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/inlinetypenode.hpp"
37 #include "opto/loopnode.hpp"
38 #include "opto/matcher.hpp"
39 #include "opto/mulnode.hpp"
40 #include "opto/movenode.hpp"
41 #include "opto/opaquenode.hpp"
42 #include "opto/rootnode.hpp"
43 #include "opto/subnode.hpp"
44 #include "opto/subtypenode.hpp"
45 #include "utilities/macros.hpp"
46
47 //=============================================================================
48 //------------------------------split_thru_phi---------------------------------
49 // Split Node 'n' through merge point if there is enough win.
50 Node* PhaseIdealLoop::split_thru_phi(Node* n, Node* region, int policy) {
51 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) {
52 // ConvI2L may have type information on it which is unsafe to push up
53 // so disable this for now
54 return NULL;
55 }
56
57 // Splitting range check CastIIs through a loop induction Phi can
58 // cause new Phis to be created that are left unrelated to the loop
59 // induction Phi and prevent optimizations (vectorization)
60 if (n->Opcode() == Op_CastII && region->is_CountedLoop() &&
61 n->in(1) == region->as_CountedLoop()->phi()) {
62 return NULL;
63 }
64
65 // Inline types should not be split through Phis because they cannot be merged
66 // through Phi nodes but each value input needs to be merged individually.
67 if (n->is_InlineType()) {
68 return NULL;
69 }
70
71 // 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
72 // 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
73 // 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
74 // of 'n', we need to bail out of one of these divisors could be zero (zero in its type range).
75 if ((n->Opcode() == Op_DivI || n->Opcode() == Op_ModI) && n->in(0) == NULL
76 && region->is_CountedLoop() && n->in(2) == region->as_CountedLoop()->phi()) {
77 Node* phi = region->as_CountedLoop()->phi();
78 for (uint i = 1; i < phi->req(); i++) {
79 if (_igvn.type(phi->in(i))->filter_speculative(TypeInt::ZERO) != Type::TOP) {
80 // 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.
81 return NULL;
82 }
83 }
84 }
85
86 int wins = 0;
87 assert(!n->is_CFG(), "");
88 assert(region->is_Region(), "");
89
90 const Type* type = n->bottom_type();
91 const TypeOopPtr* t_oop = _igvn.type(n)->isa_oopptr();
92 Node* phi;
93 if (t_oop != NULL && t_oop->is_known_instance_field()) {
94 int iid = t_oop->instance_id();
95 int index = C->get_alias_index(t_oop);
96 int offset = t_oop->offset();
97 phi = new PhiNode(region, type, NULL, iid, index, offset);
98 } else {
99 phi = PhiNode::make_blank(region, n);
100 }
101 uint old_unique = C->unique();
102 for (uint i = 1; i < region->req(); i++) {
103 Node* x;
104 Node* the_clone = NULL;
105 if (region->in(i) == C->top()) {
106 x = C->top(); // Dead path? Use a dead data op
107 } else {
108 x = n->clone(); // Else clone up the data op
109 the_clone = x; // Remember for possible deletion.
110 // Alter data node to use pre-phi inputs
111 if (n->in(0) == region)
112 x->set_req( 0, region->in(i) );
113 for (uint j = 1; j < n->req(); j++) {
114 Node* in = n->in(j);
115 if (in->is_Phi() && in->in(0) == region)
116 x->set_req(j, in->in(i)); // Use pre-Phi input for the clone
117 }
118 }
119 // Check for a 'win' on some paths
120 const Type* t = x->Value(&_igvn);
121
122 bool singleton = t->singleton();
123
124 // A TOP singleton indicates that there are no possible values incoming
125 // along a particular edge. In most cases, this is OK, and the Phi will
126 // be eliminated later in an Ideal call. However, we can't allow this to
127 // happen if the singleton occurs on loop entry, as the elimination of
128 // the PhiNode may cause the resulting node to migrate back to a previous
129 // loop iteration.
130 if (singleton && t == Type::TOP) {
131 // Is_Loop() == false does not confirm the absence of a loop (e.g., an
132 // irreducible loop may not be indicated by an affirmative is_Loop());
133 // therefore, the only top we can split thru a phi is on a backedge of
134 // a loop.
135 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
136 }
137
138 if (singleton) {
139 wins++;
140 x = ((PhaseGVN&)_igvn).makecon(t);
141 } else {
142 // We now call Identity to try to simplify the cloned node.
143 // Note that some Identity methods call phase->type(this).
144 // Make sure that the type array is big enough for
145 // our new node, even though we may throw the node away.
146 // (Note: This tweaking with igvn only works because x is a new node.)
147 _igvn.set_type(x, t);
148 // If x is a TypeNode, capture any more-precise type permanently into Node
149 // otherwise it will be not updated during igvn->transform since
150 // igvn->type(x) is set to x->Value() already.
151 x->raise_bottom_type(t);
152 Node* y = x->Identity(&_igvn);
153 if (y != x) {
154 wins++;
155 x = y;
156 } else {
157 y = _igvn.hash_find(x);
158 if (y) {
159 wins++;
160 x = y;
161 } else {
162 // Else x is a new node we are keeping
163 // We do not need register_new_node_with_optimizer
164 // because set_type has already been called.
165 _igvn._worklist.push(x);
166 }
167 }
168 }
169 if (x != the_clone && the_clone != NULL)
170 _igvn.remove_dead_node(the_clone);
171 phi->set_req( i, x );
172 }
173 // Too few wins?
174 if (wins <= policy) {
175 _igvn.remove_dead_node(phi);
176 return NULL;
177 }
178
179 // Record Phi
180 register_new_node( phi, region );
181
182 for (uint i2 = 1; i2 < phi->req(); i2++) {
183 Node *x = phi->in(i2);
184 // If we commoned up the cloned 'x' with another existing Node,
185 // the existing Node picks up a new use. We need to make the
186 // existing Node occur higher up so it dominates its uses.
187 Node *old_ctrl;
188 IdealLoopTree *old_loop;
189
190 if (x->is_Con()) {
191 // Constant's control is always root.
192 set_ctrl(x, C->root());
193 continue;
194 }
195 // The occasional new node
196 if (x->_idx >= old_unique) { // Found a new, unplaced node?
197 old_ctrl = NULL;
198 old_loop = NULL; // Not in any prior loop
199 } else {
200 old_ctrl = get_ctrl(x);
201 old_loop = get_loop(old_ctrl); // Get prior loop
202 }
203 // New late point must dominate new use
204 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2));
205 if (new_ctrl == old_ctrl) // Nothing is changed
206 continue;
207
208 IdealLoopTree *new_loop = get_loop(new_ctrl);
209
210 // Don't move x into a loop if its uses are
211 // outside of loop. Otherwise x will be cloned
212 // for each use outside of this loop.
213 IdealLoopTree *use_loop = get_loop(region);
214 if (!new_loop->is_member(use_loop) &&
215 (old_loop == NULL || !new_loop->is_member(old_loop))) {
216 // Take early control, later control will be recalculated
217 // during next iteration of loop optimizations.
218 new_ctrl = get_early_ctrl(x);
219 new_loop = get_loop(new_ctrl);
220 }
221 // Set new location
222 set_ctrl(x, new_ctrl);
223 // If changing loop bodies, see if we need to collect into new body
224 if (old_loop != new_loop) {
225 if (old_loop && !old_loop->_child)
226 old_loop->_body.yank(x);
227 if (!new_loop->_child)
228 new_loop->_body.push(x); // Collect body info
229 }
230 }
231
232 return phi;
233 }
234
235 //------------------------------dominated_by------------------------------------
236 // Replace the dominated test with an obvious true or false. Place it on the
237 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the
238 // live path up to the dominating control.
239 void PhaseIdealLoop::dominated_by(IfProjNode* prevdom, IfNode* iff, bool flip, bool exclude_loop_predicate) {
240 if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); }
241
242 // prevdom is the dominating projection of the dominating test.
243 assert(iff->Opcode() == Op_If ||
244 iff->Opcode() == Op_CountedLoopEnd ||
245 iff->Opcode() == Op_LongCountedLoopEnd ||
246 iff->Opcode() == Op_RangeCheck,
247 "Check this code when new subtype is added");
248
249 int pop = prevdom->Opcode();
250 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" );
251 if (flip) {
252 if (pop == Op_IfTrue)
253 pop = Op_IfFalse;
254 else
255 pop = Op_IfTrue;
256 }
257 // 'con' is set to true or false to kill the dominated test.
258 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO);
259 set_ctrl(con, C->root()); // Constant gets a new use
260 // Hack the dominated test
261 _igvn.replace_input_of(iff, 1, con);
262
263 // If I dont have a reachable TRUE and FALSE path following the IfNode then
264 // I can assume this path reaches an infinite loop. In this case it's not
265 // important to optimize the data Nodes - either the whole compilation will
266 // be tossed or this path (and all data Nodes) will go dead.
267 if (iff->outcnt() != 2) return;
268
269 // Make control-dependent data Nodes on the live path (path that will remain
270 // once the dominated IF is removed) become control-dependent on the
271 // dominating projection.
272 Node* dp = iff->proj_out_or_null(pop == Op_IfTrue);
273
274 // Loop predicates may have depending checks which should not
275 // be skipped. For example, range check predicate has two checks
276 // for lower and upper bounds.
277 if (dp == NULL)
278 return;
279
280 ProjNode* dp_proj = dp->as_Proj();
281 ProjNode* unc_proj = iff->proj_out(1 - dp_proj->_con)->as_Proj();
282 if (exclude_loop_predicate &&
283 (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) != NULL ||
284 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_profile_predicate) != NULL ||
285 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check) != NULL)) {
286 // If this is a range check (IfNode::is_range_check), do not
287 // reorder because Compile::allow_range_check_smearing might have
288 // changed the check.
289 return; // Let IGVN transformation change control dependence.
290 }
291
292 IdealLoopTree* old_loop = get_loop(dp);
293
294 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
295 Node* cd = dp->fast_out(i); // Control-dependent node
296 // Do not rewire Div and Mod nodes which could have a zero divisor to avoid skipping their zero check.
297 if (cd->depends_only_on_test() && _igvn.no_dependent_zero_check(cd)) {
298 assert(cd->in(0) == dp, "");
299 _igvn.replace_input_of(cd, 0, prevdom);
300 set_early_ctrl(cd, false);
301 IdealLoopTree* new_loop = get_loop(get_ctrl(cd));
302 if (old_loop != new_loop) {
303 if (!old_loop->_child) {
304 old_loop->_body.yank(cd);
305 }
306 if (!new_loop->_child) {
307 new_loop->_body.push(cd);
308 }
309 }
310 --i;
311 --imax;
312 }
313 }
314 }
315
316 //------------------------------has_local_phi_input----------------------------
317 // Return TRUE if 'n' has Phi inputs from its local block and no other
318 // block-local inputs (all non-local-phi inputs come from earlier blocks)
319 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) {
320 Node *n_ctrl = get_ctrl(n);
321 // See if some inputs come from a Phi in this block, or from before
322 // this block.
323 uint i;
324 for( i = 1; i < n->req(); i++ ) {
325 Node *phi = n->in(i);
326 if( phi->is_Phi() && phi->in(0) == n_ctrl )
327 break;
328 }
329 if( i >= n->req() )
330 return NULL; // No Phi inputs; nowhere to clone thru
331
332 // Check for inputs created between 'n' and the Phi input. These
333 // must split as well; they have already been given the chance
334 // (courtesy of a post-order visit) and since they did not we must
335 // recover the 'cost' of splitting them by being very profitable
336 // when splitting 'n'. Since this is unlikely we simply give up.
337 for( i = 1; i < n->req(); i++ ) {
338 Node *m = n->in(i);
339 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) {
340 // We allow the special case of AddP's with no local inputs.
341 // This allows us to split-up address expressions.
342 if (m->is_AddP() &&
343 get_ctrl(m->in(AddPNode::Base)) != n_ctrl &&
344 get_ctrl(m->in(AddPNode::Address)) != n_ctrl &&
345 get_ctrl(m->in(AddPNode::Offset)) != n_ctrl) {
346 // Move the AddP up to the dominating point. That's fine because control of m's inputs
347 // must dominate get_ctrl(m) == n_ctrl and we just checked that the input controls are != n_ctrl.
348 Node* c = find_non_split_ctrl(idom(n_ctrl));
349 if (c->is_OuterStripMinedLoop()) {
350 c->as_Loop()->verify_strip_mined(1);
351 c = c->in(LoopNode::EntryControl);
352 }
353 set_ctrl_and_loop(m, c);
354 continue;
355 }
356 return NULL;
357 }
358 assert(n->is_Phi() || m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control");
359 }
360
361 return n_ctrl;
362 }
363
364 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2).
365 Node* PhaseIdealLoop::remix_address_expressions_add_left_shift(Node* n, IdealLoopTree* n_loop, Node* n_ctrl, BasicType bt) {
366 assert(bt == T_INT || bt == T_LONG, "only for integers");
367 int n_op = n->Opcode();
368
369 if (n_op == Op_LShift(bt)) {
370 // Scale is loop invariant
371 Node* scale = n->in(2);
372 Node* scale_ctrl = get_ctrl(scale);
373 IdealLoopTree* scale_loop = get_loop(scale_ctrl);
374 if (n_loop == scale_loop || !scale_loop->is_member(n_loop)) {
375 return NULL;
376 }
377 const TypeInt* scale_t = scale->bottom_type()->isa_int();
378 if (scale_t != NULL && scale_t->is_con() && scale_t->get_con() >= 16) {
379 return NULL; // Dont bother with byte/short masking
380 }
381 // Add must vary with loop (else shift would be loop-invariant)
382 Node* add = n->in(1);
383 Node* add_ctrl = get_ctrl(add);
384 IdealLoopTree* add_loop = get_loop(add_ctrl);
385 if (n_loop != add_loop) {
386 return NULL; // happens w/ evil ZKM loops
387 }
388
389 // Convert I-V into I+ (0-V); same for V-I
390 if (add->Opcode() == Op_Sub(bt) &&
391 _igvn.type(add->in(1)) != TypeInteger::zero(bt)) {
392 assert(add->Opcode() == Op_SubI || add->Opcode() == Op_SubL, "");
393 Node* zero = _igvn.integercon(0, bt);
394 set_ctrl(zero, C->root());
395 Node* neg = SubNode::make(zero, add->in(2), bt);
396 register_new_node(neg, get_ctrl(add->in(2)));
397 add = AddNode::make(add->in(1), neg, bt);
398 register_new_node(add, add_ctrl);
399 }
400 if (add->Opcode() != Op_Add(bt)) return NULL;
401 assert(add->Opcode() == Op_AddI || add->Opcode() == Op_AddL, "");
402 // See if one add input is loop invariant
403 Node* add_var = add->in(1);
404 Node* add_var_ctrl = get_ctrl(add_var);
405 IdealLoopTree* add_var_loop = get_loop(add_var_ctrl);
406 Node* add_invar = add->in(2);
407 Node* add_invar_ctrl = get_ctrl(add_invar);
408 IdealLoopTree* add_invar_loop = get_loop(add_invar_ctrl);
409 if (add_invar_loop == n_loop) {
410 // Swap to find the invariant part
411 add_invar = add_var;
412 add_invar_ctrl = add_var_ctrl;
413 add_invar_loop = add_var_loop;
414 add_var = add->in(2);
415 } else if (add_var_loop != n_loop) { // Else neither input is loop invariant
416 return NULL;
417 }
418 if (n_loop == add_invar_loop || !add_invar_loop->is_member(n_loop)) {
419 return NULL; // No invariant part of the add?
420 }
421
422 // Yes! Reshape address expression!
423 Node* inv_scale = LShiftNode::make(add_invar, scale, bt);
424 Node* inv_scale_ctrl =
425 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ?
426 add_invar_ctrl : scale_ctrl;
427 register_new_node(inv_scale, inv_scale_ctrl);
428 Node* var_scale = LShiftNode::make(add_var, scale, bt);
429 register_new_node(var_scale, n_ctrl);
430 Node* var_add = AddNode::make(var_scale, inv_scale, bt);
431 register_new_node(var_add, n_ctrl);
432 _igvn.replace_node(n, var_add);
433 return var_add;
434 }
435 return NULL;
436 }
437
438 //------------------------------remix_address_expressions----------------------
439 // Rework addressing expressions to get the most loop-invariant stuff
440 // moved out. We'd like to do all associative operators, but it's especially
441 // important (common) to do address expressions.
442 Node* PhaseIdealLoop::remix_address_expressions(Node* n) {
443 if (!has_ctrl(n)) return NULL;
444 Node* n_ctrl = get_ctrl(n);
445 IdealLoopTree* n_loop = get_loop(n_ctrl);
446
447 // See if 'n' mixes loop-varying and loop-invariant inputs and
448 // itself is loop-varying.
449
450 // Only interested in binary ops (and AddP)
451 if (n->req() < 3 || n->req() > 4) return NULL;
452
453 Node* n1_ctrl = get_ctrl(n->in( 1));
454 Node* n2_ctrl = get_ctrl(n->in( 2));
455 Node* n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3));
456 IdealLoopTree* n1_loop = get_loop(n1_ctrl);
457 IdealLoopTree* n2_loop = get_loop(n2_ctrl);
458 IdealLoopTree* n3_loop = get_loop(n3_ctrl);
459
460 // Does one of my inputs spin in a tighter loop than self?
461 if ((n_loop->is_member(n1_loop) && n_loop != n1_loop) ||
462 (n_loop->is_member(n2_loop) && n_loop != n2_loop) ||
463 (n_loop->is_member(n3_loop) && n_loop != n3_loop)) {
464 return NULL; // Leave well enough alone
465 }
466
467 // Is at least one of my inputs loop-invariant?
468 if (n1_loop == n_loop &&
469 n2_loop == n_loop &&
470 n3_loop == n_loop) {
471 return NULL; // No loop-invariant inputs
472 }
473
474 Node* res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_INT);
475 if (res != NULL) {
476 return res;
477 }
478 res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_LONG);
479 if (res != NULL) {
480 return res;
481 }
482
483 int n_op = n->Opcode();
484 // Replace (I+V) with (V+I)
485 if (n_op == Op_AddI ||
486 n_op == Op_AddL ||
487 n_op == Op_AddF ||
488 n_op == Op_AddD ||
489 n_op == Op_MulI ||
490 n_op == Op_MulL ||
491 n_op == Op_MulF ||
492 n_op == Op_MulD) {
493 if (n2_loop == n_loop) {
494 assert(n1_loop != n_loop, "");
495 n->swap_edges(1, 2);
496 }
497 }
498
499 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V),
500 // but not if I2 is a constant.
