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 "utilities/macros.hpp"
44 #if INCLUDE_ZGC
45 #include "gc/z/c2/zBarrierSetC2.hpp"
46 #endif
47
48 //=============================================================================
49 //------------------------------split_thru_phi---------------------------------
50 // Split Node 'n' through merge point if there is enough win.
51 Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) {
52 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) {
53 // ConvI2L may have type information on it which is unsafe to push up
54 // so disable this for now
55 return NULL;
56 }
57
58 // Splitting range check CastIIs through a loop induction Phi can
59 // cause new Phis to be created that are left unrelated to the loop
60 // induction Phi and prevent optimizations (vectorization)
61 if (n->Opcode() == Op_CastII && n->as_CastII()->has_range_check() &&
62 region->is_CountedLoop() && n->in(1) == region->as_CountedLoop()->phi()) {
63 return NULL;
64 }
65
66 int wins = 0;
67 assert(!n->is_CFG(), "");
68 assert(region->is_Region(), "");
69
70 const Type* type = n->bottom_type();
71 const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr();
72 Node *phi;
73 if (t_oop != NULL && t_oop->is_known_instance_field()) {
74 int iid = t_oop->instance_id();
75 int index = C->get_alias_index(t_oop);
76 int offset = t_oop->offset();
77 phi = new PhiNode(region, type, NULL, iid, index, offset);
78 } else {
79 phi = PhiNode::make_blank(region, n);
80 }
81 uint old_unique = C->unique();
82 for (uint i = 1; i < region->req(); i++) {
83 Node *x;
84 Node* the_clone = NULL;
85 if (region->in(i) == C->top()) {
86 x = C->top(); // Dead path? Use a dead data op
87 } else {
88 x = n->clone(); // Else clone up the data op
89 the_clone = x; // Remember for possible deletion.
90 // Alter data node to use pre-phi inputs
91 if (n->in(0) == region)
92 x->set_req( 0, region->in(i) );
93 for (uint j = 1; j < n->req(); j++) {
94 Node *in = n->in(j);
95 if (in->is_Phi() && in->in(0) == region)
96 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone
97 }
98 }
99 // Check for a 'win' on some paths
100 const Type *t = x->Value(&_igvn);
101
102 bool singleton = t->singleton();
103
104 // A TOP singleton indicates that there are no possible values incoming
105 // along a particular edge. In most cases, this is OK, and the Phi will
106 // be eliminated later in an Ideal call. However, we can't allow this to
107 // happen if the singleton occurs on loop entry, as the elimination of
108 // the PhiNode may cause the resulting node to migrate back to a previous
109 // loop iteration.
110 if (singleton && t == Type::TOP) {
111 // Is_Loop() == false does not confirm the absence of a loop (e.g., an
112 // irreducible loop may not be indicated by an affirmative is_Loop());
113 // therefore, the only top we can split thru a phi is on a backedge of
114 // a loop.
115 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
116 }
117
118 if (singleton) {
119 wins++;
120 x = ((PhaseGVN&)_igvn).makecon(t);
121 } else {
122 // We now call Identity to try to simplify the cloned node.
123 // Note that some Identity methods call phase->type(this).
124 // Make sure that the type array is big enough for
125 // our new node, even though we may throw the node away.
126 // (Note: This tweaking with igvn only works because x is a new node.)
127 _igvn.set_type(x, t);
128 // If x is a TypeNode, capture any more-precise type permanently into Node
129 // otherwise it will be not updated during igvn->transform since
130 // igvn->type(x) is set to x->Value() already.
131 x->raise_bottom_type(t);
132 Node *y = _igvn.apply_identity(x);
133 if (y != x) {
134 wins++;
135 x = y;
136 } else {
137 y = _igvn.hash_find(x);
138 if (y) {
139 wins++;
140 x = y;
141 } else {
142 // Else x is a new node we are keeping
143 // We do not need register_new_node_with_optimizer
144 // because set_type has already been called.
145 _igvn._worklist.push(x);
146 }
147 }
148 }
149 if (x != the_clone && the_clone != NULL)
150 _igvn.remove_dead_node(the_clone);
151 phi->set_req( i, x );
152 }
153 // Too few wins?
154 if (wins <= policy) {
155 _igvn.remove_dead_node(phi);
156 return NULL;
157 }
158
159 // Record Phi
160 register_new_node( phi, region );
161
162 for (uint i2 = 1; i2 < phi->req(); i2++) {
163 Node *x = phi->in(i2);
164 // If we commoned up the cloned 'x' with another existing Node,
165 // the existing Node picks up a new use. We need to make the
166 // existing Node occur higher up so it dominates its uses.
167 Node *old_ctrl;
168 IdealLoopTree *old_loop;
169
170 if (x->is_Con()) {
171 // Constant's control is always root.
172 set_ctrl(x, C->root());
173 continue;
174 }
175 // The occasional new node
176 if (x->_idx >= old_unique) { // Found a new, unplaced node?
177 old_ctrl = NULL;
178 old_loop = NULL; // Not in any prior loop
179 } else {
180 old_ctrl = get_ctrl(x);
181 old_loop = get_loop(old_ctrl); // Get prior loop
182 }
183 // New late point must dominate new use
184 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2));
185 if (new_ctrl == old_ctrl) // Nothing is changed
186 continue;
187
188 IdealLoopTree *new_loop = get_loop(new_ctrl);
189
190 // Don't move x into a loop if its uses are
191 // outside of loop. Otherwise x will be cloned
192 // for each use outside of this loop.
193 IdealLoopTree *use_loop = get_loop(region);
194 if (!new_loop->is_member(use_loop) &&
195 (old_loop == NULL || !new_loop->is_member(old_loop))) {
196 // Take early control, later control will be recalculated
197 // during next iteration of loop optimizations.
198 new_ctrl = get_early_ctrl(x);
199 new_loop = get_loop(new_ctrl);
200 }
201 // Set new location
202 set_ctrl(x, new_ctrl);
203 // If changing loop bodies, see if we need to collect into new body
204 if (old_loop != new_loop) {
205 if (old_loop && !old_loop->_child)
206 old_loop->_body.yank(x);
207 if (!new_loop->_child)
208 new_loop->_body.push(x); // Collect body info
209 }
210 }
211
212 return phi;
213 }
214
215 //------------------------------dominated_by------------------------------------
216 // Replace the dominated test with an obvious true or false. Place it on the
217 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the
218 // live path up to the dominating control.
219 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff, bool flip, bool exclude_loop_predicate ) {
220 if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); }
221
222 // prevdom is the dominating projection of the dominating test.
