1 /*
2 * Copyright (c) 1999, 2019, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "gc/shared/barrierSet.hpp"
27 #include "gc/shared/c2/barrierSetC2.hpp"
28 #include "memory/allocation.inline.hpp"
29 #include "memory/resourceArea.hpp"
30 #include "opto/addnode.hpp"
31 #include "opto/callnode.hpp"
32 #include "opto/castnode.hpp"
33 #include "opto/connode.hpp"
34 #include "opto/castnode.hpp"
35 #include "opto/divnode.hpp"
36 #include "opto/loopnode.hpp"
37 #include "opto/matcher.hpp"
38 #include "opto/mulnode.hpp"
39 #include "opto/movenode.hpp"
40 #include "opto/opaquenode.hpp"
41 #include "opto/rootnode.hpp"
42 #include "opto/subnode.hpp"
43 #include "opto/subtypenode.hpp"
44 #include "utilities/macros.hpp"
45
46 //=============================================================================
47 //------------------------------split_thru_phi---------------------------------
48 // Split Node 'n' through merge point if there is enough win.
49 Node* PhaseIdealLoop::split_thru_phi(Node* n, Node* region, int policy) {
50 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) {
51 // ConvI2L may have type information on it which is unsafe to push up
52 // so disable this for now
53 return NULL;
54 }
55
56 // Splitting range check CastIIs through a loop induction Phi can
57 // cause new Phis to be created that are left unrelated to the loop
58 // induction Phi and prevent optimizations (vectorization)
59 if (n->Opcode() == Op_CastII && region->is_CountedLoop() &&
60 n->in(1) == region->as_CountedLoop()->phi()) {
61 return NULL;
62 }
63
64 // Bail out if 'n' is a Div or Mod node whose zero check was removed earlier (i.e. control is NULL) and its divisor is an induction variable
65 // phi p of a trip-counted (integer) loop whose inputs could be zero (include zero in their type range). p could have a more precise type
66 // range that does not necessarily include all values of its inputs. Since each of these inputs will be a divisor of the newly cloned nodes
67 // of 'n', we need to bail out of one of these divisors could be zero (zero in its type range).
68 if ((n->Opcode() == Op_DivI || n->Opcode() == Op_ModI) && n->in(0) == NULL
69 && region->is_CountedLoop() && n->in(2) == region->as_CountedLoop()->phi()) {
70 Node* phi = region->as_CountedLoop()->phi();
71 for (uint i = 1; i < phi->req(); i++) {
72 if (_igvn.type(phi->in(i))->filter_speculative(TypeInt::ZERO) != Type::TOP) {
73 // Zero could be a possible value but we already removed the zero check. Bail out to avoid a possible division by zero at a later point.
74 return NULL;
75 }
76 }
77 }
78
79 int wins = 0;
80 assert(!n->is_CFG(), "");
81 assert(region->is_Region(), "");
82
83 const Type* type = n->bottom_type();
84 const TypeOopPtr* t_oop = _igvn.type(n)->isa_oopptr();
85 Node* phi;
86 if (t_oop != NULL && t_oop->is_known_instance_field()) {
87 int iid = t_oop->instance_id();
88 int index = C->get_alias_index(t_oop);
89 int offset = t_oop->offset();
90 phi = new PhiNode(region, type, NULL, iid, index, offset);
91 } else {
92 phi = PhiNode::make_blank(region, n);
93 }
94 uint old_unique = C->unique();
95 for (uint i = 1; i < region->req(); i++) {
96 Node* x;
97 Node* the_clone = NULL;
98 if (region->in(i) == C->top()) {
99 x = C->top(); // Dead path? Use a dead data op
100 } else {
101 x = n->clone(); // Else clone up the data op
102 the_clone = x; // Remember for possible deletion.
103 // Alter data node to use pre-phi inputs
104 if (n->in(0) == region)
105 x->set_req( 0, region->in(i) );
106 for (uint j = 1; j < n->req(); j++) {
107 Node* in = n->in(j);
108 if (in->is_Phi() && in->in(0) == region)
109 x->set_req(j, in->in(i)); // Use pre-Phi input for the clone
110 }
111 }
112 // Check for a 'win' on some paths
113 const Type* t = x->Value(&_igvn);
114
115 bool singleton = t->singleton();
116
117 // A TOP singleton indicates that there are no possible values incoming
118 // along a particular edge. In most cases, this is OK, and the Phi will
119 // be eliminated later in an Ideal call. However, we can't allow this to
120 // happen if the singleton occurs on loop entry, as the elimination of
121 // the PhiNode may cause the resulting node to migrate back to a previous
122 // loop iteration.
123 if (singleton && t == Type::TOP) {
124 // Is_Loop() == false does not confirm the absence of a loop (e.g., an
125 // irreducible loop may not be indicated by an affirmative is_Loop());
126 // therefore, the only top we can split thru a phi is on a backedge of
127 // a loop.
128 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
129 }
130
131 if (singleton) {
132 wins++;
133 x = ((PhaseGVN&)_igvn).makecon(t);
134 } else {
135 // We now call Identity to try to simplify the cloned node.
136 // Note that some Identity methods call phase->type(this).
137 // Make sure that the type array is big enough for
138 // our new node, even though we may throw the node away.
139 // (Note: This tweaking with igvn only works because x is a new node.)
140 _igvn.set_type(x, t);
141 // If x is a TypeNode, capture any more-precise type permanently into Node
142 // otherwise it will be not updated during igvn->transform since
143 // igvn->type(x) is set to x->Value() already.
144 x->raise_bottom_type(t);
145 Node* y = x->Identity(&_igvn);
146 if (y != x) {
147 wins++;
148 x = y;
149 } else {
150 y = _igvn.hash_find(x);
151 if (y) {
152 wins++;
153 x = y;
154 } else {
155 // Else x is a new node we are keeping
156 // We do not need register_new_node_with_optimizer
157 // because set_type has already been called.
158 _igvn._worklist.push(x);
159 }
160 }
161 }
162 if (x != the_clone && the_clone != NULL)
163 _igvn.remove_dead_node(the_clone);
164 phi->set_req( i, x );
165 }
166 // Too few wins?
167 if (wins <= policy) {
168 _igvn.remove_dead_node(phi);
169 return NULL;
170 }
171
172 // Record Phi
173 register_new_node( phi, region );
174
175 for (uint i2 = 1; i2 < phi->req(); i2++) {
176 Node *x = phi->in(i2);
177 // If we commoned up the cloned 'x' with another existing Node,
178 // the existing Node picks up a new use. We need to make the
179 // existing Node occur higher up so it dominates its uses.
180 Node *old_ctrl;
181 IdealLoopTree *old_loop;
182
183 if (x->is_Con()) {
184 // Constant's control is always root.
185 set_ctrl(x, C->root());
186 continue;
187 }
188 // The occasional new node
189 if (x->_idx >= old_unique) { // Found a new, unplaced node?
190 old_ctrl = NULL;
191 old_loop = NULL; // Not in any prior loop
192 } else {
193 old_ctrl = get_ctrl(x);
194 old_loop = get_loop(old_ctrl); // Get prior loop
195 }
196 // New late point must dominate new use
197 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2));
198 if (new_ctrl == old_ctrl) // Nothing is changed
199 continue;
200
201 IdealLoopTree *new_loop = get_loop(new_ctrl);
202
203 // Don't move x into a loop if its uses are
204 // outside of loop. Otherwise x will be cloned
205 // for each use outside of this loop.
206 IdealLoopTree *use_loop = get_loop(region);
207 if (!new_loop->is_member(use_loop) &&
208 (old_loop == NULL || !new_loop->is_member(old_loop))) {
209 // Take early control, later control will be recalculated
210 // during next iteration of loop optimizations.
211 new_ctrl = get_early_ctrl(x);
212 new_loop = get_loop(new_ctrl);
213 }
214 // Set new location
215 set_ctrl(x, new_ctrl);
216 // If changing loop bodies, see if we need to collect into new body
217 if (old_loop != new_loop) {
218 if (old_loop && !old_loop->_child)
219 old_loop->_body.yank(x);
220 if (!new_loop->_child)
221 new_loop->_body.push(x); // Collect body info
222 }
223 }
224
225 return phi;
226 }
227
228 //------------------------------dominated_by------------------------------------
229 // Replace the dominated test with an obvious true or false. Place it on the
230 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the
231 // live path up to the dominating control.
232 void PhaseIdealLoop::dominated_by(IfProjNode* prevdom, IfNode* iff, bool flip, bool exclude_loop_predicate) {
233 if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); }
234
235 // prevdom is the dominating projection of the dominating test.
236 assert(iff->Opcode() == Op_If ||
237 iff->Opcode() == Op_CountedLoopEnd ||
238 iff->Opcode() == Op_LongCountedLoopEnd ||
239 iff->Opcode() == Op_RangeCheck,
240 "Check this code when new subtype is added");
241
242 int pop = prevdom->Opcode();
243 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" );
244 if (flip) {
245 if (pop == Op_IfTrue)
246 pop = Op_IfFalse;
247 else
248 pop = Op_IfTrue;
249 }
250 // 'con' is set to true or false to kill the dominated test.
251 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO);
252 set_ctrl(con, C->root()); // Constant gets a new use
253 // Hack the dominated test
254 _igvn.replace_input_of(iff, 1, con);
255
256 // If I dont have a reachable TRUE and FALSE path following the IfNode then
257 // I can assume this path reaches an infinite loop. In this case it's not
258 // important to optimize the data Nodes - either the whole compilation will
259 // be tossed or this path (and all data Nodes) will go dead.
260 if (iff->outcnt() != 2) return;
261
262 // Make control-dependent data Nodes on the live path (path that will remain
263 // once the dominated IF is removed) become control-dependent on the
264 // dominating projection.
265 Node* dp = iff->proj_out_or_null(pop == Op_IfTrue);
266
267 // Loop predicates may have depending checks which should not
268 // be skipped. For example, range check predicate has two checks
269 // for lower and upper bounds.
270 if (dp == NULL)
271 return;
272
273 ProjNode* dp_proj = dp->as_Proj();
274 ProjNode* unc_proj = iff->proj_out(1 - dp_proj->_con)->as_Proj();
275 if (exclude_loop_predicate &&
276 (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) != NULL ||
277 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_profile_predicate) != NULL ||
278 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check) != NULL)) {
279 // If this is a range check (IfNode::is_range_check), do not
280 // reorder because Compile::allow_range_check_smearing might have
281 // changed the check.
282 return; // Let IGVN transformation change control dependence.
283 }
284
285 IdealLoopTree* old_loop = get_loop(dp);
286
287 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
288 Node* cd = dp->fast_out(i); // Control-dependent node
289 // Do not rewire Div and Mod nodes which could have a zero divisor to avoid skipping their zero check.
290 if (cd->depends_only_on_test() && _igvn.no_dependent_zero_check(cd)) {
291 assert(cd->in(0) == dp, "");
292 _igvn.replace_input_of(cd, 0, prevdom);
293 set_early_ctrl(cd, false);
294 IdealLoopTree* new_loop = get_loop(get_ctrl(cd));
295 if (old_loop != new_loop) {
296 if (!old_loop->_child) {
297 old_loop->_body.yank(cd);
298 }
299 if (!new_loop->_child) {
300 new_loop->_body.push(cd);
301 }
302 }
303 --i;
304 --imax;
305 }
306 }
307 }
308
309 //------------------------------has_local_phi_input----------------------------
310 // Return TRUE if 'n' has Phi inputs from its local block and no other
311 // block-local inputs (all non-local-phi inputs come from earlier blocks)
312 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) {
313 Node *n_ctrl = get_ctrl(n);
314 // See if some inputs come from a Phi in this block, or from before
315 // this block.
316 uint i;
317 for( i = 1; i < n->req(); i++ ) {
318 Node *phi = n->in(i);
319 if( phi->is_Phi() && phi->in(0) == n_ctrl )
320 break;
321 }
322 if( i >= n->req() )
323 return NULL; // No Phi inputs; nowhere to clone thru
324
325 // Check for inputs created between 'n' and the Phi input. These
326 // must split as well; they have already been given the chance
327 // (courtesy of a post-order visit) and since they did not we must
328 // recover the 'cost' of splitting them by being very profitable
329 // when splitting 'n'. Since this is unlikely we simply give up.
330 for( i = 1; i < n->req(); i++ ) {
331 Node *m = n->in(i);
332 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) {
333 // We allow the special case of AddP's with no local inputs.
334 // This allows us to split-up address expressions.
335 if (m->is_AddP() &&
336 get_ctrl(m->in(2)) != n_ctrl &&
337 get_ctrl(m->in(3)) != n_ctrl) {
338 // Move the AddP up to dominating point
339 Node* c = find_non_split_ctrl(idom(n_ctrl));
340 if (c->is_OuterStripMinedLoop()) {
341 c->as_Loop()->verify_strip_mined(1);
342 c = c->in(LoopNode::EntryControl);
343 }
344 set_ctrl_and_loop(m, c);
345 continue;
346 }
347 return NULL;
348 }
349 assert(n->is_Phi() || m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control");
350 }
351
352 return n_ctrl;
353 }
354
355 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2).
356 Node* PhaseIdealLoop::remix_address_expressions_add_left_shift(Node* n, IdealLoopTree* n_loop, Node* n_ctrl, BasicType bt) {
357 assert(bt == T_INT || bt == T_LONG, "only for integers");
358 int n_op = n->Opcode();
359
360 if (n_op == Op_LShift(bt)) {
361 // Scale is loop invariant
362 Node* scale = n->in(2);
363 Node* scale_ctrl = get_ctrl(scale);
364 IdealLoopTree* scale_loop = get_loop(scale_ctrl);
365 if (n_loop == scale_loop || !scale_loop->is_member(n_loop)) {
366 return NULL;
367 }
368 const TypeInt* scale_t = scale->bottom_type()->isa_int();
369 if (scale_t != NULL && scale_t->is_con() && scale_t->get_con() >= 16) {
370 return NULL; // Dont bother with byte/short masking
371 }
372 // Add must vary with loop (else shift would be loop-invariant)
373 Node* add = n->in(1);
374 Node* add_ctrl = get_ctrl(add);
375 IdealLoopTree* add_loop = get_loop(add_ctrl);
376 if (n_loop != add_loop) {
377 return NULL; // happens w/ evil ZKM loops
378 }
379
380 // Convert I-V into I+ (0-V); same for V-I
381 if (add->Opcode() == Op_Sub(bt) &&
382 _igvn.type(add->in(1)) != TypeInteger::zero(bt)) {
383 assert(add->Opcode() == Op_SubI || add->Opcode() == Op_SubL, "");
384 Node* zero = _igvn.integercon(0, bt);
385 set_ctrl(zero, C->root());
386 Node* neg = SubNode::make(zero, add->in(2), bt);
387 register_new_node(neg, get_ctrl(add->in(2)));
388 add = AddNode::make(add->in(1), neg, bt);
389 register_new_node(add, add_ctrl);
390 }
391 if (add->Opcode() != Op_Add(bt)) return NULL;
392 assert(add->Opcode() == Op_AddI || add->Opcode() == Op_AddL, "");
393 // See if one add input is loop invariant
394 Node* add_var = add->in(1);
395 Node* add_var_ctrl = get_ctrl(add_var);
396 IdealLoopTree* add_var_loop = get_loop(add_var_ctrl);
397 Node* add_invar = add->in(2);
398 Node* add_invar_ctrl = get_ctrl(add_invar);
399 IdealLoopTree* add_invar_loop = get_loop(add_invar_ctrl);
400 if (add_invar_loop == n_loop) {
401 // Swap to find the invariant part
402 add_invar = add_var;
403 add_invar_ctrl = add_var_ctrl;
404 add_invar_loop = add_var_loop;
405 add_var = add->in(2);
406 } else if (add_var_loop != n_loop) { // Else neither input is loop invariant
407 return NULL;
408 }
409 if (n_loop == add_invar_loop || !add_invar_loop->is_member(n_loop)) {
410 return NULL; // No invariant part of the add?
411 }
412
413 // Yes! Reshape address expression!
414 Node* inv_scale = LShiftNode::make(add_invar, scale, bt);
415 Node* inv_scale_ctrl =
416 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ?
417 add_invar_ctrl : scale_ctrl;
418 register_new_node(inv_scale, inv_scale_ctrl);
419 Node* var_scale = LShiftNode::make(add_var, scale, bt);
420 register_new_node(var_scale, n_ctrl);
421 Node* var_add = AddNode::make(var_scale, inv_scale, bt);
422 register_new_node(var_add, n_ctrl);
423 _igvn.replace_node(n, var_add);
424 return var_add;
425 }
426 return NULL;
427 }
428
429 //------------------------------remix_address_expressions----------------------
430 // Rework addressing expressions to get the most loop-invariant stuff
431 // moved out. We'd like to do all associative operators, but it's especially
432 // important (common) to do address expressions.
433 Node* PhaseIdealLoop::remix_address_expressions(Node* n) {
434 if (!has_ctrl(n)) return NULL;
435 Node* n_ctrl = get_ctrl(n);
436 IdealLoopTree* n_loop = get_loop(n_ctrl);
437
438 // See if 'n' mixes loop-varying and loop-invariant inputs and
439 // itself is loop-varying.
440
441 // Only interested in binary ops (and AddP)
442 if (n->req() < 3 || n->req() > 4) return NULL;
443
444 Node* n1_ctrl = get_ctrl(n->in( 1));
445 Node* n2_ctrl = get_ctrl(n->in( 2));
446 Node* n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3));
447 IdealLoopTree* n1_loop = get_loop(n1_ctrl);
448 IdealLoopTree* n2_loop = get_loop(n2_ctrl);
449 IdealLoopTree* n3_loop = get_loop(n3_ctrl);
450
451 // Does one of my inputs spin in a tighter loop than self?
