1 /*
2 * Copyright (c) 1999, 2019, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "gc/shared/barrierSet.hpp"
27 #include "gc/shared/c2/barrierSetC2.hpp"
28 #include "memory/allocation.inline.hpp"
29 #include "memory/resourceArea.hpp"
30 #include "opto/addnode.hpp"
31 #include "opto/callnode.hpp"
32 #include "opto/castnode.hpp"
33 #include "opto/connode.hpp"
34 #include "opto/castnode.hpp"
35 #include "opto/divnode.hpp"
36 #include "opto/loopnode.hpp"
37 #include "opto/matcher.hpp"
38 #include "opto/mulnode.hpp"
39 #include "opto/movenode.hpp"
40 #include "opto/opaquenode.hpp"
41 #include "opto/rootnode.hpp"
42 #include "opto/subnode.hpp"
43 #include "opto/valuetypenode.hpp"
44 #include "utilities/macros.hpp"
45 #if INCLUDE_ZGC
46 #include "gc/z/c2/zBarrierSetC2.hpp"
47 #endif
48
49 //=============================================================================
50 //------------------------------split_thru_phi---------------------------------
51 // Split Node 'n' through merge point if there is enough win.
52 Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) {
53 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) {
54 // ConvI2L may have type information on it which is unsafe to push up
55 // so disable this for now
56 return NULL;
57 }
58
59 // Splitting range check CastIIs through a loop induction Phi can
60 // cause new Phis to be created that are left unrelated to the loop
61 // induction Phi and prevent optimizations (vectorization)
62 if (n->Opcode() == Op_CastII && n->as_CastII()->has_range_check() &&
63 region->is_CountedLoop() && n->in(1) == region->as_CountedLoop()->phi()) {
64 return NULL;
65 }
66
67 // Value types should not be split through Phis because they cannot be merged
68 // through Phi nodes but each value input needs to be merged individually.
69 if (n->is_ValueType()) {
70 return NULL;
71 }
72
73 int wins = 0;
74 assert(!n->is_CFG(), "");
75 assert(region->is_Region(), "");
76
77 const Type* type = n->bottom_type();
78 const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr();
79 Node *phi;
80 if (t_oop != NULL && t_oop->is_known_instance_field()) {
81 int iid = t_oop->instance_id();
82 int index = C->get_alias_index(t_oop);
83 int offset = t_oop->offset();
84 phi = new PhiNode(region, type, NULL, iid, index, offset);
85 } else {
86 phi = PhiNode::make_blank(region, n);
87 }
88 uint old_unique = C->unique();
89 for (uint i = 1; i < region->req(); i++) {
90 Node *x;
91 Node* the_clone = NULL;
92 if (region->in(i) == C->top()) {
93 x = C->top(); // Dead path? Use a dead data op
94 } else {
95 x = n->clone(); // Else clone up the data op
96 the_clone = x; // Remember for possible deletion.
97 // Alter data node to use pre-phi inputs
98 if (n->in(0) == region)
99 x->set_req( 0, region->in(i) );
100 for (uint j = 1; j < n->req(); j++) {
101 Node *in = n->in(j);
102 if (in->is_Phi() && in->in(0) == region)
103 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone
104 }
105 }
106 // Check for a 'win' on some paths
107 const Type *t = x->Value(&_igvn);
108
109 bool singleton = t->singleton();
110
111 // A TOP singleton indicates that there are no possible values incoming
112 // along a particular edge. In most cases, this is OK, and the Phi will
113 // be eliminated later in an Ideal call. However, we can't allow this to
114 // happen if the singleton occurs on loop entry, as the elimination of
115 // the PhiNode may cause the resulting node to migrate back to a previous
116 // loop iteration.
117 if (singleton && t == Type::TOP) {
118 // Is_Loop() == false does not confirm the absence of a loop (e.g., an
119 // irreducible loop may not be indicated by an affirmative is_Loop());
120 // therefore, the only top we can split thru a phi is on a backedge of
121 // a loop.
122 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
123 }
124
125 if (singleton) {
126 wins++;
127 x = ((PhaseGVN&)_igvn).makecon(t);
128 } else {
129 // We now call Identity to try to simplify the cloned node.
130 // Note that some Identity methods call phase->type(this).
131 // Make sure that the type array is big enough for
132 // our new node, even though we may throw the node away.
133 // (Note: This tweaking with igvn only works because x is a new node.)
134 _igvn.set_type(x, t);
135 // If x is a TypeNode, capture any more-precise type permanently into Node
136 // otherwise it will be not updated during igvn->transform since
137 // igvn->type(x) is set to x->Value() already.
138 x->raise_bottom_type(t);
139 Node *y = _igvn.apply_identity(x);
140 if (y != x) {
141 wins++;
142 x = y;
143 } else {
144 y = _igvn.hash_find(x);
145 if (y) {
146 wins++;
147 x = y;
148 } else {
149 // Else x is a new node we are keeping
150 // We do not need register_new_node_with_optimizer
151 // because set_type has already been called.
152 _igvn._worklist.push(x);
153 }
154 }
155 }
156 if (x != the_clone && the_clone != NULL)
157 _igvn.remove_dead_node(the_clone);
158 phi->set_req( i, x );
159 }
160 // Too few wins?
161 if (wins <= policy) {
162 _igvn.remove_dead_node(phi);
163 return NULL;
164 }
165
166 // Record Phi
167 register_new_node( phi, region );
168
169 for (uint i2 = 1; i2 < phi->req(); i2++) {
170 Node *x = phi->in(i2);
171 // If we commoned up the cloned 'x' with another existing Node,
172 // the existing Node picks up a new use. We need to make the
173 // existing Node occur higher up so it dominates its uses.
174 Node *old_ctrl;
175 IdealLoopTree *old_loop;
176
177 if (x->is_Con()) {
178 // Constant's control is always root.
179 set_ctrl(x, C->root());
180 continue;
181 }
182 // The occasional new node
183 if (x->_idx >= old_unique) { // Found a new, unplaced node?
184 old_ctrl = NULL;
185 old_loop = NULL; // Not in any prior loop
186 } else {
187 old_ctrl = get_ctrl(x);
188 old_loop = get_loop(old_ctrl); // Get prior loop
189 }
190 // New late point must dominate new use
191 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2));
192 if (new_ctrl == old_ctrl) // Nothing is changed
193 continue;
194
195 IdealLoopTree *new_loop = get_loop(new_ctrl);
196
197 // Don't move x into a loop if its uses are
198 // outside of loop. Otherwise x will be cloned
199 // for each use outside of this loop.
200 IdealLoopTree *use_loop = get_loop(region);
201 if (!new_loop->is_member(use_loop) &&
202 (old_loop == NULL || !new_loop->is_member(old_loop))) {
203 // Take early control, later control will be recalculated
204 // during next iteration of loop optimizations.
205 new_ctrl = get_early_ctrl(x);
206 new_loop = get_loop(new_ctrl);
207 }
208 // Set new location
209 set_ctrl(x, new_ctrl);
210 // If changing loop bodies, see if we need to collect into new body
211 if (old_loop != new_loop) {
212 if (old_loop && !old_loop->_child)
213 old_loop->_body.yank(x);
214 if (!new_loop->_child)
215 new_loop->_body.push(x); // Collect body info
216 }
217 }
218
219 return phi;
220 }
221
222 //------------------------------dominated_by------------------------------------
223 // Replace the dominated test with an obvious true or false. Place it on the
224 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the
225 // live path up to the dominating control.
226 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff, bool flip, bool exclude_loop_predicate ) {
227 if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); }
228
229 // prevdom is the dominating projection of the dominating test.
230 assert( iff->is_If(), "" );
231 assert(iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd || iff->Opcode() == Op_RangeCheck, "Check this code when new subtype is added");
232 int pop = prevdom->Opcode();
233 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" );
234 if (flip) {
235 if (pop == Op_IfTrue)
236 pop = Op_IfFalse;
237 else
238 pop = Op_IfTrue;
239 }
240 // 'con' is set to true or false to kill the dominated test.
241 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO);
242 set_ctrl(con, C->root()); // Constant gets a new use
243 // Hack the dominated test
244 _igvn.replace_input_of(iff, 1, con);
245
246 // If I dont have a reachable TRUE and FALSE path following the IfNode then
247 // I can assume this path reaches an infinite loop. In this case it's not
248 // important to optimize the data Nodes - either the whole compilation will
249 // be tossed or this path (and all data Nodes) will go dead.
250 if (iff->outcnt() != 2) return;
251
252 // Make control-dependent data Nodes on the live path (path that will remain
253 // once the dominated IF is removed) become control-dependent on the
254 // dominating projection.
255 Node* dp = iff->as_If()->proj_out_or_null(pop == Op_IfTrue);
256
257 // Loop predicates may have depending checks which should not
258 // be skipped. For example, range check predicate has two checks
259 // for lower and upper bounds.
260 if (dp == NULL)
261 return;
262
263 ProjNode* dp_proj = dp->as_Proj();
264 ProjNode* unc_proj = iff->as_If()->proj_out(1 - dp_proj->_con)->as_Proj();
265 if (exclude_loop_predicate &&
266 (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) != NULL ||
267 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_profile_predicate) != NULL ||
268 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check) != NULL)) {
269 // If this is a range check (IfNode::is_range_check), do not
270 // reorder because Compile::allow_range_check_smearing might have
271 // changed the check.
272 return; // Let IGVN transformation change control dependence.
273 }
274
275 IdealLoopTree *old_loop = get_loop(dp);
276
277 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
278 Node* cd = dp->fast_out(i); // Control-dependent node
279 if (cd->depends_only_on_test()) {
280 assert(cd->in(0) == dp, "");
281 _igvn.replace_input_of(cd, 0, prevdom);
282 set_early_ctrl(cd);
283 IdealLoopTree *new_loop = get_loop(get_ctrl(cd));
284 if (old_loop != new_loop) {
285 if (!old_loop->_child) old_loop->_body.yank(cd);
286 if (!new_loop->_child) new_loop->_body.push(cd);
287 }
288 --i;
289 --imax;
290 }
291 }
292 }
293
294 //------------------------------has_local_phi_input----------------------------
295 // Return TRUE if 'n' has Phi inputs from its local block and no other
296 // block-local inputs (all non-local-phi inputs come from earlier blocks)
297 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) {
298 Node *n_ctrl = get_ctrl(n);
299 // See if some inputs come from a Phi in this block, or from before
300 // this block.
301 uint i;
302 for( i = 1; i < n->req(); i++ ) {
303 Node *phi = n->in(i);
304 if( phi->is_Phi() && phi->in(0) == n_ctrl )
305 break;
306 }
307 if( i >= n->req() )
308 return NULL; // No Phi inputs; nowhere to clone thru
309
310 // Check for inputs created between 'n' and the Phi input. These
311 // must split as well; they have already been given the chance
312 // (courtesy of a post-order visit) and since they did not we must
313 // recover the 'cost' of splitting them by being very profitable
314 // when splitting 'n'. Since this is unlikely we simply give up.
315 for( i = 1; i < n->req(); i++ ) {
316 Node *m = n->in(i);
317 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) {
318 // We allow the special case of AddP's with no local inputs.
319 // This allows us to split-up address expressions.
320 if (m->is_AddP() &&
321 get_ctrl(m->in(2)) != n_ctrl &&
322 get_ctrl(m->in(3)) != n_ctrl) {
323 // Move the AddP up to dominating point
324 Node* c = find_non_split_ctrl(idom(n_ctrl));
325 if (c->is_OuterStripMinedLoop()) {
326 c->as_Loop()->verify_strip_mined(1);
327 c = c->in(LoopNode::EntryControl);
328 }
329 set_ctrl_and_loop(m, c);
330 continue;
331 }
332 return NULL;
333 }
334 assert(m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control");
335 }
336
337 return n_ctrl;
338 }
339
340 //------------------------------remix_address_expressions----------------------
341 // Rework addressing expressions to get the most loop-invariant stuff
342 // moved out. We'd like to do all associative operators, but it's especially
343 // important (common) to do address expressions.
344 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) {
345 if (!has_ctrl(n)) return NULL;
346 Node *n_ctrl = get_ctrl(n);
347 IdealLoopTree *n_loop = get_loop(n_ctrl);
348
349 // See if 'n' mixes loop-varying and loop-invariant inputs and
350 // itself is loop-varying.
351
352 // Only interested in binary ops (and AddP)
353 if( n->req() < 3 || n->req() > 4 ) return NULL;
354
355 Node *n1_ctrl = get_ctrl(n->in( 1));
356 Node *n2_ctrl = get_ctrl(n->in( 2));
357 Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3));
358 IdealLoopTree *n1_loop = get_loop( n1_ctrl );
359 IdealLoopTree *n2_loop = get_loop( n2_ctrl );
360 IdealLoopTree *n3_loop = get_loop( n3_ctrl );
361
362 // Does one of my inputs spin in a tighter loop than self?
363 if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) ||
364 (n_loop->is_member( n2_loop ) && n_loop != n2_loop) ||
365 (n_loop->is_member( n3_loop ) && n_loop != n3_loop) )
366 return NULL; // Leave well enough alone
367
368 // Is at least one of my inputs loop-invariant?
369 if( n1_loop == n_loop &&
370 n2_loop == n_loop &&
371 n3_loop == n_loop )
372 return NULL; // No loop-invariant inputs
373
374
375 int n_op = n->Opcode();
376
377 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2).
378 if( n_op == Op_LShiftI ) {
379 // Scale is loop invariant
380 Node *scale = n->in(2);
381 Node *scale_ctrl = get_ctrl(scale);
382 IdealLoopTree *scale_loop = get_loop(scale_ctrl );
383 if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) )
384 return NULL;
385 const TypeInt *scale_t = scale->bottom_type()->isa_int();
386 if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 )
387 return NULL; // Dont bother with byte/short masking
388 // Add must vary with loop (else shift would be loop-invariant)
389 Node *add = n->in(1);
390 Node *add_ctrl = get_ctrl(add);
391 IdealLoopTree *add_loop = get_loop(add_ctrl);
392 //assert( n_loop == add_loop, "" );
393 if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops
394
395 // Convert I-V into I+ (0-V); same for V-I
396 if( add->Opcode() == Op_SubI &&
397 _igvn.type( add->in(1) ) != TypeInt::ZERO ) {
398 Node *zero = _igvn.intcon(0);
399 set_ctrl(zero, C->root());
400 Node *neg = new SubINode( _igvn.intcon(0), add->in(2) );
401 register_new_node( neg, get_ctrl(add->in(2) ) );
402 add = new AddINode( add->in(1), neg );
403 register_new_node( add, add_ctrl );
404 }
405 if( add->Opcode() != Op_AddI ) return NULL;
406 // See if one add input is loop invariant
407 Node *add_var = add->in(1);
408 Node *add_var_ctrl = get_ctrl(add_var);
409 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
410 Node *add_invar = add->in(2);
411 Node *add_invar_ctrl = get_ctrl(add_invar);
412 IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl );
413 if( add_var_loop == n_loop ) {
414 } else if( add_invar_loop == n_loop ) {
415 // Swap to find the invariant part
416 add_invar = add_var;
417 add_invar_ctrl = add_var_ctrl;
418 add_invar_loop = add_var_loop;
419 add_var = add->in(2);
420 Node *add_var_ctrl = get_ctrl(add_var);
421 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
422 } else // Else neither input is loop invariant
423 return NULL;
424 if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) )
425 return NULL; // No invariant part of the add?
426
427 // Yes! Reshape address expression!
428 Node *inv_scale = new LShiftINode( add_invar, scale );
429 Node *inv_scale_ctrl =
430 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ?
431 add_invar_ctrl : scale_ctrl;
432 register_new_node( inv_scale, inv_scale_ctrl );
433 Node *var_scale = new LShiftINode( add_var, scale );
434 register_new_node( var_scale, n_ctrl );
435 Node *var_add = new AddINode( var_scale, inv_scale );
436 register_new_node( var_add, n_ctrl );
437 _igvn.replace_node( n, var_add );
438 return var_add;
439 }
440
441 // Replace (I+V) with (V+I)
442 if( n_op == Op_AddI ||
443 n_op == Op_AddL ||
444 n_op == Op_AddF ||
445 n_op == Op_AddD ||
446 n_op == Op_MulI ||
447 n_op == Op_MulL ||
448 n_op == Op_MulF ||
449 n_op == Op_MulD ) {
450 if( n2_loop == n_loop ) {
451 assert( n1_loop != n_loop, "" );
452 n->swap_edges(1, 2);
453 }
454 }
455
456 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V),
457 // but not if I2 is a constant.
