1 /*
2 * Copyright (c) 1999, 2022, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
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.
8 *
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).
14 *
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.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "gc/shared/barrierSet.hpp"
27 #include "gc/shared/c2/barrierSetC2.hpp"
28 #include "memory/allocation.inline.hpp"
29 #include "memory/resourceArea.hpp"
30 #include "opto/addnode.hpp"
31 #include "opto/callnode.hpp"
32 #include "opto/castnode.hpp"
33 #include "opto/connode.hpp"
34 #include "opto/castnode.hpp"
35 #include "opto/divnode.hpp"
36 #include "opto/loopnode.hpp"
37 #include "opto/matcher.hpp"
38 #include "opto/mulnode.hpp"
39 #include "opto/movenode.hpp"
40 #include "opto/opaquenode.hpp"
41 #include "opto/rootnode.hpp"
42 #include "opto/subnode.hpp"
43 #include "opto/subtypenode.hpp"
44 #include "utilities/macros.hpp"
45
46 //=============================================================================
47 //------------------------------split_thru_phi---------------------------------
48 // Split Node 'n' through merge point if there is enough win.
49 Node* PhaseIdealLoop::split_thru_phi(Node* n, Node* region, int policy) {
50 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) {
51 // ConvI2L may have type information on it which is unsafe to push up
52 // so disable this for now
53 return NULL;
54 }
55
56 // Splitting range check CastIIs through a loop induction Phi can
57 // cause new Phis to be created that are left unrelated to the loop
58 // induction Phi and prevent optimizations (vectorization)
59 if (n->Opcode() == Op_CastII && region->is_CountedLoop() &&
60 n->in(1) == region->as_CountedLoop()->phi()) {
61 return NULL;
62 }
63
64 // Bail out if 'n' is a Div or Mod node whose zero check was removed earlier (i.e. control is NULL) and its divisor is an induction variable
65 // phi p of a trip-counted (integer) loop whose inputs could be zero (include zero in their type range). p could have a more precise type
66 // range that does not necessarily include all values of its inputs. Since each of these inputs will be a divisor of the newly cloned nodes
67 // of 'n', we need to bail out of one of these divisors could be zero (zero in its type range).
68 if ((n->Opcode() == Op_DivI || n->Opcode() == Op_ModI) && n->in(0) == NULL
69 && region->is_CountedLoop() && n->in(2) == region->as_CountedLoop()->phi()) {
70 Node* phi = region->as_CountedLoop()->phi();
71 for (uint i = 1; i < phi->req(); i++) {
72 if (_igvn.type(phi->in(i))->filter_speculative(TypeInt::ZERO) != Type::TOP) {
73 // Zero could be a possible value but we already removed the zero check. Bail out to avoid a possible division by zero at a later point.
74 return NULL;
75 }
76 }
77 }
78
79 int wins = 0;
80 assert(!n->is_CFG(), "");
81 assert(region->is_Region(), "");
82
83 const Type* type = n->bottom_type();
84 const TypeOopPtr* t_oop = _igvn.type(n)->isa_oopptr();
85 Node* phi;
86 if (t_oop != NULL && t_oop->is_known_instance_field()) {
87 int iid = t_oop->instance_id();
88 int index = C->get_alias_index(t_oop);
89 int offset = t_oop->offset();
90 phi = new PhiNode(region, type, NULL, iid, index, offset);
91 } else {
92 phi = PhiNode::make_blank(region, n);
93 }
94 uint old_unique = C->unique();
95 for (uint i = 1; i < region->req(); i++) {
96 Node* x;
97 Node* the_clone = NULL;
98 if (region->in(i) == C->top()) {
99 x = C->top(); // Dead path? Use a dead data op
100 } else {
101 x = n->clone(); // Else clone up the data op
102 the_clone = x; // Remember for possible deletion.
103 // Alter data node to use pre-phi inputs
104 if (n->in(0) == region)
105 x->set_req( 0, region->in(i) );
106 for (uint j = 1; j < n->req(); j++) {
107 Node* in = n->in(j);
108 if (in->is_Phi() && in->in(0) == region)
109 x->set_req(j, in->in(i)); // Use pre-Phi input for the clone
110 }
111 }
112 // Check for a 'win' on some paths
113 const Type* t = x->Value(&_igvn);
114
115 bool singleton = t->singleton();
116
117 // A TOP singleton indicates that there are no possible values incoming
118 // along a particular edge. In most cases, this is OK, and the Phi will
119 // be eliminated later in an Ideal call. However, we can't allow this to
120 // happen if the singleton occurs on loop entry, as the elimination of
121 // the PhiNode may cause the resulting node to migrate back to a previous
122 // loop iteration.
123 if (singleton && t == Type::TOP) {
124 // Is_Loop() == false does not confirm the absence of a loop (e.g., an
125 // irreducible loop may not be indicated by an affirmative is_Loop());
126 // therefore, the only top we can split thru a phi is on a backedge of
127 // a loop.
128 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
129 }
130
131 if (singleton) {
132 wins++;
133 x = ((PhaseGVN&)_igvn).makecon(t);
134 } else {
135 // We now call Identity to try to simplify the cloned node.
136 // Note that some Identity methods call phase->type(this).
137 // Make sure that the type array is big enough for
138 // our new node, even though we may throw the node away.
139 // (Note: This tweaking with igvn only works because x is a new node.)
140 _igvn.set_type(x, t);
141 // If x is a TypeNode, capture any more-precise type permanently into Node
142 // otherwise it will be not updated during igvn->transform since
143 // igvn->type(x) is set to x->Value() already.
144 x->raise_bottom_type(t);
145 Node* y = x->Identity(&_igvn);
146 if (y != x) {
147 wins++;
148 x = y;
149 } else {
150 y = _igvn.hash_find(x);
151 if (y) {
152 wins++;
153 x = y;
154 } else {
155 // Else x is a new node we are keeping
156 // We do not need register_new_node_with_optimizer
157 // because set_type has already been called.
158 _igvn._worklist.push(x);
159 }
160 }
161 }
162 if (x != the_clone && the_clone != NULL)
163 _igvn.remove_dead_node(the_clone);
164 phi->set_req( i, x );
165 }
166 // Too few wins?
167 if (wins <= policy) {
168 _igvn.remove_dead_node(phi);
169 return NULL;
170 }
171
172 // Record Phi
173 register_new_node( phi, region );
174
175 for (uint i2 = 1; i2 < phi->req(); i2++) {
176 Node *x = phi->in(i2);
177 // If we commoned up the cloned 'x' with another existing Node,
178 // the existing Node picks up a new use. We need to make the
179 // existing Node occur higher up so it dominates its uses.
180 Node *old_ctrl;
181 IdealLoopTree *old_loop;
182
183 if (x->is_Con()) {
184 // Constant's control is always root.
185 set_ctrl(x, C->root());
186 continue;
187 }
188 // The occasional new node
189 if (x->_idx >= old_unique) { // Found a new, unplaced node?
190 old_ctrl = NULL;
191 old_loop = NULL; // Not in any prior loop
192 } else {
193 old_ctrl = get_ctrl(x);
194 old_loop = get_loop(old_ctrl); // Get prior loop
195 }
196 // New late point must dominate new use
197 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2));
198 if (new_ctrl == old_ctrl) // Nothing is changed
199 continue;
200
201 IdealLoopTree *new_loop = get_loop(new_ctrl);
202
203 // Don't move x into a loop if its uses are
204 // outside of loop. Otherwise x will be cloned
205 // for each use outside of this loop.
206 IdealLoopTree *use_loop = get_loop(region);
207 if (!new_loop->is_member(use_loop) &&
208 (old_loop == NULL || !new_loop->is_member(old_loop))) {
209 // Take early control, later control will be recalculated
210 // during next iteration of loop optimizations.
211 new_ctrl = get_early_ctrl(x);
212 new_loop = get_loop(new_ctrl);
213 }
214 // Set new location
215 set_ctrl(x, new_ctrl);
216 // If changing loop bodies, see if we need to collect into new body
217 if (old_loop != new_loop) {
218 if (old_loop && !old_loop->_child)
219 old_loop->_body.yank(x);
220 if (!new_loop->_child)
221 new_loop->_body.push(x); // Collect body info
222 }
223 }
224
225 return phi;
226 }
227
228 //------------------------------dominated_by------------------------------------
229 // Replace the dominated test with an obvious true or false. Place it on the
230 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the
231 // live path up to the dominating control.
232 void PhaseIdealLoop::dominated_by(IfProjNode* prevdom, IfNode* iff, bool flip, bool exclude_loop_predicate) {
233 if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); }
234
235 // prevdom is the dominating projection of the dominating test.
236 assert(iff->Opcode() == Op_If ||
237 iff->Opcode() == Op_CountedLoopEnd ||
238 iff->Opcode() == Op_LongCountedLoopEnd ||
239 iff->Opcode() == Op_RangeCheck,
240 "Check this code when new subtype is added");
241
242 int pop = prevdom->Opcode();
243 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" );
244 if (flip) {
245 if (pop == Op_IfTrue)
246 pop = Op_IfFalse;
247 else
248 pop = Op_IfTrue;
249 }
250 // 'con' is set to true or false to kill the dominated test.
251 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO);
252 set_ctrl(con, C->root()); // Constant gets a new use
253 // Hack the dominated test
254 _igvn.replace_input_of(iff, 1, con);
255
256 // If I dont have a reachable TRUE and FALSE path following the IfNode then
257 // I can assume this path reaches an infinite loop. In this case it's not
258 // important to optimize the data Nodes - either the whole compilation will
259 // be tossed or this path (and all data Nodes) will go dead.
260 if (iff->outcnt() != 2) return;
261
262 // Make control-dependent data Nodes on the live path (path that will remain
263 // once the dominated IF is removed) become control-dependent on the
264 // dominating projection.
265 Node* dp = iff->proj_out_or_null(pop == Op_IfTrue);
266
267 // Loop predicates may have depending checks which should not
268 // be skipped. For example, range check predicate has two checks
269 // for lower and upper bounds.
270 if (dp == NULL)
271 return;
272
273 ProjNode* dp_proj = dp->as_Proj();
274 ProjNode* unc_proj = iff->proj_out(1 - dp_proj->_con)->as_Proj();
275 if (exclude_loop_predicate &&
276 (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) != NULL ||
277 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_profile_predicate) != NULL ||
278 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check) != NULL)) {
279 // If this is a range check (IfNode::is_range_check), do not
280 // reorder because Compile::allow_range_check_smearing might have
281 // changed the check.
282 return; // Let IGVN transformation change control dependence.
283 }
284
285 IdealLoopTree* old_loop = get_loop(dp);
286
287 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
288 Node* cd = dp->fast_out(i); // Control-dependent node
289 // Do not rewire Div and Mod nodes which could have a zero divisor to avoid skipping their zero check.
290 if (cd->depends_only_on_test() && _igvn.no_dependent_zero_check(cd)) {
291 assert(cd->in(0) == dp, "");
292 _igvn.replace_input_of(cd, 0, prevdom);
293 set_early_ctrl(cd, false);
294 IdealLoopTree* new_loop = get_loop(get_ctrl(cd));
295 if (old_loop != new_loop) {
296 if (!old_loop->_child) {
297 old_loop->_body.yank(cd);
298 }
299 if (!new_loop->_child) {
300 new_loop->_body.push(cd);
301 }
302 }
303 --i;
304 --imax;
305 }
306 }
307 }
308
309 //------------------------------has_local_phi_input----------------------------
310 // Return TRUE if 'n' has Phi inputs from its local block and no other
311 // block-local inputs (all non-local-phi inputs come from earlier blocks)
312 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) {
313 Node *n_ctrl = get_ctrl(n);
314 // See if some inputs come from a Phi in this block, or from before
315 // this block.
316 uint i;
317 for( i = 1; i < n->req(); i++ ) {
318 Node *phi = n->in(i);
319 if( phi->is_Phi() && phi->in(0) == n_ctrl )
320 break;
321 }
322 if( i >= n->req() )
323 return NULL; // No Phi inputs; nowhere to clone thru
324
325 // Check for inputs created between 'n' and the Phi input. These
326 // must split as well; they have already been given the chance
327 // (courtesy of a post-order visit) and since they did not we must
328 // recover the 'cost' of splitting them by being very profitable
329 // when splitting 'n'. Since this is unlikely we simply give up.
330 for( i = 1; i < n->req(); i++ ) {
331 Node *m = n->in(i);
332 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) {
333 // We allow the special case of AddP's with no local inputs.
334 // This allows us to split-up address expressions.
335 if (m->is_AddP() &&
336 get_ctrl(m->in(AddPNode::Base)) != n_ctrl &&
337 get_ctrl(m->in(AddPNode::Address)) != n_ctrl &&
338 get_ctrl(m->in(AddPNode::Offset)) != n_ctrl) {
339 // Move the AddP up to the dominating point. That's fine because control of m's inputs
340 // must dominate get_ctrl(m) == n_ctrl and we just checked that the input controls are != n_ctrl.
341 Node* c = find_non_split_ctrl(idom(n_ctrl));
342 if (c->is_OuterStripMinedLoop()) {
343 c->as_Loop()->verify_strip_mined(1);
344 c = c->in(LoopNode::EntryControl);
345 }
346 set_ctrl_and_loop(m, c);
347 continue;
348 }
349 return NULL;
350 }
351 assert(n->is_Phi() || m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control");
352 }
353
354 return n_ctrl;
355 }
356
357 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2).
358 Node* PhaseIdealLoop::remix_address_expressions_add_left_shift(Node* n, IdealLoopTree* n_loop, Node* n_ctrl, BasicType bt) {
359 assert(bt == T_INT || bt == T_LONG, "only for integers");
360 int n_op = n->Opcode();
361
362 if (n_op == Op_LShift(bt)) {
363 // Scale is loop invariant
364 Node* scale = n->in(2);
365 Node* scale_ctrl = get_ctrl(scale);
366 IdealLoopTree* scale_loop = get_loop(scale_ctrl);
367 if (n_loop == scale_loop || !scale_loop->is_member(n_loop)) {
368 return NULL;
369 }
370 const TypeInt* scale_t = scale->bottom_type()->isa_int();
371 if (scale_t != NULL && scale_t->is_con() && scale_t->get_con() >= 16) {
372 return NULL; // Dont bother with byte/short masking
373 }
374 // Add must vary with loop (else shift would be loop-invariant)
375 Node* add = n->in(1);
376 Node* add_ctrl = get_ctrl(add);
377 IdealLoopTree* add_loop = get_loop(add_ctrl);
378 if (n_loop != add_loop) {
379 return NULL; // happens w/ evil ZKM loops
380 }
381
382 // Convert I-V into I+ (0-V); same for V-I
383 if (add->Opcode() == Op_Sub(bt) &&
384 _igvn.type(add->in(1)) != TypeInteger::zero(bt)) {
385 assert(add->Opcode() == Op_SubI || add->Opcode() == Op_SubL, "");
386 Node* zero = _igvn.integercon(0, bt);
387 set_ctrl(zero, C->root());
388 Node* neg = SubNode::make(zero, add->in(2), bt);
389 register_new_node(neg, get_ctrl(add->in(2)));
390 add = AddNode::make(add->in(1), neg, bt);
391 register_new_node(add, add_ctrl);
392 }
393 if (add->Opcode() != Op_Add(bt)) return NULL;
394 assert(add->Opcode() == Op_AddI || add->Opcode() == Op_AddL, "");
395 // See if one add input is loop invariant
396 Node* add_var = add->in(1);
397 Node* add_var_ctrl = get_ctrl(add_var);
398 IdealLoopTree* add_var_loop = get_loop(add_var_ctrl);
399 Node* add_invar = add->in(2);
400 Node* add_invar_ctrl = get_ctrl(add_invar);
401 IdealLoopTree* add_invar_loop = get_loop(add_invar_ctrl);
402 if (add_invar_loop == n_loop) {
403 // Swap to find the invariant part
404 add_invar = add_var;
405 add_invar_ctrl = add_var_ctrl;
406 add_invar_loop = add_var_loop;
407 add_var = add->in(2);
408 } else if (add_var_loop != n_loop) { // Else neither input is loop invariant
409 return NULL;
410 }
411 if (n_loop == add_invar_loop || !add_invar_loop->is_member(n_loop)) {
412 return NULL; // No invariant part of the add?
413 }
414
415 // Yes! Reshape address expression!
416 Node* inv_scale = LShiftNode::make(add_invar, scale, bt);
417 Node* inv_scale_ctrl =
418 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ?
419 add_invar_ctrl : scale_ctrl;
420 register_new_node(inv_scale, inv_scale_ctrl);
421 Node* var_scale = LShiftNode::make(add_var, scale, bt);
422 register_new_node(var_scale, n_ctrl);
423 Node* var_add = AddNode::make(var_scale, inv_scale, bt);
424 register_new_node(var_add, n_ctrl);
425 _igvn.replace_node(n, var_add);
426 return var_add;
427 }
428 return NULL;
429 }
430
431 //------------------------------remix_address_expressions----------------------
432 // Rework addressing expressions to get the most loop-invariant stuff
433 // moved out. We'd like to do all associative operators, but it's especially
434 // important (common) to do address expressions.
435 Node* PhaseIdealLoop::remix_address_expressions(Node* n) {
436 if (!has_ctrl(n)) return NULL;
437 Node* n_ctrl = get_ctrl(n);
438 IdealLoopTree* n_loop = get_loop(n_ctrl);
439
440 // See if 'n' mixes loop-varying and loop-invariant inputs and
441 // itself is loop-varying.
442
443 // Only interested in binary ops (and AddP)
444 if (n->req() < 3 || n->req() > 4) return NULL;
445
446 Node* n1_ctrl = get_ctrl(n->in( 1));
447 Node* n2_ctrl = get_ctrl(n->in( 2));
448 Node* n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3));
449 IdealLoopTree* n1_loop = get_loop(n1_ctrl);
450 IdealLoopTree* n2_loop = get_loop(n2_ctrl);
451 IdealLoopTree* n3_loop = get_loop(n3_ctrl);
452
453 // Does one of my inputs spin in a tighter loop than self?
454 if ((n_loop->is_member(n1_loop) && n_loop != n1_loop) ||
455 (n_loop->is_member(n2_loop) && n_loop != n2_loop) ||
456 (n_loop->is_member(n3_loop) && n_loop != n3_loop)) {
457 return NULL; // Leave well enough alone
458 }
459
460 // Is at least one of my inputs loop-invariant?
461 if (n1_loop == n_loop &&
462 n2_loop == n_loop &&
463 n3_loop == n_loop) {
464 return NULL; // No loop-invariant inputs
465 }
466
467 Node* res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_INT);
468 if (res != NULL) {
469 return res;
470 }
471 res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_LONG);
472 if (res != NULL) {
473 return res;
474 }
475
476 int n_op = n->Opcode();
477 // Replace (I+V) with (V+I)
478 if (n_op == Op_AddI ||
479 n_op == Op_AddL ||
480 n_op == Op_AddF ||
481 n_op == Op_AddD ||
482 n_op == Op_MulI ||
483 n_op == Op_MulL ||
484 n_op == Op_MulF ||
485 n_op == Op_MulD) {
486 if (n2_loop == n_loop) {
487 assert(n1_loop != n_loop, "");
488 n->swap_edges(1, 2);
489 }
490 }
491
492 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V),
493 // but not if I2 is a constant.
