1 /*
2 * Copyright (c) 1999, 2018, 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 #if INCLUDE_SHENANDOAHGC
1093 if (m->is_ShenandoahBarrier() && m->has_out_with(Op_FastLock)) {
1094 return false;
1095 }
1096 #endif
1097 #ifdef _LP64
1098 if (m->Opcode() == Op_ConvI2L)
1099 return false;
1100 if (m->is_CastII() && m->isa_CastII()->has_range_check()) {
1101 return false;
1102 }
1103 #endif
1104 }
1105 }
1106 }
1107 return true;
1108 }
1109
1110
1111 //------------------------------place_near_use---------------------------------
1112 // Place some computation next to use but not inside inner loops.
1113 // For inner loop uses move it to the preheader area.
1114 Node *PhaseIdealLoop::place_near_use(Node *useblock) const {
1115 IdealLoopTree *u_loop = get_loop( useblock );
1116 if (u_loop->_irreducible) {
1117 return useblock;
1118 }
1119 if (u_loop->_child) {
1120 if (useblock == u_loop->_head && u_loop->_head->is_OuterStripMinedLoop()) {
1121 return u_loop->_head->in(LoopNode::EntryControl);
1122 }
1123 return useblock;
1124 }
1125 return u_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl);
1126 }
1127
1128
1129 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) {
1130 if (!n->is_If() || n->is_CountedLoopEnd()) {
1131 return false;
1132 }
1133 if (!n->in(0)->is_Region()) {
1134 return false;
1135 }
1136 Node* region = n->in(0);
1137 Node* dom = idom(region);
1138 if (!dom->is_If() || dom->in(1) != n->in(1)) {
1139 return false;
1140 }
1141 IfNode* dom_if = dom->as_If();
1142 Node* proj_true = dom_if->proj_out(1);
1143 Node* proj_false = dom_if->proj_out(0);
1144
1145 for (uint i = 1; i < region->req(); i++) {
1146 if (is_dominator(proj_true, region->in(i))) {
1147 continue;
1148 }
1149 if (is_dominator(proj_false, region->in(i))) {
1150 continue;
1151 }
1152 return false;
1153 }
1154
1155 return true;
1156 }
1157
1158
1159 bool PhaseIdealLoop::can_split_if(Node* n_ctrl) {
1160 if (must_throttle_split_if()) {
1161 return false;
1162 }
1163
1164 // Do not do 'split-if' if irreducible loops are present.
1165 if (_has_irreducible_loops) {
1166 return false;
1167 }
1168
1169 if (merge_point_too_heavy(C, n_ctrl)) {
1170 return false;
1171 }
1172
1173 // Do not do 'split-if' if some paths are dead. First do dead code
1174 // elimination and then see if its still profitable.
1175 for (uint i = 1; i < n_ctrl->req(); i++) {
1176 if (n_ctrl->in(i) == C->top()) {
1177 return false;
1178 }
1179 }
1180
1181 // If trying to do a 'Split-If' at the loop head, it is only
1182 // profitable if the cmp folds up on BOTH paths. Otherwise we
1183 // risk peeling a loop forever.
1184
1185 // CNC - Disabled for now. Requires careful handling of loop
1186 // body selection for the cloned code. Also, make sure we check
1187 // for any input path not being in the same loop as n_ctrl. For
1188 // irreducible loops we cannot check for 'n_ctrl->is_Loop()'
1189 // because the alternative loop entry points won't be converted
1190 // into LoopNodes.
1191 IdealLoopTree *n_loop = get_loop(n_ctrl);
1192 for (uint j = 1; j < n_ctrl->req(); j++) {
1193 if (get_loop(n_ctrl->in(j)) != n_loop) {
1194 return false;
1195 }
1196 }
1197
1198 // Check for safety of the merge point.
1199 if (!merge_point_safe(n_ctrl)) {
1200 return false;
1201 }
1202
1203 return true;
1204 }
1205
1206 //------------------------------split_if_with_blocks_post----------------------
1207 // Do the real work in a non-recursive function. CFG hackery wants to be
1208 // in the post-order, so it can dirty the I-DOM info and not use the dirtied
1209 // info.
1210 void PhaseIdealLoop::split_if_with_blocks_post(Node *n, bool last_round) {
1211
1212 // Cloning Cmp through Phi's involves the split-if transform.
1213 // FastLock is not used by an If
1214 if (n->is_Cmp() && !n->is_FastLock() && !last_round) {
1215 Node *n_ctrl = get_ctrl(n);
1216 // Determine if the Node has inputs from some local Phi.
1217 // Returns the block to clone thru.
1218 Node *n_blk = has_local_phi_input(n);
1219 if (n_blk != n_ctrl) {
1220 return;
1221 }
1222
1223 if (!can_split_if(n_ctrl)) {
1224 return;
1225 }
1226
1227 if (n->outcnt() != 1) {
1228 return; // Multiple bool's from 1 compare?
1229 }
1230 Node *bol = n->unique_out();
1231 assert(bol->is_Bool(), "expect a bool here");
1232 if (bol->outcnt() != 1) {
1233 return;// Multiple branches from 1 compare?
1234 }
1235 Node *iff = bol->unique_out();
1236
1237 // Check some safety conditions
1238 if (iff->is_If()) { // Classic split-if?
1239 if (iff->in(0) != n_ctrl) {
1240 return; // Compare must be in same blk as if
1241 }
1242 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE
1243 // Can't split CMove with different control edge.
1244 if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) {
1245 return;
1246 }
1247 if (get_ctrl(iff->in(2)) == n_ctrl ||
1248 get_ctrl(iff->in(3)) == n_ctrl) {
1249 return; // Inputs not yet split-up
1250 }
1251 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) {
1252 return; // Loop-invar test gates loop-varying CMOVE
1253 }
1254 } else {
1255 return; // some other kind of node, such as an Allocate
1256 }
1257
1258 // When is split-if profitable? Every 'win' on means some control flow
1259 // goes dead, so it's almost always a win.
1260 int policy = 0;
1261 // Split compare 'n' through the merge point if it is profitable
1262 Node *phi = split_thru_phi( n, n_ctrl, policy);
1263 if (!phi) {
1264 return;
1265 }
1266
1267 // Found a Phi to split thru!
1268 // Replace 'n' with the new phi
1269 _igvn.replace_node(n, phi);
1270
1271 // Now split the bool up thru the phi
1272 Node *bolphi = split_thru_phi(bol, n_ctrl, -1);
1273 guarantee(bolphi != NULL, "null boolean phi node");
1274
1275 _igvn.replace_node(bol, bolphi);
1276 assert(iff->in(1) == bolphi, "");
1277
1278 if (bolphi->Value(&_igvn)->singleton()) {
1279 return;
1280 }
1281
1282 // Conditional-move? Must split up now
1283 if (!iff->is_If()) {
1284 Node *cmovphi = split_thru_phi(iff, n_ctrl, -1);
1285 _igvn.replace_node(iff, cmovphi);
1286 return;
1287 }
1288
1289 // Now split the IF
1290 do_split_if(iff);
1291 return;
1292 }
1293
1294 // Two identical ifs back to back can be merged
1295 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) {
1296 Node *n_ctrl = n->in(0);
1297 PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1));
1298 IfNode* dom_if = idom(n_ctrl)->as_If();
1299 Node* proj_true = dom_if->proj_out(1);
1300 Node* proj_false = dom_if->proj_out(0);
1301 Node* con_true = _igvn.makecon(TypeInt::ONE);
1302 Node* con_false = _igvn.makecon(TypeInt::ZERO);
1303
1304 for (uint i = 1; i < n_ctrl->req(); i++) {
1305 if (is_dominator(proj_true, n_ctrl->in(i))) {
1306 bolphi->init_req(i, con_true);
1307 } else {
1308 assert(is_dominator(proj_false, n_ctrl->in(i)), "bad if");
1309 bolphi->init_req(i, con_false);
1310 }
1311 }
1312 register_new_node(bolphi, n_ctrl);
1313 _igvn.replace_input_of(n, 1, bolphi);
1314
1315 // Now split the IF
1316 do_split_if(n);
1317 return;
1318 }
1319
1320 // Check for an IF ready to split; one that has its
1321 // condition codes input coming from a Phi at the block start.
1322 int n_op = n->Opcode();
1323
1324 // Check for an IF being dominated by another IF same test
1325 if (n_op == Op_If ||
1326 n_op == Op_RangeCheck) {
1327 Node *bol = n->in(1);
1328 uint max = bol->outcnt();
1329 // Check for same test used more than once?
1330 if (max > 1 && bol->is_Bool()) {
1331 // Search up IDOMs to see if this IF is dominated.
1332 Node *cutoff = get_ctrl(bol);
1333
1334 // Now search up IDOMs till cutoff, looking for a dominating test
1335 Node *prevdom = n;
1336 Node *dom = idom(prevdom);
1337 while (dom != cutoff) {
1338 if (dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom) {
1339 // Replace the dominated test with an obvious true or false.
1340 // Place it on the IGVN worklist for later cleanup.
1341 C->set_major_progress();
1342 dominated_by(prevdom, n, false, true);
1343 #ifndef PRODUCT
1344 if( VerifyLoopOptimizations ) verify();
1345 #endif
1346 return;
1347 }
1348 prevdom = dom;
1349 dom = idom(prevdom);
1350 }
1351 }
1352 }
1353
1354 // See if a shared loop-varying computation has no loop-varying uses.
1355 // Happens if something is only used for JVM state in uncommon trap exits,
1356 // like various versions of induction variable+offset. Clone the
1357 // computation per usage to allow it to sink out of the loop.
1358 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about)
1359 Node *n_ctrl = get_ctrl(n);
1360 IdealLoopTree *n_loop = get_loop(n_ctrl);
1361 if( n_loop != _ltree_root ) {
1362 DUIterator_Fast imax, i = n->fast_outs(imax);
1363 for (; i < imax; i++) {
1364 Node* u = n->fast_out(i);
1365 if( !has_ctrl(u) ) break; // Found control user
1366 IdealLoopTree *u_loop = get_loop(get_ctrl(u));
1367 if( u_loop == n_loop ) break; // Found loop-varying use
1368 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop
1369 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003
1370 }
1371 bool did_break = (i < imax); // Did we break out of the previous loop?
1372 if (!did_break && n->outcnt() > 1) { // All uses in outer loops!
1373 Node *late_load_ctrl = NULL;
1374 if (n->is_Load()) {
1375 // If n is a load, get and save the result from get_late_ctrl(),
1376 // to be later used in calculating the control for n's clones.
1377 clear_dom_lca_tags();
1378 late_load_ctrl = get_late_ctrl(n, n_ctrl);
1379 }
1380 // If n is a load, and the late control is the same as the current
1381 // control, then the cloning of n is a pointless exercise, because
1382 // GVN will ensure that we end up where we started.
