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 Node *cmp = bol->in(1);
554 Node *rc_exp = cmp->in(1);
555 Node *limit = cmp->in(2);
557 Node *limit_c = phase->get_ctrl(limit);
558 if( limit_c == phase->C->top() )
559 return false; // Found dead test on live IF? No RCE!
560 if( is_member(phase->get_loop(limit_c) ) ) {
561 // Compare might have operands swapped; commute them
564 limit_c = phase->get_ctrl(limit);
565 if( is_member(phase->get_loop(limit_c) ) )
566 continue; // Both inputs are loop varying; cannot RCE
569 if (!phase->is_scaled_iv_plus_offset(rc_exp, trip_counter, NULL, NULL)) {
572 // Yeah! Found a test like 'trip+off vs limit'
573 // Test is an IfNode, has 2 projections. If BOTH are in the loop
574 // we need loop unswitching instead of iteration splitting.
575 if( is_loop_exit(iff) )
576 return true; // Found reason to split iterations
583 //------------------------------policy_peel_only-------------------------------
584 // Return TRUE or FALSE if the loop should NEVER be RCE'd or aligned. Useful
585 // for unrolling loops with NO array accesses.
586 bool IdealLoopTree::policy_peel_only( PhaseIdealLoop *phase ) const {
588 for( uint i = 0; i < _body.size(); i++ )
589 if( _body[i]->is_Mem() )
592 // No memory accesses at all!
596 //------------------------------clone_up_backedge_goo--------------------------
597 // If Node n lives in the back_ctrl block and cannot float, we clone a private
598 // version of n in preheader_ctrl block and return that, otherwise return n.
599 Node *PhaseIdealLoop::clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n ) {
600 if( get_ctrl(n) != back_ctrl ) return n;
602 Node *x = NULL; // If required, a clone of 'n'
603 // Check for 'n' being pinned in the backedge.
604 if( n->in(0) && n->in(0) == back_ctrl ) {
605 x = n->clone(); // Clone a copy of 'n' to preheader
606 x->set_req( 0, preheader_ctrl ); // Fix x's control input to preheader
609 // Recursive fixup any other input edges into x.
610 // If there are no changes we can just return 'n', otherwise
611 // we need to clone a private copy and change it.
612 for( uint i = 1; i < n->req(); i++ ) {
613 Node *g = clone_up_backedge_goo( back_ctrl, preheader_ctrl, n->in(i) );
614 if( g != n->in(i) ) {
620 if( x ) { // x can legally float to pre-header location
621 register_new_node( x, preheader_ctrl );
623 } else { // raise n to cover LCA of uses
624 set_ctrl( n, find_non_split_ctrl(back_ctrl->in(0)) );
629 //------------------------------insert_pre_post_loops--------------------------
630 // Insert pre and post loops. If peel_only is set, the pre-loop can not have
631 // more iterations added. It acts as a 'peel' only, no lower-bound RCE, no
632 // alignment. Useful to unroll loops that do no array accesses.
633 void PhaseIdealLoop::insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ) {
635 C->set_major_progress();
637 // Find common pieces of the loop being guarded with pre & post loops
638 CountedLoopNode *main_head = loop->_head->as_CountedLoop();
639 assert( main_head->is_normal_loop(), "" );
640 CountedLoopEndNode *main_end = main_head->loopexit();
641 assert( main_end->outcnt() == 2, "1 true, 1 false path only" );
642 uint dd_main_head = dom_depth(main_head);
643 uint max = main_head->outcnt();
645 Node *pre_header= main_head->in(LoopNode::EntryControl);
646 Node *init = main_head->init_trip();
647 Node *incr = main_end ->incr();
648 Node *limit = main_end ->limit();
649 Node *stride = main_end ->stride();
650 Node *cmp = main_end ->cmp_node();
651 BoolTest::mask b_test = main_end->test_trip();
653 // Need only 1 user of 'bol' because I will be hacking the loop bounds.
654 Node *bol = main_end->in(CountedLoopEndNode::TestValue);
655 if( bol->outcnt() != 1 ) {
657 register_new_node(bol,main_end->in(CountedLoopEndNode::TestControl));
658 _igvn.hash_delete(main_end);
659 main_end->set_req(CountedLoopEndNode::TestValue, bol);
661 // Need only 1 user of 'cmp' because I will be hacking the loop bounds.
662 if( cmp->outcnt() != 1 ) {
664 register_new_node(cmp,main_end->in(CountedLoopEndNode::TestControl));
665 _igvn.hash_delete(bol);
666 bol->set_req(1, cmp);
669 //------------------------------
670 // Step A: Create Post-Loop.
671 Node* main_exit = main_end->proj_out(false);
672 assert( main_exit->Opcode() == Op_IfFalse, "" );
673 int dd_main_exit = dom_depth(main_exit);
675 // Step A1: Clone the loop body. The clone becomes the post-loop. The main
676 // loop pre-header illegally has 2 control users (old & new loops).
677 clone_loop( loop, old_new, dd_main_exit );
678 assert( old_new[main_end ->_idx]->Opcode() == Op_CountedLoopEnd, "" );
679 CountedLoopNode *post_head = old_new[main_head->_idx]->as_CountedLoop();
680 post_head->set_post_loop(main_head);
682 // Reduce the post-loop trip count.
683 CountedLoopEndNode* post_end = old_new[main_end ->_idx]->as_CountedLoopEnd();
684 post_end->_prob = PROB_FAIR;
686 // Build the main-loop normal exit.
687 IfFalseNode *new_main_exit = new (C, 1) IfFalseNode(main_end);
688 _igvn.register_new_node_with_optimizer( new_main_exit );
689 set_idom(new_main_exit, main_end, dd_main_exit );
690 set_loop(new_main_exit, loop->_parent);
692 // Step A2: Build a zero-trip guard for the post-loop. After leaving the
693 // main-loop, the post-loop may not execute at all. We 'opaque' the incr
694 // (the main-loop trip-counter exit value) because we will be changing
695 // the exit value (via unrolling) so we cannot constant-fold away the zero
696 // trip guard until all unrolling is done.