494 if (n_op == Op_AddP) {
495 if (n2_loop == n_loop && n3_loop != n_loop) {
496 if (n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con()) {
497 Node* n22_ctrl = get_ctrl(n->in(2)->in(2));
498 Node* n23_ctrl = get_ctrl(n->in(2)->in(3));
499 IdealLoopTree* n22loop = get_loop(n22_ctrl);
500 IdealLoopTree* n23_loop = get_loop(n23_ctrl);
501 if (n22loop != n_loop && n22loop->is_member(n_loop) &&
502 n23_loop == n_loop) {
503 Node* add1 = new AddPNode(n->in(1), n->in(2)->in(2), n->in(3));
504 // Stuff new AddP in the loop preheader
505 register_new_node(add1, n_loop->_head->in(LoopNode::EntryControl));
506 Node* add2 = new AddPNode(n->in(1), add1, n->in(2)->in(3));
507 register_new_node(add2, n_ctrl);
508 _igvn.replace_node(n, add2);
509 return add2;
510 }
511 }
512 }
513
514 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V)
515 if (n2_loop != n_loop && n3_loop == n_loop) {
516 if (n->in(3)->Opcode() == Op_AddX) {
517 Node* V = n->in(3)->in(1);
518 Node* I = n->in(3)->in(2);
519 if (is_member(n_loop,get_ctrl(V))) {
520 } else {
521 Node *tmp = V; V = I; I = tmp;
522 }
523 if (!is_member(n_loop,get_ctrl(I))) {
524 Node* add1 = new AddPNode(n->in(1), n->in(2), I);
525 // Stuff new AddP in the loop preheader
526 register_new_node(add1, n_loop->_head->in(LoopNode::EntryControl));
527 Node* add2 = new AddPNode(n->in(1), add1, V);
528 register_new_node(add2, n_ctrl);
529 _igvn.replace_node(n, add2);
530 return add2;
531 }
532 }
533 }
534 }
535
536 return NULL;
537 }
538
539 // Optimize ((in1[2*i] * in2[2*i]) + (in1[2*i+1] * in2[2*i+1]))
540 Node *PhaseIdealLoop::convert_add_to_muladd(Node* n) {
541 assert(n->Opcode() == Op_AddI, "sanity");
542 Node * nn = NULL;
543 Node * in1 = n->in(1);
544 Node * in2 = n->in(2);
545 if (in1->Opcode() == Op_MulI && in2->Opcode() == Op_MulI) {
546 IdealLoopTree* loop_n = get_loop(get_ctrl(n));
547 if (loop_n->is_counted() &&
548 loop_n->_head->as_Loop()->is_valid_counted_loop(T_INT) &&
549 Matcher::match_rule_supported(Op_MulAddVS2VI) &&
550 Matcher::match_rule_supported(Op_MulAddS2I)) {
551 Node* mul_in1 = in1->in(1);
552 Node* mul_in2 = in1->in(2);
553 Node* mul_in3 = in2->in(1);
554 Node* mul_in4 = in2->in(2);
555 if (mul_in1->Opcode() == Op_LoadS &&
556 mul_in2->Opcode() == Op_LoadS &&
557 mul_in3->Opcode() == Op_LoadS &&
558 mul_in4->Opcode() == Op_LoadS) {
559 IdealLoopTree* loop1 = get_loop(get_ctrl(mul_in1));
560 IdealLoopTree* loop2 = get_loop(get_ctrl(mul_in2));
561 IdealLoopTree* loop3 = get_loop(get_ctrl(mul_in3));
562 IdealLoopTree* loop4 = get_loop(get_ctrl(mul_in4));
563 IdealLoopTree* loop5 = get_loop(get_ctrl(in1));
564 IdealLoopTree* loop6 = get_loop(get_ctrl(in2));
565 // All nodes should be in the same counted loop.
566 if (loop_n == loop1 && loop_n == loop2 && loop_n == loop3 &&
567 loop_n == loop4 && loop_n == loop5 && loop_n == loop6) {
568 Node* adr1 = mul_in1->in(MemNode::Address);
569 Node* adr2 = mul_in2->in(MemNode::Address);
570 Node* adr3 = mul_in3->in(MemNode::Address);
571 Node* adr4 = mul_in4->in(MemNode::Address);
572 if (adr1->is_AddP() && adr2->is_AddP() && adr3->is_AddP() && adr4->is_AddP()) {
573 if ((adr1->in(AddPNode::Base) == adr3->in(AddPNode::Base)) &&
574 (adr2->in(AddPNode::Base) == adr4->in(AddPNode::Base))) {
575 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in3, mul_in4);
576 register_new_node(nn, get_ctrl(n));
577 _igvn.replace_node(n, nn);
578 return nn;
579 } else if ((adr1->in(AddPNode::Base) == adr4->in(AddPNode::Base)) &&
580 (adr2->in(AddPNode::Base) == adr3->in(AddPNode::Base))) {
581 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in4, mul_in3);
582 register_new_node(nn, get_ctrl(n));
583 _igvn.replace_node(n, nn);
584 return nn;
585 }
586 }
587 }
588 }
589 }
590 }
591 return nn;
592 }
593
594 //------------------------------conditional_move-------------------------------
595 // Attempt to replace a Phi with a conditional move. We have some pretty
596 // strict profitability requirements. All Phis at the merge point must
597 // be converted, so we can remove the control flow. We need to limit the
598 // number of c-moves to a small handful. All code that was in the side-arms
599 // of the CFG diamond is now speculatively executed. This code has to be
600 // "cheap enough". We are pretty much limited to CFG diamonds that merge
601 // 1 or 2 items with a total of 1 or 2 ops executed speculatively.
602 Node *PhaseIdealLoop::conditional_move( Node *region ) {
603
604 assert(region->is_Region(), "sanity check");
605 if (region->req() != 3) return NULL;
606
607 // Check for CFG diamond
608 Node *lp = region->in(1);
609 Node *rp = region->in(2);
610 if (!lp || !rp) return NULL;
611 Node *lp_c = lp->in(0);
612 if (lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If()) return NULL;
613 IfNode *iff = lp_c->as_If();
614
615 // Check for ops pinned in an arm of the diamond.
616 // Can't remove the control flow in this case
617 if (lp->outcnt() > 1) return NULL;
618 if (rp->outcnt() > 1) return NULL;
619
620 IdealLoopTree* r_loop = get_loop(region);
621 assert(r_loop == get_loop(iff), "sanity");
622 // Always convert to CMOVE if all results are used only outside this loop.
623 bool used_inside_loop = (r_loop == _ltree_root);
624
625 // Check profitability
626 int cost = 0;
627 int phis = 0;
628 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
629 Node *out = region->fast_out(i);
630 if (!out->is_Phi()) continue; // Ignore other control edges, etc
631 phis++;
632 PhiNode* phi = out->as_Phi();
633 BasicType bt = phi->type()->basic_type();
634 switch (bt) {
635 case T_DOUBLE:
636 case T_FLOAT:
637 if (C->use_cmove()) {
638 continue; //TODO: maybe we want to add some cost
639 }
640 cost += Matcher::float_cmove_cost(); // Could be very expensive
641 break;
642 case T_LONG: {
643 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's
644 }
645 case T_INT: // These all CMOV fine
646 case T_ADDRESS: { // (RawPtr)
647 cost++;
648 break;
649 }
650 case T_NARROWOOP: // Fall through
651 case T_OBJECT: { // Base oops are OK, but not derived oops
652 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr();
653 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a
654 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus
655 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we
656 // have a Phi for the base here that we convert to a CMOVE all is well
657 // and good. But if the base is dead, we'll not make a CMOVE. Later
658 // the allocator will have to produce a base by creating a CMOVE of the
659 // relevant bases. This puts the allocator in the business of
660 // manufacturing expensive instructions, generally a bad plan.
661 // Just Say No to Conditionally-Moved Derived Pointers.
662 if (tp && tp->offset() != 0)
663 return NULL;
664 cost++;
665 break;
666 }
667 default:
668 return NULL; // In particular, can't do memory or I/O
669 }
670 // Add in cost any speculative ops
671 for (uint j = 1; j < region->req(); j++) {
672 Node *proj = region->in(j);
673 Node *inp = phi->in(j);
674 if (get_ctrl(inp) == proj) { // Found local op
675 cost++;
676 // Check for a chain of dependent ops; these will all become
677 // speculative in a CMOV.
678 for (uint k = 1; k < inp->req(); k++)
679 if (get_ctrl(inp->in(k)) == proj)
680 cost += ConditionalMoveLimit; // Too much speculative goo
681 }
682 }
683 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode.
684 // This will likely Split-If, a higher-payoff operation.
685 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) {
686 Node* use = phi->fast_out(k);
687 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr())
688 cost += ConditionalMoveLimit;
689 // Is there a use inside the loop?
690 // Note: check only basic types since CMoveP is pinned.
691 if (!used_inside_loop && is_java_primitive(bt)) {
692 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use);
693 if (r_loop == u_loop || r_loop->is_member(u_loop)) {
694 used_inside_loop = true;
695 }
696 }
697 }
698 }//for
699 Node* bol = iff->in(1);
700 if (bol->Opcode() == Op_Opaque4) {
701 return NULL; // Ignore loop predicate checks (the Opaque4 ensures they will go away)
702 }
703 assert(bol->Opcode() == Op_Bool, "Unexpected node");
704 int cmp_op = bol->in(1)->Opcode();
705 if (cmp_op == Op_SubTypeCheck) { // SubTypeCheck expansion expects an IfNode
706 return NULL;
707 }
708 // It is expensive to generate flags from a float compare.
709 // Avoid duplicated float compare.
710 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL;
711
712 float infrequent_prob = PROB_UNLIKELY_MAG(3);
713 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop.
714 if (used_inside_loop) {
715 if (cost >= ConditionalMoveLimit) return NULL; // Too much goo
716
717 // BlockLayoutByFrequency optimization moves infrequent branch
718 // from hot path. No point in CMOV'ing in such case (110 is used
719 // instead of 100 to take into account not exactness of float value).
720 if (BlockLayoutByFrequency) {
721 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f);
722 }
723 }
724 // Check for highly predictable branch. No point in CMOV'ing if
725 // we are going to predict accurately all the time.
726 if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) {
727 //keep going
728 } else if (iff->_prob < infrequent_prob ||
729 iff->_prob > (1.0f - infrequent_prob))
730 return NULL;
731
732 // --------------
733 // Now replace all Phis with CMOV's
734 Node *cmov_ctrl = iff->in(0);
735 uint flip = (lp->Opcode() == Op_IfTrue);
736 Node_List wq;
737 while (1) {
738 PhiNode* phi = NULL;
739 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
740 Node *out = region->fast_out(i);
741 if (out->is_Phi()) {
742 phi = out->as_Phi();
743 break;
744 }
745 }
746 if (phi == NULL || _igvn.type(phi) == Type::TOP) {
747 break;
748 }
749 if (PrintOpto && VerifyLoopOptimizations) { tty->print_cr("CMOV"); }
750 // Move speculative ops
751 wq.push(phi);
752 while (wq.size() > 0) {
753 Node *n = wq.pop();
754 for (uint j = 1; j < n->req(); j++) {
755 Node* m = n->in(j);
756 if (m != NULL && !is_dominator(get_ctrl(m), cmov_ctrl)) {
757 #ifndef PRODUCT
758 if (PrintOpto && VerifyLoopOptimizations) {
759 tty->print(" speculate: ");
760 m->dump();
761 }
762 #endif
763 set_ctrl(m, cmov_ctrl);
764 wq.push(m);
765 }
766 }
767 }
768 Node *cmov = CMoveNode::make(cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi));
769 register_new_node( cmov, cmov_ctrl );
770 _igvn.replace_node( phi, cmov );
771 #ifndef PRODUCT
772 if (TraceLoopOpts) {
773 tty->print("CMOV ");
774 r_loop->dump_head();
775 if (Verbose) {
776 bol->in(1)->dump(1);
777 cmov->dump(1);
778 }
779 }
780 if (VerifyLoopOptimizations) verify();
781 #endif
782 }
783
784 // The useless CFG diamond will fold up later; see the optimization in
785 // RegionNode::Ideal.
786 _igvn._worklist.push(region);
787
788 return iff->in(1);
789 }
790
791 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) {
792 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
793 Node* u = m->fast_out(i);
794 if (u->is_CFG()) {
795 if (u->Opcode() == Op_NeverBranch) {
796 u = ((NeverBranchNode*)u)->proj_out(0);
797 enqueue_cfg_uses(u, wq);
798 } else {
799 wq.push(u);
800 }
801 }
802 }
803 }
804
805 // Try moving a store out of a loop, right before the loop
806 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) {
807 // Store has to be first in the loop body
808 IdealLoopTree *n_loop = get_loop(n_ctrl);
809 if (n->is_Store() && n_loop != _ltree_root &&
810 n_loop->is_loop() && n_loop->_head->is_Loop() &&
811 n->in(0) != NULL) {
812 Node* address = n->in(MemNode::Address);
813 Node* value = n->in(MemNode::ValueIn);
814 Node* mem = n->in(MemNode::Memory);
815 IdealLoopTree* address_loop = get_loop(get_ctrl(address));
816 IdealLoopTree* value_loop = get_loop(get_ctrl(value));
817
818 // - address and value must be loop invariant
819 // - memory must be a memory Phi for the loop
820 // - Store must be the only store on this memory slice in the
821 // loop: if there's another store following this one then value
822 // written at iteration i by the second store could be overwritten
823 // at iteration i+n by the first store: it's not safe to move the
824 // first store out of the loop
825 // - nothing must observe the memory Phi: it guarantees no read
826 // before the store, we are also guaranteed the store post
827 // dominates the loop head (ignoring a possible early
828 // exit). Otherwise there would be extra Phi involved between the
829 // loop's Phi and the store.
830 // - there must be no early exit from the loop before the Store
831 // (such an exit most of the time would be an extra use of the
832 // memory Phi but sometimes is a bottom memory Phi that takes the
833 // store as input).
834
835 if (!n_loop->is_member(address_loop) &&
836 !n_loop->is_member(value_loop) &&
837 mem->is_Phi() && mem->in(0) == n_loop->_head &&
838 mem->outcnt() == 1 &&
839 mem->in(LoopNode::LoopBackControl) == n) {
840
841 assert(n_loop->_tail != NULL, "need a tail");
842 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop");
843
844 // Verify that there's no early exit of the loop before the store.
845 bool ctrl_ok = false;
846 {
847 // Follow control from loop head until n, we exit the loop or
848 // we reach the tail
849 ResourceMark rm;
850 Unique_Node_List wq;
851 wq.push(n_loop->_head);
852
853 for (uint next = 0; next < wq.size(); ++next) {
854 Node *m = wq.at(next);
855 if (m == n->in(0)) {
856 ctrl_ok = true;
857 continue;
858 }
859 assert(!has_ctrl(m), "should be CFG");
860 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) {
861 ctrl_ok = false;
862 break;
863 }
864 enqueue_cfg_uses(m, wq);
865 if (wq.size() > 10) {
866 ctrl_ok = false;
867 break;
868 }
869 }
870 }
871 if (ctrl_ok) {
872 // move the Store
873 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem);
874 _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl));
875 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl));
876 // Disconnect the phi now. An empty phi can confuse other
877 // optimizations in this pass of loop opts.
878 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl));
879 n_loop->_body.yank(mem);
880
881 set_ctrl_and_loop(n, n->in(0));
882
883 return n;
884 }
885 }
886 }
887 return NULL;
888 }
889
890 // Try moving a store out of a loop, right after the loop
891 void PhaseIdealLoop::try_move_store_after_loop(Node* n) {
892 if (n->is_Store() && n->in(0) != NULL) {
893 Node *n_ctrl = get_ctrl(n);
894 IdealLoopTree *n_loop = get_loop(n_ctrl);
895 // Store must be in a loop
896 if (n_loop != _ltree_root && !n_loop->_irreducible) {
897 Node* address = n->in(MemNode::Address);
898 Node* value = n->in(MemNode::ValueIn);
899 IdealLoopTree* address_loop = get_loop(get_ctrl(address));
900 // address must be loop invariant
901 if (!n_loop->is_member(address_loop)) {
902 // Store must be last on this memory slice in the loop and
903 // nothing in the loop must observe it
904 Node* phi = NULL;
905 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
906 Node* u = n->fast_out(i);
907 if (has_ctrl(u)) { // control use?
908 IdealLoopTree *u_loop = get_loop(get_ctrl(u));
909 if (!n_loop->is_member(u_loop)) {
910 continue;
911 }
912 if (u->is_Phi() && u->in(0) == n_loop->_head) {
913 assert(_igvn.type(u) == Type::MEMORY, "bad phi");
914 // multiple phis on the same slice are possible
915 if (phi != NULL) {
916 return;
917 }
918 phi = u;
919 continue;
920 }
921 }
922 return;
923 }
924 if (phi != NULL) {
925 // Nothing in the loop before the store (next iteration)
926 // must observe the stored value
927 bool mem_ok = true;
928 {
929 ResourceMark rm;
930 Unique_Node_List wq;
931 wq.push(phi);
932 for (uint next = 0; next < wq.size() && mem_ok; ++next) {
933 Node *m = wq.at(next);
934 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) {
935 Node* u = m->fast_out(i);
936 if (u->is_Store() || u->is_Phi()) {
937 if (u != n) {
938 wq.push(u);
939 mem_ok = (wq.size() <= 10);
940 }
941 } else {
942 mem_ok = false;
943 break;
944 }
945 }
946 }
947 }
948 if (mem_ok) {
949 // Move the store out of the loop if the LCA of all
950 // users (except for the phi) is outside the loop.
951 Node* hook = new Node(1);
952 hook->init_req(0, n_ctrl); // Add an input to prevent hook from being dead
953 _igvn.rehash_node_delayed(phi);
954 int count = phi->replace_edge(n, hook, &_igvn);
955 assert(count > 0, "inconsistent phi");
956
957 // Compute latest point this store can go
958 Node* lca = get_late_ctrl(n, get_ctrl(n));
959 if (lca->is_OuterStripMinedLoop()) {
960 lca = lca->in(LoopNode::EntryControl);
961 }
962 if (n_loop->is_member(get_loop(lca))) {
963 // LCA is in the loop - bail out
964 _igvn.replace_node(hook, n);
965 return;
966 }
967 #ifdef ASSERT
968 if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) {
969 assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined");
970 n_loop->_head->as_Loop()->verify_strip_mined(1);
971 Node* outer = n_loop->_head->as_CountedLoop()->outer_loop();
972 IdealLoopTree* outer_loop = get_loop(outer);
973 assert(n_loop->_parent == outer_loop, "broken loop tree");
974 assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state");
975 }
976 #endif
977 lca = place_outside_loop(lca, n_loop);
978 assert(!n_loop->is_member(get_loop(lca)), "control must not be back in the loop");
979 assert(get_loop(lca)->_nest < n_loop->_nest || lca->in(0)->Opcode() == Op_NeverBranch, "must not be moved into inner loop");
980
981 // Move store out of the loop
982 _igvn.replace_node(hook, n->in(MemNode::Memory));
983 _igvn.replace_input_of(n, 0, lca);
984 set_ctrl_and_loop(n, lca);
985
986 // Disconnect the phi now. An empty phi can confuse other
987 // optimizations in this pass of loop opts..
988 if (phi->in(LoopNode::LoopBackControl) == phi) {
989 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl));
990 n_loop->_body.yank(phi);
991 }
992 }
993 }
994 }
995 }
996 }
997 }
998
999 //------------------------------split_if_with_blocks_pre-----------------------
1000 // Do the real work in a non-recursive function. Data nodes want to be
1001 // cloned in the pre-order so they can feed each other nicely.