1383 if (!n->is_Load() || late_load_ctrl != n_ctrl) {
1384 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1385 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) {
1386 Node *u = n->last_out(j); // Clone private computation per use
1387 _igvn.rehash_node_delayed(u);
1388 Node *x = n->clone(); // Clone computation
1389 Node *x_ctrl = NULL;
1390 if( u->is_Phi() ) {
1391 // Replace all uses of normal nodes. Replace Phi uses
1392 // individually, so the separate Nodes can sink down
1393 // different paths.
1394 uint k = 1;
1395 while( u->in(k) != n ) k++;
1396 u->set_req( k, x );
1397 // x goes next to Phi input path
1398 x_ctrl = u->in(0)->in(k);
1399 --j;
1400 } else { // Normal use
1401 // Replace all uses
1402 for( uint k = 0; k < u->req(); k++ ) {
1403 if( u->in(k) == n ) {
1404 u->set_req( k, x );
1405 --j;
1406 }
1407 }
1408 x_ctrl = get_ctrl(u);
1409 }
1410
1411 // Find control for 'x' next to use but not inside inner loops.
1412 // For inner loop uses get the preheader area.
1413 x_ctrl = place_near_use(x_ctrl);
1414
1415 if (bs->sink_node(this, n, x, x_ctrl, n_ctrl)) {
1416 continue;
1417 }
1418
1419 if (n->is_Load()) {
1420 // For loads, add a control edge to a CFG node outside of the loop
1421 // to force them to not combine and return back inside the loop
1422 // during GVN optimization (4641526).
1423 //
1424 // Because we are setting the actual control input, factor in
1425 // the result from get_late_ctrl() so we respect any
1426 // anti-dependences. (6233005).
1427 x_ctrl = dom_lca(late_load_ctrl, x_ctrl);
1428
1429 // Don't allow the control input to be a CFG splitting node.
1430 // Such nodes should only have ProjNodes as outs, e.g. IfNode
1431 // should only have IfTrueNode and IfFalseNode (4985384).
1432 x_ctrl = find_non_split_ctrl(x_ctrl);
1433 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone");
1434
1435 x->set_req(0, x_ctrl);
1436 }
1437 register_new_node(x, x_ctrl);
1438
1439 // Some institutional knowledge is needed here: 'x' is
1440 // yanked because if the optimizer runs GVN on it all the
1441 // cloned x's will common up and undo this optimization and
1442 // be forced back in the loop. This is annoying because it
1443 // makes +VerifyOpto report false-positives on progress. I
1444 // tried setting control edges on the x's to force them to
1445 // not combine, but the matching gets worried when it tries
1446 // to fold a StoreP and an AddP together (as part of an
1447 // address expression) and the AddP and StoreP have
1448 // different controls.
1449 if (!x->is_Load() && !x->is_DecodeNarrowPtr()) _igvn._worklist.yank(x);
1450 }
1451 _igvn.remove_dead_node(n);
1452 }
1453 }
1454 }
1455 }
1456
1457 try_move_store_after_loop(n);
1458
1459 // Remove multiple allocations of the same value type
1460 if (n->is_ValueType() && EliminateAllocations) {
1461 n->as_ValueType()->remove_redundant_allocations(&_igvn, this);
1462 return; // n is now dead
1463 }
1464
1465 // Check for Opaque2's who's loop has disappeared - who's input is in the
1466 // same loop nest as their output. Remove 'em, they are no longer useful.
1467 if( n_op == Op_Opaque2 &&
1468 n->in(1) != NULL &&
1469 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) {
1470 _igvn.replace_node( n, n->in(1) );
1471 }
1472
1473 #if INCLUDE_ZGC
1474 if (UseZGC) {
1475 ZBarrierSetC2::loop_optimize_gc_barrier(this, n, last_round);
1476 }
1477 #endif
1478 }
1479
1480 //------------------------------split_if_with_blocks---------------------------
1481 // Check for aggressive application of 'split-if' optimization,
1482 // using basic block level info.
1483 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack, bool last_round) {
1484 Node* root = C->root();
1485 visited.set(root->_idx); // first, mark root as visited
1486 // Do pre-visit work for root
1487 Node* n = split_if_with_blocks_pre(root);
1488 uint cnt = n->outcnt();
1489 uint i = 0;
1490
1491 while (true) {
1492 // Visit all children
1493 if (i < cnt) {
1494 Node* use = n->raw_out(i);
1495 ++i;
1496 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
1497 // Now do pre-visit work for this use
1498 use = split_if_with_blocks_pre(use);
1499 nstack.push(n, i); // Save parent and next use's index.
1500 n = use; // Process all children of current use.
1501 cnt = use->outcnt();
1502 i = 0;
1503 }
1504 }
1505 else {
1506 // All of n's children have been processed, complete post-processing.
1507 if (cnt != 0 && !n->is_Con()) {
1508 assert(has_node(n), "no dead nodes");
1509 split_if_with_blocks_post(n, last_round);
1510 }
1511 if (must_throttle_split_if()) {
1512 nstack.clear();
1513 }
1514 if (nstack.is_empty()) {
1515 // Finished all nodes on stack.
1516 break;
1517 }
1518 // Get saved parent node and next use's index. Visit the rest of uses.
1519 n = nstack.node();
1520 cnt = n->outcnt();
1521 i = nstack.index();
1522 nstack.pop();
1523 }
1524 }
1525 }
1526
1527
1528 //=============================================================================
1529 //
1530 // C L O N E A L O O P B O D Y
1531 //
1532
1533 //------------------------------clone_iff--------------------------------------
1534 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1535 // "Nearly" because all Nodes have been cloned from the original in the loop,
1536 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1537 // through the Phi recursively, and return a Bool.
1538 Node* PhaseIdealLoop::clone_iff(PhiNode *phi, IdealLoopTree *loop) {
1539
1540 // Convert this Phi into a Phi merging Bools
1541 uint i;
1542 for (i = 1; i < phi->req(); i++) {
1543 Node *b = phi->in(i);
1544 if (b->is_Phi()) {
1545 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi(), loop));
1546 } else {
1547 assert(b->is_Bool() || b->Opcode() == Op_Opaque4, "");
1548 }
1549 }
1550
1551 Node* n = phi->in(1);
1552 Node* sample_opaque = NULL;
1553 Node *sample_bool = NULL;
1554 if (n->Opcode() == Op_Opaque4) {
1555 sample_opaque = n;
1556 sample_bool = n->in(1);
1557 assert(sample_bool->is_Bool(), "wrong type");
1558 } else {
1559 sample_bool = n;
1560 }
1561 Node *sample_cmp = sample_bool->in(1);
1562
1563 // Make Phis to merge the Cmp's inputs.
1564 PhiNode *phi1 = new PhiNode(phi->in(0), Type::TOP);
1565 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP);
1566 for (i = 1; i < phi->req(); i++) {
1567 Node *n1 = sample_opaque == NULL ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1);
1568 Node *n2 = sample_opaque == NULL ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2);
1569 phi1->set_req(i, n1);
1570 phi2->set_req(i, n2);
1571 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
1572 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
1573 }
1574 // See if these Phis have been made before.
1575 // Register with optimizer
1576 Node *hit1 = _igvn.hash_find_insert(phi1);
1577 if (hit1) { // Hit, toss just made Phi
1578 _igvn.remove_dead_node(phi1); // Remove new phi
1579 assert(hit1->is_Phi(), "" );
1580 phi1 = (PhiNode*)hit1; // Use existing phi
1581 } else { // Miss
1582 _igvn.register_new_node_with_optimizer(phi1);
1583 }
1584 Node *hit2 = _igvn.hash_find_insert(phi2);
1585 if (hit2) { // Hit, toss just made Phi
1586 _igvn.remove_dead_node(phi2); // Remove new phi
1587 assert(hit2->is_Phi(), "" );
1588 phi2 = (PhiNode*)hit2; // Use existing phi
1589 } else { // Miss
1590 _igvn.register_new_node_with_optimizer(phi2);
1591 }
1592 // Register Phis with loop/block info
1593 set_ctrl(phi1, phi->in(0));
1594 set_ctrl(phi2, phi->in(0));
1595 // Make a new Cmp
1596 Node *cmp = sample_cmp->clone();
1597 cmp->set_req(1, phi1);
1598 cmp->set_req(2, phi2);
1599 _igvn.register_new_node_with_optimizer(cmp);
1600 set_ctrl(cmp, phi->in(0));
1601
1602 // Make a new Bool
1603 Node *b = sample_bool->clone();
1604 b->set_req(1,cmp);
1605 _igvn.register_new_node_with_optimizer(b);
1606 set_ctrl(b, phi->in(0));
1607
1608 if (sample_opaque != NULL) {
1609 Node* opaque = sample_opaque->clone();
1610 opaque->set_req(1, b);
1611 _igvn.register_new_node_with_optimizer(opaque);
1612 set_ctrl(opaque, phi->in(0));
1613 return opaque;
1614 }
1615
1616 assert(b->is_Bool(), "");
1617 return b;
1618 }
1619
1620 //------------------------------clone_bool-------------------------------------
1621 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1622 // "Nearly" because all Nodes have been cloned from the original in the loop,
1623 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1624 // through the Phi recursively, and return a Bool.
1625 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) {
1626 uint i;
1627 // Convert this Phi into a Phi merging Bools
1628 for( i = 1; i < phi->req(); i++ ) {
1629 Node *b = phi->in(i);
1630 if( b->is_Phi() ) {
1631 _igvn.replace_input_of(phi, i, clone_bool( b->as_Phi(), loop ));
1632 } else {
1633 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
1634 }
1635 }
1636
1637 Node *sample_cmp = phi->in(1);
1638
1639 // Make Phis to merge the Cmp's inputs.
1640 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP );
1641 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP );
1642 for( uint j = 1; j < phi->req(); j++ ) {
1643 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
1644 Node *n1, *n2;
1645 if( cmp_top->is_Cmp() ) {
1646 n1 = cmp_top->in(1);
1647 n2 = cmp_top->in(2);
1648 } else {
1649 n1 = n2 = cmp_top;
1650 }
1651 phi1->set_req( j, n1 );
1652 phi2->set_req( j, n2 );
1653 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
1654 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
1655 }
1656
1657 // See if these Phis have been made before.
1658 // Register with optimizer
1659 Node *hit1 = _igvn.hash_find_insert(phi1);
1660 if( hit1 ) { // Hit, toss just made Phi
1661 _igvn.remove_dead_node(phi1); // Remove new phi
1662 assert( hit1->is_Phi(), "" );
1663 phi1 = (PhiNode*)hit1; // Use existing phi
1664 } else { // Miss
1665 _igvn.register_new_node_with_optimizer(phi1);
1666 }
1667 Node *hit2 = _igvn.hash_find_insert(phi2);
1668 if( hit2 ) { // Hit, toss just made Phi
1669 _igvn.remove_dead_node(phi2); // Remove new phi
1670 assert( hit2->is_Phi(), "" );
1671 phi2 = (PhiNode*)hit2; // Use existing phi
1672 } else { // Miss
1673 _igvn.register_new_node_with_optimizer(phi2);
1674 }
1675 // Register Phis with loop/block info
1676 set_ctrl(phi1, phi->in(0));
1677 set_ctrl(phi2, phi->in(0));
1678 // Make a new Cmp
1679 Node *cmp = sample_cmp->clone();
1680 cmp->set_req( 1, phi1 );
1681 cmp->set_req( 2, phi2 );
1682 _igvn.register_new_node_with_optimizer(cmp);
1683 set_ctrl(cmp, phi->in(0));
1684
1685 assert( cmp->is_Cmp(), "" );
1686 return (CmpNode*)cmp;
1687 }
1688
1689 //------------------------------sink_use---------------------------------------
1690 // If 'use' was in the loop-exit block, it now needs to be sunk
1691 // below the post-loop merge point.