452 if ((n_loop->is_member(n1_loop) && n_loop != n1_loop) ||
453 (n_loop->is_member(n2_loop) && n_loop != n2_loop) ||
454 (n_loop->is_member(n3_loop) && n_loop != n3_loop)) {
455 return NULL; // Leave well enough alone
456 }
457
458 // Is at least one of my inputs loop-invariant?
459 if (n1_loop == n_loop &&
460 n2_loop == n_loop &&
461 n3_loop == n_loop) {
462 return NULL; // No loop-invariant inputs
463 }
464
465 Node* res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_INT);
466 if (res != NULL) {
467 return res;
468 }
469 res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_LONG);
470 if (res != NULL) {
471 return res;
472 }
473
474 int n_op = n->Opcode();
475 // Replace (I+V) with (V+I)
476 if (n_op == Op_AddI ||
477 n_op == Op_AddL ||
478 n_op == Op_AddF ||
479 n_op == Op_AddD ||
480 n_op == Op_MulI ||
481 n_op == Op_MulL ||
482 n_op == Op_MulF ||
483 n_op == Op_MulD) {
484 if (n2_loop == n_loop) {
485 assert(n1_loop != n_loop, "");
486 n->swap_edges(1, 2);
487 }
488 }
489
490 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V),
491 // but not if I2 is a constant.
492 if (n_op == Op_AddP) {
493 if (n2_loop == n_loop && n3_loop != n_loop) {
494 if (n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con()) {
495 Node* n22_ctrl = get_ctrl(n->in(2)->in(2));
496 Node* n23_ctrl = get_ctrl(n->in(2)->in(3));
497 IdealLoopTree* n22loop = get_loop(n22_ctrl);
498 IdealLoopTree* n23_loop = get_loop(n23_ctrl);
499 if (n22loop != n_loop && n22loop->is_member(n_loop) &&
500 n23_loop == n_loop) {
501 Node* add1 = new AddPNode(n->in(1), n->in(2)->in(2), n->in(3));
502 // Stuff new AddP in the loop preheader
503 register_new_node(add1, n_loop->_head->in(LoopNode::EntryControl));
504 Node* add2 = new AddPNode(n->in(1), add1, n->in(2)->in(3));
505 register_new_node(add2, n_ctrl);
506 _igvn.replace_node(n, add2);
507 return add2;
508 }
509 }
510 }
511
512 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V)
513 if (n2_loop != n_loop && n3_loop == n_loop) {
514 if (n->in(3)->Opcode() == Op_AddX) {
515 Node* V = n->in(3)->in(1);
516 Node* I = n->in(3)->in(2);
517 if (is_member(n_loop,get_ctrl(V))) {
518 } else {
519 Node *tmp = V; V = I; I = tmp;
520 }
521 if (!is_member(n_loop,get_ctrl(I))) {
522 Node* add1 = new AddPNode(n->in(1), n->in(2), I);
523 // Stuff new AddP in the loop preheader
524 register_new_node(add1, n_loop->_head->in(LoopNode::EntryControl));
525 Node* add2 = new AddPNode(n->in(1), add1, V);
526 register_new_node(add2, n_ctrl);
527 _igvn.replace_node(n, add2);
528 return add2;
529 }
530 }
531 }
532 }
533
534 return NULL;
535 }
536
537 // Optimize ((in1[2*i] * in2[2*i]) + (in1[2*i+1] * in2[2*i+1]))
538 Node *PhaseIdealLoop::convert_add_to_muladd(Node* n) {
539 assert(n->Opcode() == Op_AddI, "sanity");
540 Node * nn = NULL;
541 Node * in1 = n->in(1);
542 Node * in2 = n->in(2);
543 if (in1->Opcode() == Op_MulI && in2->Opcode() == Op_MulI) {
544 IdealLoopTree* loop_n = get_loop(get_ctrl(n));
545 if (loop_n->is_counted() &&
546 loop_n->_head->as_Loop()->is_valid_counted_loop(T_INT) &&
547 Matcher::match_rule_supported(Op_MulAddVS2VI) &&
548 Matcher::match_rule_supported(Op_MulAddS2I)) {
549 Node* mul_in1 = in1->in(1);
550 Node* mul_in2 = in1->in(2);
551 Node* mul_in3 = in2->in(1);
552 Node* mul_in4 = in2->in(2);
553 if (mul_in1->Opcode() == Op_LoadS &&
554 mul_in2->Opcode() == Op_LoadS &&
555 mul_in3->Opcode() == Op_LoadS &&
556 mul_in4->Opcode() == Op_LoadS) {
557 IdealLoopTree* loop1 = get_loop(get_ctrl(mul_in1));
558 IdealLoopTree* loop2 = get_loop(get_ctrl(mul_in2));
559 IdealLoopTree* loop3 = get_loop(get_ctrl(mul_in3));
560 IdealLoopTree* loop4 = get_loop(get_ctrl(mul_in4));
561 IdealLoopTree* loop5 = get_loop(get_ctrl(in1));
562 IdealLoopTree* loop6 = get_loop(get_ctrl(in2));
563 // All nodes should be in the same counted loop.
564 if (loop_n == loop1 && loop_n == loop2 && loop_n == loop3 &&
565 loop_n == loop4 && loop_n == loop5 && loop_n == loop6) {
566 Node* adr1 = mul_in1->in(MemNode::Address);
567 Node* adr2 = mul_in2->in(MemNode::Address);
568 Node* adr3 = mul_in3->in(MemNode::Address);
569 Node* adr4 = mul_in4->in(MemNode::Address);
570 if (adr1->is_AddP() && adr2->is_AddP() && adr3->is_AddP() && adr4->is_AddP()) {
571 if ((adr1->in(AddPNode::Base) == adr3->in(AddPNode::Base)) &&
572 (adr2->in(AddPNode::Base) == adr4->in(AddPNode::Base))) {
573 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in3, mul_in4);
574 register_new_node(nn, get_ctrl(n));
575 _igvn.replace_node(n, nn);
576 return nn;
577 } else if ((adr1->in(AddPNode::Base) == adr4->in(AddPNode::Base)) &&
578 (adr2->in(AddPNode::Base) == adr3->in(AddPNode::Base))) {
579 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in4, mul_in3);
580 register_new_node(nn, get_ctrl(n));
581 _igvn.replace_node(n, nn);
582 return nn;
583 }
584 }
585 }
586 }
587 }
588 }
589 return nn;
590 }
591
592 //------------------------------conditional_move-------------------------------
593 // Attempt to replace a Phi with a conditional move. We have some pretty
594 // strict profitability requirements. All Phis at the merge point must
595 // be converted, so we can remove the control flow. We need to limit the
596 // number of c-moves to a small handful. All code that was in the side-arms
597 // of the CFG diamond is now speculatively executed. This code has to be
598 // "cheap enough". We are pretty much limited to CFG diamonds that merge
599 // 1 or 2 items with a total of 1 or 2 ops executed speculatively.
600 Node *PhaseIdealLoop::conditional_move( Node *region ) {
601
602 assert(region->is_Region(), "sanity check");
603 if (region->req() != 3) return NULL;
604
605 // Check for CFG diamond
606 Node *lp = region->in(1);
607 Node *rp = region->in(2);
608 if (!lp || !rp) return NULL;
609 Node *lp_c = lp->in(0);
610 if (lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If()) return NULL;
611 IfNode *iff = lp_c->as_If();
612
613 // Check for ops pinned in an arm of the diamond.
614 // Can't remove the control flow in this case
615 if (lp->outcnt() > 1) return NULL;
616 if (rp->outcnt() > 1) return NULL;
617
618 IdealLoopTree* r_loop = get_loop(region);
619 assert(r_loop == get_loop(iff), "sanity");
620 // Always convert to CMOVE if all results are used only outside this loop.
621 bool used_inside_loop = (r_loop == _ltree_root);
622
623 // Check profitability
624 int cost = 0;
625 int phis = 0;
626 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
627 Node *out = region->fast_out(i);
628 if (!out->is_Phi()) continue; // Ignore other control edges, etc
629 phis++;
630 PhiNode* phi = out->as_Phi();
631 BasicType bt = phi->type()->basic_type();
632 switch (bt) {
633 case T_DOUBLE:
634 case T_FLOAT:
635 if (C->use_cmove()) {
636 continue; //TODO: maybe we want to add some cost
637 }
638 cost += Matcher::float_cmove_cost(); // Could be very expensive
639 break;
640 case T_LONG: {
641 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's
642 }
643 case T_INT: // These all CMOV fine
644 case T_ADDRESS: { // (RawPtr)
645 cost++;
646 break;
647 }
648 case T_NARROWOOP: // Fall through
649 case T_OBJECT: { // Base oops are OK, but not derived oops
650 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr();
651 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a
652 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus
653 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we
654 // have a Phi for the base here that we convert to a CMOVE all is well
655 // and good. But if the base is dead, we'll not make a CMOVE. Later
656 // the allocator will have to produce a base by creating a CMOVE of the
657 // relevant bases. This puts the allocator in the business of
658 // manufacturing expensive instructions, generally a bad plan.
659 // Just Say No to Conditionally-Moved Derived Pointers.
660 if (tp && tp->offset() != 0)
661 return NULL;
662 cost++;
663 break;
664 }
665 default:
666 return NULL; // In particular, can't do memory or I/O
667 }
668 // Add in cost any speculative ops
669 for (uint j = 1; j < region->req(); j++) {
670 Node *proj = region->in(j);
671 Node *inp = phi->in(j);
672 if (get_ctrl(inp) == proj) { // Found local op
673 cost++;
674 // Check for a chain of dependent ops; these will all become
675 // speculative in a CMOV.
676 for (uint k = 1; k < inp->req(); k++)
677 if (get_ctrl(inp->in(k)) == proj)
678 cost += ConditionalMoveLimit; // Too much speculative goo
679 }
680 }
681 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode.
682 // This will likely Split-If, a higher-payoff operation.
683 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) {
684 Node* use = phi->fast_out(k);
685 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr())
686 cost += ConditionalMoveLimit;
687 // Is there a use inside the loop?
688 // Note: check only basic types since CMoveP is pinned.
689 if (!used_inside_loop && is_java_primitive(bt)) {
690 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use);
691 if (r_loop == u_loop || r_loop->is_member(u_loop)) {
692 used_inside_loop = true;
693 }
694 }
695 }
696 }//for
697 Node* bol = iff->in(1);
698 if (bol->Opcode() == Op_Opaque4) {
699 return NULL; // Ignore loop predicate checks (the Opaque4 ensures they will go away)
700 }
701 assert(bol->Opcode() == Op_Bool, "Unexpected node");
702 int cmp_op = bol->in(1)->Opcode();
703 if (cmp_op == Op_SubTypeCheck) { // SubTypeCheck expansion expects an IfNode
704 return NULL;
705 }
706 // It is expensive to generate flags from a float compare.
707 // Avoid duplicated float compare.
708 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL;
709
710 float infrequent_prob = PROB_UNLIKELY_MAG(3);
711 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop.
712 if (used_inside_loop) {
713 if (cost >= ConditionalMoveLimit) return NULL; // Too much goo
714
715 // BlockLayoutByFrequency optimization moves infrequent branch
716 // from hot path. No point in CMOV'ing in such case (110 is used
717 // instead of 100 to take into account not exactness of float value).
718 if (BlockLayoutByFrequency) {
719 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f);
720 }
721 }
722 // Check for highly predictable branch. No point in CMOV'ing if
723 // we are going to predict accurately all the time.
724 if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) {
725 //keep going
726 } else if (iff->_prob < infrequent_prob ||
727 iff->_prob > (1.0f - infrequent_prob))
728 return NULL;
729
730 // --------------
731 // Now replace all Phis with CMOV's
732 Node *cmov_ctrl = iff->in(0);
733 uint flip = (lp->Opcode() == Op_IfTrue);
734 Node_List wq;
735 while (1) {
736 PhiNode* phi = NULL;
737 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
738 Node *out = region->fast_out(i);
739 if (out->is_Phi()) {
740 phi = out->as_Phi();
741 break;
742 }
743 }
744 if (phi == NULL || _igvn.type(phi) == Type::TOP) {
745 break;
746 }
747 if (PrintOpto && VerifyLoopOptimizations) { tty->print_cr("CMOV"); }
748 // Move speculative ops
749 wq.push(phi);
750 while (wq.size() > 0) {
751 Node *n = wq.pop();
752 for (uint j = 1; j < n->req(); j++) {
753 Node* m = n->in(j);
754 if (m != NULL && !is_dominator(get_ctrl(m), cmov_ctrl)) {
755 #ifndef PRODUCT
756 if (PrintOpto && VerifyLoopOptimizations) {
757 tty->print(" speculate: ");
758 m->dump();
759 }
760 #endif
761 set_ctrl(m, cmov_ctrl);
762 wq.push(m);
763 }
764 }
765 }
766 Node *cmov = CMoveNode::make(cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi));
767 register_new_node( cmov, cmov_ctrl );
768 _igvn.replace_node( phi, cmov );
769 #ifndef PRODUCT
770 if (TraceLoopOpts) {
771 tty->print("CMOV ");
772 r_loop->dump_head();
773 if (Verbose) {
774 bol->in(1)->dump(1);
775 cmov->dump(1);
776 }
777 }
778 if (VerifyLoopOptimizations) verify();
779 #endif
780 }
781
782 // The useless CFG diamond will fold up later; see the optimization in
783 // RegionNode::Ideal.
784 _igvn._worklist.push(region);
785
786 return iff->in(1);
787 }
788
789 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) {
790 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
791 Node* u = m->fast_out(i);
792 if (u->is_CFG()) {
793 if (u->Opcode() == Op_NeverBranch) {
794 u = ((NeverBranchNode*)u)->proj_out(0);
795 enqueue_cfg_uses(u, wq);
796 } else {
797 wq.push(u);
798 }
799 }
800 }
801 }
802
803 // Try moving a store out of a loop, right before the loop
804 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) {
805 // Store has to be first in the loop body
806 IdealLoopTree *n_loop = get_loop(n_ctrl);
807 if (n->is_Store() && n_loop != _ltree_root &&
808 n_loop->is_loop() && n_loop->_head->is_Loop() &&
809 n->in(0) != NULL) {
810 Node* address = n->in(MemNode::Address);
811 Node* value = n->in(MemNode::ValueIn);
812 Node* mem = n->in(MemNode::Memory);
813 IdealLoopTree* address_loop = get_loop(get_ctrl(address));
814 IdealLoopTree* value_loop = get_loop(get_ctrl(value));
815
816 // - address and value must be loop invariant
817 // - memory must be a memory Phi for the loop
818 // - Store must be the only store on this memory slice in the
819 // loop: if there's another store following this one then value
820 // written at iteration i by the second store could be overwritten
821 // at iteration i+n by the first store: it's not safe to move the
822 // first store out of the loop
823 // - nothing must observe the memory Phi: it guarantees no read
824 // before the store, we are also guaranteed the store post
825 // dominates the loop head (ignoring a possible early
826 // exit). Otherwise there would be extra Phi involved between the
827 // loop's Phi and the store.
828 // - there must be no early exit from the loop before the Store
829 // (such an exit most of the time would be an extra use of the
830 // memory Phi but sometimes is a bottom memory Phi that takes the
831 // store as input).
832
833 if (!n_loop->is_member(address_loop) &&
834 !n_loop->is_member(value_loop) &&
835 mem->is_Phi() && mem->in(0) == n_loop->_head &&
836 mem->outcnt() == 1 &&
837 mem->in(LoopNode::LoopBackControl) == n) {
838
839 assert(n_loop->_tail != NULL, "need a tail");
840 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop");
841
842 // Verify that there's no early exit of the loop before the store.
843 bool ctrl_ok = false;
844 {
845 // Follow control from loop head until n, we exit the loop or
846 // we reach the tail
847 ResourceMark rm;
848 Unique_Node_List wq;
849 wq.push(n_loop->_head);
850
851 for (uint next = 0; next < wq.size(); ++next) {
852 Node *m = wq.at(next);
853 if (m == n->in(0)) {
854 ctrl_ok = true;
855 continue;
856 }
857 assert(!has_ctrl(m), "should be CFG");
858 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) {
859 ctrl_ok = false;
860 break;
861 }
862 enqueue_cfg_uses(m, wq);
863 if (wq.size() > 10) {
864 ctrl_ok = false;
865 break;
866 }
867 }
868 }
869 if (ctrl_ok) {
870 // move the Store
871 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem);
872 _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl));
873 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl));
874 // Disconnect the phi now. An empty phi can confuse other
875 // optimizations in this pass of loop opts.
876 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl));
877 n_loop->_body.yank(mem);
878
879 set_ctrl_and_loop(n, n->in(0));
880
881 return n;
882 }
883 }
884 }
885 return NULL;
886 }
887
888 // Try moving a store out of a loop, right after the loop
889 void PhaseIdealLoop::try_move_store_after_loop(Node* n) {
890 if (n->is_Store() && n->in(0) != NULL) {
891 Node *n_ctrl = get_ctrl(n);
892 IdealLoopTree *n_loop = get_loop(n_ctrl);
893 // Store must be in a loop
894 if (n_loop != _ltree_root && !n_loop->_irreducible) {
895 Node* address = n->in(MemNode::Address);
896 Node* value = n->in(MemNode::ValueIn);
897 IdealLoopTree* address_loop = get_loop(get_ctrl(address));
898 // address must be loop invariant
899 if (!n_loop->is_member(address_loop)) {
900 // Store must be last on this memory slice in the loop and
901 // nothing in the loop must observe it
902 Node* phi = NULL;
903 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
904 Node* u = n->fast_out(i);
905 if (has_ctrl(u)) { // control use?