501 if (n_op == Op_AddP) {
502 if (n2_loop == n_loop && n3_loop != n_loop) {
503 if (n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con()) {
504 Node* n22_ctrl = get_ctrl(n->in(2)->in(2));
505 Node* n23_ctrl = get_ctrl(n->in(2)->in(3));
506 IdealLoopTree* n22loop = get_loop(n22_ctrl);
507 IdealLoopTree* n23_loop = get_loop(n23_ctrl);
508 if (n22loop != n_loop && n22loop->is_member(n_loop) &&
509 n23_loop == n_loop) {
510 Node* add1 = new AddPNode(n->in(1), n->in(2)->in(2), n->in(3));
511 // Stuff new AddP in the loop preheader
512 register_new_node(add1, n_loop->_head->in(LoopNode::EntryControl));
513 Node* add2 = new AddPNode(n->in(1), add1, n->in(2)->in(3));
514 register_new_node(add2, n_ctrl);
515 _igvn.replace_node(n, add2);
516 return add2;
517 }
518 }
519 }
520
521 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V)
522 if (n2_loop != n_loop && n3_loop == n_loop) {
523 if (n->in(3)->Opcode() == Op_AddX) {
524 Node* V = n->in(3)->in(1);
525 Node* I = n->in(3)->in(2);
526 if (is_member(n_loop,get_ctrl(V))) {
527 } else {
528 Node *tmp = V; V = I; I = tmp;
529 }
530 if (!is_member(n_loop,get_ctrl(I))) {
531 Node* add1 = new AddPNode(n->in(1), n->in(2), I);
532 // Stuff new AddP in the loop preheader
533 register_new_node(add1, n_loop->_head->in(LoopNode::EntryControl));
534 Node* add2 = new AddPNode(n->in(1), add1, V);
535 register_new_node(add2, n_ctrl);
536 _igvn.replace_node(n, add2);
537 return add2;
538 }
539 }
540 }
541 }
542
543 return NULL;
544 }
545
546 // Optimize ((in1[2*i] * in2[2*i]) + (in1[2*i+1] * in2[2*i+1]))
547 Node *PhaseIdealLoop::convert_add_to_muladd(Node* n) {
548 assert(n->Opcode() == Op_AddI, "sanity");
549 Node * nn = NULL;
550 Node * in1 = n->in(1);
551 Node * in2 = n->in(2);
552 if (in1->Opcode() == Op_MulI && in2->Opcode() == Op_MulI) {
553 IdealLoopTree* loop_n = get_loop(get_ctrl(n));
554 if (loop_n->is_counted() &&
555 loop_n->_head->as_Loop()->is_valid_counted_loop(T_INT) &&
556 Matcher::match_rule_supported(Op_MulAddVS2VI) &&
557 Matcher::match_rule_supported(Op_MulAddS2I)) {
558 Node* mul_in1 = in1->in(1);
559 Node* mul_in2 = in1->in(2);
560 Node* mul_in3 = in2->in(1);
561 Node* mul_in4 = in2->in(2);
562 if (mul_in1->Opcode() == Op_LoadS &&
563 mul_in2->Opcode() == Op_LoadS &&
564 mul_in3->Opcode() == Op_LoadS &&
565 mul_in4->Opcode() == Op_LoadS) {
566 IdealLoopTree* loop1 = get_loop(get_ctrl(mul_in1));
567 IdealLoopTree* loop2 = get_loop(get_ctrl(mul_in2));
568 IdealLoopTree* loop3 = get_loop(get_ctrl(mul_in3));
569 IdealLoopTree* loop4 = get_loop(get_ctrl(mul_in4));
570 IdealLoopTree* loop5 = get_loop(get_ctrl(in1));
571 IdealLoopTree* loop6 = get_loop(get_ctrl(in2));
572 // All nodes should be in the same counted loop.
573 if (loop_n == loop1 && loop_n == loop2 && loop_n == loop3 &&
574 loop_n == loop4 && loop_n == loop5 && loop_n == loop6) {
575 Node* adr1 = mul_in1->in(MemNode::Address);
576 Node* adr2 = mul_in2->in(MemNode::Address);
577 Node* adr3 = mul_in3->in(MemNode::Address);
578 Node* adr4 = mul_in4->in(MemNode::Address);
579 if (adr1->is_AddP() && adr2->is_AddP() && adr3->is_AddP() && adr4->is_AddP()) {
580 if ((adr1->in(AddPNode::Base) == adr3->in(AddPNode::Base)) &&
581 (adr2->in(AddPNode::Base) == adr4->in(AddPNode::Base))) {
582 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in3, mul_in4);
583 register_new_node(nn, get_ctrl(n));
584 _igvn.replace_node(n, nn);
585 return nn;
586 } else if ((adr1->in(AddPNode::Base) == adr4->in(AddPNode::Base)) &&
587 (adr2->in(AddPNode::Base) == adr3->in(AddPNode::Base))) {
588 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in4, mul_in3);
589 register_new_node(nn, get_ctrl(n));
590 _igvn.replace_node(n, nn);
591 return nn;
592 }
593 }
594 }
595 }
596 }
597 }
598 return nn;
599 }
600
601 //------------------------------conditional_move-------------------------------
602 // Attempt to replace a Phi with a conditional move. We have some pretty
603 // strict profitability requirements. All Phis at the merge point must
604 // be converted, so we can remove the control flow. We need to limit the
605 // number of c-moves to a small handful. All code that was in the side-arms
606 // of the CFG diamond is now speculatively executed. This code has to be
607 // "cheap enough". We are pretty much limited to CFG diamonds that merge
608 // 1 or 2 items with a total of 1 or 2 ops executed speculatively.
609 Node *PhaseIdealLoop::conditional_move( Node *region ) {
610
611 assert(region->is_Region(), "sanity check");
612 if (region->req() != 3) return NULL;
613
614 // Check for CFG diamond
615 Node *lp = region->in(1);
616 Node *rp = region->in(2);
617 if (!lp || !rp) return NULL;
618 Node *lp_c = lp->in(0);
619 if (lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If()) return NULL;
620 IfNode *iff = lp_c->as_If();
621
622 // Check for ops pinned in an arm of the diamond.
623 // Can't remove the control flow in this case
624 if (lp->outcnt() > 1) return NULL;
625 if (rp->outcnt() > 1) return NULL;
626
627 IdealLoopTree* r_loop = get_loop(region);
628 assert(r_loop == get_loop(iff), "sanity");
629 // Always convert to CMOVE if all results are used only outside this loop.
630 bool used_inside_loop = (r_loop == _ltree_root);
631
632 // Check profitability
633 int cost = 0;
634 int phis = 0;
635 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
636 Node *out = region->fast_out(i);
637 if (!out->is_Phi()) continue; // Ignore other control edges, etc
638 phis++;
639 PhiNode* phi = out->as_Phi();
640 BasicType bt = phi->type()->basic_type();
641 switch (bt) {
642 case T_DOUBLE:
643 case T_FLOAT:
644 if (C->use_cmove()) {
645 continue; //TODO: maybe we want to add some cost
646 }
647 cost += Matcher::float_cmove_cost(); // Could be very expensive
648 break;
649 case T_LONG: {
650 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's
651 }
652 case T_INT: // These all CMOV fine
653 case T_ADDRESS: { // (RawPtr)
654 cost++;
655 break;
656 }
657 case T_NARROWOOP: // Fall through
658 case T_OBJECT: { // Base oops are OK, but not derived oops
659 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr();
660 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a
661 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus
662 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we
663 // have a Phi for the base here that we convert to a CMOVE all is well
664 // and good. But if the base is dead, we'll not make a CMOVE. Later
665 // the allocator will have to produce a base by creating a CMOVE of the
666 // relevant bases. This puts the allocator in the business of
667 // manufacturing expensive instructions, generally a bad plan.
668 // Just Say No to Conditionally-Moved Derived Pointers.
669 if (tp && tp->offset() != 0)
670 return NULL;
671 cost++;
672 break;
673 }
674 default:
675 return NULL; // In particular, can't do memory or I/O
676 }
677 // Add in cost any speculative ops
678 for (uint j = 1; j < region->req(); j++) {
679 Node *proj = region->in(j);
680 Node *inp = phi->in(j);
681 if (get_ctrl(inp) == proj) { // Found local op
682 cost++;
683 // Check for a chain of dependent ops; these will all become
684 // speculative in a CMOV.
685 for (uint k = 1; k < inp->req(); k++)
686 if (get_ctrl(inp->in(k)) == proj)
687 cost += ConditionalMoveLimit; // Too much speculative goo
688 }
689 }
690 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode.
691 // This will likely Split-If, a higher-payoff operation.
692 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) {
693 Node* use = phi->fast_out(k);
694 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr())
695 cost += ConditionalMoveLimit;
696 // Is there a use inside the loop?
697 // Note: check only basic types since CMoveP is pinned.
698 if (!used_inside_loop && is_java_primitive(bt)) {
699 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use);
700 if (r_loop == u_loop || r_loop->is_member(u_loop)) {
701 used_inside_loop = true;
702 }
703 }
704 }
705 }//for
706 Node* bol = iff->in(1);
707 if (bol->Opcode() == Op_Opaque4) {
708 return NULL; // Ignore loop predicate checks (the Opaque4 ensures they will go away)
709 }
710 assert(bol->Opcode() == Op_Bool, "Unexpected node");
711 int cmp_op = bol->in(1)->Opcode();
712 if (cmp_op == Op_SubTypeCheck) { // SubTypeCheck expansion expects an IfNode
713 return NULL;
714 }
715 // It is expensive to generate flags from a float compare.
716 // Avoid duplicated float compare.
717 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL;
718
719 float infrequent_prob = PROB_UNLIKELY_MAG(3);
720 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop.
721 if (used_inside_loop) {
722 if (cost >= ConditionalMoveLimit) return NULL; // Too much goo
723
724 // BlockLayoutByFrequency optimization moves infrequent branch
725 // from hot path. No point in CMOV'ing in such case (110 is used
726 // instead of 100 to take into account not exactness of float value).
727 if (BlockLayoutByFrequency) {
728 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f);
729 }
730 }
731 // Check for highly predictable branch. No point in CMOV'ing if
732 // we are going to predict accurately all the time.
733 if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) {
734 //keep going
735 } else if (iff->_prob < infrequent_prob ||
736 iff->_prob > (1.0f - infrequent_prob))
737 return NULL;
738
739 // --------------
740 // Now replace all Phis with CMOV's
741 Node *cmov_ctrl = iff->in(0);
742 uint flip = (lp->Opcode() == Op_IfTrue);
743 Node_List wq;
744 while (1) {
745 PhiNode* phi = NULL;
746 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
747 Node *out = region->fast_out(i);
748 if (out->is_Phi()) {
749 phi = out->as_Phi();
750 break;
751 }
752 }
753 if (phi == NULL || _igvn.type(phi) == Type::TOP) {
754 break;
755 }
756 if (PrintOpto && VerifyLoopOptimizations) { tty->print_cr("CMOV"); }
757 // Move speculative ops
758 wq.push(phi);
759 while (wq.size() > 0) {
760 Node *n = wq.pop();
761 for (uint j = 1; j < n->req(); j++) {
762 Node* m = n->in(j);
763 if (m != NULL && !is_dominator(get_ctrl(m), cmov_ctrl)) {
764 #ifndef PRODUCT
765 if (PrintOpto && VerifyLoopOptimizations) {
766 tty->print(" speculate: ");
767 m->dump();
768 }
769 #endif
770 set_ctrl(m, cmov_ctrl);
771 wq.push(m);
772 }
773 }
774 }
775 Node *cmov = CMoveNode::make(cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi));
776 register_new_node( cmov, cmov_ctrl );
777 _igvn.replace_node( phi, cmov );
778 #ifndef PRODUCT
779 if (TraceLoopOpts) {
780 tty->print("CMOV ");
781 r_loop->dump_head();
782 if (Verbose) {
783 bol->in(1)->dump(1);
784 cmov->dump(1);
785 }
786 }
787 if (VerifyLoopOptimizations) verify();
788 #endif
789 }
790
791 // The useless CFG diamond will fold up later; see the optimization in
792 // RegionNode::Ideal.
793 _igvn._worklist.push(region);
794
795 return iff->in(1);
796 }
797
798 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) {
799 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
800 Node* u = m->fast_out(i);
801 if (u->is_CFG()) {
802 if (u->Opcode() == Op_NeverBranch) {
803 u = ((NeverBranchNode*)u)->proj_out(0);
804 enqueue_cfg_uses(u, wq);
805 } else {
806 wq.push(u);
807 }
808 }
809 }
810 }
811
812 // Try moving a store out of a loop, right before the loop
813 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) {
814 // Store has to be first in the loop body
815 IdealLoopTree *n_loop = get_loop(n_ctrl);
816 if (n->is_Store() && n_loop != _ltree_root &&
817 n_loop->is_loop() && n_loop->_head->is_Loop() &&
818 n->in(0) != NULL) {
819 Node* address = n->in(MemNode::Address);
820 Node* value = n->in(MemNode::ValueIn);
821 Node* mem = n->in(MemNode::Memory);
822 IdealLoopTree* address_loop = get_loop(get_ctrl(address));
823 IdealLoopTree* value_loop = get_loop(get_ctrl(value));
824
825 // - address and value must be loop invariant
826 // - memory must be a memory Phi for the loop
827 // - Store must be the only store on this memory slice in the
828 // loop: if there's another store following this one then value
829 // written at iteration i by the second store could be overwritten
830 // at iteration i+n by the first store: it's not safe to move the
831 // first store out of the loop
832 // - nothing must observe the memory Phi: it guarantees no read
833 // before the store, we are also guaranteed the store post
834 // dominates the loop head (ignoring a possible early
835 // exit). Otherwise there would be extra Phi involved between the
836 // loop's Phi and the store.
837 // - there must be no early exit from the loop before the Store
838 // (such an exit most of the time would be an extra use of the
839 // memory Phi but sometimes is a bottom memory Phi that takes the
840 // store as input).
841
842 if (!n_loop->is_member(address_loop) &&
843 !n_loop->is_member(value_loop) &&
844 mem->is_Phi() && mem->in(0) == n_loop->_head &&
845 mem->outcnt() == 1 &&
846 mem->in(LoopNode::LoopBackControl) == n) {
847
848 assert(n_loop->_tail != NULL, "need a tail");
849 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop");
850
851 // Verify that there's no early exit of the loop before the store.
852 bool ctrl_ok = false;
853 {
854 // Follow control from loop head until n, we exit the loop or
855 // we reach the tail
856 ResourceMark rm;
857 Unique_Node_List wq;
858 wq.push(n_loop->_head);
859
860 for (uint next = 0; next < wq.size(); ++next) {
861 Node *m = wq.at(next);
862 if (m == n->in(0)) {
863 ctrl_ok = true;
864 continue;
865 }
866 assert(!has_ctrl(m), "should be CFG");
867 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) {
868 ctrl_ok = false;
869 break;
870 }
871 enqueue_cfg_uses(m, wq);
872 if (wq.size() > 10) {
873 ctrl_ok = false;
874 break;
875 }
876 }
877 }
878 if (ctrl_ok) {
879 // move the Store
880 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem);
881 _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl));
882 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl));
883 // Disconnect the phi now. An empty phi can confuse other
884 // optimizations in this pass of loop opts.
885 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl));
886 n_loop->_body.yank(mem);
887
888 set_ctrl_and_loop(n, n->in(0));
889
890 return n;
891 }
892 }
893 }
894 return NULL;
895 }
896
897 // Try moving a store out of a loop, right after the loop
898 void PhaseIdealLoop::try_move_store_after_loop(Node* n) {
899 if (n->is_Store() && n->in(0) != NULL) {
900 Node *n_ctrl = get_ctrl(n);
901 IdealLoopTree *n_loop = get_loop(n_ctrl);
902 // Store must be in a loop
903 if (n_loop != _ltree_root && !n_loop->_irreducible) {
904 Node* address = n->in(MemNode::Address);
905 Node* value = n->in(MemNode::ValueIn);
906 IdealLoopTree* address_loop = get_loop(get_ctrl(address));
907 // address must be loop invariant
908 if (!n_loop->is_member(address_loop)) {
909 // Store must be last on this memory slice in the loop and
910 // nothing in the loop must observe it
911 Node* phi = NULL;
912 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
913 Node* u = n->fast_out(i);
914 if (has_ctrl(u)) { // control use?
915 IdealLoopTree *u_loop = get_loop(get_ctrl(u));
916 if (!n_loop->is_member(u_loop)) {
917 continue;
918 }
919 if (u->is_Phi() && u->in(0) == n_loop->_head) {
920 assert(_igvn.type(u) == Type::MEMORY, "bad phi");
921 // multiple phis on the same slice are possible
922 if (phi != NULL) {
923 return;
924 }
925 phi = u;
926 continue;
927 }
928 }
929 return;
930 }
931 if (phi != NULL) {
932 // Nothing in the loop before the store (next iteration)
933 // must observe the stored value
934 bool mem_ok = true;
935 {
936 ResourceMark rm;
937 Unique_Node_List wq;
938 wq.push(phi);
939 for (uint next = 0; next < wq.size() && mem_ok; ++next) {
940 Node *m = wq.at(next);
941 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) {
942 Node* u = m->fast_out(i);
943 if (u->is_Store() || u->is_Phi()) {
944 if (u != n) {
945 wq.push(u);
946 mem_ok = (wq.size() <= 10);
947 }
948 } else {
949 mem_ok = false;
950 break;
951 }
952 }
953 }
954 }
955 if (mem_ok) {
956 // Move the store out of the loop if the LCA of all
957 // users (except for the phi) is outside the loop.
958 Node* hook = new Node(1);
959 hook->init_req(0, n_ctrl); // Add an input to prevent hook from being dead
960 _igvn.rehash_node_delayed(phi);
961 int count = phi->replace_edge(n, hook, &_igvn);
962 assert(count > 0, "inconsistent phi");
963
964 // Compute latest point this store can go
965 Node* lca = get_late_ctrl(n, get_ctrl(n));
966 if (lca->is_OuterStripMinedLoop()) {
967 lca = lca->in(LoopNode::EntryControl);
968 }
969 if (n_loop->is_member(get_loop(lca))) {
970 // LCA is in the loop - bail out
971 _igvn.replace_node(hook, n);
972 return;
973 }
974 #ifdef ASSERT
975 if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) {
976 assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined");
977 n_loop->_head->as_Loop()->verify_strip_mined(1);
978 Node* outer = n_loop->_head->as_CountedLoop()->outer_loop();
979 IdealLoopTree* outer_loop = get_loop(outer);
980 assert(n_loop->_parent == outer_loop, "broken loop tree");
981 assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state");
982 }
983 #endif
984 lca = place_outside_loop(lca, n_loop);
985 assert(!n_loop->is_member(get_loop(lca)), "control must not be back in the loop");
986 assert(get_loop(lca)->_nest < n_loop->_nest || lca->in(0)->Opcode() == Op_NeverBranch, "must not be moved into inner loop");
987
988 // Move store out of the loop
989 _igvn.replace_node(hook, n->in(MemNode::Memory));
990 _igvn.replace_input_of(n, 0, lca);
991 set_ctrl_and_loop(n, lca);
992
993 // Disconnect the phi now. An empty phi can confuse other
994 // optimizations in this pass of loop opts..
995 if (phi->in(LoopNode::LoopBackControl) == phi) {
996 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl));
997 n_loop->_body.yank(phi);
998 }
999 }
1000 }
1001 }
1002 }
1003 }
1004 }
1005
1006 // If UseArrayMarkWordCheck is enabled, we can't use immutable memory for the flat array check
1007 // because we are loading the mark word which is mutable. Although the bits we are interested in
1008 // are immutable (we check for markWord::unlocked_value), we need to use raw memory to not break
1009 // anti dependency analysis. Below code will attempt to still move flat array checks out of loops,
1010 // mainly to enable loop unswitching.
1011 void PhaseIdealLoop::move_flat_array_check_out_of_loop(Node* n) {
1012 // Skip checks for more than one array
1013 if (n->req() > 3) {
1014 return;
1015 }
1016 Node* mem = n->in(FlatArrayCheckNode::Memory);
1017 Node* array = n->in(FlatArrayCheckNode::ArrayOrKlass)->uncast();
1018 IdealLoopTree* check_loop = get_loop(get_ctrl(n));
1019 IdealLoopTree* ary_loop = get_loop(get_ctrl(array));
1020
1021 // Check if array is loop invariant
1022 if (!check_loop->is_member(ary_loop)) {
1023 // Walk up memory graph from the check until we leave the loop
1024 VectorSet wq;
1025 wq.set(mem->_idx);
1026 while (check_loop->is_member(get_loop(ctrl_or_self(mem)))) {
1027 if (mem->is_Phi()) {
1028 mem = mem->in(1);
1029 } else if (mem->is_MergeMem()) {
1030 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1031 } else if (mem->is_Proj()) {
1032 mem = mem->in(0);
1033 } else if (mem->is_MemBar() || mem->is_SafePoint()) {
1034 mem = mem->in(TypeFunc::Memory);
1035 } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) {
1036 mem = mem->in(MemNode::Memory);
1037 } else {
1038 #ifdef ASSERT
1039 mem->dump();
1040 #endif
1041 ShouldNotReachHere();
1042 }
1043 if (wq.test_set(mem->_idx)) {
1044 return;
1045 }
1046 }
1047 // Replace memory input and re-compute ctrl to move the check out of the loop
1048 _igvn.replace_input_of(n, 1, mem);
1049 set_ctrl_and_loop(n, get_early_ctrl(n));
1050 Node* bol = n->unique_out();
1051 set_ctrl_and_loop(bol, get_early_ctrl(bol));
1052 }
1053 }
1054
1055 //------------------------------split_if_with_blocks_pre-----------------------
1056 // Do the real work in a non-recursive function. Data nodes want to be
1057 // cloned in the pre-order so they can feed each other nicely.