697 Node *zer_opaq = new (C, 2) Opaque1Node(C, incr);
698 Node *zer_cmp = new (C, 3) CmpINode( zer_opaq, limit );
699 Node *zer_bol = new (C, 2) BoolNode( zer_cmp, b_test );
700 register_new_node( zer_opaq, new_main_exit );
701 register_new_node( zer_cmp , new_main_exit );
702 register_new_node( zer_bol , new_main_exit );
705 IfNode *zer_iff = new (C, 2) IfNode( new_main_exit, zer_bol, PROB_FAIR, COUNT_UNKNOWN );
706 _igvn.register_new_node_with_optimizer( zer_iff );
707 set_idom(zer_iff, new_main_exit, dd_main_exit);
708 set_loop(zer_iff, loop->_parent);
710 // Plug in the false-path, taken if we need to skip post-loop
711 _igvn.hash_delete( main_exit );
712 main_exit->set_req(0, zer_iff);
713 _igvn._worklist.push(main_exit);
714 set_idom(main_exit, zer_iff, dd_main_exit);
715 set_idom(main_exit->unique_out(), zer_iff, dd_main_exit);
716 // Make the true-path, must enter the post loop
717 Node *zer_taken = new (C, 1) IfTrueNode( zer_iff );
718 _igvn.register_new_node_with_optimizer( zer_taken );
719 set_idom(zer_taken, zer_iff, dd_main_exit);
720 set_loop(zer_taken, loop->_parent);
721 // Plug in the true path
722 _igvn.hash_delete( post_head );
723 post_head->set_req(LoopNode::EntryControl, zer_taken);
724 set_idom(post_head, zer_taken, dd_main_exit);
726 // Step A3: Make the fall-in values to the post-loop come from the
727 // fall-out values of the main-loop.
728 for (DUIterator_Fast imax, i = main_head->fast_outs(imax); i < imax; i++) {
729 Node* main_phi = main_head->fast_out(i);
730 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() >0 ) {
731 Node *post_phi = old_new[main_phi->_idx];
732 Node *fallmain = clone_up_backedge_goo(main_head->back_control(),
733 post_head->init_control(),
734 main_phi->in(LoopNode::LoopBackControl));
735 _igvn.hash_delete(post_phi);
736 post_phi->set_req( LoopNode::EntryControl, fallmain );
740 // Update local caches for next stanza
741 main_exit = new_main_exit;
744 //------------------------------
745 // Step B: Create Pre-Loop.
747 // Step B1: Clone the loop body. The clone becomes the pre-loop. The main
748 // loop pre-header illegally has 2 control users (old & new loops).
749 clone_loop( loop, old_new, dd_main_head );
750 CountedLoopNode* pre_head = old_new[main_head->_idx]->as_CountedLoop();
751 CountedLoopEndNode* pre_end = old_new[main_end ->_idx]->as_CountedLoopEnd();
752 pre_head->set_pre_loop(main_head);
753 Node *pre_incr = old_new[incr->_idx];
755 // Reduce the pre-loop trip count.
756 pre_end->_prob = PROB_FAIR;
758 // Find the pre-loop normal exit.
759 Node* pre_exit = pre_end->proj_out(false);
760 assert( pre_exit->Opcode() == Op_IfFalse, "" );
761 IfFalseNode *new_pre_exit = new (C, 1) IfFalseNode(pre_end);
762 _igvn.register_new_node_with_optimizer( new_pre_exit );
763 set_idom(new_pre_exit, pre_end, dd_main_head);
764 set_loop(new_pre_exit, loop->_parent);
766 // Step B2: Build a zero-trip guard for the main-loop. After leaving the
767 // pre-loop, the main-loop may not execute at all. Later in life this
768 // zero-trip guard will become the minimum-trip guard when we unroll
770 Node *min_opaq = new (C, 2) Opaque1Node(C, limit);
771 Node *min_cmp = new (C, 3) CmpINode( pre_incr, min_opaq );
772 Node *min_bol = new (C, 2) BoolNode( min_cmp, b_test );
773 register_new_node( min_opaq, new_pre_exit );
774 register_new_node( min_cmp , new_pre_exit );
775 register_new_node( min_bol , new_pre_exit );
777 // Build the IfNode (assume the main-loop is executed always).
778 IfNode *min_iff = new (C, 2) IfNode( new_pre_exit, min_bol, PROB_ALWAYS, COUNT_UNKNOWN );
779 _igvn.register_new_node_with_optimizer( min_iff );
780 set_idom(min_iff, new_pre_exit, dd_main_head);
781 set_loop(min_iff, loop->_parent);
783 // Plug in the false-path, taken if we need to skip main-loop
784 _igvn.hash_delete( pre_exit );
785 pre_exit->set_req(0, min_iff);
786 set_idom(pre_exit, min_iff, dd_main_head);
787 set_idom(pre_exit->unique_out(), min_iff, dd_main_head);
788 // Make the true-path, must enter the main loop
789 Node *min_taken = new (C, 1) IfTrueNode( min_iff );
790 _igvn.register_new_node_with_optimizer( min_taken );
791 set_idom(min_taken, min_iff, dd_main_head);
792 set_loop(min_taken, loop->_parent);
793 // Plug in the true path
794 _igvn.hash_delete( main_head );
795 main_head->set_req(LoopNode::EntryControl, min_taken);
796 set_idom(main_head, min_taken, dd_main_head);
798 // Step B3: Make the fall-in values to the main-loop come from the
799 // fall-out values of the pre-loop.
800 for (DUIterator_Fast i2max, i2 = main_head->fast_outs(i2max); i2 < i2max; i2++) {
801 Node* main_phi = main_head->fast_out(i2);
802 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() > 0 ) {
803 Node *pre_phi = old_new[main_phi->_idx];
804 Node *fallpre = clone_up_backedge_goo(pre_head->back_control(),
805 main_head->init_control(),
806 pre_phi->in(LoopNode::LoopBackControl));
807 _igvn.hash_delete(main_phi);
808 main_phi->set_req( LoopNode::EntryControl, fallpre );
812 // Step B4: Shorten the pre-loop to run only 1 iteration (for now).
813 // RCE and alignment may change this later.
814 Node *cmp_end = pre_end->cmp_node();
815 assert( cmp_end->in(2) == limit, "" );
816 Node *pre_limit = new (C, 3) AddINode( init, stride );
818 // Save the original loop limit in this Opaque1 node for
819 // use by range check elimination.