1692 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) {
1693 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) {
1694 set_ctrl(use, post_loop);
1695 for (DUIterator j = use->outs(); use->has_out(j); j++)
1696 sink_use(use->out(j), post_loop);
1697 }
1698 }
1699
1700 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new,
1701 IdealLoopTree* loop, IdealLoopTree* outer_loop,
1702 Node_List*& split_if_set, Node_List*& split_bool_set,
1703 Node_List*& split_cex_set, Node_List& worklist,
1704 uint new_counter, CloneLoopMode mode) {
1705 Node* nnn = old_new[old->_idx];
1706 // Copy uses to a worklist, so I can munge the def-use info
1707 // with impunity.
1708 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
1709 worklist.push(old->fast_out(j));
1710
1711 while( worklist.size() ) {
1712 Node *use = worklist.pop();
1713 if (!has_node(use)) continue; // Ignore dead nodes
1714 if (use->in(0) == C->top()) continue;
1715 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
1716 // Check for data-use outside of loop - at least one of OLD or USE
1717 // must not be a CFG node.
1718 #ifdef ASSERT
1719 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == NULL) {
1720 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint();
1721 assert(mode == ControlAroundStripMined && use == sfpt, "missed a node");
1722 }
1723 #endif
1724 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) {
1725
1726 // If the Data use is an IF, that means we have an IF outside of the
1727 // loop that is switching on a condition that is set inside of the
1728 // loop. Happens if people set a loop-exit flag; then test the flag
1729 // in the loop to break the loop, then test is again outside of the
1730 // loop to determine which way the loop exited.
1731 // Loop predicate If node connects to Bool node through Opaque1 node.
1732 if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use) || use->Opcode() == Op_Opaque4) {
1733 // Since this code is highly unlikely, we lazily build the worklist
1734 // of such Nodes to go split.
1735 if (!split_if_set) {
1736 ResourceArea *area = Thread::current()->resource_area();
1737 split_if_set = new Node_List(area);
1738 }
1739 split_if_set->push(use);
1740 }
1741 if (use->is_Bool()) {
1742 if (!split_bool_set) {
1743 ResourceArea *area = Thread::current()->resource_area();
1744 split_bool_set = new Node_List(area);
1745 }
1746 split_bool_set->push(use);
1747 }
1748 if (use->Opcode() == Op_CreateEx) {
1749 if (!split_cex_set) {
1750 ResourceArea *area = Thread::current()->resource_area();
1751 split_cex_set = new Node_List(area);
1752 }
1753 split_cex_set->push(use);
1754 }
1755
1756
1757 // Get "block" use is in
1758 uint idx = 0;
1759 while( use->in(idx) != old ) idx++;
1760 Node *prev = use->is_CFG() ? use : get_ctrl(use);
1761 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" );
1762 Node *cfg = prev->_idx >= new_counter
1763 ? prev->in(2)
1764 : idom(prev);
1765 if( use->is_Phi() ) // Phi use is in prior block
1766 cfg = prev->in(idx); // NOT in block of Phi itself
1767 if (cfg->is_top()) { // Use is dead?
1768 _igvn.replace_input_of(use, idx, C->top());
1769 continue;
1770 }
1771
1772 // If use is referenced through control edge... (idx == 0)
1773 if (mode == IgnoreStripMined && idx == 0) {
1774 LoopNode *head = loop->_head->as_Loop();
1775 if (head->is_strip_mined() && is_dominator(head->outer_loop_exit(), prev)) {
1776 // That node is outside the inner loop, leave it outside the
1777 // outer loop as well to not confuse verification code.
1778 assert(!loop->_parent->is_member(use_loop), "should be out of the outer loop");
1779 _igvn.replace_input_of(use, 0, head->outer_loop_exit());
1780 continue;
1781 }
1782 }
1783
1784 while(!outer_loop->is_member(get_loop(cfg))) {
1785 prev = cfg;
1786 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg);
1787 }
1788 // If the use occurs after merging several exits from the loop, then
1789 // old value must have dominated all those exits. Since the same old
1790 // value was used on all those exits we did not need a Phi at this
1791 // merge point. NOW we do need a Phi here. Each loop exit value
1792 // is now merged with the peeled body exit; each exit gets its own
1793 // private Phi and those Phis need to be merged here.
1794 Node *phi;
1795 if( prev->is_Region() ) {
1796 if( idx == 0 ) { // Updating control edge?
1797 phi = prev; // Just use existing control
1798 } else { // Else need a new Phi
1799 phi = PhiNode::make( prev, old );
1800 // Now recursively fix up the new uses of old!
1801 for( uint i = 1; i < prev->req(); i++ ) {
1802 worklist.push(phi); // Onto worklist once for each 'old' input
1803 }
1804 }
1805 } else {
1806 // Get new RegionNode merging old and new loop exits
1807 prev = old_new[prev->_idx];
1808 assert( prev, "just made this in step 7" );
1809 if( idx == 0) { // Updating control edge?
1810 phi = prev; // Just use existing control
1811 } else { // Else need a new Phi
1812 // Make a new Phi merging data values properly
1813 phi = PhiNode::make( prev, old );
1814 phi->set_req( 1, nnn );
1815 }
1816 }
1817 // If inserting a new Phi, check for prior hits
1818 if( idx != 0 ) {
1819 Node *hit = _igvn.hash_find_insert(phi);
1820 if( hit == NULL ) {
1821 _igvn.register_new_node_with_optimizer(phi); // Register new phi
1822 } else { // or
1823 // Remove the new phi from the graph and use the hit
1824 _igvn.remove_dead_node(phi);
1825 phi = hit; // Use existing phi
1826 }
1827 set_ctrl(phi, prev);
1828 }
1829 // Make 'use' use the Phi instead of the old loop body exit value
1830 _igvn.replace_input_of(use, idx, phi);
1831 if( use->_idx >= new_counter ) { // If updating new phis
1832 // Not needed for correctness, but prevents a weak assert
1833 // in AddPNode from tripping (when we end up with different
1834 // base & derived Phis that will become the same after
1835 // IGVN does CSE).
1836 Node *hit = _igvn.hash_find_insert(use);
1837 if( hit ) // Go ahead and re-hash for hits.
1838 _igvn.replace_node( use, hit );
1839 }
1840
1841 // If 'use' was in the loop-exit block, it now needs to be sunk
1842 // below the post-loop merge point.
1843 sink_use( use, prev );
1844 }
1845 }
1846 }
1847
1848 static void clone_outer_loop_helper(Node* n, const IdealLoopTree *loop, const IdealLoopTree* outer_loop,
1849 const Node_List &old_new, Unique_Node_List& wq, PhaseIdealLoop* phase,
1850 bool check_old_new) {
1851 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1852 Node* u = n->fast_out(j);
1853 assert(check_old_new || old_new[u->_idx] == NULL, "shouldn't have been cloned");
1854 if (!u->is_CFG() && (!check_old_new || old_new[u->_idx] == NULL)) {
1855 Node* c = phase->get_ctrl(u);
1856 IdealLoopTree* u_loop = phase->get_loop(c);
1857 assert(!loop->is_member(u_loop), "can be in outer loop or out of both loops only");
1858 if (outer_loop->is_member(u_loop)) {
1859 wq.push(u);
1860 }
1861 }
1862 }
1863 }
1864
1865 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
1866 IdealLoopTree* outer_loop, int dd, Node_List &old_new,
1867 Node_List& extra_data_nodes) {
1868 if (head->is_strip_mined() && mode != IgnoreStripMined) {
1869 CountedLoopNode* cl = head->as_CountedLoop();
1870 Node* l = cl->outer_loop();
1871 Node* tail = cl->outer_loop_tail();
1872 IfNode* le = cl->outer_loop_end();
1873 Node* sfpt = cl->outer_safepoint();
1874 CountedLoopEndNode* cle = cl->loopexit();
1875 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop();
1876 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null();
1877 Node* cle_out = cle->proj_out(false);
1878
1879 Node* new_sfpt = NULL;
1880 Node* new_cle_out = cle_out->clone();
1881 old_new.map(cle_out->_idx, new_cle_out);
1882 if (mode == CloneIncludesStripMined) {
1883 // clone outer loop body
1884 Node* new_l = l->clone();
1885 Node* new_tail = tail->clone();
1886 IfNode* new_le = le->clone()->as_If();
1887 new_sfpt = sfpt->clone();
1888
1889 set_loop(new_l, outer_loop->_parent);
1890 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd);
1891 set_loop(new_cle_out, outer_loop->_parent);
1892 set_idom(new_cle_out, new_cle, dd);
1893 set_loop(new_sfpt, outer_loop->_parent);
1894 set_idom(new_sfpt, new_cle_out, dd);
1895 set_loop(new_le, outer_loop->_parent);
1896 set_idom(new_le, new_sfpt, dd);
1897 set_loop(new_tail, outer_loop->_parent);
1898 set_idom(new_tail, new_le, dd);
1899 set_idom(new_cl, new_l, dd);
1900
1901 old_new.map(l->_idx, new_l);
1902 old_new.map(tail->_idx, new_tail);
1903 old_new.map(le->_idx, new_le);
1904 old_new.map(sfpt->_idx, new_sfpt);
1905
1906 new_l->set_req(LoopNode::LoopBackControl, new_tail);
1907 new_l->set_req(0, new_l);
1908 new_tail->set_req(0, new_le);
1909 new_le->set_req(0, new_sfpt);
1910 new_sfpt->set_req(0, new_cle_out);
1911 new_cle_out->set_req(0, new_cle);
1912 new_cl->set_req(LoopNode::EntryControl, new_l);
1913
1914 _igvn.register_new_node_with_optimizer(new_l);
1915 _igvn.register_new_node_with_optimizer(new_tail);
1916 _igvn.register_new_node_with_optimizer(new_le);
1917 } else {
1918 Node *newhead = old_new[loop->_head->_idx];
1919 newhead->as_Loop()->clear_strip_mined();
1920 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl));
1921 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
1922 }
1923 // Look at data node that were assigned a control in the outer
1924 // loop: they are kept in the outer loop by the safepoint so start
1925 // from the safepoint node's inputs.