223 assert( iff->is_If(), "" );
224 assert(iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd || iff->Opcode() == Op_RangeCheck, "Check this code when new subtype is added");
225 int pop = prevdom->Opcode();
226 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" );
227 if (flip) {
228 if (pop == Op_IfTrue)
229 pop = Op_IfFalse;
230 else
231 pop = Op_IfTrue;
232 }
233 // 'con' is set to true or false to kill the dominated test.
234 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO);
235 set_ctrl(con, C->root()); // Constant gets a new use
236 // Hack the dominated test
237 _igvn.replace_input_of(iff, 1, con);
238
239 // If I dont have a reachable TRUE and FALSE path following the IfNode then
240 // I can assume this path reaches an infinite loop. In this case it's not
241 // important to optimize the data Nodes - either the whole compilation will
242 // be tossed or this path (and all data Nodes) will go dead.
243 if (iff->outcnt() != 2) return;
244
245 // Make control-dependent data Nodes on the live path (path that will remain
246 // once the dominated IF is removed) become control-dependent on the
247 // dominating projection.
248 Node* dp = iff->as_If()->proj_out_or_null(pop == Op_IfTrue);
249
250 // Loop predicates may have depending checks which should not
251 // be skipped. For example, range check predicate has two checks
252 // for lower and upper bounds.
253 if (dp == NULL)
254 return;
255
256 ProjNode* dp_proj = dp->as_Proj();
257 ProjNode* unc_proj = iff->as_If()->proj_out(1 - dp_proj->_con)->as_Proj();
258 if (exclude_loop_predicate &&
259 (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) != NULL ||
260 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_profile_predicate) != NULL ||
261 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check) != NULL)) {
262 // If this is a range check (IfNode::is_range_check), do not
263 // reorder because Compile::allow_range_check_smearing might have
264 // changed the check.
265 return; // Let IGVN transformation change control dependence.
266 }
267
268 IdealLoopTree *old_loop = get_loop(dp);
269
270 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
271 Node* cd = dp->fast_out(i); // Control-dependent node
272 if (cd->depends_only_on_test()) {
273 assert(cd->in(0) == dp, "");
274 _igvn.replace_input_of(cd, 0, prevdom);
275 set_early_ctrl(cd);
276 IdealLoopTree *new_loop = get_loop(get_ctrl(cd));
277 if (old_loop != new_loop) {
278 if (!old_loop->_child) old_loop->_body.yank(cd);
279 if (!new_loop->_child) new_loop->_body.push(cd);
280 }
281 --i;
282 --imax;
283 }
284 }
285 }
286
287 //------------------------------has_local_phi_input----------------------------
288 // Return TRUE if 'n' has Phi inputs from its local block and no other
289 // block-local inputs (all non-local-phi inputs come from earlier blocks)
290 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) {
291 Node *n_ctrl = get_ctrl(n);
292 // See if some inputs come from a Phi in this block, or from before
293 // this block.
294 uint i;
295 for( i = 1; i < n->req(); i++ ) {
296 Node *phi = n->in(i);
297 if( phi->is_Phi() && phi->in(0) == n_ctrl )
298 break;
299 }
300 if( i >= n->req() )
301 return NULL; // No Phi inputs; nowhere to clone thru
302
303 // Check for inputs created between 'n' and the Phi input. These
304 // must split as well; they have already been given the chance
305 // (courtesy of a post-order visit) and since they did not we must
306 // recover the 'cost' of splitting them by being very profitable
307 // when splitting 'n'. Since this is unlikely we simply give up.
308 for( i = 1; i < n->req(); i++ ) {
309 Node *m = n->in(i);
310 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) {
311 // We allow the special case of AddP's with no local inputs.
312 // This allows us to split-up address expressions.
313 if (m->is_AddP() &&
314 get_ctrl(m->in(2)) != n_ctrl &&
315 get_ctrl(m->in(3)) != n_ctrl) {
316 // Move the AddP up to dominating point
317 Node* c = find_non_split_ctrl(idom(n_ctrl));
318 if (c->is_OuterStripMinedLoop()) {
319 c->as_Loop()->verify_strip_mined(1);
320 c = c->in(LoopNode::EntryControl);
321 }
322 set_ctrl_and_loop(m, c);
323 continue;
324 }
325 return NULL;
326 }
327 assert(m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control");
328 }
329
330 return n_ctrl;
331 }
332
333 //------------------------------remix_address_expressions----------------------
334 // Rework addressing expressions to get the most loop-invariant stuff
335 // moved out. We'd like to do all associative operators, but it's especially
336 // important (common) to do address expressions.
337 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) {
338 if (!has_ctrl(n)) return NULL;
339 Node *n_ctrl = get_ctrl(n);
340 IdealLoopTree *n_loop = get_loop(n_ctrl);
341
342 // See if 'n' mixes loop-varying and loop-invariant inputs and
343 // itself is loop-varying.
344
345 // Only interested in binary ops (and AddP)
346 if( n->req() < 3 || n->req() > 4 ) return NULL;
347
348 Node *n1_ctrl = get_ctrl(n->in( 1));
349 Node *n2_ctrl = get_ctrl(n->in( 2));
350 Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3));
351 IdealLoopTree *n1_loop = get_loop( n1_ctrl );
352 IdealLoopTree *n2_loop = get_loop( n2_ctrl );
353 IdealLoopTree *n3_loop = get_loop( n3_ctrl );
354
355 // Does one of my inputs spin in a tighter loop than self?
356 if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) ||
357 (n_loop->is_member( n2_loop ) && n_loop != n2_loop) ||
358 (n_loop->is_member( n3_loop ) && n_loop != n3_loop) )
359 return NULL; // Leave well enough alone
360
361 // Is at least one of my inputs loop-invariant?
362 if( n1_loop == n_loop &&
363 n2_loop == n_loop &&
364 n3_loop == n_loop )
365 return NULL; // No loop-invariant inputs
366
367
368 int n_op = n->Opcode();
369
370 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2).