906 IdealLoopTree *u_loop = get_loop(get_ctrl(u));
907 if (!n_loop->is_member(u_loop)) {
908 continue;
909 }
910 if (u->is_Phi() && u->in(0) == n_loop->_head) {
911 assert(_igvn.type(u) == Type::MEMORY, "bad phi");
912 // multiple phis on the same slice are possible
913 if (phi != NULL) {
914 return;
915 }
916 phi = u;
917 continue;
918 }
919 }
920 return;
921 }
922 if (phi != NULL) {
923 // Nothing in the loop before the store (next iteration)
924 // must observe the stored value
925 bool mem_ok = true;
926 {
927 ResourceMark rm;
928 Unique_Node_List wq;
929 wq.push(phi);
930 for (uint next = 0; next < wq.size() && mem_ok; ++next) {
931 Node *m = wq.at(next);
932 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) {
933 Node* u = m->fast_out(i);
934 if (u->is_Store() || u->is_Phi()) {
935 if (u != n) {
936 wq.push(u);
937 mem_ok = (wq.size() <= 10);
938 }
939 } else {
940 mem_ok = false;
941 break;
942 }
943 }
944 }
945 }
946 if (mem_ok) {
947 // Move the store out of the loop if the LCA of all
948 // users (except for the phi) is outside the loop.
949 Node* hook = new Node(1);
950 hook->init_req(0, n_ctrl); // Add an input to prevent hook from being dead
951 _igvn.rehash_node_delayed(phi);
952 int count = phi->replace_edge(n, hook, &_igvn);
953 assert(count > 0, "inconsistent phi");
954
955 // Compute latest point this store can go
956 Node* lca = get_late_ctrl(n, get_ctrl(n));
957 if (lca->is_OuterStripMinedLoop()) {
958 lca = lca->in(LoopNode::EntryControl);
959 }
960 if (n_loop->is_member(get_loop(lca))) {
961 // LCA is in the loop - bail out
962 _igvn.replace_node(hook, n);
963 return;
964 }
965 #ifdef ASSERT
966 if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) {
967 assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined");
968 n_loop->_head->as_Loop()->verify_strip_mined(1);
969 Node* outer = n_loop->_head->as_CountedLoop()->outer_loop();
970 IdealLoopTree* outer_loop = get_loop(outer);
971 assert(n_loop->_parent == outer_loop, "broken loop tree");
972 assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state");
973 }
974 #endif
975 lca = place_outside_loop(lca, n_loop);
976 assert(!n_loop->is_member(get_loop(lca)), "control must not be back in the loop");
977 assert(get_loop(lca)->_nest < n_loop->_nest || lca->in(0)->Opcode() == Op_NeverBranch, "must not be moved into inner loop");
978
979 // Move store out of the loop
980 _igvn.replace_node(hook, n->in(MemNode::Memory));
981 _igvn.replace_input_of(n, 0, lca);
982 set_ctrl_and_loop(n, lca);
983
984 // Disconnect the phi now. An empty phi can confuse other
985 // optimizations in this pass of loop opts..
986 if (phi->in(LoopNode::LoopBackControl) == phi) {
987 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl));
988 n_loop->_body.yank(phi);
989 }
990 }
991 }
992 }
993 }
994 }
995 }
996
997 //------------------------------split_if_with_blocks_pre-----------------------
998 // Do the real work in a non-recursive function. Data nodes want to be
999 // cloned in the pre-order so they can feed each other nicely.
1000 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
1001 // Cloning these guys is unlikely to win
1002 int n_op = n->Opcode();
1003 if (n_op == Op_MergeMem) {
1004 return n;
1005 }
1006 if (n->is_Proj()) {
1007 return n;
1008 }
1009 // Do not clone-up CmpFXXX variations, as these are always
1010 // followed by a CmpI
1011 if (n->is_Cmp()) {
1012 return n;
1013 }
1014 // Attempt to use a conditional move instead of a phi/branch
1015 if (ConditionalMoveLimit > 0 && n_op == Op_Region) {
1016 Node *cmov = conditional_move( n );
1017 if (cmov) {
1018 return cmov;
1019 }
1020 }
1021 if (n->is_CFG() || n->is_LoadStore()) {
1022 return n;
1023 }
1024 if (n->is_Opaque1() || // Opaque nodes cannot be mod'd
1025 n_op == Op_Opaque2) {
1026 if (!C->major_progress()) { // If chance of no more loop opts...
1027 _igvn._worklist.push(n); // maybe we'll remove them
1028 }
1029 return n;
1030 }
1031
1032 if (n->is_Con()) {
1033 return n; // No cloning for Con nodes
1034 }
1035
1036 Node *n_ctrl = get_ctrl(n);
1037 if (!n_ctrl) {
1038 return n; // Dead node
1039 }
1040
1041 Node* res = try_move_store_before_loop(n, n_ctrl);
1042 if (res != NULL) {
1043 return n;
1044 }
1045
1046 // Attempt to remix address expressions for loop invariants
1047 Node *m = remix_address_expressions( n );
1048 if( m ) return m;
1049
1050 if (n_op == Op_AddI) {
1051 Node *nn = convert_add_to_muladd( n );
1052 if ( nn ) return nn;
1053 }
1054
1055 if (n->is_ConstraintCast()) {
1056 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this);
1057 // ConstraintCastNode::dominating_cast() uses node control input to determine domination.
1058 // Node control inputs don't necessarily agree with loop control info (due to
1059 // transformations happened in between), thus additional dominance check is needed
1060 // to keep loop info valid.
1061 if (dom_cast != NULL && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) {
1062 _igvn.replace_node(n, dom_cast);
1063 return dom_cast;
1064 }
1065 }
1066
1067 // Determine if the Node has inputs from some local Phi.
1068 // Returns the block to clone thru.
1069 Node *n_blk = has_local_phi_input( n );
1070 if( !n_blk ) return n;
1071
1072 // Do not clone the trip counter through on a CountedLoop
1073 // (messes up the canonical shape).
1074 if (((n_blk->is_CountedLoop() || (n_blk->is_Loop() && n_blk->as_Loop()->is_loop_nest_inner_loop())) && n->Opcode() == Op_AddI) ||
1075 (n_blk->is_LongCountedLoop() && n->Opcode() == Op_AddL)) {
1076 return n;
1077 }
1078
1079 // Check for having no control input; not pinned. Allow
1080 // dominating control.
1081 if (n->in(0)) {
1082 Node *dom = idom(n_blk);
1083 if (dom_lca(n->in(0), dom) != n->in(0)) {
1084 return n;
1085 }
1086 }
1087 // Policy: when is it profitable. You must get more wins than
1088 // policy before it is considered profitable. Policy is usually 0,
1089 // so 1 win is considered profitable. Big merges will require big
1090 // cloning, so get a larger policy.
1091 int policy = n_blk->req() >> 2;
1092
1093 // If the loop is a candidate for range check elimination,
1094 // delay splitting through it's phi until a later loop optimization
1095 if (n_blk->is_BaseCountedLoop()) {
1096 IdealLoopTree *lp = get_loop(n_blk);
1097 if (lp && lp->_rce_candidate) {
1098 return n;
1099 }
1100 }
1101
1102 if (must_throttle_split_if()) return n;
1103
1104 // Split 'n' through the merge point if it is profitable
1105 Node *phi = split_thru_phi( n, n_blk, policy );
1106 if (!phi) return n;
1107
1108 // Found a Phi to split thru!
1109 // Replace 'n' with the new phi
1110 _igvn.replace_node( n, phi );
1111 // Moved a load around the loop, 'en-registering' something.
1112 if (n_blk->is_Loop() && n->is_Load() &&
1113 !phi->in(LoopNode::LoopBackControl)->is_Load())
1114 C->set_major_progress();
1115
1116 return phi;
1117 }
1118
1119 static bool merge_point_too_heavy(Compile* C, Node* region) {
1120 // Bail out if the region and its phis have too many users.
1121 int weight = 0;
1122 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1123 weight += region->fast_out(i)->outcnt();
1124 }
1125 int nodes_left = C->max_node_limit() - C->live_nodes();
1126 if (weight * 8 > nodes_left) {
1127 if (PrintOpto) {
1128 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight);
1129 }
1130 return true;
1131 } else {
1132 return false;
1133 }
1134 }
1135
1136 static bool merge_point_safe(Node* region) {
1137 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode
1138 // having a PhiNode input. This sidesteps the dangerous case where the split
1139 // ConvI2LNode may become TOP if the input Value() does not
1140 // overlap the ConvI2L range, leaving a node which may not dominate its
1141 // uses.
1142 // A better fix for this problem can be found in the BugTraq entry, but
1143 // expediency for Mantis demands this hack.
1144 #ifdef _LP64
1145 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1146 Node* n = region->fast_out(i);
1147 if (n->is_Phi()) {
1148 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1149 Node* m = n->fast_out(j);
1150 if (m->Opcode() == Op_ConvI2L)
1151 return false;
1152 if (m->is_CastII()) {
1153 return false;
1154 }
1155 }
1156 }
1157 }
1158 #endif
1159 return true;
1160 }
1161
1162
1163 //------------------------------place_outside_loop---------------------------------
1164 // Place some computation outside of this loop on the path to the use passed as argument
1165 Node* PhaseIdealLoop::place_outside_loop(Node* useblock, IdealLoopTree* loop) const {
1166 Node* head = loop->_head;
1167 assert(!loop->is_member(get_loop(useblock)), "must be outside loop");
1168 if (head->is_Loop() && head->as_Loop()->is_strip_mined()) {
1169 loop = loop->_parent;
1170 assert(loop->_head->is_OuterStripMinedLoop(), "malformed strip mined loop");
1171 }
1172
1173 // Pick control right outside the loop
1174 for (;;) {
1175 Node* dom = idom(useblock);
1176 if (loop->is_member(get_loop(dom)) ||
1177 // NeverBranch nodes are not assigned to the loop when constructed
1178 (dom->Opcode() == Op_NeverBranch && loop->is_member(get_loop(dom->in(0))))) {
1179 break;
1180 }
1181 useblock = dom;
1182 }
1183 assert(find_non_split_ctrl(useblock) == useblock, "should be non split control");
1184 return useblock;
1185 }
1186
1187
1188 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) {
1189 if (!n->is_If() || n->is_BaseCountedLoopEnd()) {
1190 return false;
1191 }
1192 if (!n->in(0)->is_Region()) {
1193 return false;
1194 }
1195
1196 Node* region = n->in(0);
1197 Node* dom = idom(region);
1198 if (!dom->is_If() || dom->in(1) != n->in(1)) {
1199 return false;
1200 }
1201 IfNode* dom_if = dom->as_If();
1202 Node* proj_true = dom_if->proj_out(1);
1203 Node* proj_false = dom_if->proj_out(0);
1204
1205 for (uint i = 1; i < region->req(); i++) {
1206 if (is_dominator(proj_true, region->in(i))) {
1207 continue;
1208 }
1209 if (is_dominator(proj_false, region->in(i))) {
1210 continue;
1211 }
1212 return false;
1213 }
1214
1215 return true;
1216 }
1217
1218
1219 bool PhaseIdealLoop::can_split_if(Node* n_ctrl) {
1220 if (must_throttle_split_if()) {
1221 return false;
1222 }
1223
1224 // Do not do 'split-if' if irreducible loops are present.
1225 if (_has_irreducible_loops) {
1226 return false;
1227 }
1228
1229 if (merge_point_too_heavy(C, n_ctrl)) {
1230 return false;
1231 }
1232
1233 // Do not do 'split-if' if some paths are dead. First do dead code
1234 // elimination and then see if its still profitable.
1235 for (uint i = 1; i < n_ctrl->req(); i++) {
1236 if (n_ctrl->in(i) == C->top()) {
1237 return false;
1238 }
1239 }
1240
1241 // If trying to do a 'Split-If' at the loop head, it is only
1242 // profitable if the cmp folds up on BOTH paths. Otherwise we
1243 // risk peeling a loop forever.
1244
1245 // CNC - Disabled for now. Requires careful handling of loop
1246 // body selection for the cloned code. Also, make sure we check
1247 // for any input path not being in the same loop as n_ctrl. For
1248 // irreducible loops we cannot check for 'n_ctrl->is_Loop()'
1249 // because the alternative loop entry points won't be converted
1250 // into LoopNodes.
1251 IdealLoopTree *n_loop = get_loop(n_ctrl);
1252 for (uint j = 1; j < n_ctrl->req(); j++) {
1253 if (get_loop(n_ctrl->in(j)) != n_loop) {
1254 return false;
1255 }
1256 }
1257
1258 // Check for safety of the merge point.
1259 if (!merge_point_safe(n_ctrl)) {
1260 return false;
1261 }
1262
1263 return true;
1264 }
1265
1266 // Detect if the node is the inner strip-mined loop
1267 // Return: NULL if it's not the case, or the exit of outer strip-mined loop
1268 static Node* is_inner_of_stripmined_loop(const Node* out) {
1269 Node* out_le = NULL;
1270
1271 if (out->is_CountedLoopEnd()) {
1272 const CountedLoopNode* loop = out->as_CountedLoopEnd()->loopnode();
1273
1274 if (loop != NULL && loop->is_strip_mined()) {
1275 out_le = loop->in(LoopNode::EntryControl)->as_OuterStripMinedLoop()->outer_loop_exit();
1276 }
1277 }
1278
1279 return out_le;
1280 }
1281
1282 //------------------------------split_if_with_blocks_post----------------------
1283 // Do the real work in a non-recursive function. CFG hackery wants to be
1284 // in the post-order, so it can dirty the I-DOM info and not use the dirtied
1285 // info.
1286 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) {
1287
1288 // Cloning Cmp through Phi's involves the split-if transform.
1289 // FastLock is not used by an If
1290 if (n->is_Cmp() && !n->is_FastLock()) {
1291 Node *n_ctrl = get_ctrl(n);
1292 // Determine if the Node has inputs from some local Phi.
1293 // Returns the block to clone thru.
1294 Node *n_blk = has_local_phi_input(n);
1295 if (n_blk != n_ctrl) {
1296 return;
1297 }
1298
1299 if (!can_split_if(n_ctrl)) {
1300 return;
1301 }
1302
1303 if (n->outcnt() != 1) {
1304 return; // Multiple bool's from 1 compare?
1305 }
1306 Node *bol = n->unique_out();
1307 assert(bol->is_Bool(), "expect a bool here");
1308 if (bol->outcnt() != 1) {
1309 return;// Multiple branches from 1 compare?
1310 }
1311 Node *iff = bol->unique_out();
1312
1313 // Check some safety conditions
1314 if (iff->is_If()) { // Classic split-if?
1315 if (iff->in(0) != n_ctrl) {
1316 return; // Compare must be in same blk as if
1317 }
1318 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE
1319 // Can't split CMove with different control edge.
1320 if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) {
1321 return;
1322 }
1323 if (get_ctrl(iff->in(2)) == n_ctrl ||
1324 get_ctrl(iff->in(3)) == n_ctrl) {
1325 return; // Inputs not yet split-up
1326 }
1327 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) {
1328 return; // Loop-invar test gates loop-varying CMOVE
1329 }
1330 } else {
1331 return; // some other kind of node, such as an Allocate
1332 }
1333
1334 // When is split-if profitable? Every 'win' on means some control flow
1335 // goes dead, so it's almost always a win.
1336 int policy = 0;
1337 // Split compare 'n' through the merge point if it is profitable
1338 Node *phi = split_thru_phi( n, n_ctrl, policy);
1339 if (!phi) {
1340 return;
1341 }
1342
1343 // Found a Phi to split thru!
1344 // Replace 'n' with the new phi
1345 _igvn.replace_node(n, phi);
1346
1347 // Now split the bool up thru the phi
1348 Node *bolphi = split_thru_phi(bol, n_ctrl, -1);
1349 guarantee(bolphi != NULL, "null boolean phi node");
1350
1351 _igvn.replace_node(bol, bolphi);
1352 assert(iff->in(1) == bolphi, "");
1353
1354 if (bolphi->Value(&_igvn)->singleton()) {
1355 return;
1356 }
1357
1358 // Conditional-move? Must split up now
1359 if (!iff->is_If()) {
1360 Node *cmovphi = split_thru_phi(iff, n_ctrl, -1);
1361 _igvn.replace_node(iff, cmovphi);
1362 return;
1363 }
1364
1365 // Now split the IF
1366 do_split_if(iff);
1367 return;
1368 }
1369
1370 // Two identical ifs back to back can be merged
1371 if (try_merge_identical_ifs(n)) {
1372 return;
1373 }
1374
1375 // Check for an IF ready to split; one that has its
1376 // condition codes input coming from a Phi at the block start.
1377 int n_op = n->Opcode();
1378
1379 // Check for an IF being dominated by another IF same test
1380 if (n_op == Op_If ||
1381 n_op == Op_RangeCheck) {
1382 Node *bol = n->in(1);
1383 uint max = bol->outcnt();
1384 // Check for same test used more than once?
1385 if (max > 1 && bol->is_Bool()) {
1386 // Search up IDOMs to see if this IF is dominated.
1387 Node *cutoff = get_ctrl(bol);
1388
1389 // Now search up IDOMs till cutoff, looking for a dominating test
1390 Node *prevdom = n;
1391 Node *dom = idom(prevdom);
1392 while (dom != cutoff) {
1393 if (dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom &&
1394 safe_for_if_replacement(dom)) {
1395 // It's invalid to move control dependent data nodes in the inner
1396 // strip-mined loop, because:
1397 // 1) break validation of LoopNode::verify_strip_mined()
1398 // 2) move code with side-effect in strip-mined loop
1399 // Move to the exit of outer strip-mined loop in that case.
1400 Node* out_le = is_inner_of_stripmined_loop(dom);
1401 if (out_le != NULL) {
1402 prevdom = out_le;
1403 }
1404 // Replace the dominated test with an obvious true or false.
1405 // Place it on the IGVN worklist for later cleanup.
1406 C->set_major_progress();
1407 dominated_by(prevdom->as_IfProj(), n->as_If(), false, true);
1408 #ifndef PRODUCT
1409 if( VerifyLoopOptimizations ) verify();
1410 #endif
1411 return;
1412 }
1413 prevdom = dom;
1414 dom = idom(prevdom);
1415 }
1416 }
1417 }
1418
1419 try_sink_out_of_loop(n);
1420
1421 try_move_store_after_loop(n);
1422
1423 // Check for Opaque2's who's loop has disappeared - who's input is in the
1424 // same loop nest as their output. Remove 'em, they are no longer useful.