820 Node *pre_opaq = new (C, 3) Opaque1Node(C, pre_limit, limit);
822 register_new_node( pre_limit, pre_head->in(0) );
823 register_new_node( pre_opaq , pre_head->in(0) );
825 // Since no other users of pre-loop compare, I can hack limit directly
826 assert( cmp_end->outcnt() == 1, "no other users" );
827 _igvn.hash_delete(cmp_end);
828 cmp_end->set_req(2, peel_only ? pre_limit : pre_opaq);
830 // Special case for not-equal loop bounds:
831 // Change pre loop test, main loop test, and the
832 // main loop guard test to use lt or gt depending on stride
834 // positive stride use <
835 // negative stride use >
837 if (pre_end->in(CountedLoopEndNode::TestValue)->as_Bool()->_test._test == BoolTest::ne) {
839 BoolTest::mask new_test = (main_end->stride_con() > 0) ? BoolTest::lt : BoolTest::gt;
840 // Modify pre loop end condition
841 Node* pre_bol = pre_end->in(CountedLoopEndNode::TestValue)->as_Bool();
842 BoolNode* new_bol0 = new (C, 2) BoolNode(pre_bol->in(1), new_test);
843 register_new_node( new_bol0, pre_head->in(0) );
844 _igvn.hash_delete(pre_end);
845 pre_end->set_req(CountedLoopEndNode::TestValue, new_bol0);
846 // Modify main loop guard condition
847 assert(min_iff->in(CountedLoopEndNode::TestValue) == min_bol, "guard okay");
848 BoolNode* new_bol1 = new (C, 2) BoolNode(min_bol->in(1), new_test);
849 register_new_node( new_bol1, new_pre_exit );
850 _igvn.hash_delete(min_iff);
851 min_iff->set_req(CountedLoopEndNode::TestValue, new_bol1);
852 // Modify main loop end condition
853 BoolNode* main_bol = main_end->in(CountedLoopEndNode::TestValue)->as_Bool();
854 BoolNode* new_bol2 = new (C, 2) BoolNode(main_bol->in(1), new_test);
855 register_new_node( new_bol2, main_end->in(CountedLoopEndNode::TestControl) );
856 _igvn.hash_delete(main_end);
857 main_end->set_req(CountedLoopEndNode::TestValue, new_bol2);
861 main_head->set_main_loop();
862 if( peel_only ) main_head->set_main_no_pre_loop();
864 // It's difficult to be precise about the trip-counts
865 // for the pre/post loops. They are usually very short,
866 // so guess that 4 trips is a reasonable value.
867 post_head->set_profile_trip_cnt(4.0);
868 pre_head->set_profile_trip_cnt(4.0);
870 // Now force out all loop-invariant dominating tests. The optimizer
871 // finds some, but we _know_ they are all useless.
872 peeled_dom_test_elim(loop,old_new);
875 //------------------------------is_invariant-----------------------------
876 // Return true if n is invariant
877 bool IdealLoopTree::is_invariant(Node* n) const {
878 Node *n_c = _phase->get_ctrl(n);
879 if (n_c->is_top()) return false;
880 return !is_member(_phase->get_loop(n_c));
884 //------------------------------do_unroll--------------------------------------
885 // Unroll the loop body one step - make each trip do 2 iterations.
886 void PhaseIdealLoop::do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ) {
887 assert( LoopUnrollLimit, "" );
889 if( PrintOpto && VerifyLoopOptimizations ) {
890 tty->print("Unrolling ");
894 CountedLoopNode *loop_head = loop->_head->as_CountedLoop();
895 CountedLoopEndNode *loop_end = loop_head->loopexit();
896 assert( loop_end, "" );
898 // Remember loop node count before unrolling to detect
899 // if rounds of unroll,optimize are making progress
900 loop_head->set_node_count_before_unroll(loop->_body.size());
902 Node *ctrl = loop_head->in(LoopNode::EntryControl);
903 Node *limit = loop_head->limit();
904 Node *init = loop_head->init_trip();
905 Node *strid = loop_head->stride();
908 if( adjust_min_trip ) { // If not maximally unrolling, need adjustment
909 assert( loop_head->is_main_loop(), "" );
910 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );
911 Node *iff = ctrl->in(0);
912 assert( iff->Opcode() == Op_If, "" );
913 Node *bol = iff->in(1);
914 assert( bol->Opcode() == Op_Bool, "" );
915 Node *cmp = bol->in(1);
916 assert( cmp->Opcode() == Op_CmpI, "" );
918 // Occasionally it's possible for a pre-loop Opaque1 node to be
919 // optimized away and then another round of loop opts attempted.
920 // We can not optimize this particular loop in that case.
921 if( opaq->Opcode() != Op_Opaque1 )
922 return; // Cannot find pre-loop! Bail out!
925 C->set_major_progress();
927 // Adjust max trip count. The trip count is intentionally rounded
928 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll,
929 // the main, unrolled, part of the loop will never execute as it is protected
930 // by the min-trip test. See bug 4834191 for a case where we over-unrolled
931 // and later determined that part of the unrolled loop was dead.
932 loop_head->set_trip_count(loop_head->trip_count() / 2);
934 // Double the count of original iterations in the unrolled loop body.
935 loop_head->double_unrolled_count();
938 // Step 2: Cut back the trip counter for an unroll amount of 2.
939 // Loop will normally trip (limit - init)/stride_con. Since it's a
940 // CountedLoop this is exact (stride divides limit-init exactly).
941 // We are going to double the loop body, so we want to knock off any
942 // odd iteration: (trip_cnt & ~1). Then back compute a new limit.
943 Node *span = new (C, 3) SubINode( limit, init );
944 register_new_node( span, ctrl );
945 Node *trip = new (C, 3) DivINode( 0, span, strid );
946 register_new_node( trip, ctrl );
947 Node *mtwo = _igvn.intcon(-2);
948 set_ctrl(mtwo, C->root());
949 Node *rond = new (C, 3) AndINode( trip, mtwo );
950 register_new_node( rond, ctrl );
951 Node *spn2 = new (C, 3) MulINode( rond, strid );
952 register_new_node( spn2, ctrl );
953 Node *lim2 = new (C, 3) AddINode( spn2, init );
954 register_new_node( lim2, ctrl );
956 // Hammer in the new limit
957 Node *ctrl2 = loop_end->in(0);
958 Node *cmp2 = new (C, 3) CmpINode( loop_head->incr(), lim2 );
959 register_new_node( cmp2, ctrl2 );
960 Node *bol2 = new (C, 2) BoolNode( cmp2, loop_end->test_trip() );
961 register_new_node( bol2, ctrl2 );
962 _igvn.hash_delete(loop_end);
963 loop_end->set_req(CountedLoopEndNode::TestValue, bol2);
965 // Step 3: Find the min-trip test guaranteed before a 'main' loop.
