // ( if (scale_con > 0) /*common case*/
1115 // (limit-offset)/scale_con - 1 < I
1117 // (limit-offset)/scale_con + 1 > I
1119 // ( if (scale_con > 0) /*common case*/
1120 // (limit-offset)/scale_con + SGN(-scale_con) < I
1122 // (limit-offset)/scale_con + SGN(-scale_con) > I
1127 //------------------------------is_scaled_iv---------------------------------
1128 // Return true if exp is a constant times an induction var
1129 bool PhaseIdealLoop::is_scaled_iv(Node* exp, Node* iv, int* p_scale) {
1131 if (p_scale != NULL) {
1136 int opc = exp->Opcode();
1137 if (opc == Op_MulI) {
1138 if (exp->in(1) == iv && exp->in(2)->is_Con()) {
1139 if (p_scale != NULL) {
1140 *p_scale = exp->in(2)->get_int();
1144 if (exp->in(2) == iv && exp->in(1)->is_Con()) {
1145 if (p_scale != NULL) {
1146 *p_scale = exp->in(1)->get_int();
1150 } else if (opc == Op_LShiftI) {
1151 if (exp->in(1) == iv && exp->in(2)->is_Con()) {
1152 if (p_scale != NULL) {
1153 *p_scale = 1 << exp->in(2)->get_int();
1161 //-----------------------------is_scaled_iv_plus_offset------------------------------
1162 // Return true if exp is a simple induction variable expression: k1*iv + (invar + k2)
1163 bool PhaseIdealLoop::is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth) {
1164 if (is_scaled_iv(exp, iv, p_scale)) {
1165 if (p_offset != NULL) {
1166 Node *zero = _igvn.intcon(0);
1167 set_ctrl(zero, C->root());
1172 int opc = exp->Opcode();
1173 if (opc == Op_AddI) {
1174 if (is_scaled_iv(exp->in(1), iv, p_scale)) {
1175 if (p_offset != NULL) {
1176 *p_offset = exp->in(2);
1180 if (exp->in(2)->is_Con()) {
1181 Node* offset2 = NULL;
1183 is_scaled_iv_plus_offset(exp->in(1), iv, p_scale,
1184 p_offset != NULL ? &offset2 : NULL, depth+1)) {
1185 if (p_offset != NULL) {
1186 Node *ctrl_off2 = get_ctrl(offset2);
1187 Node* offset = new (C, 3) AddINode(offset2, exp->in(2));
1188 register_new_node(offset, ctrl_off2);
1194 } else if (opc == Op_SubI) {
1195 if (is_scaled_iv(exp->in(1), iv, p_scale)) {
1196 if (p_offset != NULL) {
1197 Node *zero = _igvn.intcon(0);
1198 set_ctrl(zero, C->root());
1199 Node *ctrl_off = get_ctrl(exp->in(2));
1200 Node* offset = new (C, 3) SubINode(zero, exp->in(2));
1201 register_new_node(offset, ctrl_off);
1206 if (is_scaled_iv(exp->in(2), iv, p_scale)) {
1207 if (p_offset != NULL) {
1209 *p_offset = exp->in(1);
1217 //------------------------------do_range_check---------------------------------
1218 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
1219 void PhaseIdealLoop::do_range_check( IdealLoopTree *loop, Node_List &old_new ) {
1221 if( PrintOpto && VerifyLoopOptimizations ) {
1222 tty->print("Range Check Elimination ");
1226 assert( RangeCheckElimination, "" );
1227 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1228 assert( cl->is_main_loop(), "" );
1230 // Find the trip counter; we are iteration splitting based on it
1231 Node *trip_counter = cl->phi();
1232 // Find the main loop limit; we will trim it's iterations
1233 // to not ever trip end tests
1234 Node *main_limit = cl->limit();
1235 // Find the pre-loop limit; we will expand it's iterations to
1236 // not ever trip low tests.
1237 Node *ctrl = cl->in(LoopNode::EntryControl);
1238 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );
1239 Node *iffm = ctrl->in(0);
1240 assert( iffm->Opcode() == Op_If, "" );
1241 Node *p_f = iffm->in(0);
1242 assert( p_f->Opcode() == Op_IfFalse, "" );
1243 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
1244 assert( pre_end->loopnode()->is_pre_loop(), "" );
1245 Node *pre_opaq1 = pre_end->limit();
1246 // Occasionally it's possible for a pre-loop Opaque1 node to be
1247 // optimized away and then another round of loop opts attempted.
1248 // We can not optimize this particular loop in that case.
1249 if( pre_opaq1->Opcode() != Op_Opaque1 )
1251 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
1252 Node *pre_limit = pre_opaq->in(1);
1254 // Where do we put new limit calculations
1255 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
1257 // Ensure the original loop limit is available from the
1258 // pre-loop Opaque1 node.
1259 Node *orig_limit = pre_opaq->original_loop_limit();
1260 if( orig_limit == NULL || _igvn.type(orig_limit) == Type::TOP )
1263 // Need to find the main-loop zero-trip guard
1264 Node *bolzm = iffm->in(1);
1265 assert( bolzm->Opcode() == Op_Bool, "" );
1266 Node *cmpzm = bolzm->in(1);
1267 assert( cmpzm->is_Cmp(), "" );
1268 Node *opqzm = cmpzm->in(2);
1269 if( opqzm->Opcode() != Op_Opaque1 )
1271 assert( opqzm->in(1) == main_limit, "do not understand situation" );
1273 // Must know if its a count-up or count-down loop
1275 // protect against stride not being a constant
1276 if ( !cl->stride_is_con() ) {
1279 int stride_con = cl->stride_con();
1280 Node *zero = _igvn.intcon(0);
1281 Node *one = _igvn.intcon(1);
1282 set_ctrl(zero, C->root());
1283 set_ctrl(one, C->root());
1285 // Range checks that do not dominate the loop backedge (ie.