1002 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
1003 // Cloning these guys is unlikely to win
1004 int n_op = n->Opcode();
1005 if (n_op == Op_MergeMem) {
1006 return n;
1007 }
1008 if (n->is_Proj()) {
1009 return n;
1010 }
1011 // Do not clone-up CmpFXXX variations, as these are always
1012 // followed by a CmpI
1013 if (n->is_Cmp()) {
1014 return n;
1015 }
1016 // Attempt to use a conditional move instead of a phi/branch
1017 if (ConditionalMoveLimit > 0 && n_op == Op_Region) {
1018 Node *cmov = conditional_move( n );
1019 if (cmov) {
1020 return cmov;
1021 }
1022 }
1023 if (n->is_CFG() || n->is_LoadStore()) {
1024 return n;
1025 }
1026 if (n->is_Opaque1() || // Opaque nodes cannot be mod'd
1027 n_op == Op_Opaque2) {
1028 if (!C->major_progress()) { // If chance of no more loop opts...
1029 _igvn._worklist.push(n); // maybe we'll remove them
1030 }
1031 return n;
1032 }
1033
1034 if (n->is_Con()) {
1035 return n; // No cloning for Con nodes
1036 }
1037
1038 Node *n_ctrl = get_ctrl(n);
1039 if (!n_ctrl) {
1040 return n; // Dead node
1041 }
1042
1043 Node* res = try_move_store_before_loop(n, n_ctrl);
1044 if (res != NULL) {
1045 return n;
1046 }
1047
1048 // Attempt to remix address expressions for loop invariants
1049 Node *m = remix_address_expressions( n );
1050 if( m ) return m;
1051
1052 if (n_op == Op_AddI) {
1053 Node *nn = convert_add_to_muladd( n );
1054 if ( nn ) return nn;
1055 }
1056
1057 if (n->is_ConstraintCast()) {
1058 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this);
1059 // ConstraintCastNode::dominating_cast() uses node control input to determine domination.
1060 // Node control inputs don't necessarily agree with loop control info (due to
1061 // transformations happened in between), thus additional dominance check is needed
1062 // to keep loop info valid.
1063 if (dom_cast != NULL && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) {
1064 _igvn.replace_node(n, dom_cast);
1065 return dom_cast;
1066 }
1067 }
1068
1069 // Determine if the Node has inputs from some local Phi.
1070 // Returns the block to clone thru.
1071 Node *n_blk = has_local_phi_input( n );
1072 if( !n_blk ) return n;
1073
1074 // Do not clone the trip counter through on a CountedLoop
1075 // (messes up the canonical shape).
1076 if (((n_blk->is_CountedLoop() || (n_blk->is_Loop() && n_blk->as_Loop()->is_loop_nest_inner_loop())) && n->Opcode() == Op_AddI) ||
1077 (n_blk->is_LongCountedLoop() && n->Opcode() == Op_AddL)) {
1078 return n;
1079 }
1080 // Pushing a shift through the iv Phi can get in the way of addressing optimizations or range check elimination
1081 if (n_blk->is_BaseCountedLoop() && n->Opcode() == Op_LShift(n_blk->as_BaseCountedLoop()->bt()) &&
1082 n->in(1) == n_blk->as_BaseCountedLoop()->phi()) {
1083 return n;
1084 }
1085
1086 // Check for having no control input; not pinned. Allow
1087 // dominating control.
1088 if (n->in(0)) {
1089 Node *dom = idom(n_blk);
1090 if (dom_lca(n->in(0), dom) != n->in(0)) {
1091 return n;
1092 }
1093 }
1094 // Policy: when is it profitable. You must get more wins than
1095 // policy before it is considered profitable. Policy is usually 0,
1096 // so 1 win is considered profitable. Big merges will require big
1097 // cloning, so get a larger policy.
1098 int policy = n_blk->req() >> 2;
1099
1100 // If the loop is a candidate for range check elimination,
1101 // delay splitting through it's phi until a later loop optimization
1102 if (n_blk->is_BaseCountedLoop()) {
1103 IdealLoopTree *lp = get_loop(n_blk);
1104 if (lp && lp->_rce_candidate) {
1105 return n;
1106 }
1107 }
1108
1109 if (must_throttle_split_if()) return n;
1110
1111 // Split 'n' through the merge point if it is profitable
1112 Node *phi = split_thru_phi( n, n_blk, policy );
1113 if (!phi) return n;
1114
1115 // Found a Phi to split thru!
1116 // Replace 'n' with the new phi
1117 _igvn.replace_node( n, phi );
1118 // Moved a load around the loop, 'en-registering' something.
1119 if (n_blk->is_Loop() && n->is_Load() &&
1120 !phi->in(LoopNode::LoopBackControl)->is_Load())
1121 C->set_major_progress();
1122
1123 return phi;
1124 }
1125
1126 static bool merge_point_too_heavy(Compile* C, Node* region) {
1127 // Bail out if the region and its phis have too many users.
1128 int weight = 0;
1129 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1130 weight += region->fast_out(i)->outcnt();
1131 }
1132 int nodes_left = C->max_node_limit() - C->live_nodes();
1133 if (weight * 8 > nodes_left) {
1134 if (PrintOpto) {
1135 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight);
1136 }
1137 return true;
1138 } else {
1139 return false;
1140 }
1141 }
1142
1143 static bool merge_point_safe(Node* region) {
1144 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode
1145 // having a PhiNode input. This sidesteps the dangerous case where the split
1146 // ConvI2LNode may become TOP if the input Value() does not
1147 // overlap the ConvI2L range, leaving a node which may not dominate its
1148 // uses.
1149 // A better fix for this problem can be found in the BugTraq entry, but
1150 // expediency for Mantis demands this hack.
1151 #ifdef _LP64
1152 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1153 Node* n = region->fast_out(i);
1154 if (n->is_Phi()) {
1155 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1156 Node* m = n->fast_out(j);
1157 if (m->Opcode() == Op_ConvI2L)
1158 return false;
1159 if (m->is_CastII()) {
1160 return false;
1161 }
1162 }
1163 }
1164 }
1165 #endif
1166 return true;
1167 }
1168
1169
1170 //------------------------------place_outside_loop---------------------------------
1171 // Place some computation outside of this loop on the path to the use passed as argument
1172 Node* PhaseIdealLoop::place_outside_loop(Node* useblock, IdealLoopTree* loop) const {
1173 Node* head = loop->_head;
1174 assert(!loop->is_member(get_loop(useblock)), "must be outside loop");
1175 if (head->is_Loop() && head->as_Loop()->is_strip_mined()) {
1176 loop = loop->_parent;
1177 assert(loop->_head->is_OuterStripMinedLoop(), "malformed strip mined loop");
1178 }
1179
1180 // Pick control right outside the loop
1181 for (;;) {
1182 Node* dom = idom(useblock);
1183 if (loop->is_member(get_loop(dom)) ||
1184 // NeverBranch nodes are not assigned to the loop when constructed
1185 (dom->Opcode() == Op_NeverBranch && loop->is_member(get_loop(dom->in(0))))) {
1186 break;
1187 }
1188 useblock = dom;
1189 }
1190 assert(find_non_split_ctrl(useblock) == useblock, "should be non split control");
1191 return useblock;
1192 }
1193
1194
1195 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) {
1196 if (!n->is_If() || n->is_BaseCountedLoopEnd()) {
1197 return false;
1198 }
1199 if (!n->in(0)->is_Region()) {
1200 return false;
1201 }
1202
1203 Node* region = n->in(0);
1204 Node* dom = idom(region);
1205 if (!dom->is_If() || dom->in(1) != n->in(1)) {
1206 return false;
1207 }
1208 IfNode* dom_if = dom->as_If();
1209 Node* proj_true = dom_if->proj_out(1);
1210 Node* proj_false = dom_if->proj_out(0);
1211
1212 for (uint i = 1; i < region->req(); i++) {
1213 if (is_dominator(proj_true, region->in(i))) {
1214 continue;
1215 }
1216 if (is_dominator(proj_false, region->in(i))) {
1217 continue;
1218 }
1219 return false;
1220 }
1221
1222 return true;
1223 }
1224
1225
1226 bool PhaseIdealLoop::can_split_if(Node* n_ctrl) {
1227 if (must_throttle_split_if()) {
1228 return false;
1229 }
1230
1231 // Do not do 'split-if' if irreducible loops are present.
1232 if (_has_irreducible_loops) {
1233 return false;
1234 }
1235
1236 if (merge_point_too_heavy(C, n_ctrl)) {
1237 return false;
1238 }
1239
1240 // Do not do 'split-if' if some paths are dead. First do dead code
1241 // elimination and then see if its still profitable.
1242 for (uint i = 1; i < n_ctrl->req(); i++) {
1243 if (n_ctrl->in(i) == C->top()) {
1244 return false;
1245 }
1246 }
1247
1248 // If trying to do a 'Split-If' at the loop head, it is only
1249 // profitable if the cmp folds up on BOTH paths. Otherwise we
1250 // risk peeling a loop forever.
1251
1252 // CNC - Disabled for now. Requires careful handling of loop
1253 // body selection for the cloned code. Also, make sure we check
1254 // for any input path not being in the same loop as n_ctrl. For
1255 // irreducible loops we cannot check for 'n_ctrl->is_Loop()'
1256 // because the alternative loop entry points won't be converted
1257 // into LoopNodes.
1258 IdealLoopTree *n_loop = get_loop(n_ctrl);
1259 for (uint j = 1; j < n_ctrl->req(); j++) {
1260 if (get_loop(n_ctrl->in(j)) != n_loop) {
1261 return false;
1262 }
1263 }
1264
1265 // Check for safety of the merge point.
1266 if (!merge_point_safe(n_ctrl)) {
1267 return false;
1268 }
1269
1270 return true;
1271 }
1272
1273 // Detect if the node is the inner strip-mined loop
1274 // Return: NULL if it's not the case, or the exit of outer strip-mined loop
1275 static Node* is_inner_of_stripmined_loop(const Node* out) {
1276 Node* out_le = NULL;
1277
1278 if (out->is_CountedLoopEnd()) {
1279 const CountedLoopNode* loop = out->as_CountedLoopEnd()->loopnode();
1280
1281 if (loop != NULL && loop->is_strip_mined()) {
1282 out_le = loop->in(LoopNode::EntryControl)->as_OuterStripMinedLoop()->outer_loop_exit();
1283 }
1284 }
1285
1286 return out_le;
1287 }
1288
1289 //------------------------------split_if_with_blocks_post----------------------
1290 // Do the real work in a non-recursive function. CFG hackery wants to be
1291 // in the post-order, so it can dirty the I-DOM info and not use the dirtied
1292 // info.
1293 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) {
1294
1295 // Cloning Cmp through Phi's involves the split-if transform.
1296 // FastLock is not used by an If
1297 if (n->is_Cmp() && !n->is_FastLock()) {
1298 Node *n_ctrl = get_ctrl(n);
1299 // Determine if the Node has inputs from some local Phi.
1300 // Returns the block to clone thru.
1301 Node *n_blk = has_local_phi_input(n);
1302 if (n_blk != n_ctrl) {
1303 return;
1304 }
1305
1306 if (!can_split_if(n_ctrl)) {
1307 return;
1308 }
1309
1310 if (n->outcnt() != 1) {
1311 return; // Multiple bool's from 1 compare?
1312 }
1313 Node *bol = n->unique_out();
1314 assert(bol->is_Bool(), "expect a bool here");
1315 if (bol->outcnt() != 1) {
1316 return;// Multiple branches from 1 compare?
1317 }
1318 Node *iff = bol->unique_out();
1319
1320 // Check some safety conditions
1321 if (iff->is_If()) { // Classic split-if?
1322 if (iff->in(0) != n_ctrl) {
1323 return; // Compare must be in same blk as if
1324 }
1325 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE
1326 // Can't split CMove with different control edge.
1327 if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) {
1328 return;
1329 }
1330 if (get_ctrl(iff->in(2)) == n_ctrl ||
1331 get_ctrl(iff->in(3)) == n_ctrl) {
1332 return; // Inputs not yet split-up
1333 }
1334 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) {
1335 return; // Loop-invar test gates loop-varying CMOVE
1336 }
1337 } else {
1338 return; // some other kind of node, such as an Allocate
1339 }
1340
1341 // When is split-if profitable? Every 'win' on means some control flow
1342 // goes dead, so it's almost always a win.
1343 int policy = 0;
1344 // Split compare 'n' through the merge point if it is profitable
1345 Node *phi = split_thru_phi( n, n_ctrl, policy);
1346 if (!phi) {
1347 return;
1348 }
1349
1350 // Found a Phi to split thru!
1351 // Replace 'n' with the new phi
1352 _igvn.replace_node(n, phi);
1353
1354 // Now split the bool up thru the phi
1355 Node *bolphi = split_thru_phi(bol, n_ctrl, -1);
1356 guarantee(bolphi != NULL, "null boolean phi node");
1357
1358 _igvn.replace_node(bol, bolphi);
1359 assert(iff->in(1) == bolphi, "");
1360
1361 if (bolphi->Value(&_igvn)->singleton()) {
1362 return;
1363 }
1364
1365 // Conditional-move? Must split up now
1366 if (!iff->is_If()) {
1367 Node *cmovphi = split_thru_phi(iff, n_ctrl, -1);
1368 _igvn.replace_node(iff, cmovphi);
1369 return;
1370 }
1371
1372 // Now split the IF
1373 do_split_if(iff);
1374 return;
1375 }
1376
1377 // Two identical ifs back to back can be merged
1378 if (try_merge_identical_ifs(n)) {
1379 return;
1380 }
1381
1382 // Check for an IF ready to split; one that has its
1383 // condition codes input coming from a Phi at the block start.
1384 int n_op = n->Opcode();
1385
1386 // Check for an IF being dominated by another IF same test
1387 if (n_op == Op_If ||
1388 n_op == Op_RangeCheck) {
1389 Node *bol = n->in(1);
1390 uint max = bol->outcnt();
1391 // Check for same test used more than once?
1392 if (max > 1 && bol->is_Bool()) {
1393 // Search up IDOMs to see if this IF is dominated.
1394 Node *cutoff = get_ctrl(bol);
1395
1396 // Now search up IDOMs till cutoff, looking for a dominating test
1397 Node *prevdom = n;
1398 Node *dom = idom(prevdom);
1399 while (dom != cutoff) {
1400 if (dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom &&
1401 safe_for_if_replacement(dom)) {
1402 // It's invalid to move control dependent data nodes in the inner
1403 // strip-mined loop, because:
1404 // 1) break validation of LoopNode::verify_strip_mined()
1405 // 2) move code with side-effect in strip-mined loop
1406 // Move to the exit of outer strip-mined loop in that case.
1407 Node* out_le = is_inner_of_stripmined_loop(dom);
1408 if (out_le != NULL) {
1409 prevdom = out_le;
1410 }
1411 // Replace the dominated test with an obvious true or false.
1412 // Place it on the IGVN worklist for later cleanup.
1413 C->set_major_progress();
1414 dominated_by(prevdom->as_IfProj(), n->as_If(), false, true);
1415 #ifndef PRODUCT
1416 if( VerifyLoopOptimizations ) verify();
1417 #endif
1418 return;
1419 }
1420 prevdom = dom;
1421 dom = idom(prevdom);
1422 }
1423 }
1424 }
1425
1426 try_sink_out_of_loop(n);
1427
1428 try_move_store_after_loop(n);
1429
1430 // Check for Opaque2's who's loop has disappeared - who's input is in the
1431 // same loop nest as their output. Remove 'em, they are no longer useful.
1432 if( n_op == Op_Opaque2 &&
1433 n->in(1) != NULL &&
1434 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) {
1435 _igvn.replace_node( n, n->in(1) );
1436 }
1437 }
1438
1439 // Transform:
1440 //
1441 // if (some_condition) {
1442 // // body 1
1443 // } else {
1444 // // body 2
1445 // }
1446 // if (some_condition) {
1447 // // body 3
1448 // } else {
1449 // // body 4
1450 // }
1451 //
1452 // into:
1453 //
1454 //
1455 // if (some_condition) {
1456 // // body 1
1457 // // body 3
1458 // } else {
1459 // // body 2
1460 // // body 4
1461 // }
1462 bool PhaseIdealLoop::try_merge_identical_ifs(Node* n) {
1463 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) {
1464 Node *n_ctrl = n->in(0);
1465 IfNode* dom_if = idom(n_ctrl)->as_If();
1466 ProjNode* dom_proj_true = dom_if->proj_out(1);
1467 ProjNode* dom_proj_false = dom_if->proj_out(0);
1468
1469 // Now split the IF
1470 RegionNode* new_false_region;
1471 RegionNode* new_true_region;
1472 do_split_if(n, &new_false_region, &new_true_region);
1473 assert(new_false_region->req() == new_true_region->req(), "");
1474 #ifdef ASSERT
1475 for (uint i = 1; i < new_false_region->req(); ++i) {
1476 assert(new_false_region->in(i)->in(0) == new_true_region->in(i)->in(0), "unexpected shape following split if");
1477 assert(i == new_false_region->req() - 1 || new_false_region->in(i)->in(0)->in(1) == new_false_region->in(i + 1)->in(0)->in(1), "unexpected shape following split if");
1478 }
1479 #endif
1480 assert(new_false_region->in(1)->in(0)->in(1) == dom_if->in(1), "dominating if and dominated if after split must share test");
1481
1482 // We now have:
1483 // if (some_condition) {
1484 // // body 1
1485 // if (some_condition) {
1486 // body3: // new_true_region
1487 // // body3
1488 // } else {
1489 // goto body4;
1490 // }
1491 // } else {
1492 // // body 2
1493 // if (some_condition) {
1494 // goto body3;
1495 // } else {
1496 // body4: // new_false_region
1497 // // body4;
1498 // }
1499 // }
1500 //
1501
1502 // clone pinned nodes thru the resulting regions
1503 push_pinned_nodes_thru_region(dom_if, new_true_region);
1504 push_pinned_nodes_thru_region(dom_if, new_false_region);
1505
1506 // Optimize out the cloned ifs. Because pinned nodes were cloned, this also allows a CastPP that would be dependent
1507 // on a projection of n to have the dom_if as a control dependency. We don't want the CastPP to end up with an
1508 // unrelated control dependency.
1509 for (uint i = 1; i < new_false_region->req(); i++) {
1510 if (is_dominator(dom_proj_true, new_false_region->in(i))) {
1511 dominated_by(dom_proj_true->as_IfProj(), new_false_region->in(i)->in(0)->as_If(), false, false);
1512 } else {
1513 assert(is_dominator(dom_proj_false, new_false_region->in(i)), "bad if");
1514 dominated_by(dom_proj_false->as_IfProj(), new_false_region->in(i)->in(0)->as_If(), false, false);
1515 }
1516 }
1517 return true;
1518 }
1519 return false;
1520 }
1521
1522 void PhaseIdealLoop::push_pinned_nodes_thru_region(IfNode* dom_if, Node* region) {
1523 for (DUIterator i = region->outs(); region->has_out(i); i++) {
1524 Node* u = region->out(i);
1525 if (!has_ctrl(u) || u->is_Phi() || !u->depends_only_on_test() || !_igvn.no_dependent_zero_check(u)) {
1526 continue;
1527 }
1528 assert(u->in(0) == region, "not a control dependent node?");
1529 uint j = 1;
1530 for (; j < u->req(); ++j) {
1531 Node* in = u->in(j);
1532 if (!is_dominator(ctrl_or_self(in), dom_if)) {
1533 break;
1534 }
1535 }
1536 if (j == u->req()) {
1537 Node *phi = PhiNode::make_blank(region, u);
1538 for (uint k = 1; k < region->req(); ++k) {
1539 Node* clone = u->clone();
1540 clone->set_req(0, region->in(k));
1541 register_new_node(clone, region->in(k));
1542 phi->init_req(k, clone);
1543 }
1544 register_new_node(phi, region);
1545 _igvn.replace_node(u, phi);
1546 --i;
1547 }
1548 }
1549 }
1550
1551 bool PhaseIdealLoop::safe_for_if_replacement(const Node* dom) const {
1552 if (!dom->is_CountedLoopEnd()) {
1553 return true;
1554 }
1555 CountedLoopEndNode* le = dom->as_CountedLoopEnd();
1556 CountedLoopNode* cl = le->loopnode();
1557 if (cl == NULL) {
1558 return true;
1559 }
1560 if (!cl->is_main_loop()) {
1561 return true;
1562 }
1563 if (cl->is_canonical_loop_entry() == NULL) {
1564 return true;
1565 }
1566 // Further unrolling is possible so loop exit condition might change
1567 return false;
1568 }
1569
1570 // See if a shared loop-varying computation has no loop-varying uses.