966 // Make it a 1-trip test (means at least 2 trips).
967 if( adjust_min_trip ) {
968 // Guard test uses an 'opaque' node which is not shared. Hence I
969 // can edit it's inputs directly. Hammer in the new limit for the
970 // minimum-trip guard.
971 assert( opaq->outcnt() == 1, "" );
972 _igvn.hash_delete(opaq);
973 opaq->set_req(1, lim2);
977 // Step 4: Clone the loop body. Move it inside the loop. This loop body
978 // represents the odd iterations; since the loop trips an even number of
979 // times its backedge is never taken. Kill the backedge.
980 uint dd = dom_depth(loop_head);
981 clone_loop( loop, old_new, dd );
983 // Make backedges of the clone equal to backedges of the original.
984 // Make the fall-in from the original come from the fall-out of the clone.
985 for (DUIterator_Fast jmax, j = loop_head->fast_outs(jmax); j < jmax; j++) {
986 Node* phi = loop_head->fast_out(j);
987 if( phi->is_Phi() && phi->in(0) == loop_head && phi->outcnt() > 0 ) {
988 Node *newphi = old_new[phi->_idx];
989 _igvn.hash_delete( phi );
990 _igvn.hash_delete( newphi );
992 phi ->set_req(LoopNode:: EntryControl, newphi->in(LoopNode::LoopBackControl));
993 newphi->set_req(LoopNode::LoopBackControl, phi ->in(LoopNode::LoopBackControl));
994 phi ->set_req(LoopNode::LoopBackControl, C->top());
997 Node *clone_head = old_new[loop_head->_idx];
998 _igvn.hash_delete( clone_head );
999 loop_head ->set_req(LoopNode:: EntryControl, clone_head->in(LoopNode::LoopBackControl));
1000 clone_head->set_req(LoopNode::LoopBackControl, loop_head ->in(LoopNode::LoopBackControl));
1001 loop_head ->set_req(LoopNode::LoopBackControl, C->top());
1002 loop->_head = clone_head; // New loop header
1004 set_idom(loop_head, loop_head ->in(LoopNode::EntryControl), dd);
1005 set_idom(clone_head, clone_head->in(LoopNode::EntryControl), dd);
1007 // Kill the clone's backedge
1008 Node *newcle = old_new[loop_end->_idx];
1009 _igvn.hash_delete( newcle );
1010 Node *one = _igvn.intcon(1);
1011 set_ctrl(one, C->root());
1012 newcle->set_req(1, one);
1013 // Force clone into same loop body
1014 uint max = loop->_body.size();
1015 for( uint k = 0; k < max; k++ ) {
1016 Node *old = loop->_body.at(k);
1017 Node *nnn = old_new[old->_idx];
1018 loop->_body.push(nnn);
1020 set_loop(nnn, loop);
1023 loop->record_for_igvn();
1026 //------------------------------do_maximally_unroll----------------------------
1028 void PhaseIdealLoop::do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ) {
1029 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1030 assert( cl->trip_count() > 0, "");
1032 // If loop is tripping an odd number of times, peel odd iteration
1033 if( (cl->trip_count() & 1) == 1 ) {
1034 do_peeling( loop, old_new );
1037 // Now its tripping an even number of times remaining. Double loop body.
1038 // Do not adjust pre-guards; they are not needed and do not exist.
1039 if( cl->trip_count() > 0 ) {
1040 do_unroll( loop, old_new, false );
1044 //------------------------------dominates_backedge---------------------------------
1045 // Returns true if ctrl is executed on every complete iteration
1046 bool IdealLoopTree::dominates_backedge(Node* ctrl) {
1047 assert(ctrl->is_CFG(), "must be control");
1048 Node* backedge = _head->as_Loop()->in(LoopNode::LoopBackControl);
1049 return _phase->dom_lca_internal(ctrl, backedge) == ctrl;
1052 //------------------------------add_constraint---------------------------------
1053 // Constrain the main loop iterations so the condition:
1054 // scale_con * I + offset < limit
1055 // always holds true. That is, either increase the number of iterations in
1056 // the pre-loop or the post-loop until the condition holds true in the main
1057 // loop. Stride, scale, offset and limit are all loop invariant. Further,
1058 // stride and scale are constants (offset and limit often are).
1059 void PhaseIdealLoop::add_constraint( int stride_con, int scale_con, Node *offset, Node *limit, Node *pre_ctrl, Node **pre_limit, Node **main_limit ) {
1061 // Compute "I :: (limit-offset)/scale_con"
1062 Node *con = new (C, 3) SubINode( limit, offset );
1063 register_new_node( con, pre_ctrl );
1064 Node *scale = _igvn.intcon(scale_con);
1065 set_ctrl(scale, C->root());
1066 Node *X = new (C, 3) DivINode( 0, con, scale );
1067 register_new_node( X, pre_ctrl );
1069 // For positive stride, the pre-loop limit always uses a MAX function
1070 // and the main loop a MIN function. For negative stride these are
1073 // Also for positive stride*scale the affine function is increasing, so the
1074 // pre-loop must check for underflow and the post-loop for overflow.
1075 // Negative stride*scale reverses this; pre-loop checks for overflow and
1076 // post-loop for underflow.
1077 if( stride_con*scale_con > 0 ) {
1078 // Compute I < (limit-offset)/scale_con
1079 // Adjust main-loop last iteration to be MIN/MAX(main_loop,X)
1080 *main_limit = (stride_con > 0)
1081 ? (Node*)(new (C, 3) MinINode( *main_limit, X ))
1082 : (Node*)(new (C, 3) MaxINode( *main_limit, X ));
1083 register_new_node( *main_limit, pre_ctrl );
1086 // Compute (limit-offset)/scale_con + SGN(-scale_con) <= I
1087 // Add the negation of the main-loop constraint to the pre-loop.
1088 // See footnote [++] below for a derivation of the limit expression.