371 if( n_op == Op_LShiftI ) {
372 // Scale is loop invariant
373 Node *scale = n->in(2);
374 Node *scale_ctrl = get_ctrl(scale);
375 IdealLoopTree *scale_loop = get_loop(scale_ctrl );
376 if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) )
377 return NULL;
378 const TypeInt *scale_t = scale->bottom_type()->isa_int();
379 if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 )
380 return NULL; // Dont bother with byte/short masking
381 // Add must vary with loop (else shift would be loop-invariant)
382 Node *add = n->in(1);
383 Node *add_ctrl = get_ctrl(add);
384 IdealLoopTree *add_loop = get_loop(add_ctrl);
385 //assert( n_loop == add_loop, "" );
386 if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops
387
388 // Convert I-V into I+ (0-V); same for V-I
389 if( add->Opcode() == Op_SubI &&
390 _igvn.type( add->in(1) ) != TypeInt::ZERO ) {
391 Node *zero = _igvn.intcon(0);
392 set_ctrl(zero, C->root());
393 Node *neg = new SubINode( _igvn.intcon(0), add->in(2) );
394 register_new_node( neg, get_ctrl(add->in(2) ) );
395 add = new AddINode( add->in(1), neg );
396 register_new_node( add, add_ctrl );
397 }
398 if( add->Opcode() != Op_AddI ) return NULL;
399 // See if one add input is loop invariant
400 Node *add_var = add->in(1);
401 Node *add_var_ctrl = get_ctrl(add_var);
402 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
403 Node *add_invar = add->in(2);
404 Node *add_invar_ctrl = get_ctrl(add_invar);
405 IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl );
406 if( add_var_loop == n_loop ) {
407 } else if( add_invar_loop == n_loop ) {
408 // Swap to find the invariant part
409 add_invar = add_var;
410 add_invar_ctrl = add_var_ctrl;
411 add_invar_loop = add_var_loop;
412 add_var = add->in(2);
413 Node *add_var_ctrl = get_ctrl(add_var);
414 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
415 } else // Else neither input is loop invariant
416 return NULL;
417 if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) )
418 return NULL; // No invariant part of the add?
419
420 // Yes! Reshape address expression!
421 Node *inv_scale = new LShiftINode( add_invar, scale );
422 Node *inv_scale_ctrl =
423 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ?
424 add_invar_ctrl : scale_ctrl;
425 register_new_node( inv_scale, inv_scale_ctrl );
426 Node *var_scale = new LShiftINode( add_var, scale );
427 register_new_node( var_scale, n_ctrl );
428 Node *var_add = new AddINode( var_scale, inv_scale );
429 register_new_node( var_add, n_ctrl );
430 _igvn.replace_node( n, var_add );
431 return var_add;
432 }
433
434 // Replace (I+V) with (V+I)
435 if( n_op == Op_AddI ||
436 n_op == Op_AddL ||
437 n_op == Op_AddF ||
438 n_op == Op_AddD ||
439 n_op == Op_MulI ||
440 n_op == Op_MulL ||
441 n_op == Op_MulF ||
442 n_op == Op_MulD ) {
443 if( n2_loop == n_loop ) {
444 assert( n1_loop != n_loop, "" );
445 n->swap_edges(1, 2);
446 }
447 }
448
449 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V),
450 // but not if I2 is a constant.
451 if( n_op == Op_AddP ) {
452 if( n2_loop == n_loop && n3_loop != n_loop ) {
453 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) {
454 Node *n22_ctrl = get_ctrl(n->in(2)->in(2));
455 Node *n23_ctrl = get_ctrl(n->in(2)->in(3));
456 IdealLoopTree *n22loop = get_loop( n22_ctrl );
457 IdealLoopTree *n23_loop = get_loop( n23_ctrl );
458 if( n22loop != n_loop && n22loop->is_member(n_loop) &&
459 n23_loop == n_loop ) {
460 Node *add1 = new AddPNode( n->in(1), n->in(2)->in(2), n->in(3) );
461 // Stuff new AddP in the loop preheader
462 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) );
463 Node *add2 = new AddPNode( n->in(1), add1, n->in(2)->in(3) );
464 register_new_node( add2, n_ctrl );
465 _igvn.replace_node( n, add2 );
466 return add2;
467 }
468 }
469 }
470
471 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V)
472 if (n2_loop != n_loop && n3_loop == n_loop) {
473 if (n->in(3)->Opcode() == Op_AddX) {
474 Node *V = n->in(3)->in(1);
475 Node *I = n->in(3)->in(2);
476 if (is_member(n_loop,get_ctrl(V))) {
477 } else {
478 Node *tmp = V; V = I; I = tmp;
479 }
480 if (!is_member(n_loop,get_ctrl(I))) {
481 Node *add1 = new AddPNode(n->in(1), n->in(2), I);
482 // Stuff new AddP in the loop preheader
483 register_new_node(add1, n_loop->_head->in(LoopNode::EntryControl));
484 Node *add2 = new AddPNode(n->in(1), add1, V);
485 register_new_node(add2, n_ctrl);
486 _igvn.replace_node(n, add2);
487 return add2;
488 }
489 }
490 }
491 }
492
493 return NULL;
494 }
495
496 // Optimize ((in1[2*i] * in2[2*i]) + (in1[2*i+1] * in2[2*i+1]))
497 Node *PhaseIdealLoop::convert_add_to_muladd(Node* n) {
498 assert(n->Opcode() == Op_AddI, "sanity");
499 Node * nn = NULL;
500 Node * in1 = n->in(1);
501 Node * in2 = n->in(2);
502 if (in1->Opcode() == Op_MulI && in2->Opcode() == Op_MulI) {
503 IdealLoopTree* loop_n = get_loop(get_ctrl(n));
504 if (loop_n->_head->as_Loop()->is_valid_counted_loop() &&
505 Matcher::match_rule_supported(Op_MulAddS2I) &&
506 Matcher::match_rule_supported(Op_MulAddVS2VI)) {
507 Node* mul_in1 = in1->in(1);
508 Node* mul_in2 = in1->in(2);
509 Node* mul_in3 = in2->in(1);
510 Node* mul_in4 = in2->in(2);
511 if (mul_in1->Opcode() == Op_LoadS &&
512 mul_in2->Opcode() == Op_LoadS &&
513 mul_in3->Opcode() == Op_LoadS &&
514 mul_in4->Opcode() == Op_LoadS) {
515 IdealLoopTree* loop1 = get_loop(get_ctrl(mul_in1));
516 IdealLoopTree* loop2 = get_loop(get_ctrl(mul_in2));
517 IdealLoopTree* loop3 = get_loop(get_ctrl(mul_in3));
518 IdealLoopTree* loop4 = get_loop(get_ctrl(mul_in4));
519 IdealLoopTree* loop5 = get_loop(get_ctrl(in1));
520 IdealLoopTree* loop6 = get_loop(get_ctrl(in2));
521 // All nodes should be in the same counted loop.