1286 // conditionally executed) can lengthen the pre loop limit beyond
1287 // the original loop limit. To prevent this, the pre limit is
1288 // (for stride > 0) MINed with the original loop limit (MAXed
1289 // stride < 0) when some range_check (rc) is conditionally
1291 bool conditional_rc = false;
1293 // Check loop body for tests of trip-counter plus loop-invariant vs
1295 for( uint i = 0; i < loop->_body.size(); i++ ) {
1296 Node *iff = loop->_body[i];
1297 if( iff->Opcode() == Op_If ) { // Test?
1299 // Test is an IfNode, has 2 projections. If BOTH are in the loop
1300 // we need loop unswitching instead of iteration splitting.
1301 Node *exit = loop->is_loop_exit(iff);
1302 if( !exit ) continue;
1303 int flip = (exit->Opcode() == Op_IfTrue) ? 1 : 0;
1305 // Get boolean condition to test
1306 Node *i1 = iff->in(1);
1307 if( !i1->is_Bool() ) continue;
1308 BoolNode *bol = i1->as_Bool();
1309 BoolTest b_test = bol->_test;
1310 // Flip sense of test if exit condition is flipped
1312 b_test = b_test.negate();
1315 Node *cmp = bol->in(1);
1317 // Look for trip_counter + offset vs limit
1318 Node *rc_exp = cmp->in(1);
1319 Node *limit = cmp->in(2);
1320 jint scale_con= 1; // Assume trip counter not scaled
1322 Node *limit_c = get_ctrl(limit);
1323 if( loop->is_member(get_loop(limit_c) ) ) {
1324 // Compare might have operands swapped; commute them
1325 b_test = b_test.commute();
1326 rc_exp = cmp->in(2);
1328 limit_c = get_ctrl(limit);
1329 if( loop->is_member(get_loop(limit_c) ) )
1330 continue; // Both inputs are loop varying; cannot RCE
1332 // Here we know 'limit' is loop invariant
1334 // 'limit' maybe pinned below the zero trip test (probably from a
1335 // previous round of rce), in which case, it can't be used in the
1336 // zero trip test expression which must occur before the zero test's if.
1337 if( limit_c == ctrl ) {
1338 continue; // Don't rce this check but continue looking for other candidates.
1341 // Check for scaled induction variable plus an offset
1342 Node *offset = NULL;
1344 if (!is_scaled_iv_plus_offset(rc_exp, trip_counter, &scale_con, &offset)) {
1348 Node *offset_c = get_ctrl(offset);
1349 if( loop->is_member( get_loop(offset_c) ) )
1350 continue; // Offset is not really loop invariant
1351 // Here we know 'offset' is loop invariant.
1353 // As above for the 'limit', the 'offset' maybe pinned below the
1355 if( offset_c == ctrl ) {
1356 continue; // Don't rce this check but continue looking for other candidates.
1359 // At this point we have the expression as:
1360 // scale_con * trip_counter + offset :: limit
1361 // where scale_con, offset and limit are loop invariant. Trip_counter
1362 // monotonically increases by stride_con, a constant. Both (or either)
1363 // stride_con and scale_con can be negative which will flip about the
1364 // sense of the test.
1366 // Adjust pre and main loop limits to guard the correct iteration set
1367 if( cmp->Opcode() == Op_CmpU ) {// Unsigned compare is really 2 tests
1368 if( b_test._test == BoolTest::lt ) { // Range checks always use lt
1369 // The overflow limit: scale*I+offset < limit
1370 add_constraint( stride_con, scale_con, offset, limit, pre_ctrl, &pre_limit, &main_limit );
1371 // The underflow limit: 0 <= scale*I+offset.
1372 // Some math yields: -scale*I-(offset+1) < 0
1373 Node *plus_one = new (C, 3) AddINode( offset, one );
1374 register_new_node( plus_one, pre_ctrl );
1375 Node *neg_offset = new (C, 3) SubINode( zero, plus_one );
1376 register_new_node( neg_offset, pre_ctrl );
1377 add_constraint( stride_con, -scale_con, neg_offset, zero, pre_ctrl, &pre_limit, &main_limit );
1378 if (!conditional_rc) {
1379 conditional_rc = !loop->dominates_backedge(iff);
1384 tty->print_cr("missed RCE opportunity");
1386 continue; // In release mode, ignore it
1388 } else { // Otherwise work on normal compares
1389 switch( b_test._test ) {
1390 case BoolTest::ge: // Convert X >= Y to -X <= -Y
1391 scale_con = -scale_con;
1392 offset = new (C, 3) SubINode( zero, offset );
1393 register_new_node( offset, pre_ctrl );
1394 limit = new (C, 3) SubINode( zero, limit );
1395 register_new_node( limit, pre_ctrl );
1396 // Fall into LE case
1397 case BoolTest::le: // Convert X <= Y to X < Y+1
1398 limit = new (C, 3) AddINode( limit, one );
1399 register_new_node( limit, pre_ctrl );
1400 // Fall into LT case
1402 add_constraint( stride_con, scale_con, offset, limit, pre_ctrl, &pre_limit, &main_limit );
1403 if (!conditional_rc) {
1404 conditional_rc = !loop->dominates_backedge(iff);
1410 tty->print_cr("missed RCE opportunity");
1412 continue; // Unhandled case
1416 // Kill the eliminated test
1417 C->set_major_progress();
1418 Node *kill_con = _igvn.intcon( 1-flip );
1419 set_ctrl(kill_con, C->root());
1420 _igvn.hash_delete(iff);
1421 iff->set_req(1, kill_con);
1422 _igvn._worklist.push(iff);
1423 // Find surviving projection
1424 assert(iff->is_If(), "");
1425 ProjNode* dp = ((IfNode*)iff)->proj_out(1-flip);
1426 // Find loads off the surviving projection; remove their control edge
1427 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
1428 Node* cd = dp->fast_out(i); // Control-dependent node
1429 if( cd->is_Load() ) { // Loads can now float around in the loop
1430 _igvn.hash_delete(cd);
1431 // Allow the load to float around in the loop, or before it
1432 // but NOT before the pre-loop.