1058 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
1059 // Cloning these guys is unlikely to win
1060 int n_op = n->Opcode();
1061 if (n_op == Op_MergeMem) {
1062 return n;
1063 }
1064 if (n->is_Proj()) {
1065 return n;
1066 }
1067
1068 if (UseArrayMarkWordCheck && n->isa_FlatArrayCheck()) {
1069 move_flat_array_check_out_of_loop(n);
1070 return n;
1071 }
1072
1073 // Do not clone-up CmpFXXX variations, as these are always
1074 // followed by a CmpI
1075 if (n->is_Cmp()) {
1076 return n;
1077 }
1078 // Attempt to use a conditional move instead of a phi/branch
1079 if (ConditionalMoveLimit > 0 && n_op == Op_Region) {
1080 Node *cmov = conditional_move( n );
1081 if (cmov) {
1082 return cmov;
1083 }
1084 }
1085 if (n->is_CFG() || n->is_LoadStore()) {
1086 return n;
1087 }
1088 if (n->is_Opaque1() || // Opaque nodes cannot be mod'd
1089 n_op == Op_Opaque2) {
1090 if (!C->major_progress()) { // If chance of no more loop opts...
1091 _igvn._worklist.push(n); // maybe we'll remove them
1092 }
1093 return n;
1094 }
1095
1096 if (n->is_Con()) {
1097 return n; // No cloning for Con nodes
1098 }
1099
1100 Node *n_ctrl = get_ctrl(n);
1101 if (!n_ctrl) {
1102 return n; // Dead node
1103 }
1104
1105 Node* res = try_move_store_before_loop(n, n_ctrl);
1106 if (res != NULL) {
1107 return n;
1108 }
1109
1110 // Attempt to remix address expressions for loop invariants
1111 Node *m = remix_address_expressions( n );
1112 if( m ) return m;
1113
1114 if (n_op == Op_AddI) {
1115 Node *nn = convert_add_to_muladd( n );
1116 if ( nn ) return nn;
1117 }
1118
1119 if (n->is_ConstraintCast()) {
1120 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this);
1121 // ConstraintCastNode::dominating_cast() uses node control input to determine domination.
1122 // Node control inputs don't necessarily agree with loop control info (due to
1123 // transformations happened in between), thus additional dominance check is needed
1124 // to keep loop info valid.
1125 if (dom_cast != NULL && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) {
1126 _igvn.replace_node(n, dom_cast);
1127 return dom_cast;
1128 }
1129 }
1130
1131 // Determine if the Node has inputs from some local Phi.
1132 // Returns the block to clone thru.
1133 Node *n_blk = has_local_phi_input( n );
1134 if( !n_blk ) return n;
1135
1136 // Do not clone the trip counter through on a CountedLoop
1137 // (messes up the canonical shape).
1138 if (((n_blk->is_CountedLoop() || (n_blk->is_Loop() && n_blk->as_Loop()->is_loop_nest_inner_loop())) && n->Opcode() == Op_AddI) ||
1139 (n_blk->is_LongCountedLoop() && n->Opcode() == Op_AddL)) {
1140 return n;
1141 }
1142 // Pushing a shift through the iv Phi can get in the way of addressing optimizations or range check elimination
1143 if (n_blk->is_BaseCountedLoop() && n->Opcode() == Op_LShift(n_blk->as_BaseCountedLoop()->bt()) &&
1144 n->in(1) == n_blk->as_BaseCountedLoop()->phi()) {
1145 return n;
1146 }
1147
1148 // Check for having no control input; not pinned. Allow
1149 // dominating control.
1150 if (n->in(0)) {
1151 Node *dom = idom(n_blk);
1152 if (dom_lca(n->in(0), dom) != n->in(0)) {
1153 return n;
1154 }
1155 }
1156 // Policy: when is it profitable. You must get more wins than
1157 // policy before it is considered profitable. Policy is usually 0,
1158 // so 1 win is considered profitable. Big merges will require big
1159 // cloning, so get a larger policy.
1160 int policy = n_blk->req() >> 2;
1161
1162 // If the loop is a candidate for range check elimination,
1163 // delay splitting through it's phi until a later loop optimization
1164 if (n_blk->is_BaseCountedLoop()) {
1165 IdealLoopTree *lp = get_loop(n_blk);
1166 if (lp && lp->_rce_candidate) {
1167 return n;
1168 }
1169 }
1170
1171 if (must_throttle_split_if()) return n;
1172
1173 // Split 'n' through the merge point if it is profitable
1174 Node *phi = split_thru_phi( n, n_blk, policy );
1175 if (!phi) return n;
1176
1177 // Found a Phi to split thru!
1178 // Replace 'n' with the new phi
1179 _igvn.replace_node( n, phi );
1180 // Moved a load around the loop, 'en-registering' something.
1181 if (n_blk->is_Loop() && n->is_Load() &&
1182 !phi->in(LoopNode::LoopBackControl)->is_Load())
1183 C->set_major_progress();
1184
1185 return phi;
1186 }
1187
1188 static bool merge_point_too_heavy(Compile* C, Node* region) {
1189 // Bail out if the region and its phis have too many users.
1190 int weight = 0;
1191 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1192 weight += region->fast_out(i)->outcnt();
1193 }
1194 int nodes_left = C->max_node_limit() - C->live_nodes();
1195 if (weight * 8 > nodes_left) {
1196 if (PrintOpto) {
1197 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight);
1198 }
1199 return true;
1200 } else {
1201 return false;
1202 }
1203 }
1204
1205 static bool merge_point_safe(Node* region) {
1206 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode
1207 // having a PhiNode input. This sidesteps the dangerous case where the split
1208 // ConvI2LNode may become TOP if the input Value() does not
1209 // overlap the ConvI2L range, leaving a node which may not dominate its
1210 // uses.
1211 // A better fix for this problem can be found in the BugTraq entry, but
1212 // expediency for Mantis demands this hack.
1213 #ifdef _LP64
1214 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1215 Node* n = region->fast_out(i);
1216 if (n->is_Phi()) {
1217 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1218 Node* m = n->fast_out(j);
1219 if (m->Opcode() == Op_ConvI2L)
1220 return false;
1221 if (m->is_CastII()) {
1222 return false;
1223 }
1224 }
1225 }
1226 }
1227 #endif
1228 return true;
1229 }
1230
1231
1232 //------------------------------place_outside_loop---------------------------------
1233 // Place some computation outside of this loop on the path to the use passed as argument
1234 Node* PhaseIdealLoop::place_outside_loop(Node* useblock, IdealLoopTree* loop) const {
1235 Node* head = loop->_head;
1236 assert(!loop->is_member(get_loop(useblock)), "must be outside loop");
1237 if (head->is_Loop() && head->as_Loop()->is_strip_mined()) {
1238 loop = loop->_parent;
1239 assert(loop->_head->is_OuterStripMinedLoop(), "malformed strip mined loop");
1240 }
1241
1242 // Pick control right outside the loop
1243 for (;;) {
1244 Node* dom = idom(useblock);
1245 if (loop->is_member(get_loop(dom)) ||
1246 // NeverBranch nodes are not assigned to the loop when constructed
1247 (dom->Opcode() == Op_NeverBranch && loop->is_member(get_loop(dom->in(0))))) {
1248 break;
1249 }
1250 useblock = dom;
1251 }
1252 assert(find_non_split_ctrl(useblock) == useblock, "should be non split control");
1253 return useblock;
1254 }
1255
1256
1257 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) {
1258 if (!n->is_If() || n->is_BaseCountedLoopEnd()) {
1259 return false;
1260 }
1261 if (!n->in(0)->is_Region()) {
1262 return false;
1263 }
1264
1265 Node* region = n->in(0);
1266 Node* dom = idom(region);
1267 if (!dom->is_If() || dom->in(1) != n->in(1)) {
1268 return false;
1269 }
1270 IfNode* dom_if = dom->as_If();
1271 Node* proj_true = dom_if->proj_out(1);
1272 Node* proj_false = dom_if->proj_out(0);
1273
1274 for (uint i = 1; i < region->req(); i++) {
1275 if (is_dominator(proj_true, region->in(i))) {
1276 continue;
1277 }
1278 if (is_dominator(proj_false, region->in(i))) {
1279 continue;
1280 }
1281 return false;
1282 }
1283
1284 return true;
1285 }
1286
1287
1288 bool PhaseIdealLoop::can_split_if(Node* n_ctrl) {
1289 if (must_throttle_split_if()) {
1290 return false;
1291 }
1292
1293 // Do not do 'split-if' if irreducible loops are present.
1294 if (_has_irreducible_loops) {
1295 return false;
1296 }
1297
1298 if (merge_point_too_heavy(C, n_ctrl)) {
1299 return false;
1300 }
1301
1302 // Do not do 'split-if' if some paths are dead. First do dead code
1303 // elimination and then see if its still profitable.
1304 for (uint i = 1; i < n_ctrl->req(); i++) {
1305 if (n_ctrl->in(i) == C->top()) {
1306 return false;
1307 }
1308 }
1309
1310 // If trying to do a 'Split-If' at the loop head, it is only
1311 // profitable if the cmp folds up on BOTH paths. Otherwise we
1312 // risk peeling a loop forever.
1313
1314 // CNC - Disabled for now. Requires careful handling of loop
1315 // body selection for the cloned code. Also, make sure we check
1316 // for any input path not being in the same loop as n_ctrl. For
1317 // irreducible loops we cannot check for 'n_ctrl->is_Loop()'
1318 // because the alternative loop entry points won't be converted
1319 // into LoopNodes.
1320 IdealLoopTree *n_loop = get_loop(n_ctrl);
1321 for (uint j = 1; j < n_ctrl->req(); j++) {
1322 if (get_loop(n_ctrl->in(j)) != n_loop) {
1323 return false;
1324 }
1325 }
1326
1327 // Check for safety of the merge point.
1328 if (!merge_point_safe(n_ctrl)) {
1329 return false;
1330 }
1331
1332 return true;
1333 }
1334
1335 // Detect if the node is the inner strip-mined loop
1336 // Return: NULL if it's not the case, or the exit of outer strip-mined loop
1337 static Node* is_inner_of_stripmined_loop(const Node* out) {
1338 Node* out_le = NULL;
1339
1340 if (out->is_CountedLoopEnd()) {
1341 const CountedLoopNode* loop = out->as_CountedLoopEnd()->loopnode();
1342
1343 if (loop != NULL && loop->is_strip_mined()) {
1344 out_le = loop->in(LoopNode::EntryControl)->as_OuterStripMinedLoop()->outer_loop_exit();
1345 }
1346 }
1347
1348 return out_le;
1349 }
1350
1351 bool PhaseIdealLoop::flatten_array_element_type_check(Node *n) {
1352 // If the CmpP is a subtype check for a value that has just been
1353 // loaded from an array, the subtype check guarantees the value
1354 // can't be stored in a flattened array and the load of the value
1355 // happens with a flattened array check then: push the type check
1356 // through the phi of the flattened array check. This needs special
1357 // logic because the subtype check's input is not a phi but a
1358 // LoadKlass that must first be cloned through the phi.
1359 if (n->Opcode() != Op_CmpP) {
1360 return false;
1361 }
1362
1363 Node* klassptr = n->in(1);
1364 Node* klasscon = n->in(2);
1365
1366 if (klassptr->is_DecodeNarrowPtr()) {
1367 klassptr = klassptr->in(1);
1368 }
1369
1370 if (klassptr->Opcode() != Op_LoadKlass && klassptr->Opcode() != Op_LoadNKlass) {
1371 return false;
1372 }
1373
1374 if (!klasscon->is_Con()) {
1375 return false;
1376 }
1377
1378 Node* addr = klassptr->in(MemNode::Address);
1379
1380 if (!addr->is_AddP()) {
1381 return false;
1382 }
1383
1384 intptr_t offset;
1385 Node* obj = AddPNode::Ideal_base_and_offset(addr, &_igvn, offset);
1386
1387 if (obj == NULL) {
1388 return false;
1389 }
1390
1391 assert(obj != NULL && addr->in(AddPNode::Base) == addr->in(AddPNode::Address), "malformed AddP?");
1392 if (obj->Opcode() == Op_CastPP) {
1393 obj = obj->in(1);
1394 }
1395
1396 if (!obj->is_Phi()) {
1397 return false;
1398 }
1399
1400 Node* region = obj->in(0);
1401
1402 Node* phi = PhiNode::make_blank(region, n->in(1));
1403 for (uint i = 1; i < region->req(); i++) {
1404 Node* in = obj->in(i);
1405 Node* ctrl = region->in(i);
1406 if (addr->in(AddPNode::Base) != obj) {
1407 Node* cast = addr->in(AddPNode::Base);
1408 assert(cast->Opcode() == Op_CastPP && cast->in(0) != NULL, "inconsistent subgraph");
1409 Node* cast_clone = cast->clone();
1410 cast_clone->set_req(0, ctrl);
1411 cast_clone->set_req(1, in);
1412 register_new_node(cast_clone, ctrl);
1413 _igvn.set_type(cast_clone, cast_clone->Value(&_igvn));
1414 in = cast_clone;
1415 }
1416 Node* addr_clone = addr->clone();
1417 addr_clone->set_req(AddPNode::Base, in);
1418 addr_clone->set_req(AddPNode::Address, in);
1419 register_new_node(addr_clone, ctrl);
1420 _igvn.set_type(addr_clone, addr_clone->Value(&_igvn));
1421 Node* klassptr_clone = klassptr->clone();
1422 klassptr_clone->set_req(2, addr_clone);
1423 register_new_node(klassptr_clone, ctrl);
1424 _igvn.set_type(klassptr_clone, klassptr_clone->Value(&_igvn));
1425 if (klassptr != n->in(1)) {
1426 Node* decode = n->in(1);
1427 assert(decode->is_DecodeNarrowPtr(), "inconsistent subgraph");
1428 Node* decode_clone = decode->clone();
1429 decode_clone->set_req(1, klassptr_clone);
1430 register_new_node(decode_clone, ctrl);
1431 _igvn.set_type(decode_clone, decode_clone->Value(&_igvn));
1432 klassptr_clone = decode_clone;
1433 }
1434 phi->set_req(i, klassptr_clone);
1435 }
1436 register_new_node(phi, region);
1437 Node* orig = n->in(1);
1438 _igvn.replace_input_of(n, 1, phi);
1439 split_if_with_blocks_post(n);
1440 if (n->outcnt() != 0) {
1441 _igvn.replace_input_of(n, 1, orig);
1442 _igvn.remove_dead_node(phi);
1443 }
1444 return true;
1445 }
1446
1447 //------------------------------split_if_with_blocks_post----------------------
1448 // Do the real work in a non-recursive function. CFG hackery wants to be
1449 // in the post-order, so it can dirty the I-DOM info and not use the dirtied
1450 // info.
1451 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) {
1452
1453 if (flatten_array_element_type_check(n)) {
1454 return;
1455 }
1456
1457 // Cloning Cmp through Phi's involves the split-if transform.
1458 // FastLock is not used by an If
1459 if (n->is_Cmp() && !n->is_FastLock()) {
1460 Node *n_ctrl = get_ctrl(n);
1461 // Determine if the Node has inputs from some local Phi.
1462 // Returns the block to clone thru.
1463 Node *n_blk = has_local_phi_input(n);
1464 if (n_blk != n_ctrl) {
1465 return;
1466 }
1467
1468 if (!can_split_if(n_ctrl)) {
1469 return;
1470 }
1471
1472 if (n->outcnt() != 1) {
1473 return; // Multiple bool's from 1 compare?
1474 }
1475 Node *bol = n->unique_out();
1476 assert(bol->is_Bool(), "expect a bool here");
1477 if (bol->outcnt() != 1) {
1478 return;// Multiple branches from 1 compare?
1479 }
1480 Node *iff = bol->unique_out();
1481
1482 // Check some safety conditions
1483 if (iff->is_If()) { // Classic split-if?
1484 if (iff->in(0) != n_ctrl) {
1485 return; // Compare must be in same blk as if
1486 }
1487 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE
1488 // Can't split CMove with different control edge.
1489 if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) {
1490 return;
1491 }
1492 if (get_ctrl(iff->in(2)) == n_ctrl ||
1493 get_ctrl(iff->in(3)) == n_ctrl) {
1494 return; // Inputs not yet split-up
1495 }
1496 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) {
1497 return; // Loop-invar test gates loop-varying CMOVE
1498 }
1499 } else {
1500 return; // some other kind of node, such as an Allocate
1501 }
1502
1503 // When is split-if profitable? Every 'win' on means some control flow
1504 // goes dead, so it's almost always a win.
1505 int policy = 0;
1506 // Split compare 'n' through the merge point if it is profitable
1507 Node *phi = split_thru_phi( n, n_ctrl, policy);
1508 if (!phi) {
1509 return;
1510 }
1511
1512 // Found a Phi to split thru!
1513 // Replace 'n' with the new phi
1514 _igvn.replace_node(n, phi);
1515
1516 // Now split the bool up thru the phi
1517 Node *bolphi = split_thru_phi(bol, n_ctrl, -1);
1518 guarantee(bolphi != NULL, "null boolean phi node");
1519
1520 _igvn.replace_node(bol, bolphi);
1521 assert(iff->in(1) == bolphi, "");
1522
1523 if (bolphi->Value(&_igvn)->singleton()) {
1524 return;
1525 }
1526
1527 // Conditional-move? Must split up now
1528 if (!iff->is_If()) {
1529 Node *cmovphi = split_thru_phi(iff, n_ctrl, -1);
1530 _igvn.replace_node(iff, cmovphi);
1531 return;
1532 }
1533
1534 // Now split the IF
1535 do_split_if(iff);
1536 return;
1537 }
1538
1539 // Two identical ifs back to back can be merged
1540 if (try_merge_identical_ifs(n)) {
1541 return;
1542 }
1543
1544 // Check for an IF ready to split; one that has its
1545 // condition codes input coming from a Phi at the block start.
1546 int n_op = n->Opcode();
1547
1548 // Check for an IF being dominated by another IF same test
1549 if (n_op == Op_If ||
1550 n_op == Op_RangeCheck) {
1551 Node *bol = n->in(1);
1552 uint max = bol->outcnt();
1553 // Check for same test used more than once?
1554 if (max > 1 && bol->is_Bool()) {
1555 // Search up IDOMs to see if this IF is dominated.
1556 Node *cutoff = get_ctrl(bol);
1557
1558 // Now search up IDOMs till cutoff, looking for a dominating test
1559 Node *prevdom = n;
1560 Node *dom = idom(prevdom);
1561 while (dom != cutoff) {
1562 if (dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom &&
1563 safe_for_if_replacement(dom)) {
1564 // It's invalid to move control dependent data nodes in the inner
1565 // strip-mined loop, because:
1566 // 1) break validation of LoopNode::verify_strip_mined()
1567 // 2) move code with side-effect in strip-mined loop
1568 // Move to the exit of outer strip-mined loop in that case.
1569 Node* out_le = is_inner_of_stripmined_loop(dom);
1570 if (out_le != NULL) {
1571 prevdom = out_le;
1572 }
1573 // Replace the dominated test with an obvious true or false.
1574 // Place it on the IGVN worklist for later cleanup.
1575 C->set_major_progress();
1576 dominated_by(prevdom->as_IfProj(), n->as_If(), false, true);
1577 #ifndef PRODUCT
1578 if( VerifyLoopOptimizations ) verify();
1579 #endif
1580 return;
1581 }
1582 prevdom = dom;
1583 dom = idom(prevdom);
1584 }
1585 }
1586 }
1587
1588 try_sink_out_of_loop(n);
1589
1590 try_move_store_after_loop(n);
1591
1592 // Remove multiple allocations of the same inline type
1593 if (n->is_InlineType()) {
1594 n->as_InlineType()->remove_redundant_allocations(&_igvn, this);
1595 }
1596
1597 // Check for Opaque2's who's loop has disappeared - who's input is in the
1598 // same loop nest as their output. Remove 'em, they are no longer useful.