1571 // Happens if something is only used for JVM state in uncommon trap exits,
1572 // like various versions of induction variable+offset. Clone the
1573 // computation per usage to allow it to sink out of the loop.
1574 void PhaseIdealLoop::try_sink_out_of_loop(Node* n) {
1575 bool is_raw_to_oop_cast = n->is_ConstraintCast() &&
1576 n->in(1)->bottom_type()->isa_rawptr() &&
1577 !n->bottom_type()->isa_rawptr();
1578 if (has_ctrl(n) &&
1579 !n->is_Phi() &&
1580 !n->is_Bool() &&
1581 !n->is_Proj() &&
1582 !n->is_MergeMem() &&
1583 !n->is_CMove() &&
1584 !is_raw_to_oop_cast && // don't extend live ranges of raw oops
1585 n->Opcode() != Op_Opaque4 &&
1586 !n->is_Type()) {
1587 Node *n_ctrl = get_ctrl(n);
1588 IdealLoopTree *n_loop = get_loop(n_ctrl);
1589
1590 if (n->in(0) != NULL) {
1591 IdealLoopTree* loop_ctrl = get_loop(n->in(0));
1592 if (n_loop != loop_ctrl && n_loop->is_member(loop_ctrl)) {
1593 // n has a control input inside a loop but get_ctrl() is member of an outer loop. This could happen, for example,
1594 // for Div nodes inside a loop (control input inside loop) without a use except for an UCT (outside the loop).
1595 // Rewire control of n to right outside of the loop, regardless if its input(s) are later sunk or not.
1596 _igvn.replace_input_of(n, 0, place_outside_loop(n_ctrl, loop_ctrl));
1597 }
1598 }
1599 if (n_loop != _ltree_root && n->outcnt() > 1) {
1600 // Compute early control: needed for anti-dependence analysis. It's also possible that as a result of
1601 // previous transformations in this loop opts round, the node can be hoisted now: early control will tell us.
1602 Node* early_ctrl = compute_early_ctrl(n, n_ctrl);
1603 if (n_loop->is_member(get_loop(early_ctrl)) && // check that this one can't be hoisted now
1604 ctrl_of_all_uses_out_of_loop(n, early_ctrl, n_loop)) { // All uses in outer loops!
1605 assert(!n->is_Store() && !n->is_LoadStore(), "no node with a side effect");
1606 Node* outer_loop_clone = NULL;
1607 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin;) {
1608 Node* u = n->last_out(j); // Clone private computation per use
1609 _igvn.rehash_node_delayed(u);
1610 Node* x = n->clone(); // Clone computation
1611 Node* x_ctrl = NULL;
1612 if (u->is_Phi()) {
1613 // Replace all uses of normal nodes. Replace Phi uses
1614 // individually, so the separate Nodes can sink down
1615 // different paths.
1616 uint k = 1;
1617 while (u->in(k) != n) k++;
1618 u->set_req(k, x);
1619 // x goes next to Phi input path
1620 x_ctrl = u->in(0)->in(k);
1621 // Find control for 'x' next to use but not inside inner loops.
1622 x_ctrl = place_outside_loop(x_ctrl, n_loop);
1623 --j;
1624 } else { // Normal use
1625 if (has_ctrl(u)) {
1626 x_ctrl = get_ctrl(u);
1627 } else {
1628 x_ctrl = u->in(0);
1629 }
1630 // Find control for 'x' next to use but not inside inner loops.
1631 x_ctrl = place_outside_loop(x_ctrl, n_loop);
1632 // Replace all uses
1633 if (u->is_ConstraintCast() && u->bottom_type()->higher_equal(_igvn.type(n)) && u->in(0) == x_ctrl) {
1634 // If we're sinking a chain of data nodes, we might have inserted a cast to pin the use which is not necessary
1635 // anymore now that we're going to pin n as well
1636 _igvn.replace_node(u, x);
1637 --j;
1638 } else {
1639 int nb = u->replace_edge(n, x, &_igvn);
1640 j -= nb;
1641 }
1642 }
1643
1644 if (n->is_Load()) {
1645 // For loads, add a control edge to a CFG node outside of the loop
1646 // to force them to not combine and return back inside the loop
1647 // during GVN optimization (4641526).
1648 assert(x_ctrl == get_late_ctrl_with_anti_dep(x->as_Load(), early_ctrl, x_ctrl), "anti-dependences were already checked");
1649
1650 IdealLoopTree* x_loop = get_loop(x_ctrl);
1651 Node* x_head = x_loop->_head;
1652 if (x_head->is_Loop() && x_head->is_OuterStripMinedLoop()) {
1653 // Do not add duplicate LoadNodes to the outer strip mined loop
1654 if (outer_loop_clone != NULL) {
1655 _igvn.replace_node(x, outer_loop_clone);
1656 continue;
1657 }
1658 outer_loop_clone = x;
1659 }
1660 x->set_req(0, x_ctrl);
1661 } else if (n->in(0) != NULL){
1662 x->set_req(0, x_ctrl);
1663 }
1664 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone");
1665 assert(!n_loop->is_member(get_loop(x_ctrl)), "should have moved out of loop");
1666 register_new_node(x, x_ctrl);
1667
1668 // Chain of AddP: (AddP base (AddP base )) must keep the same base after sinking so:
1669 // 1- We don't add a CastPP here when the first one is sunk so if the second one is not, their bases remain
1670 // the same.
1671 // (see 2- below)
1672 assert(!x->is_AddP() || !x->in(AddPNode::Address)->is_AddP() ||
1673 x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base) ||
1674 !x->in(AddPNode::Address)->in(AddPNode::Base)->eqv_uncast(x->in(AddPNode::Base)), "unexpected AddP shape");
1675 if (x->in(0) == NULL && !x->is_DecodeNarrowPtr() &&
1676 !(x->is_AddP() && x->in(AddPNode::Address)->is_AddP() && x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base))) {
1677 assert(!x->is_Load(), "load should be pinned");
1678 // Use a cast node to pin clone out of loop
1679 Node* cast = NULL;
1680 for (uint k = 0; k < x->req(); k++) {
1681 Node* in = x->in(k);
1682 if (in != NULL && n_loop->is_member(get_loop(get_ctrl(in)))) {
1683 const Type* in_t = _igvn.type(in);
1684 cast = ConstraintCastNode::make_cast_for_type(x_ctrl, in, in_t, ConstraintCastNode::UnconditionalDependency);
1685 }
1686 if (cast != NULL) {
1687 register_new_node(cast, x_ctrl);
1688 x->replace_edge(in, cast);
1689 // Chain of AddP:
1690 // 2- A CastPP of the base is only added now that both AddP nodes are sunk
1691 if (x->is_AddP() && k == AddPNode::Base) {
1692 for (DUIterator_Fast imax, i = x->fast_outs(imax); i < imax; i++) {
1693 Node* u = x->fast_out(i);
1694 if (u->is_AddP() && u->in(AddPNode::Base) == n->in(AddPNode::Base)) {
1695 _igvn.replace_input_of(u, AddPNode::Base, cast);
1696 assert(u->find_out_with(Op_AddP) == NULL, "more than 2 chained AddP nodes?");
1697 }
1698 }
1699 }
1700 break;
1701 }
1702 }
1703 assert(cast != NULL, "must have added a cast to pin the node");
1704 }
1705 }
1706 _igvn.remove_dead_node(n);
1707 }
1708 _dom_lca_tags_round = 0;
1709 }
1710 }
1711 }
1712
1713 // Compute the early control of a node by following its inputs until we reach
1714 // nodes that are pinned. Then compute the LCA of the control of all pinned nodes.
1715 Node* PhaseIdealLoop::compute_early_ctrl(Node* n, Node* n_ctrl) {
1716 Node* early_ctrl = NULL;
1717 ResourceMark rm;
1718 Unique_Node_List wq;
1719 wq.push(n);
1720 for (uint i = 0; i < wq.size(); i++) {
1721 Node* m = wq.at(i);
1722 Node* c = NULL;
1723 if (m->is_CFG()) {
1724 c = m;
1725 } else if (m->pinned()) {
1726 c = m->in(0);
1727 } else {
1728 for (uint j = 0; j < m->req(); j++) {
1729 Node* in = m->in(j);
1730 if (in != NULL) {
1731 wq.push(in);
1732 }
1733 }
1734 }
1735 if (c != NULL) {
1736 assert(is_dominator(c, n_ctrl), "control input must dominate current control");
1737 if (early_ctrl == NULL || is_dominator(early_ctrl, c)) {
1738 early_ctrl = c;
1739 }
1740 }
1741 }
1742 assert(is_dominator(early_ctrl, n_ctrl), "early control must dominate current control");
1743 return early_ctrl;
1744 }
1745
1746 bool PhaseIdealLoop::ctrl_of_all_uses_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop) {
1747 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1748 Node* u = n->fast_out(i);
1749 if (u->Opcode() == Op_Opaque1) {
1750 return false; // Found loop limit, bugfix for 4677003
1751 }
1752 // We can't reuse tags in PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal() so make sure calls to
1753 // get_late_ctrl_with_anti_dep() use their own tag
1754 _dom_lca_tags_round++;
1755 assert(_dom_lca_tags_round != 0, "shouldn't wrap around");
1756
1757 if (u->is_Phi()) {
1758 for (uint j = 1; j < u->req(); ++j) {
1759 if (u->in(j) == n && !ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, u->in(0)->in(j))) {
1760 return false;
1761 }
1762 }
1763 } else {
1764 Node* ctrl = has_ctrl(u) ? get_ctrl(u) : u->in(0);
1765 if (!ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, ctrl)) {
1766 return false;
1767 }
1768 }
1769 }
1770 return true;
1771 }
1772
1773 bool PhaseIdealLoop::ctrl_of_use_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop, Node* ctrl) {
1774 if (n->is_Load()) {
1775 ctrl = get_late_ctrl_with_anti_dep(n->as_Load(), n_ctrl, ctrl);
1776 }
1777 IdealLoopTree *u_loop = get_loop(ctrl);
1778 if (u_loop == n_loop) {
1779 return false; // Found loop-varying use
1780 }
1781 if (n_loop->is_member(u_loop)) {
1782 return false; // Found use in inner loop
1783 }
1784 return true;
1785 }
1786
1787 //------------------------------split_if_with_blocks---------------------------
1788 // Check for aggressive application of 'split-if' optimization,
1789 // using basic block level info.
1790 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack) {
1791 Node* root = C->root();
1792 visited.set(root->_idx); // first, mark root as visited
1793 // Do pre-visit work for root
1794 Node* n = split_if_with_blocks_pre(root);
1795 uint cnt = n->outcnt();
1796 uint i = 0;
1797
1798 while (true) {
1799 // Visit all children
1800 if (i < cnt) {
1801 Node* use = n->raw_out(i);
1802 ++i;
1803 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
1804 // Now do pre-visit work for this use
1805 use = split_if_with_blocks_pre(use);
1806 nstack.push(n, i); // Save parent and next use's index.
1807 n = use; // Process all children of current use.
1808 cnt = use->outcnt();
1809 i = 0;
1810 }
1811 }
1812 else {
1813 // All of n's children have been processed, complete post-processing.
1814 if (cnt != 0 && !n->is_Con()) {
1815 assert(has_node(n), "no dead nodes");
1816 split_if_with_blocks_post(n);
1817 }
1818 if (must_throttle_split_if()) {
1819 nstack.clear();
1820 }
1821 if (nstack.is_empty()) {
1822 // Finished all nodes on stack.
1823 break;
1824 }
1825 // Get saved parent node and next use's index. Visit the rest of uses.
1826 n = nstack.node();
1827 cnt = n->outcnt();
1828 i = nstack.index();
1829 nstack.pop();
1830 }
1831 }
1832 }
1833
1834
1835 //=============================================================================
1836 //
1837 // C L O N E A L O O P B O D Y
1838 //
1839
1840 //------------------------------clone_iff--------------------------------------
1841 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1842 // "Nearly" because all Nodes have been cloned from the original in the loop,
1843 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1844 // through the Phi recursively, and return a Bool.
1845 Node* PhaseIdealLoop::clone_iff(PhiNode* phi) {
1846
1847 // Convert this Phi into a Phi merging Bools
1848 uint i;
1849 for (i = 1; i < phi->req(); i++) {
1850 Node *b = phi->in(i);
1851 if (b->is_Phi()) {
1852 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi()));
1853 } else {
1854 assert(b->is_Bool() || b->Opcode() == Op_Opaque4, "");
1855 }
1856 }
1857
1858 Node* n = phi->in(1);
1859 Node* sample_opaque = NULL;
1860 Node *sample_bool = NULL;
1861 if (n->Opcode() == Op_Opaque4) {
1862 sample_opaque = n;
1863 sample_bool = n->in(1);
1864 assert(sample_bool->is_Bool(), "wrong type");
1865 } else {
1866 sample_bool = n;
1867 }
1868 Node *sample_cmp = sample_bool->in(1);
1869
1870 // Make Phis to merge the Cmp's inputs.
1871 PhiNode *phi1 = new PhiNode(phi->in(0), Type::TOP);
1872 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP);
1873 for (i = 1; i < phi->req(); i++) {
1874 Node *n1 = sample_opaque == NULL ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1);
1875 Node *n2 = sample_opaque == NULL ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2);
1876 phi1->set_req(i, n1);
1877 phi2->set_req(i, n2);
1878 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
1879 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
1880 }
1881 // See if these Phis have been made before.
1882 // Register with optimizer
1883 Node *hit1 = _igvn.hash_find_insert(phi1);
1884 if (hit1) { // Hit, toss just made Phi
1885 _igvn.remove_dead_node(phi1); // Remove new phi
1886 assert(hit1->is_Phi(), "" );
1887 phi1 = (PhiNode*)hit1; // Use existing phi
1888 } else { // Miss
1889 _igvn.register_new_node_with_optimizer(phi1);
1890 }
1891 Node *hit2 = _igvn.hash_find_insert(phi2);
1892 if (hit2) { // Hit, toss just made Phi
1893 _igvn.remove_dead_node(phi2); // Remove new phi
1894 assert(hit2->is_Phi(), "" );
1895 phi2 = (PhiNode*)hit2; // Use existing phi
1896 } else { // Miss
1897 _igvn.register_new_node_with_optimizer(phi2);
1898 }
1899 // Register Phis with loop/block info
1900 set_ctrl(phi1, phi->in(0));
1901 set_ctrl(phi2, phi->in(0));
1902 // Make a new Cmp
1903 Node *cmp = sample_cmp->clone();
1904 cmp->set_req(1, phi1);
1905 cmp->set_req(2, phi2);
1906 _igvn.register_new_node_with_optimizer(cmp);
1907 set_ctrl(cmp, phi->in(0));
1908
1909 // Make a new Bool
1910 Node *b = sample_bool->clone();
1911 b->set_req(1,cmp);
1912 _igvn.register_new_node_with_optimizer(b);
1913 set_ctrl(b, phi->in(0));
1914
1915 if (sample_opaque != NULL) {
1916 Node* opaque = sample_opaque->clone();
1917 opaque->set_req(1, b);
1918 _igvn.register_new_node_with_optimizer(opaque);
1919 set_ctrl(opaque, phi->in(0));
1920 return opaque;
1921 }
1922
1923 assert(b->is_Bool(), "");
1924 return b;
1925 }
1926
1927 //------------------------------clone_bool-------------------------------------
1928 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1929 // "Nearly" because all Nodes have been cloned from the original in the loop,
1930 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1931 // through the Phi recursively, and return a Bool.
1932 CmpNode*PhaseIdealLoop::clone_bool(PhiNode* phi) {
1933 uint i;
1934 // Convert this Phi into a Phi merging Bools
1935 for( i = 1; i < phi->req(); i++ ) {
1936 Node *b = phi->in(i);
1937 if( b->is_Phi() ) {
1938 _igvn.replace_input_of(phi, i, clone_bool(b->as_Phi()));
1939 } else {
1940 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
1941 }
1942 }
1943
1944 Node *sample_cmp = phi->in(1);
1945
1946 // Make Phis to merge the Cmp's inputs.
1947 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP );
1948 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP );
1949 for( uint j = 1; j < phi->req(); j++ ) {
1950 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
1951 Node *n1, *n2;
1952 if( cmp_top->is_Cmp() ) {
1953 n1 = cmp_top->in(1);
1954 n2 = cmp_top->in(2);
1955 } else {
1956 n1 = n2 = cmp_top;
1957 }
1958 phi1->set_req( j, n1 );
1959 phi2->set_req( j, n2 );
1960 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
1961 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
1962 }
1963
1964 // See if these Phis have been made before.
1965 // Register with optimizer
1966 Node *hit1 = _igvn.hash_find_insert(phi1);
1967 if( hit1 ) { // Hit, toss just made Phi
1968 _igvn.remove_dead_node(phi1); // Remove new phi
1969 assert( hit1->is_Phi(), "" );
1970 phi1 = (PhiNode*)hit1; // Use existing phi
1971 } else { // Miss
1972 _igvn.register_new_node_with_optimizer(phi1);
1973 }
1974 Node *hit2 = _igvn.hash_find_insert(phi2);
1975 if( hit2 ) { // Hit, toss just made Phi
1976 _igvn.remove_dead_node(phi2); // Remove new phi
1977 assert( hit2->is_Phi(), "" );
1978 phi2 = (PhiNode*)hit2; // Use existing phi
1979 } else { // Miss
1980 _igvn.register_new_node_with_optimizer(phi2);
1981 }
1982 // Register Phis with loop/block info
1983 set_ctrl(phi1, phi->in(0));
1984 set_ctrl(phi2, phi->in(0));
1985 // Make a new Cmp
1986 Node *cmp = sample_cmp->clone();
1987 cmp->set_req( 1, phi1 );
1988 cmp->set_req( 2, phi2 );
1989 _igvn.register_new_node_with_optimizer(cmp);
1990 set_ctrl(cmp, phi->in(0));
1991
1992 assert( cmp->is_Cmp(), "" );
1993 return (CmpNode*)cmp;
1994 }
1995
1996 //------------------------------sink_use---------------------------------------
1997 // If 'use' was in the loop-exit block, it now needs to be sunk
1998 // below the post-loop merge point.
1999 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) {
2000 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) {
2001 set_ctrl(use, post_loop);
2002 for (DUIterator j = use->outs(); use->has_out(j); j++)
2003 sink_use(use->out(j), post_loop);
2004 }
2005 }
2006
2007 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new,
2008 IdealLoopTree* loop, IdealLoopTree* outer_loop,
2009 Node_List*& split_if_set, Node_List*& split_bool_set,
2010 Node_List*& split_cex_set, Node_List& worklist,
2011 uint new_counter, CloneLoopMode mode) {
2012 Node* nnn = old_new[old->_idx];
2013 // Copy uses to a worklist, so I can munge the def-use info
2014 // with impunity.