522 if (loop_n == loop1 && loop_n == loop2 && loop_n == loop3 &&
523 loop_n == loop4 && loop_n == loop5 && loop_n == loop6) {
524 Node* adr1 = mul_in1->in(MemNode::Address);
525 Node* adr2 = mul_in2->in(MemNode::Address);
526 Node* adr3 = mul_in3->in(MemNode::Address);
527 Node* adr4 = mul_in4->in(MemNode::Address);
528 if (adr1->is_AddP() && adr2->is_AddP() && adr3->is_AddP() && adr4->is_AddP()) {
529 if ((adr1->in(AddPNode::Base) == adr3->in(AddPNode::Base)) &&
530 (adr2->in(AddPNode::Base) == adr4->in(AddPNode::Base))) {
531 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in3, mul_in4);
532 register_new_node(nn, get_ctrl(n));
533 _igvn.replace_node(n, nn);
534 return nn;
535 } else if ((adr1->in(AddPNode::Base) == adr4->in(AddPNode::Base)) &&
536 (adr2->in(AddPNode::Base) == adr3->in(AddPNode::Base))) {
537 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in4, mul_in3);
538 register_new_node(nn, get_ctrl(n));
539 _igvn.replace_node(n, nn);
540 return nn;
541 }
542 }
543 }
544 }
545 }
546 }
547 return nn;
548 }
549
550 //------------------------------conditional_move-------------------------------
551 // Attempt to replace a Phi with a conditional move. We have some pretty
552 // strict profitability requirements. All Phis at the merge point must
553 // be converted, so we can remove the control flow. We need to limit the
554 // number of c-moves to a small handful. All code that was in the side-arms
555 // of the CFG diamond is now speculatively executed. This code has to be
556 // "cheap enough". We are pretty much limited to CFG diamonds that merge
557 // 1 or 2 items with a total of 1 or 2 ops executed speculatively.
558 Node *PhaseIdealLoop::conditional_move( Node *region ) {
559
560 assert(region->is_Region(), "sanity check");
561 if (region->req() != 3) return NULL;
562
563 // Check for CFG diamond
564 Node *lp = region->in(1);
565 Node *rp = region->in(2);
566 if (!lp || !rp) return NULL;
567 Node *lp_c = lp->in(0);
568 if (lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If()) return NULL;
569 IfNode *iff = lp_c->as_If();
570
571 // Check for ops pinned in an arm of the diamond.
572 // Can't remove the control flow in this case
573 if (lp->outcnt() > 1) return NULL;
574 if (rp->outcnt() > 1) return NULL;
575
576 IdealLoopTree* r_loop = get_loop(region);
577 assert(r_loop == get_loop(iff), "sanity");
578 // Always convert to CMOVE if all results are used only outside this loop.
579 bool used_inside_loop = (r_loop == _ltree_root);
580
581 // Check profitability
582 int cost = 0;
583 int phis = 0;
584 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
585 Node *out = region->fast_out(i);
586 if (!out->is_Phi()) continue; // Ignore other control edges, etc
587 phis++;
588 PhiNode* phi = out->as_Phi();
589 BasicType bt = phi->type()->basic_type();
590 switch (bt) {
591 case T_DOUBLE:
592 case T_FLOAT:
593 if (C->use_cmove()) {
594 continue; //TODO: maybe we want to add some cost
595 }
596 cost += Matcher::float_cmove_cost(); // Could be very expensive
597 break;
598 case T_LONG: {
599 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's
600 }
601 case T_INT: // These all CMOV fine
602 case T_ADDRESS: { // (RawPtr)
603 cost++;
604 break;
605 }
606 case T_NARROWOOP: // Fall through
607 case T_OBJECT: { // Base oops are OK, but not derived oops
608 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr();
609 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a
610 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus
611 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we
612 // have a Phi for the base here that we convert to a CMOVE all is well
613 // and good. But if the base is dead, we'll not make a CMOVE. Later
614 // the allocator will have to produce a base by creating a CMOVE of the
615 // relevant bases. This puts the allocator in the business of
616 // manufacturing expensive instructions, generally a bad plan.
617 // Just Say No to Conditionally-Moved Derived Pointers.
618 if (tp && tp->offset() != 0)
619 return NULL;
620 cost++;
621 break;
622 }
623 default:
624 return NULL; // In particular, can't do memory or I/O
625 }
626 // Add in cost any speculative ops
627 for (uint j = 1; j < region->req(); j++) {
628 Node *proj = region->in(j);
629 Node *inp = phi->in(j);
630 if (get_ctrl(inp) == proj) { // Found local op
631 cost++;
632 // Check for a chain of dependent ops; these will all become
633 // speculative in a CMOV.
634 for (uint k = 1; k < inp->req(); k++)
635 if (get_ctrl(inp->in(k)) == proj)
636 cost += ConditionalMoveLimit; // Too much speculative goo
637 }
638 }
639 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode.
640 // This will likely Split-If, a higher-payoff operation.
641 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) {
642 Node* use = phi->fast_out(k);
643 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr())
644 cost += ConditionalMoveLimit;
645 // Is there a use inside the loop?
646 // Note: check only basic types since CMoveP is pinned.
647 if (!used_inside_loop && is_java_primitive(bt)) {
648 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use);
649 if (r_loop == u_loop || r_loop->is_member(u_loop)) {
650 used_inside_loop = true;
651 }
652 }
653 }
654 }//for
655 Node* bol = iff->in(1);
656 assert(bol->Opcode() == Op_Bool, "");
657 int cmp_op = bol->in(1)->Opcode();
658 // It is expensive to generate flags from a float compare.
659 // Avoid duplicated float compare.
660 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL;
661
662 float infrequent_prob = PROB_UNLIKELY_MAG(3);
663 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop.
664 if (used_inside_loop) {
665 if (cost >= ConditionalMoveLimit) return NULL; // Too much goo
666
667 // BlockLayoutByFrequency optimization moves infrequent branch
668 // from hot path. No point in CMOV'ing in such case (110 is used
669 // instead of 100 to take into account not exactness of float value).
670 if (BlockLayoutByFrequency) {
671 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f);
672 }
673 }
674 // Check for highly predictable branch. No point in CMOV'ing if
675 // we are going to predict accurately all the time.