1089 Node *incr = _igvn.intcon(scale_con > 0 ? -1 : 1);
1090 set_ctrl(incr, C->root());
1091 Node *adj = new (C, 3) AddINode( X, incr );
1092 register_new_node( adj, pre_ctrl );
1093 *pre_limit = (scale_con > 0)
1094 ? (Node*)new (C, 3) MinINode( *pre_limit, adj )
1095 : (Node*)new (C, 3) MaxINode( *pre_limit, adj );
1096 register_new_node( *pre_limit, pre_ctrl );
1098 // [++] Here's the algebra that justifies the pre-loop limit expression:
1100 // NOT( scale_con * I + offset < limit )
1102 // scale_con * I + offset >= limit
1104 // SGN(scale_con) * I >= (limit-offset)/|scale_con|
1106 // (limit-offset)/|scale_con| <= I * SGN(scale_con)
1108 // (limit-offset)/|scale_con|-1 < I * SGN(scale_con)
1110 // ( if (scale_con > 0) /*common case*/
1111 // (limit-offset)/scale_con - 1 < I
1113 // (limit-offset)/scale_con + 1 > I
1115 // ( if (scale_con > 0) /*common case*/
1116 // (limit-offset)/scale_con + SGN(-scale_con) < I
1118 // (limit-offset)/scale_con + SGN(-scale_con) > I
1123 //------------------------------is_scaled_iv---------------------------------
1124 // Return true if exp is a constant times an induction var
1125 bool PhaseIdealLoop::is_scaled_iv(Node* exp, Node* iv, int* p_scale) {
1127 if (p_scale != NULL) {
1132 int opc = exp->Opcode();
1133 if (opc == Op_MulI) {
1134 if (exp->in(1) == iv && exp->in(2)->is_Con()) {
1135 if (p_scale != NULL) {
1136 *p_scale = exp->in(2)->get_int();
1140 if (exp->in(2) == iv && exp->in(1)->is_Con()) {
1141 if (p_scale != NULL) {
1142 *p_scale = exp->in(1)->get_int();
1146 } else if (opc == Op_LShiftI) {
1147 if (exp->in(1) == iv && exp->in(2)->is_Con()) {
1148 if (p_scale != NULL) {
1149 *p_scale = 1 << exp->in(2)->get_int();
1157 //-----------------------------is_scaled_iv_plus_offset------------------------------
1158 // Return true if exp is a simple induction variable expression: k1*iv + (invar + k2)
1159 bool PhaseIdealLoop::is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth) {
1160 if (is_scaled_iv(exp, iv, p_scale)) {
1161 if (p_offset != NULL) {
1162 Node *zero = _igvn.intcon(0);
1163 set_ctrl(zero, C->root());
1168 int opc = exp->Opcode();
1169 if (opc == Op_AddI) {
1170 if (is_scaled_iv(exp->in(1), iv, p_scale)) {
1171 if (p_offset != NULL) {
1172 *p_offset = exp->in(2);
1176 if (exp->in(2)->is_Con()) {
1177 Node* offset2 = NULL;
1179 is_scaled_iv_plus_offset(exp->in(1), iv, p_scale,
1180 p_offset != NULL ? &offset2 : NULL, depth+1)) {
1181 if (p_offset != NULL) {
1182 Node *ctrl_off2 = get_ctrl(offset2);
1183 Node* offset = new (C, 3) AddINode(offset2, exp->in(2));
1184 register_new_node(offset, ctrl_off2);
1190 } else if (opc == Op_SubI) {
1191 if (is_scaled_iv(exp->in(1), iv, p_scale)) {
1192 if (p_offset != NULL) {
1193 Node *zero = _igvn.intcon(0);
1194 set_ctrl(zero, C->root());
1195 Node *ctrl_off = get_ctrl(exp->in(2));
1196 Node* offset = new (C, 3) SubINode(zero, exp->in(2));
1197 register_new_node(offset, ctrl_off);
1202 if (is_scaled_iv(exp->in(2), iv, p_scale)) {
1203 if (p_offset != NULL) {
1205 *p_offset = exp->in(1);
1213 //------------------------------do_range_check---------------------------------
1214 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
1215 void PhaseIdealLoop::do_range_check( IdealLoopTree *loop, Node_List &old_new ) {
1217 if( PrintOpto && VerifyLoopOptimizations ) {
1218 tty->print("Range Check Elimination ");
1222 assert( RangeCheckElimination, "" );
1223 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1224 assert( cl->is_main_loop(), "" );
1226 // Find the trip counter; we are iteration splitting based on it
1227 Node *trip_counter = cl->phi();
1228 // Find the main loop limit; we will trim it's iterations
1229 // to not ever trip end tests
1230 Node *main_limit = cl->limit();
1231 // Find the pre-loop limit; we will expand it's iterations to
1232 // not ever trip low tests.
1233 Node *ctrl = cl->in(LoopNode::EntryControl);
1234 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );
1235 Node *iffm = ctrl->in(0);
1236 assert( iffm->Opcode() == Op_If, "" );
1237 Node *p_f = iffm->in(0);
1238 assert( p_f->Opcode() == Op_IfFalse, "" );
1239 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
1240 assert( pre_end->loopnode()->is_pre_loop(), "" );
1241 Node *pre_opaq1 = pre_end->limit();
1242 // Occasionally it's possible for a pre-loop Opaque1 node to be
1243 // optimized away and then another round of loop opts attempted.
1244 // We can not optimize this particular loop in that case.
1245 if( pre_opaq1->Opcode() != Op_Opaque1 )
1247 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
1248 Node *pre_limit = pre_opaq->in(1);
1250 // Where do we put new limit calculations
1251 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
1253 // Ensure the original loop limit is available from the
1254 // pre-loop Opaque1 node.
1255 Node *orig_limit = pre_opaq->original_loop_limit();
1256 if( orig_limit == NULL || _igvn.type(orig_limit) == Type::TOP )
1259 // Need to find the main-loop zero-trip guard
1260 Node *bolzm = iffm->in(1);
1261 assert( bolzm->Opcode() == Op_Bool, "" );
1262 Node *cmpzm = bolzm->in(1);
1263 assert( cmpzm->is_Cmp(), "" );
1264 Node *opqzm = cmpzm->in(2);
1265 if( opqzm->Opcode() != Op_Opaque1 )
1267 assert( opqzm->in(1) == main_limit, "do not understand situation" );
1269 // Must know if its a count-up or count-down loop
1271 // protect against stride not being a constant
1272 if ( !cl->stride_is_con() ) {
1275 int stride_con = cl->stride_con();
1276 Node *zero = _igvn.intcon(0);
1277 Node *one = _igvn.intcon(1);
1278 set_ctrl(zero, C->root());
1279 set_ctrl(one, C->root());
1281 // Range checks that do not dominate the loop backedge (ie.