1433 cd->set_req(0, ctrl); // ctrl, not NULL
1434 _igvn._worklist.push(cd);
1444 // Update loop limits
1445 if (conditional_rc) {
1446 pre_limit = (stride_con > 0) ? (Node*)new (C,3) MinINode(pre_limit, orig_limit)
1447 : (Node*)new (C,3) MaxINode(pre_limit, orig_limit);
1448 register_new_node(pre_limit, pre_ctrl);
1450 _igvn.hash_delete(pre_opaq);
1451 pre_opaq->set_req(1, pre_limit);
1453 // Note:: we are making the main loop limit no longer precise;
1454 // need to round up based on stride.
1455 if( stride_con != 1 && stride_con != -1 ) { // Cutout for common case
1456 // "Standard" round-up logic: ([main_limit-init+(y-1)]/y)*y+init
1457 // Hopefully, compiler will optimize for powers of 2.
1458 Node *ctrl = get_ctrl(main_limit);
1459 Node *stride = cl->stride();
1460 Node *init = cl->init_trip();
1461 Node *span = new (C, 3) SubINode(main_limit,init);
1462 register_new_node(span,ctrl);
1463 Node *rndup = _igvn.intcon(stride_con + ((stride_con>0)?-1:1));
1464 Node *add = new (C, 3) AddINode(span,rndup);
1465 register_new_node(add,ctrl);
1466 Node *div = new (C, 3) DivINode(0,add,stride);
1467 register_new_node(div,ctrl);
1468 Node *mul = new (C, 3) MulINode(div,stride);
1469 register_new_node(mul,ctrl);
1470 Node *newlim = new (C, 3) AddINode(mul,init);
1471 register_new_node(newlim,ctrl);
1472 main_limit = newlim;
1475 Node *main_cle = cl->loopexit();
1476 Node *main_bol = main_cle->in(1);
1477 // Hacking loop bounds; need private copies of exit test
1478 if( main_bol->outcnt() > 1 ) {// BoolNode shared?
1479 _igvn.hash_delete(main_cle);
1480 main_bol = main_bol->clone();// Clone a private BoolNode
1481 register_new_node( main_bol, main_cle->in(0) );
1482 main_cle->set_req(1,main_bol);
1484 Node *main_cmp = main_bol->in(1);
1485 if( main_cmp->outcnt() > 1 ) { // CmpNode shared?
1486 _igvn.hash_delete(main_bol);
1487 main_cmp = main_cmp->clone();// Clone a private CmpNode
1488 register_new_node( main_cmp, main_cle->in(0) );
1489 main_bol->set_req(1,main_cmp);
1491 // Hack the now-private loop bounds
1492 _igvn.hash_delete(main_cmp);
1493 main_cmp->set_req(2, main_limit);
1494 _igvn._worklist.push(main_cmp);
1495 // The OpaqueNode is unshared by design
1496 _igvn.hash_delete(opqzm);
1497 assert( opqzm->outcnt() == 1, "cannot hack shared node" );
1498 opqzm->set_req(1,main_limit);
1499 _igvn._worklist.push(opqzm);
1502 //------------------------------DCE_loop_body----------------------------------
1503 // Remove simplistic dead code from loop body
1504 void IdealLoopTree::DCE_loop_body() {
1505 for( uint i = 0; i < _body.size(); i++ )
1506 if( _body.at(i)->outcnt() == 0 )
1507 _body.map( i--, _body.pop() );
1511 //------------------------------adjust_loop_exit_prob--------------------------
1512 // Look for loop-exit tests with the 50/50 (or worse) guesses from the parsing stage.
1513 // Replace with a 1-in-10 exit guess.
1514 void IdealLoopTree::adjust_loop_exit_prob( PhaseIdealLoop *phase ) {
1515 Node *test = tail();
1516 while( test != _head ) {
1517 uint top = test->Opcode();
1518 if( top == Op_IfTrue || top == Op_IfFalse ) {
1519 int test_con = ((ProjNode*)test)->_con;
1520 assert(top == (uint)(test_con? Op_IfTrue: Op_IfFalse), "sanity");
1521 IfNode *iff = test->in(0)->as_If();
1522 if( iff->outcnt() == 2 ) { // Ignore dead tests
1523 Node *bol = iff->in(1);
1524 if( bol && bol->req() > 1 && bol->in(1) &&
1525 ((bol->in(1)->Opcode() == Op_StorePConditional ) ||
1526 (bol->in(1)->Opcode() == Op_StoreIConditional ) ||
1527 (bol->in(1)->Opcode() == Op_StoreLConditional ) ||
1528 (bol->in(1)->Opcode() == Op_CompareAndSwapI ) ||
1529 (bol->in(1)->Opcode() == Op_CompareAndSwapL ) ||
1530 (bol->in(1)->Opcode() == Op_CompareAndSwapP ) ||
1531 (bol->in(1)->Opcode() == Op_CompareAndSwapN )))
1532 return; // Allocation loops RARELY take backedge
1533 // Find the OTHER exit path from the IF
1534 Node* ex = iff->proj_out(1-test_con);
1535 float p = iff->_prob;
1536 if( !phase->is_member( this, ex ) && iff->_fcnt == COUNT_UNKNOWN ) {
1537 if( top == Op_IfTrue ) {
1538 if( p < (PROB_FAIR + PROB_UNLIKELY_MAG(3))) {
1539 iff->_prob = PROB_STATIC_FREQUENT;
1542 if( p > (PROB_FAIR - PROB_UNLIKELY_MAG(3))) {
1543 iff->_prob = PROB_STATIC_INFREQUENT;
1549 test = phase->idom(test);
1554 //------------------------------policy_do_remove_empty_loop--------------------
1555 // Micro-benchmark spamming. Policy is to always remove empty loops.