1926 IdealLoopTree* outer_loop = get_loop(l);
1927 Node_Stack stack(2);
1928 stack.push(sfpt, 1);
1929 uint new_counter = C->unique();
1930 while (stack.size() > 0) {
1931 Node* n = stack.node();
1932 uint i = stack.index();
1933 while (i < n->req() &&
1934 (n->in(i) == NULL ||
1935 !has_ctrl(n->in(i)) ||
1936 get_loop(get_ctrl(n->in(i))) != outer_loop ||
1937 (old_new[n->in(i)->_idx] != NULL && old_new[n->in(i)->_idx]->_idx >= new_counter))) {
1938 i++;
1939 }
1940 if (i < n->req()) {
1941 stack.set_index(i+1);
1942 stack.push(n->in(i), 0);
1943 } else {
1944 assert(old_new[n->_idx] == NULL || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet");
1945 Node* m = n == sfpt ? new_sfpt : n->clone();
1946 if (m != NULL) {
1947 for (uint i = 0; i < n->req(); i++) {
1948 if (m->in(i) != NULL && old_new[m->in(i)->_idx] != NULL) {
1949 m->set_req(i, old_new[m->in(i)->_idx]);
1950 }
1951 }
1952 } else {
1953 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?");
1954 }
1955 if (n != sfpt) {
1956 extra_data_nodes.push(n);
1957 _igvn.register_new_node_with_optimizer(m);
1958 assert(get_ctrl(n) == cle_out, "what other control?");
1959 set_ctrl(m, new_cle_out);
1960 old_new.map(n->_idx, m);
1961 }
1962 stack.pop();
1963 }
1964 }
1965 if (mode == CloneIncludesStripMined) {
1966 _igvn.register_new_node_with_optimizer(new_sfpt);
1967 _igvn.register_new_node_with_optimizer(new_cle_out);
1968 }
1969 // Some other transformation may have pessimistically assign some
1970 // data nodes to the outer loop. Set their control so they are out
1971 // of the outer loop.
1972 ResourceMark rm;
1973 Unique_Node_List wq;
1974 for (uint i = 0; i < extra_data_nodes.size(); i++) {
1975 Node* old = extra_data_nodes.at(i);
1976 clone_outer_loop_helper(old, loop, outer_loop, old_new, wq, this, true);
1977 }
1978 Node* new_ctrl = cl->outer_loop_exit();
1979 assert(get_loop(new_ctrl) != outer_loop, "must be out of the loop nest");
1980 for (uint i = 0; i < wq.size(); i++) {
1981 Node* n = wq.at(i);
1982 set_ctrl(n, new_ctrl);
1983 clone_outer_loop_helper(n, loop, outer_loop, old_new, wq, this, false);
1984 }
1985 } else {
1986 Node *newhead = old_new[loop->_head->_idx];
1987 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
1988 }
1989 }
1990
1991 //------------------------------clone_loop-------------------------------------
1992 //
1993 // C L O N E A L O O P B O D Y
1994 //
1995 // This is the basic building block of the loop optimizations. It clones an
1996 // entire loop body. It makes an old_new loop body mapping; with this mapping
1997 // you can find the new-loop equivalent to an old-loop node. All new-loop
1998 // nodes are exactly equal to their old-loop counterparts, all edges are the
1999 // same. All exits from the old-loop now have a RegionNode that merges the
2000 // equivalent new-loop path. This is true even for the normal "loop-exit"
2001 // condition. All uses of loop-invariant old-loop values now come from (one
2002 // or more) Phis that merge their new-loop equivalents.
2003 //
2004 // This operation leaves the graph in an illegal state: there are two valid
2005 // control edges coming from the loop pre-header to both loop bodies. I'll
2006 // definitely have to hack the graph after running this transform.
2007 //
2008 // From this building block I will further edit edges to perform loop peeling
2009 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
2010 //
2011 // Parameter side_by_size_idom:
2012 // When side_by_size_idom is NULL, the dominator tree is constructed for
2013 // the clone loop to dominate the original. Used in construction of
2014 // pre-main-post loop sequence.
2015 // When nonnull, the clone and original are side-by-side, both are
2016 // dominated by the side_by_side_idom node. Used in construction of
2017 // unswitched loops.
2018 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
2019 CloneLoopMode mode, Node* side_by_side_idom) {
2020
2021 LoopNode* head = loop->_head->as_Loop();
2022 head->verify_strip_mined(1);
2023
2024 if (C->do_vector_loop() && PrintOpto) {
2025 const char* mname = C->method()->name()->as_quoted_ascii();
2026 if (mname != NULL) {
2027 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname);
2028 }
2029 }
2030
2031 CloneMap& cm = C->clone_map();
2032 Dict* dict = cm.dict();
2033 if (C->do_vector_loop()) {
2034 cm.set_clone_idx(cm.max_gen()+1);
2035 #ifndef PRODUCT
2036 if (PrintOpto) {
2037 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx());
2038 loop->dump_head();
2039 }
2040 #endif
2041 }
2042
2043 // Step 1: Clone the loop body. Make the old->new mapping.
2044 uint i;
2045 for( i = 0; i < loop->_body.size(); i++ ) {
2046 Node *old = loop->_body.at(i);
2047 Node *nnn = old->clone();
2048 old_new.map( old->_idx, nnn );
2049 if (C->do_vector_loop()) {
2050 cm.verify_insert_and_clone(old, nnn, cm.clone_idx());
2051 }
2052 _igvn.register_new_node_with_optimizer(nnn);
2053 }
2054
2055 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop;
2056
2057 // Step 2: Fix the edges in the new body. If the old input is outside the
2058 // loop use it. If the old input is INside the loop, use the corresponding
2059 // new node instead.
2060 for( i = 0; i < loop->_body.size(); i++ ) {
2061 Node *old = loop->_body.at(i);
2062 Node *nnn = old_new[old->_idx];
2063 // Fix CFG/Loop controlling the new node
2064 if (has_ctrl(old)) {
2065 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
2066 } else {
2067 set_loop(nnn, outer_loop->_parent);
2068 if (old->outcnt() > 0) {
2069 set_idom( nnn, old_new[idom(old)->_idx], dd );
2070 }
2071 }
2072 // Correct edges to the new node
2073 for( uint j = 0; j < nnn->req(); j++ ) {
2074 Node *n = nnn->in(j);
2075 if( n ) {
2076 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n );
2077 if( loop->is_member( old_in_loop ) )
2078 nnn->set_req(j, old_new[n->_idx]);
2079 }
2080 }
2081 _igvn.hash_find_insert(nnn);
2082 }
2083
2084 ResourceArea *area = Thread::current()->resource_area();
2085 Node_List extra_data_nodes(area); // data nodes in the outer strip mined loop
2086 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes);
2087
2088 // Step 3: Now fix control uses. Loop varying control uses have already
2089 // been fixed up (as part of all input edges in Step 2). Loop invariant
2090 // control uses must be either an IfFalse or an IfTrue. Make a merge
2091 // point to merge the old and new IfFalse/IfTrue nodes; make the use
2092 // refer to this.
2093 Node_List worklist(area);
2094 uint new_counter = C->unique();
2095 for( i = 0; i < loop->_body.size(); i++ ) {
2096 Node* old = loop->_body.at(i);
2097 if( !old->is_CFG() ) continue;
2098
2099 // Copy uses to a worklist, so I can munge the def-use info
2100 // with impunity.
2101 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
2102 worklist.push(old->fast_out(j));
2103
2104 while( worklist.size() ) { // Visit all uses
2105 Node *use = worklist.pop();
2106 if (!has_node(use)) continue; // Ignore dead nodes
2107 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
2108 if( !loop->is_member( use_loop ) && use->is_CFG() ) {
2109 // Both OLD and USE are CFG nodes here.
2110 assert( use->is_Proj(), "" );
2111 Node* nnn = old_new[old->_idx];
2112
2113 Node* newuse = NULL;
2114 if (head->is_strip_mined() && mode != IgnoreStripMined) {
2115 CountedLoopNode* cl = head->as_CountedLoop();
2116 CountedLoopEndNode* cle = cl->loopexit();
2117 Node* cle_out = cle->proj_out_or_null(false);
2118 if (use == cle_out) {
2119 IfNode* le = cl->outer_loop_end();
2120 use = le->proj_out(false);
2121 use_loop = get_loop(use);
2122 if (mode == CloneIncludesStripMined) {
2123 nnn = old_new[le->_idx];
2124 } else {
2125 newuse = old_new[cle_out->_idx];
2126 }
2127 }
2128 }
2129 if (newuse == NULL) {
2130 newuse = use->clone();
2131 }
2132
2133 // Clone the loop exit control projection
2134 if (C->do_vector_loop()) {
2135 cm.verify_insert_and_clone(use, newuse, cm.clone_idx());
2136 }
2137 newuse->set_req(0,nnn);
2138 _igvn.register_new_node_with_optimizer(newuse);
2139 set_loop(newuse, use_loop);
2140 set_idom(newuse, nnn, dom_depth(nnn) + 1 );
2141
2142 // We need a Region to merge the exit from the peeled body and the
2143 // exit from the old loop body.
2144 RegionNode *r = new RegionNode(3);
2145 // Map the old use to the new merge point
2146 old_new.map( use->_idx, r );
2147 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use));
2148 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" );
2149
2150 // The original user of 'use' uses 'r' instead.
2151 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
2152 Node* useuse = use->last_out(l);
2153 _igvn.rehash_node_delayed(useuse);
2154 uint uses_found = 0;
2155 if( useuse->in(0) == use ) {
2156 useuse->set_req(0, r);
2157 uses_found++;
2158 if( useuse->is_CFG() ) {
2159 assert( dom_depth(useuse) > dd_r, "" );
2160 set_idom(useuse, r, dom_depth(useuse));
2161 }
2162 }
2163 for( uint k = 1; k < useuse->req(); k++ ) {
2164 if( useuse->in(k) == use ) {
2165 useuse->set_req(k, r);
2166 uses_found++;
2167 if (useuse->is_Loop() && k == LoopNode::EntryControl) {
2168 assert(dom_depth(useuse) > dd_r , "");
2169 set_idom(useuse, r, dom_depth(useuse));
2170 }
2171 }
2172 }
2173 l -= uses_found; // we deleted 1 or more copies of this edge
2174 }
2175
2176 // Now finish up 'r'
2177 r->set_req( 1, newuse );
2178 r->set_req( 2, use );
2179 _igvn.register_new_node_with_optimizer(r);
2180 set_loop(r, use_loop);
2181 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r);
2182 } // End of if a loop-exit test
2183 }
2184 }
2185
2186 // Step 4: If loop-invariant use is not control, it must be dominated by a
2187 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region
2188 // there if needed. Make a Phi there merging old and new used values.