2015 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
2016 worklist.push(old->fast_out(j));
2017
2018 while( worklist.size() ) {
2019 Node *use = worklist.pop();
2020 if (!has_node(use)) continue; // Ignore dead nodes
2021 if (use->in(0) == C->top()) continue;
2022 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
2023 // Check for data-use outside of loop - at least one of OLD or USE
2024 // must not be a CFG node.
2025 #ifdef ASSERT
2026 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == NULL) {
2027 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint();
2028 assert(mode != IgnoreStripMined, "incorrect cloning mode");
2029 assert((mode == ControlAroundStripMined && use == sfpt) || !use->is_reachable_from_root(), "missed a node");
2030 }
2031 #endif
2032 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) {
2033
2034 // If the Data use is an IF, that means we have an IF outside of the
2035 // loop that is switching on a condition that is set inside of the
2036 // loop. Happens if people set a loop-exit flag; then test the flag
2037 // in the loop to break the loop, then test is again outside of the
2038 // loop to determine which way the loop exited.
2039 // Loop predicate If node connects to Bool node through Opaque1 node.
2040 if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use) || use->Opcode() == Op_Opaque4) {
2041 // Since this code is highly unlikely, we lazily build the worklist
2042 // of such Nodes to go split.
2043 if (!split_if_set) {
2044 split_if_set = new Node_List();
2045 }
2046 split_if_set->push(use);
2047 }
2048 if (use->is_Bool()) {
2049 if (!split_bool_set) {
2050 split_bool_set = new Node_List();
2051 }
2052 split_bool_set->push(use);
2053 }
2054 if (use->Opcode() == Op_CreateEx) {
2055 if (!split_cex_set) {
2056 split_cex_set = new Node_List();
2057 }
2058 split_cex_set->push(use);
2059 }
2060
2061
2062 // Get "block" use is in
2063 uint idx = 0;
2064 while( use->in(idx) != old ) idx++;
2065 Node *prev = use->is_CFG() ? use : get_ctrl(use);
2066 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" );
2067 Node *cfg = prev->_idx >= new_counter
2068 ? prev->in(2)
2069 : idom(prev);
2070 if( use->is_Phi() ) // Phi use is in prior block
2071 cfg = prev->in(idx); // NOT in block of Phi itself
2072 if (cfg->is_top()) { // Use is dead?
2073 _igvn.replace_input_of(use, idx, C->top());
2074 continue;
2075 }
2076
2077 // If use is referenced through control edge... (idx == 0)
2078 if (mode == IgnoreStripMined && idx == 0) {
2079 LoopNode *head = loop->_head->as_Loop();
2080 if (head->is_strip_mined() && is_dominator(head->outer_loop_exit(), prev)) {
2081 // That node is outside the inner loop, leave it outside the
2082 // outer loop as well to not confuse verification code.
2083 assert(!loop->_parent->is_member(use_loop), "should be out of the outer loop");
2084 _igvn.replace_input_of(use, 0, head->outer_loop_exit());
2085 continue;
2086 }
2087 }
2088
2089 while(!outer_loop->is_member(get_loop(cfg))) {
2090 prev = cfg;
2091 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg);
2092 }
2093 // If the use occurs after merging several exits from the loop, then
2094 // old value must have dominated all those exits. Since the same old
2095 // value was used on all those exits we did not need a Phi at this
2096 // merge point. NOW we do need a Phi here. Each loop exit value
2097 // is now merged with the peeled body exit; each exit gets its own
2098 // private Phi and those Phis need to be merged here.
2099 Node *phi;
2100 if( prev->is_Region() ) {
2101 if( idx == 0 ) { // Updating control edge?
2102 phi = prev; // Just use existing control
2103 } else { // Else need a new Phi
2104 phi = PhiNode::make( prev, old );
2105 // Now recursively fix up the new uses of old!
2106 for( uint i = 1; i < prev->req(); i++ ) {
2107 worklist.push(phi); // Onto worklist once for each 'old' input
2108 }
2109 }
2110 } else {
2111 // Get new RegionNode merging old and new loop exits
2112 prev = old_new[prev->_idx];
2113 assert( prev, "just made this in step 7" );
2114 if( idx == 0) { // Updating control edge?
2115 phi = prev; // Just use existing control
2116 } else { // Else need a new Phi
2117 // Make a new Phi merging data values properly
2118 phi = PhiNode::make( prev, old );
2119 phi->set_req( 1, nnn );
2120 }
2121 }
2122 // If inserting a new Phi, check for prior hits
2123 if( idx != 0 ) {
2124 Node *hit = _igvn.hash_find_insert(phi);
2125 if( hit == NULL ) {
2126 _igvn.register_new_node_with_optimizer(phi); // Register new phi
2127 } else { // or
2128 // Remove the new phi from the graph and use the hit
2129 _igvn.remove_dead_node(phi);
2130 phi = hit; // Use existing phi
2131 }
2132 set_ctrl(phi, prev);
2133 }
2134 // Make 'use' use the Phi instead of the old loop body exit value
2135 _igvn.replace_input_of(use, idx, phi);
2136 if( use->_idx >= new_counter ) { // If updating new phis
2137 // Not needed for correctness, but prevents a weak assert
2138 // in AddPNode from tripping (when we end up with different
2139 // base & derived Phis that will become the same after
2140 // IGVN does CSE).
2141 Node *hit = _igvn.hash_find_insert(use);
2142 if( hit ) // Go ahead and re-hash for hits.
2143 _igvn.replace_node( use, hit );
2144 }
2145
2146 // If 'use' was in the loop-exit block, it now needs to be sunk
2147 // below the post-loop merge point.
2148 sink_use( use, prev );
2149 }
2150 }
2151 }
2152
2153 static void clone_outer_loop_helper(Node* n, const IdealLoopTree *loop, const IdealLoopTree* outer_loop,
2154 const Node_List &old_new, Unique_Node_List& wq, PhaseIdealLoop* phase,
2155 bool check_old_new) {
2156 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2157 Node* u = n->fast_out(j);
2158 assert(check_old_new || old_new[u->_idx] == NULL, "shouldn't have been cloned");
2159 if (!u->is_CFG() && (!check_old_new || old_new[u->_idx] == NULL)) {
2160 Node* c = phase->get_ctrl(u);
2161 IdealLoopTree* u_loop = phase->get_loop(c);
2162 assert(!loop->is_member(u_loop), "can be in outer loop or out of both loops only");
2163 if (outer_loop->is_member(u_loop)) {
2164 wq.push(u);
2165 } else {
2166 // nodes pinned with control in the outer loop but not referenced from the safepoint must be moved out of
2167 // the outer loop too
2168 Node* u_c = u->in(0);
2169 if (u_c != NULL) {
2170 IdealLoopTree* u_c_loop = phase->get_loop(u_c);
2171 if (outer_loop->is_member(u_c_loop) && !loop->is_member(u_c_loop)) {
2172 wq.push(u);
2173 }
2174 }
2175 }
2176 }
2177 }
2178 }
2179
2180 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
2181 IdealLoopTree* outer_loop, int dd, Node_List &old_new,
2182 Node_List& extra_data_nodes) {
2183 if (head->is_strip_mined() && mode != IgnoreStripMined) {
2184 CountedLoopNode* cl = head->as_CountedLoop();
2185 Node* l = cl->outer_loop();
2186 Node* tail = cl->outer_loop_tail();
2187 IfNode* le = cl->outer_loop_end();
2188 Node* sfpt = cl->outer_safepoint();
2189 CountedLoopEndNode* cle = cl->loopexit();
2190 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop();
2191 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null();
2192 Node* cle_out = cle->proj_out(false);
2193
2194 Node* new_sfpt = NULL;
2195 Node* new_cle_out = cle_out->clone();
2196 old_new.map(cle_out->_idx, new_cle_out);
2197 if (mode == CloneIncludesStripMined) {
2198 // clone outer loop body
2199 Node* new_l = l->clone();
2200 Node* new_tail = tail->clone();
2201 IfNode* new_le = le->clone()->as_If();
2202 new_sfpt = sfpt->clone();
2203
2204 set_loop(new_l, outer_loop->_parent);
2205 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd);
2206 set_loop(new_cle_out, outer_loop->_parent);
2207 set_idom(new_cle_out, new_cle, dd);
2208 set_loop(new_sfpt, outer_loop->_parent);
2209 set_idom(new_sfpt, new_cle_out, dd);
2210 set_loop(new_le, outer_loop->_parent);
2211 set_idom(new_le, new_sfpt, dd);
2212 set_loop(new_tail, outer_loop->_parent);
2213 set_idom(new_tail, new_le, dd);
2214 set_idom(new_cl, new_l, dd);
2215
2216 old_new.map(l->_idx, new_l);
2217 old_new.map(tail->_idx, new_tail);
2218 old_new.map(le->_idx, new_le);
2219 old_new.map(sfpt->_idx, new_sfpt);
2220
2221 new_l->set_req(LoopNode::LoopBackControl, new_tail);
2222 new_l->set_req(0, new_l);
2223 new_tail->set_req(0, new_le);
2224 new_le->set_req(0, new_sfpt);
2225 new_sfpt->set_req(0, new_cle_out);
2226 new_cle_out->set_req(0, new_cle);
2227 new_cl->set_req(LoopNode::EntryControl, new_l);
2228
2229 _igvn.register_new_node_with_optimizer(new_l);
2230 _igvn.register_new_node_with_optimizer(new_tail);
2231 _igvn.register_new_node_with_optimizer(new_le);
2232 } else {
2233 Node *newhead = old_new[loop->_head->_idx];
2234 newhead->as_Loop()->clear_strip_mined();
2235 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl));
2236 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
2237 }
2238 // Look at data node that were assigned a control in the outer
2239 // loop: they are kept in the outer loop by the safepoint so start
2240 // from the safepoint node's inputs.
2241 IdealLoopTree* outer_loop = get_loop(l);
2242 Node_Stack stack(2);
2243 stack.push(sfpt, 1);
2244 uint new_counter = C->unique();
2245 while (stack.size() > 0) {
2246 Node* n = stack.node();
2247 uint i = stack.index();
2248 while (i < n->req() &&
2249 (n->in(i) == NULL ||
2250 !has_ctrl(n->in(i)) ||
2251 get_loop(get_ctrl(n->in(i))) != outer_loop ||
2252 (old_new[n->in(i)->_idx] != NULL && old_new[n->in(i)->_idx]->_idx >= new_counter))) {
2253 i++;
2254 }
2255 if (i < n->req()) {
2256 stack.set_index(i+1);
2257 stack.push(n->in(i), 0);
2258 } else {
2259 assert(old_new[n->_idx] == NULL || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet");
2260 Node* m = n == sfpt ? new_sfpt : n->clone();
2261 if (m != NULL) {
2262 for (uint i = 0; i < n->req(); i++) {
2263 if (m->in(i) != NULL && old_new[m->in(i)->_idx] != NULL) {
2264 m->set_req(i, old_new[m->in(i)->_idx]);
2265 }
2266 }
2267 } else {
2268 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?");
2269 }
2270 if (n != sfpt) {
2271 extra_data_nodes.push(n);
2272 _igvn.register_new_node_with_optimizer(m);
2273 assert(get_ctrl(n) == cle_out, "what other control?");
2274 set_ctrl(m, new_cle_out);
2275 old_new.map(n->_idx, m);
2276 }
2277 stack.pop();
2278 }
2279 }
2280 if (mode == CloneIncludesStripMined) {
2281 _igvn.register_new_node_with_optimizer(new_sfpt);
2282 _igvn.register_new_node_with_optimizer(new_cle_out);
2283 }
2284 // Some other transformation may have pessimistically assigned some
2285 // data nodes to the outer loop. Set their control so they are out
2286 // of the outer loop.
2287 ResourceMark rm;
2288 Unique_Node_List wq;
2289 for (uint i = 0; i < extra_data_nodes.size(); i++) {
2290 Node* old = extra_data_nodes.at(i);
2291 clone_outer_loop_helper(old, loop, outer_loop, old_new, wq, this, true);
2292 }
2293
2294 Node* inner_out = sfpt->in(0);
2295 if (inner_out->outcnt() > 1) {
2296 clone_outer_loop_helper(inner_out, loop, outer_loop, old_new, wq, this, true);
2297 }
2298
2299 Node* new_ctrl = cl->outer_loop_exit();
2300 assert(get_loop(new_ctrl) != outer_loop, "must be out of the loop nest");
2301 for (uint i = 0; i < wq.size(); i++) {
2302 Node* n = wq.at(i);
2303 set_ctrl(n, new_ctrl);
2304 if (n->in(0) != NULL) {
2305 _igvn.replace_input_of(n, 0, new_ctrl);
2306 }
2307 clone_outer_loop_helper(n, loop, outer_loop, old_new, wq, this, false);
2308 }
2309 } else {
2310 Node *newhead = old_new[loop->_head->_idx];
2311 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
2312 }
2313 }
2314
2315 //------------------------------clone_loop-------------------------------------
2316 //
2317 // C L O N E A L O O P B O D Y
2318 //
2319 // This is the basic building block of the loop optimizations. It clones an
2320 // entire loop body. It makes an old_new loop body mapping; with this mapping
2321 // you can find the new-loop equivalent to an old-loop node. All new-loop
2322 // nodes are exactly equal to their old-loop counterparts, all edges are the
2323 // same. All exits from the old-loop now have a RegionNode that merges the
2324 // equivalent new-loop path. This is true even for the normal "loop-exit"
2325 // condition. All uses of loop-invariant old-loop values now come from (one
2326 // or more) Phis that merge their new-loop equivalents.
2327 //
2328 // This operation leaves the graph in an illegal state: there are two valid
2329 // control edges coming from the loop pre-header to both loop bodies. I'll
2330 // definitely have to hack the graph after running this transform.
2331 //
2332 // From this building block I will further edit edges to perform loop peeling
2333 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
2334 //
2335 // Parameter side_by_size_idom:
2336 // When side_by_size_idom is NULL, the dominator tree is constructed for
2337 // the clone loop to dominate the original. Used in construction of
2338 // pre-main-post loop sequence.
2339 // When nonnull, the clone and original are side-by-side, both are
2340 // dominated by the side_by_side_idom node. Used in construction of
2341 // unswitched loops.
2342 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
2343 CloneLoopMode mode, Node* side_by_side_idom) {
2344
2345 LoopNode* head = loop->_head->as_Loop();
2346 head->verify_strip_mined(1);
2347
2348 if (C->do_vector_loop() && PrintOpto) {
2349 const char* mname = C->method()->name()->as_quoted_ascii();
2350 if (mname != NULL) {
2351 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname);
2352 }
2353 }
2354
2355 CloneMap& cm = C->clone_map();
2356 if (C->do_vector_loop()) {
2357 cm.set_clone_idx(cm.max_gen()+1);
2358 #ifndef PRODUCT
2359 if (PrintOpto) {
2360 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx());
2361 loop->dump_head();
2362 }
2363 #endif
2364 }
2365
2366 // Step 1: Clone the loop body. Make the old->new mapping.
2367 clone_loop_body(loop->_body, old_new, &cm);
2368
2369 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop;
2370
2371 // Step 2: Fix the edges in the new body. If the old input is outside the
2372 // loop use it. If the old input is INside the loop, use the corresponding
2373 // new node instead.
2374 fix_body_edges(loop->_body, loop, old_new, dd, outer_loop->_parent, false);
2375
2376 Node_List extra_data_nodes; // data nodes in the outer strip mined loop
2377 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes);
2378
2379 // Step 3: Now fix control uses. Loop varying control uses have already
2380 // been fixed up (as part of all input edges in Step 2). Loop invariant
2381 // control uses must be either an IfFalse or an IfTrue. Make a merge
2382 // point to merge the old and new IfFalse/IfTrue nodes; make the use
2383 // refer to this.
2384 Node_List worklist;
2385 uint new_counter = C->unique();
2386 fix_ctrl_uses(loop->_body, loop, old_new, mode, side_by_side_idom, &cm, worklist);
2387
2388 // Step 4: If loop-invariant use is not control, it must be dominated by a
2389 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region
2390 // there if needed. Make a Phi there merging old and new used values.
2391 Node_List *split_if_set = NULL;
2392 Node_List *split_bool_set = NULL;
2393 Node_List *split_cex_set = NULL;
2394 fix_data_uses(loop->_body, loop, mode, outer_loop, new_counter, old_new, worklist, split_if_set, split_bool_set, split_cex_set);
2395
2396 for (uint i = 0; i < extra_data_nodes.size(); i++) {
2397 Node* old = extra_data_nodes.at(i);
2398 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2399 split_bool_set, split_cex_set, worklist, new_counter,
2400 mode);
2401 }
2402
2403 // Check for IFs that need splitting/cloning. Happens if an IF outside of
2404 // the loop uses a condition set in the loop. The original IF probably
2405 // takes control from one or more OLD Regions (which in turn get from NEW
2406 // Regions). In any case, there will be a set of Phis for each merge point
2407 // from the IF up to where the original BOOL def exists the loop.
2408 finish_clone_loop(split_if_set, split_bool_set, split_cex_set);
2409
2410 }
2411
2412 void PhaseIdealLoop::finish_clone_loop(Node_List* split_if_set, Node_List* split_bool_set, Node_List* split_cex_set) {
2413 if (split_if_set) {
2414 while (split_if_set->size()) {
2415 Node *iff = split_if_set->pop();
2416 if (iff->in(1)->is_Phi()) {
2417 Node *b = clone_iff(iff->in(1)->as_Phi());
2418 _igvn.replace_input_of(iff, 1, b);
2419 }
2420 }
2421 }
2422 if (split_bool_set) {
2423 while (split_bool_set->size()) {
2424 Node *b = split_bool_set->pop();
2425 Node *phi = b->in(1);
2426 assert(phi->is_Phi(), "");
2427 CmpNode *cmp = clone_bool((PhiNode*) phi);
2428 _igvn.replace_input_of(b, 1, cmp);
2429 }
2430 }
2431 if (split_cex_set) {
2432 while (split_cex_set->size()) {
2433 Node *b = split_cex_set->pop();
2434 assert(b->in(0)->is_Region(), "");
2435 assert(b->in(1)->is_Phi(), "");
2436 assert(b->in(0)->in(0) == b->in(1)->in(0), "");
2437 split_up(b, b->in(0), NULL);
2438 }
2439 }
2440 }
2441
2442 void PhaseIdealLoop::fix_data_uses(Node_List& body, IdealLoopTree* loop, CloneLoopMode mode, IdealLoopTree* outer_loop,
2443 uint new_counter, Node_List &old_new, Node_List &worklist, Node_List*& split_if_set,
2444 Node_List*& split_bool_set, Node_List*& split_cex_set) {
2445 for(uint i = 0; i < body.size(); i++ ) {
2446 Node* old = body.at(i);
2447 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2448 split_bool_set, split_cex_set, worklist, new_counter,
2449 mode);
2450 }
2451 }
2452
2453 void PhaseIdealLoop::fix_ctrl_uses(const Node_List& body, const IdealLoopTree* loop, Node_List &old_new, CloneLoopMode mode,
2454 Node* side_by_side_idom, CloneMap* cm, Node_List &worklist) {
2455 LoopNode* head = loop->_head->as_Loop();
2456 for(uint i = 0; i < body.size(); i++ ) {
2457 Node* old = body.at(i);
2458 if( !old->is_CFG() ) continue;
2459
2460 // Copy uses to a worklist, so I can munge the def-use info
2461 // with impunity.