1282 // conditionally executed) can lengthen the pre loop limit beyond
1283 // the original loop limit. To prevent this, the pre limit is
1284 // (for stride > 0) MINed with the original loop limit (MAXed
1285 // stride < 0) when some range_check (rc) is conditionally
1287 bool conditional_rc = false;
1289 // Check loop body for tests of trip-counter plus loop-invariant vs
1291 for( uint i = 0; i < loop->_body.size(); i++ ) {
1292 Node *iff = loop->_body[i];
1293 if( iff->Opcode() == Op_If ) { // Test?
1295 // Test is an IfNode, has 2 projections. If BOTH are in the loop
1296 // we need loop unswitching instead of iteration splitting.
1297 Node *exit = loop->is_loop_exit(iff);
1298 if( !exit ) continue;
1299 int flip = (exit->Opcode() == Op_IfTrue) ? 1 : 0;
1301 // Get boolean condition to test
1302 Node *i1 = iff->in(1);
1303 if( !i1->is_Bool() ) continue;
1304 BoolNode *bol = i1->as_Bool();
1305 BoolTest b_test = bol->_test;
1306 // Flip sense of test if exit condition is flipped
1308 b_test = b_test.negate();
1311 Node *cmp = bol->in(1);
1313 // Look for trip_counter + offset vs limit
1314 Node *rc_exp = cmp->in(1);
1315 Node *limit = cmp->in(2);
1316 jint scale_con= 1; // Assume trip counter not scaled
1318 Node *limit_c = get_ctrl(limit);
1319 if( loop->is_member(get_loop(limit_c) ) ) {
1320 // Compare might have operands swapped; commute them
1321 b_test = b_test.commute();
1322 rc_exp = cmp->in(2);
1324 limit_c = get_ctrl(limit);
1325 if( loop->is_member(get_loop(limit_c) ) )
1326 continue; // Both inputs are loop varying; cannot RCE
1328 // Here we know 'limit' is loop invariant
1330 // 'limit' maybe pinned below the zero trip test (probably from a
1331 // previous round of rce), in which case, it can't be used in the
1332 // zero trip test expression which must occur before the zero test's if.
1333 if( limit_c == ctrl ) {
1334 continue; // Don't rce this check but continue looking for other candidates.
1337 // Check for scaled induction variable plus an offset
1338 Node *offset = NULL;
1340 if (!is_scaled_iv_plus_offset(rc_exp, trip_counter, &scale_con, &offset)) {
1344 Node *offset_c = get_ctrl(offset);
1345 if( loop->is_member( get_loop(offset_c) ) )
1346 continue; // Offset is not really loop invariant
1347 // Here we know 'offset' is loop invariant.
1349 // As above for the 'limit', the 'offset' maybe pinned below the
1351 if( offset_c == ctrl ) {
1352 continue; // Don't rce this check but continue looking for other candidates.
1355 // At this point we have the expression as:
1356 // scale_con * trip_counter + offset :: limit
1357 // where scale_con, offset and limit are loop invariant. Trip_counter
1358 // monotonically increases by stride_con, a constant. Both (or either)
1359 // stride_con and scale_con can be negative which will flip about the
1360 // sense of the test.
1362 // Adjust pre and main loop limits to guard the correct iteration set
1363 if( cmp->Opcode() == Op_CmpU ) {// Unsigned compare is really 2 tests
1364 if( b_test._test == BoolTest::lt ) { // Range checks always use lt
1365 // The overflow limit: scale*I+offset < limit
1366 add_constraint( stride_con, scale_con, offset, limit, pre_ctrl, &pre_limit, &main_limit );
1367 // The underflow limit: 0 <= scale*I+offset.
1368 // Some math yields: -scale*I-(offset+1) < 0
1369 Node *plus_one = new (C, 3) AddINode( offset, one );
1370 register_new_node( plus_one, pre_ctrl );
1371 Node *neg_offset = new (C, 3) SubINode( zero, plus_one );
1372 register_new_node( neg_offset, pre_ctrl );
1373 add_constraint( stride_con, -scale_con, neg_offset, zero, pre_ctrl, &pre_limit, &main_limit );
1374 if (!conditional_rc) {
1375 conditional_rc = !loop->dominates_backedge(iff);
1380 tty->print_cr("missed RCE opportunity");
1382 continue; // In release mode, ignore it
1384 } else { // Otherwise work on normal compares
1385 switch( b_test._test ) {
1386 case BoolTest::ge: // Convert X >= Y to -X <= -Y
1387 scale_con = -scale_con;
1388 offset = new (C, 3) SubINode( zero, offset );
1389 register_new_node( offset, pre_ctrl );
1390 limit = new (C, 3) SubINode( zero, limit );
1391 register_new_node( limit, pre_ctrl );
1392 // Fall into LE case
1393 case BoolTest::le: // Convert X <= Y to X < Y+1
1394 limit = new (C, 3) AddINode( limit, one );
1395 register_new_node( limit, pre_ctrl );
1396 // Fall into LT case
1398 add_constraint( stride_con, scale_con, offset, limit, pre_ctrl, &pre_limit, &main_limit );
1399 if (!conditional_rc) {
1400 conditional_rc = !loop->dominates_backedge(iff);
1406 tty->print_cr("missed RCE opportunity");
1408 continue; // Unhandled case
1412 // Kill the eliminated test
1413 C->set_major_progress();
1414 Node *kill_con = _igvn.intcon( 1-flip );
1415 set_ctrl(kill_con, C->root());
1416 _igvn.hash_delete(iff);
1417 iff->set_req(1, kill_con);
1418 _igvn._worklist.push(iff);
1419 // Find surviving projection
1420 assert(iff->is_If(), "");
1421 ProjNode* dp = ((IfNode*)iff)->proj_out(1-flip);
1422 // Find loads off the surviving projection; remove their control edge
1423 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
1424 Node* cd = dp->fast_out(i); // Control-dependent node
1425 if( cd->is_Load() ) { // Loads can now float around in the loop
1426 _igvn.hash_delete(cd);
1427 // Allow the load to float around in the loop, or before it
1428 // but NOT before the pre-loop.