1556 // The 'DO' part is to replace the trip counter with the value it will
1557 // have on the last iteration. This will break the loop.
1558 bool IdealLoopTree::policy_do_remove_empty_loop( PhaseIdealLoop *phase ) {
1559 // Minimum size must be empty loop
1560 if( _body.size() > 7/*number of nodes in an empty loop*/ ) return false;
1562 if( !_head->is_CountedLoop() ) return false; // Dead loop
1563 CountedLoopNode *cl = _head->as_CountedLoop();
1564 if( !cl->loopexit() ) return false; // Malformed loop
1565 if( !phase->is_member(this,phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue)) ) )
1566 return false; // Infinite loop
1569 tty->print_cr("Removing empty loop");
1572 // Ensure only one phi which is the iv.
1574 for (DUIterator_Fast imax, i = cl->fast_outs(imax); i < imax; i++) {
1575 Node* n = cl->fast_out(i);
1576 if (n->Opcode() == Op_Phi) {
1577 assert(iv == NULL, "Too many phis" );
1581 assert(iv == cl->phi(), "Wrong phi" );
1583 // Replace the phi at loop head with the final value of the last
1584 // iteration. Then the CountedLoopEnd will collapse (backedge never
1585 // taken) and all loop-invariant uses of the exit values will be correct.
1586 Node *phi = cl->phi();
1587 Node *final = new (phase->C, 3) SubINode( cl->limit(), cl->stride() );
1588 phase->register_new_node(final,cl->in(LoopNode::EntryControl));
1589 phase->_igvn.hash_delete(phi);
1590 phase->_igvn.subsume_node(phi,final);
1591 phase->C->set_major_progress();
1596 //=============================================================================
1597 //------------------------------iteration_split_impl---------------------------
1598 bool IdealLoopTree::iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ) {
1599 // Check and remove empty loops (spam micro-benchmarks)
1600 if( policy_do_remove_empty_loop(phase) )
1601 return true; // Here we removed an empty loop
1603 bool should_peel = policy_peeling(phase); // Should we peel?
1605 bool should_unswitch = policy_unswitching(phase);
1607 // Non-counted loops may be peeled; exactly 1 iteration is peeled.
1608 // This removes loop-invariant tests (usually null checks).
1609 if( !_head->is_CountedLoop() ) { // Non-counted loop
1610 if (PartialPeelLoop && phase->partial_peel(this, old_new)) {
1611 // Partial peel succeeded so terminate this round of loop opts
1614 if( should_peel ) { // Should we peel?
1616 if (PrintOpto) tty->print_cr("should_peel");
1618 phase->do_peeling(this,old_new);
1619 } else if( should_unswitch ) {
1620 phase->do_unswitching(this, old_new);
1624 CountedLoopNode *cl = _head->as_CountedLoop();
1626 if( !cl->loopexit() ) return true; // Ignore various kinds of broken loops
1628 // Do nothing special to pre- and post- loops
1629 if( cl->is_pre_loop() || cl->is_post_loop() ) return true;
1631 // Compute loop trip count from profile data
1632 compute_profile_trip_cnt(phase);
1634 // Before attempting fancy unrolling, RCE or alignment, see if we want
1635 // to completely unroll this loop or do loop unswitching.
1636 if( cl->is_normal_loop() ) {
1637 if (should_unswitch) {
1638 phase->do_unswitching(this, old_new);
1641 bool should_maximally_unroll = policy_maximally_unroll(phase);
1642 if( should_maximally_unroll ) {
1643 // Here we did some unrolling and peeling. Eventually we will
1644 // completely unroll this loop and it will no longer be a loop.
1645 phase->do_maximally_unroll(this,old_new);
1651 // Counted loops may be peeled, may need some iterations run up
1652 // front for RCE, and may want to align loop refs to a cache
1653 // line. Thus we clone a full loop up front whose trip count is
1654 // at least 1 (if peeling), but may be several more.
1656 // The main loop will start cache-line aligned with at least 1
1657 // iteration of the unrolled body (zero-trip test required) and
1658 // will have some range checks removed.
1660 // A post-loop will finish any odd iterations (leftover after
1661 // unrolling), plus any needed for RCE purposes.
1663 bool should_unroll = policy_unroll(phase);
1665 bool should_rce = policy_range_check(phase);
1667 bool should_align = policy_align(phase);
1669 // If not RCE'ing (iteration splitting) or Aligning, then we do not
1670 // need a pre-loop. We may still need to peel an initial iteration but
1671 // we will not be needing an unknown number of pre-iterations.
1673 // Basically, if may_rce_align reports FALSE first time through,
1674 // we will not be able to later do RCE or Aligning on this loop.
1675 bool may_rce_align = !policy_peel_only(phase) || should_rce || should_align;
1677 // If we have any of these conditions (RCE, alignment, unrolling) met, then
1678 // we switch to the pre-/main-/post-loop model. This model also covers
1680 if( should_rce || should_align || should_unroll ) {
1681 if( cl->is_normal_loop() ) // Convert to 'pre/main/post' loops
1682 phase->insert_pre_post_loops(this,old_new, !may_rce_align);
1684 // Adjust the pre- and main-loop limits to let the pre and post loops run
1685 // with full checks, but the main-loop with no checks. Remove said
1686 // checks from the main body.