2462 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) {
2463 worklist.push(old->fast_out(j));
2464 }
2465
2466 while (worklist.size()) { // Visit all uses
2467 Node *use = worklist.pop();
2468 if (!has_node(use)) continue; // Ignore dead nodes
2469 IdealLoopTree *use_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use );
2470 if (!loop->is_member(use_loop) && use->is_CFG()) {
2471 // Both OLD and USE are CFG nodes here.
2472 assert(use->is_Proj(), "" );
2473 Node* nnn = old_new[old->_idx];
2474
2475 Node* newuse = NULL;
2476 if (head->is_strip_mined() && mode != IgnoreStripMined) {
2477 CountedLoopNode* cl = head->as_CountedLoop();
2478 CountedLoopEndNode* cle = cl->loopexit();
2479 Node* cle_out = cle->proj_out_or_null(false);
2480 if (use == cle_out) {
2481 IfNode* le = cl->outer_loop_end();
2482 use = le->proj_out(false);
2483 use_loop = get_loop(use);
2484 if (mode == CloneIncludesStripMined) {
2485 nnn = old_new[le->_idx];
2486 } else {
2487 newuse = old_new[cle_out->_idx];
2488 }
2489 }
2490 }
2491 if (newuse == NULL) {
2492 newuse = use->clone();
2493 }
2494
2495 // Clone the loop exit control projection
2496 if (C->do_vector_loop() && cm != NULL) {
2497 cm->verify_insert_and_clone(use, newuse, cm->clone_idx());
2498 }
2499 newuse->set_req(0,nnn);
2500 _igvn.register_new_node_with_optimizer(newuse);
2501 set_loop(newuse, use_loop);
2502 set_idom(newuse, nnn, dom_depth(nnn) + 1 );
2503
2504 // We need a Region to merge the exit from the peeled body and the
2505 // exit from the old loop body.
2506 RegionNode *r = new RegionNode(3);
2507 // Map the old use to the new merge point
2508 old_new.map( use->_idx, r );
2509 uint dd_r = MIN2(dom_depth(newuse), dom_depth(use));
2510 assert(dd_r >= dom_depth(dom_lca(newuse, use)), "" );
2511
2512 // The original user of 'use' uses 'r' instead.
2513 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
2514 Node* useuse = use->last_out(l);
2515 _igvn.rehash_node_delayed(useuse);
2516 uint uses_found = 0;
2517 if (useuse->in(0) == use) {
2518 useuse->set_req(0, r);
2519 uses_found++;
2520 if (useuse->is_CFG()) {
2521 // This is not a dom_depth > dd_r because when new
2522 // control flow is constructed by a loop opt, a node and
2523 // its dominator can end up at the same dom_depth
2524 assert(dom_depth(useuse) >= dd_r, "");
2525 set_idom(useuse, r, dom_depth(useuse));
2526 }
2527 }
2528 for (uint k = 1; k < useuse->req(); k++) {
2529 if( useuse->in(k) == use ) {
2530 useuse->set_req(k, r);
2531 uses_found++;
2532 if (useuse->is_Loop() && k == LoopNode::EntryControl) {
2533 // This is not a dom_depth > dd_r because when new
2534 // control flow is constructed by a loop opt, a node
2535 // and its dominator can end up at the same dom_depth
2536 assert(dom_depth(useuse) >= dd_r , "");
2537 set_idom(useuse, r, dom_depth(useuse));
2538 }
2539 }
2540 }
2541 l -= uses_found; // we deleted 1 or more copies of this edge
2542 }
2543
2544 // Now finish up 'r'
2545 r->set_req(1, newuse);
2546 r->set_req(2, use);
2547 _igvn.register_new_node_with_optimizer(r);
2548 set_loop(r, use_loop);
2549 set_idom(r, (side_by_side_idom == NULL) ? newuse->in(0) : side_by_side_idom, dd_r);
2550 } // End of if a loop-exit test
2551 }
2552 }
2553 }
2554
2555 void PhaseIdealLoop::fix_body_edges(const Node_List &body, IdealLoopTree* loop, const Node_List &old_new, int dd,
2556 IdealLoopTree* parent, bool partial) {
2557 for(uint i = 0; i < body.size(); i++ ) {
2558 Node *old = body.at(i);
2559 Node *nnn = old_new[old->_idx];
2560 // Fix CFG/Loop controlling the new node
2561 if (has_ctrl(old)) {
2562 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
2563 } else {
2564 set_loop(nnn, parent);
2565 if (old->outcnt() > 0) {
2566 Node* dom = idom(old);
2567 if (old_new[dom->_idx] != NULL) {
2568 dom = old_new[dom->_idx];
2569 set_idom(nnn, dom, dd );
2570 }
2571 }
2572 }
2573 // Correct edges to the new node
2574 for (uint j = 0; j < nnn->req(); j++) {
2575 Node *n = nnn->in(j);
2576 if (n != NULL) {
2577 IdealLoopTree *old_in_loop = get_loop(has_ctrl(n) ? get_ctrl(n) : n);
2578 if (loop->is_member(old_in_loop)) {
2579 if (old_new[n->_idx] != NULL) {
2580 nnn->set_req(j, old_new[n->_idx]);
2581 } else {
2582 assert(!body.contains(n), "");
2583 assert(partial, "node not cloned");
2584 }
2585 }
2586 }
2587 }
2588 _igvn.hash_find_insert(nnn);
2589 }
2590 }
2591
2592 void PhaseIdealLoop::clone_loop_body(const Node_List& body, Node_List &old_new, CloneMap* cm) {
2593 for (uint i = 0; i < body.size(); i++) {
2594 Node* old = body.at(i);
2595 Node* nnn = old->clone();
2596 old_new.map(old->_idx, nnn);
2597 if (old->is_reduction()) {
2598 // Reduction flag is not copied by default. Copy it here when cloning the entire loop body.
2599 nnn->add_flag(Node::Flag_is_reduction);
2600 }
2601 if (C->do_vector_loop() && cm != NULL) {
2602 cm->verify_insert_and_clone(old, nnn, cm->clone_idx());
2603 }
2604 _igvn.register_new_node_with_optimizer(nnn);
2605 }
2606 }
2607
2608
2609 //---------------------- stride_of_possible_iv -------------------------------------
2610 // Looks for an iff/bool/comp with one operand of the compare
2611 // being a cycle involving an add and a phi,
2612 // with an optional truncation (left-shift followed by a right-shift)
2613 // of the add. Returns zero if not an iv.
2614 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
2615 Node* trunc1 = NULL;
2616 Node* trunc2 = NULL;
2617 const TypeInteger* ttype = NULL;
2618 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) {
2619 return 0;
2620 }
2621 BoolNode* bl = iff->in(1)->as_Bool();
2622 Node* cmp = bl->in(1);
2623 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) {
2624 return 0;
2625 }
2626 // Must have an invariant operand
2627 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) {
2628 return 0;
2629 }
2630 Node* add2 = NULL;
2631 Node* cmp1 = cmp->in(1);
2632 if (cmp1->is_Phi()) {
2633 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
2634 Node* phi = cmp1;
2635 for (uint i = 1; i < phi->req(); i++) {
2636 Node* in = phi->in(i);
2637 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in,
2638 &trunc1, &trunc2, &ttype, T_INT);
2639 if (add && add->in(1) == phi) {
2640 add2 = add->in(2);
2641 break;
2642 }
2643 }
2644 } else {
2645 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
2646 Node* addtrunc = cmp1;
2647 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc,
2648 &trunc1, &trunc2, &ttype, T_INT);
2649 if (add && add->in(1)->is_Phi()) {
2650 Node* phi = add->in(1);
2651 for (uint i = 1; i < phi->req(); i++) {
2652 if (phi->in(i) == addtrunc) {
2653 add2 = add->in(2);
2654 break;
2655 }
2656 }
2657 }
2658 }
2659 if (add2 != NULL) {
2660 const TypeInt* add2t = _igvn.type(add2)->is_int();
2661 if (add2t->is_con()) {
2662 return add2t->get_con();
2663 }
2664 }
2665 return 0;
2666 }
2667
2668
2669 //---------------------- stay_in_loop -------------------------------------
2670 // Return the (unique) control output node that's in the loop (if it exists.)
2671 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
2672 Node* unique = NULL;
2673 if (!n) return NULL;
2674 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2675 Node* use = n->fast_out(i);
2676 if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
2677 if (unique != NULL) {
2678 return NULL;
2679 }
2680 unique = use;
2681 }
2682 }
2683 return unique;
2684 }
2685
2686 //------------------------------ register_node -------------------------------------
2687 // Utility to register node "n" with PhaseIdealLoop
2688 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) {
2689 _igvn.register_new_node_with_optimizer(n);
2690 loop->_body.push(n);
2691 if (n->is_CFG()) {
2692 set_loop(n, loop);
2693 set_idom(n, pred, ddepth);
2694 } else {
2695 set_ctrl(n, pred);
2696 }
2697 }
2698
2699 //------------------------------ proj_clone -------------------------------------
2700 // Utility to create an if-projection
2701 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
2702 ProjNode* c = p->clone()->as_Proj();
2703 c->set_req(0, iff);
2704 return c;
2705 }
2706
2707 //------------------------------ short_circuit_if -------------------------------------
2708 // Force the iff control output to be the live_proj
2709 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
2710 guarantee(live_proj != NULL, "null projection");
2711 int proj_con = live_proj->_con;
2712 assert(proj_con == 0 || proj_con == 1, "false or true projection");
2713 Node *con = _igvn.intcon(proj_con);
2714 set_ctrl(con, C->root());
2715 if (iff) {
2716 iff->set_req(1, con);
2717 }
2718 return con;
2719 }
2720
2721 //------------------------------ insert_if_before_proj -------------------------------------
2722 // Insert a new if before an if projection (* - new node)
2723 //
2724 // before
2725 // if(test)
2726 // / \
2727 // v v
2728 // other-proj proj (arg)
2729 //
2730 // after
2731 // if(test)
2732 // / \
2733 // / v
2734 // | * proj-clone
2735 // v |
2736 // other-proj v
2737 // * new_if(relop(cmp[IU](left,right)))
2738 // / \
2739 // v v
2740 // * new-proj proj
2741 // (returned)
2742 //
2743 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) {
2744 IfNode* iff = proj->in(0)->as_If();
2745 IdealLoopTree *loop = get_loop(proj);
2746 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2747 int ddepth = dom_depth(proj);
2748
2749 _igvn.rehash_node_delayed(iff);
2750 _igvn.rehash_node_delayed(proj);
2751
2752 proj->set_req(0, NULL); // temporary disconnect
2753 ProjNode* proj2 = proj_clone(proj, iff);
2754 register_node(proj2, loop, iff, ddepth);
2755
2756 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right);
2757 register_node(cmp, loop, proj2, ddepth);
2758
2759 BoolNode* bol = new BoolNode(cmp, relop);
2760 register_node(bol, loop, proj2, ddepth);
2761
2762 int opcode = iff->Opcode();
2763 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode");
2764 IfNode* new_if = (opcode == Op_If) ? new IfNode(proj2, bol, iff->_prob, iff->_fcnt):
2765 new RangeCheckNode(proj2, bol, iff->_prob, iff->_fcnt);
2766 register_node(new_if, loop, proj2, ddepth);
2767
2768 proj->set_req(0, new_if); // reattach
2769 set_idom(proj, new_if, ddepth);
2770
2771 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
2772 guarantee(new_exit != NULL, "null exit node");
2773 register_node(new_exit, get_loop(other_proj), new_if, ddepth);
2774
2775 return new_exit;
2776 }
2777
2778 //------------------------------ insert_region_before_proj -------------------------------------
2779 // Insert a region before an if projection (* - new node)
2780 //
2781 // before
2782 // if(test)
2783 // / |
2784 // v |
2785 // proj v
2786 // other-proj
2787 //
2788 // after
2789 // if(test)
2790 // / |
2791 // v |
2792 // * proj-clone v
2793 // | other-proj
2794 // v
2795 // * new-region
2796 // |
2797 // v
2798 // * dum_if
2799 // / \
2800 // v \
2801 // * dum-proj v
2802 // proj
2803 //
2804 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
2805 IfNode* iff = proj->in(0)->as_If();
2806 IdealLoopTree *loop = get_loop(proj);
2807 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2808 int ddepth = dom_depth(proj);
2809
2810 _igvn.rehash_node_delayed(iff);
2811 _igvn.rehash_node_delayed(proj);
2812
2813 proj->set_req(0, NULL); // temporary disconnect
2814 ProjNode* proj2 = proj_clone(proj, iff);
2815 register_node(proj2, loop, iff, ddepth);
2816
2817 RegionNode* reg = new RegionNode(2);
2818 reg->set_req(1, proj2);
2819 register_node(reg, loop, iff, ddepth);
2820
2821 IfNode* dum_if = new IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt);
2822 register_node(dum_if, loop, reg, ddepth);
2823
2824 proj->set_req(0, dum_if); // reattach
2825 set_idom(proj, dum_if, ddepth);
2826
2827 ProjNode* dum_proj = proj_clone(other_proj, dum_if);
2828 register_node(dum_proj, loop, dum_if, ddepth);
2829
2830 return reg;
2831 }
2832
2833 //------------------------------ insert_cmpi_loop_exit -------------------------------------
2834 // Clone a signed compare loop exit from an unsigned compare and
2835 // insert it before the unsigned cmp on the stay-in-loop path.
2836 // All new nodes inserted in the dominator tree between the original
2837 // if and it's projections. The original if test is replaced with
2838 // a constant to force the stay-in-loop path.
2839 //
2840 // This is done to make sure that the original if and it's projections
2841 // still dominate the same set of control nodes, that the ctrl() relation
2842 // from data nodes to them is preserved, and that their loop nesting is
2843 // preserved.
2844 //
2845 // before
2846 // if(i <u limit) unsigned compare loop exit
2847 // / |
2848 // v v
2849 // exit-proj stay-in-loop-proj
2850 //
2851 // after
2852 // if(stay-in-loop-const) original if
2853 // / |
2854 // / v
2855 // / if(i < limit) new signed test
2856 // / / |
2857 // / / v
2858 // / / if(i <u limit) new cloned unsigned test
2859 // / / / |
2860 // v v v |
2861 // region |
2862 // | |
2863 // dum-if |
2864 // / | |
2865 // ether | |
2866 // v v
2867 // exit-proj stay-in-loop-proj
2868 //
2869 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) {
2870 const bool Signed = true;
2871 const bool Unsigned = false;
2872
2873 BoolNode* bol = if_cmpu->in(1)->as_Bool();
2874 if (bol->_test._test != BoolTest::lt) return NULL;
2875 CmpNode* cmpu = bol->in(1)->as_Cmp();
2876 if (cmpu->Opcode() != Op_CmpU) return NULL;
2877 int stride = stride_of_possible_iv(if_cmpu);
2878 if (stride == 0) return NULL;
2879
2880 Node* lp_proj = stay_in_loop(if_cmpu, loop);
2881 guarantee(lp_proj != NULL, "null loop node");
2882
2883 ProjNode* lp_continue = lp_proj->as_Proj();
2884 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
2885 if (!lp_exit->is_IfFalse()) {
2886 // The loop exit condition is (i <u limit) ==> (i >= 0 && i < limit).
2887 // We therefore can't add a single exit condition.
2888 return NULL;
2889 }
2890 // The loop exit condition is !(i <u limit) ==> (i < 0 || i >= limit).
2891 // Split out the exit condition (i < 0) for stride < 0 or (i >= limit) for stride > 0.
2892 Node* limit = NULL;
2893 if (stride > 0) {
2894 limit = cmpu->in(2);
2895 } else {
2896 limit = _igvn.makecon(TypeInt::ZERO);
2897 set_ctrl(limit, C->root());
2898 }
2899 // Create a new region on the exit path
2900 RegionNode* reg = insert_region_before_proj(lp_exit);
2901 guarantee(reg != NULL, "null region node");
2902
2903 // Clone the if-cmpu-true-false using a signed compare
2904 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
2905 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue);
2906 reg->add_req(cmpi_exit);
2907
2908 // Clone the if-cmpu-true-false
2909 BoolTest::mask rel_u = bol->_test._test;
2910 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
2911 reg->add_req(cmpu_exit);
2912
2913 // Force original if to stay in loop.
2914 short_circuit_if(if_cmpu, lp_continue);
2915
2916 return cmpi_exit->in(0)->as_If();
2917 }
2918
2919 //------------------------------ remove_cmpi_loop_exit -------------------------------------
2920 // Remove a previously inserted signed compare loop exit.
2921 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
2922 Node* lp_proj = stay_in_loop(if_cmp, loop);
2923 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
2924 stay_in_loop(lp_proj, loop)->is_If() &&
2925 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
2926 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
2927 set_ctrl(con, C->root());
2928 if_cmp->set_req(1, con);
2929 }
2930
2931 //------------------------------ scheduled_nodelist -------------------------------------
2932 // Create a post order schedule of nodes that are in the
2933 // "member" set. The list is returned in "sched".
2934 // The first node in "sched" is the loop head, followed by
2935 // nodes which have no inputs in the "member" set, and then
2936 // followed by the nodes that have an immediate input dependence
2937 // on a node in "sched".
2938 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
2939
2940 assert(member.test(loop->_head->_idx), "loop head must be in member set");
2941 VectorSet visited;
2942 Node_Stack nstack(loop->_body.size());
2943
2944 Node* n = loop->_head; // top of stack is cached in "n"
2945 uint idx = 0;
2946 visited.set(n->_idx);
2947
2948 // Initially push all with no inputs from within member set
2949 for(uint i = 0; i < loop->_body.size(); i++ ) {
2950 Node *elt = loop->_body.at(i);
2951 if (member.test(elt->_idx)) {
2952 bool found = false;
2953 for (uint j = 0; j < elt->req(); j++) {
2954 Node* def = elt->in(j);
2955 if (def && member.test(def->_idx) && def != elt) {
2956 found = true;
2957 break;
2958 }
2959 }
2960 if (!found && elt != loop->_head) {
2961 nstack.push(n, idx);
2962 n = elt;
2963 assert(!visited.test(n->_idx), "not seen yet");
2964 visited.set(n->_idx);
2965 }
2966 }
2967 }
2968
2969 // traverse out's that are in the member set
2970 while (true) {
2971 if (idx < n->outcnt()) {
2972 Node* use = n->raw_out(idx);
2973 idx++;
2974 if (!visited.test_set(use->_idx)) {
2975 if (member.test(use->_idx)) {
2976 nstack.push(n, idx);
2977 n = use;
2978 idx = 0;
2979 }
2980 }
2981 } else {
2982 // All outputs processed
2983 sched.push(n);
2984 if (nstack.is_empty()) break;
2985 n = nstack.node();
2986 idx = nstack.index();
2987 nstack.pop();
2988 }
2989 }
2990 }
2991
2992
2993 //------------------------------ has_use_in_set -------------------------------------
2994 // Has a use in the vector set
2995 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
2996 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2997 Node* use = n->fast_out(j);
2998 if (vset.test(use->_idx)) {
2999 return true;
3000 }
3001 }
3002 return false;
3003 }
3004
3005
3006 //------------------------------ has_use_internal_to_set -------------------------------------
3007 // Has use internal to the vector set (ie. not in a phi at the loop head)
3008 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
3009 Node* head = loop->_head;
3010 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3011 Node* use = n->fast_out(j);
3012 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
3013 return true;
3014 }
3015 }
3016 return false;
3017 }
3018
3019
3020 //------------------------------ clone_for_use_outside_loop -------------------------------------
3021 // clone "n" for uses that are outside of loop
3022 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
3023 int cloned = 0;
3024 assert(worklist.size() == 0, "should be empty");
3025 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3026 Node* use = n->fast_out(j);
3027 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
3028 worklist.push(use);
3029 }
3030 }
3031
3032 if (C->check_node_count(worklist.size() + NodeLimitFudgeFactor,
3033 "Too many clones required in clone_for_use_outside_loop in partial peeling")) {
3034 return -1;
3035 }
3036
3037 while( worklist.size() ) {
3038 Node *use = worklist.pop();
3039 if (!has_node(use) || use->in(0) == C->top()) continue;
3040 uint j;
3041 for (j = 0; j < use->req(); j++) {
3042 if (use->in(j) == n) break;
3043 }
3044 assert(j < use->req(), "must be there");
3045
3046 // clone "n" and insert it between the inputs of "n" and the use outside the loop
3047 Node* n_clone = n->clone();
3048 _igvn.replace_input_of(use, j, n_clone);
3049 cloned++;
3050 Node* use_c;
3051 if (!use->is_Phi()) {
3052 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
3053 } else {
3054 // Use in a phi is considered a use in the associated predecessor block
3055 use_c = use->in(0)->in(j);
3056 }
3057 set_ctrl(n_clone, use_c);
3058 assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
3059 get_loop(use_c)->_body.push(n_clone);
3060 _igvn.register_new_node_with_optimizer(n_clone);
3061 #ifndef PRODUCT
3062 if (TracePartialPeeling) {
3063 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
3064 }
3065 #endif
3066 }
3067 return cloned;
3068 }
3069
3070
3071 //------------------------------ clone_for_special_use_inside_loop -------------------------------------
3072 // clone "n" for special uses that are in the not_peeled region.