1599 if( n_op == Op_Opaque2 &&
1600 n->in(1) != NULL &&
1601 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) {
1602 _igvn.replace_node( n, n->in(1) );
1603 }
1604 }
1605
1606 // Transform:
1607 //
1608 // if (some_condition) {
1609 // // body 1
1610 // } else {
1611 // // body 2
1612 // }
1613 // if (some_condition) {
1614 // // body 3
1615 // } else {
1616 // // body 4
1617 // }
1618 //
1619 // into:
1620 //
1621 //
1622 // if (some_condition) {
1623 // // body 1
1624 // // body 3
1625 // } else {
1626 // // body 2
1627 // // body 4
1628 // }
1629 bool PhaseIdealLoop::try_merge_identical_ifs(Node* n) {
1630 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) {
1631 Node *n_ctrl = n->in(0);
1632 IfNode* dom_if = idom(n_ctrl)->as_If();
1633 ProjNode* dom_proj_true = dom_if->proj_out(1);
1634 ProjNode* dom_proj_false = dom_if->proj_out(0);
1635
1636 // Now split the IF
1637 RegionNode* new_false_region;
1638 RegionNode* new_true_region;
1639 do_split_if(n, &new_false_region, &new_true_region);
1640 assert(new_false_region->req() == new_true_region->req(), "");
1641 #ifdef ASSERT
1642 for (uint i = 1; i < new_false_region->req(); ++i) {
1643 assert(new_false_region->in(i)->in(0) == new_true_region->in(i)->in(0), "unexpected shape following split if");
1644 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");
1645 }
1646 #endif
1647 assert(new_false_region->in(1)->in(0)->in(1) == dom_if->in(1), "dominating if and dominated if after split must share test");
1648
1649 // We now have:
1650 // if (some_condition) {
1651 // // body 1
1652 // if (some_condition) {
1653 // body3: // new_true_region
1654 // // body3
1655 // } else {
1656 // goto body4;
1657 // }
1658 // } else {
1659 // // body 2
1660 // if (some_condition) {
1661 // goto body3;
1662 // } else {
1663 // body4: // new_false_region
1664 // // body4;
1665 // }
1666 // }
1667 //
1668
1669 // clone pinned nodes thru the resulting regions
1670 push_pinned_nodes_thru_region(dom_if, new_true_region);
1671 push_pinned_nodes_thru_region(dom_if, new_false_region);
1672
1673 // Optimize out the cloned ifs. Because pinned nodes were cloned, this also allows a CastPP that would be dependent
1674 // 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
1675 // unrelated control dependency.
1676 for (uint i = 1; i < new_false_region->req(); i++) {
1677 if (is_dominator(dom_proj_true, new_false_region->in(i))) {
1678 dominated_by(dom_proj_true->as_IfProj(), new_false_region->in(i)->in(0)->as_If(), false, false);
1679 } else {
1680 assert(is_dominator(dom_proj_false, new_false_region->in(i)), "bad if");
1681 dominated_by(dom_proj_false->as_IfProj(), new_false_region->in(i)->in(0)->as_If(), false, false);
1682 }
1683 }
1684 return true;
1685 }
1686 return false;
1687 }
1688
1689 void PhaseIdealLoop::push_pinned_nodes_thru_region(IfNode* dom_if, Node* region) {
1690 for (DUIterator i = region->outs(); region->has_out(i); i++) {
1691 Node* u = region->out(i);
1692 if (!has_ctrl(u) || u->is_Phi() || !u->depends_only_on_test() || !_igvn.no_dependent_zero_check(u)) {
1693 continue;
1694 }
1695 assert(u->in(0) == region, "not a control dependent node?");
1696 uint j = 1;
1697 for (; j < u->req(); ++j) {
1698 Node* in = u->in(j);
1699 if (!is_dominator(ctrl_or_self(in), dom_if)) {
1700 break;
1701 }
1702 }
1703 if (j == u->req()) {
1704 Node *phi = PhiNode::make_blank(region, u);
1705 for (uint k = 1; k < region->req(); ++k) {
1706 Node* clone = u->clone();
1707 clone->set_req(0, region->in(k));
1708 register_new_node(clone, region->in(k));
1709 phi->init_req(k, clone);
1710 }
1711 register_new_node(phi, region);
1712 _igvn.replace_node(u, phi);
1713 --i;
1714 }
1715 }
1716 }
1717
1718 bool PhaseIdealLoop::safe_for_if_replacement(const Node* dom) const {
1719 if (!dom->is_CountedLoopEnd()) {
1720 return true;
1721 }
1722 CountedLoopEndNode* le = dom->as_CountedLoopEnd();
1723 CountedLoopNode* cl = le->loopnode();
1724 if (cl == NULL) {
1725 return true;
1726 }
1727 if (!cl->is_main_loop()) {
1728 return true;
1729 }
1730 if (cl->is_canonical_loop_entry() == NULL) {
1731 return true;
1732 }
1733 // Further unrolling is possible so loop exit condition might change
1734 return false;
1735 }
1736
1737 // See if a shared loop-varying computation has no loop-varying uses.
1738 // Happens if something is only used for JVM state in uncommon trap exits,
1739 // like various versions of induction variable+offset. Clone the
1740 // computation per usage to allow it to sink out of the loop.
1741 void PhaseIdealLoop::try_sink_out_of_loop(Node* n) {
1742 bool is_raw_to_oop_cast = n->is_ConstraintCast() &&
1743 n->in(1)->bottom_type()->isa_rawptr() &&
1744 !n->bottom_type()->isa_rawptr();
1745 if (has_ctrl(n) &&
1746 !n->is_Phi() &&
1747 !n->is_Bool() &&
1748 !n->is_Proj() &&
1749 !n->is_MergeMem() &&
1750 !n->is_CMove() &&
1751 !is_raw_to_oop_cast && // don't extend live ranges of raw oops
1752 n->Opcode() != Op_Opaque4 &&
1753 !n->is_Type()) {
1754 Node *n_ctrl = get_ctrl(n);
1755 IdealLoopTree *n_loop = get_loop(n_ctrl);
1756
1757 if (n->in(0) != NULL) {
1758 IdealLoopTree* loop_ctrl = get_loop(n->in(0));
1759 if (n_loop != loop_ctrl && n_loop->is_member(loop_ctrl)) {
1760 // n has a control input inside a loop but get_ctrl() is member of an outer loop. This could happen, for example,
1761 // for Div nodes inside a loop (control input inside loop) without a use except for an UCT (outside the loop).
1762 // Rewire control of n to right outside of the loop, regardless if its input(s) are later sunk or not.
1763 _igvn.replace_input_of(n, 0, place_outside_loop(n_ctrl, loop_ctrl));
1764 }
1765 }
1766 if (n_loop != _ltree_root && n->outcnt() > 1) {
1767 // Compute early control: needed for anti-dependence analysis. It's also possible that as a result of
1768 // previous transformations in this loop opts round, the node can be hoisted now: early control will tell us.
1769 Node* early_ctrl = compute_early_ctrl(n, n_ctrl);
1770 if (n_loop->is_member(get_loop(early_ctrl)) && // check that this one can't be hoisted now
1771 ctrl_of_all_uses_out_of_loop(n, early_ctrl, n_loop)) { // All uses in outer loops!
1772 assert(!n->is_Store() && !n->is_LoadStore(), "no node with a side effect");
1773 Node* outer_loop_clone = NULL;
1774 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin;) {
1775 Node* u = n->last_out(j); // Clone private computation per use
1776 _igvn.rehash_node_delayed(u);
1777 Node* x = n->clone(); // Clone computation
1778 Node* x_ctrl = NULL;
1779 if (u->is_Phi()) {
1780 // Replace all uses of normal nodes. Replace Phi uses
1781 // individually, so the separate Nodes can sink down
1782 // different paths.
1783 uint k = 1;
1784 while (u->in(k) != n) k++;
1785 u->set_req(k, x);
1786 // x goes next to Phi input path
1787 x_ctrl = u->in(0)->in(k);
1788 // Find control for 'x' next to use but not inside inner loops.
1789 x_ctrl = place_outside_loop(x_ctrl, n_loop);
1790 --j;
1791 } else { // Normal use
1792 if (has_ctrl(u)) {
1793 x_ctrl = get_ctrl(u);
1794 } else {
1795 x_ctrl = u->in(0);
1796 }
1797 // Find control for 'x' next to use but not inside inner loops.
1798 x_ctrl = place_outside_loop(x_ctrl, n_loop);
1799 // Replace all uses
1800 if (u->is_ConstraintCast() && u->bottom_type()->higher_equal(_igvn.type(n)) && u->in(0) == x_ctrl) {
1801 // If we're sinking a chain of data nodes, we might have inserted a cast to pin the use which is not necessary
1802 // anymore now that we're going to pin n as well
1803 _igvn.replace_node(u, x);
1804 --j;
1805 } else {
1806 int nb = u->replace_edge(n, x, &_igvn);
1807 j -= nb;
1808 }
1809 }
1810
1811 if (n->is_Load()) {
1812 // For loads, add a control edge to a CFG node outside of the loop
1813 // to force them to not combine and return back inside the loop
1814 // during GVN optimization (4641526).
1815 assert(x_ctrl == get_late_ctrl_with_anti_dep(x->as_Load(), early_ctrl, x_ctrl), "anti-dependences were already checked");
1816
1817 IdealLoopTree* x_loop = get_loop(x_ctrl);
1818 Node* x_head = x_loop->_head;
1819 if (x_head->is_Loop() && x_head->is_OuterStripMinedLoop()) {
1820 // Do not add duplicate LoadNodes to the outer strip mined loop
1821 if (outer_loop_clone != NULL) {
1822 _igvn.replace_node(x, outer_loop_clone);
1823 continue;
1824 }
1825 outer_loop_clone = x;
1826 }
1827 x->set_req(0, x_ctrl);
1828 } else if (n->in(0) != NULL){
1829 x->set_req(0, x_ctrl);
1830 }
1831 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone");
1832 assert(!n_loop->is_member(get_loop(x_ctrl)), "should have moved out of loop");
1833 register_new_node(x, x_ctrl);
1834
1835 // Chain of AddP: (AddP base (AddP base )) must keep the same base after sinking so:
1836 // 1- We don't add a CastPP here when the first one is sunk so if the second one is not, their bases remain
1837 // the same.
1838 // (see 2- below)
1839 assert(!x->is_AddP() || !x->in(AddPNode::Address)->is_AddP() ||
1840 x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base) ||
1841 !x->in(AddPNode::Address)->in(AddPNode::Base)->eqv_uncast(x->in(AddPNode::Base)), "unexpected AddP shape");
1842 if (x->in(0) == NULL && !x->is_DecodeNarrowPtr() &&
1843 !(x->is_AddP() && x->in(AddPNode::Address)->is_AddP() && x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base))) {
1844 assert(!x->is_Load(), "load should be pinned");
1845 // Use a cast node to pin clone out of loop
1846 Node* cast = NULL;
1847 for (uint k = 0; k < x->req(); k++) {
1848 Node* in = x->in(k);
1849 if (in != NULL && n_loop->is_member(get_loop(get_ctrl(in)))) {
1850 const Type* in_t = _igvn.type(in);
1851 cast = ConstraintCastNode::make_cast_for_type(x_ctrl, in, in_t, ConstraintCastNode::UnconditionalDependency);
1852 }
1853 if (cast != NULL) {
1854 register_new_node(cast, x_ctrl);
1855 x->replace_edge(in, cast);
1856 // Chain of AddP:
1857 // 2- A CastPP of the base is only added now that both AddP nodes are sunk
1858 if (x->is_AddP() && k == AddPNode::Base) {
1859 for (DUIterator_Fast imax, i = x->fast_outs(imax); i < imax; i++) {
1860 Node* u = x->fast_out(i);
1861 if (u->is_AddP() && u->in(AddPNode::Base) == n->in(AddPNode::Base)) {
1862 _igvn.replace_input_of(u, AddPNode::Base, cast);
1863 assert(u->find_out_with(Op_AddP) == NULL, "more than 2 chained AddP nodes?");
1864 }
1865 }
1866 }
1867 break;
1868 }
1869 }
1870 assert(cast != NULL, "must have added a cast to pin the node");
1871 }
1872 }
1873 _igvn.remove_dead_node(n);
1874 }
1875 _dom_lca_tags_round = 0;
1876 }
1877 }
1878 }
1879
1880 // Compute the early control of a node by following its inputs until we reach
1881 // nodes that are pinned. Then compute the LCA of the control of all pinned nodes.
1882 Node* PhaseIdealLoop::compute_early_ctrl(Node* n, Node* n_ctrl) {
1883 Node* early_ctrl = NULL;
1884 ResourceMark rm;
1885 Unique_Node_List wq;
1886 wq.push(n);
1887 for (uint i = 0; i < wq.size(); i++) {
1888 Node* m = wq.at(i);
1889 Node* c = NULL;
1890 if (m->is_CFG()) {
1891 c = m;
1892 } else if (m->pinned()) {
1893 c = m->in(0);
1894 } else {
1895 for (uint j = 0; j < m->req(); j++) {
1896 Node* in = m->in(j);
1897 if (in != NULL) {
1898 wq.push(in);
1899 }
1900 }
1901 }
1902 if (c != NULL) {
1903 assert(is_dominator(c, n_ctrl), "control input must dominate current control");
1904 if (early_ctrl == NULL || is_dominator(early_ctrl, c)) {
1905 early_ctrl = c;
1906 }
1907 }
1908 }
1909 assert(is_dominator(early_ctrl, n_ctrl), "early control must dominate current control");
1910 return early_ctrl;
1911 }
1912
1913 bool PhaseIdealLoop::ctrl_of_all_uses_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop) {
1914 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1915 Node* u = n->fast_out(i);
1916 if (u->Opcode() == Op_Opaque1) {
1917 return false; // Found loop limit, bugfix for 4677003
1918 }
1919 // We can't reuse tags in PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal() so make sure calls to
1920 // get_late_ctrl_with_anti_dep() use their own tag
1921 _dom_lca_tags_round++;
1922 assert(_dom_lca_tags_round != 0, "shouldn't wrap around");
1923
1924 if (u->is_Phi()) {
1925 for (uint j = 1; j < u->req(); ++j) {
1926 if (u->in(j) == n && !ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, u->in(0)->in(j))) {
1927 return false;
1928 }
1929 }
1930 } else {
1931 Node* ctrl = has_ctrl(u) ? get_ctrl(u) : u->in(0);
1932 if (!ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, ctrl)) {
1933 return false;
1934 }
1935 }
1936 }
1937 return true;
1938 }
1939
1940 bool PhaseIdealLoop::ctrl_of_use_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop, Node* ctrl) {
1941 if (n->is_Load()) {
1942 ctrl = get_late_ctrl_with_anti_dep(n->as_Load(), n_ctrl, ctrl);
1943 }
1944 IdealLoopTree *u_loop = get_loop(ctrl);
1945 if (u_loop == n_loop) {
1946 return false; // Found loop-varying use
1947 }
1948 if (n_loop->is_member(u_loop)) {
1949 return false; // Found use in inner loop
1950 }
1951 return true;
1952 }
1953
1954 //------------------------------split_if_with_blocks---------------------------
1955 // Check for aggressive application of 'split-if' optimization,
1956 // using basic block level info.
1957 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack) {
1958 Node* root = C->root();
1959 visited.set(root->_idx); // first, mark root as visited
1960 // Do pre-visit work for root
1961 Node* n = split_if_with_blocks_pre(root);
1962 uint cnt = n->outcnt();
1963 uint i = 0;
1964
1965 while (true) {
1966 // Visit all children
1967 if (i < cnt) {
1968 Node* use = n->raw_out(i);
1969 ++i;
1970 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
1971 // Now do pre-visit work for this use
1972 use = split_if_with_blocks_pre(use);
1973 nstack.push(n, i); // Save parent and next use's index.
1974 n = use; // Process all children of current use.
1975 cnt = use->outcnt();
1976 i = 0;
1977 }
1978 }
1979 else {
1980 // All of n's children have been processed, complete post-processing.
1981 if (cnt != 0 && !n->is_Con()) {
1982 assert(has_node(n), "no dead nodes");
1983 split_if_with_blocks_post(n);
1984 }
1985 if (must_throttle_split_if()) {
1986 nstack.clear();
1987 }
1988 if (nstack.is_empty()) {
1989 // Finished all nodes on stack.
1990 break;
1991 }
1992 // Get saved parent node and next use's index. Visit the rest of uses.
1993 n = nstack.node();
1994 cnt = n->outcnt();
1995 i = nstack.index();
1996 nstack.pop();
1997 }
1998 }
1999 }
2000
2001
2002 //=============================================================================
2003 //
2004 // C L O N E A L O O P B O D Y
2005 //
2006
2007 //------------------------------clone_iff--------------------------------------
2008 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
2009 // "Nearly" because all Nodes have been cloned from the original in the loop,
2010 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
2011 // through the Phi recursively, and return a Bool.
2012 Node* PhaseIdealLoop::clone_iff(PhiNode* phi) {
2013
2014 // Convert this Phi into a Phi merging Bools
2015 uint i;
2016 for (i = 1; i < phi->req(); i++) {
2017 Node *b = phi->in(i);
2018 if (b->is_Phi()) {
2019 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi()));
2020 } else {
2021 assert(b->is_Bool() || b->Opcode() == Op_Opaque4, "");
2022 }
2023 }
2024
2025 Node* n = phi->in(1);
2026 Node* sample_opaque = NULL;
2027 Node *sample_bool = NULL;
2028 if (n->Opcode() == Op_Opaque4) {
2029 sample_opaque = n;
2030 sample_bool = n->in(1);
2031 assert(sample_bool->is_Bool(), "wrong type");
2032 } else {
2033 sample_bool = n;
2034 }
2035 Node* sample_cmp = sample_bool->in(1);
2036 const Type* t = Type::TOP;
2037 const TypePtr* at = NULL;
2038 if (sample_cmp->is_FlatArrayCheck()) {
2039 // Left input of a FlatArrayCheckNode is memory, set the (adr) type of the phi accordingly
2040 assert(sample_cmp->in(1)->bottom_type() == Type::MEMORY, "unexpected input type");
2041 t = Type::MEMORY;
2042 at = TypeRawPtr::BOTTOM;
2043 }
2044
2045 // Make Phis to merge the Cmp's inputs.
2046 PhiNode *phi1 = new PhiNode(phi->in(0), t, at);
2047 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP);
2048 for (i = 1; i < phi->req(); i++) {
2049 Node *n1 = sample_opaque == NULL ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1);
2050 Node *n2 = sample_opaque == NULL ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2);
2051 phi1->set_req(i, n1);
2052 phi2->set_req(i, n2);
2053 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
2054 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
2055 }
2056 // See if these Phis have been made before.
2057 // Register with optimizer
2058 Node *hit1 = _igvn.hash_find_insert(phi1);
2059 if (hit1) { // Hit, toss just made Phi
2060 _igvn.remove_dead_node(phi1); // Remove new phi
2061 assert(hit1->is_Phi(), "" );
2062 phi1 = (PhiNode*)hit1; // Use existing phi
2063 } else { // Miss
2064 _igvn.register_new_node_with_optimizer(phi1);
2065 }
2066 Node *hit2 = _igvn.hash_find_insert(phi2);
2067 if (hit2) { // Hit, toss just made Phi
2068 _igvn.remove_dead_node(phi2); // Remove new phi
2069 assert(hit2->is_Phi(), "" );
2070 phi2 = (PhiNode*)hit2; // Use existing phi
2071 } else { // Miss
2072 _igvn.register_new_node_with_optimizer(phi2);
2073 }
2074 // Register Phis with loop/block info
2075 set_ctrl(phi1, phi->in(0));
2076 set_ctrl(phi2, phi->in(0));
2077 // Make a new Cmp
2078 Node *cmp = sample_cmp->clone();
2079 cmp->set_req(1, phi1);
2080 cmp->set_req(2, phi2);
2081 _igvn.register_new_node_with_optimizer(cmp);
2082 set_ctrl(cmp, phi->in(0));
2083
2084 // Make a new Bool
2085 Node *b = sample_bool->clone();
2086 b->set_req(1,cmp);
2087 _igvn.register_new_node_with_optimizer(b);
2088 set_ctrl(b, phi->in(0));
2089
2090 if (sample_opaque != NULL) {
2091 Node* opaque = sample_opaque->clone();
2092 opaque->set_req(1, b);
2093 _igvn.register_new_node_with_optimizer(opaque);
2094 set_ctrl(opaque, phi->in(0));
2095 return opaque;
2096 }
2097
2098 assert(b->is_Bool(), "");
2099 return b;
2100 }
2101
2102 //------------------------------clone_bool-------------------------------------
2103 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
2104 // "Nearly" because all Nodes have been cloned from the original in the loop,
2105 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
2106 // through the Phi recursively, and return a Bool.