1688 phase->do_range_check(this,old_new);
1690 // Double loop body for unrolling. Adjust the minimum-trip test (will do
1691 // twice as many iterations as before) and the main body limit (only do
1692 // an even number of trips). If we are peeling, we might enable some RCE
1693 // and we'd rather unroll the post-RCE'd loop SO... do not unroll if
1695 if( should_unroll && !should_peel )
1696 phase->do_unroll(this,old_new, true);
1698 // Adjust the pre-loop limits to align the main body
1703 } else { // Else we have an unchanged counted loop
1704 if( should_peel ) // Might want to peel but do nothing else
1705 phase->do_peeling(this,old_new);
1711 //=============================================================================
1712 //------------------------------iteration_split--------------------------------
1713 bool IdealLoopTree::iteration_split( PhaseIdealLoop *phase, Node_List &old_new ) {
1714 // Recursively iteration split nested loops
1715 if( _child && !_child->iteration_split( phase, old_new ))
1718 // Clean out prior deadwood
1722 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
1723 // Replace with a 1-in-10 exit guess.
1724 if( _parent /*not the root loop*/ &&
1726 // Also ignore the occasional dead backedge
1727 !tail()->is_top() ) {
1728 adjust_loop_exit_prob(phase);
1732 // Gate unrolling, RCE and peeling efforts.
1733 if( !_child && // If not an inner loop, do not split
1735 _allow_optimizations &&
1736 !tail()->is_top() ) { // Also ignore the occasional dead backedge
1738 if (!iteration_split_impl( phase, old_new )) {
1741 } else if (policy_unswitching(phase)) {
1742 phase->do_unswitching(this, old_new);
1746 // Minor offset re-organization to remove loop-fallout uses of
1748 if( _head->is_CountedLoop() ) phase->reorg_offsets( this );
1749 if( _next && !_next->iteration_split( phase, old_new ))
1754 //-------------------------------is_uncommon_trap_proj----------------------------
1755 // Return true if proj is the form of "proj->[region->..]call_uct"
1756 bool PhaseIdealLoop::is_uncommon_trap_proj(ProjNode* proj, bool must_reason_predicate) {
1757 int path_limit = 10;
1758 assert(proj, "invalid argument");
1760 for (int ct = 0; ct < path_limit; ct++) {
1761 out = out->unique_ctrl_out();
1762 if (out == NULL || out->is_Root() || out->is_Start())
1764 if (out->is_CallStaticJava()) {
1765 int req = out->as_CallStaticJava()->uncommon_trap_request();
1767 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(req);
1768 if (!must_reason_predicate || reason == Deoptimization::Reason_predicate){
1772 return false; // don't do further after call
1778 //-------------------------------is_uncommon_trap_if_pattern-------------------------
1779 // Return true for "if(test)-> proj -> ...
1782 // other_proj->[region->..]call_uct"
1784 // "must_reason_predicate" means the uct reason must be Reason_predicate
1785 bool PhaseIdealLoop::is_uncommon_trap_if_pattern(ProjNode *proj, bool must_reason_predicate) {
1786 Node *in0 = proj->in(0);
1787 if (!in0->is_If()) return false;
1788 IfNode* iff = in0->as_If();
1790 // we need "If(Conv2B(Opaque1(...)))" pattern for must_reason_predicate
1791 if (must_reason_predicate) {
1792 if (iff->in(1)->Opcode() != Op_Conv2B ||
1793 iff->in(1)->in(1)->Opcode() != Op_Opaque1) {
1798 ProjNode* other_proj = iff->proj_out(1-proj->_con)->as_Proj();
1799 return is_uncommon_trap_proj(other_proj, must_reason_predicate);
1802 //------------------------------create_new_if_for_predicate------------------------
1803 // create a new if above the uct_if_pattern for the predicate to be promoted.
1806 // ---------- ----------
1815 // uncommon_proj cont_proj if_uct if_cont
1823 // uncommon_trap | uncommon_proj cont_proj
1834 // We will create a region to guard the uct call if there is no one there.
1835 // The true projecttion (if_cont) of the new_iff is returned.
1836 ProjNode* PhaseIdealLoop::create_new_if_for_predicate(ProjNode* cont_proj) {
1837 assert(is_uncommon_trap_if_pattern(cont_proj, true), "must be a uct if pattern!");
1838 IfNode* iff = cont_proj->in(0)->as_If();
1840 ProjNode *uncommon_proj = iff->proj_out(1 - cont_proj->_con);
1841 Node *rgn = uncommon_proj->unique_ctrl_out();
1842 assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct");
1844 if (!rgn->is_Region()) { // create a region to guard the call
1845 assert(rgn->is_Call(), "must be call uct");
1846 CallNode* call = rgn->as_Call();
1847 rgn = new (C, 1) RegionNode(1);
1848 _igvn.set_type(rgn, rgn->bottom_type());
1849 rgn->add_req(uncommon_proj);
1850 set_idom(rgn, idom(uncommon_proj), dom_depth(uncommon_proj)+1);
1851 _igvn.hash_delete(call);
1852 call->set_req(0, rgn);
1856 uint iffdd = dom_depth(iff);
1857 IdealLoopTree* lp = get_loop(iff);
1858 IfNode *new_iff = new (C, 2) IfNode(iff->in(0), NULL, iff->_prob, iff->_fcnt);
1859 register_node(new_iff, lp, idom(iff), iffdd);
1860 Node *if_cont = new (C, 1) IfTrueNode(new_iff);
1861 Node *if_uct = new (C, 1) IfFalseNode(new_iff);
1862 if (cont_proj->is_IfFalse()) {
1864 Node* tmp = if_uct; if_uct = if_cont; if_cont = tmp;
1866 register_node(if_cont, lp, new_iff, iffdd);
1867 register_node(if_uct, get_loop(rgn), new_iff, iffdd);
1870 _igvn.hash_delete(iff);
1871 iff->set_req(0, if_cont);
1872 set_idom(iff, if_cont, dom_depth(iff));
1875 _igvn.hash_delete(rgn);
1876 rgn->add_req(if_uct);
1877 Node* ridom = idom(rgn);
1878 Node* nrdom = dom_lca(ridom, new_iff);
1879 set_idom(rgn, nrdom, dom_depth(rgn));
1881 // rgn must have no phis
1882 assert(!rgn->as_Region()->has_phi(), "region must have no phis");
1884 return if_cont->as_Proj();
1887 //------------------------------find_predicate_insertion_point--------------------------
1888 // Find a good location to insert a predicate
1889 ProjNode* PhaseIdealLoop::find_predicate_insertion_point(Node* start_c) {
1890 if (start_c == C->root() || !start_c->is_Proj())
1892 if (is_uncommon_trap_if_pattern(start_c->as_Proj(), true/*Reason_Predicate*/)) {
1893 return start_c->as_Proj();
1898 //------------------------------Invariance-----------------------------------
1899 // Helper class for loop_predication_impl to compute invariance on the fly and
1900 // clone invariants.