3073 // If these def-uses occur in separate blocks, the code generator
3074 // marks the method as not compilable. For example, if a "BoolNode"
3075 // is in a different basic block than the "IfNode" that uses it, then
3076 // the compilation is aborted in the code generator.
3077 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
3078 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
3079 if (n->is_Phi() || n->is_Load()) {
3080 return;
3081 }
3082 assert(worklist.size() == 0, "should be empty");
3083 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3084 Node* use = n->fast_out(j);
3085 if ( not_peel.test(use->_idx) &&
3086 (use->is_If() || use->is_CMove() || use->is_Bool()) &&
3087 use->in(1) == n) {
3088 worklist.push(use);
3089 }
3090 }
3091 if (worklist.size() > 0) {
3092 // clone "n" and insert it between inputs of "n" and the use
3093 Node* n_clone = n->clone();
3094 loop->_body.push(n_clone);
3095 _igvn.register_new_node_with_optimizer(n_clone);
3096 set_ctrl(n_clone, get_ctrl(n));
3097 sink_list.push(n_clone);
3098 not_peel.set(n_clone->_idx);
3099 #ifndef PRODUCT
3100 if (TracePartialPeeling) {
3101 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
3102 }
3103 #endif
3104 while( worklist.size() ) {
3105 Node *use = worklist.pop();
3106 _igvn.rehash_node_delayed(use);
3107 for (uint j = 1; j < use->req(); j++) {
3108 if (use->in(j) == n) {
3109 use->set_req(j, n_clone);
3110 }
3111 }
3112 }
3113 }
3114 }
3115
3116
3117 //------------------------------ insert_phi_for_loop -------------------------------------
3118 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
3119 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
3120 Node *phi = PhiNode::make(lp, back_edge_val);
3121 phi->set_req(LoopNode::EntryControl, lp_entry_val);
3122 // Use existing phi if it already exists
3123 Node *hit = _igvn.hash_find_insert(phi);
3124 if( hit == NULL ) {
3125 _igvn.register_new_node_with_optimizer(phi);
3126 set_ctrl(phi, lp);
3127 } else {
3128 // Remove the new phi from the graph and use the hit
3129 _igvn.remove_dead_node(phi);
3130 phi = hit;
3131 }
3132 _igvn.replace_input_of(use, idx, phi);
3133 }
3134
3135 #ifdef ASSERT
3136 //------------------------------ is_valid_loop_partition -------------------------------------
3137 // Validate the loop partition sets: peel and not_peel
3138 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
3139 VectorSet& not_peel ) {
3140 uint i;
3141 // Check that peel_list entries are in the peel set
3142 for (i = 0; i < peel_list.size(); i++) {
3143 if (!peel.test(peel_list.at(i)->_idx)) {
3144 return false;
3145 }
3146 }
3147 // Check at loop members are in one of peel set or not_peel set
3148 for (i = 0; i < loop->_body.size(); i++ ) {
3149 Node *def = loop->_body.at(i);
3150 uint di = def->_idx;
3151 // Check that peel set elements are in peel_list
3152 if (peel.test(di)) {
3153 if (not_peel.test(di)) {
3154 return false;
3155 }
3156 // Must be in peel_list also
3157 bool found = false;
3158 for (uint j = 0; j < peel_list.size(); j++) {
3159 if (peel_list.at(j)->_idx == di) {
3160 found = true;
3161 break;
3162 }
3163 }
3164 if (!found) {
3165 return false;
3166 }
3167 } else if (not_peel.test(di)) {
3168 if (peel.test(di)) {
3169 return false;
3170 }
3171 } else {
3172 return false;
3173 }
3174 }
3175 return true;
3176 }
3177
3178 //------------------------------ is_valid_clone_loop_exit_use -------------------------------------
3179 // Ensure a use outside of loop is of the right form
3180 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
3181 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
3182 return (use->is_Phi() &&
3183 use_c->is_Region() && use_c->req() == 3 &&
3184 (use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
3185 use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
3186 use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
3187 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
3188 }
3189
3190 //------------------------------ is_valid_clone_loop_form -------------------------------------
3191 // Ensure that all uses outside of loop are of the right form
3192 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
3193 uint orig_exit_idx, uint clone_exit_idx) {
3194 uint len = peel_list.size();
3195 for (uint i = 0; i < len; i++) {
3196 Node *def = peel_list.at(i);
3197
3198 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
3199 Node *use = def->fast_out(j);
3200 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
3201 if (!loop->is_member(get_loop(use_c))) {
3202 // use is not in the loop, check for correct structure
3203 if (use->in(0) == def) {
3204 // Okay
3205 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
3206 return false;
3207 }
3208 }
3209 }
3210 }
3211 return true;
3212 }
3213 #endif
3214
3215 //------------------------------ partial_peel -------------------------------------
3216 // Partially peel (aka loop rotation) the top portion of a loop (called
3217 // the peel section below) by cloning it and placing one copy just before
3218 // the new loop head and the other copy at the bottom of the new loop.
3219 //
3220 // before after where it came from
3221 //
3222 // stmt1 stmt1
3223 // loop: stmt2 clone
3224 // stmt2 if condA goto exitA clone
3225 // if condA goto exitA new_loop: new
3226 // stmt3 stmt3 clone
3227 // if !condB goto loop if condB goto exitB clone
3228 // exitB: stmt2 orig
3229 // stmt4 if !condA goto new_loop orig
3230 // exitA: goto exitA
3231 // exitB:
3232 // stmt4
3233 // exitA:
3234 //
3235 // Step 1: find the cut point: an exit test on probable
3236 // induction variable.
3237 // Step 2: schedule (with cloning) operations in the peel
3238 // section that can be executed after the cut into
3239 // the section that is not peeled. This may need
3240 // to clone operations into exit blocks. For
3241 // instance, a reference to A[i] in the not-peel
3242 // section and a reference to B[i] in an exit block
3243 // may cause a left-shift of i by 2 to be placed
3244 // in the peel block. This step will clone the left
3245 // shift into the exit block and sink the left shift
3246 // from the peel to the not-peel section.
3247 // Step 3: clone the loop, retarget the control, and insert
3248 // phis for values that are live across the new loop
3249 // head. This is very dependent on the graph structure
3250 // from clone_loop. It creates region nodes for
3251 // exit control and associated phi nodes for values
3252 // flow out of the loop through that exit. The region
3253 // node is dominated by the clone's control projection.
3254 // So the clone's peel section is placed before the
3255 // new loop head, and the clone's not-peel section is
3256 // forms the top part of the new loop. The original
3257 // peel section forms the tail of the new loop.
3258 // Step 4: update the dominator tree and recompute the
3259 // dominator depth.
3260 //
3261 // orig
3262 //
3263 // stmt1
3264 // |
3265 // v
3266 // loop predicate
3267 // |
3268 // v
3269 // loop<----+
3270 // | |
3271 // stmt2 |
3272 // | |
3273 // v |
3274 // ifA |
3275 // / | |
3276 // v v |
3277 // false true ^ <-- last_peel
3278 // / | |
3279 // / ===|==cut |
3280 // / stmt3 | <-- first_not_peel
3281 // / | |
3282 // | v |
3283 // v ifB |
3284 // exitA: / \ |
3285 // / \ |
3286 // v v |
3287 // false true |
3288 // / \ |
3289 // / ----+
3290 // |
3291 // v
3292 // exitB:
3293 // stmt4
3294 //
3295 //
3296 // after clone loop
3297 //
3298 // stmt1
3299 // |
3300 // v
3301 // loop predicate
3302 // / \
3303 // clone / \ orig
3304 // / \
3305 // / \
3306 // v v
3307 // +---->loop loop<----+
3308 // | | | |
3309 // | stmt2 stmt2 |
3310 // | | | |
3311 // | v v |
3312 // | ifA ifA |
3313 // | | \ / | |
3314 // | v v v v |
3315 // ^ true false false true ^ <-- last_peel
3316 // | | ^ \ / | |
3317 // | cut==|== \ \ / ===|==cut |
3318 // | stmt3 \ \ / stmt3 | <-- first_not_peel
3319 // | | dom | | | |
3320 // | v \ 1v v2 v |
3321 // | ifB regionA ifB |
3322 // | / \ | / \ |
3323 // | / \ v / \ |
3324 // | v v exitA: v v |
3325 // | true false false true |
3326 // | / ^ \ / \ |
3327 // +---- \ \ / ----+
3328 // dom \ /
3329 // \ 1v v2
3330 // regionB
3331 // |
3332 // v
3333 // exitB:
3334 // stmt4
3335 //
3336 //
3337 // after partial peel
3338 //
3339 // stmt1
3340 // |
3341 // v
3342 // loop predicate
3343 // /
3344 // clone / orig
3345 // / TOP
3346 // / \
3347 // v v
3348 // TOP->loop loop----+
3349 // | | |
3350 // stmt2 stmt2 |
3351 // | | |
3352 // v v |
3353 // ifA ifA |
3354 // | \ / | |
3355 // v v v v |
3356 // true false false true | <-- last_peel
3357 // | ^ \ / +------|---+
3358 // +->newloop \ \ / === ==cut | |
3359 // | stmt3 \ \ / TOP | |
3360 // | | dom | | stmt3 | | <-- first_not_peel
3361 // | v \ 1v v2 v | |
3362 // | ifB regionA ifB ^ v
3363 // | / \ | / \ | |
3364 // | / \ v / \ | |
3365 // | v v exitA: v v | |
3366 // | true false false true | |
3367 // | / ^ \ / \ | |
3368 // | | \ \ / v | |
3369 // | | dom \ / TOP | |
3370 // | | \ 1v v2 | |
3371 // ^ v regionB | |
3372 // | | | | |
3373 // | | v ^ v
3374 // | | exitB: | |
3375 // | | stmt4 | |
3376 // | +------------>-----------------+ |
3377 // | |
3378 // +-----------------<---------------------+
3379 //
3380 //
3381 // final graph
3382 //
3383 // stmt1
3384 // |
3385 // v
3386 // loop predicate
3387 // |
3388 // v
3389 // stmt2 clone
3390 // |
3391 // v
3392 // ........> ifA clone
3393 // : / |
3394 // dom / |
3395 // : v v
3396 // : false true
3397 // : | |
3398 // : | v
3399 // : | newloop<-----+
3400 // : | | |
3401 // : | stmt3 clone |
3402 // : | | |
3403 // : | v |
3404 // : | ifB |
3405 // : | / \ |
3406 // : | v v |
3407 // : | false true |
3408 // : | | | |
3409 // : | v stmt2 |
3410 // : | exitB: | |
3411 // : | stmt4 v |
3412 // : | ifA orig |
3413 // : | / \ |
3414 // : | / \ |
3415 // : | v v |
3416 // : | false true |
3417 // : | / \ |
3418 // : v v -----+
3419 // RegionA
3420 // |
3421 // v
3422 // exitA
3423 //
3424 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
3425
3426 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only");
3427 if (!loop->_head->is_Loop()) {
3428 return false;
3429 }
3430 LoopNode *head = loop->_head->as_Loop();
3431
3432 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
3433 return false;
3434 }
3435
3436 // Check for complex exit control
3437 for (uint ii = 0; ii < loop->_body.size(); ii++) {
3438 Node *n = loop->_body.at(ii);
3439 int opc = n->Opcode();
3440 if (n->is_Call() ||
3441 opc == Op_Catch ||
3442 opc == Op_CatchProj ||
3443 opc == Op_Jump ||
3444 opc == Op_JumpProj) {
3445 #ifndef PRODUCT
3446 if (TracePartialPeeling) {
3447 tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
3448 }
3449 #endif
3450 return false;
3451 }
3452 }
3453
3454 int dd = dom_depth(head);
3455
3456 // Step 1: find cut point
3457
3458 // Walk up dominators to loop head looking for first loop exit
3459 // which is executed on every path thru loop.
3460 IfNode *peel_if = NULL;
3461 IfNode *peel_if_cmpu = NULL;
3462
3463 Node *iff = loop->tail();
3464 while (iff != head) {
3465 if (iff->is_If()) {
3466 Node *ctrl = get_ctrl(iff->in(1));
3467 if (ctrl->is_top()) return false; // Dead test on live IF.
3468 // If loop-varying exit-test, check for induction variable
3469 if (loop->is_member(get_loop(ctrl)) &&
3470 loop->is_loop_exit(iff) &&
3471 is_possible_iv_test(iff)) {
3472 Node* cmp = iff->in(1)->in(1);
3473 if (cmp->Opcode() == Op_CmpI) {
3474 peel_if = iff->as_If();
3475 } else {
3476 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
3477 peel_if_cmpu = iff->as_If();
3478 }
3479 }
3480 }
3481 iff = idom(iff);
3482 }
3483
3484 // Prefer signed compare over unsigned compare.
3485 IfNode* new_peel_if = NULL;
3486 if (peel_if == NULL) {
3487 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) {
3488 return false; // No peel point found
3489 }
3490 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
3491 if (new_peel_if == NULL) {
3492 return false; // No peel point found
3493 }
3494 peel_if = new_peel_if;
3495 }
3496 Node* last_peel = stay_in_loop(peel_if, loop);
3497 Node* first_not_peeled = stay_in_loop(last_peel, loop);
3498 if (first_not_peeled == NULL || first_not_peeled == head) {
3499 return false;
3500 }
3501
3502 #ifndef PRODUCT
3503 if (TraceLoopOpts) {
3504 tty->print("PartialPeel ");
3505 loop->dump_head();
3506 }
3507
3508 if (TracePartialPeeling) {
3509 tty->print_cr("before partial peel one iteration");
3510 Node_List wl;
3511 Node* t = head->in(2);
3512 while (true) {
3513 wl.push(t);
3514 if (t == head) break;
3515 t = idom(t);
3516 }
3517 while (wl.size() > 0) {
3518 Node* tt = wl.pop();
3519 tt->dump();
3520 if (tt == last_peel) tty->print_cr("-- cut --");
3521 }
3522 }
3523 #endif
3524 VectorSet peel;
3525 VectorSet not_peel;
3526 Node_List peel_list;
3527 Node_List worklist;
3528 Node_List sink_list;
3529
3530 uint estimate = loop->est_loop_clone_sz(1);
3531 if (exceeding_node_budget(estimate)) {
3532 return false;
3533 }
3534
3535 // Set of cfg nodes to peel are those that are executable from
3536 // the head through last_peel.
3537 assert(worklist.size() == 0, "should be empty");
3538 worklist.push(head);
3539 peel.set(head->_idx);
3540 while (worklist.size() > 0) {
3541 Node *n = worklist.pop();
3542 if (n != last_peel) {
3543 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3544 Node* use = n->fast_out(j);
3545 if (use->is_CFG() &&
3546 loop->is_member(get_loop(use)) &&
3547 !peel.test_set(use->_idx)) {
3548 worklist.push(use);
3549 }
3550 }
3551 }
3552 }
3553
3554 // Set of non-cfg nodes to peel are those that are control
3555 // dependent on the cfg nodes.
3556 for (uint i = 0; i < loop->_body.size(); i++) {
3557 Node *n = loop->_body.at(i);
3558 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
3559 if (peel.test(n_c->_idx)) {
3560 peel.set(n->_idx);
3561 } else {
3562 not_peel.set(n->_idx);
3563 }
3564 }
3565
3566 // Step 2: move operations from the peeled section down into the
3567 // not-peeled section
3568
3569 // Get a post order schedule of nodes in the peel region
3570 // Result in right-most operand.
3571 scheduled_nodelist(loop, peel, peel_list);
3572
3573 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
3574
3575 // For future check for too many new phis
3576 uint old_phi_cnt = 0;
3577 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
3578 Node* use = head->fast_out(j);
3579 if (use->is_Phi()) old_phi_cnt++;
3580 }
3581
3582 #ifndef PRODUCT
3583 if (TracePartialPeeling) {
3584 tty->print_cr("\npeeled list");
3585 }
3586 #endif
3587
3588 // Evacuate nodes in peel region into the not_peeled region if possible
3589 bool too_many_clones = false;
3590 uint new_phi_cnt = 0;
3591 uint cloned_for_outside_use = 0;
3592 for (uint i = 0; i < peel_list.size();) {
3593 Node* n = peel_list.at(i);
3594 #ifndef PRODUCT
3595 if (TracePartialPeeling) n->dump();
3596 #endif
3597 bool incr = true;
3598 if (!n->is_CFG()) {
3599 if (has_use_in_set(n, not_peel)) {
3600 // If not used internal to the peeled region,
3601 // move "n" from peeled to not_peeled region.
3602 if (!has_use_internal_to_set(n, peel, loop)) {
3603 // if not pinned and not a load (which maybe anti-dependent on a store)
3604 // and not a CMove (Matcher expects only bool->cmove).
3605 if (n->in(0) == NULL && !n->is_Load() && !n->is_CMove()) {
3606 int new_clones = clone_for_use_outside_loop(loop, n, worklist);
3607 if (new_clones == -1) {
3608 too_many_clones = true;
3609 break;
3610 }
3611 cloned_for_outside_use += new_clones;
3612 sink_list.push(n);
3613 peel.remove(n->_idx);
3614 not_peel.set(n->_idx);
3615 peel_list.remove(i);
3616 incr = false;
3617 #ifndef PRODUCT
3618 if (TracePartialPeeling) {
3619 tty->print_cr("sink to not_peeled region: %d newbb: %d",
3620 n->_idx, get_ctrl(n)->_idx);
3621 }
3622 #endif
3623 }
3624 } else {
3625 // Otherwise check for special def-use cases that span
3626 // the peel/not_peel boundary such as bool->if
3627 clone_for_special_use_inside_loop(loop, n, not_peel, sink_list, worklist);
3628 new_phi_cnt++;
3629 }
3630 }
3631 }
3632 if (incr) i++;
3633 }
3634
3635 estimate += cloned_for_outside_use + new_phi_cnt;
3636 bool exceed_node_budget = !may_require_nodes(estimate);
3637 bool exceed_phi_limit = new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta;
3638
3639 if (too_many_clones || exceed_node_budget || exceed_phi_limit) {
3640 #ifndef PRODUCT
3641 if (TracePartialPeeling && exceed_phi_limit) {
3642 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c",
3643 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F');
3644 }
3645 #endif
3646 if (new_peel_if != NULL) {
3647 remove_cmpi_loop_exit(new_peel_if, loop);
3648 }
3649 // Inhibit more partial peeling on this loop
3650 assert(!head->is_partial_peel_loop(), "not partial peeled");
3651 head->mark_partial_peel_failed();
3652 if (cloned_for_outside_use > 0) {
3653 // Terminate this round of loop opts because
3654 // the graph outside this loop was changed.