1425 if( n_op == Op_Opaque2 &&
1426 n->in(1) != NULL &&
1427 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) {
1428 _igvn.replace_node( n, n->in(1) );
1429 }
1430 }
1431
1432 // Tranform:
1433 //
1434 // if (some_condition) {
1435 // // body 1
1436 // } else {
1437 // // body 2
1438 // }
1439 // if (some_condition) {
1440 // // body 3
1441 // } else {
1442 // // body 4
1443 // }
1444 //
1445 // into:
1446 //
1447 //
1448 // if (some_condition) {
1449 // // body 1
1450 // // body 3
1451 // } else {
1452 // // body 2
1453 // // body 4
1454 // }
1455 bool PhaseIdealLoop::try_merge_identical_ifs(Node* n) {
1456 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) {
1457 Node *n_ctrl = n->in(0);
1458 IfNode* dom_if = idom(n_ctrl)->as_If();
1459 ProjNode* dom_proj_true = dom_if->proj_out(1);
1460 ProjNode* dom_proj_false = dom_if->proj_out(0);
1461
1462 // Now split the IF
1463 RegionNode* new_false_region;
1464 RegionNode* new_true_region;
1465 do_split_if(n, &new_false_region, &new_true_region);
1466 assert(new_false_region->req() == new_true_region->req(), "");
1467 #ifdef ASSERT
1468 for (uint i = 1; i < new_false_region->req(); ++i) {
1469 assert(new_false_region->in(i)->in(0) == new_true_region->in(i)->in(0), "unexpected shape following split if");
1470 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");
1471 }
1472 #endif
1473 assert(new_false_region->in(1)->in(0)->in(1) == dom_if->in(1), "dominating if and dominated if after split must share test");
1474
1475 // We now have:
1476 // if (some_condition) {
1477 // // body 1
1478 // if (some_condition) {
1479 // body3: // new_true_region
1480 // // body3
1481 // } else {
1482 // goto body4;
1483 // }
1484 // } else {
1485 // // body 2
1486 // if (some_condition) {
1487 // goto body3;
1488 // } else {
1489 // body4: // new_false_region
1490 // // body4;
1491 // }
1492 // }
1493 //
1494
1495 // clone pinned nodes thru the resulting regions
1496 push_pinned_nodes_thru_region(dom_if, new_true_region);
1497 push_pinned_nodes_thru_region(dom_if, new_false_region);
1498
1499 // Optimize out the cloned ifs. Because pinned nodes were cloned, this also allows a CastPP that would be dependent
1500 // 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
1501 // unrelated control dependency.
1502 for (uint i = 1; i < new_false_region->req(); i++) {
1503 if (is_dominator(dom_proj_true, new_false_region->in(i))) {
1504 dominated_by(dom_proj_true->as_IfProj(), new_false_region->in(i)->in(0)->as_If(), false, false);
1505 } else {
1506 assert(is_dominator(dom_proj_false, new_false_region->in(i)), "bad if");
1507 dominated_by(dom_proj_false->as_IfProj(), new_false_region->in(i)->in(0)->as_If(), false, false);
1508 }
1509 }
1510 return true;
1511 }
1512 return false;
1513 }
1514
1515 void PhaseIdealLoop::push_pinned_nodes_thru_region(IfNode* dom_if, Node* region) {
1516 for (DUIterator i = region->outs(); region->has_out(i); i++) {
1517 Node* u = region->out(i);
1518 if (!has_ctrl(u) || u->is_Phi() || !u->depends_only_on_test() || !_igvn.no_dependent_zero_check(u)) {
1519 continue;
1520 }
1521 assert(u->in(0) == region, "not a control dependent node?");
1522 uint j = 1;
1523 for (; j < u->req(); ++j) {
1524 Node* in = u->in(j);
1525 if (!is_dominator(ctrl_or_self(in), dom_if)) {
1526 break;
1527 }
1528 }
1529 if (j == u->req()) {
1530 Node *phi = PhiNode::make_blank(region, u);
1531 for (uint k = 1; k < region->req(); ++k) {
1532 Node* clone = u->clone();
1533 clone->set_req(0, region->in(k));
1534 register_new_node(clone, region->in(k));
1535 phi->init_req(k, clone);
1536 }
1537 register_new_node(phi, region);
1538 _igvn.replace_node(u, phi);
1539 --i;
1540 }
1541 }
1542 }
1543
1544 bool PhaseIdealLoop::safe_for_if_replacement(const Node* dom) const {
1545 if (!dom->is_CountedLoopEnd()) {
1546 return true;
1547 }
1548 CountedLoopEndNode* le = dom->as_CountedLoopEnd();
1549 CountedLoopNode* cl = le->loopnode();
1550 if (cl == NULL) {
1551 return true;
1552 }
1553 if (!cl->is_main_loop()) {
1554 return true;
1555 }
1556 if (cl->is_canonical_loop_entry() == NULL) {
1557 return true;
1558 }
1559 // Further unrolling is possible so loop exit condition might change
1560 return false;
1561 }
1562
1563 // See if a shared loop-varying computation has no loop-varying uses.
1564 // Happens if something is only used for JVM state in uncommon trap exits,
1565 // like various versions of induction variable+offset. Clone the
1566 // computation per usage to allow it to sink out of the loop.
1567 void PhaseIdealLoop::try_sink_out_of_loop(Node* n) {
1568 bool is_raw_to_oop_cast = n->is_ConstraintCast() &&
1569 n->in(1)->bottom_type()->isa_rawptr() &&
1570 !n->bottom_type()->isa_rawptr();
1571 if (has_ctrl(n) &&
1572 !n->is_Phi() &&
1573 !n->is_Bool() &&
1574 !n->is_Proj() &&
1575 !n->is_MergeMem() &&
1576 !n->is_CMove() &&
1577 !is_raw_to_oop_cast && // don't extend live ranges of raw oops
1578 n->Opcode() != Op_Opaque4 &&
1579 !n->is_Type()) {
1580 Node *n_ctrl = get_ctrl(n);
1581 IdealLoopTree *n_loop = get_loop(n_ctrl);
1582
1583 if (n->in(0) != NULL) {
1584 IdealLoopTree* loop_ctrl = get_loop(n->in(0));
1585 if (n_loop != loop_ctrl && n_loop->is_member(loop_ctrl)) {
1586 // n has a control input inside a loop but get_ctrl() is member of an outer loop. This could happen, for example,
1587 // for Div nodes inside a loop (control input inside loop) without a use except for an UCT (outside the loop).
1588 // Rewire control of n to right outside of the loop, regardless if its input(s) are later sunk or not.
1589 _igvn.replace_input_of(n, 0, place_outside_loop(n_ctrl, loop_ctrl));
1590 }
1591 }
1592 if (n_loop != _ltree_root && n->outcnt() > 1) {
1593 // Compute early control: needed for anti-dependence analysis. It's also possible that as a result of
1594 // previous transformations in this loop opts round, the node can be hoisted now: early control will tell us.
1595 Node* early_ctrl = compute_early_ctrl(n, n_ctrl);
1596 if (n_loop->is_member(get_loop(early_ctrl)) && // check that this one can't be hoisted now
1597 ctrl_of_all_uses_out_of_loop(n, early_ctrl, n_loop)) { // All uses in outer loops!
1598 assert(!n->is_Store() && !n->is_LoadStore(), "no node with a side effect");
1599 Node* outer_loop_clone = NULL;
1600 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin;) {
1601 Node* u = n->last_out(j); // Clone private computation per use
1602 _igvn.rehash_node_delayed(u);
1603 Node* x = n->clone(); // Clone computation
1604 Node* x_ctrl = NULL;
1605 if (u->is_Phi()) {
1606 // Replace all uses of normal nodes. Replace Phi uses
1607 // individually, so the separate Nodes can sink down
1608 // different paths.
1609 uint k = 1;
1610 while (u->in(k) != n) k++;
1611 u->set_req(k, x);
1612 // x goes next to Phi input path
1613 x_ctrl = u->in(0)->in(k);
1614 // Find control for 'x' next to use but not inside inner loops.
1615 x_ctrl = place_outside_loop(x_ctrl, n_loop);
1616 --j;
1617 } else { // Normal use
1618 if (has_ctrl(u)) {
1619 x_ctrl = get_ctrl(u);
1620 } else {
1621 x_ctrl = u->in(0);
1622 }
1623 // Find control for 'x' next to use but not inside inner loops.
1624 x_ctrl = place_outside_loop(x_ctrl, n_loop);
1625 // Replace all uses
1626 if (u->is_ConstraintCast() && u->bottom_type()->higher_equal(_igvn.type(n)) && u->in(0) == x_ctrl) {
1627 // If we're sinking a chain of data nodes, we might have inserted a cast to pin the use which is not necessary
1628 // anymore now that we're going to pin n as well
1629 _igvn.replace_node(u, x);
1630 --j;
1631 } else {
1632 int nb = u->replace_edge(n, x, &_igvn);
1633 j -= nb;
1634 }
1635 }
1636
1637 if (n->is_Load()) {
1638 // For loads, add a control edge to a CFG node outside of the loop
1639 // to force them to not combine and return back inside the loop
1640 // during GVN optimization (4641526).
1641 assert(x_ctrl == get_late_ctrl_with_anti_dep(x->as_Load(), early_ctrl, x_ctrl), "anti-dependences were already checked");
1642
1643 IdealLoopTree* x_loop = get_loop(x_ctrl);
1644 Node* x_head = x_loop->_head;
1645 if (x_head->is_Loop() && x_head->is_OuterStripMinedLoop()) {
1646 // Do not add duplicate LoadNodes to the outer strip mined loop
1647 if (outer_loop_clone != NULL) {
1648 _igvn.replace_node(x, outer_loop_clone);
1649 continue;
1650 }
1651 outer_loop_clone = x;
1652 }
1653 x->set_req(0, x_ctrl);
1654 } else if (n->in(0) != NULL){
1655 x->set_req(0, x_ctrl);
1656 }
1657 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone");
1658 assert(!n_loop->is_member(get_loop(x_ctrl)), "should have moved out of loop");
1659 register_new_node(x, x_ctrl);
1660
1661 // Chain of AddP: (AddP base (AddP base )) must keep the same base after sinking so:
1662 // 1- We don't add a CastPP here when the first one is sunk so if the second one is not, their bases remain
1663 // the same.
1664 // (see 2- below)
1665 assert(!x->is_AddP() || !x->in(AddPNode::Address)->is_AddP() ||
1666 x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base) ||
1667 !x->in(AddPNode::Address)->in(AddPNode::Base)->eqv_uncast(x->in(AddPNode::Base)), "unexpected AddP shape");
1668 if (x->in(0) == NULL && !x->is_DecodeNarrowPtr() &&
1669 !(x->is_AddP() && x->in(AddPNode::Address)->is_AddP() && x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base))) {
1670 assert(!x->is_Load(), "load should be pinned");
1671 // Use a cast node to pin clone out of loop
1672 Node* cast = NULL;
1673 for (uint k = 0; k < x->req(); k++) {
1674 Node* in = x->in(k);
1675 if (in != NULL && n_loop->is_member(get_loop(get_ctrl(in)))) {
1676 const Type* in_t = _igvn.type(in);
1677 cast = ConstraintCastNode::make_cast_for_type(x_ctrl, in, in_t, ConstraintCastNode::UnconditionalDependency);
1678 }
1679 if (cast != NULL) {
1680 register_new_node(cast, x_ctrl);
1681 x->replace_edge(in, cast);
1682 // Chain of AddP:
1683 // 2- A CastPP of the base is only added now that both AddP nodes are sunk
1684 if (x->is_AddP() && k == AddPNode::Base) {
1685 for (DUIterator_Fast imax, i = x->fast_outs(imax); i < imax; i++) {
1686 Node* u = x->fast_out(i);
1687 if (u->is_AddP() && u->in(AddPNode::Base) == n->in(AddPNode::Base)) {
1688 _igvn.replace_input_of(u, AddPNode::Base, cast);
1689 assert(u->find_out_with(Op_AddP) == NULL, "more than 2 chained AddP nodes?");
1690 }
1691 }
1692 }
1693 break;
1694 }
1695 }
1696 assert(cast != NULL, "must have added a cast to pin the node");
1697 }
1698 }
1699 _igvn.remove_dead_node(n);
1700 }
1701 _dom_lca_tags_round = 0;
1702 }
1703 }
1704 }
1705
1706 Node* PhaseIdealLoop::compute_early_ctrl(Node* n, Node* n_ctrl) {
1707 Node* early_ctrl = NULL;
1708 ResourceMark rm;
1709 Unique_Node_List wq;
1710 wq.push(n);
1711 for (uint i = 0; i < wq.size(); i++) {
1712 Node* m = wq.at(i);
1713 Node* c = NULL;
1714 if (m->is_CFG()) {
1715 c = m;
1716 } else if (m->pinned()) {
1717 c = m->in(0);
1718 } else {
1719 for (uint j = 0; j < m->req(); j++) {
1720 Node* in = m->in(j);
1721 if (in == NULL) {
1722 continue;
1723 }
1724 wq.push(in);
1725 }
1726 }
1727 if (c != NULL) {
1728 assert(is_dominator(c, n_ctrl), "");
1729 if (early_ctrl == NULL) {
1730 early_ctrl = c;
1731 } else if (is_dominator(early_ctrl, c)) {
1732 early_ctrl = c;
1733 }
1734 }
1735 }
1736 assert(is_dominator(early_ctrl, n_ctrl), "early control must dominate current control");
1737 return early_ctrl;
1738 }
1739
1740 bool PhaseIdealLoop::ctrl_of_all_uses_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop) {
1741 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1742 Node* u = n->fast_out(i);
1743 if (u->Opcode() == Op_Opaque1) {
1744 return false; // Found loop limit, bugfix for 4677003
1745 }
1746 // We can't reuse tags in PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal() so make sure calls to
1747 // get_late_ctrl_with_anti_dep() use their own tag
1748 _dom_lca_tags_round++;
1749 assert(_dom_lca_tags_round != 0, "shouldn't wrap around");
1750
1751 if (u->is_Phi()) {
1752 for (uint j = 1; j < u->req(); ++j) {
1753 if (u->in(j) == n && !ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, u->in(0)->in(j))) {
1754 return false;
1755 }
1756 }
1757 } else {
1758 Node* ctrl = has_ctrl(u) ? get_ctrl(u) : u->in(0);
1759 if (!ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, ctrl)) {
1760 return false;
1761 }
1762 }
1763 }
1764 return true;
1765 }
1766
1767 bool PhaseIdealLoop::ctrl_of_use_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop, Node* ctrl) {
1768 if (n->is_Load()) {
1769 ctrl = get_late_ctrl_with_anti_dep(n->as_Load(), n_ctrl, ctrl);
1770 }
1771 IdealLoopTree *u_loop = get_loop(ctrl);
1772 if (u_loop == n_loop) {
1773 return false; // Found loop-varying use
1774 }
1775 if (n_loop->is_member(u_loop)) {
1776 return false; // Found use in inner loop
1777 }
1778 return true;
1779 }
1780
1781 //------------------------------split_if_with_blocks---------------------------
1782 // Check for aggressive application of 'split-if' optimization,
1783 // using basic block level info.
1784 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack) {
1785 Node* root = C->root();
1786 visited.set(root->_idx); // first, mark root as visited
1787 // Do pre-visit work for root
1788 Node* n = split_if_with_blocks_pre(root);
1789 uint cnt = n->outcnt();
1790 uint i = 0;
1791
1792 while (true) {
1793 // Visit all children
1794 if (i < cnt) {
1795 Node* use = n->raw_out(i);
1796 ++i;
1797 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
1798 // Now do pre-visit work for this use
1799 use = split_if_with_blocks_pre(use);
1800 nstack.push(n, i); // Save parent and next use's index.
1801 n = use; // Process all children of current use.
1802 cnt = use->outcnt();
1803 i = 0;
1804 }
1805 }
1806 else {
1807 // All of n's children have been processed, complete post-processing.
1808 if (cnt != 0 && !n->is_Con()) {
1809 assert(has_node(n), "no dead nodes");
1810 split_if_with_blocks_post(n);
1811 }
1812 if (must_throttle_split_if()) {
1813 nstack.clear();
1814 }
1815 if (nstack.is_empty()) {
1816 // Finished all nodes on stack.
1817 break;
1818 }
1819 // Get saved parent node and next use's index. Visit the rest of uses.
1820 n = nstack.node();
1821 cnt = n->outcnt();
1822 i = nstack.index();
1823 nstack.pop();
1824 }
1825 }
1826 }
1827
1828
1829 //=============================================================================
1830 //
1831 // C L O N E A L O O P B O D Y
1832 //
1833
1834 //------------------------------clone_iff--------------------------------------
1835 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1836 // "Nearly" because all Nodes have been cloned from the original in the loop,
1837 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1838 // through the Phi recursively, and return a Bool.
1839 Node* PhaseIdealLoop::clone_iff(PhiNode *phi, IdealLoopTree *loop) {
1840
1841 // Convert this Phi into a Phi merging Bools
1842 uint i;
1843 for (i = 1; i < phi->req(); i++) {
1844 Node *b = phi->in(i);
1845 if (b->is_Phi()) {
1846 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi(), loop));
1847 } else {
1848 assert(b->is_Bool() || b->Opcode() == Op_Opaque4, "");
1849 }
1850 }
1851
1852 Node* n = phi->in(1);
1853 Node* sample_opaque = NULL;
1854 Node *sample_bool = NULL;
1855 if (n->Opcode() == Op_Opaque4) {
1856 sample_opaque = n;
1857 sample_bool = n->in(1);
1858 assert(sample_bool->is_Bool(), "wrong type");
1859 } else {
1860 sample_bool = n;
1861 }
1862 Node *sample_cmp = sample_bool->in(1);
1863
1864 // Make Phis to merge the Cmp's inputs.