2189 Node_List *split_if_set = NULL;
2190 Node_List *split_bool_set = NULL;
2191 Node_List *split_cex_set = NULL;
2192 for( i = 0; i < loop->_body.size(); i++ ) {
2193 Node* old = loop->_body.at(i);
2194 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2195 split_bool_set, split_cex_set, worklist, new_counter,
2196 mode);
2197 }
2198
2199 for (i = 0; i < extra_data_nodes.size(); i++) {
2200 Node* old = extra_data_nodes.at(i);
2201 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2202 split_bool_set, split_cex_set, worklist, new_counter,
2203 mode);
2204 }
2205
2206 // Check for IFs that need splitting/cloning. Happens if an IF outside of
2207 // the loop uses a condition set in the loop. The original IF probably
2208 // takes control from one or more OLD Regions (which in turn get from NEW
2209 // Regions). In any case, there will be a set of Phis for each merge point
2210 // from the IF up to where the original BOOL def exists the loop.
2211 if (split_if_set) {
2212 while (split_if_set->size()) {
2213 Node *iff = split_if_set->pop();
2214 if (iff->in(1)->is_Phi()) {
2215 Node *b = clone_iff(iff->in(1)->as_Phi(), loop);
2216 _igvn.replace_input_of(iff, 1, b);
2217 }
2218 }
2219 }
2220 if (split_bool_set) {
2221 while (split_bool_set->size()) {
2222 Node *b = split_bool_set->pop();
2223 Node *phi = b->in(1);
2224 assert(phi->is_Phi(), "");
2225 CmpNode *cmp = clone_bool((PhiNode*)phi, loop);
2226 _igvn.replace_input_of(b, 1, cmp);
2227 }
2228 }
2229 if (split_cex_set) {
2230 while (split_cex_set->size()) {
2231 Node *b = split_cex_set->pop();
2232 assert(b->in(0)->is_Region(), "");
2233 assert(b->in(1)->is_Phi(), "");
2234 assert(b->in(0)->in(0) == b->in(1)->in(0), "");
2235 split_up(b, b->in(0), NULL);
2236 }
2237 }
2238
2239 }
2240
2241
2242 //---------------------- stride_of_possible_iv -------------------------------------
2243 // Looks for an iff/bool/comp with one operand of the compare
2244 // being a cycle involving an add and a phi,
2245 // with an optional truncation (left-shift followed by a right-shift)
2246 // of the add. Returns zero if not an iv.
2247 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
2248 Node* trunc1 = NULL;
2249 Node* trunc2 = NULL;
2250 const TypeInt* ttype = NULL;
2251 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) {
2252 return 0;
2253 }
2254 BoolNode* bl = iff->in(1)->as_Bool();
2255 Node* cmp = bl->in(1);
2256 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) {
2257 return 0;
2258 }
2259 // Must have an invariant operand
2260 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) {
2261 return 0;
2262 }
2263 Node* add2 = NULL;
2264 Node* cmp1 = cmp->in(1);
2265 if (cmp1->is_Phi()) {
2266 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
2267 Node* phi = cmp1;
2268 for (uint i = 1; i < phi->req(); i++) {
2269 Node* in = phi->in(i);
2270 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in,
2271 &trunc1, &trunc2, &ttype);
2272 if (add && add->in(1) == phi) {
2273 add2 = add->in(2);
2274 break;
2275 }
2276 }
2277 } else {
2278 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
2279 Node* addtrunc = cmp1;
2280 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc,
2281 &trunc1, &trunc2, &ttype);
2282 if (add && add->in(1)->is_Phi()) {
2283 Node* phi = add->in(1);
2284 for (uint i = 1; i < phi->req(); i++) {
2285 if (phi->in(i) == addtrunc) {
2286 add2 = add->in(2);
2287 break;
2288 }
2289 }
2290 }
2291 }
2292 if (add2 != NULL) {
2293 const TypeInt* add2t = _igvn.type(add2)->is_int();
2294 if (add2t->is_con()) {
2295 return add2t->get_con();
2296 }
2297 }
2298 return 0;
2299 }
2300
2301
2302 //---------------------- stay_in_loop -------------------------------------
2303 // Return the (unique) control output node that's in the loop (if it exists.)
2304 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
2305 Node* unique = NULL;
2306 if (!n) return NULL;
2307 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2308 Node* use = n->fast_out(i);
2309 if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
2310 if (unique != NULL) {
2311 return NULL;
2312 }
2313 unique = use;
2314 }
2315 }
2316 return unique;
2317 }
2318
2319 //------------------------------ register_node -------------------------------------
2320 // Utility to register node "n" with PhaseIdealLoop
2321 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) {
2322 _igvn.register_new_node_with_optimizer(n);
2323 loop->_body.push(n);
2324 if (n->is_CFG()) {
2325 set_loop(n, loop);
2326 set_idom(n, pred, ddepth);
2327 } else {
2328 set_ctrl(n, pred);
2329 }
2330 }
2331
2332 //------------------------------ proj_clone -------------------------------------
2333 // Utility to create an if-projection
2334 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
2335 ProjNode* c = p->clone()->as_Proj();
2336 c->set_req(0, iff);
2337 return c;
2338 }
2339
2340 //------------------------------ short_circuit_if -------------------------------------
2341 // Force the iff control output to be the live_proj
2342 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
2343 guarantee(live_proj != NULL, "null projection");
2344 int proj_con = live_proj->_con;
2345 assert(proj_con == 0 || proj_con == 1, "false or true projection");
2346 Node *con = _igvn.intcon(proj_con);
2347 set_ctrl(con, C->root());
2348 if (iff) {
2349 iff->set_req(1, con);
2350 }
2351 return con;
2352 }
2353
2354 //------------------------------ insert_if_before_proj -------------------------------------
2355 // Insert a new if before an if projection (* - new node)
2356 //
2357 // before
2358 // if(test)
2359 // / \
2360 // v v
2361 // other-proj proj (arg)
2362 //
2363 // after
2364 // if(test)
2365 // / \
2366 // / v
2367 // | * proj-clone
2368 // v |
2369 // other-proj v
2370 // * new_if(relop(cmp[IU](left,right)))
2371 // / \
2372 // v v
2373 // * new-proj proj
2374 // (returned)
2375 //
2376 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) {
2377 IfNode* iff = proj->in(0)->as_If();
2378 IdealLoopTree *loop = get_loop(proj);
2379 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2380 int ddepth = dom_depth(proj);
2381
2382 _igvn.rehash_node_delayed(iff);
2383 _igvn.rehash_node_delayed(proj);
2384
2385 proj->set_req(0, NULL); // temporary disconnect
2386 ProjNode* proj2 = proj_clone(proj, iff);
2387 register_node(proj2, loop, iff, ddepth);
2388
2389 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right);
2390 register_node(cmp, loop, proj2, ddepth);
2391
2392 BoolNode* bol = new BoolNode(cmp, relop);
2393 register_node(bol, loop, proj2, ddepth);
2394
2395 int opcode = iff->Opcode();
2396 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode");
2397 IfNode* new_if = (opcode == Op_If) ? new IfNode(proj2, bol, iff->_prob, iff->_fcnt):
2398 new RangeCheckNode(proj2, bol, iff->_prob, iff->_fcnt);
2399 register_node(new_if, loop, proj2, ddepth);
2400
2401 proj->set_req(0, new_if); // reattach
2402 set_idom(proj, new_if, ddepth);
2403
2404 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
2405 guarantee(new_exit != NULL, "null exit node");
2406 register_node(new_exit, get_loop(other_proj), new_if, ddepth);
2407
2408 return new_exit;
2409 }
2410
2411 //------------------------------ insert_region_before_proj -------------------------------------
2412 // Insert a region before an if projection (* - new node)
2413 //
2414 // before
2415 // if(test)
2416 // / |
2417 // v |
2418 // proj v
2419 // other-proj
2420 //
2421 // after
2422 // if(test)
2423 // / |
2424 // v |
2425 // * proj-clone v
2426 // | other-proj
2427 // v
2428 // * new-region
2429 // |
2430 // v
2431 // * dum_if
2432 // / \
2433 // v \
2434 // * dum-proj v
2435 // proj
2436 //
2437 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
2438 IfNode* iff = proj->in(0)->as_If();
2439 IdealLoopTree *loop = get_loop(proj);
2440 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2441 int ddepth = dom_depth(proj);
2442
2443 _igvn.rehash_node_delayed(iff);
2444 _igvn.rehash_node_delayed(proj);
2445
2446 proj->set_req(0, NULL); // temporary disconnect
2447 ProjNode* proj2 = proj_clone(proj, iff);
2448 register_node(proj2, loop, iff, ddepth);
2449
2450 RegionNode* reg = new RegionNode(2);
2451 reg->set_req(1, proj2);
2452 register_node(reg, loop, iff, ddepth);
2453
2454 IfNode* dum_if = new IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt);
2455 register_node(dum_if, loop, reg, ddepth);
2456
2457 proj->set_req(0, dum_if); // reattach
2458 set_idom(proj, dum_if, ddepth);
2459
2460 ProjNode* dum_proj = proj_clone(other_proj, dum_if);
2461 register_node(dum_proj, loop, dum_if, ddepth);
2462
2463 return reg;
2464 }
2465
2466 //------------------------------ insert_cmpi_loop_exit -------------------------------------
2467 // Clone a signed compare loop exit from an unsigned compare and
2468 // insert it before the unsigned cmp on the stay-in-loop path.
2469 // All new nodes inserted in the dominator tree between the original
2470 // if and it's projections. The original if test is replaced with
2471 // a constant to force the stay-in-loop path.
2472 //
2473 // This is done to make sure that the original if and it's projections
2474 // still dominate the same set of control nodes, that the ctrl() relation
2475 // from data nodes to them is preserved, and that their loop nesting is
2476 // preserved.
2477 //
2478 // before
2479 // if(i <u limit) unsigned compare loop exit
2480 // / |
2481 // v v
2482 // exit-proj stay-in-loop-proj
2483 //
2484 // after
2485 // if(stay-in-loop-const) original if
2486 // / |
2487 // / v
2488 // / if(i < limit) new signed test
2489 // / / |
2490 // / / v
2491 // / / if(i <u limit) new cloned unsigned test
2492 // / / / |
2493 // v v v |
2494 // region |
2495 // | |
2496 // dum-if |
2497 // / | |
2498 // ether | |
2499 // v v
2500 // exit-proj stay-in-loop-proj
2501 //
2502 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) {
2503 const bool Signed = true;
2504 const bool Unsigned = false;
2505
2506 BoolNode* bol = if_cmpu->in(1)->as_Bool();
2507 if (bol->_test._test != BoolTest::lt) return NULL;
2508 CmpNode* cmpu = bol->in(1)->as_Cmp();
2509 if (cmpu->Opcode() != Op_CmpU) return NULL;
2510 int stride = stride_of_possible_iv(if_cmpu);
2511 if (stride == 0) return NULL;
2512
2513 Node* lp_proj = stay_in_loop(if_cmpu, loop);
2514 guarantee(lp_proj != NULL, "null loop node");
2515
2516 ProjNode* lp_continue = lp_proj->as_Proj();
2517 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
2518
2519 Node* limit = NULL;
2520 if (stride > 0) {
2521 limit = cmpu->in(2);
2522 } else {
2523 limit = _igvn.makecon(TypeInt::ZERO);
2524 set_ctrl(limit, C->root());
2525 }
2526 // Create a new region on the exit path
2527 RegionNode* reg = insert_region_before_proj(lp_exit);
2528 guarantee(reg != NULL, "null region node");
2529
2530 // Clone the if-cmpu-true-false using a signed compare
2531 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
2532 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue);
2533 reg->add_req(cmpi_exit);
2534
2535 // Clone the if-cmpu-true-false
2536 BoolTest::mask rel_u = bol->_test._test;
2537 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
2538 reg->add_req(cmpu_exit);
2539
2540 // Force original if to stay in loop.