1429 cd->set_req(0, ctrl); // ctrl, not NULL
1430 _igvn._worklist.push(cd);
1440 // Update loop limits
1441 if (conditional_rc) {
1442 pre_limit = (stride_con > 0) ? (Node*)new (C,3) MinINode(pre_limit, orig_limit)
1443 : (Node*)new (C,3) MaxINode(pre_limit, orig_limit);
1444 register_new_node(pre_limit, pre_ctrl);
1446 _igvn.hash_delete(pre_opaq);
1447 pre_opaq->set_req(1, pre_limit);
1449 // Note:: we are making the main loop limit no longer precise;
1450 // need to round up based on stride.
1451 if( stride_con != 1 && stride_con != -1 ) { // Cutout for common case
1452 // "Standard" round-up logic: ([main_limit-init+(y-1)]/y)*y+init
1453 // Hopefully, compiler will optimize for powers of 2.
1454 Node *ctrl = get_ctrl(main_limit);
1455 Node *stride = cl->stride();
1456 Node *init = cl->init_trip();
1457 Node *span = new (C, 3) SubINode(main_limit,init);
1458 register_new_node(span,ctrl);
1459 Node *rndup = _igvn.intcon(stride_con + ((stride_con>0)?-1:1));
1460 Node *add = new (C, 3) AddINode(span,rndup);
1461 register_new_node(add,ctrl);
1462 Node *div = new (C, 3) DivINode(0,add,stride);
1463 register_new_node(div,ctrl);
1464 Node *mul = new (C, 3) MulINode(div,stride);
1465 register_new_node(mul,ctrl);
1466 Node *newlim = new (C, 3) AddINode(mul,init);
1467 register_new_node(newlim,ctrl);
1468 main_limit = newlim;
1471 Node *main_cle = cl->loopexit();
1472 Node *main_bol = main_cle->in(1);
1473 // Hacking loop bounds; need private copies of exit test
1474 if( main_bol->outcnt() > 1 ) {// BoolNode shared?
1475 _igvn.hash_delete(main_cle);
1476 main_bol = main_bol->clone();// Clone a private BoolNode
1477 register_new_node( main_bol, main_cle->in(0) );
1478 main_cle->set_req(1,main_bol);
1480 Node *main_cmp = main_bol->in(1);
1481 if( main_cmp->outcnt() > 1 ) { // CmpNode shared?
1482 _igvn.hash_delete(main_bol);
1483 main_cmp = main_cmp->clone();// Clone a private CmpNode
1484 register_new_node( main_cmp, main_cle->in(0) );
1485 main_bol->set_req(1,main_cmp);
1487 // Hack the now-private loop bounds
1488 _igvn.hash_delete(main_cmp);
1489 main_cmp->set_req(2, main_limit);
1490 _igvn._worklist.push(main_cmp);
1491 // The OpaqueNode is unshared by design
1492 _igvn.hash_delete(opqzm);
1493 assert( opqzm->outcnt() == 1, "cannot hack shared node" );
1494 opqzm->set_req(1,main_limit);
1495 _igvn._worklist.push(opqzm);
1498 //------------------------------DCE_loop_body----------------------------------
1499 // Remove simplistic dead code from loop body
1500 void IdealLoopTree::DCE_loop_body() {
1501 for( uint i = 0; i < _body.size(); i++ )
1502 if( _body.at(i)->outcnt() == 0 )
1503 _body.map( i--, _body.pop() );
1507 //------------------------------adjust_loop_exit_prob--------------------------
1508 // Look for loop-exit tests with the 50/50 (or worse) guesses from the parsing stage.
1509 // Replace with a 1-in-10 exit guess.
1510 void IdealLoopTree::adjust_loop_exit_prob( PhaseIdealLoop *phase ) {
1511 Node *test = tail();
1512 while( test != _head ) {
1513 uint top = test->Opcode();
1514 if( top == Op_IfTrue || top == Op_IfFalse ) {
1515 int test_con = ((ProjNode*)test)->_con;
1516 assert(top == (uint)(test_con? Op_IfTrue: Op_IfFalse), "sanity");
1517 IfNode *iff = test->in(0)->as_If();
1518 if( iff->outcnt() == 2 ) { // Ignore dead tests
1519 Node *bol = iff->in(1);
1520 if( bol && bol->req() > 1 && bol->in(1) &&
1521 ((bol->in(1)->Opcode() == Op_StorePConditional ) ||
1522 (bol->in(1)->Opcode() == Op_StoreIConditional ) ||
1523 (bol->in(1)->Opcode() == Op_StoreLConditional ) ||
1524 (bol->in(1)->Opcode() == Op_CompareAndSwapI ) ||
1525 (bol->in(1)->Opcode() == Op_CompareAndSwapL ) ||
1526 (bol->in(1)->Opcode() == Op_CompareAndSwapP ) ||
1527 (bol->in(1)->Opcode() == Op_CompareAndSwapN )))
1528 return; // Allocation loops RARELY take backedge
1529 // Find the OTHER exit path from the IF
1530 Node* ex = iff->proj_out(1-test_con);
1531 float p = iff->_prob;
1532 if( !phase->is_member( this, ex ) && iff->_fcnt == COUNT_UNKNOWN ) {
1533 if( top == Op_IfTrue ) {
1534 if( p < (PROB_FAIR + PROB_UNLIKELY_MAG(3))) {
1535 iff->_prob = PROB_STATIC_FREQUENT;
1538 if( p > (PROB_FAIR - PROB_UNLIKELY_MAG(3))) {
1539 iff->_prob = PROB_STATIC_INFREQUENT;
1545 test = phase->idom(test);
1550 //------------------------------policy_do_remove_empty_loop--------------------
1551 // Micro-benchmark spamming. Policy is to always remove empty loops.
1552 // The 'DO' part is to replace the trip counter with the value it will
1553 // have on the last iteration. This will break the loop.
1554 bool IdealLoopTree::policy_do_remove_empty_loop( PhaseIdealLoop *phase ) {
1555 // Minimum size must be empty loop
1556 if( _body.size() > 7/*number of nodes in an empty loop*/ ) return false;
1558 if( !_head->is_CountedLoop() ) return false; // Dead loop
1559 CountedLoopNode *cl = _head->as_CountedLoop();
1560 if( !cl->loopexit() ) return false; // Malformed loop
1561 if( !phase->is_member(this,phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue)) ) )
1562 return false; // Infinite loop
1565 tty->print_cr("Removing empty loop");
1568 // Ensure only one phi which is the iv.