1865 PhiNode *phi1 = new PhiNode(phi->in(0), Type::TOP);
1866 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP);
1867 for (i = 1; i < phi->req(); i++) {
1868 Node *n1 = sample_opaque == NULL ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1);
1869 Node *n2 = sample_opaque == NULL ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2);
1870 phi1->set_req(i, n1);
1871 phi2->set_req(i, n2);
1872 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
1873 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
1874 }
1875 // See if these Phis have been made before.
1876 // Register with optimizer
1877 Node *hit1 = _igvn.hash_find_insert(phi1);
1878 if (hit1) { // Hit, toss just made Phi
1879 _igvn.remove_dead_node(phi1); // Remove new phi
1880 assert(hit1->is_Phi(), "" );
1881 phi1 = (PhiNode*)hit1; // Use existing phi
1882 } else { // Miss
1883 _igvn.register_new_node_with_optimizer(phi1);
1884 }
1885 Node *hit2 = _igvn.hash_find_insert(phi2);
1886 if (hit2) { // Hit, toss just made Phi
1887 _igvn.remove_dead_node(phi2); // Remove new phi
1888 assert(hit2->is_Phi(), "" );
1889 phi2 = (PhiNode*)hit2; // Use existing phi
1890 } else { // Miss
1891 _igvn.register_new_node_with_optimizer(phi2);
1892 }
1893 // Register Phis with loop/block info
1894 set_ctrl(phi1, phi->in(0));
1895 set_ctrl(phi2, phi->in(0));
1896 // Make a new Cmp
1897 Node *cmp = sample_cmp->clone();
1898 cmp->set_req(1, phi1);
1899 cmp->set_req(2, phi2);
1900 _igvn.register_new_node_with_optimizer(cmp);
1901 set_ctrl(cmp, phi->in(0));
1902
1903 // Make a new Bool
1904 Node *b = sample_bool->clone();
1905 b->set_req(1,cmp);
1906 _igvn.register_new_node_with_optimizer(b);
1907 set_ctrl(b, phi->in(0));
1908
1909 if (sample_opaque != NULL) {
1910 Node* opaque = sample_opaque->clone();
1911 opaque->set_req(1, b);
1912 _igvn.register_new_node_with_optimizer(opaque);
1913 set_ctrl(opaque, phi->in(0));
1914 return opaque;
1915 }
1916
1917 assert(b->is_Bool(), "");
1918 return b;
1919 }
1920
1921 //------------------------------clone_bool-------------------------------------
1922 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1923 // "Nearly" because all Nodes have been cloned from the original in the loop,
1924 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1925 // through the Phi recursively, and return a Bool.
1926 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) {
1927 uint i;
1928 // Convert this Phi into a Phi merging Bools
1929 for( i = 1; i < phi->req(); i++ ) {
1930 Node *b = phi->in(i);
1931 if( b->is_Phi() ) {
1932 _igvn.replace_input_of(phi, i, clone_bool( b->as_Phi(), loop ));
1933 } else {
1934 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
1935 }
1936 }
1937
1938 Node *sample_cmp = phi->in(1);
1939
1940 // Make Phis to merge the Cmp's inputs.
1941 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP );
1942 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP );
1943 for( uint j = 1; j < phi->req(); j++ ) {
1944 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
1945 Node *n1, *n2;
1946 if( cmp_top->is_Cmp() ) {
1947 n1 = cmp_top->in(1);
1948 n2 = cmp_top->in(2);
1949 } else {
1950 n1 = n2 = cmp_top;
1951 }
1952 phi1->set_req( j, n1 );
1953 phi2->set_req( j, n2 );
1954 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
1955 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
1956 }
1957
1958 // See if these Phis have been made before.
1959 // Register with optimizer
1960 Node *hit1 = _igvn.hash_find_insert(phi1);
1961 if( hit1 ) { // Hit, toss just made Phi
1962 _igvn.remove_dead_node(phi1); // Remove new phi
1963 assert( hit1->is_Phi(), "" );
1964 phi1 = (PhiNode*)hit1; // Use existing phi
1965 } else { // Miss
1966 _igvn.register_new_node_with_optimizer(phi1);
1967 }
1968 Node *hit2 = _igvn.hash_find_insert(phi2);
1969 if( hit2 ) { // Hit, toss just made Phi
1970 _igvn.remove_dead_node(phi2); // Remove new phi
1971 assert( hit2->is_Phi(), "" );
1972 phi2 = (PhiNode*)hit2; // Use existing phi
1973 } else { // Miss
1974 _igvn.register_new_node_with_optimizer(phi2);
1975 }
1976 // Register Phis with loop/block info
1977 set_ctrl(phi1, phi->in(0));
1978 set_ctrl(phi2, phi->in(0));
1979 // Make a new Cmp
1980 Node *cmp = sample_cmp->clone();
1981 cmp->set_req( 1, phi1 );
1982 cmp->set_req( 2, phi2 );
1983 _igvn.register_new_node_with_optimizer(cmp);
1984 set_ctrl(cmp, phi->in(0));
1985
1986 assert( cmp->is_Cmp(), "" );
1987 return (CmpNode*)cmp;
1988 }
1989
1990 //------------------------------sink_use---------------------------------------
1991 // If 'use' was in the loop-exit block, it now needs to be sunk
1992 // below the post-loop merge point.
1993 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) {
1994 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) {
1995 set_ctrl(use, post_loop);
1996 for (DUIterator j = use->outs(); use->has_out(j); j++)
1997 sink_use(use->out(j), post_loop);
1998 }
1999 }
2000
2001 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new,
2002 IdealLoopTree* loop, IdealLoopTree* outer_loop,
2003 Node_List*& split_if_set, Node_List*& split_bool_set,
2004 Node_List*& split_cex_set, Node_List& worklist,
2005 uint new_counter, CloneLoopMode mode) {
2006 Node* nnn = old_new[old->_idx];
2007 // Copy uses to a worklist, so I can munge the def-use info
2008 // with impunity.
2009 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
2010 worklist.push(old->fast_out(j));
2011
2012 while( worklist.size() ) {
2013 Node *use = worklist.pop();
2014 if (!has_node(use)) continue; // Ignore dead nodes
2015 if (use->in(0) == C->top()) continue;
2016 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
2017 // Check for data-use outside of loop - at least one of OLD or USE
2018 // must not be a CFG node.
2019 #ifdef ASSERT
2020 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == NULL) {
2021 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint();
2022 assert(mode != IgnoreStripMined, "incorrect cloning mode");
2023 assert((mode == ControlAroundStripMined && use == sfpt) || !use->is_reachable_from_root(), "missed a node");
2024 }
2025 #endif
2026 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) {
2027
2028 // If the Data use is an IF, that means we have an IF outside of the
2029 // loop that is switching on a condition that is set inside of the
2030 // loop. Happens if people set a loop-exit flag; then test the flag
2031 // in the loop to break the loop, then test is again outside of the
2032 // loop to determine which way the loop exited.
2033 // Loop predicate If node connects to Bool node through Opaque1 node.
2034 if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use) || use->Opcode() == Op_Opaque4) {
2035 // Since this code is highly unlikely, we lazily build the worklist
2036 // of such Nodes to go split.
2037 if (!split_if_set) {
2038 split_if_set = new Node_List();
2039 }
2040 split_if_set->push(use);
2041 }
2042 if (use->is_Bool()) {
2043 if (!split_bool_set) {
2044 split_bool_set = new Node_List();
2045 }
2046 split_bool_set->push(use);
2047 }
2048 if (use->Opcode() == Op_CreateEx) {
2049 if (!split_cex_set) {
2050 split_cex_set = new Node_List();
2051 }
2052 split_cex_set->push(use);
2053 }
2054
2055
2056 // Get "block" use is in
2057 uint idx = 0;
2058 while( use->in(idx) != old ) idx++;
2059 Node *prev = use->is_CFG() ? use : get_ctrl(use);
2060 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" );
2061 Node *cfg = prev->_idx >= new_counter
2062 ? prev->in(2)
2063 : idom(prev);
2064 if( use->is_Phi() ) // Phi use is in prior block
2065 cfg = prev->in(idx); // NOT in block of Phi itself
2066 if (cfg->is_top()) { // Use is dead?
2067 _igvn.replace_input_of(use, idx, C->top());
2068 continue;
2069 }
2070
2071 // If use is referenced through control edge... (idx == 0)
2072 if (mode == IgnoreStripMined && idx == 0) {
2073 LoopNode *head = loop->_head->as_Loop();
2074 if (head->is_strip_mined() && is_dominator(head->outer_loop_exit(), prev)) {
2075 // That node is outside the inner loop, leave it outside the
2076 // outer loop as well to not confuse verification code.
2077 assert(!loop->_parent->is_member(use_loop), "should be out of the outer loop");
2078 _igvn.replace_input_of(use, 0, head->outer_loop_exit());
2079 continue;
2080 }
2081 }
2082
2083 while(!outer_loop->is_member(get_loop(cfg))) {
2084 prev = cfg;
2085 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg);
2086 }
2087 // If the use occurs after merging several exits from the loop, then
2088 // old value must have dominated all those exits. Since the same old
2089 // value was used on all those exits we did not need a Phi at this
2090 // merge point. NOW we do need a Phi here. Each loop exit value
2091 // is now merged with the peeled body exit; each exit gets its own
2092 // private Phi and those Phis need to be merged here.
2093 Node *phi;
2094 if( prev->is_Region() ) {
2095 if( idx == 0 ) { // Updating control edge?
2096 phi = prev; // Just use existing control
2097 } else { // Else need a new Phi
2098 phi = PhiNode::make( prev, old );
2099 // Now recursively fix up the new uses of old!
2100 for( uint i = 1; i < prev->req(); i++ ) {
2101 worklist.push(phi); // Onto worklist once for each 'old' input
2102 }
2103 }
2104 } else {
2105 // Get new RegionNode merging old and new loop exits
2106 prev = old_new[prev->_idx];
2107 assert( prev, "just made this in step 7" );
2108 if( idx == 0) { // Updating control edge?
2109 phi = prev; // Just use existing control
2110 } else { // Else need a new Phi
2111 // Make a new Phi merging data values properly
2112 phi = PhiNode::make( prev, old );
2113 phi->set_req( 1, nnn );
2114 }
2115 }
2116 // If inserting a new Phi, check for prior hits
2117 if( idx != 0 ) {
2118 Node *hit = _igvn.hash_find_insert(phi);
2119 if( hit == NULL ) {
2120 _igvn.register_new_node_with_optimizer(phi); // Register new phi
2121 } else { // or
2122 // Remove the new phi from the graph and use the hit
2123 _igvn.remove_dead_node(phi);
2124 phi = hit; // Use existing phi
2125 }
2126 set_ctrl(phi, prev);
2127 }
2128 // Make 'use' use the Phi instead of the old loop body exit value
2129 _igvn.replace_input_of(use, idx, phi);
2130 if( use->_idx >= new_counter ) { // If updating new phis
2131 // Not needed for correctness, but prevents a weak assert
2132 // in AddPNode from tripping (when we end up with different
2133 // base & derived Phis that will become the same after
2134 // IGVN does CSE).
2135 Node *hit = _igvn.hash_find_insert(use);
2136 if( hit ) // Go ahead and re-hash for hits.
2137 _igvn.replace_node( use, hit );
2138 }
2139
2140 // If 'use' was in the loop-exit block, it now needs to be sunk
2141 // below the post-loop merge point.
2142 sink_use( use, prev );
2143 }
2144 }
2145 }
2146
2147 static void clone_outer_loop_helper(Node* n, const IdealLoopTree *loop, const IdealLoopTree* outer_loop,
2148 const Node_List &old_new, Unique_Node_List& wq, PhaseIdealLoop* phase,
2149 bool check_old_new) {
2150 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2151 Node* u = n->fast_out(j);
2152 assert(check_old_new || old_new[u->_idx] == NULL, "shouldn't have been cloned");
2153 if (!u->is_CFG() && (!check_old_new || old_new[u->_idx] == NULL)) {
2154 Node* c = phase->get_ctrl(u);
2155 IdealLoopTree* u_loop = phase->get_loop(c);
2156 assert(!loop->is_member(u_loop), "can be in outer loop or out of both loops only");
2157 if (outer_loop->is_member(u_loop) ||
2158 // nodes pinned with control in the outer loop but not referenced from the safepoint must be moved out of
2159 // the outer loop too
2160 (u->in(0) != NULL && outer_loop->is_member(phase->get_loop(u->in(0))))) {
2161 wq.push(u);
2162 }
2163 }
2164 }
2165 }
2166
2167 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
2168 IdealLoopTree* outer_loop, int dd, Node_List &old_new,
2169 Node_List& extra_data_nodes) {
2170 if (head->is_strip_mined() && mode != IgnoreStripMined) {
2171 CountedLoopNode* cl = head->as_CountedLoop();
2172 Node* l = cl->outer_loop();
2173 Node* tail = cl->outer_loop_tail();
2174 IfNode* le = cl->outer_loop_end();
2175 Node* sfpt = cl->outer_safepoint();
2176 CountedLoopEndNode* cle = cl->loopexit();
2177 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop();
2178 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null();
2179 Node* cle_out = cle->proj_out(false);
2180
2181 Node* new_sfpt = NULL;
2182 Node* new_cle_out = cle_out->clone();
2183 old_new.map(cle_out->_idx, new_cle_out);
2184 if (mode == CloneIncludesStripMined) {
2185 // clone outer loop body
2186 Node* new_l = l->clone();
2187 Node* new_tail = tail->clone();
2188 IfNode* new_le = le->clone()->as_If();
2189 new_sfpt = sfpt->clone();
2190
2191 set_loop(new_l, outer_loop->_parent);
2192 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd);
2193 set_loop(new_cle_out, outer_loop->_parent);
2194 set_idom(new_cle_out, new_cle, dd);
2195 set_loop(new_sfpt, outer_loop->_parent);
2196 set_idom(new_sfpt, new_cle_out, dd);
2197 set_loop(new_le, outer_loop->_parent);
2198 set_idom(new_le, new_sfpt, dd);
2199 set_loop(new_tail, outer_loop->_parent);
2200 set_idom(new_tail, new_le, dd);
2201 set_idom(new_cl, new_l, dd);
2202
2203 old_new.map(l->_idx, new_l);
2204 old_new.map(tail->_idx, new_tail);
2205 old_new.map(le->_idx, new_le);
2206 old_new.map(sfpt->_idx, new_sfpt);
2207
2208 new_l->set_req(LoopNode::LoopBackControl, new_tail);
2209 new_l->set_req(0, new_l);
2210 new_tail->set_req(0, new_le);
2211 new_le->set_req(0, new_sfpt);
2212 new_sfpt->set_req(0, new_cle_out);
2213 new_cle_out->set_req(0, new_cle);
2214 new_cl->set_req(LoopNode::EntryControl, new_l);
2215
2216 _igvn.register_new_node_with_optimizer(new_l);
2217 _igvn.register_new_node_with_optimizer(new_tail);
2218 _igvn.register_new_node_with_optimizer(new_le);
2219 } else {
2220 Node *newhead = old_new[loop->_head->_idx];
2221 newhead->as_Loop()->clear_strip_mined();
2222 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl));
2223 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
2224 }
2225 // Look at data node that were assigned a control in the outer
2226 // loop: they are kept in the outer loop by the safepoint so start
2227 // from the safepoint node's inputs.
2228 IdealLoopTree* outer_loop = get_loop(l);
2229 Node_Stack stack(2);
2230 stack.push(sfpt, 1);
2231 uint new_counter = C->unique();
2232 while (stack.size() > 0) {
2233 Node* n = stack.node();
2234 uint i = stack.index();
2235 while (i < n->req() &&
2236 (n->in(i) == NULL ||
2237 !has_ctrl(n->in(i)) ||
2238 get_loop(get_ctrl(n->in(i))) != outer_loop ||
2239 (old_new[n->in(i)->_idx] != NULL && old_new[n->in(i)->_idx]->_idx >= new_counter))) {
2240 i++;
2241 }
2242 if (i < n->req()) {
2243 stack.set_index(i+1);
2244 stack.push(n->in(i), 0);
2245 } else {
2246 assert(old_new[n->_idx] == NULL || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet");
2247 Node* m = n == sfpt ? new_sfpt : n->clone();
2248 if (m != NULL) {
2249 for (uint i = 0; i < n->req(); i++) {
2250 if (m->in(i) != NULL && old_new[m->in(i)->_idx] != NULL) {
2251 m->set_req(i, old_new[m->in(i)->_idx]);
2252 }
2253 }
2254 } else {
2255 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?");
2256 }
2257 if (n != sfpt) {
2258 extra_data_nodes.push(n);
2259 _igvn.register_new_node_with_optimizer(m);
2260 assert(get_ctrl(n) == cle_out, "what other control?");
2261 set_ctrl(m, new_cle_out);
2262 old_new.map(n->_idx, m);
2263 }
2264 stack.pop();
2265 }
2266 }
2267 if (mode == CloneIncludesStripMined) {
2268 _igvn.register_new_node_with_optimizer(new_sfpt);
2269 _igvn.register_new_node_with_optimizer(new_cle_out);
2270 }
2271 // Some other transformation may have pessimistically assign some
2272 // data nodes to the outer loop. Set their control so they are out
2273 // of the outer loop.