458 if( n_op == Op_AddP ) {
459 if( n2_loop == n_loop && n3_loop != n_loop ) {
460 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) {
461 Node *n22_ctrl = get_ctrl(n->in(2)->in(2));
462 Node *n23_ctrl = get_ctrl(n->in(2)->in(3));
463 IdealLoopTree *n22loop = get_loop( n22_ctrl );
464 IdealLoopTree *n23_loop = get_loop( n23_ctrl );
465 if( n22loop != n_loop && n22loop->is_member(n_loop) &&
466 n23_loop == n_loop ) {
467 Node *add1 = new AddPNode( n->in(1), n->in(2)->in(2), n->in(3) );
468 // Stuff new AddP in the loop preheader
469 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) );
470 Node *add2 = new AddPNode( n->in(1), add1, n->in(2)->in(3) );
471 register_new_node( add2, n_ctrl );
472 _igvn.replace_node( n, add2 );
473 return add2;
474 }
475 }
476 }
477
478 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V)
479 if (n2_loop != n_loop && n3_loop == n_loop) {
480 if (n->in(3)->Opcode() == Op_AddX) {
481 Node *V = n->in(3)->in(1);
482 Node *I = n->in(3)->in(2);
483 if (is_member(n_loop,get_ctrl(V))) {
484 } else {
485 Node *tmp = V; V = I; I = tmp;
486 }
487 if (!is_member(n_loop,get_ctrl(I))) {
488 Node *add1 = new AddPNode(n->in(1), n->in(2), I);
489 // Stuff new AddP in the loop preheader
490 register_new_node(add1, n_loop->_head->in(LoopNode::EntryControl));
491 Node *add2 = new AddPNode(n->in(1), add1, V);
492 register_new_node(add2, n_ctrl);
493 _igvn.replace_node(n, add2);
494 return add2;
495 }
496 }
497 }
498 }
499
500 return NULL;
501 }
502
503 // Optimize ((in1[2*i] * in2[2*i]) + (in1[2*i+1] * in2[2*i+1]))
504 Node *PhaseIdealLoop::convert_add_to_muladd(Node* n) {
505 assert(n->Opcode() == Op_AddI, "sanity");
506 Node * nn = NULL;
507 Node * in1 = n->in(1);
508 Node * in2 = n->in(2);
509 if (in1->Opcode() == Op_MulI && in2->Opcode() == Op_MulI) {
510 IdealLoopTree* loop_n = get_loop(get_ctrl(n));
511 if (loop_n->_head->as_Loop()->is_valid_counted_loop() &&
512 Matcher::match_rule_supported(Op_MulAddS2I) &&
513 Matcher::match_rule_supported(Op_MulAddVS2VI)) {
514 Node* mul_in1 = in1->in(1);
515 Node* mul_in2 = in1->in(2);
516 Node* mul_in3 = in2->in(1);
517 Node* mul_in4 = in2->in(2);
518 if (mul_in1->Opcode() == Op_LoadS &&
519 mul_in2->Opcode() == Op_LoadS &&
520 mul_in3->Opcode() == Op_LoadS &&
521 mul_in4->Opcode() == Op_LoadS) {
522 IdealLoopTree* loop1 = get_loop(get_ctrl(mul_in1));
523 IdealLoopTree* loop2 = get_loop(get_ctrl(mul_in2));
524 IdealLoopTree* loop3 = get_loop(get_ctrl(mul_in3));
525 IdealLoopTree* loop4 = get_loop(get_ctrl(mul_in4));
526 IdealLoopTree* loop5 = get_loop(get_ctrl(in1));
527 IdealLoopTree* loop6 = get_loop(get_ctrl(in2));
528 // All nodes should be in the same counted loop.
529 if (loop_n == loop1 && loop_n == loop2 && loop_n == loop3 &&
530 loop_n == loop4 && loop_n == loop5 && loop_n == loop6) {
531 Node* adr1 = mul_in1->in(MemNode::Address);
532 Node* adr2 = mul_in2->in(MemNode::Address);
533 Node* adr3 = mul_in3->in(MemNode::Address);
534 Node* adr4 = mul_in4->in(MemNode::Address);
535 if (adr1->is_AddP() && adr2->is_AddP() && adr3->is_AddP() && adr4->is_AddP()) {
536 if ((adr1->in(AddPNode::Base) == adr3->in(AddPNode::Base)) &&
537 (adr2->in(AddPNode::Base) == adr4->in(AddPNode::Base))) {
538 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in3, mul_in4);
539 register_new_node(nn, get_ctrl(n));
540 _igvn.replace_node(n, nn);
541 return nn;
542 } else if ((adr1->in(AddPNode::Base) == adr4->in(AddPNode::Base)) &&
543 (adr2->in(AddPNode::Base) == adr3->in(AddPNode::Base))) {
544 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in4, mul_in3);
545 register_new_node(nn, get_ctrl(n));
546 _igvn.replace_node(n, nn);
547 return nn;
548 }
549 }
550 }
551 }
552 }
553 }
554 return nn;
555 }
556
557 //------------------------------conditional_move-------------------------------
558 // Attempt to replace a Phi with a conditional move. We have some pretty
559 // strict profitability requirements. All Phis at the merge point must
560 // be converted, so we can remove the control flow. We need to limit the
561 // number of c-moves to a small handful. All code that was in the side-arms
562 // of the CFG diamond is now speculatively executed. This code has to be
563 // "cheap enough". We are pretty much limited to CFG diamonds that merge
564 // 1 or 2 items with a total of 1 or 2 ops executed speculatively.
565 Node *PhaseIdealLoop::conditional_move( Node *region ) {
566
567 assert(region->is_Region(), "sanity check");
568 if (region->req() != 3) return NULL;
569
570 // Check for CFG diamond
571 Node *lp = region->in(1);
572 Node *rp = region->in(2);
573 if (!lp || !rp) return NULL;
574 Node *lp_c = lp->in(0);
575 if (lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If()) return NULL;
576 IfNode *iff = lp_c->as_If();
577
578 // Check for ops pinned in an arm of the diamond.
579 // Can't remove the control flow in this case
580 if (lp->outcnt() > 1) return NULL;
581 if (rp->outcnt() > 1) return NULL;
582
583 IdealLoopTree* r_loop = get_loop(region);
584 assert(r_loop == get_loop(iff), "sanity");
585 // Always convert to CMOVE if all results are used only outside this loop.
586 bool used_inside_loop = (r_loop == _ltree_root);
587
588 // Check profitability
589 int cost = 0;
590 int phis = 0;
591 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
592 Node *out = region->fast_out(i);
593 if (!out->is_Phi()) continue; // Ignore other control edges, etc
594 phis++;
595 PhiNode* phi = out->as_Phi();
596 BasicType bt = phi->type()->basic_type();
597 switch (bt) {
598 case T_DOUBLE:
599 case T_FLOAT:
600 if (C->use_cmove()) {
601 continue; //TODO: maybe we want to add some cost
602 }
603 cost += Matcher::float_cmove_cost(); // Could be very expensive
604 break;
605 case T_LONG: {
606 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's
607 }
608 case T_INT: // These all CMOV fine
609 case T_ADDRESS: { // (RawPtr)
610 cost++;
611 break;
612 }
613 case T_NARROWOOP: // Fall through
614 case T_OBJECT: { // Base oops are OK, but not derived oops
615 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr();
616 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a
617 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus
618 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we
619 // have a Phi for the base here that we convert to a CMOVE all is well
620 // and good. But if the base is dead, we'll not make a CMOVE. Later
621 // the allocator will have to produce a base by creating a CMOVE of the
622 // relevant bases. This puts the allocator in the business of
623 // manufacturing expensive instructions, generally a bad plan.
624 // Just Say No to Conditionally-Moved Derived Pointers.
625 if (tp && tp->offset() != 0)
626 return NULL;
627 cost++;
628 break;
629 }
630 default:
631 return NULL; // In particular, can't do memory or I/O
632 }
633 // Add in cost any speculative ops
634 for (uint j = 1; j < region->req(); j++) {
635 Node *proj = region->in(j);
636 Node *inp = phi->in(j);
637 if (get_ctrl(inp) == proj) { // Found local op
638 cost++;
639 // Check for a chain of dependent ops; these will all become
640 // speculative in a CMOV.
641 for (uint k = 1; k < inp->req(); k++)
642 if (get_ctrl(inp->in(k)) == proj)
643 cost += ConditionalMoveLimit; // Too much speculative goo
644 }
645 }
646 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode.
647 // This will likely Split-If, a higher-payoff operation.
648 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) {
649 Node* use = phi->fast_out(k);
650 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr())
651 cost += ConditionalMoveLimit;
652 // Is there a use inside the loop?
653 // Note: check only basic types since CMoveP is pinned.
654 if (!used_inside_loop && is_java_primitive(bt)) {
655 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use);
656 if (r_loop == u_loop || r_loop->is_member(u_loop)) {
657 used_inside_loop = true;
658 }
659 }
660 }
661 }//for
662 Node* bol = iff->in(1);
663 assert(bol->Opcode() == Op_Bool, "");
664 int cmp_op = bol->in(1)->Opcode();
665 // It is expensive to generate flags from a float compare.
666 // Avoid duplicated float compare.
667 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL;
668
669 float infrequent_prob = PROB_UNLIKELY_MAG(3);
670 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop.
671 if (used_inside_loop) {
672 if (cost >= ConditionalMoveLimit) return NULL; // Too much goo
673
674 // BlockLayoutByFrequency optimization moves infrequent branch
675 // from hot path. No point in CMOV'ing in such case (110 is used
676 // instead of 100 to take into account not exactness of float value).
677 if (BlockLayoutByFrequency) {
678 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f);
679 }
680 }
681 // Check for highly predictable branch. No point in CMOV'ing if
682 // we are going to predict accurately all the time.
683 if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) {
684 //keep going
685 } else if (iff->_prob < infrequent_prob ||
686 iff->_prob > (1.0f - infrequent_prob))
687 return NULL;
688
689 // --------------
690 // Now replace all Phis with CMOV's
691 Node *cmov_ctrl = iff->in(0);
692 uint flip = (lp->Opcode() == Op_IfTrue);
693 Node_List wq;
694 while (1) {
695 PhiNode* phi = NULL;
696 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
697 Node *out = region->fast_out(i);
698 if (out->is_Phi()) {
699 phi = out->as_Phi();
700 break;
701 }
702 }
703 if (phi == NULL) break;
704 if (PrintOpto && VerifyLoopOptimizations) { tty->print_cr("CMOV"); }
705 // Move speculative ops
706 wq.push(phi);
707 while (wq.size() > 0) {
708 Node *n = wq.pop();
709 for (uint j = 1; j < n->req(); j++) {
710 Node* m = n->in(j);
711 if (m != NULL && !is_dominator(get_ctrl(m), cmov_ctrl)) {
712 #ifndef PRODUCT
713 if (PrintOpto && VerifyLoopOptimizations) {
714 tty->print(" speculate: ");
715 m->dump();
716 }
717 #endif
718 set_ctrl(m, cmov_ctrl);
719 wq.push(m);
720 }
721 }
722 }
723 Node *cmov = CMoveNode::make(cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi));
724 register_new_node( cmov, cmov_ctrl );
725 _igvn.replace_node( phi, cmov );
726 #ifndef PRODUCT
727 if (TraceLoopOpts) {
728 tty->print("CMOV ");
729 r_loop->dump_head();
730 if (Verbose) {
731 bol->in(1)->dump(1);
732 cmov->dump(1);
733 }
734 }
735 if (VerifyLoopOptimizations) verify();
736 #endif
737 }
738
739 // The useless CFG diamond will fold up later; see the optimization in
740 // RegionNode::Ideal.
741 _igvn._worklist.push(region);
742
743 return iff->in(1);
744 }
745
746 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) {
747 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
748 Node* u = m->fast_out(i);
749 if (u->is_CFG()) {
750 if (u->Opcode() == Op_NeverBranch) {
751 u = ((NeverBranchNode*)u)->proj_out(0);
752 enqueue_cfg_uses(u, wq);
753 } else {
754 wq.push(u);
755 }
756 }
757 }
758 }
759
760 // Try moving a store out of a loop, right before the loop
761 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) {
762 // Store has to be first in the loop body
763 IdealLoopTree *n_loop = get_loop(n_ctrl);
764 if (n->is_Store() && n_loop != _ltree_root &&
765 n_loop->is_loop() && n_loop->_head->is_Loop() &&
766 n->in(0) != NULL) {
767 Node* address = n->in(MemNode::Address);
768 Node* value = n->in(MemNode::ValueIn);
769 Node* mem = n->in(MemNode::Memory);
770 IdealLoopTree* address_loop = get_loop(get_ctrl(address));
771 IdealLoopTree* value_loop = get_loop(get_ctrl(value));
772
773 // - address and value must be loop invariant
774 // - memory must be a memory Phi for the loop
775 // - Store must be the only store on this memory slice in the
776 // loop: if there's another store following this one then value
777 // written at iteration i by the second store could be overwritten
778 // at iteration i+n by the first store: it's not safe to move the
779 // first store out of the loop
780 // - nothing must observe the memory Phi: it guarantees no read
781 // before the store, we are also guaranteed the store post
782 // dominates the loop head (ignoring a possible early
783 // exit). Otherwise there would be extra Phi involved between the
784 // loop's Phi and the store.
785 // - there must be no early exit from the loop before the Store
786 // (such an exit most of the time would be an extra use of the
787 // memory Phi but sometimes is a bottom memory Phi that takes the
788 // store as input).
789
790 if (!n_loop->is_member(address_loop) &&
791 !n_loop->is_member(value_loop) &&
792 mem->is_Phi() && mem->in(0) == n_loop->_head &&
793 mem->outcnt() == 1 &&
794 mem->in(LoopNode::LoopBackControl) == n) {
795
796 assert(n_loop->_tail != NULL, "need a tail");
797 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop");
798
799 // Verify that there's no early exit of the loop before the store.
800 bool ctrl_ok = false;
801 {
802 // Follow control from loop head until n, we exit the loop or
803 // we reach the tail
804 ResourceMark rm;
805 Unique_Node_List wq;
806 wq.push(n_loop->_head);
807
808 for (uint next = 0; next < wq.size(); ++next) {
809 Node *m = wq.at(next);
810 if (m == n->in(0)) {
811 ctrl_ok = true;
812 continue;
813 }
814 assert(!has_ctrl(m), "should be CFG");
815 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) {
816 ctrl_ok = false;
817 break;
818 }
819 enqueue_cfg_uses(m, wq);
820 if (wq.size() > 10) {
821 ctrl_ok = false;
822 break;
823 }
824 }
825 }
826 if (ctrl_ok) {
827 // move the Store
828 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem);
829 _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl));
830 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl));
831 // Disconnect the phi now. An empty phi can confuse other
832 // optimizations in this pass of loop opts.
833 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl));
834 n_loop->_body.yank(mem);
835
836 set_ctrl_and_loop(n, n->in(0));
837
838 return n;
839 }
840 }
841 }
842 return NULL;
843 }
844
845 // Try moving a store out of a loop, right after the loop
846 void PhaseIdealLoop::try_move_store_after_loop(Node* n) {
847 if (n->is_Store() && n->in(0) != NULL) {
848 Node *n_ctrl = get_ctrl(n);
849 IdealLoopTree *n_loop = get_loop(n_ctrl);
850 // Store must be in a loop
851 if (n_loop != _ltree_root && !n_loop->_irreducible) {
852 Node* address = n->in(MemNode::Address);
853 Node* value = n->in(MemNode::ValueIn);
854 IdealLoopTree* address_loop = get_loop(get_ctrl(address));
855 // address must be loop invariant
856 if (!n_loop->is_member(address_loop)) {
857 // Store must be last on this memory slice in the loop and
858 // nothing in the loop must observe it
859 Node* phi = NULL;
860 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
861 Node* u = n->fast_out(i);
862 if (has_ctrl(u)) { // control use?