676 if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) {
677 //keep going
678 } else if (iff->_prob < infrequent_prob ||
679 iff->_prob > (1.0f - infrequent_prob))
680 return NULL;
681
682 // --------------
683 // Now replace all Phis with CMOV's
684 Node *cmov_ctrl = iff->in(0);
685 uint flip = (lp->Opcode() == Op_IfTrue);
686 Node_List wq;
687 while (1) {
688 PhiNode* phi = NULL;
689 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
690 Node *out = region->fast_out(i);
691 if (out->is_Phi()) {
692 phi = out->as_Phi();
693 break;
694 }
695 }
696 if (phi == NULL) break;
697 if (PrintOpto && VerifyLoopOptimizations) { tty->print_cr("CMOV"); }
698 // Move speculative ops
699 wq.push(phi);
700 while (wq.size() > 0) {
701 Node *n = wq.pop();
702 for (uint j = 1; j < n->req(); j++) {
703 Node* m = n->in(j);
704 if (m != NULL && !is_dominator(get_ctrl(m), cmov_ctrl)) {
705 #ifndef PRODUCT
706 if (PrintOpto && VerifyLoopOptimizations) {
707 tty->print(" speculate: ");
708 m->dump();
709 }
710 #endif
711 set_ctrl(m, cmov_ctrl);
712 wq.push(m);
713 }
714 }
715 }
716 Node *cmov = CMoveNode::make(cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi));
717 register_new_node( cmov, cmov_ctrl );
718 _igvn.replace_node( phi, cmov );
719 #ifndef PRODUCT
720 if (TraceLoopOpts) {
721 tty->print("CMOV ");
722 r_loop->dump_head();
723 if (Verbose) {
724 bol->in(1)->dump(1);
725 cmov->dump(1);
726 }
727 }
728 if (VerifyLoopOptimizations) verify();
729 #endif
730 }
731
732 // The useless CFG diamond will fold up later; see the optimization in
733 // RegionNode::Ideal.
734 _igvn._worklist.push(region);
735
736 return iff->in(1);
737 }
738
739 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) {
740 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
741 Node* u = m->fast_out(i);
742 if (u->is_CFG()) {
743 if (u->Opcode() == Op_NeverBranch) {
744 u = ((NeverBranchNode*)u)->proj_out(0);
745 enqueue_cfg_uses(u, wq);
746 } else {
747 wq.push(u);
748 }
749 }
750 }
751 }
752
753 // Try moving a store out of a loop, right before the loop
754 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) {
755 // Store has to be first in the loop body
756 IdealLoopTree *n_loop = get_loop(n_ctrl);
757 if (n->is_Store() && n_loop != _ltree_root &&
758 n_loop->is_loop() && n_loop->_head->is_Loop() &&
759 n->in(0) != NULL) {
760 Node* address = n->in(MemNode::Address);
761 Node* value = n->in(MemNode::ValueIn);
762 Node* mem = n->in(MemNode::Memory);
763 IdealLoopTree* address_loop = get_loop(get_ctrl(address));
764 IdealLoopTree* value_loop = get_loop(get_ctrl(value));
765
766 // - address and value must be loop invariant
767 // - memory must be a memory Phi for the loop
768 // - Store must be the only store on this memory slice in the
769 // loop: if there's another store following this one then value
770 // written at iteration i by the second store could be overwritten
771 // at iteration i+n by the first store: it's not safe to move the
772 // first store out of the loop
773 // - nothing must observe the memory Phi: it guarantees no read
774 // before the store, we are also guaranteed the store post
775 // dominates the loop head (ignoring a possible early
776 // exit). Otherwise there would be extra Phi involved between the
777 // loop's Phi and the store.
778 // - there must be no early exit from the loop before the Store
779 // (such an exit most of the time would be an extra use of the
780 // memory Phi but sometimes is a bottom memory Phi that takes the
781 // store as input).
782
783 if (!n_loop->is_member(address_loop) &&
784 !n_loop->is_member(value_loop) &&
785 mem->is_Phi() && mem->in(0) == n_loop->_head &&
786 mem->outcnt() == 1 &&
787 mem->in(LoopNode::LoopBackControl) == n) {
788
789 assert(n_loop->_tail != NULL, "need a tail");
790 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop");
791
792 // Verify that there's no early exit of the loop before the store.
793 bool ctrl_ok = false;
794 {
795 // Follow control from loop head until n, we exit the loop or
796 // we reach the tail
797 ResourceMark rm;
798 Unique_Node_List wq;
799 wq.push(n_loop->_head);
800
801 for (uint next = 0; next < wq.size(); ++next) {
802 Node *m = wq.at(next);
803 if (m == n->in(0)) {
804 ctrl_ok = true;
805 continue;
806 }
807 assert(!has_ctrl(m), "should be CFG");
808 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) {
809 ctrl_ok = false;
810 break;
811 }
812 enqueue_cfg_uses(m, wq);
813 if (wq.size() > 10) {
814 ctrl_ok = false;
815 break;
816 }
817 }
818 }
819 if (ctrl_ok) {
820 // move the Store
821 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem);
822 _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl));
823 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl));
824 // Disconnect the phi now. An empty phi can confuse other
825 // optimizations in this pass of loop opts.
826 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl));
827 n_loop->_body.yank(mem);
828
829 set_ctrl_and_loop(n, n->in(0));
830
831 return n;
832 }
833 }
834 }
835 return NULL;
836 }
837
838 // Try moving a store out of a loop, right after the loop
839 void PhaseIdealLoop::try_move_store_after_loop(Node* n) {
840 if (n->is_Store() && n->in(0) != NULL) {
841 Node *n_ctrl = get_ctrl(n);
842 IdealLoopTree *n_loop = get_loop(n_ctrl);
843 // Store must be in a loop
844 if (n_loop != _ltree_root && !n_loop->_irreducible) {
845 Node* address = n->in(MemNode::Address);
846 Node* value = n->in(MemNode::ValueIn);
847 IdealLoopTree* address_loop = get_loop(get_ctrl(address));
848 // address must be loop invariant
849 if (!n_loop->is_member(address_loop)) {
850 // Store must be last on this memory slice in the loop and
851 // nothing in the loop must observe it
852 Node* phi = NULL;
853 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
854 Node* u = n->fast_out(i);
855 if (has_ctrl(u)) { // control use?
856 IdealLoopTree *u_loop = get_loop(get_ctrl(u));
857 if (!n_loop->is_member(u_loop)) {
858 continue;
859 }
860 if (u->is_Phi() && u->in(0) == n_loop->_head) {
861 assert(_igvn.type(u) == Type::MEMORY, "bad phi");
862 // multiple phis on the same slice are possible
863 if (phi != NULL) {
864 return;
865 }
866 phi = u;
867 continue;
868 }
869 }
870 return;
871 }
872 if (phi != NULL) {
873 // Nothing in the loop before the store (next iteration)
874 // must observe the stored value
875 bool mem_ok = true;
876 {
877 ResourceMark rm;
878 Unique_Node_List wq;
879 wq.push(phi);
880 for (uint next = 0; next < wq.size() && mem_ok; ++next) {
881 Node *m = wq.at(next);
882 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) {
883 Node* u = m->fast_out(i);
884 if (u->is_Store() || u->is_Phi()) {
885 if (u != n) {
886 wq.push(u);
887 mem_ok = (wq.size() <= 10);
888 }
889 } else {
890 mem_ok = false;
891 break;
892 }
893 }
894 }
895 }
896 if (mem_ok) {
897 // Move the store out of the loop if the LCA of all
898 // users (except for the phi) is outside the loop.