1901 class Invariance : public StackObj {
1902 VectorSet _visited, _invariant;
1904 VectorSet _clone_visited;
1905 Node_List _old_new; // map of old to new (clone)
1906 IdealLoopTree* _lpt;
1907 PhaseIdealLoop* _phase;
1909 // Helper function to set up the invariance for invariance computation
1910 // If n is a known invariant, set up directly. Otherwise, look up the
1911 // the possibility to push n onto the stack for further processing.
1912 void visit(Node* use, Node* n) {
1913 if (_lpt->is_invariant(n)) { // known invariant
1914 _invariant.set(n->_idx);
1915 } else if (!n->is_CFG()) {
1916 Node *n_ctrl = _phase->ctrl_or_self(n);
1917 Node *u_ctrl = _phase->ctrl_or_self(use); // self if use is a CFG
1918 if (_phase->is_dominator(n_ctrl, u_ctrl)) {
1919 _stack.push(n, n->in(0) == NULL ? 1 : 0);
1924 // Compute invariance for "the_node" and (possibly) all its inputs recursively
1926 void compute_invariance(Node* n) {
1927 assert(_visited.test(n->_idx), "must be");
1929 while (_stack.is_nonempty()) {
1930 Node* n = _stack.node();
1931 uint idx = _stack.index();
1932 if (idx == n->req()) { // all inputs are processed
1934 // n is invariant if it's inputs are all invariant
1935 bool all_inputs_invariant = true;
1936 for (uint i = 0; i < n->req(); i++) {
1937 Node* in = n->in(i);
1938 if (in == NULL) continue;
1939 assert(_visited.test(in->_idx), "must have visited input");
1940 if (!_invariant.test(in->_idx)) { // bad guy
1941 all_inputs_invariant = false;
1945 if (all_inputs_invariant) {
1946 _invariant.set(n->_idx); // I am a invariant too
1948 } else { // process next input
1949 _stack.set_index(idx + 1);
1950 Node* m = n->in(idx);
1951 if (m != NULL && !_visited.test_set(m->_idx)) {
1958 // Helper function to set up _old_new map for clone_nodes.
1959 // If n is a known invariant, set up directly ("clone" of n == n).
1960 // Otherwise, push n onto the stack for real cloning.
1961 void clone_visit(Node* n) {
1962 assert(_invariant.test(n->_idx), "must be invariant");
1963 if (_lpt->is_invariant(n)) { // known invariant
1964 _old_new.map(n->_idx, n);
1965 } else{ // to be cloned
1966 assert (!n->is_CFG(), "should not see CFG here");
1967 _stack.push(n, n->in(0) == NULL ? 1 : 0);
1971 // Clone "n" and (possibly) all its inputs recursively
1972 void clone_nodes(Node* n, Node* ctrl) {
1974 while (_stack.is_nonempty()) {
1975 Node* n = _stack.node();
1976 uint idx = _stack.index();
1977 if (idx == n->req()) { // all inputs processed, clone n!