863 IdealLoopTree *u_loop = get_loop(get_ctrl(u));
864 if (!n_loop->is_member(u_loop)) {
865 continue;
866 }
867 if (u->is_Phi() && u->in(0) == n_loop->_head) {
868 assert(_igvn.type(u) == Type::MEMORY, "bad phi");
869 // multiple phis on the same slice are possible
870 if (phi != NULL) {
871 return;
872 }
873 phi = u;
874 continue;
875 }
876 }
877 return;
878 }
879 if (phi != NULL) {
880 // Nothing in the loop before the store (next iteration)
881 // must observe the stored value
882 bool mem_ok = true;
883 {
884 ResourceMark rm;
885 Unique_Node_List wq;
886 wq.push(phi);
887 for (uint next = 0; next < wq.size() && mem_ok; ++next) {
888 Node *m = wq.at(next);
889 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) {
890 Node* u = m->fast_out(i);
891 if (u->is_Store() || u->is_Phi()) {
892 if (u != n) {
893 wq.push(u);
894 mem_ok = (wq.size() <= 10);
895 }
896 } else {
897 mem_ok = false;
898 break;
899 }
900 }
901 }
902 }
903 if (mem_ok) {
904 // Move the store out of the loop if the LCA of all
905 // users (except for the phi) is outside the loop.
906 Node* hook = new Node(1);
907 _igvn.rehash_node_delayed(phi);
908 int count = phi->replace_edge(n, hook);
909 assert(count > 0, "inconsistent phi");
910
911 // Compute latest point this store can go
912 Node* lca = get_late_ctrl(n, get_ctrl(n));
913 if (n_loop->is_member(get_loop(lca))) {
914 // LCA is in the loop - bail out
915 _igvn.replace_node(hook, n);
916 return;
917 }
918 #ifdef ASSERT
919 if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) {
920 assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined");
921 n_loop->_head->as_Loop()->verify_strip_mined(1);
922 Node* outer = n_loop->_head->as_CountedLoop()->outer_loop();
923 IdealLoopTree* outer_loop = get_loop(outer);
924 assert(n_loop->_parent == outer_loop, "broken loop tree");
925 assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state");
926 }
927 #endif
928
929 // Move store out of the loop
930 _igvn.replace_node(hook, n->in(MemNode::Memory));
931 _igvn.replace_input_of(n, 0, lca);
932 set_ctrl_and_loop(n, lca);
933
934 // Disconnect the phi now. An empty phi can confuse other
935 // optimizations in this pass of loop opts..
936 if (phi->in(LoopNode::LoopBackControl) == phi) {
937 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl));
938 n_loop->_body.yank(phi);
939 }
940 }
941 }
942 }
943 }
944 }
945 }
946
947 //------------------------------split_if_with_blocks_pre-----------------------
948 // Do the real work in a non-recursive function. Data nodes want to be
949 // cloned in the pre-order so they can feed each other nicely.
950 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
951 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
952 Node* bs_res = bs->split_if_pre(this, n);
953 if (bs_res != NULL) {
954 return bs_res;
955 }
956 // Cloning these guys is unlikely to win
957 int n_op = n->Opcode();
958 if( n_op == Op_MergeMem ) return n;
959 if( n->is_Proj() ) return n;
960 // Do not clone-up CmpFXXX variations, as these are always
961 // followed by a CmpI
962 if( n->is_Cmp() ) return n;
963 // Attempt to use a conditional move instead of a phi/branch
964 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) {
965 Node *cmov = conditional_move( n );
966 if( cmov ) return cmov;
967 }
968 if( n->is_CFG() || n->is_LoadStore() )
969 return n;
970 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd
971 n_op == Op_Opaque2 ) {
972 if( !C->major_progress() ) // If chance of no more loop opts...
973 _igvn._worklist.push(n); // maybe we'll remove them
974 return n;
975 }
976
977 if( n->is_Con() ) return n; // No cloning for Con nodes
978
979 Node *n_ctrl = get_ctrl(n);
980 if( !n_ctrl ) return n; // Dead node
981
982 Node* res = try_move_store_before_loop(n, n_ctrl);
983 if (res != NULL) {
984 return n;
985 }
986
987 // Attempt to remix address expressions for loop invariants
988 Node *m = remix_address_expressions( n );
989 if( m ) return m;
990
991 if (n_op == Op_AddI) {
992 Node *nn = convert_add_to_muladd( n );
993 if ( nn ) return nn;
994 }
995
996 if (n->is_ConstraintCast()) {
997 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this);
998 // ConstraintCastNode::dominating_cast() uses node control input to determine domination.
999 // Node control inputs don't necessarily agree with loop control info (due to
1000 // transformations happened in between), thus additional dominance check is needed
1001 // to keep loop info valid.
1002 if (dom_cast != NULL && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) {
1003 _igvn.replace_node(n, dom_cast);
1004 return dom_cast;
1005 }
1006 }
1007
1008 // Determine if the Node has inputs from some local Phi.
1009 // Returns the block to clone thru.
1010 Node *n_blk = has_local_phi_input( n );
1011 if( !n_blk ) return n;
1012
1013 // Do not clone the trip counter through on a CountedLoop
1014 // (messes up the canonical shape).
1015 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n;
1016
1017 // Check for having no control input; not pinned. Allow
1018 // dominating control.
1019 if (n->in(0)) {
1020 Node *dom = idom(n_blk);
1021 if (dom_lca(n->in(0), dom) != n->in(0)) {
1022 return n;
1023 }
1024 }
1025 // Policy: when is it profitable. You must get more wins than
1026 // policy before it is considered profitable. Policy is usually 0,
1027 // so 1 win is considered profitable. Big merges will require big
1028 // cloning, so get a larger policy.
1029 int policy = n_blk->req() >> 2;
1030
1031 // If the loop is a candidate for range check elimination,
1032 // delay splitting through it's phi until a later loop optimization
1033 if (n_blk->is_CountedLoop()) {
1034 IdealLoopTree *lp = get_loop(n_blk);
1035 if (lp && lp->_rce_candidate) {
1036 return n;
1037 }
1038 }
1039
1040 if (must_throttle_split_if()) return n;
1041
1042 // Split 'n' through the merge point if it is profitable
1043 Node *phi = split_thru_phi( n, n_blk, policy );
1044 if (!phi) return n;
1045
1046 // Found a Phi to split thru!
1047 // Replace 'n' with the new phi
1048 _igvn.replace_node( n, phi );
1049 // Moved a load around the loop, 'en-registering' something.
1050 if (n_blk->is_Loop() && n->is_Load() &&
1051 !phi->in(LoopNode::LoopBackControl)->is_Load())
1052 C->set_major_progress();
1053
1054 return phi;
1055 }
1056
1057 static bool merge_point_too_heavy(Compile* C, Node* region) {
1058 // Bail out if the region and its phis have too many users.
1059 int weight = 0;
1060 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1061 weight += region->fast_out(i)->outcnt();
1062 }
1063 int nodes_left = C->max_node_limit() - C->live_nodes();
1064 if (weight * 8 > nodes_left) {
1065 if (PrintOpto) {
1066 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight);
1067 }
1068 return true;
1069 } else {
1070 return false;
1071 }
1072 }
1073
1074 static bool merge_point_safe(Node* region) {
1075 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode
1076 // having a PhiNode input. This sidesteps the dangerous case where the split
1077 // ConvI2LNode may become TOP if the input Value() does not
1078 // overlap the ConvI2L range, leaving a node which may not dominate its
1079 // uses.
1080 // A better fix for this problem can be found in the BugTraq entry, but
1081 // expediency for Mantis demands this hack.
1082 // 6855164: If the merge point has a FastLockNode with a PhiNode input, we stop
1083 // split_if_with_blocks from splitting a block because we could not move around
1084 // the FastLockNode.
1085 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1086 Node* n = region->fast_out(i);
1087 if (n->is_Phi()) {
1088 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1089 Node* m = n->fast_out(j);
1090 if (m->is_FastLock())
1091 return false;
1092 #ifdef _LP64
1093 if (m->Opcode() == Op_ConvI2L)
1094 return false;
1095 if (m->is_CastII() && m->isa_CastII()->has_range_check()) {
1096 return false;
1097 }
1098 #endif
1099 }
1100 }
1101 }
1102 return true;
1103 }
1104
1105
1106 //------------------------------place_near_use---------------------------------
1107 // Place some computation next to use but not inside inner loops.
1108 // For inner loop uses move it to the preheader area.
1109 Node *PhaseIdealLoop::place_near_use(Node *useblock) const {
1110 IdealLoopTree *u_loop = get_loop( useblock );
1111 if (u_loop->_irreducible) {
1112 return useblock;
1113 }
1114 if (u_loop->_child) {
1115 if (useblock == u_loop->_head && u_loop->_head->is_OuterStripMinedLoop()) {
1116 return u_loop->_head->in(LoopNode::EntryControl);
1117 }
1118 return useblock;
1119 }
1120 return u_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl);
1121 }
1122
1123
1124 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) {
1125 if (!n->is_If() || n->is_CountedLoopEnd()) {
1126 return false;
1127 }
1128 if (!n->in(0)->is_Region()) {
1129 return false;
1130 }
1131 Node* region = n->in(0);
1132 Node* dom = idom(region);
1133 if (!dom->is_If() || dom->in(1) != n->in(1)) {
1134 return false;
1135 }
1136 IfNode* dom_if = dom->as_If();
1137 Node* proj_true = dom_if->proj_out(1);
1138 Node* proj_false = dom_if->proj_out(0);
1139
1140 for (uint i = 1; i < region->req(); i++) {
1141 if (is_dominator(proj_true, region->in(i))) {
1142 continue;
1143 }
1144 if (is_dominator(proj_false, region->in(i))) {
1145 continue;
1146 }
1147 return false;
1148 }
1149
1150 return true;
1151 }
1152
1153
1154 bool PhaseIdealLoop::can_split_if(Node* n_ctrl) {
1155 if (must_throttle_split_if()) {
1156 return false;
1157 }
1158
1159 // Do not do 'split-if' if irreducible loops are present.
1160 if (_has_irreducible_loops) {
1161 return false;
1162 }
1163
1164 if (merge_point_too_heavy(C, n_ctrl)) {
1165 return false;
1166 }
1167
1168 // Do not do 'split-if' if some paths are dead. First do dead code
1169 // elimination and then see if its still profitable.
1170 for (uint i = 1; i < n_ctrl->req(); i++) {
1171 if (n_ctrl->in(i) == C->top()) {
1172 return false;
1173 }
1174 }
1175
1176 // If trying to do a 'Split-If' at the loop head, it is only
1177 // profitable if the cmp folds up on BOTH paths. Otherwise we
1178 // risk peeling a loop forever.
1179
1180 // CNC - Disabled for now. Requires careful handling of loop
1181 // body selection for the cloned code. Also, make sure we check
1182 // for any input path not being in the same loop as n_ctrl. For
1183 // irreducible loops we cannot check for 'n_ctrl->is_Loop()'
1184 // because the alternative loop entry points won't be converted
1185 // into LoopNodes.
1186 IdealLoopTree *n_loop = get_loop(n_ctrl);
1187 for (uint j = 1; j < n_ctrl->req(); j++) {
1188 if (get_loop(n_ctrl->in(j)) != n_loop) {
1189 return false;
1190 }
1191 }
1192
1193 // Check for safety of the merge point.
1194 if (!merge_point_safe(n_ctrl)) {
1195 return false;
1196 }
1197
1198 return true;
1199 }
1200
1201 bool PhaseIdealLoop::flatten_array_element_type_check(Node *n) {
1202 // If the CmpP is a subtype check for a value that has just been
1203 // loaded from an array, the subtype checks guarantees the value
1204 // can't be stored in a flattened array and the load of the value
1205 // happens with a flattened array check then: push the type check
1206 // through the phi of the flattened array check. This needs special
1207 // logic because the subtype check's input is not a phi but a
1208 // LoadKlass that must first be cloned through the phi.
1209 if (n->Opcode() != Op_CmpP) {
1210 return false;
1211 }
1212
1213 Node* klassptr = n->in(1);
1214 Node* klasscon = n->in(2);
1215
1216 if (klassptr->is_DecodeNarrowPtr()) {
1217 klassptr = klassptr->in(1);
1218 }
1219
1220 if (klassptr->Opcode() != Op_LoadKlass && klassptr->Opcode() != Op_LoadNKlass) {
1221 return false;
1222 }
1223
1224 if (!klasscon->is_Con()) {
1225 return false;
1226 }
1227
1228 Node* addr = klassptr->in(MemNode::Address);
1229
1230 if (!addr->is_AddP()) {
1231 return false;
1232 }
1233
1234 intptr_t offset;
1235 Node* obj = AddPNode::Ideal_base_and_offset(addr, &_igvn, offset);
1236 assert(obj != NULL && addr->in(AddPNode::Base) == addr->in(AddPNode::Address), "malformed AddP?");
1237 if (obj->Opcode() == Op_CastPP) {
1238 obj = obj->in(1);
1239 }
1240
1241 if (!obj->is_Phi()) {
1242 return false;
1243 }
1244
1245 Node* region = obj->in(0);
1246
1247 Node* phi = PhiNode::make_blank(region, n->in(1));
1248 for (uint i = 1; i < region->req(); i++) {
1249 Node* in = obj->in(i);
1250 Node* ctrl = get_ctrl(in);
1251 if (addr->in(AddPNode::Base) != obj) {
1252 Node* cast = addr->in(AddPNode::Base);
1253 assert(cast->Opcode() == Op_CastPP && cast->in(0) != NULL, "inconsistent subgraph");
1254 Node* cast_clone = cast->clone();
1255 cast_clone->set_req(0, region->in(i));
1256 cast_clone->set_req(1, in);
1257 register_new_node(cast_clone, region->in(i));
1258 _igvn.set_type(cast_clone, cast_clone->Value(&_igvn));
1259 in = cast_clone;
1260 }
1261 Node* addr_clone = addr->clone();
1262 addr_clone->set_req(AddPNode::Base, in);
1263 addr_clone->set_req(AddPNode::Address, in);
1264 register_new_node(addr_clone, ctrl);
1265 _igvn.set_type(addr_clone, addr_clone->Value(&_igvn));
1266 Node* klassptr_clone = klassptr->clone();
1267 klassptr_clone->set_req(2, addr_clone);
1268 register_new_node(klassptr_clone, ctrl);
1269 _igvn.set_type(klassptr_clone, klassptr_clone->Value(&_igvn));
1270 if (klassptr != n->in(1)) {
1271 Node* decode = n->in(1);
1272 assert(decode->is_DecodeNarrowPtr(), "inconcistent subgraph");
1273 Node* decode_clone = decode->clone();
1274 decode_clone->set_req(1, klassptr_clone);
1275 register_new_node(decode_clone, ctrl);
1276 _igvn.set_type(decode_clone, decode_clone->Value(&_igvn));
1277 klassptr_clone = decode_clone;
1278 }
1279 phi->set_req(i, klassptr_clone);
1280 }
1281 register_new_node(phi, region);
1282 Node* orig = n->in(1);
1283 _igvn.replace_input_of(n, 1, phi);
1284 split_if_with_blocks_post(n);
1285 if (n->outcnt() != 0) {
1286 _igvn.replace_input_of(n, 1, orig);
1287 _igvn._worklist.push(phi);
1288 } else {
1289 }
1290 return true;
1291 }
1292
1293 //------------------------------split_if_with_blocks_post----------------------
1294 // Do the real work in a non-recursive function. CFG hackery wants to be
1295 // in the post-order, so it can dirty the I-DOM info and not use the dirtied
1296 // info.
1297 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) {
1298
1299 if (flatten_array_element_type_check(n)) {
1300 return;
1301 }
1302
1303 // Cloning Cmp through Phi's involves the split-if transform.
1304 // FastLock is not used by an If
1305 if (n->is_Cmp() && !n->is_FastLock()) {
1306 Node *n_ctrl = get_ctrl(n);
1307 // Determine if the Node has inputs from some local Phi.
1308 // Returns the block to clone thru.
1309 Node *n_blk = has_local_phi_input(n);
1310 if (n_blk != n_ctrl) {
1311 return;
1312 }
1313
1314 if (!can_split_if(n_ctrl)) {
1315 return;
1316 }
1317
1318 if (n->outcnt() != 1) {
1319 return; // Multiple bool's from 1 compare?
1320 }
1321 Node *bol = n->unique_out();
1322 assert(bol->is_Bool(), "expect a bool here");
1323 if (bol->outcnt() != 1) {
1324 return;// Multiple branches from 1 compare?