899 Node* hook = new Node(1);
900 _igvn.rehash_node_delayed(phi);
901 int count = phi->replace_edge(n, hook);
902 assert(count > 0, "inconsistent phi");
903
904 // Compute latest point this store can go
905 Node* lca = get_late_ctrl(n, get_ctrl(n));
906 if (n_loop->is_member(get_loop(lca))) {
907 // LCA is in the loop - bail out
908 _igvn.replace_node(hook, n);
909 return;
910 }
911 #ifdef ASSERT
912 if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) {
913 assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined");
914 n_loop->_head->as_Loop()->verify_strip_mined(1);
915 Node* outer = n_loop->_head->as_CountedLoop()->outer_loop();
916 IdealLoopTree* outer_loop = get_loop(outer);
917 assert(n_loop->_parent == outer_loop, "broken loop tree");
918 assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state");
919 }
920 #endif
921
922 // Move store out of the loop
923 _igvn.replace_node(hook, n->in(MemNode::Memory));
924 _igvn.replace_input_of(n, 0, lca);
925 set_ctrl_and_loop(n, lca);
926
927 // Disconnect the phi now. An empty phi can confuse other
928 // optimizations in this pass of loop opts..
929 if (phi->in(LoopNode::LoopBackControl) == phi) {
930 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl));
931 n_loop->_body.yank(phi);
932 }
933 }
934 }
935 }
936 }
937 }
938 }
939
940 //------------------------------split_if_with_blocks_pre-----------------------
941 // Do the real work in a non-recursive function. Data nodes want to be
942 // cloned in the pre-order so they can feed each other nicely.
943 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
944 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
945 Node* bs_res = bs->split_if_pre(this, n);
946 if (bs_res != NULL) {
947 return bs_res;
948 }
949 // Cloning these guys is unlikely to win
950 int n_op = n->Opcode();
951 if( n_op == Op_MergeMem ) return n;
952 if( n->is_Proj() ) return n;
953 // Do not clone-up CmpFXXX variations, as these are always
954 // followed by a CmpI
955 if( n->is_Cmp() ) return n;
956 // Attempt to use a conditional move instead of a phi/branch
957 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) {
958 Node *cmov = conditional_move( n );
959 if( cmov ) return cmov;
960 }
961 if( n->is_CFG() || n->is_LoadStore() )
962 return n;
963 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd
964 n_op == Op_Opaque2 ) {
965 if( !C->major_progress() ) // If chance of no more loop opts...
966 _igvn._worklist.push(n); // maybe we'll remove them
967 return n;
968 }
969
970 if( n->is_Con() ) return n; // No cloning for Con nodes
971
972 Node *n_ctrl = get_ctrl(n);
973 if( !n_ctrl ) return n; // Dead node
974
975 Node* res = try_move_store_before_loop(n, n_ctrl);
976 if (res != NULL) {
977 return n;
978 }
979
980 // Attempt to remix address expressions for loop invariants
981 Node *m = remix_address_expressions( n );
982 if( m ) return m;
983
984 if (n_op == Op_AddI) {
985 Node *nn = convert_add_to_muladd( n );
986 if ( nn ) return nn;
987 }
988
989 if (n->is_ConstraintCast()) {
990 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this);
991 // ConstraintCastNode::dominating_cast() uses node control input to determine domination.
992 // Node control inputs don't necessarily agree with loop control info (due to
993 // transformations happened in between), thus additional dominance check is needed
994 // to keep loop info valid.
995 if (dom_cast != NULL && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) {
996 _igvn.replace_node(n, dom_cast);
997 return dom_cast;
998 }
999 }
1000
1001 // Determine if the Node has inputs from some local Phi.
1002 // Returns the block to clone thru.
1003 Node *n_blk = has_local_phi_input( n );
1004 if( !n_blk ) return n;
1005
1006 // Do not clone the trip counter through on a CountedLoop
1007 // (messes up the canonical shape).
1008 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n;
1009
1010 // Check for having no control input; not pinned. Allow
1011 // dominating control.
1012 if (n->in(0)) {
1013 Node *dom = idom(n_blk);
1014 if (dom_lca(n->in(0), dom) != n->in(0)) {
1015 return n;
1016 }
1017 }
1018 // Policy: when is it profitable. You must get more wins than
1019 // policy before it is considered profitable. Policy is usually 0,
1020 // so 1 win is considered profitable. Big merges will require big
1021 // cloning, so get a larger policy.
1022 int policy = n_blk->req() >> 2;
1023
1024 // If the loop is a candidate for range check elimination,
1025 // delay splitting through it's phi until a later loop optimization
1026 if (n_blk->is_CountedLoop()) {
1027 IdealLoopTree *lp = get_loop(n_blk);
1028 if (lp && lp->_rce_candidate) {
1029 return n;
1030 }
1031 }
1032
1033 if (must_throttle_split_if()) return n;
1034
1035 // Split 'n' through the merge point if it is profitable
1036 Node *phi = split_thru_phi( n, n_blk, policy );
1037 if (!phi) return n;
1038
1039 // Found a Phi to split thru!
1040 // Replace 'n' with the new phi
1041 _igvn.replace_node( n, phi );
1042 // Moved a load around the loop, 'en-registering' something.
1043 if (n_blk->is_Loop() && n->is_Load() &&
1044 !phi->in(LoopNode::LoopBackControl)->is_Load())
1045 C->set_major_progress();
1046
1047 return phi;
1048 }
1049
1050 static bool merge_point_too_heavy(Compile* C, Node* region) {
1051 // Bail out if the region and its phis have too many users.
1052 int weight = 0;
1053 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1054 weight += region->fast_out(i)->outcnt();
1055 }
1056 int nodes_left = C->max_node_limit() - C->live_nodes();
1057 if (weight * 8 > nodes_left) {
1058 if (PrintOpto) {
1059 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight);
1060 }
1061 return true;
1062 } else {
1063 return false;
1064 }
1065 }
1066
1067 static bool merge_point_safe(Node* region) {
1068 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode
1069 // having a PhiNode input. This sidesteps the dangerous case where the split
1070 // ConvI2LNode may become TOP if the input Value() does not
1071 // overlap the ConvI2L range, leaving a node which may not dominate its
1072 // uses.
1073 // A better fix for this problem can be found in the BugTraq entry, but
1074 // expediency for Mantis demands this hack.
1075 // 6855164: If the merge point has a FastLockNode with a PhiNode input, we stop
1076 // split_if_with_blocks from splitting a block because we could not move around
1077 // the FastLockNode.