1570 for (DUIterator_Fast imax, i = cl->fast_outs(imax); i < imax; i++) {
1571 Node* n = cl->fast_out(i);
1572 if (n->Opcode() == Op_Phi) {
1573 assert(iv == NULL, "Too many phis" );
1577 assert(iv == cl->phi(), "Wrong phi" );
1579 // Replace the phi at loop head with the final value of the last
1580 // iteration. Then the CountedLoopEnd will collapse (backedge never
1581 // taken) and all loop-invariant uses of the exit values will be correct.
1582 Node *phi = cl->phi();
1583 Node *final = new (phase->C, 3) SubINode( cl->limit(), cl->stride() );
1584 phase->register_new_node(final,cl->in(LoopNode::EntryControl));
1585 phase->_igvn.hash_delete(phi);
1586 phase->_igvn.subsume_node(phi,final);
1587 phase->C->set_major_progress();
1592 //=============================================================================
1593 //------------------------------iteration_split_impl---------------------------
1594 bool IdealLoopTree::iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ) {
1595 // Check and remove empty loops (spam micro-benchmarks)
1596 if( policy_do_remove_empty_loop(phase) )
1597 return true; // Here we removed an empty loop
1599 bool should_peel = policy_peeling(phase); // Should we peel?
1601 bool should_unswitch = policy_unswitching(phase);
1603 // Non-counted loops may be peeled; exactly 1 iteration is peeled.
1604 // This removes loop-invariant tests (usually null checks).
1605 if( !_head->is_CountedLoop() ) { // Non-counted loop
1606 if (PartialPeelLoop && phase->partial_peel(this, old_new)) {
1607 // Partial peel succeeded so terminate this round of loop opts
1610 if( should_peel ) { // Should we peel?
1612 if (PrintOpto) tty->print_cr("should_peel");
1614 phase->do_peeling(this,old_new);
1615 } else if( should_unswitch ) {
1616 phase->do_unswitching(this, old_new);
1620 CountedLoopNode *cl = _head->as_CountedLoop();
1622 if( !cl->loopexit() ) return true; // Ignore various kinds of broken loops
1624 // Do nothing special to pre- and post- loops
1625 if( cl->is_pre_loop() || cl->is_post_loop() ) return true;
1627 // Compute loop trip count from profile data
1628 compute_profile_trip_cnt(phase);
1630 // Before attempting fancy unrolling, RCE or alignment, see if we want
1631 // to completely unroll this loop or do loop unswitching.
1632 if( cl->is_normal_loop() ) {
1633 if (should_unswitch) {
1634 phase->do_unswitching(this, old_new);
1637 bool should_maximally_unroll = policy_maximally_unroll(phase);
1638 if( should_maximally_unroll ) {
1639 // Here we did some unrolling and peeling. Eventually we will
1640 // completely unroll this loop and it will no longer be a loop.
1641 phase->do_maximally_unroll(this,old_new);
1647 // Counted loops may be peeled, may need some iterations run up
1648 // front for RCE, and may want to align loop refs to a cache
1649 // line. Thus we clone a full loop up front whose trip count is
1650 // at least 1 (if peeling), but may be several more.
1652 // The main loop will start cache-line aligned with at least 1
1653 // iteration of the unrolled body (zero-trip test required) and
1654 // will have some range checks removed.
1656 // A post-loop will finish any odd iterations (leftover after
1657 // unrolling), plus any needed for RCE purposes.
1659 bool should_unroll = policy_unroll(phase);
1661 bool should_rce = policy_range_check(phase);
1663 bool should_align = policy_align(phase);
1665 // If not RCE'ing (iteration splitting) or Aligning, then we do not
1666 // need a pre-loop. We may still need to peel an initial iteration but
1667 // we will not be needing an unknown number of pre-iterations.
1669 // Basically, if may_rce_align reports FALSE first time through,
1670 // we will not be able to later do RCE or Aligning on this loop.
1671 bool may_rce_align = !policy_peel_only(phase) || should_rce || should_align;
1673 // If we have any of these conditions (RCE, alignment, unrolling) met, then
1674 // we switch to the pre-/main-/post-loop model. This model also covers
1676 if( should_rce || should_align || should_unroll ) {
1677 if( cl->is_normal_loop() ) // Convert to 'pre/main/post' loops
1678 phase->insert_pre_post_loops(this,old_new, !may_rce_align);
1680 // Adjust the pre- and main-loop limits to let the pre and post loops run
1681 // with full checks, but the main-loop with no checks. Remove said
1682 // checks from the main body.
1684 phase->do_range_check(this,old_new);
1686 // Double loop body for unrolling. Adjust the minimum-trip test (will do
1687 // twice as many iterations as before) and the main body limit (only do
1688 // an even number of trips). If we are peeling, we might enable some RCE
1689 // and we'd rather unroll the post-RCE'd loop SO... do not unroll if
1691 if( should_unroll && !should_peel )
1692 phase->do_unroll(this,old_new, true);
1694 // Adjust the pre-loop limits to align the main body
1699 } else { // Else we have an unchanged counted loop
1700 if( should_peel ) // Might want to peel but do nothing else
1701 phase->do_peeling(this,old_new);
1707 //=============================================================================
1708 //------------------------------iteration_split--------------------------------
1709 bool IdealLoopTree::iteration_split( PhaseIdealLoop *phase, Node_List &old_new ) {
1710 // Recursively iteration split nested loops
1711 if( _child && !_child->iteration_split( phase, old_new ))
1714 // Clean out prior deadwood
1718 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
1719 // Replace with a 1-in-10 exit guess.
1720 if( _parent /*not the root loop*/ &&
1722 // Also ignore the occasional dead backedge
1723 !tail()->is_top() ) {
1724 adjust_loop_exit_prob(phase);
1728 // Gate unrolling, RCE and peeling efforts.
1729 if( !_child && // If not an inner loop, do not split
1731 _allow_optimizations &&
1732 !tail()->is_top() ) { // Also ignore the occasional dead backedge
1734 if (!iteration_split_impl( phase, old_new )) {
1737 } else if (policy_unswitching(phase)) {
1738 phase->do_unswitching(this, old_new);
1742 // Minor offset re-organization to remove loop-fallout uses of
1744 if( _head->is_CountedLoop() ) phase->reorg_offsets( this );
1745 if( _next && !_next->iteration_split( phase, old_new ))