1325 }
1326 Node *iff = bol->unique_out();
1327
1328 // Check some safety conditions
1329 if (iff->is_If()) { // Classic split-if?
1330 if (iff->in(0) != n_ctrl) {
1331 return; // Compare must be in same blk as if
1332 }
1333 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE
1334 // Can't split CMove with different control edge.
1335 if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) {
1336 return;
1337 }
1338 if (get_ctrl(iff->in(2)) == n_ctrl ||
1339 get_ctrl(iff->in(3)) == n_ctrl) {
1340 return; // Inputs not yet split-up
1341 }
1342 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) {
1343 return; // Loop-invar test gates loop-varying CMOVE
1344 }
1345 } else {
1346 return; // some other kind of node, such as an Allocate
1347 }
1348
1349 // When is split-if profitable? Every 'win' on means some control flow
1350 // goes dead, so it's almost always a win.
1351 int policy = 0;
1352 // Split compare 'n' through the merge point if it is profitable
1353 Node *phi = split_thru_phi( n, n_ctrl, policy);
1354 if (!phi) {
1355 return;
1356 }
1357
1358 // Found a Phi to split thru!
1359 // Replace 'n' with the new phi
1360 _igvn.replace_node(n, phi);
1361
1362 // Now split the bool up thru the phi
1363 Node *bolphi = split_thru_phi(bol, n_ctrl, -1);
1364 guarantee(bolphi != NULL, "null boolean phi node");
1365
1366 _igvn.replace_node(bol, bolphi);
1367 assert(iff->in(1) == bolphi, "");
1368
1369 if (bolphi->Value(&_igvn)->singleton()) {
1370 return;
1371 }
1372
1373 // Conditional-move? Must split up now
1374 if (!iff->is_If()) {
1375 Node *cmovphi = split_thru_phi(iff, n_ctrl, -1);
1376 _igvn.replace_node(iff, cmovphi);
1377 return;
1378 }
1379
1380 // Now split the IF
1381 do_split_if(iff);
1382 return;
1383 }
1384
1385 // Two identical ifs back to back can be merged
1386 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) {
1387 Node *n_ctrl = n->in(0);
1388 PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1));
1389 IfNode* dom_if = idom(n_ctrl)->as_If();
1390 Node* proj_true = dom_if->proj_out(1);
1391 Node* proj_false = dom_if->proj_out(0);
1392 Node* con_true = _igvn.makecon(TypeInt::ONE);
1393 Node* con_false = _igvn.makecon(TypeInt::ZERO);
1394
1395 for (uint i = 1; i < n_ctrl->req(); i++) {
1396 if (is_dominator(proj_true, n_ctrl->in(i))) {
1397 bolphi->init_req(i, con_true);
1398 } else {
1399 assert(is_dominator(proj_false, n_ctrl->in(i)), "bad if");
1400 bolphi->init_req(i, con_false);
1401 }
1402 }
1403 register_new_node(bolphi, n_ctrl);
1404 _igvn.replace_input_of(n, 1, bolphi);
1405
1406 // Now split the IF
1407 do_split_if(n);
1408 return;
1409 }
1410
1411 // Check for an IF ready to split; one that has its
1412 // condition codes input coming from a Phi at the block start.
1413 int n_op = n->Opcode();
1414
1415 // Check for an IF being dominated by another IF same test
1416 if (n_op == Op_If ||
1417 n_op == Op_RangeCheck) {
1418 Node *bol = n->in(1);
1419 uint max = bol->outcnt();
1420 // Check for same test used more than once?
1421 if (max > 1 && bol->is_Bool()) {
1422 // Search up IDOMs to see if this IF is dominated.
1423 Node *cutoff = get_ctrl(bol);
1424
1425 // Now search up IDOMs till cutoff, looking for a dominating test
1426 Node *prevdom = n;
1427 Node *dom = idom(prevdom);
1428 while (dom != cutoff) {
1429 if (dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom) {
1430 // Replace the dominated test with an obvious true or false.
1431 // Place it on the IGVN worklist for later cleanup.
1432 C->set_major_progress();
1433 dominated_by(prevdom, n, false, true);
1434 #ifndef PRODUCT
1435 if( VerifyLoopOptimizations ) verify();
1436 #endif
1437 return;
1438 }
1439 prevdom = dom;
1440 dom = idom(prevdom);
1441 }
1442 }
1443 }
1444
1445 // See if a shared loop-varying computation has no loop-varying uses.
1446 // Happens if something is only used for JVM state in uncommon trap exits,
1447 // like various versions of induction variable+offset. Clone the
1448 // computation per usage to allow it to sink out of the loop.
1449 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about)
1450 Node *n_ctrl = get_ctrl(n);
1451 IdealLoopTree *n_loop = get_loop(n_ctrl);
1452 if( n_loop != _ltree_root ) {
1453 DUIterator_Fast imax, i = n->fast_outs(imax);
1454 for (; i < imax; i++) {
1455 Node* u = n->fast_out(i);
1456 if( !has_ctrl(u) ) break; // Found control user
1457 IdealLoopTree *u_loop = get_loop(get_ctrl(u));
1458 if( u_loop == n_loop ) break; // Found loop-varying use
1459 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop
1460 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003
1461 }
1462 bool did_break = (i < imax); // Did we break out of the previous loop?
1463 if (!did_break && n->outcnt() > 1) { // All uses in outer loops!
1464 Node *late_load_ctrl = NULL;
1465 if (n->is_Load()) {
1466 // If n is a load, get and save the result from get_late_ctrl(),
1467 // to be later used in calculating the control for n's clones.
1468 clear_dom_lca_tags();
1469 late_load_ctrl = get_late_ctrl(n, n_ctrl);
1470 }
1471 // If n is a load, and the late control is the same as the current
1472 // control, then the cloning of n is a pointless exercise, because
1473 // GVN will ensure that we end up where we started.
1474 if (!n->is_Load() || late_load_ctrl != n_ctrl) {
1475 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1476 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) {
1477 Node *u = n->last_out(j); // Clone private computation per use
1478 _igvn.rehash_node_delayed(u);
1479 Node *x = n->clone(); // Clone computation
1480 Node *x_ctrl = NULL;
1481 if( u->is_Phi() ) {
1482 // Replace all uses of normal nodes. Replace Phi uses
1483 // individually, so the separate Nodes can sink down
1484 // different paths.
1485 uint k = 1;
1486 while( u->in(k) != n ) k++;
1487 u->set_req( k, x );
1488 // x goes next to Phi input path
1489 x_ctrl = u->in(0)->in(k);
1490 --j;
1491 } else { // Normal use
1492 // Replace all uses
1493 for( uint k = 0; k < u->req(); k++ ) {
1494 if( u->in(k) == n ) {
1495 u->set_req( k, x );
1496 --j;
1497 }
1498 }
1499 x_ctrl = get_ctrl(u);
1500 }
1501
1502 // Find control for 'x' next to use but not inside inner loops.
1503 // For inner loop uses get the preheader area.
1504 x_ctrl = place_near_use(x_ctrl);
1505
1506 if (bs->sink_node(this, n, x, x_ctrl, n_ctrl)) {
1507 continue;
1508 }
1509
1510 if (n->is_Load()) {
1511 // For loads, add a control edge to a CFG node outside of the loop
1512 // to force them to not combine and return back inside the loop
1513 // during GVN optimization (4641526).
1514 //
1515 // Because we are setting the actual control input, factor in
1516 // the result from get_late_ctrl() so we respect any
1517 // anti-dependences. (6233005).
1518 x_ctrl = dom_lca(late_load_ctrl, x_ctrl);
1519
1520 // Don't allow the control input to be a CFG splitting node.
1521 // Such nodes should only have ProjNodes as outs, e.g. IfNode
1522 // should only have IfTrueNode and IfFalseNode (4985384).
1523 x_ctrl = find_non_split_ctrl(x_ctrl);
1524 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone");
1525
1526 x->set_req(0, x_ctrl);
1527 }
1528 register_new_node(x, x_ctrl);
1529
1530 // Some institutional knowledge is needed here: 'x' is
1531 // yanked because if the optimizer runs GVN on it all the
1532 // cloned x's will common up and undo this optimization and
1533 // be forced back in the loop. This is annoying because it
1534 // makes +VerifyOpto report false-positives on progress. I
1535 // tried setting control edges on the x's to force them to
1536 // not combine, but the matching gets worried when it tries
1537 // to fold a StoreP and an AddP together (as part of an
1538 // address expression) and the AddP and StoreP have
1539 // different controls.
1540 if (!x->is_Load() && !x->is_DecodeNarrowPtr()) _igvn._worklist.yank(x);
1541 }
1542 _igvn.remove_dead_node(n);
1543 }
1544 }
1545 }
1546 }
1547
1548 try_move_store_after_loop(n);
1549
1550 // Remove multiple allocations of the same value type
1551 if (n->is_ValueType() && EliminateAllocations) {
1552 n->as_ValueType()->remove_redundant_allocations(&_igvn, this);
1553 return; // n is now dead
1554 }
1555
1556 // Check for Opaque2's who's loop has disappeared - who's input is in the
1557 // same loop nest as their output. Remove 'em, they are no longer useful.
1558 if( n_op == Op_Opaque2 &&
1559 n->in(1) != NULL &&
1560 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) {
1561 _igvn.replace_node( n, n->in(1) );
1562 }
1563 }
1564
1565 //------------------------------split_if_with_blocks---------------------------
1566 // Check for aggressive application of 'split-if' optimization,
1567 // using basic block level info.
1568 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack) {
1569 Node* root = C->root();
1570 visited.set(root->_idx); // first, mark root as visited
1571 // Do pre-visit work for root
1572 Node* n = split_if_with_blocks_pre(root);
1573 uint cnt = n->outcnt();
1574 uint i = 0;
1575
1576 while (true) {
1577 // Visit all children
1578 if (i < cnt) {
1579 Node* use = n->raw_out(i);
1580 ++i;
1581 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
1582 // Now do pre-visit work for this use
1583 use = split_if_with_blocks_pre(use);
1584 nstack.push(n, i); // Save parent and next use's index.
1585 n = use; // Process all children of current use.
1586 cnt = use->outcnt();
1587 i = 0;
1588 }
1589 }
1590 else {
1591 // All of n's children have been processed, complete post-processing.
1592 if (cnt != 0 && !n->is_Con()) {
1593 assert(has_node(n), "no dead nodes");
1594 split_if_with_blocks_post(n);
1595 }
1596 if (must_throttle_split_if()) {
1597 nstack.clear();
1598 }
1599 if (nstack.is_empty()) {
1600 // Finished all nodes on stack.
1601 break;
1602 }
1603 // Get saved parent node and next use's index. Visit the rest of uses.
1604 n = nstack.node();
1605 cnt = n->outcnt();
1606 i = nstack.index();
1607 nstack.pop();
1608 }
1609 }
1610 }
1611
1612
1613 //=============================================================================
1614 //
1615 // C L O N E A L O O P B O D Y
1616 //
1617
1618 //------------------------------clone_iff--------------------------------------
1619 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1620 // "Nearly" because all Nodes have been cloned from the original in the loop,
1621 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1622 // through the Phi recursively, and return a Bool.
1623 Node* PhaseIdealLoop::clone_iff(PhiNode *phi, IdealLoopTree *loop) {
1624
1625 // Convert this Phi into a Phi merging Bools
1626 uint i;
1627 for (i = 1; i < phi->req(); i++) {
1628 Node *b = phi->in(i);
1629 if (b->is_Phi()) {
1630 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi(), loop));
1631 } else {
1632 assert(b->is_Bool() || b->Opcode() == Op_Opaque4, "");
1633 }
1634 }
1635
1636 Node* n = phi->in(1);
1637 Node* sample_opaque = NULL;
1638 Node *sample_bool = NULL;
1639 if (n->Opcode() == Op_Opaque4) {
1640 sample_opaque = n;
1641 sample_bool = n->in(1);
1642 assert(sample_bool->is_Bool(), "wrong type");
1643 } else {
1644 sample_bool = n;
1645 }
1646 Node *sample_cmp = sample_bool->in(1);
1647
1648 // Make Phis to merge the Cmp's inputs.
1649 PhiNode *phi1 = new PhiNode(phi->in(0), Type::TOP);
1650 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP);
1651 for (i = 1; i < phi->req(); i++) {
1652 Node *n1 = sample_opaque == NULL ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1);
1653 Node *n2 = sample_opaque == NULL ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2);
1654 phi1->set_req(i, n1);
1655 phi2->set_req(i, n2);
1656 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
1657 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
1658 }
1659 // See if these Phis have been made before.
1660 // Register with optimizer
1661 Node *hit1 = _igvn.hash_find_insert(phi1);
1662 if (hit1) { // Hit, toss just made Phi
1663 _igvn.remove_dead_node(phi1); // Remove new phi
1664 assert(hit1->is_Phi(), "" );
1665 phi1 = (PhiNode*)hit1; // Use existing phi
1666 } else { // Miss
1667 _igvn.register_new_node_with_optimizer(phi1);
1668 }
1669 Node *hit2 = _igvn.hash_find_insert(phi2);
1670 if (hit2) { // Hit, toss just made Phi
1671 _igvn.remove_dead_node(phi2); // Remove new phi
1672 assert(hit2->is_Phi(), "" );
1673 phi2 = (PhiNode*)hit2; // Use existing phi
1674 } else { // Miss
1675 _igvn.register_new_node_with_optimizer(phi2);
1676 }
1677 // Register Phis with loop/block info
1678 set_ctrl(phi1, phi->in(0));
1679 set_ctrl(phi2, phi->in(0));
1680 // Make a new Cmp
1681 Node *cmp = sample_cmp->clone();
1682 cmp->set_req(1, phi1);
1683 cmp->set_req(2, phi2);
1684 _igvn.register_new_node_with_optimizer(cmp);
1685 set_ctrl(cmp, phi->in(0));
1686
1687 // Make a new Bool
1688 Node *b = sample_bool->clone();
1689 b->set_req(1,cmp);
1690 _igvn.register_new_node_with_optimizer(b);
1691 set_ctrl(b, phi->in(0));
1692
1693 if (sample_opaque != NULL) {
1694 Node* opaque = sample_opaque->clone();
1695 opaque->set_req(1, b);
1696 _igvn.register_new_node_with_optimizer(opaque);
1697 set_ctrl(opaque, phi->in(0));
1698 return opaque;
1699 }
1700
1701 assert(b->is_Bool(), "");
1702 return b;
1703 }
1704
1705 //------------------------------clone_bool-------------------------------------
1706 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1707 // "Nearly" because all Nodes have been cloned from the original in the loop,
1708 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1709 // through the Phi recursively, and return a Bool.
1710 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) {
1711 uint i;
1712 // Convert this Phi into a Phi merging Bools
1713 for( i = 1; i < phi->req(); i++ ) {
1714 Node *b = phi->in(i);
1715 if( b->is_Phi() ) {
1716 _igvn.replace_input_of(phi, i, clone_bool( b->as_Phi(), loop ));
1717 } else {
1718 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
1719 }
1720 }
1721
1722 Node *sample_cmp = phi->in(1);
1723
1724 // Make Phis to merge the Cmp's inputs.
1725 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP );
1726 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP );
1727 for( uint j = 1; j < phi->req(); j++ ) {
1728 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
1729 Node *n1, *n2;
1730 if( cmp_top->is_Cmp() ) {
1731 n1 = cmp_top->in(1);
1732 n2 = cmp_top->in(2);
1733 } else {
1734 n1 = n2 = cmp_top;
1735 }
1736 phi1->set_req( j, n1 );
1737 phi2->set_req( j, n2 );
1738 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
1739 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
1740 }
1741
1742 // See if these Phis have been made before.