2107 CmpNode*PhaseIdealLoop::clone_bool(PhiNode* phi) {
2108 uint i;
2109 // Convert this Phi into a Phi merging Bools
2110 for( i = 1; i < phi->req(); i++ ) {
2111 Node *b = phi->in(i);
2112 if( b->is_Phi() ) {
2113 _igvn.replace_input_of(phi, i, clone_bool(b->as_Phi()));
2114 } else {
2115 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
2116 }
2117 }
2118
2119 Node *sample_cmp = phi->in(1);
2120
2121 // Make Phis to merge the Cmp's inputs.
2122 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP );
2123 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP );
2124 for( uint j = 1; j < phi->req(); j++ ) {
2125 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
2126 Node *n1, *n2;
2127 if( cmp_top->is_Cmp() ) {
2128 n1 = cmp_top->in(1);
2129 n2 = cmp_top->in(2);
2130 } else {
2131 n1 = n2 = cmp_top;
2132 }
2133 phi1->set_req( j, n1 );
2134 phi2->set_req( j, n2 );
2135 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
2136 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
2137 }
2138
2139 // See if these Phis have been made before.
2140 // Register with optimizer
2141 Node *hit1 = _igvn.hash_find_insert(phi1);
2142 if( hit1 ) { // Hit, toss just made Phi
2143 _igvn.remove_dead_node(phi1); // Remove new phi
2144 assert( hit1->is_Phi(), "" );
2145 phi1 = (PhiNode*)hit1; // Use existing phi
2146 } else { // Miss
2147 _igvn.register_new_node_with_optimizer(phi1);
2148 }
2149 Node *hit2 = _igvn.hash_find_insert(phi2);
2150 if( hit2 ) { // Hit, toss just made Phi
2151 _igvn.remove_dead_node(phi2); // Remove new phi
2152 assert( hit2->is_Phi(), "" );
2153 phi2 = (PhiNode*)hit2; // Use existing phi
2154 } else { // Miss
2155 _igvn.register_new_node_with_optimizer(phi2);
2156 }
2157 // Register Phis with loop/block info
2158 set_ctrl(phi1, phi->in(0));
2159 set_ctrl(phi2, phi->in(0));
2160 // Make a new Cmp
2161 Node *cmp = sample_cmp->clone();
2162 cmp->set_req( 1, phi1 );
2163 cmp->set_req( 2, phi2 );
2164 _igvn.register_new_node_with_optimizer(cmp);
2165 set_ctrl(cmp, phi->in(0));
2166
2167 assert( cmp->is_Cmp(), "" );
2168 return (CmpNode*)cmp;
2169 }
2170
2171 //------------------------------sink_use---------------------------------------
2172 // If 'use' was in the loop-exit block, it now needs to be sunk
2173 // below the post-loop merge point.
2174 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) {
2175 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) {
2176 set_ctrl(use, post_loop);
2177 for (DUIterator j = use->outs(); use->has_out(j); j++)
2178 sink_use(use->out(j), post_loop);
2179 }
2180 }
2181
2182 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new,
2183 IdealLoopTree* loop, IdealLoopTree* outer_loop,
2184 Node_List*& split_if_set, Node_List*& split_bool_set,
2185 Node_List*& split_cex_set, Node_List& worklist,
2186 uint new_counter, CloneLoopMode mode) {
2187 Node* nnn = old_new[old->_idx];
2188 // Copy uses to a worklist, so I can munge the def-use info
2189 // with impunity.
2190 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
2191 worklist.push(old->fast_out(j));
2192
2193 while( worklist.size() ) {
2194 Node *use = worklist.pop();
2195 if (!has_node(use)) continue; // Ignore dead nodes
2196 if (use->in(0) == C->top()) continue;
2197 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
2198 // Check for data-use outside of loop - at least one of OLD or USE
2199 // must not be a CFG node.
2200 #ifdef ASSERT
2201 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == NULL) {
2202 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint();
2203 assert(mode != IgnoreStripMined, "incorrect cloning mode");
2204 assert((mode == ControlAroundStripMined && use == sfpt) || !use->is_reachable_from_root(), "missed a node");
2205 }
2206 #endif
2207 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) {
2208
2209 // If the Data use is an IF, that means we have an IF outside of the
2210 // loop that is switching on a condition that is set inside of the
2211 // loop. Happens if people set a loop-exit flag; then test the flag
2212 // in the loop to break the loop, then test is again outside of the
2213 // loop to determine which way the loop exited.
2214 // Loop predicate If node connects to Bool node through Opaque1 node.
2215 if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use) || use->Opcode() == Op_Opaque4) {
2216 // Since this code is highly unlikely, we lazily build the worklist
2217 // of such Nodes to go split.
2218 if (!split_if_set) {
2219 split_if_set = new Node_List();
2220 }
2221 split_if_set->push(use);
2222 }
2223 if (use->is_Bool()) {
2224 if (!split_bool_set) {
2225 split_bool_set = new Node_List();
2226 }
2227 split_bool_set->push(use);
2228 }
2229 if (use->Opcode() == Op_CreateEx) {
2230 if (!split_cex_set) {
2231 split_cex_set = new Node_List();
2232 }
2233 split_cex_set->push(use);
2234 }
2235
2236
2237 // Get "block" use is in
2238 uint idx = 0;
2239 while( use->in(idx) != old ) idx++;
2240 Node *prev = use->is_CFG() ? use : get_ctrl(use);
2241 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" );
2242 Node *cfg = prev->_idx >= new_counter
2243 ? prev->in(2)
2244 : idom(prev);
2245 if( use->is_Phi() ) // Phi use is in prior block
2246 cfg = prev->in(idx); // NOT in block of Phi itself
2247 if (cfg->is_top()) { // Use is dead?
2248 _igvn.replace_input_of(use, idx, C->top());
2249 continue;
2250 }
2251
2252 // If use is referenced through control edge... (idx == 0)
2253 if (mode == IgnoreStripMined && idx == 0) {
2254 LoopNode *head = loop->_head->as_Loop();
2255 if (head->is_strip_mined() && is_dominator(head->outer_loop_exit(), prev)) {
2256 // That node is outside the inner loop, leave it outside the
2257 // outer loop as well to not confuse verification code.
2258 assert(!loop->_parent->is_member(use_loop), "should be out of the outer loop");
2259 _igvn.replace_input_of(use, 0, head->outer_loop_exit());
2260 continue;
2261 }
2262 }
2263
2264 while(!outer_loop->is_member(get_loop(cfg))) {
2265 prev = cfg;
2266 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg);
2267 }
2268 // If the use occurs after merging several exits from the loop, then
2269 // old value must have dominated all those exits. Since the same old
2270 // value was used on all those exits we did not need a Phi at this
2271 // merge point. NOW we do need a Phi here. Each loop exit value
2272 // is now merged with the peeled body exit; each exit gets its own
2273 // private Phi and those Phis need to be merged here.
2274 Node *phi;
2275 if( prev->is_Region() ) {
2276 if( idx == 0 ) { // Updating control edge?
2277 phi = prev; // Just use existing control
2278 } else { // Else need a new Phi
2279 phi = PhiNode::make( prev, old );
2280 // Now recursively fix up the new uses of old!
2281 for( uint i = 1; i < prev->req(); i++ ) {
2282 worklist.push(phi); // Onto worklist once for each 'old' input
2283 }
2284 }
2285 } else {
2286 // Get new RegionNode merging old and new loop exits
2287 prev = old_new[prev->_idx];
2288 assert( prev, "just made this in step 7" );
2289 if( idx == 0) { // Updating control edge?
2290 phi = prev; // Just use existing control
2291 } else { // Else need a new Phi
2292 // Make a new Phi merging data values properly
2293 phi = PhiNode::make( prev, old );
2294 phi->set_req( 1, nnn );
2295 }
2296 }
2297 // If inserting a new Phi, check for prior hits
2298 if( idx != 0 ) {
2299 Node *hit = _igvn.hash_find_insert(phi);
2300 if( hit == NULL ) {
2301 _igvn.register_new_node_with_optimizer(phi); // Register new phi
2302 } else { // or
2303 // Remove the new phi from the graph and use the hit
2304 _igvn.remove_dead_node(phi);
2305 phi = hit; // Use existing phi
2306 }
2307 set_ctrl(phi, prev);
2308 }
2309 // Make 'use' use the Phi instead of the old loop body exit value
2310 _igvn.replace_input_of(use, idx, phi);
2311 if( use->_idx >= new_counter ) { // If updating new phis
2312 // Not needed for correctness, but prevents a weak assert
2313 // in AddPNode from tripping (when we end up with different
2314 // base & derived Phis that will become the same after
2315 // IGVN does CSE).
2316 Node *hit = _igvn.hash_find_insert(use);
2317 if( hit ) // Go ahead and re-hash for hits.
2318 _igvn.replace_node( use, hit );
2319 }
2320
2321 // If 'use' was in the loop-exit block, it now needs to be sunk
2322 // below the post-loop merge point.
2323 sink_use( use, prev );
2324 }
2325 }
2326 }
2327
2328 static void clone_outer_loop_helper(Node* n, const IdealLoopTree *loop, const IdealLoopTree* outer_loop,
2329 const Node_List &old_new, Unique_Node_List& wq, PhaseIdealLoop* phase,
2330 bool check_old_new) {
2331 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2332 Node* u = n->fast_out(j);
2333 assert(check_old_new || old_new[u->_idx] == NULL, "shouldn't have been cloned");
2334 if (!u->is_CFG() && (!check_old_new || old_new[u->_idx] == NULL)) {
2335 Node* c = phase->get_ctrl(u);
2336 IdealLoopTree* u_loop = phase->get_loop(c);
2337 assert(!loop->is_member(u_loop), "can be in outer loop or out of both loops only");
2338 if (outer_loop->is_member(u_loop)) {
2339 wq.push(u);
2340 } else {
2341 // nodes pinned with control in the outer loop but not referenced from the safepoint must be moved out of
2342 // the outer loop too
2343 Node* u_c = u->in(0);
2344 if (u_c != NULL) {
2345 IdealLoopTree* u_c_loop = phase->get_loop(u_c);
2346 if (outer_loop->is_member(u_c_loop) && !loop->is_member(u_c_loop)) {
2347 wq.push(u);
2348 }
2349 }
2350 }
2351 }
2352 }
2353 }
2354
2355 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
2356 IdealLoopTree* outer_loop, int dd, Node_List &old_new,
2357 Node_List& extra_data_nodes) {
2358 if (head->is_strip_mined() && mode != IgnoreStripMined) {
2359 CountedLoopNode* cl = head->as_CountedLoop();
2360 Node* l = cl->outer_loop();
2361 Node* tail = cl->outer_loop_tail();
2362 IfNode* le = cl->outer_loop_end();
2363 Node* sfpt = cl->outer_safepoint();
2364 CountedLoopEndNode* cle = cl->loopexit();
2365 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop();
2366 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null();
2367 Node* cle_out = cle->proj_out(false);
2368
2369 Node* new_sfpt = NULL;
2370 Node* new_cle_out = cle_out->clone();
2371 old_new.map(cle_out->_idx, new_cle_out);
2372 if (mode == CloneIncludesStripMined) {
2373 // clone outer loop body
2374 Node* new_l = l->clone();
2375 Node* new_tail = tail->clone();
2376 IfNode* new_le = le->clone()->as_If();
2377 new_sfpt = sfpt->clone();
2378
2379 set_loop(new_l, outer_loop->_parent);
2380 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd);
2381 set_loop(new_cle_out, outer_loop->_parent);
2382 set_idom(new_cle_out, new_cle, dd);
2383 set_loop(new_sfpt, outer_loop->_parent);
2384 set_idom(new_sfpt, new_cle_out, dd);
2385 set_loop(new_le, outer_loop->_parent);
2386 set_idom(new_le, new_sfpt, dd);
2387 set_loop(new_tail, outer_loop->_parent);
2388 set_idom(new_tail, new_le, dd);
2389 set_idom(new_cl, new_l, dd);
2390
2391 old_new.map(l->_idx, new_l);
2392 old_new.map(tail->_idx, new_tail);
2393 old_new.map(le->_idx, new_le);
2394 old_new.map(sfpt->_idx, new_sfpt);
2395
2396 new_l->set_req(LoopNode::LoopBackControl, new_tail);
2397 new_l->set_req(0, new_l);
2398 new_tail->set_req(0, new_le);
2399 new_le->set_req(0, new_sfpt);
2400 new_sfpt->set_req(0, new_cle_out);
2401 new_cle_out->set_req(0, new_cle);
2402 new_cl->set_req(LoopNode::EntryControl, new_l);
2403
2404 _igvn.register_new_node_with_optimizer(new_l);
2405 _igvn.register_new_node_with_optimizer(new_tail);
2406 _igvn.register_new_node_with_optimizer(new_le);
2407 } else {
2408 Node *newhead = old_new[loop->_head->_idx];
2409 newhead->as_Loop()->clear_strip_mined();
2410 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl));
2411 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
2412 }
2413 // Look at data node that were assigned a control in the outer
2414 // loop: they are kept in the outer loop by the safepoint so start
2415 // from the safepoint node's inputs.
2416 IdealLoopTree* outer_loop = get_loop(l);
2417 Node_Stack stack(2);
2418 stack.push(sfpt, 1);
2419 uint new_counter = C->unique();
2420 while (stack.size() > 0) {
2421 Node* n = stack.node();
2422 uint i = stack.index();
2423 while (i < n->req() &&
2424 (n->in(i) == NULL ||
2425 !has_ctrl(n->in(i)) ||
2426 get_loop(get_ctrl(n->in(i))) != outer_loop ||
2427 (old_new[n->in(i)->_idx] != NULL && old_new[n->in(i)->_idx]->_idx >= new_counter))) {
2428 i++;
2429 }
2430 if (i < n->req()) {
2431 stack.set_index(i+1);
2432 stack.push(n->in(i), 0);
2433 } else {
2434 assert(old_new[n->_idx] == NULL || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet");
2435 Node* m = n == sfpt ? new_sfpt : n->clone();
2436 if (m != NULL) {
2437 for (uint i = 0; i < n->req(); i++) {
2438 if (m->in(i) != NULL && old_new[m->in(i)->_idx] != NULL) {
2439 m->set_req(i, old_new[m->in(i)->_idx]);
2440 }
2441 }
2442 } else {
2443 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?");
2444 }
2445 if (n != sfpt) {
2446 extra_data_nodes.push(n);
2447 _igvn.register_new_node_with_optimizer(m);
2448 assert(get_ctrl(n) == cle_out, "what other control?");
2449 set_ctrl(m, new_cle_out);
2450 old_new.map(n->_idx, m);
2451 }
2452 stack.pop();
2453 }
2454 }
2455 if (mode == CloneIncludesStripMined) {
2456 _igvn.register_new_node_with_optimizer(new_sfpt);
2457 _igvn.register_new_node_with_optimizer(new_cle_out);
2458 }
2459 // Some other transformation may have pessimistically assigned some
2460 // data nodes to the outer loop. Set their control so they are out
2461 // of the outer loop.
2462 ResourceMark rm;
2463 Unique_Node_List wq;
2464 for (uint i = 0; i < extra_data_nodes.size(); i++) {
2465 Node* old = extra_data_nodes.at(i);
2466 clone_outer_loop_helper(old, loop, outer_loop, old_new, wq, this, true);
2467 }
2468
2469 Node* inner_out = sfpt->in(0);
2470 if (inner_out->outcnt() > 1) {
2471 clone_outer_loop_helper(inner_out, loop, outer_loop, old_new, wq, this, true);
2472 }
2473
2474 Node* new_ctrl = cl->outer_loop_exit();
2475 assert(get_loop(new_ctrl) != outer_loop, "must be out of the loop nest");
2476 for (uint i = 0; i < wq.size(); i++) {
2477 Node* n = wq.at(i);
2478 set_ctrl(n, new_ctrl);
2479 if (n->in(0) != NULL) {
2480 _igvn.replace_input_of(n, 0, new_ctrl);
2481 }
2482 clone_outer_loop_helper(n, loop, outer_loop, old_new, wq, this, false);
2483 }
2484 } else {
2485 Node *newhead = old_new[loop->_head->_idx];
2486 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
2487 }
2488 }
2489
2490 //------------------------------clone_loop-------------------------------------
2491 //
2492 // C L O N E A L O O P B O D Y
2493 //
2494 // This is the basic building block of the loop optimizations. It clones an
2495 // entire loop body. It makes an old_new loop body mapping; with this mapping
2496 // you can find the new-loop equivalent to an old-loop node. All new-loop
2497 // nodes are exactly equal to their old-loop counterparts, all edges are the
2498 // same. All exits from the old-loop now have a RegionNode that merges the
2499 // equivalent new-loop path. This is true even for the normal "loop-exit"
2500 // condition. All uses of loop-invariant old-loop values now come from (one
2501 // or more) Phis that merge their new-loop equivalents.
2502 //
2503 // This operation leaves the graph in an illegal state: there are two valid
2504 // control edges coming from the loop pre-header to both loop bodies. I'll
2505 // definitely have to hack the graph after running this transform.
2506 //
2507 // From this building block I will further edit edges to perform loop peeling
2508 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
2509 //
2510 // Parameter side_by_size_idom:
2511 // When side_by_size_idom is NULL, the dominator tree is constructed for
2512 // the clone loop to dominate the original. Used in construction of
2513 // pre-main-post loop sequence.
2514 // When nonnull, the clone and original are side-by-side, both are
2515 // dominated by the side_by_side_idom node. Used in construction of
2516 // unswitched loops.
2517 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
2518 CloneLoopMode mode, Node* side_by_side_idom) {
2519
2520 LoopNode* head = loop->_head->as_Loop();
2521 head->verify_strip_mined(1);
2522
2523 if (C->do_vector_loop() && PrintOpto) {
2524 const char* mname = C->method()->name()->as_quoted_ascii();
2525 if (mname != NULL) {
2526 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname);
2527 }
2528 }
2529
2530 CloneMap& cm = C->clone_map();
2531 if (C->do_vector_loop()) {
2532 cm.set_clone_idx(cm.max_gen()+1);
2533 #ifndef PRODUCT
2534 if (PrintOpto) {
2535 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx());
2536 loop->dump_head();
2537 }
2538 #endif
2539 }
2540
2541 // Step 1: Clone the loop body. Make the old->new mapping.
2542 clone_loop_body(loop->_body, old_new, &cm);
2543
2544 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop;
2545
2546 // Step 2: Fix the edges in the new body. If the old input is outside the
2547 // loop use it. If the old input is INside the loop, use the corresponding
2548 // new node instead.
2549 fix_body_edges(loop->_body, loop, old_new, dd, outer_loop->_parent, false);
2550
2551 Node_List extra_data_nodes; // data nodes in the outer strip mined loop
2552 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes);
2553
2554 // Step 3: Now fix control uses. Loop varying control uses have already
2555 // been fixed up (as part of all input edges in Step 2). Loop invariant
2556 // control uses must be either an IfFalse or an IfTrue. Make a merge
2557 // point to merge the old and new IfFalse/IfTrue nodes; make the use
2558 // refer to this.
2559 Node_List worklist;
2560 uint new_counter = C->unique();
2561 fix_ctrl_uses(loop->_body, loop, old_new, mode, side_by_side_idom, &cm, worklist);
2562
2563 // Step 4: If loop-invariant use is not control, it must be dominated by a
2564 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region
2565 // there if needed. Make a Phi there merging old and new used values.
2566 Node_List *split_if_set = NULL;
2567 Node_List *split_bool_set = NULL;
2568 Node_List *split_cex_set = NULL;
2569 fix_data_uses(loop->_body, loop, mode, outer_loop, new_counter, old_new, worklist, split_if_set, split_bool_set, split_cex_set);
2570
2571 for (uint i = 0; i < extra_data_nodes.size(); i++) {
2572 Node* old = extra_data_nodes.at(i);
2573 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2574 split_bool_set, split_cex_set, worklist, new_counter,
2575 mode);
2576 }
2577
2578 // Check for IFs that need splitting/cloning. Happens if an IF outside of
2579 // the loop uses a condition set in the loop. The original IF probably
2580 // takes control from one or more OLD Regions (which in turn get from NEW
2581 // Regions). In any case, there will be a set of Phis for each merge point
2582 // from the IF up to where the original BOOL def exists the loop.