2541 short_circuit_if(if_cmpu, lp_continue);
2542
2543 return cmpi_exit->in(0)->as_If();
2544 }
2545
2546 //------------------------------ remove_cmpi_loop_exit -------------------------------------
2547 // Remove a previously inserted signed compare loop exit.
2548 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
2549 Node* lp_proj = stay_in_loop(if_cmp, loop);
2550 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
2551 stay_in_loop(lp_proj, loop)->is_If() &&
2552 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
2553 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
2554 set_ctrl(con, C->root());
2555 if_cmp->set_req(1, con);
2556 }
2557
2558 //------------------------------ scheduled_nodelist -------------------------------------
2559 // Create a post order schedule of nodes that are in the
2560 // "member" set. The list is returned in "sched".
2561 // The first node in "sched" is the loop head, followed by
2562 // nodes which have no inputs in the "member" set, and then
2563 // followed by the nodes that have an immediate input dependence
2564 // on a node in "sched".
2565 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
2566
2567 assert(member.test(loop->_head->_idx), "loop head must be in member set");
2568 Arena *a = Thread::current()->resource_area();
2569 VectorSet visited(a);
2570 Node_Stack nstack(a, loop->_body.size());
2571
2572 Node* n = loop->_head; // top of stack is cached in "n"
2573 uint idx = 0;
2574 visited.set(n->_idx);
2575
2576 // Initially push all with no inputs from within member set
2577 for(uint i = 0; i < loop->_body.size(); i++ ) {
2578 Node *elt = loop->_body.at(i);
2579 if (member.test(elt->_idx)) {
2580 bool found = false;
2581 for (uint j = 0; j < elt->req(); j++) {
2582 Node* def = elt->in(j);
2583 if (def && member.test(def->_idx) && def != elt) {
2584 found = true;
2585 break;
2586 }
2587 }
2588 if (!found && elt != loop->_head) {
2589 nstack.push(n, idx);
2590 n = elt;
2591 assert(!visited.test(n->_idx), "not seen yet");
2592 visited.set(n->_idx);
2593 }
2594 }
2595 }
2596
2597 // traverse out's that are in the member set
2598 while (true) {
2599 if (idx < n->outcnt()) {
2600 Node* use = n->raw_out(idx);
2601 idx++;
2602 if (!visited.test_set(use->_idx)) {
2603 if (member.test(use->_idx)) {
2604 nstack.push(n, idx);
2605 n = use;
2606 idx = 0;
2607 }
2608 }
2609 } else {
2610 // All outputs processed
2611 sched.push(n);
2612 if (nstack.is_empty()) break;
2613 n = nstack.node();
2614 idx = nstack.index();
2615 nstack.pop();
2616 }
2617 }
2618 }
2619
2620
2621 //------------------------------ has_use_in_set -------------------------------------
2622 // Has a use in the vector set
2623 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
2624 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2625 Node* use = n->fast_out(j);
2626 if (vset.test(use->_idx)) {
2627 return true;
2628 }
2629 }
2630 return false;
2631 }
2632
2633
2634 //------------------------------ has_use_internal_to_set -------------------------------------
2635 // Has use internal to the vector set (ie. not in a phi at the loop head)
2636 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
2637 Node* head = loop->_head;
2638 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2639 Node* use = n->fast_out(j);
2640 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
2641 return true;
2642 }
2643 }
2644 return false;
2645 }
2646
2647
2648 //------------------------------ clone_for_use_outside_loop -------------------------------------
2649 // clone "n" for uses that are outside of loop
2650 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
2651 int cloned = 0;
2652 assert(worklist.size() == 0, "should be empty");
2653 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2654 Node* use = n->fast_out(j);
2655 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
2656 worklist.push(use);
2657 }
2658 }
2659 while( worklist.size() ) {
2660 Node *use = worklist.pop();
2661 if (!has_node(use) || use->in(0) == C->top()) continue;
2662 uint j;
2663 for (j = 0; j < use->req(); j++) {
2664 if (use->in(j) == n) break;
2665 }
2666 assert(j < use->req(), "must be there");
2667
2668 // clone "n" and insert it between the inputs of "n" and the use outside the loop
2669 Node* n_clone = n->clone();
2670 _igvn.replace_input_of(use, j, n_clone);
2671 cloned++;
2672 Node* use_c;
2673 if (!use->is_Phi()) {
2674 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
2675 } else {
2676 // Use in a phi is considered a use in the associated predecessor block
2677 use_c = use->in(0)->in(j);
2678 }
2679 set_ctrl(n_clone, use_c);
2680 assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
2681 get_loop(use_c)->_body.push(n_clone);
2682 _igvn.register_new_node_with_optimizer(n_clone);
2683 #if !defined(PRODUCT)
2684 if (TracePartialPeeling) {
2685 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
2686 }
2687 #endif
2688 }
2689 return cloned;
2690 }
2691
2692
2693 //------------------------------ clone_for_special_use_inside_loop -------------------------------------
2694 // clone "n" for special uses that are in the not_peeled region.
2695 // If these def-uses occur in separate blocks, the code generator
2696 // marks the method as not compilable. For example, if a "BoolNode"
2697 // is in a different basic block than the "IfNode" that uses it, then
2698 // the compilation is aborted in the code generator.
2699 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
2700 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
2701 if (n->is_Phi() || n->is_Load()) {
2702 return;
2703 }
2704 assert(worklist.size() == 0, "should be empty");
2705 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2706 Node* use = n->fast_out(j);
2707 if ( not_peel.test(use->_idx) &&
2708 (use->is_If() || use->is_CMove() || use->is_Bool()) &&
2709 use->in(1) == n) {
2710 worklist.push(use);
2711 }
2712 }
2713 if (worklist.size() > 0) {
2714 // clone "n" and insert it between inputs of "n" and the use
2715 Node* n_clone = n->clone();
2716 loop->_body.push(n_clone);
2717 _igvn.register_new_node_with_optimizer(n_clone);
2718 set_ctrl(n_clone, get_ctrl(n));
2719 sink_list.push(n_clone);
2720 not_peel <<= n_clone->_idx; // add n_clone to not_peel set.
2721 #if !defined(PRODUCT)
2722 if (TracePartialPeeling) {
2723 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
2724 }
2725 #endif
2726 while( worklist.size() ) {
2727 Node *use = worklist.pop();
2728 _igvn.rehash_node_delayed(use);
2729 for (uint j = 1; j < use->req(); j++) {
2730 if (use->in(j) == n) {
2731 use->set_req(j, n_clone);
2732 }
2733 }
2734 }
2735 }
2736 }
2737
2738
2739 //------------------------------ insert_phi_for_loop -------------------------------------
2740 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
2741 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
2742 Node *phi = PhiNode::make(lp, back_edge_val);
2743 phi->set_req(LoopNode::EntryControl, lp_entry_val);
2744 // Use existing phi if it already exists
2745 Node *hit = _igvn.hash_find_insert(phi);
2746 if( hit == NULL ) {
2747 _igvn.register_new_node_with_optimizer(phi);
2748 set_ctrl(phi, lp);
2749 } else {
2750 // Remove the new phi from the graph and use the hit
2751 _igvn.remove_dead_node(phi);
2752 phi = hit;
2753 }
2754 _igvn.replace_input_of(use, idx, phi);
2755 }
2756
2757 #ifdef ASSERT
2758 //------------------------------ is_valid_loop_partition -------------------------------------
2759 // Validate the loop partition sets: peel and not_peel
2760 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
2761 VectorSet& not_peel ) {
2762 uint i;
2763 // Check that peel_list entries are in the peel set
2764 for (i = 0; i < peel_list.size(); i++) {
2765 if (!peel.test(peel_list.at(i)->_idx)) {
2766 return false;
2767 }
2768 }
2769 // Check at loop members are in one of peel set or not_peel set
2770 for (i = 0; i < loop->_body.size(); i++ ) {
2771 Node *def = loop->_body.at(i);
2772 uint di = def->_idx;
2773 // Check that peel set elements are in peel_list
2774 if (peel.test(di)) {
2775 if (not_peel.test(di)) {
2776 return false;
2777 }
2778 // Must be in peel_list also
2779 bool found = false;
2780 for (uint j = 0; j < peel_list.size(); j++) {
2781 if (peel_list.at(j)->_idx == di) {
2782 found = true;
2783 break;
2784 }
2785 }
2786 if (!found) {
2787 return false;
2788 }
2789 } else if (not_peel.test(di)) {
2790 if (peel.test(di)) {
2791 return false;
2792 }
2793 } else {
2794 return false;
2795 }
2796 }
2797 return true;
2798 }
2799
2800 //------------------------------ is_valid_clone_loop_exit_use -------------------------------------
2801 // Ensure a use outside of loop is of the right form
2802 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
2803 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2804 return (use->is_Phi() &&
2805 use_c->is_Region() && use_c->req() == 3 &&
2806 (use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
2807 use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
2808 use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
2809 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
2810 }
2811
2812 //------------------------------ is_valid_clone_loop_form -------------------------------------
2813 // Ensure that all uses outside of loop are of the right form
2814 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
2815 uint orig_exit_idx, uint clone_exit_idx) {
2816 uint len = peel_list.size();
2817 for (uint i = 0; i < len; i++) {
2818 Node *def = peel_list.at(i);
2819
2820 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2821 Node *use = def->fast_out(j);
2822 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2823 if (!loop->is_member(get_loop(use_c))) {
2824 // use is not in the loop, check for correct structure
2825 if (use->in(0) == def) {
2826 // Okay
2827 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
2828 return false;
2829 }
2830 }
2831 }
2832 }
2833 return true;
2834 }
2835 #endif
2836
2837 //------------------------------ partial_peel -------------------------------------
2838 // Partially peel (aka loop rotation) the top portion of a loop (called
2839 // the peel section below) by cloning it and placing one copy just before
2840 // the new loop head and the other copy at the bottom of the new loop.