1743 // Register with optimizer
1744 Node *hit1 = _igvn.hash_find_insert(phi1);
1745 if( hit1 ) { // Hit, toss just made Phi
1746 _igvn.remove_dead_node(phi1); // Remove new phi
1747 assert( hit1->is_Phi(), "" );
1748 phi1 = (PhiNode*)hit1; // Use existing phi
1749 } else { // Miss
1750 _igvn.register_new_node_with_optimizer(phi1);
1751 }
1752 Node *hit2 = _igvn.hash_find_insert(phi2);
1753 if( hit2 ) { // Hit, toss just made Phi
1754 _igvn.remove_dead_node(phi2); // Remove new phi
1755 assert( hit2->is_Phi(), "" );
1756 phi2 = (PhiNode*)hit2; // Use existing phi
1757 } else { // Miss
1758 _igvn.register_new_node_with_optimizer(phi2);
1759 }
1760 // Register Phis with loop/block info
1761 set_ctrl(phi1, phi->in(0));
1762 set_ctrl(phi2, phi->in(0));
1763 // Make a new Cmp
1764 Node *cmp = sample_cmp->clone();
1765 cmp->set_req( 1, phi1 );
1766 cmp->set_req( 2, phi2 );
1767 _igvn.register_new_node_with_optimizer(cmp);
1768 set_ctrl(cmp, phi->in(0));
1769
1770 assert( cmp->is_Cmp(), "" );
1771 return (CmpNode*)cmp;
1772 }
1773
1774 //------------------------------sink_use---------------------------------------
1775 // If 'use' was in the loop-exit block, it now needs to be sunk
1776 // below the post-loop merge point.
1777 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) {
1778 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) {
1779 set_ctrl(use, post_loop);
1780 for (DUIterator j = use->outs(); use->has_out(j); j++)
1781 sink_use(use->out(j), post_loop);
1782 }
1783 }
1784
1785 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new,
1786 IdealLoopTree* loop, IdealLoopTree* outer_loop,
1787 Node_List*& split_if_set, Node_List*& split_bool_set,
1788 Node_List*& split_cex_set, Node_List& worklist,
1789 uint new_counter, CloneLoopMode mode) {
1790 Node* nnn = old_new[old->_idx];
1791 // Copy uses to a worklist, so I can munge the def-use info
1792 // with impunity.
1793 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
1794 worklist.push(old->fast_out(j));
1795
1796 while( worklist.size() ) {
1797 Node *use = worklist.pop();
1798 if (!has_node(use)) continue; // Ignore dead nodes
1799 if (use->in(0) == C->top()) continue;
1800 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
1801 // Check for data-use outside of loop - at least one of OLD or USE
1802 // must not be a CFG node.
1803 #ifdef ASSERT
1804 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == NULL) {
1805 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint();
1806 assert(mode == ControlAroundStripMined && use == sfpt, "missed a node");
1807 }
1808 #endif
1809 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) {
1810
1811 // If the Data use is an IF, that means we have an IF outside of the
1812 // loop that is switching on a condition that is set inside of the
1813 // loop. Happens if people set a loop-exit flag; then test the flag
1814 // in the loop to break the loop, then test is again outside of the
1815 // loop to determine which way the loop exited.
1816 // Loop predicate If node connects to Bool node through Opaque1 node.
1817 if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use) || use->Opcode() == Op_Opaque4) {
1818 // Since this code is highly unlikely, we lazily build the worklist
1819 // of such Nodes to go split.
1820 if (!split_if_set) {
1821 ResourceArea *area = Thread::current()->resource_area();
1822 split_if_set = new Node_List(area);
1823 }
1824 split_if_set->push(use);
1825 }
1826 if (use->is_Bool()) {
1827 if (!split_bool_set) {
1828 ResourceArea *area = Thread::current()->resource_area();
1829 split_bool_set = new Node_List(area);
1830 }
1831 split_bool_set->push(use);
1832 }
1833 if (use->Opcode() == Op_CreateEx) {
1834 if (!split_cex_set) {
1835 ResourceArea *area = Thread::current()->resource_area();
1836 split_cex_set = new Node_List(area);
1837 }
1838 split_cex_set->push(use);
1839 }
1840
1841
1842 // Get "block" use is in
1843 uint idx = 0;
1844 while( use->in(idx) != old ) idx++;
1845 Node *prev = use->is_CFG() ? use : get_ctrl(use);
1846 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" );
1847 Node *cfg = prev->_idx >= new_counter
1848 ? prev->in(2)
1849 : idom(prev);
1850 if( use->is_Phi() ) // Phi use is in prior block
1851 cfg = prev->in(idx); // NOT in block of Phi itself
1852 if (cfg->is_top()) { // Use is dead?
1853 _igvn.replace_input_of(use, idx, C->top());
1854 continue;
1855 }
1856
1857 // If use is referenced through control edge... (idx == 0)
1858 if (mode == IgnoreStripMined && idx == 0) {
1859 LoopNode *head = loop->_head->as_Loop();
1860 if (head->is_strip_mined() && is_dominator(head->outer_loop_exit(), prev)) {
1861 // That node is outside the inner loop, leave it outside the
1862 // outer loop as well to not confuse verification code.
1863 assert(!loop->_parent->is_member(use_loop), "should be out of the outer loop");
1864 _igvn.replace_input_of(use, 0, head->outer_loop_exit());
1865 continue;
1866 }
1867 }
1868
1869 while(!outer_loop->is_member(get_loop(cfg))) {
1870 prev = cfg;
1871 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg);
1872 }
1873 // If the use occurs after merging several exits from the loop, then
1874 // old value must have dominated all those exits. Since the same old
1875 // value was used on all those exits we did not need a Phi at this
1876 // merge point. NOW we do need a Phi here. Each loop exit value
1877 // is now merged with the peeled body exit; each exit gets its own
1878 // private Phi and those Phis need to be merged here.
1879 Node *phi;
1880 if( prev->is_Region() ) {
1881 if( idx == 0 ) { // Updating control edge?
1882 phi = prev; // Just use existing control
1883 } else { // Else need a new Phi
1884 phi = PhiNode::make( prev, old );
1885 // Now recursively fix up the new uses of old!
1886 for( uint i = 1; i < prev->req(); i++ ) {
1887 worklist.push(phi); // Onto worklist once for each 'old' input
1888 }
1889 }
1890 } else {
1891 // Get new RegionNode merging old and new loop exits
1892 prev = old_new[prev->_idx];
1893 assert( prev, "just made this in step 7" );
1894 if( idx == 0) { // Updating control edge?
1895 phi = prev; // Just use existing control
1896 } else { // Else need a new Phi
1897 // Make a new Phi merging data values properly
1898 phi = PhiNode::make( prev, old );
1899 phi->set_req( 1, nnn );
1900 }
1901 }
1902 // If inserting a new Phi, check for prior hits
1903 if( idx != 0 ) {
1904 Node *hit = _igvn.hash_find_insert(phi);
1905 if( hit == NULL ) {
1906 _igvn.register_new_node_with_optimizer(phi); // Register new phi
1907 } else { // or
1908 // Remove the new phi from the graph and use the hit
1909 _igvn.remove_dead_node(phi);
1910 phi = hit; // Use existing phi
1911 }
1912 set_ctrl(phi, prev);
1913 }
1914 // Make 'use' use the Phi instead of the old loop body exit value
1915 _igvn.replace_input_of(use, idx, phi);
1916 if( use->_idx >= new_counter ) { // If updating new phis
1917 // Not needed for correctness, but prevents a weak assert
1918 // in AddPNode from tripping (when we end up with different
1919 // base & derived Phis that will become the same after
1920 // IGVN does CSE).
1921 Node *hit = _igvn.hash_find_insert(use);
1922 if( hit ) // Go ahead and re-hash for hits.
1923 _igvn.replace_node( use, hit );
1924 }
1925
1926 // If 'use' was in the loop-exit block, it now needs to be sunk
1927 // below the post-loop merge point.
1928 sink_use( use, prev );
1929 }
1930 }
1931 }
1932
1933 static void clone_outer_loop_helper(Node* n, const IdealLoopTree *loop, const IdealLoopTree* outer_loop,
1934 const Node_List &old_new, Unique_Node_List& wq, PhaseIdealLoop* phase,
1935 bool check_old_new) {
1936 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1937 Node* u = n->fast_out(j);
1938 assert(check_old_new || old_new[u->_idx] == NULL, "shouldn't have been cloned");
1939 if (!u->is_CFG() && (!check_old_new || old_new[u->_idx] == NULL)) {
1940 Node* c = phase->get_ctrl(u);
1941 IdealLoopTree* u_loop = phase->get_loop(c);
1942 assert(!loop->is_member(u_loop), "can be in outer loop or out of both loops only");
1943 if (outer_loop->is_member(u_loop)) {
1944 wq.push(u);
1945 }
1946 }
1947 }
1948 }
1949
1950 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
1951 IdealLoopTree* outer_loop, int dd, Node_List &old_new,
1952 Node_List& extra_data_nodes) {
1953 if (head->is_strip_mined() && mode != IgnoreStripMined) {
1954 CountedLoopNode* cl = head->as_CountedLoop();
1955 Node* l = cl->outer_loop();
1956 Node* tail = cl->outer_loop_tail();
1957 IfNode* le = cl->outer_loop_end();
1958 Node* sfpt = cl->outer_safepoint();
1959 CountedLoopEndNode* cle = cl->loopexit();
1960 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop();
1961 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null();
1962 Node* cle_out = cle->proj_out(false);
1963
1964 Node* new_sfpt = NULL;
1965 Node* new_cle_out = cle_out->clone();
1966 old_new.map(cle_out->_idx, new_cle_out);
1967 if (mode == CloneIncludesStripMined) {
1968 // clone outer loop body
1969 Node* new_l = l->clone();
1970 Node* new_tail = tail->clone();
1971 IfNode* new_le = le->clone()->as_If();
1972 new_sfpt = sfpt->clone();
1973
1974 set_loop(new_l, outer_loop->_parent);
1975 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd);
1976 set_loop(new_cle_out, outer_loop->_parent);
1977 set_idom(new_cle_out, new_cle, dd);
1978 set_loop(new_sfpt, outer_loop->_parent);
1979 set_idom(new_sfpt, new_cle_out, dd);
1980 set_loop(new_le, outer_loop->_parent);
1981 set_idom(new_le, new_sfpt, dd);
1982 set_loop(new_tail, outer_loop->_parent);
1983 set_idom(new_tail, new_le, dd);
1984 set_idom(new_cl, new_l, dd);
1985
1986 old_new.map(l->_idx, new_l);
1987 old_new.map(tail->_idx, new_tail);
1988 old_new.map(le->_idx, new_le);
1989 old_new.map(sfpt->_idx, new_sfpt);
1990
1991 new_l->set_req(LoopNode::LoopBackControl, new_tail);
1992 new_l->set_req(0, new_l);
1993 new_tail->set_req(0, new_le);
1994 new_le->set_req(0, new_sfpt);
1995 new_sfpt->set_req(0, new_cle_out);
1996 new_cle_out->set_req(0, new_cle);
1997 new_cl->set_req(LoopNode::EntryControl, new_l);
1998
1999 _igvn.register_new_node_with_optimizer(new_l);
2000 _igvn.register_new_node_with_optimizer(new_tail);
2001 _igvn.register_new_node_with_optimizer(new_le);
2002 } else {
2003 Node *newhead = old_new[loop->_head->_idx];
2004 newhead->as_Loop()->clear_strip_mined();
2005 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl));
2006 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
2007 }
2008 // Look at data node that were assigned a control in the outer
2009 // loop: they are kept in the outer loop by the safepoint so start
2010 // from the safepoint node's inputs.
2011 IdealLoopTree* outer_loop = get_loop(l);
2012 Node_Stack stack(2);
2013 stack.push(sfpt, 1);
2014 uint new_counter = C->unique();
2015 while (stack.size() > 0) {
2016 Node* n = stack.node();
2017 uint i = stack.index();
2018 while (i < n->req() &&
2019 (n->in(i) == NULL ||
2020 !has_ctrl(n->in(i)) ||
2021 get_loop(get_ctrl(n->in(i))) != outer_loop ||
2022 (old_new[n->in(i)->_idx] != NULL && old_new[n->in(i)->_idx]->_idx >= new_counter))) {
2023 i++;
2024 }
2025 if (i < n->req()) {
2026 stack.set_index(i+1);
2027 stack.push(n->in(i), 0);
2028 } else {
2029 assert(old_new[n->_idx] == NULL || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet");
2030 Node* m = n == sfpt ? new_sfpt : n->clone();
2031 if (m != NULL) {
2032 for (uint i = 0; i < n->req(); i++) {
2033 if (m->in(i) != NULL && old_new[m->in(i)->_idx] != NULL) {
2034 m->set_req(i, old_new[m->in(i)->_idx]);
2035 }
2036 }
2037 } else {
2038 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?");
2039 }
2040 if (n != sfpt) {
2041 extra_data_nodes.push(n);
2042 _igvn.register_new_node_with_optimizer(m);
2043 assert(get_ctrl(n) == cle_out, "what other control?");
2044 set_ctrl(m, new_cle_out);
2045 old_new.map(n->_idx, m);
2046 }
2047 stack.pop();
2048 }
2049 }
2050 if (mode == CloneIncludesStripMined) {
2051 _igvn.register_new_node_with_optimizer(new_sfpt);
2052 _igvn.register_new_node_with_optimizer(new_cle_out);
2053 }
2054 // Some other transformation may have pessimistically assign some
2055 // data nodes to the outer loop. Set their control so they are out
2056 // of the outer loop.
2057 ResourceMark rm;
2058 Unique_Node_List wq;
2059 for (uint i = 0; i < extra_data_nodes.size(); i++) {
2060 Node* old = extra_data_nodes.at(i);
2061 clone_outer_loop_helper(old, loop, outer_loop, old_new, wq, this, true);
2062 }
2063 Node* new_ctrl = cl->outer_loop_exit();
2064 assert(get_loop(new_ctrl) != outer_loop, "must be out of the loop nest");
2065 for (uint i = 0; i < wq.size(); i++) {
2066 Node* n = wq.at(i);
2067 set_ctrl(n, new_ctrl);
2068 clone_outer_loop_helper(n, loop, outer_loop, old_new, wq, this, false);
2069 }
2070 } else {
2071 Node *newhead = old_new[loop->_head->_idx];
2072 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
2073 }
2074 }
2075
2076 //------------------------------clone_loop-------------------------------------
2077 //
2078 // C L O N E A L O O P B O D Y
2079 //
2080 // This is the basic building block of the loop optimizations. It clones an
2081 // entire loop body. It makes an old_new loop body mapping; with this mapping
2082 // you can find the new-loop equivalent to an old-loop node. All new-loop
2083 // nodes are exactly equal to their old-loop counterparts, all edges are the
2084 // same. All exits from the old-loop now have a RegionNode that merges the
2085 // equivalent new-loop path. This is true even for the normal "loop-exit"
2086 // condition. All uses of loop-invariant old-loop values now come from (one
2087 // or more) Phis that merge their new-loop equivalents.
2088 //
2089 // This operation leaves the graph in an illegal state: there are two valid
2090 // control edges coming from the loop pre-header to both loop bodies. I'll
2091 // definitely have to hack the graph after running this transform.
2092 //
2093 // From this building block I will further edit edges to perform loop peeling
2094 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
2095 //
2096 // Parameter side_by_size_idom:
2097 // When side_by_size_idom is NULL, the dominator tree is constructed for
2098 // the clone loop to dominate the original. Used in construction of
2099 // pre-main-post loop sequence.
2100 // When nonnull, the clone and original are side-by-side, both are
2101 // dominated by the side_by_side_idom node. Used in construction of
2102 // unswitched loops.
2103 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
2104 CloneLoopMode mode, Node* side_by_side_idom) {
2105
2106 LoopNode* head = loop->_head->as_Loop();
2107 head->verify_strip_mined(1);
2108
2109 if (C->do_vector_loop() && PrintOpto) {
2110 const char* mname = C->method()->name()->as_quoted_ascii();
2111 if (mname != NULL) {
2112 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname);
2113 }
2114 }
2115
2116 CloneMap& cm = C->clone_map();
2117 Dict* dict = cm.dict();
2118 if (C->do_vector_loop()) {
2119 cm.set_clone_idx(cm.max_gen()+1);
2120 #ifndef PRODUCT
2121 if (PrintOpto) {
2122 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx());
2123 loop->dump_head();
2124 }
2125 #endif
2126 }
2127
2128 // Step 1: Clone the loop body. Make the old->new mapping.
2129 uint i;
2130 for( i = 0; i < loop->_body.size(); i++ ) {
2131 Node *old = loop->_body.at(i);
2132 Node *nnn = old->clone();
2133 old_new.map( old->_idx, nnn );
2134 if (C->do_vector_loop()) {
2135 cm.verify_insert_and_clone(old, nnn, cm.clone_idx());
2136 }
2137 _igvn.register_new_node_with_optimizer(nnn);
2138 }
2139
2140 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop;
2141
2142 // Step 2: Fix the edges in the new body. If the old input is outside the
2143 // loop use it. If the old input is INside the loop, use the corresponding
2144 // new node instead.