3655 C->set_major_progress();
3656 return true;
3657 }
3658 return false;
3659 }
3660
3661 // Step 3: clone loop, retarget control, and insert new phis
3662
3663 // Create new loop head for new phis and to hang
3664 // the nodes being moved (sinked) from the peel region.
3665 LoopNode* new_head = new LoopNode(last_peel, last_peel);
3666 new_head->set_unswitch_count(head->unswitch_count()); // Preserve
3667 _igvn.register_new_node_with_optimizer(new_head);
3668 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
3669 _igvn.replace_input_of(first_not_peeled, 0, new_head);
3670 set_loop(new_head, loop);
3671 loop->_body.push(new_head);
3672 not_peel.set(new_head->_idx);
3673 set_idom(new_head, last_peel, dom_depth(first_not_peeled));
3674 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
3675
3676 while (sink_list.size() > 0) {
3677 Node* n = sink_list.pop();
3678 set_ctrl(n, new_head);
3679 }
3680
3681 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
3682
3683 clone_loop(loop, old_new, dd, IgnoreStripMined);
3684
3685 const uint clone_exit_idx = 1;
3686 const uint orig_exit_idx = 2;
3687 assert(is_valid_clone_loop_form(loop, peel_list, orig_exit_idx, clone_exit_idx), "bad clone loop");
3688
3689 Node* head_clone = old_new[head->_idx];
3690 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop();
3691 Node* orig_tail_clone = head_clone->in(2);
3692
3693 // Add phi if "def" node is in peel set and "use" is not
3694
3695 for (uint i = 0; i < peel_list.size(); i++) {
3696 Node *def = peel_list.at(i);
3697 if (!def->is_CFG()) {
3698 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
3699 Node *use = def->fast_out(j);
3700 if (has_node(use) && use->in(0) != C->top() &&
3701 (!peel.test(use->_idx) ||
3702 (use->is_Phi() && use->in(0) == head)) ) {
3703 worklist.push(use);
3704 }
3705 }
3706 while( worklist.size() ) {
3707 Node *use = worklist.pop();
3708 for (uint j = 1; j < use->req(); j++) {
3709 Node* n = use->in(j);
3710 if (n == def) {
3711
3712 // "def" is in peel set, "use" is not in peel set
3713 // or "use" is in the entry boundary (a phi) of the peel set
3714
3715 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
3716
3717 if ( loop->is_member(get_loop( use_c )) ) {
3718 // use is in loop
3719 if (old_new[use->_idx] != NULL) { // null for dead code
3720 Node* use_clone = old_new[use->_idx];
3721 _igvn.replace_input_of(use, j, C->top());
3722 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
3723 }
3724 } else {
3725 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
3726 // use is not in the loop, check if the live range includes the cut
3727 Node* lp_if = use_c->in(orig_exit_idx)->in(0);
3728 if (not_peel.test(lp_if->_idx)) {
3729 assert(j == orig_exit_idx, "use from original loop");
3730 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
3731 }
3732 }
3733 }
3734 }
3735 }
3736 }
3737 }
3738
3739 // Step 3b: retarget control
3740
3741 // Redirect control to the new loop head if a cloned node in
3742 // the not_peeled region has control that points into the peeled region.
3743 // This necessary because the cloned peeled region will be outside
3744 // the loop.
3745 // from to
3746 // cloned-peeled <---+
3747 // new_head_clone: | <--+
3748 // cloned-not_peeled in(0) in(0)
3749 // orig-peeled
3750
3751 for (uint i = 0; i < loop->_body.size(); i++) {
3752 Node *n = loop->_body.at(i);
3753 if (!n->is_CFG() && n->in(0) != NULL &&
3754 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
3755 Node* n_clone = old_new[n->_idx];
3756 _igvn.replace_input_of(n_clone, 0, new_head_clone);
3757 }
3758 }
3759
3760 // Backedge of the surviving new_head (the clone) is original last_peel
3761 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel);
3762
3763 // Cut first node in original not_peel set
3764 _igvn.rehash_node_delayed(new_head); // Multiple edge updates:
3765 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of
3766 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls
3767
3768 // Copy head_clone back-branch info to original head
3769 // and remove original head's loop entry and
3770 // clone head's back-branch
3771 _igvn.rehash_node_delayed(head); // Multiple edge updates
3772 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl));
3773 head->set_req(LoopNode::LoopBackControl, C->top());
3774 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top());
3775
3776 // Similarly modify the phis
3777 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
3778 Node* use = head->fast_out(k);
3779 if (use->is_Phi() && use->outcnt() > 0) {
3780 Node* use_clone = old_new[use->_idx];
3781 _igvn.rehash_node_delayed(use); // Multiple edge updates
3782 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl));
3783 use->set_req(LoopNode::LoopBackControl, C->top());
3784 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top());
3785 }
3786 }
3787
3788 // Step 4: update dominator tree and dominator depth
3789
3790 set_idom(head, orig_tail_clone, dd);
3791 recompute_dom_depth();
3792
3793 // Inhibit more partial peeling on this loop
3794 new_head_clone->set_partial_peel_loop();
3795 C->set_major_progress();
3796 loop->record_for_igvn();
3797
3798 #ifndef PRODUCT
3799 if (TracePartialPeeling) {
3800 tty->print_cr("\nafter partial peel one iteration");
3801 Node_List wl;
3802 Node* t = last_peel;
3803 while (true) {
3804 wl.push(t);
3805 if (t == head_clone) break;
3806 t = idom(t);
3807 }
3808 while (wl.size() > 0) {
3809 Node* tt = wl.pop();
3810 if (tt == head) tty->print_cr("orig head");
3811 else if (tt == new_head_clone) tty->print_cr("new head");
3812 else if (tt == head_clone) tty->print_cr("clone head");
3813 tt->dump();
3814 }
3815 }
3816 #endif
3817 return true;
3818 }
3819
3820 // Transform:
3821 //
3822 // loop<-----------------+
3823 // | |
3824 // stmt1 stmt2 .. stmtn |
3825 // | | | |
3826 // \ | / |
3827 // v v v |
3828 // region |
3829 // | |
3830 // shared_stmt |
3831 // | |
3832 // v |
3833 // if |
3834 // / \ |
3835 // | -----------+
3836 // v
3837 //
3838 // into:
3839 //
3840 // loop<-------------------+
3841 // | |
3842 // v |
3843 // +->loop |
3844 // | | |
3845 // | stmt1 stmt2 .. stmtn |
3846 // | | | | |
3847 // | | \ / |
3848 // | | v v |
3849 // | | region1 |
3850 // | | | |
3851 // | shared_stmt shared_stmt |
3852 // | | | |
3853 // | v v |
3854 // | if if |
3855 // | /\ / \ |
3856 // +-- | | -------+
3857 // \ /
3858 // v v
3859 // region2
3860 //
3861 // (region2 is shown to merge mirrored projections of the loop exit
3862 // ifs to make the diagram clearer but they really merge the same
3863 // projection)
3864 //
3865 // Conditions for this transformation to trigger:
3866 // - the path through stmt1 is frequent enough
3867 // - the inner loop will be turned into a counted loop after transformation
3868 bool PhaseIdealLoop::duplicate_loop_backedge(IdealLoopTree *loop, Node_List &old_new) {
3869 if (!DuplicateBackedge) {
3870 return false;
3871 }
3872 assert(!loop->_head->is_CountedLoop() || StressDuplicateBackedge, "Non-counted loop only");
3873 if (!loop->_head->is_Loop()) {
3874 return false;
3875 }
3876
3877 uint estimate = loop->est_loop_clone_sz(1);
3878 if (exceeding_node_budget(estimate)) {
3879 return false;
3880 }
3881
3882 LoopNode *head = loop->_head->as_Loop();
3883
3884 Node* region = NULL;
3885 IfNode* exit_test = NULL;
3886 uint inner;
3887 float f;
3888 if (StressDuplicateBackedge) {
3889 if (head->is_strip_mined()) {
3890 return false;
3891 }
3892 Node* c = head->in(LoopNode::LoopBackControl);
3893
3894 while (c != head) {
3895 if (c->is_Region()) {
3896 region = c;
3897 }
3898 c = idom(c);
3899 }
3900
3901 if (region == NULL) {
3902 return false;
3903 }
3904
3905 inner = 1;
3906 } else {
3907 // Is the shape of the loop that of a counted loop...
3908 Node* back_control = loop_exit_control(head, loop);
3909 if (back_control == NULL) {
3910 return false;
3911 }
3912
3913 BoolTest::mask bt = BoolTest::illegal;
3914 float cl_prob = 0;
3915 Node* incr = NULL;
3916 Node* limit = NULL;
3917 Node* cmp = loop_exit_test(back_control, loop, incr, limit, bt, cl_prob);
3918 if (cmp == NULL || cmp->Opcode() != Op_CmpI) {
3919 return false;
3920 }
3921
3922 // With an extra phi for the candidate iv?
3923 // Or the region node is the loop head
3924 if (!incr->is_Phi() || incr->in(0) == head) {
3925 return false;
3926 }
3927
3928 PathFrequency pf(head, this);
3929 region = incr->in(0);
3930
3931 // Go over all paths for the extra phi's region and see if that
3932 // path is frequent enough and would match the expected iv shape
3933 // if the extra phi is removed
3934 inner = 0;
3935 for (uint i = 1; i < incr->req(); ++i) {
3936 Node* in = incr->in(i);
3937 Node* trunc1 = NULL;
3938 Node* trunc2 = NULL;
3939 const TypeInteger* iv_trunc_t = NULL;
3940 Node* orig_in = in;
3941 if (!(in = CountedLoopNode::match_incr_with_optional_truncation(in, &trunc1, &trunc2, &iv_trunc_t, T_INT))) {
3942 continue;
3943 }
3944 assert(in->Opcode() == Op_AddI, "wrong increment code");
3945 Node* xphi = NULL;
3946 Node* stride = loop_iv_stride(in, loop, xphi);
3947
3948 if (stride == NULL) {
3949 continue;
3950 }
3951
3952 PhiNode* phi = loop_iv_phi(xphi, NULL, head, loop);
3953 if (phi == NULL ||
3954 (trunc1 == NULL && phi->in(LoopNode::LoopBackControl) != incr) ||
3955 (trunc1 != NULL && phi->in(LoopNode::LoopBackControl) != trunc1)) {
3956 return false;
3957 }
3958
3959 f = pf.to(region->in(i));
3960 if (f > 0.5) {
3961 inner = i;
3962 break;
3963 }
3964 }
3965
3966 if (inner == 0) {
3967 return false;
3968 }
3969
3970 exit_test = back_control->in(0)->as_If();
3971 }
3972
3973 if (idom(region)->is_Catch()) {
3974 return false;
3975 }
3976
3977 // Collect all control nodes that need to be cloned (shared_stmt in the diagram)
3978 Unique_Node_List wq;
3979 wq.push(head->in(LoopNode::LoopBackControl));
3980 for (uint i = 0; i < wq.size(); i++) {
3981 Node* c = wq.at(i);
3982 assert(get_loop(c) == loop, "not in the right loop?");
3983 if (c->is_Region()) {
3984 if (c != region) {
3985 for (uint j = 1; j < c->req(); ++j) {
3986 wq.push(c->in(j));
3987 }
3988 }
3989 } else {
3990 wq.push(c->in(0));
3991 }
3992 assert(!is_dominator(c, region) || c == region, "shouldn't go above region");
3993 }
3994
3995 Node* region_dom = idom(region);
3996
3997 // Can't do the transformation if this would cause a membar pair to
3998 // be split
3999 for (uint i = 0; i < wq.size(); i++) {
4000 Node* c = wq.at(i);
4001 if (c->is_MemBar() && (c->as_MemBar()->trailing_store() || c->as_MemBar()->trailing_load_store())) {
4002 assert(c->as_MemBar()->leading_membar()->trailing_membar() == c, "bad membar pair");
4003 if (!wq.member(c->as_MemBar()->leading_membar())) {
4004 return false;
4005 }
4006 }
4007 }
4008
4009 // Collect data nodes that need to be clones as well
4010 int dd = dom_depth(head);
4011
4012 for (uint i = 0; i < loop->_body.size(); ++i) {
4013 Node* n = loop->_body.at(i);
4014 if (has_ctrl(n)) {
4015 Node* c = get_ctrl(n);
4016 if (wq.member(c)) {
4017 wq.push(n);
4018 }
4019 } else {
4020 set_idom(n, idom(n), dd);
4021 }
4022 }
4023
4024 // clone shared_stmt
4025 clone_loop_body(wq, old_new, NULL);
4026
4027 Node* region_clone = old_new[region->_idx];
4028 region_clone->set_req(inner, C->top());
4029 set_idom(region, region->in(inner), dd);
4030
4031 // Prepare the outer loop
4032 Node* outer_head = new LoopNode(head->in(LoopNode::EntryControl), old_new[head->in(LoopNode::LoopBackControl)->_idx]);
4033 register_control(outer_head, loop->_parent, outer_head->in(LoopNode::EntryControl));
4034 _igvn.replace_input_of(head, LoopNode::EntryControl, outer_head);
4035 set_idom(head, outer_head, dd);
4036
4037 fix_body_edges(wq, loop, old_new, dd, loop->_parent, true);
4038
4039 // Make one of the shared_stmt copies only reachable from stmt1, the
4040 // other only from stmt2..stmtn.
4041 Node* dom = NULL;
4042 for (uint i = 1; i < region->req(); ++i) {
4043 if (i != inner) {
4044 _igvn.replace_input_of(region, i, C->top());
4045 }
4046 Node* in = region_clone->in(i);
4047 if (in->is_top()) {
4048 continue;
4049 }
4050 if (dom == NULL) {
4051 dom = in;
4052 } else {
4053 dom = dom_lca(dom, in);
4054 }
4055 }
4056
4057 set_idom(region_clone, dom, dd);
4058
4059 // Set up the outer loop
4060 for (uint i = 0; i < head->outcnt(); i++) {
4061 Node* u = head->raw_out(i);
4062 if (u->is_Phi()) {
4063 Node* outer_phi = u->clone();
4064 outer_phi->set_req(0, outer_head);
4065 Node* backedge = old_new[u->in(LoopNode::LoopBackControl)->_idx];
4066 if (backedge == NULL) {
4067 backedge = u->in(LoopNode::LoopBackControl);
4068 }
4069 outer_phi->set_req(LoopNode::LoopBackControl, backedge);
4070 register_new_node(outer_phi, outer_head);
4071 _igvn.replace_input_of(u, LoopNode::EntryControl, outer_phi);
4072 }
4073 }
4074
4075 // create control and data nodes for out of loop uses (including region2)
4076 Node_List worklist;
4077 uint new_counter = C->unique();
4078 fix_ctrl_uses(wq, loop, old_new, ControlAroundStripMined, outer_head, NULL, worklist);
4079
4080 Node_List *split_if_set = NULL;
4081 Node_List *split_bool_set = NULL;
4082 Node_List *split_cex_set = NULL;
4083 fix_data_uses(wq, loop, ControlAroundStripMined, head->is_strip_mined() ? loop->_parent : loop, new_counter, old_new, worklist, split_if_set, split_bool_set, split_cex_set);
4084
4085 finish_clone_loop(split_if_set, split_bool_set, split_cex_set);
4086
4087 if (exit_test != NULL) {
4088 float cnt = exit_test->_fcnt;
4089 if (cnt != COUNT_UNKNOWN) {
4090 exit_test->_fcnt = cnt * f;
4091 old_new[exit_test->_idx]->as_If()->_fcnt = cnt * (1 - f);
4092 }
4093 }
4094
4095 C->set_major_progress();
4096
4097 return true;
4098 }
4099
4100 //------------------------------reorg_offsets----------------------------------
4101 // Reorganize offset computations to lower register pressure. Mostly
4102 // prevent loop-fallout uses of the pre-incremented trip counter (which are
4103 // then alive with the post-incremented trip counter forcing an extra
4104 // register move):
4105 //
4106 // iv Phi iv Phi
4107 // | |
4108 // | AddI (+stride)
4109 // | |
4110 // | Opaque2 # Blocks IGVN from folding these nodes until loop opts are over.
4111 // | ====> |
4112 // | AddI (-stride)
4113 // | |
4114 // | CastII # Preserve type of iv Phi
4115 // | |
4116 // Outside Use Outside Use
4117 //
4118 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) {
4119 // Perform it only for canonical counted loops.
4120 // Loop's shape could be messed up by iteration_split_impl.
4121 if (!loop->_head->is_CountedLoop())
4122 return;
4123 if (!loop->_head->as_Loop()->is_valid_counted_loop(T_INT))
4124 return;
4125
4126 CountedLoopNode *cl = loop->_head->as_CountedLoop();
4127 CountedLoopEndNode *cle = cl->loopexit();
4128 Node *exit = cle->proj_out(false);
4129 Node *phi = cl->phi();
4130
4131 // Check for the special case when using the pre-incremented trip-counter on
4132 // the fall-out path (forces the pre-incremented and post-incremented trip
4133 // counter to be live at the same time). Fix this by adjusting to use the
4134 // post-increment trip counter.
4135
4136 bool progress = true;
4137 while (progress) {
4138 progress = false;
4139 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
4140 Node* use = phi->fast_out(i); // User of trip-counter
4141 if (!has_ctrl(use)) continue;
4142 Node *u_ctrl = get_ctrl(use);
4143 if (use->is_Phi()) {
4144 u_ctrl = NULL;
4145 for (uint j = 1; j < use->req(); j++)
4146 if (use->in(j) == phi)
4147 u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j));
4148 }
4149 IdealLoopTree *u_loop = get_loop(u_ctrl);
4150 // Look for loop-invariant use
4151 if (u_loop == loop) continue;
4152 if (loop->is_member(u_loop)) continue;
4153 // Check that use is live out the bottom. Assuming the trip-counter
4154 // update is right at the bottom, uses of of the loop middle are ok.
4155 if (dom_lca(exit, u_ctrl) != exit) continue;
4156 // Hit! Refactor use to use the post-incremented tripcounter.
4157 // Compute a post-increment tripcounter.
4158 Node* c = exit;
4159 if (cl->is_strip_mined()) {
4160 IdealLoopTree* outer_loop = get_loop(cl->outer_loop());
4161 if (!outer_loop->is_member(u_loop)) {
4162 c = cl->outer_loop_exit();
4163 }
4164 }
4165 Node *opaq = new Opaque2Node(C, cle->incr());
4166 register_new_node(opaq, c);
4167 Node *neg_stride = _igvn.intcon(-cle->stride_con());
4168 set_ctrl(neg_stride, C->root());
4169 Node *post = new AddINode(opaq, neg_stride);
4170 register_new_node(post, c);
4171 post = new CastIINode(post, phi->bottom_type()); // preserve the iv phi's type
4172 register_new_node(post, c);
4173 _igvn.rehash_node_delayed(use);
4174 for (uint j = 1; j < use->req(); j++) {
4175 if (use->in(j) == phi)
4176 use->set_req(j, post);
4177 }
4178 // Since DU info changed, rerun loop
4179 progress = true;
4180 break;
4181 }
4182 }
4183
4184 }