2841 //
2842 // before after where it came from
2843 //
2844 // stmt1 stmt1
2845 // loop: stmt2 clone
2846 // stmt2 if condA goto exitA clone
2847 // if condA goto exitA new_loop: new
2848 // stmt3 stmt3 clone
2849 // if !condB goto loop if condB goto exitB clone
2850 // exitB: stmt2 orig
2851 // stmt4 if !condA goto new_loop orig
2852 // exitA: goto exitA
2853 // exitB:
2854 // stmt4
2855 // exitA:
2856 //
2857 // Step 1: find the cut point: an exit test on probable
2858 // induction variable.
2859 // Step 2: schedule (with cloning) operations in the peel
2860 // section that can be executed after the cut into
2861 // the section that is not peeled. This may need
2862 // to clone operations into exit blocks. For
2863 // instance, a reference to A[i] in the not-peel
2864 // section and a reference to B[i] in an exit block
2865 // may cause a left-shift of i by 2 to be placed
2866 // in the peel block. This step will clone the left
2867 // shift into the exit block and sink the left shift
2868 // from the peel to the not-peel section.
2869 // Step 3: clone the loop, retarget the control, and insert
2870 // phis for values that are live across the new loop
2871 // head. This is very dependent on the graph structure
2872 // from clone_loop. It creates region nodes for
2873 // exit control and associated phi nodes for values
2874 // flow out of the loop through that exit. The region
2875 // node is dominated by the clone's control projection.
2876 // So the clone's peel section is placed before the
2877 // new loop head, and the clone's not-peel section is
2878 // forms the top part of the new loop. The original
2879 // peel section forms the tail of the new loop.
2880 // Step 4: update the dominator tree and recompute the
2881 // dominator depth.
2882 //
2883 // orig
2884 //
2885 // stmt1
2886 // |
2887 // v
2888 // loop predicate
2889 // |
2890 // v
2891 // loop<----+
2892 // | |
2893 // stmt2 |
2894 // | |
2895 // v |
2896 // ifA |
2897 // / | |
2898 // v v |
2899 // false true ^ <-- last_peel
2900 // / | |
2901 // / ===|==cut |
2902 // / stmt3 | <-- first_not_peel
2903 // / | |
2904 // | v |
2905 // v ifB |
2906 // exitA: / \ |
2907 // / \ |
2908 // v v |
2909 // false true |
2910 // / \ |
2911 // / ----+
2912 // |
2913 // v
2914 // exitB:
2915 // stmt4
2916 //
2917 //
2918 // after clone loop
2919 //
2920 // stmt1
2921 // |
2922 // v
2923 // loop predicate
2924 // / \
2925 // clone / \ orig
2926 // / \
2927 // / \
2928 // v v
2929 // +---->loop loop<----+
2930 // | | | |
2931 // | stmt2 stmt2 |
2932 // | | | |
2933 // | v v |
2934 // | ifA ifA |
2935 // | | \ / | |
2936 // | v v v v |
2937 // ^ true false false true ^ <-- last_peel
2938 // | | ^ \ / | |
2939 // | cut==|== \ \ / ===|==cut |
2940 // | stmt3 \ \ / stmt3 | <-- first_not_peel
2941 // | | dom | | | |
2942 // | v \ 1v v2 v |
2943 // | ifB regionA ifB |
2944 // | / \ | / \ |
2945 // | / \ v / \ |
2946 // | v v exitA: v v |
2947 // | true false false true |
2948 // | / ^ \ / \ |
2949 // +---- \ \ / ----+
2950 // dom \ /
2951 // \ 1v v2
2952 // regionB
2953 // |
2954 // v
2955 // exitB:
2956 // stmt4
2957 //
2958 //
2959 // after partial peel
2960 //
2961 // stmt1
2962 // |
2963 // v
2964 // loop predicate
2965 // /
2966 // clone / orig
2967 // / TOP
2968 // / \
2969 // v v
2970 // TOP->loop loop----+
2971 // | | |
2972 // stmt2 stmt2 |
2973 // | | |
2974 // v v |
2975 // ifA ifA |
2976 // | \ / | |
2977 // v v v v |
2978 // true false false true | <-- last_peel
2979 // | ^ \ / +------|---+
2980 // +->newloop \ \ / === ==cut | |
2981 // | stmt3 \ \ / TOP | |
2982 // | | dom | | stmt3 | | <-- first_not_peel
2983 // | v \ 1v v2 v | |
2984 // | ifB regionA ifB ^ v
2985 // | / \ | / \ | |
2986 // | / \ v / \ | |
2987 // | v v exitA: v v | |
2988 // | true false false true | |
2989 // | / ^ \ / \ | |
2990 // | | \ \ / v | |
2991 // | | dom \ / TOP | |
2992 // | | \ 1v v2 | |
2993 // ^ v regionB | |
2994 // | | | | |
2995 // | | v ^ v
2996 // | | exitB: | |
2997 // | | stmt4 | |
2998 // | +------------>-----------------+ |
2999 // | |
3000 // +-----------------<---------------------+
3001 //
3002 //
3003 // final graph
3004 //
3005 // stmt1
3006 // |
3007 // v
3008 // loop predicate
3009 // |
3010 // v
3011 // stmt2 clone
3012 // |
3013 // v
3014 // ........> ifA clone
3015 // : / |
3016 // dom / |
3017 // : v v
3018 // : false true
3019 // : | |
3020 // : | v
3021 // : | newloop<-----+
3022 // : | | |
3023 // : | stmt3 clone |
3024 // : | | |
3025 // : | v |
3026 // : | ifB |
3027 // : | / \ |
3028 // : | v v |
3029 // : | false true |
3030 // : | | | |
3031 // : | v stmt2 |
3032 // : | exitB: | |
3033 // : | stmt4 v |
3034 // : | ifA orig |
3035 // : | / \ |
3036 // : | / \ |
3037 // : | v v |
3038 // : | false true |
3039 // : | / \ |
3040 // : v v -----+
3041 // RegionA
3042 // |
3043 // v
3044 // exitA
3045 //
3046 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
3047
3048 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only");
3049 if (!loop->_head->is_Loop()) {
3050 return false; }
3051
3052 LoopNode *head = loop->_head->as_Loop();
3053
3054 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
3055 return false;
3056 }
3057
3058 // Check for complex exit control
3059 for(uint ii = 0; ii < loop->_body.size(); ii++ ) {
3060 Node *n = loop->_body.at(ii);
3061 int opc = n->Opcode();
3062 if (n->is_Call() ||
3063 opc == Op_Catch ||
3064 opc == Op_CatchProj ||
3065 opc == Op_Jump ||
3066 opc == Op_JumpProj) {
3067 #if !defined(PRODUCT)
3068 if (TracePartialPeeling) {
3069 tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
3070 }
3071 #endif
3072 return false;
3073 }
3074 }
3075
3076 int dd = dom_depth(head);
3077
3078 // Step 1: find cut point
3079
3080 // Walk up dominators to loop head looking for first loop exit
3081 // which is executed on every path thru loop.
3082 IfNode *peel_if = NULL;
3083 IfNode *peel_if_cmpu = NULL;
3084
3085 Node *iff = loop->tail();
3086 while( iff != head ) {
3087 if( iff->is_If() ) {
3088 Node *ctrl = get_ctrl(iff->in(1));
3089 if (ctrl->is_top()) return false; // Dead test on live IF.
3090 // If loop-varying exit-test, check for induction variable
3091 if( loop->is_member(get_loop(ctrl)) &&
3092 loop->is_loop_exit(iff) &&
3093 is_possible_iv_test(iff)) {
3094 Node* cmp = iff->in(1)->in(1);
3095 if (cmp->Opcode() == Op_CmpI) {
3096 peel_if = iff->as_If();
3097 } else {
3098 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
3099 peel_if_cmpu = iff->as_If();
3100 }
3101 }
3102 }
3103 iff = idom(iff);
3104 }
3105 // Prefer signed compare over unsigned compare.
3106 IfNode* new_peel_if = NULL;
3107 if (peel_if == NULL) {
3108 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) {
3109 return false; // No peel point found
3110 }
3111 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
3112 if (new_peel_if == NULL) {
3113 return false; // No peel point found
3114 }
3115 peel_if = new_peel_if;
3116 }
3117 Node* last_peel = stay_in_loop(peel_if, loop);
3118 Node* first_not_peeled = stay_in_loop(last_peel, loop);
3119 if (first_not_peeled == NULL || first_not_peeled == head) {
3120 return false;
3121 }
3122
3123 #if !defined(PRODUCT)
3124 if (TraceLoopOpts) {
3125 tty->print("PartialPeel ");
3126 loop->dump_head();
3127 }
3128
3129 if (TracePartialPeeling) {
3130 tty->print_cr("before partial peel one iteration");
3131 Node_List wl;
3132 Node* t = head->in(2);
3133 while (true) {
3134 wl.push(t);
3135 if (t == head) break;
3136 t = idom(t);
3137 }
3138 while (wl.size() > 0) {
3139 Node* tt = wl.pop();
3140 tt->dump();
3141 if (tt == last_peel) tty->print_cr("-- cut --");
3142 }
3143 }
3144 #endif
3145 ResourceArea *area = Thread::current()->resource_area();
3146 VectorSet peel(area);
3147 VectorSet not_peel(area);
3148 Node_List peel_list(area);
3149 Node_List worklist(area);
3150 Node_List sink_list(area);
3151
3152 // Set of cfg nodes to peel are those that are executable from
3153 // the head through last_peel.
3154 assert(worklist.size() == 0, "should be empty");
3155 worklist.push(head);
3156 peel.set(head->_idx);
3157 while (worklist.size() > 0) {
3158 Node *n = worklist.pop();
3159 if (n != last_peel) {
3160 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3161 Node* use = n->fast_out(j);
3162 if (use->is_CFG() &&
3163 loop->is_member(get_loop(use)) &&
3164 !peel.test_set(use->_idx)) {
3165 worklist.push(use);
3166 }
3167 }
3168 }
3169 }
3170
3171 // Set of non-cfg nodes to peel are those that are control
3172 // dependent on the cfg nodes.
3173 uint i;
3174 for(i = 0; i < loop->_body.size(); i++ ) {
3175 Node *n = loop->_body.at(i);
3176 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
3177 if (peel.test(n_c->_idx)) {
3178 peel.set(n->_idx);
3179 } else {
3180 not_peel.set(n->_idx);
3181 }
3182 }
3183
3184 // Step 2: move operations from the peeled section down into the
3185 // not-peeled section
3186
3187 // Get a post order schedule of nodes in the peel region
3188 // Result in right-most operand.
3189 scheduled_nodelist(loop, peel, peel_list );
3190
3191 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
3192
3193 // For future check for too many new phis
3194 uint old_phi_cnt = 0;
3195 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
3196 Node* use = head->fast_out(j);
3197 if (use->is_Phi()) old_phi_cnt++;
3198 }
3199
3200 #if !defined(PRODUCT)
3201 if (TracePartialPeeling) {
3202 tty->print_cr("\npeeled list");
3203 }
3204 #endif
3205
3206 // Evacuate nodes in peel region into the not_peeled region if possible
3207 uint new_phi_cnt = 0;
3208 uint cloned_for_outside_use = 0;
3209 for (i = 0; i < peel_list.size();) {
3210 Node* n = peel_list.at(i);
3211 #if !defined(PRODUCT)
3212 if (TracePartialPeeling) n->dump();
3213 #endif
3214 bool incr = true;
3215 if ( !n->is_CFG() ) {
3216
3217 if ( has_use_in_set(n, not_peel) ) {
3218
3219 // If not used internal to the peeled region,
3220 // move "n" from peeled to not_peeled region.