2145 for( i = 0; i < loop->_body.size(); i++ ) {
2146 Node *old = loop->_body.at(i);
2147 Node *nnn = old_new[old->_idx];
2148 // Fix CFG/Loop controlling the new node
2149 if (has_ctrl(old)) {
2150 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
2151 } else {
2152 set_loop(nnn, outer_loop->_parent);
2153 if (old->outcnt() > 0) {
2154 set_idom( nnn, old_new[idom(old)->_idx], dd );
2155 }
2156 }
2157 // Correct edges to the new node
2158 for( uint j = 0; j < nnn->req(); j++ ) {
2159 Node *n = nnn->in(j);
2160 if( n ) {
2161 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n );
2162 if( loop->is_member( old_in_loop ) )
2163 nnn->set_req(j, old_new[n->_idx]);
2164 }
2165 }
2166 _igvn.hash_find_insert(nnn);
2167 }
2168
2169 ResourceArea *area = Thread::current()->resource_area();
2170 Node_List extra_data_nodes(area); // data nodes in the outer strip mined loop
2171 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes);
2172
2173 // Step 3: Now fix control uses. Loop varying control uses have already
2174 // been fixed up (as part of all input edges in Step 2). Loop invariant
2175 // control uses must be either an IfFalse or an IfTrue. Make a merge
2176 // point to merge the old and new IfFalse/IfTrue nodes; make the use
2177 // refer to this.
2178 Node_List worklist(area);
2179 uint new_counter = C->unique();
2180 for( i = 0; i < loop->_body.size(); i++ ) {
2181 Node* old = loop->_body.at(i);
2182 if( !old->is_CFG() ) continue;
2183
2184 // Copy uses to a worklist, so I can munge the def-use info
2185 // with impunity.
2186 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
2187 worklist.push(old->fast_out(j));
2188
2189 while( worklist.size() ) { // Visit all uses
2190 Node *use = worklist.pop();
2191 if (!has_node(use)) continue; // Ignore dead nodes
2192 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
2193 if( !loop->is_member( use_loop ) && use->is_CFG() ) {
2194 // Both OLD and USE are CFG nodes here.
2195 assert( use->is_Proj(), "" );
2196 Node* nnn = old_new[old->_idx];
2197
2198 Node* newuse = NULL;
2199 if (head->is_strip_mined() && mode != IgnoreStripMined) {
2200 CountedLoopNode* cl = head->as_CountedLoop();
2201 CountedLoopEndNode* cle = cl->loopexit();
2202 Node* cle_out = cle->proj_out_or_null(false);
2203 if (use == cle_out) {
2204 IfNode* le = cl->outer_loop_end();
2205 use = le->proj_out(false);
2206 use_loop = get_loop(use);
2207 if (mode == CloneIncludesStripMined) {
2208 nnn = old_new[le->_idx];
2209 } else {
2210 newuse = old_new[cle_out->_idx];
2211 }
2212 }
2213 }
2214 if (newuse == NULL) {
2215 newuse = use->clone();
2216 }
2217
2218 // Clone the loop exit control projection
2219 if (C->do_vector_loop()) {
2220 cm.verify_insert_and_clone(use, newuse, cm.clone_idx());
2221 }
2222 newuse->set_req(0,nnn);
2223 _igvn.register_new_node_with_optimizer(newuse);
2224 set_loop(newuse, use_loop);
2225 set_idom(newuse, nnn, dom_depth(nnn) + 1 );
2226
2227 // We need a Region to merge the exit from the peeled body and the
2228 // exit from the old loop body.
2229 RegionNode *r = new RegionNode(3);
2230 // Map the old use to the new merge point
2231 old_new.map( use->_idx, r );
2232 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use));
2233 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" );
2234
2235 // The original user of 'use' uses 'r' instead.
2236 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
2237 Node* useuse = use->last_out(l);
2238 _igvn.rehash_node_delayed(useuse);
2239 uint uses_found = 0;
2240 if( useuse->in(0) == use ) {
2241 useuse->set_req(0, r);
2242 uses_found++;
2243 if( useuse->is_CFG() ) {
2244 assert( dom_depth(useuse) > dd_r, "" );
2245 set_idom(useuse, r, dom_depth(useuse));
2246 }
2247 }
2248 for( uint k = 1; k < useuse->req(); k++ ) {
2249 if( useuse->in(k) == use ) {
2250 useuse->set_req(k, r);
2251 uses_found++;
2252 if (useuse->is_Loop() && k == LoopNode::EntryControl) {
2253 assert(dom_depth(useuse) > dd_r , "");
2254 set_idom(useuse, r, dom_depth(useuse));
2255 }
2256 }
2257 }
2258 l -= uses_found; // we deleted 1 or more copies of this edge
2259 }
2260
2261 // Now finish up 'r'
2262 r->set_req( 1, newuse );
2263 r->set_req( 2, use );
2264 _igvn.register_new_node_with_optimizer(r);
2265 set_loop(r, use_loop);
2266 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r);
2267 } // End of if a loop-exit test
2268 }
2269 }
2270
2271 // Step 4: If loop-invariant use is not control, it must be dominated by a
2272 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region
2273 // there if needed. Make a Phi there merging old and new used values.
2274 Node_List *split_if_set = NULL;
2275 Node_List *split_bool_set = NULL;
2276 Node_List *split_cex_set = NULL;
2277 for( i = 0; i < loop->_body.size(); i++ ) {
2278 Node* old = loop->_body.at(i);
2279 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2280 split_bool_set, split_cex_set, worklist, new_counter,
2281 mode);
2282 }
2283
2284 for (i = 0; i < extra_data_nodes.size(); i++) {
2285 Node* old = extra_data_nodes.at(i);
2286 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2287 split_bool_set, split_cex_set, worklist, new_counter,
2288 mode);
2289 }
2290
2291 // Check for IFs that need splitting/cloning. Happens if an IF outside of
2292 // the loop uses a condition set in the loop. The original IF probably
2293 // takes control from one or more OLD Regions (which in turn get from NEW
2294 // Regions). In any case, there will be a set of Phis for each merge point
2295 // from the IF up to where the original BOOL def exists the loop.
2296 if (split_if_set) {
2297 while (split_if_set->size()) {
2298 Node *iff = split_if_set->pop();
2299 if (iff->in(1)->is_Phi()) {
2300 Node *b = clone_iff(iff->in(1)->as_Phi(), loop);
2301 _igvn.replace_input_of(iff, 1, b);
2302 }
2303 }
2304 }
2305 if (split_bool_set) {
2306 while (split_bool_set->size()) {
2307 Node *b = split_bool_set->pop();
2308 Node *phi = b->in(1);
2309 assert(phi->is_Phi(), "");
2310 CmpNode *cmp = clone_bool((PhiNode*)phi, loop);
2311 _igvn.replace_input_of(b, 1, cmp);
2312 }
2313 }
2314 if (split_cex_set) {
2315 while (split_cex_set->size()) {
2316 Node *b = split_cex_set->pop();
2317 assert(b->in(0)->is_Region(), "");
2318 assert(b->in(1)->is_Phi(), "");
2319 assert(b->in(0)->in(0) == b->in(1)->in(0), "");
2320 split_up(b, b->in(0), NULL);
2321 }
2322 }
2323
2324 }
2325
2326
2327 //---------------------- stride_of_possible_iv -------------------------------------
2328 // Looks for an iff/bool/comp with one operand of the compare
2329 // being a cycle involving an add and a phi,
2330 // with an optional truncation (left-shift followed by a right-shift)
2331 // of the add. Returns zero if not an iv.
2332 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
2333 Node* trunc1 = NULL;
2334 Node* trunc2 = NULL;
2335 const TypeInt* ttype = NULL;
2336 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) {
2337 return 0;
2338 }
2339 BoolNode* bl = iff->in(1)->as_Bool();
2340 Node* cmp = bl->in(1);
2341 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) {
2342 return 0;
2343 }
2344 // Must have an invariant operand
2345 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) {
2346 return 0;
2347 }
2348 Node* add2 = NULL;
2349 Node* cmp1 = cmp->in(1);
2350 if (cmp1->is_Phi()) {
2351 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
2352 Node* phi = cmp1;
2353 for (uint i = 1; i < phi->req(); i++) {
2354 Node* in = phi->in(i);
2355 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in,
2356 &trunc1, &trunc2, &ttype);
2357 if (add && add->in(1) == phi) {
2358 add2 = add->in(2);
2359 break;
2360 }
2361 }
2362 } else {
2363 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
2364 Node* addtrunc = cmp1;
2365 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc,
2366 &trunc1, &trunc2, &ttype);
2367 if (add && add->in(1)->is_Phi()) {
2368 Node* phi = add->in(1);
2369 for (uint i = 1; i < phi->req(); i++) {
2370 if (phi->in(i) == addtrunc) {
2371 add2 = add->in(2);
2372 break;
2373 }
2374 }
2375 }
2376 }
2377 if (add2 != NULL) {
2378 const TypeInt* add2t = _igvn.type(add2)->is_int();
2379 if (add2t->is_con()) {
2380 return add2t->get_con();
2381 }
2382 }
2383 return 0;
2384 }
2385
2386
2387 //---------------------- stay_in_loop -------------------------------------
2388 // Return the (unique) control output node that's in the loop (if it exists.)
2389 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
2390 Node* unique = NULL;
2391 if (!n) return NULL;
2392 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2393 Node* use = n->fast_out(i);
2394 if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
2395 if (unique != NULL) {
2396 return NULL;
2397 }
2398 unique = use;
2399 }
2400 }
2401 return unique;
2402 }
2403
2404 //------------------------------ register_node -------------------------------------
2405 // Utility to register node "n" with PhaseIdealLoop
2406 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) {
2407 _igvn.register_new_node_with_optimizer(n);
2408 loop->_body.push(n);
2409 if (n->is_CFG()) {
2410 set_loop(n, loop);
2411 set_idom(n, pred, ddepth);
2412 } else {
2413 set_ctrl(n, pred);
2414 }
2415 }
2416
2417 //------------------------------ proj_clone -------------------------------------
2418 // Utility to create an if-projection
2419 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
2420 ProjNode* c = p->clone()->as_Proj();
2421 c->set_req(0, iff);
2422 return c;
2423 }
2424
2425 //------------------------------ short_circuit_if -------------------------------------
2426 // Force the iff control output to be the live_proj
2427 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
2428 guarantee(live_proj != NULL, "null projection");
2429 int proj_con = live_proj->_con;
2430 assert(proj_con == 0 || proj_con == 1, "false or true projection");
2431 Node *con = _igvn.intcon(proj_con);
2432 set_ctrl(con, C->root());
2433 if (iff) {
2434 iff->set_req(1, con);
2435 }
2436 return con;
2437 }
2438
2439 //------------------------------ insert_if_before_proj -------------------------------------
2440 // Insert a new if before an if projection (* - new node)
2441 //
2442 // before
2443 // if(test)
2444 // / \
2445 // v v
2446 // other-proj proj (arg)
2447 //
2448 // after
2449 // if(test)
2450 // / \
2451 // / v
2452 // | * proj-clone
2453 // v |
2454 // other-proj v
2455 // * new_if(relop(cmp[IU](left,right)))
2456 // / \
2457 // v v
2458 // * new-proj proj
2459 // (returned)
2460 //
2461 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) {
2462 IfNode* iff = proj->in(0)->as_If();
2463 IdealLoopTree *loop = get_loop(proj);
2464 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2465 int ddepth = dom_depth(proj);
2466
2467 _igvn.rehash_node_delayed(iff);
2468 _igvn.rehash_node_delayed(proj);
2469
2470 proj->set_req(0, NULL); // temporary disconnect
2471 ProjNode* proj2 = proj_clone(proj, iff);
2472 register_node(proj2, loop, iff, ddepth);
2473
2474 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right);
2475 register_node(cmp, loop, proj2, ddepth);
2476
2477 BoolNode* bol = new BoolNode(cmp, relop);
2478 register_node(bol, loop, proj2, ddepth);
2479
2480 int opcode = iff->Opcode();
2481 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode");
2482 IfNode* new_if = (opcode == Op_If) ? new IfNode(proj2, bol, iff->_prob, iff->_fcnt):
2483 new RangeCheckNode(proj2, bol, iff->_prob, iff->_fcnt);
2484 register_node(new_if, loop, proj2, ddepth);
2485
2486 proj->set_req(0, new_if); // reattach
2487 set_idom(proj, new_if, ddepth);
2488
2489 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
2490 guarantee(new_exit != NULL, "null exit node");
2491 register_node(new_exit, get_loop(other_proj), new_if, ddepth);
2492
2493 return new_exit;
2494 }
2495
2496 //------------------------------ insert_region_before_proj -------------------------------------
2497 // Insert a region before an if projection (* - new node)
2498 //
2499 // before
2500 // if(test)
2501 // / |
2502 // v |
2503 // proj v
2504 // other-proj
2505 //
2506 // after
2507 // if(test)
2508 // / |
2509 // v |
2510 // * proj-clone v
2511 // | other-proj
2512 // v
2513 // * new-region
2514 // |
2515 // v
2516 // * dum_if
2517 // / \
2518 // v \
2519 // * dum-proj v
2520 // proj
2521 //
2522 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
2523 IfNode* iff = proj->in(0)->as_If();
2524 IdealLoopTree *loop = get_loop(proj);
2525 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2526 int ddepth = dom_depth(proj);
2527
2528 _igvn.rehash_node_delayed(iff);
2529 _igvn.rehash_node_delayed(proj);
2530
2531 proj->set_req(0, NULL); // temporary disconnect
2532 ProjNode* proj2 = proj_clone(proj, iff);
2533 register_node(proj2, loop, iff, ddepth);
2534
2535 RegionNode* reg = new RegionNode(2);
2536 reg->set_req(1, proj2);
2537 register_node(reg, loop, iff, ddepth);
2538
2539 IfNode* dum_if = new IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt);
2540 register_node(dum_if, loop, reg, ddepth);
2541
2542 proj->set_req(0, dum_if); // reattach
2543 set_idom(proj, dum_if, ddepth);
2544
2545 ProjNode* dum_proj = proj_clone(other_proj, dum_if);
2546 register_node(dum_proj, loop, dum_if, ddepth);
2547
2548 return reg;
2549 }
2550
2551 //------------------------------ insert_cmpi_loop_exit -------------------------------------
2552 // Clone a signed compare loop exit from an unsigned compare and
2553 // insert it before the unsigned cmp on the stay-in-loop path.
2554 // All new nodes inserted in the dominator tree between the original
2555 // if and it's projections. The original if test is replaced with
2556 // a constant to force the stay-in-loop path.
2557 //
2558 // This is done to make sure that the original if and it's projections
2559 // still dominate the same set of control nodes, that the ctrl() relation
2560 // from data nodes to them is preserved, and that their loop nesting is
2561 // preserved.
2562 //
2563 // before
2564 // if(i <u limit) unsigned compare loop exit
2565 // / |
2566 // v v
2567 // exit-proj stay-in-loop-proj
2568 //
2569 // after
2570 // if(stay-in-loop-const) original if
2571 // / |
2572 // / v
2573 // / if(i < limit) new signed test
2574 // / / |
2575 // / / v
2576 // / / if(i <u limit) new cloned unsigned test
2577 // / / / |
2578 // v v v |
2579 // region |
2580 // | |
2581 // dum-if |
2582 // / | |
2583 // ether | |
2584 // v v
2585 // exit-proj stay-in-loop-proj
2586 //
2587 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) {
2588 const bool Signed = true;
2589 const bool Unsigned = false;
2590
2591 BoolNode* bol = if_cmpu->in(1)->as_Bool();
2592 if (bol->_test._test != BoolTest::lt) return NULL;
2593 CmpNode* cmpu = bol->in(1)->as_Cmp();
2594 if (cmpu->Opcode() != Op_CmpU) return NULL;
2595 int stride = stride_of_possible_iv(if_cmpu);
2596 if (stride == 0) return NULL;
2597
2598 Node* lp_proj = stay_in_loop(if_cmpu, loop);
2599 guarantee(lp_proj != NULL, "null loop node");
2600
2601 ProjNode* lp_continue = lp_proj->as_Proj();
2602 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
2603
2604 Node* limit = NULL;
2605 if (stride > 0) {
2606 limit = cmpu->in(2);
2607 } else {
2608 limit = _igvn.makecon(TypeInt::ZERO);
2609 set_ctrl(limit, C->root());
2610 }
2611 // Create a new region on the exit path
2612 RegionNode* reg = insert_region_before_proj(lp_exit);
2613 guarantee(reg != NULL, "null region node");
2614
2615 // Clone the if-cmpu-true-false using a signed compare
2616 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
2617 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue);
2618 reg->add_req(cmpi_exit);
2619
2620 // Clone the if-cmpu-true-false
2621 BoolTest::mask rel_u = bol->_test._test;
2622 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
2623 reg->add_req(cmpu_exit);
2624
2625 // Force original if to stay in loop.