2583 finish_clone_loop(split_if_set, split_bool_set, split_cex_set);
2584
2585 }
2586
2587 void PhaseIdealLoop::finish_clone_loop(Node_List* split_if_set, Node_List* split_bool_set, Node_List* split_cex_set) {
2588 if (split_if_set) {
2589 while (split_if_set->size()) {
2590 Node *iff = split_if_set->pop();
2591 if (iff->in(1)->is_Phi()) {
2592 Node *b = clone_iff(iff->in(1)->as_Phi());
2593 _igvn.replace_input_of(iff, 1, b);
2594 }
2595 }
2596 }
2597 if (split_bool_set) {
2598 while (split_bool_set->size()) {
2599 Node *b = split_bool_set->pop();
2600 Node *phi = b->in(1);
2601 assert(phi->is_Phi(), "");
2602 CmpNode *cmp = clone_bool((PhiNode*) phi);
2603 _igvn.replace_input_of(b, 1, cmp);
2604 }
2605 }
2606 if (split_cex_set) {
2607 while (split_cex_set->size()) {
2608 Node *b = split_cex_set->pop();
2609 assert(b->in(0)->is_Region(), "");
2610 assert(b->in(1)->is_Phi(), "");
2611 assert(b->in(0)->in(0) == b->in(1)->in(0), "");
2612 split_up(b, b->in(0), NULL);
2613 }
2614 }
2615 }
2616
2617 void PhaseIdealLoop::fix_data_uses(Node_List& body, IdealLoopTree* loop, CloneLoopMode mode, IdealLoopTree* outer_loop,
2618 uint new_counter, Node_List &old_new, Node_List &worklist, Node_List*& split_if_set,
2619 Node_List*& split_bool_set, Node_List*& split_cex_set) {
2620 for(uint i = 0; i < body.size(); i++ ) {
2621 Node* old = body.at(i);
2622 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2623 split_bool_set, split_cex_set, worklist, new_counter,
2624 mode);
2625 }
2626 }
2627
2628 void PhaseIdealLoop::fix_ctrl_uses(const Node_List& body, const IdealLoopTree* loop, Node_List &old_new, CloneLoopMode mode,
2629 Node* side_by_side_idom, CloneMap* cm, Node_List &worklist) {
2630 LoopNode* head = loop->_head->as_Loop();
2631 for(uint i = 0; i < body.size(); i++ ) {
2632 Node* old = body.at(i);
2633 if( !old->is_CFG() ) continue;
2634
2635 // Copy uses to a worklist, so I can munge the def-use info
2636 // with impunity.
2637 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) {
2638 worklist.push(old->fast_out(j));
2639 }
2640
2641 while (worklist.size()) { // Visit all uses
2642 Node *use = worklist.pop();
2643 if (!has_node(use)) continue; // Ignore dead nodes
2644 IdealLoopTree *use_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use );
2645 if (!loop->is_member(use_loop) && use->is_CFG()) {
2646 // Both OLD and USE are CFG nodes here.
2647 assert(use->is_Proj(), "" );
2648 Node* nnn = old_new[old->_idx];
2649
2650 Node* newuse = NULL;
2651 if (head->is_strip_mined() && mode != IgnoreStripMined) {
2652 CountedLoopNode* cl = head->as_CountedLoop();
2653 CountedLoopEndNode* cle = cl->loopexit();
2654 Node* cle_out = cle->proj_out_or_null(false);
2655 if (use == cle_out) {
2656 IfNode* le = cl->outer_loop_end();
2657 use = le->proj_out(false);
2658 use_loop = get_loop(use);
2659 if (mode == CloneIncludesStripMined) {
2660 nnn = old_new[le->_idx];
2661 } else {
2662 newuse = old_new[cle_out->_idx];
2663 }
2664 }
2665 }
2666 if (newuse == NULL) {
2667 newuse = use->clone();
2668 }
2669
2670 // Clone the loop exit control projection
2671 if (C->do_vector_loop() && cm != NULL) {
2672 cm->verify_insert_and_clone(use, newuse, cm->clone_idx());
2673 }
2674 newuse->set_req(0,nnn);
2675 _igvn.register_new_node_with_optimizer(newuse);
2676 set_loop(newuse, use_loop);
2677 set_idom(newuse, nnn, dom_depth(nnn) + 1 );
2678
2679 // We need a Region to merge the exit from the peeled body and the
2680 // exit from the old loop body.
2681 RegionNode *r = new RegionNode(3);
2682 // Map the old use to the new merge point
2683 old_new.map( use->_idx, r );
2684 uint dd_r = MIN2(dom_depth(newuse), dom_depth(use));
2685 assert(dd_r >= dom_depth(dom_lca(newuse, use)), "" );
2686
2687 // The original user of 'use' uses 'r' instead.
2688 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
2689 Node* useuse = use->last_out(l);
2690 _igvn.rehash_node_delayed(useuse);
2691 uint uses_found = 0;
2692 if (useuse->in(0) == use) {
2693 useuse->set_req(0, r);
2694 uses_found++;
2695 if (useuse->is_CFG()) {
2696 // This is not a dom_depth > dd_r because when new
2697 // control flow is constructed by a loop opt, a node and
2698 // its dominator can end up at the same dom_depth
2699 assert(dom_depth(useuse) >= dd_r, "");
2700 set_idom(useuse, r, dom_depth(useuse));
2701 }
2702 }
2703 for (uint k = 1; k < useuse->req(); k++) {
2704 if( useuse->in(k) == use ) {
2705 useuse->set_req(k, r);
2706 uses_found++;
2707 if (useuse->is_Loop() && k == LoopNode::EntryControl) {
2708 // This is not a dom_depth > dd_r because when new
2709 // control flow is constructed by a loop opt, a node
2710 // and its dominator can end up at the same dom_depth
2711 assert(dom_depth(useuse) >= dd_r , "");
2712 set_idom(useuse, r, dom_depth(useuse));
2713 }
2714 }
2715 }
2716 l -= uses_found; // we deleted 1 or more copies of this edge
2717 }
2718
2719 // Now finish up 'r'
2720 r->set_req(1, newuse);
2721 r->set_req(2, use);
2722 _igvn.register_new_node_with_optimizer(r);
2723 set_loop(r, use_loop);
2724 set_idom(r, (side_by_side_idom == NULL) ? newuse->in(0) : side_by_side_idom, dd_r);
2725 } // End of if a loop-exit test
2726 }
2727 }
2728 }
2729
2730 void PhaseIdealLoop::fix_body_edges(const Node_List &body, IdealLoopTree* loop, const Node_List &old_new, int dd,
2731 IdealLoopTree* parent, bool partial) {
2732 for(uint i = 0; i < body.size(); i++ ) {
2733 Node *old = body.at(i);
2734 Node *nnn = old_new[old->_idx];
2735 // Fix CFG/Loop controlling the new node
2736 if (has_ctrl(old)) {
2737 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
2738 } else {
2739 set_loop(nnn, parent);
2740 if (old->outcnt() > 0) {
2741 Node* dom = idom(old);
2742 if (old_new[dom->_idx] != NULL) {
2743 dom = old_new[dom->_idx];
2744 set_idom(nnn, dom, dd );
2745 }
2746 }
2747 }
2748 // Correct edges to the new node
2749 for (uint j = 0; j < nnn->req(); j++) {
2750 Node *n = nnn->in(j);
2751 if (n != NULL) {
2752 IdealLoopTree *old_in_loop = get_loop(has_ctrl(n) ? get_ctrl(n) : n);
2753 if (loop->is_member(old_in_loop)) {
2754 if (old_new[n->_idx] != NULL) {
2755 nnn->set_req(j, old_new[n->_idx]);
2756 } else {
2757 assert(!body.contains(n), "");
2758 assert(partial, "node not cloned");
2759 }
2760 }
2761 }
2762 }
2763 _igvn.hash_find_insert(nnn);
2764 }
2765 }
2766
2767 void PhaseIdealLoop::clone_loop_body(const Node_List& body, Node_List &old_new, CloneMap* cm) {
2768 for (uint i = 0; i < body.size(); i++) {
2769 Node* old = body.at(i);
2770 Node* nnn = old->clone();
2771 old_new.map(old->_idx, nnn);
2772 if (old->is_reduction()) {
2773 // Reduction flag is not copied by default. Copy it here when cloning the entire loop body.
2774 nnn->add_flag(Node::Flag_is_reduction);
2775 }
2776 if (C->do_vector_loop() && cm != NULL) {
2777 cm->verify_insert_and_clone(old, nnn, cm->clone_idx());
2778 }
2779 _igvn.register_new_node_with_optimizer(nnn);
2780 }
2781 }
2782
2783
2784 //---------------------- stride_of_possible_iv -------------------------------------
2785 // Looks for an iff/bool/comp with one operand of the compare
2786 // being a cycle involving an add and a phi,
2787 // with an optional truncation (left-shift followed by a right-shift)
2788 // of the add. Returns zero if not an iv.
2789 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
2790 Node* trunc1 = NULL;
2791 Node* trunc2 = NULL;
2792 const TypeInteger* ttype = NULL;
2793 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) {
2794 return 0;
2795 }
2796 BoolNode* bl = iff->in(1)->as_Bool();
2797 Node* cmp = bl->in(1);
2798 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) {
2799 return 0;
2800 }
2801 // Must have an invariant operand
2802 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) {
2803 return 0;
2804 }
2805 Node* add2 = NULL;
2806 Node* cmp1 = cmp->in(1);
2807 if (cmp1->is_Phi()) {
2808 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
2809 Node* phi = cmp1;
2810 for (uint i = 1; i < phi->req(); i++) {
2811 Node* in = phi->in(i);
2812 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in,
2813 &trunc1, &trunc2, &ttype, T_INT);
2814 if (add && add->in(1) == phi) {
2815 add2 = add->in(2);
2816 break;
2817 }
2818 }
2819 } else {
2820 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
2821 Node* addtrunc = cmp1;
2822 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc,
2823 &trunc1, &trunc2, &ttype, T_INT);
2824 if (add && add->in(1)->is_Phi()) {
2825 Node* phi = add->in(1);
2826 for (uint i = 1; i < phi->req(); i++) {
2827 if (phi->in(i) == addtrunc) {
2828 add2 = add->in(2);
2829 break;
2830 }
2831 }
2832 }
2833 }
2834 if (add2 != NULL) {
2835 const TypeInt* add2t = _igvn.type(add2)->is_int();
2836 if (add2t->is_con()) {
2837 return add2t->get_con();
2838 }
2839 }
2840 return 0;
2841 }
2842
2843
2844 //---------------------- stay_in_loop -------------------------------------
2845 // Return the (unique) control output node that's in the loop (if it exists.)
2846 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
2847 Node* unique = NULL;
2848 if (!n) return NULL;
2849 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2850 Node* use = n->fast_out(i);
2851 if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
2852 if (unique != NULL) {
2853 return NULL;
2854 }
2855 unique = use;
2856 }
2857 }
2858 return unique;
2859 }
2860
2861 //------------------------------ register_node -------------------------------------
2862 // Utility to register node "n" with PhaseIdealLoop
2863 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) {
2864 _igvn.register_new_node_with_optimizer(n);
2865 loop->_body.push(n);
2866 if (n->is_CFG()) {
2867 set_loop(n, loop);
2868 set_idom(n, pred, ddepth);
2869 } else {
2870 set_ctrl(n, pred);
2871 }
2872 }
2873
2874 //------------------------------ proj_clone -------------------------------------
2875 // Utility to create an if-projection
2876 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
2877 ProjNode* c = p->clone()->as_Proj();
2878 c->set_req(0, iff);
2879 return c;
2880 }
2881
2882 //------------------------------ short_circuit_if -------------------------------------
2883 // Force the iff control output to be the live_proj
2884 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
2885 guarantee(live_proj != NULL, "null projection");
2886 int proj_con = live_proj->_con;
2887 assert(proj_con == 0 || proj_con == 1, "false or true projection");
2888 Node *con = _igvn.intcon(proj_con);
2889 set_ctrl(con, C->root());
2890 if (iff) {
2891 iff->set_req(1, con);
2892 }
2893 return con;
2894 }
2895
2896 //------------------------------ insert_if_before_proj -------------------------------------
2897 // Insert a new if before an if projection (* - new node)
2898 //
2899 // before
2900 // if(test)
2901 // / \
2902 // v v
2903 // other-proj proj (arg)
2904 //
2905 // after
2906 // if(test)
2907 // / \
2908 // / v
2909 // | * proj-clone
2910 // v |
2911 // other-proj v
2912 // * new_if(relop(cmp[IU](left,right)))
2913 // / \
2914 // v v
2915 // * new-proj proj
2916 // (returned)
2917 //
2918 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) {
2919 IfNode* iff = proj->in(0)->as_If();
2920 IdealLoopTree *loop = get_loop(proj);
2921 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2922 int ddepth = dom_depth(proj);
2923
2924 _igvn.rehash_node_delayed(iff);
2925 _igvn.rehash_node_delayed(proj);
2926
2927 proj->set_req(0, NULL); // temporary disconnect
2928 ProjNode* proj2 = proj_clone(proj, iff);
2929 register_node(proj2, loop, iff, ddepth);
2930
2931 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right);
2932 register_node(cmp, loop, proj2, ddepth);
2933
2934 BoolNode* bol = new BoolNode(cmp, relop);
2935 register_node(bol, loop, proj2, ddepth);
2936
2937 int opcode = iff->Opcode();
2938 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode");
2939 IfNode* new_if = (opcode == Op_If) ? new IfNode(proj2, bol, iff->_prob, iff->_fcnt):
2940 new RangeCheckNode(proj2, bol, iff->_prob, iff->_fcnt);
2941 register_node(new_if, loop, proj2, ddepth);
2942
2943 proj->set_req(0, new_if); // reattach
2944 set_idom(proj, new_if, ddepth);
2945
2946 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
2947 guarantee(new_exit != NULL, "null exit node");
2948 register_node(new_exit, get_loop(other_proj), new_if, ddepth);
2949
2950 return new_exit;
2951 }
2952
2953 //------------------------------ insert_region_before_proj -------------------------------------
2954 // Insert a region before an if projection (* - new node)
2955 //
2956 // before
2957 // if(test)
2958 // / |
2959 // v |
2960 // proj v
2961 // other-proj
2962 //
2963 // after
2964 // if(test)
2965 // / |
2966 // v |
2967 // * proj-clone v
2968 // | other-proj
2969 // v
2970 // * new-region
2971 // |
2972 // v
2973 // * dum_if
2974 // / \
2975 // v \
2976 // * dum-proj v
2977 // proj
2978 //
2979 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
2980 IfNode* iff = proj->in(0)->as_If();
2981 IdealLoopTree *loop = get_loop(proj);
2982 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2983 int ddepth = dom_depth(proj);
2984
2985 _igvn.rehash_node_delayed(iff);
2986 _igvn.rehash_node_delayed(proj);
2987
2988 proj->set_req(0, NULL); // temporary disconnect
2989 ProjNode* proj2 = proj_clone(proj, iff);
2990 register_node(proj2, loop, iff, ddepth);
2991
2992 RegionNode* reg = new RegionNode(2);
2993 reg->set_req(1, proj2);
2994 register_node(reg, loop, iff, ddepth);
2995
2996 IfNode* dum_if = new IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt);
2997 register_node(dum_if, loop, reg, ddepth);
2998
2999 proj->set_req(0, dum_if); // reattach
3000 set_idom(proj, dum_if, ddepth);
3001
3002 ProjNode* dum_proj = proj_clone(other_proj, dum_if);
3003 register_node(dum_proj, loop, dum_if, ddepth);
3004
3005 return reg;
3006 }
3007
3008 //------------------------------ insert_cmpi_loop_exit -------------------------------------
3009 // Clone a signed compare loop exit from an unsigned compare and
3010 // insert it before the unsigned cmp on the stay-in-loop path.
3011 // All new nodes inserted in the dominator tree between the original
3012 // if and it's projections. The original if test is replaced with
3013 // a constant to force the stay-in-loop path.
3014 //
3015 // This is done to make sure that the original if and it's projections
3016 // still dominate the same set of control nodes, that the ctrl() relation
3017 // from data nodes to them is preserved, and that their loop nesting is
3018 // preserved.
3019 //
3020 // before
3021 // if(i <u limit) unsigned compare loop exit
3022 // / |
3023 // v v
3024 // exit-proj stay-in-loop-proj
3025 //
3026 // after
3027 // if(stay-in-loop-const) original if
3028 // / |
3029 // / v
3030 // / if(i < limit) new signed test
3031 // / / |
3032 // / / v
3033 // / / if(i <u limit) new cloned unsigned test
3034 // / / / |
3035 // v v v |
3036 // region |
3037 // | |
3038 // dum-if |
3039 // / | |
3040 // ether | |
3041 // v v
3042 // exit-proj stay-in-loop-proj
3043 //
3044 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) {
3045 const bool Signed = true;
3046 const bool Unsigned = false;
3047
3048 BoolNode* bol = if_cmpu->in(1)->as_Bool();
3049 if (bol->_test._test != BoolTest::lt) return NULL;
3050 CmpNode* cmpu = bol->in(1)->as_Cmp();
3051 if (cmpu->Opcode() != Op_CmpU) return NULL;
3052 int stride = stride_of_possible_iv(if_cmpu);
3053 if (stride == 0) return NULL;
3054
3055 Node* lp_proj = stay_in_loop(if_cmpu, loop);
3056 guarantee(lp_proj != NULL, "null loop node");
3057
3058 ProjNode* lp_continue = lp_proj->as_Proj();
3059 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
3060 if (!lp_exit->is_IfFalse()) {
3061 // The loop exit condition is (i <u limit) ==> (i >= 0 && i < limit).
3062 // We therefore can't add a single exit condition.
3063 return NULL;
3064 }
3065 // The loop exit condition is !(i <u limit) ==> (i < 0 || i >= limit).
3066 // Split out the exit condition (i < 0) for stride < 0 or (i >= limit) for stride > 0.
3067 Node* limit = NULL;
3068 if (stride > 0) {
3069 limit = cmpu->in(2);
3070 } else {
3071 limit = _igvn.makecon(TypeInt::ZERO);
3072 set_ctrl(limit, C->root());
3073 }
3074 // Create a new region on the exit path
3075 RegionNode* reg = insert_region_before_proj(lp_exit);
3076 guarantee(reg != NULL, "null region node");
3077
3078 // Clone the if-cmpu-true-false using a signed compare
3079 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
3080 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue);
3081 reg->add_req(cmpi_exit);
3082
3083 // Clone the if-cmpu-true-false
3084 BoolTest::mask rel_u = bol->_test._test;
3085 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
3086 reg->add_req(cmpu_exit);
3087
3088 // Force original if to stay in loop.
3089 short_circuit_if(if_cmpu, lp_continue);
3090
3091 return cmpi_exit->in(0)->as_If();
3092 }
3093
3094 //------------------------------ remove_cmpi_loop_exit -------------------------------------
3095 // Remove a previously inserted signed compare loop exit.
3096 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
3097 Node* lp_proj = stay_in_loop(if_cmp, loop);
3098 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
3099 stay_in_loop(lp_proj, loop)->is_If() &&
3100 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
3101 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
3102 set_ctrl(con, C->root());
3103 if_cmp->set_req(1, con);
3104 }
3105
3106 //------------------------------ scheduled_nodelist -------------------------------------
3107 // Create a post order schedule of nodes that are in the
3108 // "member" set. The list is returned in "sched".
3109 // The first node in "sched" is the loop head, followed by
3110 // nodes which have no inputs in the "member" set, and then
3111 // followed by the nodes that have an immediate input dependence
3112 // on a node in "sched".