2274 ResourceMark rm;
2275 Unique_Node_List wq;
2276 for (uint i = 0; i < extra_data_nodes.size(); i++) {
2277 Node* old = extra_data_nodes.at(i);
2278 clone_outer_loop_helper(old, loop, outer_loop, old_new, wq, this, true);
2279 }
2280
2281 Node* inner_out = sfpt->in(0);
2282 if (inner_out->outcnt() > 1) {
2283 clone_outer_loop_helper(inner_out, loop, outer_loop, old_new, wq, this, true);
2284 }
2285
2286 Node* new_ctrl = cl->outer_loop_exit();
2287 assert(get_loop(new_ctrl) != outer_loop, "must be out of the loop nest");
2288 for (uint i = 0; i < wq.size(); i++) {
2289 Node* n = wq.at(i);
2290 set_ctrl(n, new_ctrl);
2291 if (n->in(0) != NULL) {
2292 _igvn.replace_input_of(n, 0, new_ctrl);
2293 }
2294 clone_outer_loop_helper(n, loop, outer_loop, old_new, wq, this, false);
2295 }
2296 } else {
2297 Node *newhead = old_new[loop->_head->_idx];
2298 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
2299 }
2300 }
2301
2302 //------------------------------clone_loop-------------------------------------
2303 //
2304 // C L O N E A L O O P B O D Y
2305 //
2306 // This is the basic building block of the loop optimizations. It clones an
2307 // entire loop body. It makes an old_new loop body mapping; with this mapping
2308 // you can find the new-loop equivalent to an old-loop node. All new-loop
2309 // nodes are exactly equal to their old-loop counterparts, all edges are the
2310 // same. All exits from the old-loop now have a RegionNode that merges the
2311 // equivalent new-loop path. This is true even for the normal "loop-exit"
2312 // condition. All uses of loop-invariant old-loop values now come from (one
2313 // or more) Phis that merge their new-loop equivalents.
2314 //
2315 // This operation leaves the graph in an illegal state: there are two valid
2316 // control edges coming from the loop pre-header to both loop bodies. I'll
2317 // definitely have to hack the graph after running this transform.
2318 //
2319 // From this building block I will further edit edges to perform loop peeling
2320 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
2321 //
2322 // Parameter side_by_size_idom:
2323 // When side_by_size_idom is NULL, the dominator tree is constructed for
2324 // the clone loop to dominate the original. Used in construction of
2325 // pre-main-post loop sequence.
2326 // When nonnull, the clone and original are side-by-side, both are
2327 // dominated by the side_by_side_idom node. Used in construction of
2328 // unswitched loops.
2329 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
2330 CloneLoopMode mode, Node* side_by_side_idom) {
2331
2332 LoopNode* head = loop->_head->as_Loop();
2333 head->verify_strip_mined(1);
2334
2335 if (C->do_vector_loop() && PrintOpto) {
2336 const char* mname = C->method()->name()->as_quoted_ascii();
2337 if (mname != NULL) {
2338 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname);
2339 }
2340 }
2341
2342 CloneMap& cm = C->clone_map();
2343 Dict* dict = cm.dict();
2344 if (C->do_vector_loop()) {
2345 cm.set_clone_idx(cm.max_gen()+1);
2346 #ifndef PRODUCT
2347 if (PrintOpto) {
2348 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx());
2349 loop->dump_head();
2350 }
2351 #endif
2352 }
2353
2354 // Step 1: Clone the loop body. Make the old->new mapping.
2355 uint i;
2356 for (i = 0; i < loop->_body.size(); i++) {
2357 Node* old = loop->_body.at(i);
2358 Node* nnn = old->clone();
2359 old_new.map(old->_idx, nnn);
2360 if (old->is_reduction()) {
2361 // Reduction flag is not copied by default. Copy it here when cloning the entire loop body.
2362 nnn->add_flag(Node::Flag_is_reduction);
2363 }
2364 if (C->do_vector_loop()) {
2365 cm.verify_insert_and_clone(old, nnn, cm.clone_idx());
2366 }
2367 _igvn.register_new_node_with_optimizer(nnn);
2368 }
2369
2370 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop;
2371
2372 // Step 2: Fix the edges in the new body. If the old input is outside the
2373 // loop use it. If the old input is INside the loop, use the corresponding
2374 // new node instead.
2375 for( i = 0; i < loop->_body.size(); i++ ) {
2376 Node *old = loop->_body.at(i);
2377 Node *nnn = old_new[old->_idx];
2378 // Fix CFG/Loop controlling the new node
2379 if (has_ctrl(old)) {
2380 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
2381 } else {
2382 set_loop(nnn, outer_loop->_parent);
2383 if (old->outcnt() > 0) {
2384 set_idom( nnn, old_new[idom(old)->_idx], dd );
2385 }
2386 }
2387 // Correct edges to the new node
2388 for( uint j = 0; j < nnn->req(); j++ ) {
2389 Node *n = nnn->in(j);
2390 if( n ) {
2391 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n );
2392 if( loop->is_member( old_in_loop ) )
2393 nnn->set_req(j, old_new[n->_idx]);
2394 }
2395 }
2396 _igvn.hash_find_insert(nnn);
2397 }
2398
2399 Node_List extra_data_nodes; // data nodes in the outer strip mined loop
2400 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes);
2401
2402 // Step 3: Now fix control uses. Loop varying control uses have already
2403 // been fixed up (as part of all input edges in Step 2). Loop invariant
2404 // control uses must be either an IfFalse or an IfTrue. Make a merge
2405 // point to merge the old and new IfFalse/IfTrue nodes; make the use
2406 // refer to this.
2407 Node_List worklist;
2408 uint new_counter = C->unique();
2409 for( i = 0; i < loop->_body.size(); i++ ) {
2410 Node* old = loop->_body.at(i);
2411 if( !old->is_CFG() ) continue;
2412
2413 // Copy uses to a worklist, so I can munge the def-use info
2414 // with impunity.
2415 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
2416 worklist.push(old->fast_out(j));
2417
2418 while( worklist.size() ) { // Visit all uses
2419 Node *use = worklist.pop();
2420 if (!has_node(use)) continue; // Ignore dead nodes
2421 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
2422 if( !loop->is_member( use_loop ) && use->is_CFG() ) {
2423 // Both OLD and USE are CFG nodes here.
2424 assert( use->is_Proj(), "" );
2425 Node* nnn = old_new[old->_idx];
2426
2427 Node* newuse = NULL;
2428 if (head->is_strip_mined() && mode != IgnoreStripMined) {
2429 CountedLoopNode* cl = head->as_CountedLoop();
2430 CountedLoopEndNode* cle = cl->loopexit();
2431 Node* cle_out = cle->proj_out_or_null(false);
2432 if (use == cle_out) {
2433 IfNode* le = cl->outer_loop_end();
2434 use = le->proj_out(false);
2435 use_loop = get_loop(use);
2436 if (mode == CloneIncludesStripMined) {
2437 nnn = old_new[le->_idx];
2438 } else {
2439 newuse = old_new[cle_out->_idx];
2440 }
2441 }
2442 }
2443 if (newuse == NULL) {
2444 newuse = use->clone();
2445 }
2446
2447 // Clone the loop exit control projection
2448 if (C->do_vector_loop()) {
2449 cm.verify_insert_and_clone(use, newuse, cm.clone_idx());
2450 }
2451 newuse->set_req(0,nnn);
2452 _igvn.register_new_node_with_optimizer(newuse);
2453 set_loop(newuse, use_loop);
2454 set_idom(newuse, nnn, dom_depth(nnn) + 1 );
2455
2456 // We need a Region to merge the exit from the peeled body and the
2457 // exit from the old loop body.
2458 RegionNode *r = new RegionNode(3);
2459 // Map the old use to the new merge point
2460 old_new.map( use->_idx, r );
2461 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use));
2462 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" );
2463
2464 // The original user of 'use' uses 'r' instead.
2465 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
2466 Node* useuse = use->last_out(l);
2467 _igvn.rehash_node_delayed(useuse);
2468 uint uses_found = 0;
2469 if( useuse->in(0) == use ) {
2470 useuse->set_req(0, r);
2471 uses_found++;
2472 if( useuse->is_CFG() ) {
2473 // This is not a dom_depth > dd_r because when new
2474 // control flow is constructed by a loop opt, a node and
2475 // its dominator can end up at the same dom_depth
2476 assert(dom_depth(useuse) >= dd_r, "");
2477 set_idom(useuse, r, dom_depth(useuse));
2478 }
2479 }
2480 for( uint k = 1; k < useuse->req(); k++ ) {
2481 if( useuse->in(k) == use ) {
2482 useuse->set_req(k, r);
2483 uses_found++;
2484 if (useuse->is_Loop() && k == LoopNode::EntryControl) {
2485 // This is not a dom_depth > dd_r because when new
2486 // control flow is constructed by a loop opt, a node
2487 // and its dominator can end up at the same dom_depth
2488 assert(dom_depth(useuse) >= dd_r , "");
2489 set_idom(useuse, r, dom_depth(useuse));
2490 }
2491 }
2492 }
2493 l -= uses_found; // we deleted 1 or more copies of this edge
2494 }
2495
2496 // Now finish up 'r'
2497 r->set_req( 1, newuse );
2498 r->set_req( 2, use );
2499 _igvn.register_new_node_with_optimizer(r);
2500 set_loop(r, use_loop);
2501 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r);
2502 } // End of if a loop-exit test
2503 }
2504 }
2505
2506 // Step 4: If loop-invariant use is not control, it must be dominated by a
2507 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region
2508 // there if needed. Make a Phi there merging old and new used values.
2509 Node_List *split_if_set = NULL;
2510 Node_List *split_bool_set = NULL;
2511 Node_List *split_cex_set = NULL;
2512 for( i = 0; i < loop->_body.size(); i++ ) {
2513 Node* old = loop->_body.at(i);
2514 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2515 split_bool_set, split_cex_set, worklist, new_counter,
2516 mode);
2517 }
2518
2519 for (i = 0; i < extra_data_nodes.size(); i++) {
2520 Node* old = extra_data_nodes.at(i);
2521 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2522 split_bool_set, split_cex_set, worklist, new_counter,
2523 mode);
2524 }
2525
2526 // Check for IFs that need splitting/cloning. Happens if an IF outside of
2527 // the loop uses a condition set in the loop. The original IF probably
2528 // takes control from one or more OLD Regions (which in turn get from NEW
2529 // Regions). In any case, there will be a set of Phis for each merge point
2530 // from the IF up to where the original BOOL def exists the loop.
2531 if (split_if_set) {
2532 while (split_if_set->size()) {
2533 Node *iff = split_if_set->pop();
2534 if (iff->in(1)->is_Phi()) {
2535 Node *b = clone_iff(iff->in(1)->as_Phi(), loop);
2536 _igvn.replace_input_of(iff, 1, b);
2537 }
2538 }
2539 }
2540 if (split_bool_set) {
2541 while (split_bool_set->size()) {
2542 Node *b = split_bool_set->pop();
2543 Node *phi = b->in(1);
2544 assert(phi->is_Phi(), "");
2545 CmpNode *cmp = clone_bool((PhiNode*)phi, loop);
2546 _igvn.replace_input_of(b, 1, cmp);
2547 }
2548 }
2549 if (split_cex_set) {
2550 while (split_cex_set->size()) {
2551 Node *b = split_cex_set->pop();
2552 assert(b->in(0)->is_Region(), "");
2553 assert(b->in(1)->is_Phi(), "");
2554 assert(b->in(0)->in(0) == b->in(1)->in(0), "");
2555 split_up(b, b->in(0), NULL);
2556 }
2557 }
2558
2559 }
2560
2561
2562 //---------------------- stride_of_possible_iv -------------------------------------
2563 // Looks for an iff/bool/comp with one operand of the compare
2564 // being a cycle involving an add and a phi,
2565 // with an optional truncation (left-shift followed by a right-shift)
2566 // of the add. Returns zero if not an iv.
2567 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
2568 Node* trunc1 = NULL;
2569 Node* trunc2 = NULL;
2570 const TypeInteger* ttype = NULL;
2571 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) {
2572 return 0;
2573 }
2574 BoolNode* bl = iff->in(1)->as_Bool();
2575 Node* cmp = bl->in(1);
2576 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) {
2577 return 0;
2578 }
2579 // Must have an invariant operand
2580 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) {
2581 return 0;
2582 }
2583 Node* add2 = NULL;
2584 Node* cmp1 = cmp->in(1);
2585 if (cmp1->is_Phi()) {
2586 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
2587 Node* phi = cmp1;
2588 for (uint i = 1; i < phi->req(); i++) {
2589 Node* in = phi->in(i);
2590 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in,
2591 &trunc1, &trunc2, &ttype, T_INT);
2592 if (add && add->in(1) == phi) {
2593 add2 = add->in(2);
2594 break;
2595 }
2596 }
2597 } else {
2598 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
2599 Node* addtrunc = cmp1;
2600 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc,
2601 &trunc1, &trunc2, &ttype, T_INT);
2602 if (add && add->in(1)->is_Phi()) {
2603 Node* phi = add->in(1);
2604 for (uint i = 1; i < phi->req(); i++) {
2605 if (phi->in(i) == addtrunc) {
2606 add2 = add->in(2);
2607 break;
2608 }
2609 }
2610 }
2611 }
2612 if (add2 != NULL) {
2613 const TypeInt* add2t = _igvn.type(add2)->is_int();
2614 if (add2t->is_con()) {
2615 return add2t->get_con();
2616 }
2617 }
2618 return 0;
2619 }
2620
2621
2622 //---------------------- stay_in_loop -------------------------------------
2623 // Return the (unique) control output node that's in the loop (if it exists.)
2624 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
2625 Node* unique = NULL;
2626 if (!n) return NULL;
2627 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2628 Node* use = n->fast_out(i);
2629 if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
2630 if (unique != NULL) {
2631 return NULL;
2632 }
2633 unique = use;
2634 }
2635 }
2636 return unique;
2637 }
2638
2639 //------------------------------ register_node -------------------------------------
2640 // Utility to register node "n" with PhaseIdealLoop
2641 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) {
2642 _igvn.register_new_node_with_optimizer(n);
2643 loop->_body.push(n);
2644 if (n->is_CFG()) {
2645 set_loop(n, loop);
2646 set_idom(n, pred, ddepth);
2647 } else {
2648 set_ctrl(n, pred);
2649 }
2650 }
2651
2652 //------------------------------ proj_clone -------------------------------------
2653 // Utility to create an if-projection
2654 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
2655 ProjNode* c = p->clone()->as_Proj();
2656 c->set_req(0, iff);
2657 return c;
2658 }
2659
2660 //------------------------------ short_circuit_if -------------------------------------
2661 // Force the iff control output to be the live_proj
2662 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
2663 guarantee(live_proj != NULL, "null projection");
2664 int proj_con = live_proj->_con;
2665 assert(proj_con == 0 || proj_con == 1, "false or true projection");
2666 Node *con = _igvn.intcon(proj_con);
2667 set_ctrl(con, C->root());
2668 if (iff) {
2669 iff->set_req(1, con);
2670 }
2671 return con;
2672 }
2673
2674 //------------------------------ insert_if_before_proj -------------------------------------
2675 // Insert a new if before an if projection (* - new node)
2676 //
2677 // before
2678 // if(test)
2679 // / \
2680 // v v
2681 // other-proj proj (arg)
2682 //
2683 // after
2684 // if(test)
2685 // / \
2686 // / v
2687 // | * proj-clone
2688 // v |
2689 // other-proj v
2690 // * new_if(relop(cmp[IU](left,right)))
2691 // / \
2692 // v v
2693 // * new-proj proj
2694 // (returned)
2695 //
2696 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) {
2697 IfNode* iff = proj->in(0)->as_If();
2698 IdealLoopTree *loop = get_loop(proj);
2699 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2700 int ddepth = dom_depth(proj);
2701
2702 _igvn.rehash_node_delayed(iff);
2703 _igvn.rehash_node_delayed(proj);
2704
2705 proj->set_req(0, NULL); // temporary disconnect
2706 ProjNode* proj2 = proj_clone(proj, iff);
2707 register_node(proj2, loop, iff, ddepth);
2708
2709 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right);
2710 register_node(cmp, loop, proj2, ddepth);
2711
2712 BoolNode* bol = new BoolNode(cmp, relop);
2713 register_node(bol, loop, proj2, ddepth);
2714
2715 int opcode = iff->Opcode();
2716 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode");
2717 IfNode* new_if = (opcode == Op_If) ? new IfNode(proj2, bol, iff->_prob, iff->_fcnt):
2718 new RangeCheckNode(proj2, bol, iff->_prob, iff->_fcnt);
2719 register_node(new_if, loop, proj2, ddepth);
2720
2721 proj->set_req(0, new_if); // reattach
2722 set_idom(proj, new_if, ddepth);
2723
2724 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
2725 guarantee(new_exit != NULL, "null exit node");
2726 register_node(new_exit, get_loop(other_proj), new_if, ddepth);
2727
2728 return new_exit;
2729 }
2730
2731 //------------------------------ insert_region_before_proj -------------------------------------
2732 // Insert a region before an if projection (* - new node)
2733 //
2734 // before
2735 // if(test)
2736 // / |
2737 // v |
2738 // proj v
2739 // other-proj
2740 //
2741 // after
2742 // if(test)
2743 // / |
2744 // v |
2745 // * proj-clone v
2746 // | other-proj
2747 // v
2748 // * new-region
2749 // |
2750 // v
2751 // * dum_if
2752 // / \
2753 // v \
2754 // * dum-proj v
2755 // proj
2756 //
2757 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
2758 IfNode* iff = proj->in(0)->as_If();
2759 IdealLoopTree *loop = get_loop(proj);
2760 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2761 int ddepth = dom_depth(proj);
2762
2763 _igvn.rehash_node_delayed(iff);
2764 _igvn.rehash_node_delayed(proj);
2765
2766 proj->set_req(0, NULL); // temporary disconnect
2767 ProjNode* proj2 = proj_clone(proj, iff);
2768 register_node(proj2, loop, iff, ddepth);
2769
2770 RegionNode* reg = new RegionNode(2);
2771 reg->set_req(1, proj2);
2772 register_node(reg, loop, iff, ddepth);
2773
2774 IfNode* dum_if = new IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt);
2775 register_node(dum_if, loop, reg, ddepth);
2776
2777 proj->set_req(0, dum_if); // reattach
2778 set_idom(proj, dum_if, ddepth);
2779
2780 ProjNode* dum_proj = proj_clone(other_proj, dum_if);
2781 register_node(dum_proj, loop, dum_if, ddepth);
2782
2783 return reg;
2784 }
2785
2786 //------------------------------ insert_cmpi_loop_exit -------------------------------------
2787 // Clone a signed compare loop exit from an unsigned compare and
2788 // insert it before the unsigned cmp on the stay-in-loop path.