2626 short_circuit_if(if_cmpu, lp_continue);
2627
2628 return cmpi_exit->in(0)->as_If();
2629 }
2630
2631 //------------------------------ remove_cmpi_loop_exit -------------------------------------
2632 // Remove a previously inserted signed compare loop exit.
2633 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
2634 Node* lp_proj = stay_in_loop(if_cmp, loop);
2635 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
2636 stay_in_loop(lp_proj, loop)->is_If() &&
2637 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
2638 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
2639 set_ctrl(con, C->root());
2640 if_cmp->set_req(1, con);
2641 }
2642
2643 //------------------------------ scheduled_nodelist -------------------------------------
2644 // Create a post order schedule of nodes that are in the
2645 // "member" set. The list is returned in "sched".
2646 // The first node in "sched" is the loop head, followed by
2647 // nodes which have no inputs in the "member" set, and then
2648 // followed by the nodes that have an immediate input dependence
2649 // on a node in "sched".
2650 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
2651
2652 assert(member.test(loop->_head->_idx), "loop head must be in member set");
2653 Arena *a = Thread::current()->resource_area();
2654 VectorSet visited(a);
2655 Node_Stack nstack(a, loop->_body.size());
2656
2657 Node* n = loop->_head; // top of stack is cached in "n"
2658 uint idx = 0;
2659 visited.set(n->_idx);
2660
2661 // Initially push all with no inputs from within member set
2662 for(uint i = 0; i < loop->_body.size(); i++ ) {
2663 Node *elt = loop->_body.at(i);
2664 if (member.test(elt->_idx)) {
2665 bool found = false;
2666 for (uint j = 0; j < elt->req(); j++) {
2667 Node* def = elt->in(j);
2668 if (def && member.test(def->_idx) && def != elt) {
2669 found = true;
2670 break;
2671 }
2672 }
2673 if (!found && elt != loop->_head) {
2674 nstack.push(n, idx);
2675 n = elt;
2676 assert(!visited.test(n->_idx), "not seen yet");
2677 visited.set(n->_idx);
2678 }
2679 }
2680 }
2681
2682 // traverse out's that are in the member set
2683 while (true) {
2684 if (idx < n->outcnt()) {
2685 Node* use = n->raw_out(idx);
2686 idx++;
2687 if (!visited.test_set(use->_idx)) {
2688 if (member.test(use->_idx)) {
2689 nstack.push(n, idx);
2690 n = use;
2691 idx = 0;
2692 }
2693 }
2694 } else {
2695 // All outputs processed
2696 sched.push(n);
2697 if (nstack.is_empty()) break;
2698 n = nstack.node();
2699 idx = nstack.index();
2700 nstack.pop();
2701 }
2702 }
2703 }
2704
2705
2706 //------------------------------ has_use_in_set -------------------------------------
2707 // Has a use in the vector set
2708 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
2709 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2710 Node* use = n->fast_out(j);
2711 if (vset.test(use->_idx)) {
2712 return true;
2713 }
2714 }
2715 return false;
2716 }
2717
2718
2719 //------------------------------ has_use_internal_to_set -------------------------------------
2720 // Has use internal to the vector set (ie. not in a phi at the loop head)
2721 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
2722 Node* head = loop->_head;
2723 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2724 Node* use = n->fast_out(j);
2725 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
2726 return true;
2727 }
2728 }
2729 return false;
2730 }
2731
2732
2733 //------------------------------ clone_for_use_outside_loop -------------------------------------
2734 // clone "n" for uses that are outside of loop
2735 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
2736 int cloned = 0;
2737 assert(worklist.size() == 0, "should be empty");
2738 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2739 Node* use = n->fast_out(j);
2740 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
2741 worklist.push(use);
2742 }
2743 }
2744 while( worklist.size() ) {
2745 Node *use = worklist.pop();
2746 if (!has_node(use) || use->in(0) == C->top()) continue;
2747 uint j;
2748 for (j = 0; j < use->req(); j++) {
2749 if (use->in(j) == n) break;
2750 }
2751 assert(j < use->req(), "must be there");
2752
2753 // clone "n" and insert it between the inputs of "n" and the use outside the loop
2754 Node* n_clone = n->clone();
2755 _igvn.replace_input_of(use, j, n_clone);
2756 cloned++;
2757 Node* use_c;
2758 if (!use->is_Phi()) {
2759 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
2760 } else {
2761 // Use in a phi is considered a use in the associated predecessor block
2762 use_c = use->in(0)->in(j);
2763 }
2764 set_ctrl(n_clone, use_c);
2765 assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
2766 get_loop(use_c)->_body.push(n_clone);
2767 _igvn.register_new_node_with_optimizer(n_clone);
2768 #ifndef PRODUCT
2769 if (TracePartialPeeling) {
2770 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
2771 }
2772 #endif
2773 }
2774 return cloned;
2775 }
2776
2777
2778 //------------------------------ clone_for_special_use_inside_loop -------------------------------------
2779 // clone "n" for special uses that are in the not_peeled region.
2780 // If these def-uses occur in separate blocks, the code generator
2781 // marks the method as not compilable. For example, if a "BoolNode"
2782 // is in a different basic block than the "IfNode" that uses it, then
2783 // the compilation is aborted in the code generator.
2784 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
2785 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
2786 if (n->is_Phi() || n->is_Load()) {
2787 return;
2788 }
2789 assert(worklist.size() == 0, "should be empty");
2790 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2791 Node* use = n->fast_out(j);
2792 if ( not_peel.test(use->_idx) &&
2793 (use->is_If() || use->is_CMove() || use->is_Bool()) &&
2794 use->in(1) == n) {
2795 worklist.push(use);
2796 }
2797 }
2798 if (worklist.size() > 0) {
2799 // clone "n" and insert it between inputs of "n" and the use
2800 Node* n_clone = n->clone();
2801 loop->_body.push(n_clone);
2802 _igvn.register_new_node_with_optimizer(n_clone);
2803 set_ctrl(n_clone, get_ctrl(n));
2804 sink_list.push(n_clone);
2805 not_peel <<= n_clone->_idx; // add n_clone to not_peel set.
2806 #ifndef PRODUCT
2807 if (TracePartialPeeling) {
2808 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
2809 }
2810 #endif
2811 while( worklist.size() ) {
2812 Node *use = worklist.pop();
2813 _igvn.rehash_node_delayed(use);
2814 for (uint j = 1; j < use->req(); j++) {
2815 if (use->in(j) == n) {
2816 use->set_req(j, n_clone);
2817 }
2818 }
2819 }
2820 }
2821 }
2822
2823
2824 //------------------------------ insert_phi_for_loop -------------------------------------
2825 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
2826 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
2827 Node *phi = PhiNode::make(lp, back_edge_val);
2828 phi->set_req(LoopNode::EntryControl, lp_entry_val);
2829 // Use existing phi if it already exists
2830 Node *hit = _igvn.hash_find_insert(phi);
2831 if( hit == NULL ) {
2832 _igvn.register_new_node_with_optimizer(phi);
2833 set_ctrl(phi, lp);
2834 } else {
2835 // Remove the new phi from the graph and use the hit
2836 _igvn.remove_dead_node(phi);
2837 phi = hit;
2838 }
2839 _igvn.replace_input_of(use, idx, phi);
2840 }
2841
2842 #ifdef ASSERT
2843 //------------------------------ is_valid_loop_partition -------------------------------------
2844 // Validate the loop partition sets: peel and not_peel
2845 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
2846 VectorSet& not_peel ) {
2847 uint i;
2848 // Check that peel_list entries are in the peel set
2849 for (i = 0; i < peel_list.size(); i++) {
2850 if (!peel.test(peel_list.at(i)->_idx)) {
2851 return false;
2852 }
2853 }
2854 // Check at loop members are in one of peel set or not_peel set
2855 for (i = 0; i < loop->_body.size(); i++ ) {
2856 Node *def = loop->_body.at(i);
2857 uint di = def->_idx;
2858 // Check that peel set elements are in peel_list
2859 if (peel.test(di)) {
2860 if (not_peel.test(di)) {
2861 return false;
2862 }
2863 // Must be in peel_list also
2864 bool found = false;
2865 for (uint j = 0; j < peel_list.size(); j++) {
2866 if (peel_list.at(j)->_idx == di) {
2867 found = true;
2868 break;
2869 }
2870 }
2871 if (!found) {
2872 return false;
2873 }
2874 } else if (not_peel.test(di)) {
2875 if (peel.test(di)) {
2876 return false;
2877 }
2878 } else {
2879 return false;
2880 }
2881 }
2882 return true;
2883 }
2884
2885 //------------------------------ is_valid_clone_loop_exit_use -------------------------------------
2886 // Ensure a use outside of loop is of the right form
2887 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
2888 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2889 return (use->is_Phi() &&
2890 use_c->is_Region() && use_c->req() == 3 &&
2891 (use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
2892 use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
2893 use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
2894 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
2895 }
2896
2897 //------------------------------ is_valid_clone_loop_form -------------------------------------
2898 // Ensure that all uses outside of loop are of the right form
2899 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
2900 uint orig_exit_idx, uint clone_exit_idx) {
2901 uint len = peel_list.size();
2902 for (uint i = 0; i < len; i++) {
2903 Node *def = peel_list.at(i);
2904
2905 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2906 Node *use = def->fast_out(j);
2907 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2908 if (!loop->is_member(get_loop(use_c))) {
2909 // use is not in the loop, check for correct structure
2910 if (use->in(0) == def) {
2911 // Okay
2912 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
2913 return false;
2914 }
2915 }
2916 }
2917 }
2918 return true;
2919 }
2920 #endif
2921
2922 //------------------------------ partial_peel -------------------------------------
2923 // Partially peel (aka loop rotation) the top portion of a loop (called
2924 // the peel section below) by cloning it and placing one copy just before
2925 // the new loop head and the other copy at the bottom of the new loop.
2926 //
2927 // before after where it came from
2928 //
2929 // stmt1 stmt1
2930 // loop: stmt2 clone
2931 // stmt2 if condA goto exitA clone
2932 // if condA goto exitA new_loop: new
2933 // stmt3 stmt3 clone
2934 // if !condB goto loop if condB goto exitB clone
2935 // exitB: stmt2 orig
2936 // stmt4 if !condA goto new_loop orig
2937 // exitA: goto exitA
2938 // exitB:
2939 // stmt4
2940 // exitA:
2941 //
2942 // Step 1: find the cut point: an exit test on probable
2943 // induction variable.
2944 // Step 2: schedule (with cloning) operations in the peel
2945 // section that can be executed after the cut into
2946 // the section that is not peeled. This may need
2947 // to clone operations into exit blocks. For
2948 // instance, a reference to A[i] in the not-peel
2949 // section and a reference to B[i] in an exit block
2950 // may cause a left-shift of i by 2 to be placed
2951 // in the peel block. This step will clone the left
2952 // shift into the exit block and sink the left shift
2953 // from the peel to the not-peel section.
2954 // Step 3: clone the loop, retarget the control, and insert
2955 // phis for values that are live across the new loop
2956 // head. This is very dependent on the graph structure
2957 // from clone_loop. It creates region nodes for
2958 // exit control and associated phi nodes for values
2959 // flow out of the loop through that exit. The region
2960 // node is dominated by the clone's control projection.
2961 // So the clone's peel section is placed before the
2962 // new loop head, and the clone's not-peel section is
2963 // forms the top part of the new loop. The original
2964 // peel section forms the tail of the new loop.
2965 // Step 4: update the dominator tree and recompute the
2966 // dominator depth.
2967 //
2968 // orig
2969 //
2970 // stmt1
2971 // |
2972 // v
2973 // loop predicate
2974 // |
2975 // v
2976 // loop<----+
2977 // | |
2978 // stmt2 |
2979 // | |
2980 // v |
2981 // ifA |
2982 // / | |
2983 // v v |
2984 // false true ^ <-- last_peel
2985 // / | |
2986 // / ===|==cut |
2987 // / stmt3 | <-- first_not_peel
2988 // / | |
2989 // | v |
2990 // v ifB |
2991 // exitA: / \ |
2992 // / \ |
2993 // v v |
2994 // false true |
2995 // / \ |
2996 // / ----+
2997 // |
2998 // v
2999 // exitB:
3000 // stmt4
3001 //
3002 //
3003 // after clone loop
3004 //
3005 // stmt1
3006 // |
3007 // v
3008 // loop predicate
3009 // / \
3010 // clone / \ orig
3011 // / \
3012 // / \
3013 // v v
3014 // +---->loop loop<----+
3015 // | | | |
3016 // | stmt2 stmt2 |
3017 // | | | |
3018 // | v v |
3019 // | ifA ifA |
3020 // | | \ / | |
3021 // | v v v v |
3022 // ^ true false false true ^ <-- last_peel
3023 // | | ^ \ / | |
3024 // | cut==|== \ \ / ===|==cut |
3025 // | stmt3 \ \ / stmt3 | <-- first_not_peel
3026 // | | dom | | | |
3027 // | v \ 1v v2 v |
3028 // | ifB regionA ifB |
3029 // | / \ | / \ |
3030 // | / \ v / \ |
3031 // | v v exitA: v v |
3032 // | true false false true |
3033 // | / ^ \ / \ |
3034 // +---- \ \ / ----+
3035 // dom \ /
3036 // \ 1v v2
3037 // regionB
3038 // |
3039 // v
3040 // exitB:
3041 // stmt4
3042 //
3043 //
3044 // after partial peel
3045 //
3046 // stmt1
3047 // |
3048 // v
3049 // loop predicate
3050 // /
3051 // clone / orig
3052 // / TOP
3053 // / \
3054 // v v
3055 // TOP->loop loop----+
3056 // | | |
3057 // stmt2 stmt2 |
3058 // | | |
3059 // v v |
3060 // ifA ifA |
3061 // | \ / | |
3062 // v v v v |
3063 // true false false true | <-- last_peel
3064 // | ^ \ / +------|---+
3065 // +->newloop \ \ / === ==cut | |
3066 // | stmt3 \ \ / TOP | |
3067 // | | dom | | stmt3 | | <-- first_not_peel
3068 // | v \ 1v v2 v | |
3069 // | ifB regionA ifB ^ v
3070 // | / \ | / \ | |
3071 // | / \ v / \ | |
3072 // | v v exitA: v v | |
3073 // | true false false true | |
3074 // | / ^ \ / \ | |
3075 // | | \ \ / v | |
3076 // | | dom \ / TOP | |
3077 // | | \ 1v v2 | |
3078 // ^ v regionB | |
3079 // | | | | |
3080 // | | v ^ v
3081 // | | exitB: | |
3082 // | | stmt4 | |
3083 // | +------------>-----------------+ |
3084 // | |
3085 // +-----------------<---------------------+
3086 //
3087 //
3088 // final graph
3089 //
3090 // stmt1
3091 // |
3092 // v
3093 // loop predicate
3094 // |
3095 // v
3096 // stmt2 clone
3097 // |
3098 // v
3099 // ........> ifA clone
3100 // : / |
3101 // dom / |
3102 // : v v
3103 // : false true
3104 // : | |
3105 // : | v
3106 // : | newloop<-----+
3107 // : | | |
3108 // : | stmt3 clone |
3109 // : | | |
3110 // : | v |
3111 // : | ifB |
3112 // : | / \ |
3113 // : | v v |
3114 // : | false true |
3115 // : | | | |
3116 // : | v stmt2 |
3117 // : | exitB: | |
3118 // : | stmt4 v |
3119 // : | ifA orig |
3120 // : | / \ |
3121 // : | / \ |
3122 // : | v v |
3123 // : | false true |
3124 // : | / \ |
3125 // : v v -----+
3126 // RegionA
3127 // |
3128 // v
3129 // exitA
3130 //
3131 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
3132
3133 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only");
3134 if (!loop->_head->is_Loop()) {
3135 return false;
3136 }
3137 LoopNode *head = loop->_head->as_Loop();
3138
3139 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
3140 return false;
3141 }
3142
3143 // Check for complex exit control
3144 for (uint ii = 0; ii < loop->_body.size(); ii++) {
3145 Node *n = loop->_body.at(ii);
3146 int opc = n->Opcode();
3147 if (n->is_Call() ||
3148 opc == Op_Catch ||
3149 opc == Op_CatchProj ||
3150 opc == Op_Jump ||
3151 opc == Op_JumpProj) {
3152 #ifndef PRODUCT
3153 if (TracePartialPeeling) {
3154 tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
3155 }
3156 #endif
3157 return false;
3158 }
3159 }
3160
3161 int dd = dom_depth(head);
3162
3163 // Step 1: find cut point
3164
3165 // Walk up dominators to loop head looking for first loop exit
3166 // which is executed on every path thru loop.