3113 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
3114
3115 assert(member.test(loop->_head->_idx), "loop head must be in member set");
3116 VectorSet visited;
3117 Node_Stack nstack(loop->_body.size());
3118
3119 Node* n = loop->_head; // top of stack is cached in "n"
3120 uint idx = 0;
3121 visited.set(n->_idx);
3122
3123 // Initially push all with no inputs from within member set
3124 for(uint i = 0; i < loop->_body.size(); i++ ) {
3125 Node *elt = loop->_body.at(i);
3126 if (member.test(elt->_idx)) {
3127 bool found = false;
3128 for (uint j = 0; j < elt->req(); j++) {
3129 Node* def = elt->in(j);
3130 if (def && member.test(def->_idx) && def != elt) {
3131 found = true;
3132 break;
3133 }
3134 }
3135 if (!found && elt != loop->_head) {
3136 nstack.push(n, idx);
3137 n = elt;
3138 assert(!visited.test(n->_idx), "not seen yet");
3139 visited.set(n->_idx);
3140 }
3141 }
3142 }
3143
3144 // traverse out's that are in the member set
3145 while (true) {
3146 if (idx < n->outcnt()) {
3147 Node* use = n->raw_out(idx);
3148 idx++;
3149 if (!visited.test_set(use->_idx)) {
3150 if (member.test(use->_idx)) {
3151 nstack.push(n, idx);
3152 n = use;
3153 idx = 0;
3154 }
3155 }
3156 } else {
3157 // All outputs processed
3158 sched.push(n);
3159 if (nstack.is_empty()) break;
3160 n = nstack.node();
3161 idx = nstack.index();
3162 nstack.pop();
3163 }
3164 }
3165 }
3166
3167
3168 //------------------------------ has_use_in_set -------------------------------------
3169 // Has a use in the vector set
3170 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
3171 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3172 Node* use = n->fast_out(j);
3173 if (vset.test(use->_idx)) {
3174 return true;
3175 }
3176 }
3177 return false;
3178 }
3179
3180
3181 //------------------------------ has_use_internal_to_set -------------------------------------
3182 // Has use internal to the vector set (ie. not in a phi at the loop head)
3183 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
3184 Node* head = loop->_head;
3185 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3186 Node* use = n->fast_out(j);
3187 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
3188 return true;
3189 }
3190 }
3191 return false;
3192 }
3193
3194
3195 //------------------------------ clone_for_use_outside_loop -------------------------------------
3196 // clone "n" for uses that are outside of loop
3197 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
3198 int cloned = 0;
3199 assert(worklist.size() == 0, "should be empty");
3200 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3201 Node* use = n->fast_out(j);
3202 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
3203 worklist.push(use);
3204 }
3205 }
3206
3207 if (C->check_node_count(worklist.size() + NodeLimitFudgeFactor,
3208 "Too many clones required in clone_for_use_outside_loop in partial peeling")) {
3209 return -1;
3210 }
3211
3212 while( worklist.size() ) {
3213 Node *use = worklist.pop();
3214 if (!has_node(use) || use->in(0) == C->top()) continue;
3215 uint j;
3216 for (j = 0; j < use->req(); j++) {
3217 if (use->in(j) == n) break;
3218 }
3219 assert(j < use->req(), "must be there");
3220
3221 // clone "n" and insert it between the inputs of "n" and the use outside the loop
3222 Node* n_clone = n->clone();
3223 _igvn.replace_input_of(use, j, n_clone);
3224 cloned++;
3225 Node* use_c;
3226 if (!use->is_Phi()) {
3227 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
3228 } else {
3229 // Use in a phi is considered a use in the associated predecessor block
3230 use_c = use->in(0)->in(j);
3231 }
3232 set_ctrl(n_clone, use_c);
3233 assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
3234 get_loop(use_c)->_body.push(n_clone);
3235 _igvn.register_new_node_with_optimizer(n_clone);
3236 #ifndef PRODUCT
3237 if (TracePartialPeeling) {
3238 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
3239 }
3240 #endif
3241 }
3242 return cloned;
3243 }
3244
3245
3246 //------------------------------ clone_for_special_use_inside_loop -------------------------------------
3247 // clone "n" for special uses that are in the not_peeled region.
3248 // If these def-uses occur in separate blocks, the code generator
3249 // marks the method as not compilable. For example, if a "BoolNode"
3250 // is in a different basic block than the "IfNode" that uses it, then
3251 // the compilation is aborted in the code generator.
3252 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
3253 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
3254 if (n->is_Phi() || n->is_Load()) {
3255 return;
3256 }
3257 assert(worklist.size() == 0, "should be empty");
3258 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3259 Node* use = n->fast_out(j);
3260 if ( not_peel.test(use->_idx) &&
3261 (use->is_If() || use->is_CMove() || use->is_Bool()) &&
3262 use->in(1) == n) {
3263 worklist.push(use);
3264 }
3265 }
3266 if (worklist.size() > 0) {
3267 // clone "n" and insert it between inputs of "n" and the use
3268 Node* n_clone = n->clone();
3269 loop->_body.push(n_clone);
3270 _igvn.register_new_node_with_optimizer(n_clone);
3271 set_ctrl(n_clone, get_ctrl(n));
3272 sink_list.push(n_clone);
3273 not_peel.set(n_clone->_idx);
3274 #ifndef PRODUCT
3275 if (TracePartialPeeling) {
3276 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
3277 }
3278 #endif
3279 while( worklist.size() ) {
3280 Node *use = worklist.pop();
3281 _igvn.rehash_node_delayed(use);
3282 for (uint j = 1; j < use->req(); j++) {
3283 if (use->in(j) == n) {
3284 use->set_req(j, n_clone);
3285 }
3286 }
3287 }
3288 }
3289 }
3290
3291
3292 //------------------------------ insert_phi_for_loop -------------------------------------
3293 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
3294 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
3295 Node *phi = PhiNode::make(lp, back_edge_val);
3296 phi->set_req(LoopNode::EntryControl, lp_entry_val);
3297 // Use existing phi if it already exists
3298 Node *hit = _igvn.hash_find_insert(phi);
3299 if( hit == NULL ) {
3300 _igvn.register_new_node_with_optimizer(phi);
3301 set_ctrl(phi, lp);
3302 } else {
3303 // Remove the new phi from the graph and use the hit
3304 _igvn.remove_dead_node(phi);
3305 phi = hit;
3306 }
3307 _igvn.replace_input_of(use, idx, phi);
3308 }
3309
3310 #ifdef ASSERT
3311 //------------------------------ is_valid_loop_partition -------------------------------------
3312 // Validate the loop partition sets: peel and not_peel
3313 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
3314 VectorSet& not_peel ) {
3315 uint i;
3316 // Check that peel_list entries are in the peel set
3317 for (i = 0; i < peel_list.size(); i++) {
3318 if (!peel.test(peel_list.at(i)->_idx)) {
3319 return false;
3320 }
3321 }
3322 // Check at loop members are in one of peel set or not_peel set
3323 for (i = 0; i < loop->_body.size(); i++ ) {
3324 Node *def = loop->_body.at(i);
3325 uint di = def->_idx;
3326 // Check that peel set elements are in peel_list
3327 if (peel.test(di)) {
3328 if (not_peel.test(di)) {
3329 return false;
3330 }
3331 // Must be in peel_list also
3332 bool found = false;
3333 for (uint j = 0; j < peel_list.size(); j++) {
3334 if (peel_list.at(j)->_idx == di) {
3335 found = true;
3336 break;
3337 }
3338 }
3339 if (!found) {
3340 return false;
3341 }
3342 } else if (not_peel.test(di)) {
3343 if (peel.test(di)) {
3344 return false;
3345 }
3346 } else {
3347 return false;
3348 }
3349 }
3350 return true;
3351 }
3352
3353 //------------------------------ is_valid_clone_loop_exit_use -------------------------------------
3354 // Ensure a use outside of loop is of the right form
3355 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
3356 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
3357 return (use->is_Phi() &&
3358 use_c->is_Region() && use_c->req() == 3 &&
3359 (use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
3360 use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
3361 use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
3362 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
3363 }
3364
3365 //------------------------------ is_valid_clone_loop_form -------------------------------------
3366 // Ensure that all uses outside of loop are of the right form
3367 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
3368 uint orig_exit_idx, uint clone_exit_idx) {
3369 uint len = peel_list.size();
3370 for (uint i = 0; i < len; i++) {
3371 Node *def = peel_list.at(i);
3372
3373 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
3374 Node *use = def->fast_out(j);
3375 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
3376 if (!loop->is_member(get_loop(use_c))) {
3377 // use is not in the loop, check for correct structure
3378 if (use->in(0) == def) {
3379 // Okay
3380 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
3381 return false;
3382 }
3383 }
3384 }
3385 }
3386 return true;
3387 }
3388 #endif
3389
3390 //------------------------------ partial_peel -------------------------------------
3391 // Partially peel (aka loop rotation) the top portion of a loop (called
3392 // the peel section below) by cloning it and placing one copy just before
3393 // the new loop head and the other copy at the bottom of the new loop.
3394 //
3395 // before after where it came from
3396 //
3397 // stmt1 stmt1
3398 // loop: stmt2 clone
3399 // stmt2 if condA goto exitA clone
3400 // if condA goto exitA new_loop: new
3401 // stmt3 stmt3 clone
3402 // if !condB goto loop if condB goto exitB clone
3403 // exitB: stmt2 orig
3404 // stmt4 if !condA goto new_loop orig
3405 // exitA: goto exitA
3406 // exitB:
3407 // stmt4
3408 // exitA:
3409 //
3410 // Step 1: find the cut point: an exit test on probable
3411 // induction variable.
3412 // Step 2: schedule (with cloning) operations in the peel
3413 // section that can be executed after the cut into
3414 // the section that is not peeled. This may need
3415 // to clone operations into exit blocks. For
3416 // instance, a reference to A[i] in the not-peel
3417 // section and a reference to B[i] in an exit block
3418 // may cause a left-shift of i by 2 to be placed
3419 // in the peel block. This step will clone the left
3420 // shift into the exit block and sink the left shift
3421 // from the peel to the not-peel section.
3422 // Step 3: clone the loop, retarget the control, and insert
3423 // phis for values that are live across the new loop
3424 // head. This is very dependent on the graph structure
3425 // from clone_loop. It creates region nodes for
3426 // exit control and associated phi nodes for values
3427 // flow out of the loop through that exit. The region
3428 // node is dominated by the clone's control projection.
3429 // So the clone's peel section is placed before the
3430 // new loop head, and the clone's not-peel section is
3431 // forms the top part of the new loop. The original
3432 // peel section forms the tail of the new loop.
3433 // Step 4: update the dominator tree and recompute the
3434 // dominator depth.
3435 //
3436 // orig
3437 //
3438 // stmt1
3439 // |
3440 // v
3441 // loop predicate
3442 // |
3443 // v
3444 // loop<----+
3445 // | |
3446 // stmt2 |
3447 // | |
3448 // v |
3449 // ifA |
3450 // / | |
3451 // v v |
3452 // false true ^ <-- last_peel
3453 // / | |
3454 // / ===|==cut |
3455 // / stmt3 | <-- first_not_peel
3456 // / | |
3457 // | v |
3458 // v ifB |
3459 // exitA: / \ |
3460 // / \ |
3461 // v v |
3462 // false true |
3463 // / \ |
3464 // / ----+
3465 // |
3466 // v
3467 // exitB:
3468 // stmt4
3469 //
3470 //
3471 // after clone loop
3472 //
3473 // stmt1
3474 // |
3475 // v
3476 // loop predicate
3477 // / \
3478 // clone / \ orig
3479 // / \
3480 // / \
3481 // v v
3482 // +---->loop loop<----+
3483 // | | | |
3484 // | stmt2 stmt2 |
3485 // | | | |
3486 // | v v |
3487 // | ifA ifA |
3488 // | | \ / | |
3489 // | v v v v |
3490 // ^ true false false true ^ <-- last_peel
3491 // | | ^ \ / | |
3492 // | cut==|== \ \ / ===|==cut |
3493 // | stmt3 \ \ / stmt3 | <-- first_not_peel
3494 // | | dom | | | |
3495 // | v \ 1v v2 v |
3496 // | ifB regionA ifB |
3497 // | / \ | / \ |
3498 // | / \ v / \ |
3499 // | v v exitA: v v |
3500 // | true false false true |
3501 // | / ^ \ / \ |
3502 // +---- \ \ / ----+
3503 // dom \ /
3504 // \ 1v v2
3505 // regionB
3506 // |
3507 // v
3508 // exitB:
3509 // stmt4
3510 //
3511 //
3512 // after partial peel
3513 //
3514 // stmt1
3515 // |
3516 // v
3517 // loop predicate
3518 // /
3519 // clone / orig
3520 // / TOP
3521 // / \
3522 // v v
3523 // TOP->loop loop----+
3524 // | | |
3525 // stmt2 stmt2 |
3526 // | | |
3527 // v v |
3528 // ifA ifA |
3529 // | \ / | |
3530 // v v v v |
3531 // true false false true | <-- last_peel
3532 // | ^ \ / +------|---+
3533 // +->newloop \ \ / === ==cut | |
3534 // | stmt3 \ \ / TOP | |
3535 // | | dom | | stmt3 | | <-- first_not_peel
3536 // | v \ 1v v2 v | |
3537 // | ifB regionA ifB ^ v
3538 // | / \ | / \ | |
3539 // | / \ v / \ | |
3540 // | v v exitA: v v | |
3541 // | true false false true | |
3542 // | / ^ \ / \ | |
3543 // | | \ \ / v | |
3544 // | | dom \ / TOP | |
3545 // | | \ 1v v2 | |
3546 // ^ v regionB | |
3547 // | | | | |
3548 // | | v ^ v
3549 // | | exitB: | |
3550 // | | stmt4 | |
3551 // | +------------>-----------------+ |
3552 // | |
3553 // +-----------------<---------------------+
3554 //
3555 //
3556 // final graph
3557 //
3558 // stmt1
3559 // |
3560 // v
3561 // loop predicate
3562 // |
3563 // v
3564 // stmt2 clone
3565 // |
3566 // v
3567 // ........> ifA clone
3568 // : / |
3569 // dom / |
3570 // : v v
3571 // : false true
3572 // : | |
3573 // : | v
3574 // : | newloop<-----+
3575 // : | | |
3576 // : | stmt3 clone |
3577 // : | | |
3578 // : | v |
3579 // : | ifB |
3580 // : | / \ |
3581 // : | v v |
3582 // : | false true |
3583 // : | | | |
3584 // : | v stmt2 |
3585 // : | exitB: | |
3586 // : | stmt4 v |
3587 // : | ifA orig |
3588 // : | / \ |
3589 // : | / \ |
3590 // : | v v |
3591 // : | false true |
3592 // : | / \ |
3593 // : v v -----+
3594 // RegionA
3595 // |
3596 // v
3597 // exitA
3598 //
3599 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
3600
3601 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only");
3602 if (!loop->_head->is_Loop()) {
3603 return false;
3604 }
3605 LoopNode *head = loop->_head->as_Loop();
3606
3607 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
3608 return false;
3609 }
3610
3611 // Check for complex exit control
3612 for (uint ii = 0; ii < loop->_body.size(); ii++) {
3613 Node *n = loop->_body.at(ii);
3614 int opc = n->Opcode();
3615 if (n->is_Call() ||
3616 opc == Op_Catch ||
3617 opc == Op_CatchProj ||
3618 opc == Op_Jump ||
3619 opc == Op_JumpProj) {
3620 #ifndef PRODUCT
3621 if (TracePartialPeeling) {
3622 tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
3623 }
3624 #endif
3625 return false;
3626 }
3627 }
3628
3629 int dd = dom_depth(head);
3630
3631 // Step 1: find cut point
3632
3633 // Walk up dominators to loop head looking for first loop exit
3634 // which is executed on every path thru loop.
3635 IfNode *peel_if = NULL;
3636 IfNode *peel_if_cmpu = NULL;
3637
3638 Node *iff = loop->tail();
3639 while (iff != head) {
3640 if (iff->is_If()) {
3641 Node *ctrl = get_ctrl(iff->in(1));
3642 if (ctrl->is_top()) return false; // Dead test on live IF.
3643 // If loop-varying exit-test, check for induction variable
3644 if (loop->is_member(get_loop(ctrl)) &&
3645 loop->is_loop_exit(iff) &&
3646 is_possible_iv_test(iff)) {
3647 Node* cmp = iff->in(1)->in(1);
3648 if (cmp->Opcode() == Op_CmpI) {
3649 peel_if = iff->as_If();
3650 } else {
3651 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
3652 peel_if_cmpu = iff->as_If();
3653 }
3654 }
3655 }
3656 iff = idom(iff);
3657 }
3658
3659 // Prefer signed compare over unsigned compare.
3660 IfNode* new_peel_if = NULL;
3661 if (peel_if == NULL) {
3662 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) {
3663 return false; // No peel point found
3664 }
3665 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
3666 if (new_peel_if == NULL) {
3667 return false; // No peel point found
3668 }
3669 peel_if = new_peel_if;
3670 }
3671 Node* last_peel = stay_in_loop(peel_if, loop);
3672 Node* first_not_peeled = stay_in_loop(last_peel, loop);
3673 if (first_not_peeled == NULL || first_not_peeled == head) {
3674 return false;
3675 }
3676
3677 #ifndef PRODUCT
3678 if (TraceLoopOpts) {
3679 tty->print("PartialPeel ");
3680 loop->dump_head();
3681 }
3682
3683 if (TracePartialPeeling) {
3684 tty->print_cr("before partial peel one iteration");
3685 Node_List wl;
3686 Node* t = head->in(2);
3687 while (true) {
3688 wl.push(t);
3689 if (t == head) break;
3690 t = idom(t);
3691 }
3692 while (wl.size() > 0) {
3693 Node* tt = wl.pop();
3694 tt->dump();
3695 if (tt == last_peel) tty->print_cr("-- cut --");
3696 }
3697 }
3698 #endif
3699 VectorSet peel;
3700 VectorSet not_peel;
3701 Node_List peel_list;
3702 Node_List worklist;
3703 Node_List sink_list;
3704
3705 uint estimate = loop->est_loop_clone_sz(1);
3706 if (exceeding_node_budget(estimate)) {
3707 return false;
3708 }
3709
3710 // Set of cfg nodes to peel are those that are executable from
3711 // the head through last_peel.
3712 assert(worklist.size() == 0, "should be empty");
3713 worklist.push(head);
3714 peel.set(head->_idx);
3715 while (worklist.size() > 0) {
3716 Node *n = worklist.pop();
3717 if (n != last_peel) {
3718 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3719 Node* use = n->fast_out(j);
3720 if (use->is_CFG() &&
3721 loop->is_member(get_loop(use)) &&
3722 !peel.test_set(use->_idx)) {
3723 worklist.push(use);
3724 }
3725 }
3726 }
3727 }
3728
3729 // Set of non-cfg nodes to peel are those that are control
3730 // dependent on the cfg nodes.
3731 for (uint i = 0; i < loop->_body.size(); i++) {
3732 Node *n = loop->_body.at(i);
3733 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
3734 if (peel.test(n_c->_idx)) {
3735 peel.set(n->_idx);
3736 } else {
3737 not_peel.set(n->_idx);
3738 }
3739 }
3740
3741 // Step 2: move operations from the peeled section down into the
3742 // not-peeled section
3743
3744 // Get a post order schedule of nodes in the peel region
3745 // Result in right-most operand.
3746 scheduled_nodelist(loop, peel, peel_list);
3747
3748 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
3749
3750 // For future check for too many new phis
3751 uint old_phi_cnt = 0;
3752 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
3753 Node* use = head->fast_out(j);
3754 if (use->is_Phi()) old_phi_cnt++;
3755 }
3756
3757 #ifndef PRODUCT
3758 if (TracePartialPeeling) {
3759 tty->print_cr("\npeeled list");
3760 }
3761 #endif
3762
3763 // Evacuate nodes in peel region into the not_peeled region if possible
3764 bool too_many_clones = false;
3765 uint new_phi_cnt = 0;
3766 uint cloned_for_outside_use = 0;
3767 for (uint i = 0; i < peel_list.size();) {
3768 Node* n = peel_list.at(i);
3769 #ifndef PRODUCT
3770 if (TracePartialPeeling) n->dump();
3771 #endif
3772 bool incr = true;
3773 if (!n->is_CFG()) {
3774 if (has_use_in_set(n, not_peel)) {
3775 // If not used internal to the peeled region,
3776 // move "n" from peeled to not_peeled region.
3777 if (!has_use_internal_to_set(n, peel, loop)) {
3778 // if not pinned and not a load (which maybe anti-dependent on a store)
3779 // and not a CMove (Matcher expects only bool->cmove).