2789 // All new nodes inserted in the dominator tree between the original
2790 // if and it's projections. The original if test is replaced with
2791 // a constant to force the stay-in-loop path.
2792 //
2793 // This is done to make sure that the original if and it's projections
2794 // still dominate the same set of control nodes, that the ctrl() relation
2795 // from data nodes to them is preserved, and that their loop nesting is
2796 // preserved.
2797 //
2798 // before
2799 // if(i <u limit) unsigned compare loop exit
2800 // / |
2801 // v v
2802 // exit-proj stay-in-loop-proj
2803 //
2804 // after
2805 // if(stay-in-loop-const) original if
2806 // / |
2807 // / v
2808 // / if(i < limit) new signed test
2809 // / / |
2810 // / / v
2811 // / / if(i <u limit) new cloned unsigned test
2812 // / / / |
2813 // v v v |
2814 // region |
2815 // | |
2816 // dum-if |
2817 // / | |
2818 // ether | |
2819 // v v
2820 // exit-proj stay-in-loop-proj
2821 //
2822 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) {
2823 const bool Signed = true;
2824 const bool Unsigned = false;
2825
2826 BoolNode* bol = if_cmpu->in(1)->as_Bool();
2827 if (bol->_test._test != BoolTest::lt) return NULL;
2828 CmpNode* cmpu = bol->in(1)->as_Cmp();
2829 if (cmpu->Opcode() != Op_CmpU) return NULL;
2830 int stride = stride_of_possible_iv(if_cmpu);
2831 if (stride == 0) return NULL;
2832
2833 Node* lp_proj = stay_in_loop(if_cmpu, loop);
2834 guarantee(lp_proj != NULL, "null loop node");
2835
2836 ProjNode* lp_continue = lp_proj->as_Proj();
2837 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
2838 if (!lp_exit->is_IfFalse()) {
2839 // The loop exit condition is (i <u limit) ==> (i >= 0 && i < limit).
2840 // We therefore can't add a single exit condition.
2841 return NULL;
2842 }
2843 // The loop exit condition is !(i <u limit) ==> (i < 0 || i >= limit).
2844 // Split out the exit condition (i < 0) for stride < 0 or (i >= limit) for stride > 0.
2845 Node* limit = NULL;
2846 if (stride > 0) {
2847 limit = cmpu->in(2);
2848 } else {
2849 limit = _igvn.makecon(TypeInt::ZERO);
2850 set_ctrl(limit, C->root());
2851 }
2852 // Create a new region on the exit path
2853 RegionNode* reg = insert_region_before_proj(lp_exit);
2854 guarantee(reg != NULL, "null region node");
2855
2856 // Clone the if-cmpu-true-false using a signed compare
2857 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
2858 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue);
2859 reg->add_req(cmpi_exit);
2860
2861 // Clone the if-cmpu-true-false
2862 BoolTest::mask rel_u = bol->_test._test;
2863 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
2864 reg->add_req(cmpu_exit);
2865
2866 // Force original if to stay in loop.
2867 short_circuit_if(if_cmpu, lp_continue);
2868
2869 return cmpi_exit->in(0)->as_If();
2870 }
2871
2872 //------------------------------ remove_cmpi_loop_exit -------------------------------------
2873 // Remove a previously inserted signed compare loop exit.
2874 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
2875 Node* lp_proj = stay_in_loop(if_cmp, loop);
2876 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
2877 stay_in_loop(lp_proj, loop)->is_If() &&
2878 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
2879 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
2880 set_ctrl(con, C->root());
2881 if_cmp->set_req(1, con);
2882 }
2883
2884 //------------------------------ scheduled_nodelist -------------------------------------
2885 // Create a post order schedule of nodes that are in the
2886 // "member" set. The list is returned in "sched".
2887 // The first node in "sched" is the loop head, followed by
2888 // nodes which have no inputs in the "member" set, and then
2889 // followed by the nodes that have an immediate input dependence
2890 // on a node in "sched".
2891 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
2892
2893 assert(member.test(loop->_head->_idx), "loop head must be in member set");
2894 VectorSet visited;
2895 Node_Stack nstack(loop->_body.size());
2896
2897 Node* n = loop->_head; // top of stack is cached in "n"
2898 uint idx = 0;
2899 visited.set(n->_idx);
2900
2901 // Initially push all with no inputs from within member set
2902 for(uint i = 0; i < loop->_body.size(); i++ ) {
2903 Node *elt = loop->_body.at(i);
2904 if (member.test(elt->_idx)) {
2905 bool found = false;
2906 for (uint j = 0; j < elt->req(); j++) {
2907 Node* def = elt->in(j);
2908 if (def && member.test(def->_idx) && def != elt) {
2909 found = true;
2910 break;
2911 }
2912 }
2913 if (!found && elt != loop->_head) {
2914 nstack.push(n, idx);
2915 n = elt;
2916 assert(!visited.test(n->_idx), "not seen yet");
2917 visited.set(n->_idx);
2918 }
2919 }
2920 }
2921
2922 // traverse out's that are in the member set
2923 while (true) {
2924 if (idx < n->outcnt()) {
2925 Node* use = n->raw_out(idx);
2926 idx++;
2927 if (!visited.test_set(use->_idx)) {
2928 if (member.test(use->_idx)) {
2929 nstack.push(n, idx);
2930 n = use;
2931 idx = 0;
2932 }
2933 }
2934 } else {
2935 // All outputs processed
2936 sched.push(n);
2937 if (nstack.is_empty()) break;
2938 n = nstack.node();
2939 idx = nstack.index();
2940 nstack.pop();
2941 }
2942 }
2943 }
2944
2945
2946 //------------------------------ has_use_in_set -------------------------------------
2947 // Has a use in the vector set
2948 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
2949 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2950 Node* use = n->fast_out(j);
2951 if (vset.test(use->_idx)) {
2952 return true;
2953 }
2954 }
2955 return false;
2956 }
2957
2958
2959 //------------------------------ has_use_internal_to_set -------------------------------------
2960 // Has use internal to the vector set (ie. not in a phi at the loop head)
2961 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
2962 Node* head = loop->_head;
2963 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2964 Node* use = n->fast_out(j);
2965 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
2966 return true;
2967 }
2968 }
2969 return false;
2970 }
2971
2972
2973 //------------------------------ clone_for_use_outside_loop -------------------------------------
2974 // clone "n" for uses that are outside of loop
2975 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
2976 int cloned = 0;
2977 assert(worklist.size() == 0, "should be empty");
2978 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2979 Node* use = n->fast_out(j);
2980 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
2981 worklist.push(use);
2982 }
2983 }
2984
2985 if (C->check_node_count(worklist.size() + NodeLimitFudgeFactor,
2986 "Too many clones required in clone_for_use_outside_loop in partial peeling")) {
2987 return -1;
2988 }
2989
2990 while( worklist.size() ) {
2991 Node *use = worklist.pop();
2992 if (!has_node(use) || use->in(0) == C->top()) continue;
2993 uint j;
2994 for (j = 0; j < use->req(); j++) {
2995 if (use->in(j) == n) break;
2996 }
2997 assert(j < use->req(), "must be there");
2998
2999 // clone "n" and insert it between the inputs of "n" and the use outside the loop
3000 Node* n_clone = n->clone();
3001 _igvn.replace_input_of(use, j, n_clone);
3002 cloned++;
3003 Node* use_c;
3004 if (!use->is_Phi()) {
3005 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
3006 } else {
3007 // Use in a phi is considered a use in the associated predecessor block
3008 use_c = use->in(0)->in(j);
3009 }
3010 set_ctrl(n_clone, use_c);
3011 assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
3012 get_loop(use_c)->_body.push(n_clone);
3013 _igvn.register_new_node_with_optimizer(n_clone);
3014 #ifndef PRODUCT
3015 if (TracePartialPeeling) {
3016 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
3017 }
3018 #endif
3019 }
3020 return cloned;
3021 }
3022
3023
3024 //------------------------------ clone_for_special_use_inside_loop -------------------------------------
3025 // clone "n" for special uses that are in the not_peeled region.
3026 // If these def-uses occur in separate blocks, the code generator
3027 // marks the method as not compilable. For example, if a "BoolNode"
3028 // is in a different basic block than the "IfNode" that uses it, then
3029 // the compilation is aborted in the code generator.
3030 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
3031 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
3032 if (n->is_Phi() || n->is_Load()) {
3033 return;
3034 }
3035 assert(worklist.size() == 0, "should be empty");
3036 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3037 Node* use = n->fast_out(j);
3038 if ( not_peel.test(use->_idx) &&
3039 (use->is_If() || use->is_CMove() || use->is_Bool()) &&
3040 use->in(1) == n) {
3041 worklist.push(use);
3042 }
3043 }
3044 if (worklist.size() > 0) {
3045 // clone "n" and insert it between inputs of "n" and the use
3046 Node* n_clone = n->clone();
3047 loop->_body.push(n_clone);
3048 _igvn.register_new_node_with_optimizer(n_clone);
3049 set_ctrl(n_clone, get_ctrl(n));
3050 sink_list.push(n_clone);
3051 not_peel.set(n_clone->_idx);
3052 #ifndef PRODUCT
3053 if (TracePartialPeeling) {
3054 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
3055 }
3056 #endif
3057 while( worklist.size() ) {
3058 Node *use = worklist.pop();
3059 _igvn.rehash_node_delayed(use);
3060 for (uint j = 1; j < use->req(); j++) {
3061 if (use->in(j) == n) {
3062 use->set_req(j, n_clone);
3063 }
3064 }
3065 }
3066 }
3067 }
3068
3069
3070 //------------------------------ insert_phi_for_loop -------------------------------------
3071 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
3072 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
3073 Node *phi = PhiNode::make(lp, back_edge_val);
3074 phi->set_req(LoopNode::EntryControl, lp_entry_val);
3075 // Use existing phi if it already exists
3076 Node *hit = _igvn.hash_find_insert(phi);
3077 if( hit == NULL ) {
3078 _igvn.register_new_node_with_optimizer(phi);
3079 set_ctrl(phi, lp);
3080 } else {
3081 // Remove the new phi from the graph and use the hit
3082 _igvn.remove_dead_node(phi);
3083 phi = hit;
3084 }
3085 _igvn.replace_input_of(use, idx, phi);
3086 }
3087
3088 #ifdef ASSERT
3089 //------------------------------ is_valid_loop_partition -------------------------------------
3090 // Validate the loop partition sets: peel and not_peel
3091 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
3092 VectorSet& not_peel ) {
3093 uint i;
3094 // Check that peel_list entries are in the peel set
3095 for (i = 0; i < peel_list.size(); i++) {
3096 if (!peel.test(peel_list.at(i)->_idx)) {
3097 return false;
3098 }
3099 }
3100 // Check at loop members are in one of peel set or not_peel set
3101 for (i = 0; i < loop->_body.size(); i++ ) {
3102 Node *def = loop->_body.at(i);
3103 uint di = def->_idx;
3104 // Check that peel set elements are in peel_list
3105 if (peel.test(di)) {
3106 if (not_peel.test(di)) {
3107 return false;
3108 }
3109 // Must be in peel_list also
3110 bool found = false;
3111 for (uint j = 0; j < peel_list.size(); j++) {
3112 if (peel_list.at(j)->_idx == di) {
3113 found = true;
3114 break;
3115 }
3116 }
3117 if (!found) {
3118 return false;
3119 }
3120 } else if (not_peel.test(di)) {
3121 if (peel.test(di)) {
3122 return false;
3123 }
3124 } else {
3125 return false;
3126 }
3127 }
3128 return true;
3129 }
3130
3131 //------------------------------ is_valid_clone_loop_exit_use -------------------------------------
3132 // Ensure a use outside of loop is of the right form
3133 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
3134 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
3135 return (use->is_Phi() &&
3136 use_c->is_Region() && use_c->req() == 3 &&
3137 (use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
3138 use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
3139 use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
3140 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
3141 }
3142
3143 //------------------------------ is_valid_clone_loop_form -------------------------------------
3144 // Ensure that all uses outside of loop are of the right form
3145 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
3146 uint orig_exit_idx, uint clone_exit_idx) {
3147 uint len = peel_list.size();
3148 for (uint i = 0; i < len; i++) {
3149 Node *def = peel_list.at(i);
3150
3151 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
3152 Node *use = def->fast_out(j);
3153 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
3154 if (!loop->is_member(get_loop(use_c))) {
3155 // use is not in the loop, check for correct structure
3156 if (use->in(0) == def) {
3157 // Okay
3158 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
3159 return false;
3160 }
3161 }
3162 }
3163 }
3164 return true;
3165 }
3166 #endif
3167
3168 //------------------------------ partial_peel -------------------------------------
3169 // Partially peel (aka loop rotation) the top portion of a loop (called
3170 // the peel section below) by cloning it and placing one copy just before
3171 // the new loop head and the other copy at the bottom of the new loop.
3172 //
3173 // before after where it came from
3174 //
3175 // stmt1 stmt1
3176 // loop: stmt2 clone
3177 // stmt2 if condA goto exitA clone
3178 // if condA goto exitA new_loop: new
3179 // stmt3 stmt3 clone
3180 // if !condB goto loop if condB goto exitB clone
3181 // exitB: stmt2 orig
3182 // stmt4 if !condA goto new_loop orig
3183 // exitA: goto exitA
3184 // exitB:
3185 // stmt4
3186 // exitA:
3187 //
3188 // Step 1: find the cut point: an exit test on probable
3189 // induction variable.
3190 // Step 2: schedule (with cloning) operations in the peel
3191 // section that can be executed after the cut into
3192 // the section that is not peeled. This may need
3193 // to clone operations into exit blocks. For
3194 // instance, a reference to A[i] in the not-peel
3195 // section and a reference to B[i] in an exit block
3196 // may cause a left-shift of i by 2 to be placed
3197 // in the peel block. This step will clone the left
3198 // shift into the exit block and sink the left shift
3199 // from the peel to the not-peel section.
3200 // Step 3: clone the loop, retarget the control, and insert
3201 // phis for values that are live across the new loop
3202 // head. This is very dependent on the graph structure
3203 // from clone_loop. It creates region nodes for
3204 // exit control and associated phi nodes for values
3205 // flow out of the loop through that exit. The region
3206 // node is dominated by the clone's control projection.
3207 // So the clone's peel section is placed before the
3208 // new loop head, and the clone's not-peel section is
3209 // forms the top part of the new loop. The original
3210 // peel section forms the tail of the new loop.
3211 // Step 4: update the dominator tree and recompute the
3212 // dominator depth.
3213 //
3214 // orig
3215 //
3216 // stmt1
3217 // |
3218 // v
3219 // loop predicate
3220 // |
3221 // v
3222 // loop<----+
3223 // | |
3224 // stmt2 |
3225 // | |
3226 // v |
3227 // ifA |
3228 // / | |
3229 // v v |
3230 // false true ^ <-- last_peel
3231 // / | |
3232 // / ===|==cut |
3233 // / stmt3 | <-- first_not_peel
3234 // / | |
3235 // | v |
3236 // v ifB |
3237 // exitA: / \ |
3238 // / \ |
3239 // v v |
3240 // false true |
3241 // / \ |
3242 // / ----+
3243 // |
3244 // v
3245 // exitB:
3246 // stmt4
3247 //
3248 //
3249 // after clone loop
3250 //
3251 // stmt1
3252 // |
3253 // v
3254 // loop predicate
3255 // / \
3256 // clone / \ orig
3257 // / \
3258 // / \
3259 // v v
3260 // +---->loop loop<----+
3261 // | | | |
3262 // | stmt2 stmt2 |
3263 // | | | |
3264 // | v v |
3265 // | ifA ifA |
3266 // | | \ / | |
3267 // | v v v v |
3268 // ^ true false false true ^ <-- last_peel
3269 // | | ^ \ / | |
3270 // | cut==|== \ \ / ===|==cut |
3271 // | stmt3 \ \ / stmt3 | <-- first_not_peel
3272 // | | dom | | | |
3273 // | v \ 1v v2 v |
3274 // | ifB regionA ifB |
3275 // | / \ | / \ |
3276 // | / \ v / \ |
3277 // | v v exitA: v v |
3278 // | true false false true |
3279 // | / ^ \ / \ |
3280 // +---- \ \ / ----+
3281 // dom \ /
3282 // \ 1v v2
3283 // regionB
3284 // |
3285 // v
3286 // exitB:
3287 // stmt4
3288 //
3289 //
3290 // after partial peel
3291 //
3292 // stmt1
3293 // |
3294 // v
3295 // loop predicate
3296 // /
3297 // clone / orig
3298 // / TOP
3299 // / \
3300 // v v
3301 // TOP->loop loop----+
3302 // | | |
3303 // stmt2 stmt2 |
3304 // | | |
3305 // v v |
3306 // ifA ifA |
3307 // | \ / | |
3308 // v v v v |
3309 // true false false true | <-- last_peel
3310 // | ^ \ / +------|---+
3311 // +->newloop \ \ / === ==cut | |
3312 // | stmt3 \ \ / TOP | |
3313 // | | dom | | stmt3 | | <-- first_not_peel
3314 // | v \ 1v v2 v | |
3315 // | ifB regionA ifB ^ v
3316 // | / \ | / \ | |
3317 // | / \ v / \ | |
3318 // | v v exitA: v v | |
3319 // | true false false true | |
3320 // | / ^ \ / \ | |
3321 // | | \ \ / v | |
3322 // | | dom \ / TOP | |
3323 // | | \ 1v v2 | |
3324 // ^ v regionB | |
3325 // | | | | |
3326 // | | v ^ v
3327 // | | exitB: | |
3328 // | | stmt4 | |
3329 // | +------------>-----------------+ |
3330 // | |
3331 // +-----------------<---------------------+
3332 //
3333 //
3334 // final graph
3335 //
3336 // stmt1
3337 // |
3338 // v
3339 // loop predicate
3340 // |
3341 // v
3342 // stmt2 clone
3343 // |
3344 // v
3345 // ........> ifA clone
3346 // : / |
3347 // dom / |
3348 // : v v
3349 // : false true
3350 // : | |
3351 // : | v
3352 // : | newloop<-----+
3353 // : | | |
3354 // : | stmt3 clone |
3355 // : | | |
3356 // : | v |
3357 // : | ifB |
3358 // : | / \ |
3359 // : | v v |
3360 // : | false true |
3361 // : | | | |
3362 // : | v stmt2 |
3363 // : | exitB: | |
3364 // : | stmt4 v |
3365 // : | ifA orig |
3366 // : | / \ |
3367 // : | / \ |
3368 // : | v v |
3369 // : | false true |
3370 // : | / \ |
3371 // : v v -----+
3372 // RegionA
3373 // |
3374 // v
3375 // exitA
3376 //
3377 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
3378
3379 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only");
3380 if (!loop->_head->is_Loop()) {
3381 return false;
3382 }
3383 LoopNode *head = loop->_head->as_Loop();
3384
3385 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
3386 return false;
3387 }
3388
3389 // Check for complex exit control
3390 for (uint ii = 0; ii < loop->_body.size(); ii++) {
3391 Node *n = loop->_body.at(ii);
3392 int opc = n->Opcode();
3393 if (n->is_Call() ||
3394 opc == Op_Catch ||
3395 opc == Op_CatchProj ||
3396 opc == Op_Jump ||
3397 opc == Op_JumpProj) {
3398 #ifndef PRODUCT
3399 if (TracePartialPeeling) {
3400 tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
3401 }
3402 #endif
3403 return false;
3404 }
3405 }
3406
3407 int dd = dom_depth(head);
3408
3409 // Step 1: find cut point
3410
3411 // Walk up dominators to loop head looking for first loop exit
3412 // which is executed on every path thru loop.