3167 IfNode *peel_if = NULL;
3168 IfNode *peel_if_cmpu = NULL;
3169
3170 Node *iff = loop->tail();
3171 while (iff != head) {
3172 if (iff->is_If()) {
3173 Node *ctrl = get_ctrl(iff->in(1));
3174 if (ctrl->is_top()) return false; // Dead test on live IF.
3175 // If loop-varying exit-test, check for induction variable
3176 if (loop->is_member(get_loop(ctrl)) &&
3177 loop->is_loop_exit(iff) &&
3178 is_possible_iv_test(iff)) {
3179 Node* cmp = iff->in(1)->in(1);
3180 if (cmp->Opcode() == Op_CmpI) {
3181 peel_if = iff->as_If();
3182 } else {
3183 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
3184 peel_if_cmpu = iff->as_If();
3185 }
3186 }
3187 }
3188 iff = idom(iff);
3189 }
3190
3191 // Prefer signed compare over unsigned compare.
3192 IfNode* new_peel_if = NULL;
3193 if (peel_if == NULL) {
3194 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) {
3195 return false; // No peel point found
3196 }
3197 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
3198 if (new_peel_if == NULL) {
3199 return false; // No peel point found
3200 }
3201 peel_if = new_peel_if;
3202 }
3203 Node* last_peel = stay_in_loop(peel_if, loop);
3204 Node* first_not_peeled = stay_in_loop(last_peel, loop);
3205 if (first_not_peeled == NULL || first_not_peeled == head) {
3206 return false;
3207 }
3208
3209 #ifndef PRODUCT
3210 if (TraceLoopOpts) {
3211 tty->print("PartialPeel ");
3212 loop->dump_head();
3213 }
3214
3215 if (TracePartialPeeling) {
3216 tty->print_cr("before partial peel one iteration");
3217 Node_List wl;
3218 Node* t = head->in(2);
3219 while (true) {
3220 wl.push(t);
3221 if (t == head) break;
3222 t = idom(t);
3223 }
3224 while (wl.size() > 0) {
3225 Node* tt = wl.pop();
3226 tt->dump();
3227 if (tt == last_peel) tty->print_cr("-- cut --");
3228 }
3229 }
3230 #endif
3231 ResourceArea *area = Thread::current()->resource_area();
3232 VectorSet peel(area);
3233 VectorSet not_peel(area);
3234 Node_List peel_list(area);
3235 Node_List worklist(area);
3236 Node_List sink_list(area);
3237
3238 if (!may_require_nodes(loop->est_loop_clone_sz(2))) {
3239 return false;
3240 }
3241
3242 // Set of cfg nodes to peel are those that are executable from
3243 // the head through last_peel.
3244 assert(worklist.size() == 0, "should be empty");
3245 worklist.push(head);
3246 peel.set(head->_idx);
3247 while (worklist.size() > 0) {
3248 Node *n = worklist.pop();
3249 if (n != last_peel) {
3250 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3251 Node* use = n->fast_out(j);
3252 if (use->is_CFG() &&
3253 loop->is_member(get_loop(use)) &&
3254 !peel.test_set(use->_idx)) {
3255 worklist.push(use);
3256 }
3257 }
3258 }
3259 }
3260
3261 // Set of non-cfg nodes to peel are those that are control
3262 // dependent on the cfg nodes.
3263 uint i;
3264 for(i = 0; i < loop->_body.size(); i++ ) {
3265 Node *n = loop->_body.at(i);
3266 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
3267 if (peel.test(n_c->_idx)) {
3268 peel.set(n->_idx);
3269 } else {
3270 not_peel.set(n->_idx);
3271 }
3272 }
3273
3274 // Step 2: move operations from the peeled section down into the
3275 // not-peeled section
3276
3277 // Get a post order schedule of nodes in the peel region
3278 // Result in right-most operand.
3279 scheduled_nodelist(loop, peel, peel_list );
3280
3281 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
3282
3283 // For future check for too many new phis
3284 uint old_phi_cnt = 0;
3285 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
3286 Node* use = head->fast_out(j);
3287 if (use->is_Phi()) old_phi_cnt++;
3288 }
3289
3290 #ifndef PRODUCT
3291 if (TracePartialPeeling) {
3292 tty->print_cr("\npeeled list");
3293 }
3294 #endif
3295
3296 // Evacuate nodes in peel region into the not_peeled region if possible
3297 uint new_phi_cnt = 0;
3298 uint cloned_for_outside_use = 0;
3299 for (i = 0; i < peel_list.size();) {
3300 Node* n = peel_list.at(i);
3301 #ifndef PRODUCT
3302 if (TracePartialPeeling) n->dump();
3303 #endif
3304 bool incr = true;
3305 if ( !n->is_CFG() ) {
3306
3307 if ( has_use_in_set(n, not_peel) ) {
3308
3309 // If not used internal to the peeled region,
3310 // move "n" from peeled to not_peeled region.
3311
3312 if ( !has_use_internal_to_set(n, peel, loop) ) {
3313
3314 // if not pinned and not a load (which maybe anti-dependent on a store)
3315 // and not a CMove (Matcher expects only bool->cmove).
3316 if (n->in(0) == NULL && !n->is_Load() && !n->is_CMove()) {
3317 cloned_for_outside_use += clone_for_use_outside_loop( loop, n, worklist );
3318 sink_list.push(n);
3319 peel >>= n->_idx; // delete n from peel set.
3320 not_peel <<= n->_idx; // add n to not_peel set.
3321 peel_list.remove(i);
3322 incr = false;
3323 #ifndef PRODUCT
3324 if (TracePartialPeeling) {
3325 tty->print_cr("sink to not_peeled region: %d newbb: %d",
3326 n->_idx, get_ctrl(n)->_idx);
3327 }
3328 #endif
3329 }
3330 } else {
3331 // Otherwise check for special def-use cases that span
3332 // the peel/not_peel boundary such as bool->if
3333 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist );
3334 new_phi_cnt++;
3335 }
3336 }
3337 }
3338 if (incr) i++;
3339 }
3340
3341 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) {
3342 #ifndef PRODUCT
3343 if (TracePartialPeeling) {
3344 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c",
3345 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F');
3346 }
3347 #endif
3348 if (new_peel_if != NULL) {
3349 remove_cmpi_loop_exit(new_peel_if, loop);
3350 }
3351 // Inhibit more partial peeling on this loop
3352 assert(!head->is_partial_peel_loop(), "not partial peeled");
3353 head->mark_partial_peel_failed();
3354 if (cloned_for_outside_use > 0) {
3355 // Terminate this round of loop opts because
3356 // the graph outside this loop was changed.
3357 C->set_major_progress();
3358 return true;
3359 }
3360 return false;
3361 }
3362
3363 // Step 3: clone loop, retarget control, and insert new phis
3364
3365 // Create new loop head for new phis and to hang
3366 // the nodes being moved (sinked) from the peel region.
3367 LoopNode* new_head = new LoopNode(last_peel, last_peel);
3368 new_head->set_unswitch_count(head->unswitch_count()); // Preserve
3369 _igvn.register_new_node_with_optimizer(new_head);
3370 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
3371 _igvn.replace_input_of(first_not_peeled, 0, new_head);
3372 set_loop(new_head, loop);
3373 loop->_body.push(new_head);
3374 not_peel.set(new_head->_idx);
3375 set_idom(new_head, last_peel, dom_depth(first_not_peeled));
3376 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
3377
3378 while (sink_list.size() > 0) {
3379 Node* n = sink_list.pop();
3380 set_ctrl(n, new_head);
3381 }
3382
3383 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
3384
3385 clone_loop(loop, old_new, dd, IgnoreStripMined);
3386
3387 const uint clone_exit_idx = 1;
3388 const uint orig_exit_idx = 2;
3389 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop");
3390
3391 Node* head_clone = old_new[head->_idx];
3392 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop();
3393 Node* orig_tail_clone = head_clone->in(2);
3394
3395 // Add phi if "def" node is in peel set and "use" is not
3396
3397 for(i = 0; i < peel_list.size(); i++ ) {
3398 Node *def = peel_list.at(i);
3399 if (!def->is_CFG()) {
3400 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
3401 Node *use = def->fast_out(j);
3402 if (has_node(use) && use->in(0) != C->top() &&
3403 (!peel.test(use->_idx) ||
3404 (use->is_Phi() && use->in(0) == head)) ) {
3405 worklist.push(use);
3406 }
3407 }
3408 while( worklist.size() ) {
3409 Node *use = worklist.pop();
3410 for (uint j = 1; j < use->req(); j++) {
3411 Node* n = use->in(j);
3412 if (n == def) {
3413
3414 // "def" is in peel set, "use" is not in peel set
3415 // or "use" is in the entry boundary (a phi) of the peel set
3416
3417 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
3418
3419 if ( loop->is_member(get_loop( use_c )) ) {
3420 // use is in loop
3421 if (old_new[use->_idx] != NULL) { // null for dead code
3422 Node* use_clone = old_new[use->_idx];
3423 _igvn.replace_input_of(use, j, C->top());
3424 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
3425 }
3426 } else {
3427 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
3428 // use is not in the loop, check if the live range includes the cut
3429 Node* lp_if = use_c->in(orig_exit_idx)->in(0);
3430 if (not_peel.test(lp_if->_idx)) {
3431 assert(j == orig_exit_idx, "use from original loop");
3432 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
3433 }
3434 }
3435 }
3436 }
3437 }
3438 }
3439 }
3440
3441 // Step 3b: retarget control
3442
3443 // Redirect control to the new loop head if a cloned node in
3444 // the not_peeled region has control that points into the peeled region.
3445 // This necessary because the cloned peeled region will be outside
3446 // the loop.
3447 // from to
3448 // cloned-peeled <---+
3449 // new_head_clone: | <--+
3450 // cloned-not_peeled in(0) in(0)
3451 // orig-peeled
3452
3453 for(i = 0; i < loop->_body.size(); i++ ) {
3454 Node *n = loop->_body.at(i);
3455 if (!n->is_CFG() && n->in(0) != NULL &&
3456 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
3457 Node* n_clone = old_new[n->_idx];
3458 _igvn.replace_input_of(n_clone, 0, new_head_clone);
3459 }
3460 }
3461
3462 // Backedge of the surviving new_head (the clone) is original last_peel
3463 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel);
3464
3465 // Cut first node in original not_peel set
3466 _igvn.rehash_node_delayed(new_head); // Multiple edge updates:
3467 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of
3468 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls
3469
3470 // Copy head_clone back-branch info to original head
3471 // and remove original head's loop entry and
3472 // clone head's back-branch
3473 _igvn.rehash_node_delayed(head); // Multiple edge updates
3474 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl));
3475 head->set_req(LoopNode::LoopBackControl, C->top());
3476 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top());
3477
3478 // Similarly modify the phis
3479 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
3480 Node* use = head->fast_out(k);
3481 if (use->is_Phi() && use->outcnt() > 0) {
3482 Node* use_clone = old_new[use->_idx];
3483 _igvn.rehash_node_delayed(use); // Multiple edge updates
3484 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl));
3485 use->set_req(LoopNode::LoopBackControl, C->top());
3486 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top());
3487 }
3488 }
3489
3490 // Step 4: update dominator tree and dominator depth
3491
3492 set_idom(head, orig_tail_clone, dd);
3493 recompute_dom_depth();
3494
3495 // Inhibit more partial peeling on this loop
3496 new_head_clone->set_partial_peel_loop();
3497 C->set_major_progress();
3498 loop->record_for_igvn();
3499
3500 #ifndef PRODUCT
3501 if (TracePartialPeeling) {
3502 tty->print_cr("\nafter partial peel one iteration");
3503 Node_List wl(area);
3504 Node* t = last_peel;
3505 while (true) {
3506 wl.push(t);
3507 if (t == head_clone) break;
3508 t = idom(t);
3509 }
3510 while (wl.size() > 0) {
3511 Node* tt = wl.pop();
3512 if (tt == head) tty->print_cr("orig head");
3513 else if (tt == new_head_clone) tty->print_cr("new head");
3514 else if (tt == head_clone) tty->print_cr("clone head");
3515 tt->dump();
3516 }
3517 }
3518 #endif
3519 return true;
3520 }
3521
3522 //------------------------------reorg_offsets----------------------------------
3523 // Reorganize offset computations to lower register pressure. Mostly
3524 // prevent loop-fallout uses of the pre-incremented trip counter (which are
3525 // then alive with the post-incremented trip counter forcing an extra
3526 // register move)
3527 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) {
3528 // Perform it only for canonical counted loops.
3529 // Loop's shape could be messed up by iteration_split_impl.
3530 if (!loop->_head->is_CountedLoop())
3531 return;
3532 if (!loop->_head->as_Loop()->is_valid_counted_loop())
3533 return;
3534
3535 CountedLoopNode *cl = loop->_head->as_CountedLoop();
3536 CountedLoopEndNode *cle = cl->loopexit();
3537 Node *exit = cle->proj_out(false);
3538 Node *phi = cl->phi();
3539
3540 // Check for the special case when using the pre-incremented trip-counter on
3541 // the fall-out path (forces the pre-incremented and post-incremented trip
3542 // counter to be live at the same time). Fix this by adjusting to use the
3543 // post-increment trip counter.
3544
3545 bool progress = true;
3546 while (progress) {
3547 progress = false;
3548 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
3549 Node* use = phi->fast_out(i); // User of trip-counter
3550 if (!has_ctrl(use)) continue;
3551 Node *u_ctrl = get_ctrl(use);
3552 if (use->is_Phi()) {
3553 u_ctrl = NULL;
3554 for (uint j = 1; j < use->req(); j++)
3555 if (use->in(j) == phi)
3556 u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j));
3557 }
3558 IdealLoopTree *u_loop = get_loop(u_ctrl);
3559 // Look for loop-invariant use
3560 if (u_loop == loop) continue;
3561 if (loop->is_member(u_loop)) continue;
3562 // Check that use is live out the bottom. Assuming the trip-counter
3563 // update is right at the bottom, uses of of the loop middle are ok.
3564 if (dom_lca(exit, u_ctrl) != exit) continue;
3565 // Hit! Refactor use to use the post-incremented tripcounter.
3566 // Compute a post-increment tripcounter.
3567 Node* c = exit;
3568 if (cl->is_strip_mined()) {
3569 IdealLoopTree* outer_loop = get_loop(cl->outer_loop());
3570 if (!outer_loop->is_member(u_loop)) {
3571 c = cl->outer_loop_exit();
3572 }
3573 }
3574 Node *opaq = new Opaque2Node(C, cle->incr());
3575 register_new_node(opaq, c);
3576 Node *neg_stride = _igvn.intcon(-cle->stride_con());
3577 set_ctrl(neg_stride, C->root());
3578 Node *post = new AddINode(opaq, neg_stride);
3579 register_new_node(post, c);
3580 _igvn.rehash_node_delayed(use);
3581 for (uint j = 1; j < use->req(); j++) {
3582 if (use->in(j) == phi)
3583 use->set_req(j, post);
3584 }
3585 // Since DU info changed, rerun loop
3586 progress = true;
3587 break;
3588 }
3589 }
3590
3591 }