1979 // clone invariant node
1980 Node* n_cl = n->clone();
1981 _old_new.map(n->_idx, n_cl);
1982 _phase->register_new_node(n_cl, ctrl);
1983 for (uint i = 0; i < n->req(); i++) {
1984 Node* in = n_cl->in(i);
1985 if (in == NULL) continue;
1986 n_cl->set_req(i, _old_new[in->_idx]);
1988 } else { // process next input
1989 _stack.set_index(idx + 1);
1990 Node* m = n->in(idx);
1991 if (m != NULL && !_clone_visited.test_set(m->_idx)) {
1992 clone_visit(m); // visit the input
1999 Invariance(Arena* area, IdealLoopTree* lpt) :
2000 _lpt(lpt), _phase(lpt->_phase),
2001 _visited(area), _invariant(area), _stack(area, 10 /* guess */),
2002 _clone_visited(area), _old_new(area)
2005 // Map old to n for invariance computation and clone
2006 void map_ctrl(Node* old, Node* n) {
2007 assert(old->is_CFG() && n->is_CFG(), "must be");
2008 _old_new.map(old->_idx, n); // "clone" of old is n
2009 _invariant.set(old->_idx); // old is invariant
2010 _clone_visited.set(old->_idx);
2013 // Driver function to compute invariance
2014 bool is_invariant(Node* n) {
2015 if (!_visited.test_set(n->_idx))
2016 compute_invariance(n);
2017 return (_invariant.test(n->_idx) != 0);
2020 // Driver function to clone invariant
2021 Node* clone(Node* n, Node* ctrl) {
2022 assert(ctrl->is_CFG(), "must be");
2023 assert(_invariant.test(n->_idx), "must be an invariant");
2024 if (!_clone_visited.test(n->_idx))
2025 clone_nodes(n, ctrl);
2026 return _old_new[n->_idx];
2030 //------------------------------is_range_check_if -----------------------------------
2031 // Returns true if the predicate of iff is in "scale*iv + offset u< load_range(ptr)" format
2032 // Note: this function is particularly designed for loop predication. We require load_range
2033 // and offset to be loop invariant computed on the fly by "invar"
2034 bool IdealLoopTree::is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const {
2035 if (!is_loop_exit(iff)) {
2038 if (!iff->in(1)->is_Bool()) {
2041 const BoolNode *bol = iff->in(1)->as_Bool();
2042 if (bol->_test._test != BoolTest::lt) {
2045 if (!bol->in(1)->is_Cmp()) {
2048 const CmpNode *cmp = bol->in(1)->as_Cmp();
2049 if (cmp->Opcode() != Op_CmpU ) {
2052 if (cmp->in(2)->Opcode() != Op_LoadRange) {
2055 LoadRangeNode* lr = (LoadRangeNode*)cmp->in(2);
2056 if (!invar.is_invariant(lr)) { // loadRange must be invariant
2059 Node *iv = _head->as_CountedLoop()->phi();
2061 Node *offset = NULL;
2062 if (!phase->is_scaled_iv_plus_offset(cmp->in(1), iv, &scale, &offset)) {
2065 if(offset && !invar.is_invariant(offset)) { // offset must be invariant
2071 //------------------------------rc_predicate-----------------------------------
2072 // Create a range check predicate
2074 // for (i = init; i < limit; i += stride) {
2075 // a[scale*i+offset]
2078 // Compute max(scale*i + offset) for init <= i < limit and build the predicate
2079 // as "max(scale*i + offset) u< a.length".
2081 // There are two cases for max(scale*i + offset):
2082 // (1) stride*scale > 0
2083 // max(scale*i + offset) = scale*(limit-stride) + offset
2084 // (2) stride*scale < 0
2085 // max(scale*i + offset) = scale*init + offset
2086 BoolNode* PhaseIdealLoop::rc_predicate(Node* ctrl,
2087 int scale, Node* offset,
2088 Node* init, Node* limit, Node* stride,
2090 Node* max_idx_expr = init;
2091 int stride_con = stride->get_int();
2092 if ((stride_con > 0) == (scale > 0)) {
2093 max_idx_expr = new (C, 3) SubINode(limit, stride);
2094 register_new_node(max_idx_expr, ctrl);
2098 ConNode* con_scale = _igvn.intcon(scale);
2099 max_idx_expr = new (C, 3) MulINode(max_idx_expr, con_scale);
2100 register_new_node(max_idx_expr, ctrl);
2103 if (offset && (!offset->is_Con() || offset->get_int() != 0)){
2104 max_idx_expr = new (C, 3) AddINode(max_idx_expr, offset);
2105 register_new_node(max_idx_expr, ctrl);
2108 CmpUNode* cmp = new (C, 3) CmpUNode(max_idx_expr, range);
2109 register_new_node(cmp, ctrl);
2110 BoolNode* bol = new (C, 2) BoolNode(cmp, BoolTest::lt);
2111 register_new_node(bol, ctrl);
2115 //------------------------------ loop_predication_impl--------------------------
2116 // Insert loop predicates for null checks and range checks
2117 bool PhaseIdealLoop::loop_predication_impl(IdealLoopTree *loop) {
2118 if (!UseLoopPredicate) return false;
2120 // Too many traps seen?
2121 bool tmt = C->too_many_traps(C->method(), 0, Deoptimization::Reason_predicate);
2122 int tc = C->trap_count(Deoptimization::Reason_predicate);
2123 if (tmt || tc > 0) {
2124 if (TraceLoopPredicate) {
2125 tty->print_cr("too many predicate traps: %d", tc);
2126 C->method()->print(); // which method has too many predicate traps
2132 CountedLoopNode *cl = NULL;
2133 if (loop->_head->is_CountedLoop()) {
2134 cl = loop->_head->as_CountedLoop();
2135 // do nothing for iteration-splitted loops
2136 if(!cl->is_normal_loop()) return false;
2139 LoopNode *lpn = loop->_head->as_Loop();
2140 Node* entry = lpn->in(LoopNode::EntryControl);
2142 ProjNode *predicate_proj = find_predicate_insertion_point(entry);
2143 if (!predicate_proj){
2145 if (TraceLoopPredicate) {
2146 tty->print("missing predicate:");
2153 ConNode* zero = _igvn.intcon(0);
2154 set_ctrl(zero, C->root());
2155 Node *cond_false = new (C, 2) Conv2BNode(zero);
2156 register_new_node(cond_false, C->root());
2157 ConNode* one = _igvn.intcon(1);
2158 set_ctrl(one, C->root());
2159 Node *cond_true = new (C, 2) Conv2BNode(one);
2160 register_new_node(cond_true, C->root());
2162 ResourceArea *area = Thread::current()->resource_area();
2163 Invariance invar(area, loop);
2165 // Create list of if-projs such that a newer proj dominates all older
2166 // projs in the list, and they all dominate loop->tail()
2167 Node_List if_proj_list(area);
2168 LoopNode *head = loop->_head->as_Loop();
2169 Node *current_proj = loop->tail(); //start from tail
2170 while ( current_proj != head ) {
2171 if (loop == get_loop(current_proj) && // still in the loop ?
2172 current_proj->is_Proj() && // is a projection ?
2173 current_proj->in(0)->Opcode() == Op_If) { // is a if projection ?