3413 IfNode *peel_if = NULL;
3414 IfNode *peel_if_cmpu = NULL;
3415
3416 Node *iff = loop->tail();
3417 while (iff != head) {
3418 if (iff->is_If()) {
3419 Node *ctrl = get_ctrl(iff->in(1));
3420 if (ctrl->is_top()) return false; // Dead test on live IF.
3421 // If loop-varying exit-test, check for induction variable
3422 if (loop->is_member(get_loop(ctrl)) &&
3423 loop->is_loop_exit(iff) &&
3424 is_possible_iv_test(iff)) {
3425 Node* cmp = iff->in(1)->in(1);
3426 if (cmp->Opcode() == Op_CmpI) {
3427 peel_if = iff->as_If();
3428 } else {
3429 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
3430 peel_if_cmpu = iff->as_If();
3431 }
3432 }
3433 }
3434 iff = idom(iff);
3435 }
3436
3437 // Prefer signed compare over unsigned compare.
3438 IfNode* new_peel_if = NULL;
3439 if (peel_if == NULL) {
3440 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) {
3441 return false; // No peel point found
3442 }
3443 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
3444 if (new_peel_if == NULL) {
3445 return false; // No peel point found
3446 }
3447 peel_if = new_peel_if;
3448 }
3449 Node* last_peel = stay_in_loop(peel_if, loop);
3450 Node* first_not_peeled = stay_in_loop(last_peel, loop);
3451 if (first_not_peeled == NULL || first_not_peeled == head) {
3452 return false;
3453 }
3454
3455 #ifndef PRODUCT
3456 if (TraceLoopOpts) {
3457 tty->print("PartialPeel ");
3458 loop->dump_head();
3459 }
3460
3461 if (TracePartialPeeling) {
3462 tty->print_cr("before partial peel one iteration");
3463 Node_List wl;
3464 Node* t = head->in(2);
3465 while (true) {
3466 wl.push(t);
3467 if (t == head) break;
3468 t = idom(t);
3469 }
3470 while (wl.size() > 0) {
3471 Node* tt = wl.pop();
3472 tt->dump();
3473 if (tt == last_peel) tty->print_cr("-- cut --");
3474 }
3475 }
3476 #endif
3477 VectorSet peel;
3478 VectorSet not_peel;
3479 Node_List peel_list;
3480 Node_List worklist;
3481 Node_List sink_list;
3482
3483 uint estimate = loop->est_loop_clone_sz(1);
3484 if (exceeding_node_budget(estimate)) {
3485 return false;
3486 }
3487
3488 // Set of cfg nodes to peel are those that are executable from
3489 // the head through last_peel.
3490 assert(worklist.size() == 0, "should be empty");
3491 worklist.push(head);
3492 peel.set(head->_idx);
3493 while (worklist.size() > 0) {
3494 Node *n = worklist.pop();
3495 if (n != last_peel) {
3496 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3497 Node* use = n->fast_out(j);
3498 if (use->is_CFG() &&
3499 loop->is_member(get_loop(use)) &&
3500 !peel.test_set(use->_idx)) {
3501 worklist.push(use);
3502 }
3503 }
3504 }
3505 }
3506
3507 // Set of non-cfg nodes to peel are those that are control
3508 // dependent on the cfg nodes.
3509 for (uint i = 0; i < loop->_body.size(); i++) {
3510 Node *n = loop->_body.at(i);
3511 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
3512 if (peel.test(n_c->_idx)) {
3513 peel.set(n->_idx);
3514 } else {
3515 not_peel.set(n->_idx);
3516 }
3517 }
3518
3519 // Step 2: move operations from the peeled section down into the
3520 // not-peeled section
3521
3522 // Get a post order schedule of nodes in the peel region
3523 // Result in right-most operand.
3524 scheduled_nodelist(loop, peel, peel_list);
3525
3526 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
3527
3528 // For future check for too many new phis
3529 uint old_phi_cnt = 0;
3530 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
3531 Node* use = head->fast_out(j);
3532 if (use->is_Phi()) old_phi_cnt++;
3533 }
3534
3535 #ifndef PRODUCT
3536 if (TracePartialPeeling) {
3537 tty->print_cr("\npeeled list");
3538 }
3539 #endif
3540
3541 // Evacuate nodes in peel region into the not_peeled region if possible
3542 bool too_many_clones = false;
3543 uint new_phi_cnt = 0;
3544 uint cloned_for_outside_use = 0;
3545 for (uint i = 0; i < peel_list.size();) {
3546 Node* n = peel_list.at(i);
3547 #ifndef PRODUCT
3548 if (TracePartialPeeling) n->dump();
3549 #endif
3550 bool incr = true;
3551 if (!n->is_CFG()) {
3552 if (has_use_in_set(n, not_peel)) {
3553 // If not used internal to the peeled region,
3554 // move "n" from peeled to not_peeled region.
3555 if (!has_use_internal_to_set(n, peel, loop)) {
3556 // if not pinned and not a load (which maybe anti-dependent on a store)
3557 // and not a CMove (Matcher expects only bool->cmove).
3558 if (n->in(0) == NULL && !n->is_Load() && !n->is_CMove()) {
3559 int new_clones = clone_for_use_outside_loop(loop, n, worklist);
3560 if (new_clones == -1) {
3561 too_many_clones = true;
3562 break;
3563 }
3564 cloned_for_outside_use += new_clones;
3565 sink_list.push(n);
3566 peel.remove(n->_idx);
3567 not_peel.set(n->_idx);
3568 peel_list.remove(i);
3569 incr = false;
3570 #ifndef PRODUCT
3571 if (TracePartialPeeling) {
3572 tty->print_cr("sink to not_peeled region: %d newbb: %d",
3573 n->_idx, get_ctrl(n)->_idx);
3574 }
3575 #endif
3576 }
3577 } else {
3578 // Otherwise check for special def-use cases that span
3579 // the peel/not_peel boundary such as bool->if
3580 clone_for_special_use_inside_loop(loop, n, not_peel, sink_list, worklist);
3581 new_phi_cnt++;
3582 }
3583 }
3584 }
3585 if (incr) i++;
3586 }
3587
3588 estimate += cloned_for_outside_use + new_phi_cnt;
3589 bool exceed_node_budget = !may_require_nodes(estimate);
3590 bool exceed_phi_limit = new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta;
3591
3592 if (too_many_clones || exceed_node_budget || exceed_phi_limit) {
3593 #ifndef PRODUCT
3594 if (TracePartialPeeling && exceed_phi_limit) {
3595 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c",
3596 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F');
3597 }
3598 #endif
3599 if (new_peel_if != NULL) {
3600 remove_cmpi_loop_exit(new_peel_if, loop);
3601 }
3602 // Inhibit more partial peeling on this loop
3603 assert(!head->is_partial_peel_loop(), "not partial peeled");
3604 head->mark_partial_peel_failed();
3605 if (cloned_for_outside_use > 0) {
3606 // Terminate this round of loop opts because
3607 // the graph outside this loop was changed.
3608 C->set_major_progress();
3609 return true;
3610 }
3611 return false;
3612 }
3613
3614 // Step 3: clone loop, retarget control, and insert new phis
3615
3616 // Create new loop head for new phis and to hang
3617 // the nodes being moved (sinked) from the peel region.
3618 LoopNode* new_head = new LoopNode(last_peel, last_peel);
3619 new_head->set_unswitch_count(head->unswitch_count()); // Preserve
3620 _igvn.register_new_node_with_optimizer(new_head);
3621 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
3622 _igvn.replace_input_of(first_not_peeled, 0, new_head);
3623 set_loop(new_head, loop);
3624 loop->_body.push(new_head);
3625 not_peel.set(new_head->_idx);
3626 set_idom(new_head, last_peel, dom_depth(first_not_peeled));
3627 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
3628
3629 while (sink_list.size() > 0) {
3630 Node* n = sink_list.pop();
3631 set_ctrl(n, new_head);
3632 }
3633
3634 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
3635
3636 clone_loop(loop, old_new, dd, IgnoreStripMined);
3637
3638 const uint clone_exit_idx = 1;
3639 const uint orig_exit_idx = 2;
3640 assert(is_valid_clone_loop_form(loop, peel_list, orig_exit_idx, clone_exit_idx), "bad clone loop");
3641
3642 Node* head_clone = old_new[head->_idx];
3643 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop();
3644 Node* orig_tail_clone = head_clone->in(2);
3645
3646 // Add phi if "def" node is in peel set and "use" is not
3647
3648 for (uint i = 0; i < peel_list.size(); i++) {
3649 Node *def = peel_list.at(i);
3650 if (!def->is_CFG()) {
3651 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
3652 Node *use = def->fast_out(j);
3653 if (has_node(use) && use->in(0) != C->top() &&
3654 (!peel.test(use->_idx) ||
3655 (use->is_Phi() && use->in(0) == head)) ) {
3656 worklist.push(use);
3657 }
3658 }
3659 while( worklist.size() ) {
3660 Node *use = worklist.pop();
3661 for (uint j = 1; j < use->req(); j++) {
3662 Node* n = use->in(j);
3663 if (n == def) {
3664
3665 // "def" is in peel set, "use" is not in peel set
3666 // or "use" is in the entry boundary (a phi) of the peel set
3667
3668 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
3669
3670 if ( loop->is_member(get_loop( use_c )) ) {
3671 // use is in loop
3672 if (old_new[use->_idx] != NULL) { // null for dead code
3673 Node* use_clone = old_new[use->_idx];
3674 _igvn.replace_input_of(use, j, C->top());
3675 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
3676 }
3677 } else {
3678 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
3679 // use is not in the loop, check if the live range includes the cut
3680 Node* lp_if = use_c->in(orig_exit_idx)->in(0);
3681 if (not_peel.test(lp_if->_idx)) {
3682 assert(j == orig_exit_idx, "use from original loop");
3683 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
3684 }
3685 }
3686 }
3687 }
3688 }
3689 }
3690 }
3691
3692 // Step 3b: retarget control
3693
3694 // Redirect control to the new loop head if a cloned node in
3695 // the not_peeled region has control that points into the peeled region.
3696 // This necessary because the cloned peeled region will be outside
3697 // the loop.
3698 // from to
3699 // cloned-peeled <---+
3700 // new_head_clone: | <--+
3701 // cloned-not_peeled in(0) in(0)
3702 // orig-peeled
3703
3704 for (uint i = 0; i < loop->_body.size(); i++) {
3705 Node *n = loop->_body.at(i);
3706 if (!n->is_CFG() && n->in(0) != NULL &&
3707 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
3708 Node* n_clone = old_new[n->_idx];
3709 _igvn.replace_input_of(n_clone, 0, new_head_clone);
3710 }
3711 }
3712
3713 // Backedge of the surviving new_head (the clone) is original last_peel
3714 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel);
3715
3716 // Cut first node in original not_peel set
3717 _igvn.rehash_node_delayed(new_head); // Multiple edge updates:
3718 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of
3719 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls
3720
3721 // Copy head_clone back-branch info to original head
3722 // and remove original head's loop entry and
3723 // clone head's back-branch
3724 _igvn.rehash_node_delayed(head); // Multiple edge updates
3725 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl));
3726 head->set_req(LoopNode::LoopBackControl, C->top());
3727 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top());
3728
3729 // Similarly modify the phis
3730 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
3731 Node* use = head->fast_out(k);
3732 if (use->is_Phi() && use->outcnt() > 0) {
3733 Node* use_clone = old_new[use->_idx];
3734 _igvn.rehash_node_delayed(use); // Multiple edge updates
3735 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl));
3736 use->set_req(LoopNode::LoopBackControl, C->top());
3737 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top());
3738 }
3739 }
3740
3741 // Step 4: update dominator tree and dominator depth
3742
3743 set_idom(head, orig_tail_clone, dd);
3744 recompute_dom_depth();
3745
3746 // Inhibit more partial peeling on this loop
3747 new_head_clone->set_partial_peel_loop();
3748 C->set_major_progress();
3749 loop->record_for_igvn();
3750
3751 #ifndef PRODUCT
3752 if (TracePartialPeeling) {
3753 tty->print_cr("\nafter partial peel one iteration");
3754 Node_List wl;
3755 Node* t = last_peel;
3756 while (true) {
3757 wl.push(t);
3758 if (t == head_clone) break;
3759 t = idom(t);
3760 }
3761 while (wl.size() > 0) {
3762 Node* tt = wl.pop();
3763 if (tt == head) tty->print_cr("orig head");
3764 else if (tt == new_head_clone) tty->print_cr("new head");
3765 else if (tt == head_clone) tty->print_cr("clone head");
3766 tt->dump();
3767 }
3768 }
3769 #endif
3770 return true;
3771 }
3772
3773 //------------------------------reorg_offsets----------------------------------
3774 // Reorganize offset computations to lower register pressure. Mostly
3775 // prevent loop-fallout uses of the pre-incremented trip counter (which are
3776 // then alive with the post-incremented trip counter forcing an extra
3777 // register move)
3778 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) {
3779 // Perform it only for canonical counted loops.
3780 // Loop's shape could be messed up by iteration_split_impl.
3781 if (!loop->_head->is_CountedLoop())
3782 return;
3783 if (!loop->_head->as_Loop()->is_valid_counted_loop(T_INT))
3784 return;
3785
3786 CountedLoopNode *cl = loop->_head->as_CountedLoop();
3787 CountedLoopEndNode *cle = cl->loopexit();
3788 Node *exit = cle->proj_out(false);
3789 Node *phi = cl->phi();
3790
3791 // Check for the special case when using the pre-incremented trip-counter on
3792 // the fall-out path (forces the pre-incremented and post-incremented trip
3793 // counter to be live at the same time). Fix this by adjusting to use the
3794 // post-increment trip counter.
3795
3796 bool progress = true;
3797 while (progress) {
3798 progress = false;
3799 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
3800 Node* use = phi->fast_out(i); // User of trip-counter
3801 if (!has_ctrl(use)) continue;
3802 Node *u_ctrl = get_ctrl(use);
3803 if (use->is_Phi()) {
3804 u_ctrl = NULL;
3805 for (uint j = 1; j < use->req(); j++)
3806 if (use->in(j) == phi)
3807 u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j));
3808 }
3809 IdealLoopTree *u_loop = get_loop(u_ctrl);
3810 // Look for loop-invariant use
3811 if (u_loop == loop) continue;
3812 if (loop->is_member(u_loop)) continue;
3813 // Check that use is live out the bottom. Assuming the trip-counter
3814 // update is right at the bottom, uses of of the loop middle are ok.
3815 if (dom_lca(exit, u_ctrl) != exit) continue;
3816 // Hit! Refactor use to use the post-incremented tripcounter.
3817 // Compute a post-increment tripcounter.
3818 Node* c = exit;
3819 if (cl->is_strip_mined()) {
3820 IdealLoopTree* outer_loop = get_loop(cl->outer_loop());
3821 if (!outer_loop->is_member(u_loop)) {
3822 c = cl->outer_loop_exit();
3823 }
3824 }
3825 Node *opaq = new Opaque2Node(C, cle->incr());
3826 register_new_node(opaq, c);
3827 Node *neg_stride = _igvn.intcon(-cle->stride_con());
3828 set_ctrl(neg_stride, C->root());
3829 Node *post = new AddINode(opaq, neg_stride);
3830 register_new_node(post, c);
3831 post = new CastIINode(post, phi->bottom_type()); // preserve the iv phi's type
3832 register_new_node(post, c);
3833 _igvn.rehash_node_delayed(use);
3834 for (uint j = 1; j < use->req(); j++) {
3835 if (use->in(j) == phi)
3836 use->set_req(j, post);
3837 }
3838 // Since DU info changed, rerun loop
3839 progress = true;
3840 break;
3841 }
3842 }
3843
3844 }