1078 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1079 Node* n = region->fast_out(i);
1080 if (n->is_Phi()) {
1081 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1082 Node* m = n->fast_out(j);
1083 if (m->is_FastLock())
1084 return false;
1085 #ifdef _LP64
1086 if (m->Opcode() == Op_ConvI2L)
1087 return false;
1088 if (m->is_CastII() && m->isa_CastII()->has_range_check()) {
1089 return false;
1090 }
1091 #endif
1092 }
1093 }
1094 }
1095 return true;
1096 }
1097
1098
1099 //------------------------------place_near_use---------------------------------
1100 // Place some computation next to use but not inside inner loops.
1101 // For inner loop uses move it to the preheader area.
1102 Node *PhaseIdealLoop::place_near_use(Node *useblock) const {
1103 IdealLoopTree *u_loop = get_loop( useblock );
1104 if (u_loop->_irreducible) {
1105 return useblock;
1106 }
1107 if (u_loop->_child) {
1108 if (useblock == u_loop->_head && u_loop->_head->is_OuterStripMinedLoop()) {
1109 return u_loop->_head->in(LoopNode::EntryControl);
1110 }
1111 return useblock;
1112 }
1113 return u_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl);
1114 }
1115
1116
1117 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) {
1118 if (!n->is_If() || n->is_CountedLoopEnd()) {
1119 return false;
1120 }
1121 if (!n->in(0)->is_Region()) {
1122 return false;
1123 }
1124 Node* region = n->in(0);
1125 Node* dom = idom(region);
1126 if (!dom->is_If() || dom->in(1) != n->in(1)) {
1127 return false;
1128 }
1129 IfNode* dom_if = dom->as_If();
1130 Node* proj_true = dom_if->proj_out(1);
1131 Node* proj_false = dom_if->proj_out(0);
1132
1133 for (uint i = 1; i < region->req(); i++) {
1134 if (is_dominator(proj_true, region->in(i))) {
1135 continue;
1136 }
1137 if (is_dominator(proj_false, region->in(i))) {
1138 continue;
1139 }
1140 return false;
1141 }
1142
1143 return true;
1144 }
1145
1146
1147 bool PhaseIdealLoop::can_split_if(Node* n_ctrl) {
1148 if (must_throttle_split_if()) {
1149 return false;
1150 }
1151
1152 // Do not do 'split-if' if irreducible loops are present.
1153 if (_has_irreducible_loops) {
1154 return false;
1155 }
1156
1157 if (merge_point_too_heavy(C, n_ctrl)) {
1158 return false;
1159 }
1160
1161 // Do not do 'split-if' if some paths are dead. First do dead code
1162 // elimination and then see if its still profitable.
1163 for (uint i = 1; i < n_ctrl->req(); i++) {
1164 if (n_ctrl->in(i) == C->top()) {
1165 return false;
1166 }
1167 }
1168
1169 // If trying to do a 'Split-If' at the loop head, it is only
1170 // profitable if the cmp folds up on BOTH paths. Otherwise we
1171 // risk peeling a loop forever.
1172
1173 // CNC - Disabled for now. Requires careful handling of loop
1174 // body selection for the cloned code. Also, make sure we check
1175 // for any input path not being in the same loop as n_ctrl. For
1176 // irreducible loops we cannot check for 'n_ctrl->is_Loop()'
1177 // because the alternative loop entry points won't be converted
1178 // into LoopNodes.
1179 IdealLoopTree *n_loop = get_loop(n_ctrl);
1180 for (uint j = 1; j < n_ctrl->req(); j++) {
1181 if (get_loop(n_ctrl->in(j)) != n_loop) {
1182 return false;
1183 }
1184 }
1185
1186 // Check for safety of the merge point.
1187 if (!merge_point_safe(n_ctrl)) {
1188 return false;
1189 }
1190
1191 return true;
1192 }
1193
1194 //------------------------------split_if_with_blocks_post----------------------
1195 // Do the real work in a non-recursive function. CFG hackery wants to be
1196 // in the post-order, so it can dirty the I-DOM info and not use the dirtied
1197 // info.
1198 void PhaseIdealLoop::split_if_with_blocks_post(Node *n, bool last_round) {
1199
1200 // Cloning Cmp through Phi's involves the split-if transform.
1201 // FastLock is not used by an If
1202 if (n->is_Cmp() && !n->is_FastLock() && !last_round) {
1203 Node *n_ctrl = get_ctrl(n);
1204 // Determine if the Node has inputs from some local Phi.
1205 // Returns the block to clone thru.
1206 Node *n_blk = has_local_phi_input(n);
1207 if (n_blk != n_ctrl) {
1208 return;
1209 }
1210
1211 if (!can_split_if(n_ctrl)) {
1212 return;
1213 }
1214
1215 if (n->outcnt() != 1) {
1216 return; // Multiple bool's from 1 compare?
1217 }
1218 Node *bol = n->unique_out();
1219 assert(bol->is_Bool(), "expect a bool here");
1220 if (bol->outcnt() != 1) {
1221 return;// Multiple branches from 1 compare?
1222 }
1223 Node *iff = bol->unique_out();
1224
1225 // Check some safety conditions
1226 if (iff->is_If()) { // Classic split-if?
1227 if (iff->in(0) != n_ctrl) {
1228 return; // Compare must be in same blk as if
1229 }
1230 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE
1231 // Can't split CMove with different control edge.
1232 if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) {
1233 return;
1234 }
1235 if (get_ctrl(iff->in(2)) == n_ctrl ||
1236 get_ctrl(iff->in(3)) == n_ctrl) {
1237 return; // Inputs not yet split-up
1238 }
1239 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) {
1240 return; // Loop-invar test gates loop-varying CMOVE
1241 }
1242 } else {
1243 return; // some other kind of node, such as an Allocate
1244 }
1245
1246 // When is split-if profitable? Every 'win' on means some control flow
1247 // goes dead, so it's almost always a win.
1248 int policy = 0;
1249 // Split compare 'n' through the merge point if it is profitable
1250 Node *phi = split_thru_phi( n, n_ctrl, policy);
1251 if (!phi) {
1252 return;
1253 }
1254
1255 // Found a Phi to split thru!