2174 if_proj_list.push(current_proj);
2176 current_proj = idom(current_proj);
2179 bool hoisted = false; // true if at least one proj is promoted
2180 while (if_proj_list.size() > 0) {
2181 // Following are changed to nonnull when a predicate can be hoisted
2182 ProjNode* new_predicate_proj = NULL;
2183 BoolNode* new_predicate_bol = NULL;
2185 ProjNode* proj = if_proj_list.pop()->as_Proj();
2186 IfNode* iff = proj->in(0)->as_If();
2188 if (!is_uncommon_trap_if_pattern(proj)) {
2189 if (loop->is_loop_exit(iff)) {
2190 // stop processing the remaining projs in the list because the execution of them
2191 // depends on the condition of "iff" (iff->in(1)).
2194 // Both arms are inside the loop. There are two cases:
2195 // (1) there is one backward branch. In this case, any remaining proj
2196 // in the if_proj list post-dominates "iff". So, the condition of "iff"
2197 // does not determine the execution the remining projs directly, and we
2198 // can safely continue.
2199 // (2) both arms are forwarded, i.e. a diamond shape. In this case, "proj"
2200 // does not dominate loop->tail(), so it can not be in the if_proj list.
2205 Node* test = iff->in(1);
2206 if (!test->is_Bool()){ //Conv2B, ...
2209 BoolNode* bol = test->as_Bool();
2210 if (invar.is_invariant(bol)) {
2212 new_predicate_proj = create_new_if_for_predicate(predicate_proj);
2213 Node* ctrl = new_predicate_proj->in(0)->as_If()->in(0);
2214 new_predicate_bol = invar.clone(bol, ctrl)->as_Bool();
2215 if (TraceLoopPredicate) tty->print("invariant");
2216 } else if (cl != NULL && loop->is_range_check_if(iff, this, invar)) {
2217 // Range check (only for counted loops)
2218 new_predicate_proj = create_new_if_for_predicate(predicate_proj);
2219 Node *ctrl = new_predicate_proj->in(0)->as_If()->in(0);
2220 const Node* cmp = bol->in(1)->as_Cmp();
2221 Node* idx = cmp->in(1);
2222 assert(!invar.is_invariant(idx), "index is variant");
2223 assert(cmp->in(2)->Opcode() == Op_LoadRange, "must be");
2224 LoadRangeNode* ld_rng = (LoadRangeNode*)cmp->in(2); // LoadRangeNode
2225 assert(invar.is_invariant(ld_rng), "load range must be invariant");
2226 ld_rng = (LoadRangeNode*)invar.clone(ld_rng, ctrl);
2228 Node* offset = zero;
2229 bool ok = is_scaled_iv_plus_offset(idx, cl->phi(), &scale, &offset);
2230 assert(ok, "must be index expression");
2231 if (offset && offset != zero) {
2232 assert(invar.is_invariant(offset), "offset must be loop invariant");
2233 offset = invar.clone(offset, ctrl);
2235 Node* init = cl->init_trip();
2236 Node* limit = cl->limit();
2237 Node* stride = cl->stride();
2238 new_predicate_bol = rc_predicate(ctrl, scale, offset, init, limit, stride, ld_rng);
2239 if (TraceLoopPredicate) tty->print("range check");
2242 if (new_predicate_proj == NULL) {
2243 // The other proj of the "iff" is a uncommon trap projection, and we can assume
2244 // the other proj will not be executed ("executed" means uct raised).
2247 // Success - attach condition (new_predicate_bol) to predicate if
2248 invar.map_ctrl(proj, new_predicate_proj); // so that invariance test can be appropriate
2249 IfNode* new_iff = new_predicate_proj->in(0)->as_If();
2251 // Negate test if necessary
2252 if (proj->_con != predicate_proj->_con) {
2253 new_predicate_bol = new (C, 2) BoolNode(new_predicate_bol->in(1), new_predicate_bol->_test.negate());
2254 register_new_node(new_predicate_bol, new_iff->in(0));
2255 if (TraceLoopPredicate) tty->print_cr(" if negated: %d", iff->_idx);
2257 if (TraceLoopPredicate) tty->print_cr(" if: %d", iff->_idx);
2260 _igvn.hash_delete(new_iff);
2261 new_iff->set_req(1, new_predicate_bol);
2263 _igvn.hash_delete(iff);
2264 iff->set_req(1, proj->is_IfFalse() ? cond_false : cond_true);
2266 Node* ctrl = new_predicate_proj; // new control
2267 ProjNode* dp = proj; // old control
2268 assert(get_loop(dp) == loop, "guarenteed at the time of collecting proj");
2269 // Find nodes (depends only on the test) off the surviving projection;
2270 // move them outside the loop with the control of proj_clone
2271 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
2272 Node* cd = dp->fast_out(i); // Control-dependent node
2273 if (cd->depends_only_on_test()) {
2274 assert(cd->in(0) == dp, "");
2275 _igvn.hash_delete(cd);
2276 cd->set_req(0, ctrl); // ctrl, not NULL
2278 _igvn._worklist.push(cd);
2279 IdealLoopTree *new_loop = get_loop(get_ctrl(cd));
2280 if (new_loop != loop) {
2281 if (!loop->_child) loop->_body.yank(cd);
2282 if (!new_loop->_child ) new_loop->_body.push(cd);
2290 C->set_major_progress();
2295 // report that the loop predication has been actually performed
2297 if (TraceLoopPredicate && hoisted) {
2298 tty->print("Loop Predication Performed:");
2306 //------------------------------loop_predication--------------------------------
2307 // driver routine for loop predication optimization
2308 bool IdealLoopTree::loop_predication( PhaseIdealLoop *phase) {
2309 bool hoisted = false;
2310 // Recursively promote predicates
2312 hoisted = _child->loop_predication( phase);
2316 if (!_irreducible && !tail()->is_top()) {
2317 hoisted |= phase->loop_predication_impl(this);
2320 if ( _next ) { //sibling
2321 hoisted |= _next->loop_predication( phase);