3780 if (n->in(0) == NULL && !n->is_Load() && !n->is_CMove()) {
3781 int new_clones = clone_for_use_outside_loop(loop, n, worklist);
3782 if (new_clones == -1) {
3783 too_many_clones = true;
3784 break;
3785 }
3786 cloned_for_outside_use += new_clones;
3787 sink_list.push(n);
3788 peel.remove(n->_idx);
3789 not_peel.set(n->_idx);
3790 peel_list.remove(i);
3791 incr = false;
3792 #ifndef PRODUCT
3793 if (TracePartialPeeling) {
3794 tty->print_cr("sink to not_peeled region: %d newbb: %d",
3795 n->_idx, get_ctrl(n)->_idx);
3796 }
3797 #endif
3798 }
3799 } else {
3800 // Otherwise check for special def-use cases that span
3801 // the peel/not_peel boundary such as bool->if
3802 clone_for_special_use_inside_loop(loop, n, not_peel, sink_list, worklist);
3803 new_phi_cnt++;
3804 }
3805 }
3806 }
3807 if (incr) i++;
3808 }
3809
3810 estimate += cloned_for_outside_use + new_phi_cnt;
3811 bool exceed_node_budget = !may_require_nodes(estimate);
3812 bool exceed_phi_limit = new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta;
3813
3814 if (too_many_clones || exceed_node_budget || exceed_phi_limit) {
3815 #ifndef PRODUCT
3816 if (TracePartialPeeling && exceed_phi_limit) {
3817 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c",
3818 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F');
3819 }
3820 #endif
3821 if (new_peel_if != NULL) {
3822 remove_cmpi_loop_exit(new_peel_if, loop);
3823 }
3824 // Inhibit more partial peeling on this loop
3825 assert(!head->is_partial_peel_loop(), "not partial peeled");
3826 head->mark_partial_peel_failed();
3827 if (cloned_for_outside_use > 0) {
3828 // Terminate this round of loop opts because
3829 // the graph outside this loop was changed.
3830 C->set_major_progress();
3831 return true;
3832 }
3833 return false;
3834 }
3835
3836 // Step 3: clone loop, retarget control, and insert new phis
3837
3838 // Create new loop head for new phis and to hang
3839 // the nodes being moved (sinked) from the peel region.
3840 LoopNode* new_head = new LoopNode(last_peel, last_peel);
3841 new_head->set_unswitch_count(head->unswitch_count()); // Preserve
3842 _igvn.register_new_node_with_optimizer(new_head);
3843 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
3844 _igvn.replace_input_of(first_not_peeled, 0, new_head);
3845 set_loop(new_head, loop);
3846 loop->_body.push(new_head);
3847 not_peel.set(new_head->_idx);
3848 set_idom(new_head, last_peel, dom_depth(first_not_peeled));
3849 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
3850
3851 while (sink_list.size() > 0) {
3852 Node* n = sink_list.pop();
3853 set_ctrl(n, new_head);
3854 }
3855
3856 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
3857
3858 clone_loop(loop, old_new, dd, IgnoreStripMined);
3859
3860 const uint clone_exit_idx = 1;
3861 const uint orig_exit_idx = 2;
3862 assert(is_valid_clone_loop_form(loop, peel_list, orig_exit_idx, clone_exit_idx), "bad clone loop");
3863
3864 Node* head_clone = old_new[head->_idx];
3865 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop();
3866 Node* orig_tail_clone = head_clone->in(2);
3867
3868 // Add phi if "def" node is in peel set and "use" is not
3869
3870 for (uint i = 0; i < peel_list.size(); i++) {
3871 Node *def = peel_list.at(i);
3872 if (!def->is_CFG()) {
3873 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
3874 Node *use = def->fast_out(j);
3875 if (has_node(use) && use->in(0) != C->top() &&
3876 (!peel.test(use->_idx) ||
3877 (use->is_Phi() && use->in(0) == head)) ) {
3878 worklist.push(use);
3879 }
3880 }
3881 while( worklist.size() ) {
3882 Node *use = worklist.pop();
3883 for (uint j = 1; j < use->req(); j++) {
3884 Node* n = use->in(j);
3885 if (n == def) {
3886
3887 // "def" is in peel set, "use" is not in peel set
3888 // or "use" is in the entry boundary (a phi) of the peel set
3889
3890 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
3891
3892 if ( loop->is_member(get_loop( use_c )) ) {
3893 // use is in loop
3894 if (old_new[use->_idx] != NULL) { // null for dead code
3895 Node* use_clone = old_new[use->_idx];
3896 _igvn.replace_input_of(use, j, C->top());
3897 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
3898 }
3899 } else {
3900 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
3901 // use is not in the loop, check if the live range includes the cut
3902 Node* lp_if = use_c->in(orig_exit_idx)->in(0);
3903 if (not_peel.test(lp_if->_idx)) {
3904 assert(j == orig_exit_idx, "use from original loop");
3905 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
3906 }
3907 }
3908 }
3909 }
3910 }
3911 }
3912 }
3913
3914 // Step 3b: retarget control
3915
3916 // Redirect control to the new loop head if a cloned node in
3917 // the not_peeled region has control that points into the peeled region.
3918 // This necessary because the cloned peeled region will be outside
3919 // the loop.
3920 // from to
3921 // cloned-peeled <---+
3922 // new_head_clone: | <--+
3923 // cloned-not_peeled in(0) in(0)
3924 // orig-peeled
3925
3926 for (uint i = 0; i < loop->_body.size(); i++) {
3927 Node *n = loop->_body.at(i);
3928 if (!n->is_CFG() && n->in(0) != NULL &&
3929 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
3930 Node* n_clone = old_new[n->_idx];
3931 _igvn.replace_input_of(n_clone, 0, new_head_clone);
3932 }
3933 }
3934
3935 // Backedge of the surviving new_head (the clone) is original last_peel
3936 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel);
3937
3938 // Cut first node in original not_peel set
3939 _igvn.rehash_node_delayed(new_head); // Multiple edge updates:
3940 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of
3941 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls
3942
3943 // Copy head_clone back-branch info to original head
3944 // and remove original head's loop entry and
3945 // clone head's back-branch
3946 _igvn.rehash_node_delayed(head); // Multiple edge updates
3947 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl));
3948 head->set_req(LoopNode::LoopBackControl, C->top());
3949 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top());
3950
3951 // Similarly modify the phis
3952 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
3953 Node* use = head->fast_out(k);
3954 if (use->is_Phi() && use->outcnt() > 0) {
3955 Node* use_clone = old_new[use->_idx];
3956 _igvn.rehash_node_delayed(use); // Multiple edge updates
3957 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl));
3958 use->set_req(LoopNode::LoopBackControl, C->top());
3959 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top());
3960 }
3961 }
3962
3963 // Step 4: update dominator tree and dominator depth
3964
3965 set_idom(head, orig_tail_clone, dd);
3966 recompute_dom_depth();
3967
3968 // Inhibit more partial peeling on this loop
3969 new_head_clone->set_partial_peel_loop();
3970 C->set_major_progress();
3971 loop->record_for_igvn();
3972
3973 #ifndef PRODUCT
3974 if (TracePartialPeeling) {
3975 tty->print_cr("\nafter partial peel one iteration");
3976 Node_List wl;
3977 Node* t = last_peel;
3978 while (true) {
3979 wl.push(t);
3980 if (t == head_clone) break;
3981 t = idom(t);
3982 }
3983 while (wl.size() > 0) {
3984 Node* tt = wl.pop();
3985 if (tt == head) tty->print_cr("orig head");
3986 else if (tt == new_head_clone) tty->print_cr("new head");
3987 else if (tt == head_clone) tty->print_cr("clone head");
3988 tt->dump();
3989 }
3990 }
3991 #endif
3992 return true;
3993 }
3994
3995 // Transform:
3996 //
3997 // loop<-----------------+
3998 // | |
3999 // stmt1 stmt2 .. stmtn |
4000 // | | | |
4001 // \ | / |
4002 // v v v |
4003 // region |
4004 // | |
4005 // shared_stmt |
4006 // | |
4007 // v |
4008 // if |
4009 // / \ |
4010 // | -----------+
4011 // v
4012 //
4013 // into:
4014 //
4015 // loop<-------------------+
4016 // | |
4017 // v |
4018 // +->loop |
4019 // | | |
4020 // | stmt1 stmt2 .. stmtn |
4021 // | | | | |
4022 // | | \ / |
4023 // | | v v |
4024 // | | region1 |
4025 // | | | |
4026 // | shared_stmt shared_stmt |
4027 // | | | |
4028 // | v v |
4029 // | if if |
4030 // | /\ / \ |
4031 // +-- | | -------+
4032 // \ /
4033 // v v
4034 // region2
4035 //
4036 // (region2 is shown to merge mirrored projections of the loop exit
4037 // ifs to make the diagram clearer but they really merge the same
4038 // projection)
4039 //
4040 // Conditions for this transformation to trigger:
4041 // - the path through stmt1 is frequent enough
4042 // - the inner loop will be turned into a counted loop after transformation
4043 bool PhaseIdealLoop::duplicate_loop_backedge(IdealLoopTree *loop, Node_List &old_new) {
4044 if (!DuplicateBackedge) {
4045 return false;
4046 }
4047 assert(!loop->_head->is_CountedLoop() || StressDuplicateBackedge, "Non-counted loop only");
4048 if (!loop->_head->is_Loop()) {
4049 return false;
4050 }
4051
4052 uint estimate = loop->est_loop_clone_sz(1);
4053 if (exceeding_node_budget(estimate)) {
4054 return false;
4055 }
4056
4057 LoopNode *head = loop->_head->as_Loop();
4058
4059 Node* region = NULL;
4060 IfNode* exit_test = NULL;
4061 uint inner;
4062 float f;
4063 if (StressDuplicateBackedge) {
4064 if (head->is_strip_mined()) {
4065 return false;
4066 }
4067 Node* c = head->in(LoopNode::LoopBackControl);
4068
4069 while (c != head) {
4070 if (c->is_Region()) {
4071 region = c;
4072 }
4073 c = idom(c);
4074 }
4075
4076 if (region == NULL) {
4077 return false;
4078 }
4079
4080 inner = 1;
4081 } else {
4082 // Is the shape of the loop that of a counted loop...
4083 Node* back_control = loop_exit_control(head, loop);
4084 if (back_control == NULL) {
4085 return false;
4086 }
4087
4088 BoolTest::mask bt = BoolTest::illegal;
4089 float cl_prob = 0;
4090 Node* incr = NULL;
4091 Node* limit = NULL;
4092 Node* cmp = loop_exit_test(back_control, loop, incr, limit, bt, cl_prob);
4093 if (cmp == NULL || cmp->Opcode() != Op_CmpI) {
4094 return false;
4095 }
4096
4097 // With an extra phi for the candidate iv?
4098 // Or the region node is the loop head
4099 if (!incr->is_Phi() || incr->in(0) == head) {
4100 return false;
4101 }
4102
4103 PathFrequency pf(head, this);
4104 region = incr->in(0);
4105
4106 // Go over all paths for the extra phi's region and see if that
4107 // path is frequent enough and would match the expected iv shape
4108 // if the extra phi is removed
4109 inner = 0;
4110 for (uint i = 1; i < incr->req(); ++i) {
4111 Node* in = incr->in(i);
4112 Node* trunc1 = NULL;
4113 Node* trunc2 = NULL;
4114 const TypeInteger* iv_trunc_t = NULL;
4115 Node* orig_in = in;
4116 if (!(in = CountedLoopNode::match_incr_with_optional_truncation(in, &trunc1, &trunc2, &iv_trunc_t, T_INT))) {
4117 continue;
4118 }
4119 assert(in->Opcode() == Op_AddI, "wrong increment code");
4120 Node* xphi = NULL;
4121 Node* stride = loop_iv_stride(in, loop, xphi);
4122
4123 if (stride == NULL) {
4124 continue;
4125 }
4126
4127 PhiNode* phi = loop_iv_phi(xphi, NULL, head, loop);
4128 if (phi == NULL ||
4129 (trunc1 == NULL && phi->in(LoopNode::LoopBackControl) != incr) ||
4130 (trunc1 != NULL && phi->in(LoopNode::LoopBackControl) != trunc1)) {
4131 return false;
4132 }
4133
4134 f = pf.to(region->in(i));
4135 if (f > 0.5) {
4136 inner = i;
4137 break;
4138 }
4139 }
4140
4141 if (inner == 0) {
4142 return false;
4143 }
4144
4145 exit_test = back_control->in(0)->as_If();
4146 }
4147
4148 if (idom(region)->is_Catch()) {
4149 return false;
4150 }
4151
4152 // Collect all control nodes that need to be cloned (shared_stmt in the diagram)
4153 Unique_Node_List wq;
4154 wq.push(head->in(LoopNode::LoopBackControl));
4155 for (uint i = 0; i < wq.size(); i++) {
4156 Node* c = wq.at(i);
4157 assert(get_loop(c) == loop, "not in the right loop?");
4158 if (c->is_Region()) {
4159 if (c != region) {
4160 for (uint j = 1; j < c->req(); ++j) {
4161 wq.push(c->in(j));
4162 }
4163 }
4164 } else {
4165 wq.push(c->in(0));
4166 }
4167 assert(!is_dominator(c, region) || c == region, "shouldn't go above region");
4168 }
4169
4170 Node* region_dom = idom(region);
4171
4172 // Can't do the transformation if this would cause a membar pair to
4173 // be split
4174 for (uint i = 0; i < wq.size(); i++) {
4175 Node* c = wq.at(i);
4176 if (c->is_MemBar() && (c->as_MemBar()->trailing_store() || c->as_MemBar()->trailing_load_store())) {
4177 assert(c->as_MemBar()->leading_membar()->trailing_membar() == c, "bad membar pair");
4178 if (!wq.member(c->as_MemBar()->leading_membar())) {
4179 return false;
4180 }
4181 }
4182 }
4183
4184 // Collect data nodes that need to be clones as well
4185 int dd = dom_depth(head);
4186
4187 for (uint i = 0; i < loop->_body.size(); ++i) {
4188 Node* n = loop->_body.at(i);
4189 if (has_ctrl(n)) {
4190 Node* c = get_ctrl(n);
4191 if (wq.member(c)) {
4192 wq.push(n);
4193 }
4194 } else {
4195 set_idom(n, idom(n), dd);
4196 }
4197 }
4198
4199 // clone shared_stmt
4200 clone_loop_body(wq, old_new, NULL);
4201
4202 Node* region_clone = old_new[region->_idx];
4203 region_clone->set_req(inner, C->top());
4204 set_idom(region, region->in(inner), dd);
4205
4206 // Prepare the outer loop
4207 Node* outer_head = new LoopNode(head->in(LoopNode::EntryControl), old_new[head->in(LoopNode::LoopBackControl)->_idx]);
4208 register_control(outer_head, loop->_parent, outer_head->in(LoopNode::EntryControl));
4209 _igvn.replace_input_of(head, LoopNode::EntryControl, outer_head);
4210 set_idom(head, outer_head, dd);
4211
4212 fix_body_edges(wq, loop, old_new, dd, loop->_parent, true);
4213
4214 // Make one of the shared_stmt copies only reachable from stmt1, the
4215 // other only from stmt2..stmtn.
4216 Node* dom = NULL;
4217 for (uint i = 1; i < region->req(); ++i) {
4218 if (i != inner) {
4219 _igvn.replace_input_of(region, i, C->top());
4220 }
4221 Node* in = region_clone->in(i);
4222 if (in->is_top()) {
4223 continue;
4224 }
4225 if (dom == NULL) {
4226 dom = in;
4227 } else {
4228 dom = dom_lca(dom, in);
4229 }
4230 }
4231
4232 set_idom(region_clone, dom, dd);
4233
4234 // Set up the outer loop
4235 for (uint i = 0; i < head->outcnt(); i++) {
4236 Node* u = head->raw_out(i);
4237 if (u->is_Phi()) {
4238 Node* outer_phi = u->clone();
4239 outer_phi->set_req(0, outer_head);
4240 Node* backedge = old_new[u->in(LoopNode::LoopBackControl)->_idx];
4241 if (backedge == NULL) {
4242 backedge = u->in(LoopNode::LoopBackControl);
4243 }
4244 outer_phi->set_req(LoopNode::LoopBackControl, backedge);
4245 register_new_node(outer_phi, outer_head);
4246 _igvn.replace_input_of(u, LoopNode::EntryControl, outer_phi);
4247 }
4248 }
4249
4250 // create control and data nodes for out of loop uses (including region2)
4251 Node_List worklist;
4252 uint new_counter = C->unique();
4253 fix_ctrl_uses(wq, loop, old_new, ControlAroundStripMined, outer_head, NULL, worklist);
4254
4255 Node_List *split_if_set = NULL;
4256 Node_List *split_bool_set = NULL;
4257 Node_List *split_cex_set = NULL;
4258 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);
4259
4260 finish_clone_loop(split_if_set, split_bool_set, split_cex_set);
4261
4262 if (exit_test != NULL) {
4263 float cnt = exit_test->_fcnt;
4264 if (cnt != COUNT_UNKNOWN) {
4265 exit_test->_fcnt = cnt * f;
4266 old_new[exit_test->_idx]->as_If()->_fcnt = cnt * (1 - f);
4267 }
4268 }
4269
4270 C->set_major_progress();
4271
4272 return true;
4273 }
4274
4275 //------------------------------reorg_offsets----------------------------------
4276 // Reorganize offset computations to lower register pressure. Mostly
4277 // prevent loop-fallout uses of the pre-incremented trip counter (which are
4278 // then alive with the post-incremented trip counter forcing an extra
4279 // register move):
4280 //
4281 // iv Phi iv Phi
4282 // | |
4283 // | AddI (+stride)
4284 // | |
4285 // | Opaque2 # Blocks IGVN from folding these nodes until loop opts are over.
4286 // | ====> |
4287 // | AddI (-stride)
4288 // | |
4289 // | CastII # Preserve type of iv Phi
4290 // | |
4291 // Outside Use Outside Use
4292 //
4293 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) {
4294 // Perform it only for canonical counted loops.
4295 // Loop's shape could be messed up by iteration_split_impl.
4296 if (!loop->_head->is_CountedLoop())
4297 return;
4298 if (!loop->_head->as_Loop()->is_valid_counted_loop(T_INT))
4299 return;
4300
4301 CountedLoopNode *cl = loop->_head->as_CountedLoop();
4302 CountedLoopEndNode *cle = cl->loopexit();
4303 Node *exit = cle->proj_out(false);
4304 Node *phi = cl->phi();
4305
4306 // Check for the special case when using the pre-incremented trip-counter on
4307 // the fall-out path (forces the pre-incremented and post-incremented trip
4308 // counter to be live at the same time). Fix this by adjusting to use the
4309 // post-increment trip counter.
4310
4311 bool progress = true;
4312 while (progress) {
4313 progress = false;
4314 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
4315 Node* use = phi->fast_out(i); // User of trip-counter
4316 if (!has_ctrl(use)) continue;
4317 Node *u_ctrl = get_ctrl(use);
4318 if (use->is_Phi()) {
4319 u_ctrl = NULL;
4320 for (uint j = 1; j < use->req(); j++)
4321 if (use->in(j) == phi)
4322 u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j));
4323 }
4324 IdealLoopTree *u_loop = get_loop(u_ctrl);
4325 // Look for loop-invariant use
4326 if (u_loop == loop) continue;
4327 if (loop->is_member(u_loop)) continue;
4328 // Check that use is live out the bottom. Assuming the trip-counter
4329 // update is right at the bottom, uses of of the loop middle are ok.
4330 if (dom_lca(exit, u_ctrl) != exit) continue;
4331 // Hit! Refactor use to use the post-incremented tripcounter.
4332 // Compute a post-increment tripcounter.
4333 Node* c = exit;
4334 if (cl->is_strip_mined()) {
4335 IdealLoopTree* outer_loop = get_loop(cl->outer_loop());
4336 if (!outer_loop->is_member(u_loop)) {
4337 c = cl->outer_loop_exit();
4338 }
4339 }
4340 Node *opaq = new Opaque2Node(C, cle->incr());
4341 register_new_node(opaq, c);
4342 Node *neg_stride = _igvn.intcon(-cle->stride_con());
4343 set_ctrl(neg_stride, C->root());
4344 Node *post = new AddINode(opaq, neg_stride);
4345 register_new_node(post, c);
4346 post = new CastIINode(post, phi->bottom_type()); // preserve the iv phi's type
4347 register_new_node(post, c);
4348 _igvn.rehash_node_delayed(use);
4349 for (uint j = 1; j < use->req(); j++) {
4350 if (use->in(j) == phi)
4351 use->set_req(j, post);
4352 }
4353 // Since DU info changed, rerun loop
4354 progress = true;
4355 break;
4356 }
4357 }
4358
4359 }