3221
3222 if ( !has_use_internal_to_set(n, peel, loop) ) {
3223
3224 // if not pinned and not a load (which maybe anti-dependent on a store)
3225 // and not a CMove (Matcher expects only bool->cmove).
3226 if (n->in(0) == NULL && !n->is_Load() && !n->is_CMove() && n->Opcode() != Op_ShenandoahWBMemProj) {
3227 cloned_for_outside_use += clone_for_use_outside_loop( loop, n, worklist );
3228 sink_list.push(n);
3229 peel >>= n->_idx; // delete n from peel set.
3230 not_peel <<= n->_idx; // add n to not_peel set.
3231 peel_list.remove(i);
3232 incr = false;
3233 #if !defined(PRODUCT)
3234 if (TracePartialPeeling) {
3235 tty->print_cr("sink to not_peeled region: %d newbb: %d",
3236 n->_idx, get_ctrl(n)->_idx);
3237 }
3238 #endif
3239 }
3240 } else {
3241 // Otherwise check for special def-use cases that span
3242 // the peel/not_peel boundary such as bool->if
3243 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist );
3244 new_phi_cnt++;
3245 }
3246 }
3247 }
3248 if (incr) i++;
3249 }
3250
3251 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) {
3252 #if !defined(PRODUCT)
3253 if (TracePartialPeeling) {
3254 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c",
3255 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F');
3256 }
3257 #endif
3258 if (new_peel_if != NULL) {
3259 remove_cmpi_loop_exit(new_peel_if, loop);
3260 }
3261 // Inhibit more partial peeling on this loop
3262 assert(!head->is_partial_peel_loop(), "not partial peeled");
3263 head->mark_partial_peel_failed();
3264 if (cloned_for_outside_use > 0) {
3265 // Terminate this round of loop opts because
3266 // the graph outside this loop was changed.
3267 C->set_major_progress();
3268 return true;
3269 }
3270 return false;
3271 }
3272
3273 // Step 3: clone loop, retarget control, and insert new phis
3274
3275 // Create new loop head for new phis and to hang
3276 // the nodes being moved (sinked) from the peel region.
3277 LoopNode* new_head = new LoopNode(last_peel, last_peel);
3278 new_head->set_unswitch_count(head->unswitch_count()); // Preserve
3279 _igvn.register_new_node_with_optimizer(new_head);
3280 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
3281 _igvn.replace_input_of(first_not_peeled, 0, new_head);
3282 set_loop(new_head, loop);
3283 loop->_body.push(new_head);
3284 not_peel.set(new_head->_idx);
3285 set_idom(new_head, last_peel, dom_depth(first_not_peeled));
3286 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
3287
3288 while (sink_list.size() > 0) {
3289 Node* n = sink_list.pop();
3290 set_ctrl(n, new_head);
3291 }
3292
3293 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
3294
3295 clone_loop(loop, old_new, dd, IgnoreStripMined);
3296
3297 const uint clone_exit_idx = 1;
3298 const uint orig_exit_idx = 2;
3299 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop");
3300
3301 Node* head_clone = old_new[head->_idx];
3302 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop();
3303 Node* orig_tail_clone = head_clone->in(2);
3304
3305 // Add phi if "def" node is in peel set and "use" is not
3306
3307 for(i = 0; i < peel_list.size(); i++ ) {
3308 Node *def = peel_list.at(i);
3309 if (!def->is_CFG()) {
3310 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
3311 Node *use = def->fast_out(j);
3312 if (has_node(use) && use->in(0) != C->top() &&
3313 (!peel.test(use->_idx) ||
3314 (use->is_Phi() && use->in(0) == head)) ) {
3315 worklist.push(use);
3316 }
3317 }
3318 while( worklist.size() ) {
3319 Node *use = worklist.pop();
3320 for (uint j = 1; j < use->req(); j++) {
3321 Node* n = use->in(j);
3322 if (n == def) {
3323
3324 // "def" is in peel set, "use" is not in peel set
3325 // or "use" is in the entry boundary (a phi) of the peel set
3326
3327 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
3328
3329 if ( loop->is_member(get_loop( use_c )) ) {
3330 // use is in loop
3331 if (old_new[use->_idx] != NULL) { // null for dead code
3332 Node* use_clone = old_new[use->_idx];
3333 _igvn.replace_input_of(use, j, C->top());
3334 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
3335 }
3336 } else {
3337 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
3338 // use is not in the loop, check if the live range includes the cut
3339 Node* lp_if = use_c->in(orig_exit_idx)->in(0);
3340 if (not_peel.test(lp_if->_idx)) {
3341 assert(j == orig_exit_idx, "use from original loop");
3342 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
3343 }
3344 }
3345 }
3346 }
3347 }
3348 }
3349 }
3350
3351 // Step 3b: retarget control
3352
3353 // Redirect control to the new loop head if a cloned node in
3354 // the not_peeled region has control that points into the peeled region.
3355 // This necessary because the cloned peeled region will be outside
3356 // the loop.
3357 // from to
3358 // cloned-peeled <---+
3359 // new_head_clone: | <--+
3360 // cloned-not_peeled in(0) in(0)
3361 // orig-peeled
3362
3363 for(i = 0; i < loop->_body.size(); i++ ) {
3364 Node *n = loop->_body.at(i);
3365 if (!n->is_CFG() && n->in(0) != NULL &&
3366 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
3367 Node* n_clone = old_new[n->_idx];
3368 _igvn.replace_input_of(n_clone, 0, new_head_clone);
3369 }
3370 }
3371
3372 // Backedge of the surviving new_head (the clone) is original last_peel
3373 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel);
3374
3375 // Cut first node in original not_peel set
3376 _igvn.rehash_node_delayed(new_head); // Multiple edge updates:
3377 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of
3378 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls
3379
3380 // Copy head_clone back-branch info to original head
3381 // and remove original head's loop entry and
3382 // clone head's back-branch
3383 _igvn.rehash_node_delayed(head); // Multiple edge updates
3384 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl));
3385 head->set_req(LoopNode::LoopBackControl, C->top());
3386 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top());
3387
3388 // Similarly modify the phis
3389 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
3390 Node* use = head->fast_out(k);
3391 if (use->is_Phi() && use->outcnt() > 0) {
3392 Node* use_clone = old_new[use->_idx];
3393 _igvn.rehash_node_delayed(use); // Multiple edge updates
3394 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl));
3395 use->set_req(LoopNode::LoopBackControl, C->top());
3396 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top());
3397 }
3398 }
3399
3400 // Step 4: update dominator tree and dominator depth
3401
3402 set_idom(head, orig_tail_clone, dd);
3403 recompute_dom_depth();
3404
3405 // Inhibit more partial peeling on this loop
3406 new_head_clone->set_partial_peel_loop();
3407 C->set_major_progress();
3408 loop->record_for_igvn();
3409
3410 #if !defined(PRODUCT)
3411 if (TracePartialPeeling) {
3412 tty->print_cr("\nafter partial peel one iteration");
3413 Node_List wl(area);
3414 Node* t = last_peel;
3415 while (true) {
3416 wl.push(t);
3417 if (t == head_clone) break;
3418 t = idom(t);
3419 }
3420 while (wl.size() > 0) {
3421 Node* tt = wl.pop();
3422 if (tt == head) tty->print_cr("orig head");
3423 else if (tt == new_head_clone) tty->print_cr("new head");
3424 else if (tt == head_clone) tty->print_cr("clone head");
3425 tt->dump();
3426 }
3427 }
3428 #endif
3429 return true;
3430 }
3431
3432 //------------------------------reorg_offsets----------------------------------
3433 // Reorganize offset computations to lower register pressure. Mostly
3434 // prevent loop-fallout uses of the pre-incremented trip counter (which are
3435 // then alive with the post-incremented trip counter forcing an extra
3436 // register move)
3437 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) {
3438 // Perform it only for canonical counted loops.
3439 // Loop's shape could be messed up by iteration_split_impl.
3440 if (!loop->_head->is_CountedLoop())
3441 return;
3442 if (!loop->_head->as_Loop()->is_valid_counted_loop())
3443 return;
3444
3445 CountedLoopNode *cl = loop->_head->as_CountedLoop();
3446 CountedLoopEndNode *cle = cl->loopexit();
3447 Node *exit = cle->proj_out(false);
3448 Node *phi = cl->phi();
3449
3450 // Check for the special case of folks using the pre-incremented
3451 // trip-counter on the fall-out path (forces the pre-incremented
3452 // and post-incremented trip counter to be live at the same time).
3453 // Fix this by adjusting to use the post-increment trip counter.
3454
3455 bool progress = true;
3456 while (progress) {
3457 progress = false;
3458 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
3459 Node* use = phi->fast_out(i); // User of trip-counter
3460 if (!has_ctrl(use)) continue;
3461 Node *u_ctrl = get_ctrl(use);
3462 if (use->is_Phi()) {
3463 u_ctrl = NULL;
3464 for (uint j = 1; j < use->req(); j++)
3465 if (use->in(j) == phi)
3466 u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j));
3467 }
3468 IdealLoopTree *u_loop = get_loop(u_ctrl);
3469 // Look for loop-invariant use
3470 if (u_loop == loop) continue;
3471 if (loop->is_member(u_loop)) continue;
3472 // Check that use is live out the bottom. Assuming the trip-counter
3473 // update is right at the bottom, uses of of the loop middle are ok.
3474 if (dom_lca(exit, u_ctrl) != exit) continue;
3475 // Hit! Refactor use to use the post-incremented tripcounter.
3476 // Compute a post-increment tripcounter.
3477 Node* c = exit;
3478 if (cl->is_strip_mined()) {
3479 IdealLoopTree* outer_loop = get_loop(cl->outer_loop());
3480 if (!outer_loop->is_member(u_loop)) {
3481 c = cl->outer_loop_exit();
3482 }
3483 }
3484 Node *opaq = new Opaque2Node(C, cle->incr());
3485 register_new_node(opaq, c);
3486 Node *neg_stride = _igvn.intcon(-cle->stride_con());
3487 set_ctrl(neg_stride, C->root());
3488 Node *post = new AddINode(opaq, neg_stride);
3489 register_new_node(post, c);
3490 _igvn.rehash_node_delayed(use);
3491 for (uint j = 1; j < use->req(); j++) {
3492 if (use->in(j) == phi)
3493 use->set_req(j, post);
3494 }
3495 // Since DU info changed, rerun loop
3496 progress = true;
3497 break;
3498 }
3499 }
3500
3501 }