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