1256 // Replace 'n' with the new phi
1257 _igvn.replace_node(n, phi);
1258
1259 // Now split the bool up thru the phi
1260 Node *bolphi = split_thru_phi(bol, n_ctrl, -1);
1261 guarantee(bolphi != NULL, "null boolean phi node");
1262
1263 _igvn.replace_node(bol, bolphi);
1264 assert(iff->in(1) == bolphi, "");
1265
1266 if (bolphi->Value(&_igvn)->singleton()) {
1267 return;
1268 }
1269
1270 // Conditional-move? Must split up now
1271 if (!iff->is_If()) {
1272 Node *cmovphi = split_thru_phi(iff, n_ctrl, -1);
1273 _igvn.replace_node(iff, cmovphi);
1274 return;
1275 }
1276
1277 // Now split the IF
1278 do_split_if(iff);
1279 return;
1280 }
1281
1282 // Two identical ifs back to back can be merged
1283 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) {
1284 Node *n_ctrl = n->in(0);
1285 PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1));
1286 IfNode* dom_if = idom(n_ctrl)->as_If();
1287 Node* proj_true = dom_if->proj_out(1);
1288 Node* proj_false = dom_if->proj_out(0);
1289 Node* con_true = _igvn.makecon(TypeInt::ONE);
1290 Node* con_false = _igvn.makecon(TypeInt::ZERO);
1291
1292 for (uint i = 1; i < n_ctrl->req(); i++) {
1293 if (is_dominator(proj_true, n_ctrl->in(i))) {
1294 bolphi->init_req(i, con_true);
1295 } else {
1296 assert(is_dominator(proj_false, n_ctrl->in(i)), "bad if");
1297 bolphi->init_req(i, con_false);
1298 }
1299 }
1300 register_new_node(bolphi, n_ctrl);
1301 _igvn.replace_input_of(n, 1, bolphi);
1302
1303 // Now split the IF
1304 do_split_if(n);
1305 return;
1306 }
1307
1308 // Check for an IF ready to split; one that has its
1309 // condition codes input coming from a Phi at the block start.
1310 int n_op = n->Opcode();
1311
1312 // Check for an IF being dominated by another IF same test
1313 if (n_op == Op_If ||
1314 n_op == Op_RangeCheck) {
1315 Node *bol = n->in(1);
1316 uint max = bol->outcnt();
1317 // Check for same test used more than once?
1318 if (max > 1 && bol->is_Bool()) {
1319 // Search up IDOMs to see if this IF is dominated.
1320 Node *cutoff = get_ctrl(bol);
1321
1322 // Now search up IDOMs till cutoff, looking for a dominating test
1323 Node *prevdom = n;
1324 Node *dom = idom(prevdom);
1325 while (dom != cutoff) {
1326 if (dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom) {
1327 // Replace the dominated test with an obvious true or false.
1328 // Place it on the IGVN worklist for later cleanup.
1329 C->set_major_progress();
1330 dominated_by(prevdom, n, false, true);
1331 #ifndef PRODUCT
1332 if( VerifyLoopOptimizations ) verify();
1333 #endif
1334 return;
1335 }
1336 prevdom = dom;
1337 dom = idom(prevdom);
1338 }
1339 }
1340 }
1341
1342 // See if a shared loop-varying computation has no loop-varying uses.
1343 // Happens if something is only used for JVM state in uncommon trap exits,
1344 // like various versions of induction variable+offset. Clone the
1345 // computation per usage to allow it to sink out of the loop.
1346 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about)
1347 Node *n_ctrl = get_ctrl(n);
1348 IdealLoopTree *n_loop = get_loop(n_ctrl);
1349 if( n_loop != _ltree_root ) {
1350 DUIterator_Fast imax, i = n->fast_outs(imax);
1351 for (; i < imax; i++) {
1352 Node* u = n->fast_out(i);
1353 if( !has_ctrl(u) ) break; // Found control user
1354 IdealLoopTree *u_loop = get_loop(get_ctrl(u));
1355 if( u_loop == n_loop ) break; // Found loop-varying use
1356 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop
1357 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003
1358 }
1359 bool did_break = (i < imax); // Did we break out of the previous loop?
1360 if (!did_break && n->outcnt() > 1) { // All uses in outer loops!
1361 Node *late_load_ctrl = NULL;
1362 if (n->is_Load()) {
1363 // If n is a load, get and save the result from get_late_ctrl(),
1364 // to be later used in calculating the control for n's clones.
1365 clear_dom_lca_tags();
1366 late_load_ctrl = get_late_ctrl(n, n_ctrl);
1367 }
1368 // If n is a load, and the late control is the same as the current
1369 // control, then the cloning of n is a pointless exercise, because
1370 // GVN will ensure that we end up where we started.
1371 if (!n->is_Load() || late_load_ctrl != n_ctrl) {
1372 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1373 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) {
1374 Node *u = n->last_out(j); // Clone private computation per use
1375 _igvn.rehash_node_delayed(u);
1376 Node *x = n->clone(); // Clone computation
1377 Node *x_ctrl = NULL;
1378 if( u->is_Phi() ) {
1379 // Replace all uses of normal nodes. Replace Phi uses
1380 // individually, so the separate Nodes can sink down
1381 // different paths.
1382 uint k = 1;
1383 while( u->in(k) != n ) k++;
1384 u->set_req( k, x );
1385 // x goes next to Phi input path
1386 x_ctrl = u->in(0)->in(k);
1387 --j;
1388 } else { // Normal use
1389 // Replace all uses
1390 for( uint k = 0; k < u->req(); k++ ) {
1391 if( u->in(k) == n ) {
1392 u->set_req( k, x );
1393 --j;
1394 }
1395 }
1396 x_ctrl = get_ctrl(u);
1397 }
1398
1399 // Find control for 'x' next to use but not inside inner loops.
1400 // For inner loop uses get the preheader area.
1401 x_ctrl = place_near_use(x_ctrl);
1402
1403 if (bs->sink_node(this, n, x, x_ctrl, n_ctrl)) {
1404 continue;
1405 }
1406
1407 if (n->is_Load()) {
1408 // For loads, add a control edge to a CFG node outside of the loop
1409 // to force them to not combine and return back inside the loop
1410 // during GVN optimization (4641526).
1411 //
1412 // Because we are setting the actual control input, factor in
1413 // the result from get_late_ctrl() so we respect any
1414 // anti-dependences. (6233005).
1415 x_ctrl = dom_lca(late_load_ctrl, x_ctrl);
1416
1417 // Don't allow the control input to be a CFG splitting node.
1418 // Such nodes should only have ProjNodes as outs, e.g. IfNode
1419 // should only have IfTrueNode and IfFalseNode (4985384).
1420 x_ctrl = find_non_split_ctrl(x_ctrl);
1421 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone");
1422
1423 x->set_req(0, x_ctrl);
1424 }
1425 register_new_node(x, x_ctrl);
1426
1427 // Some institutional knowledge is needed here: 'x' is
1428 // yanked because if the optimizer runs GVN on it all the
1429 // cloned x's will common up and undo this optimization and
1430 // be forced back in the loop. This is annoying because it
1431 // makes +VerifyOpto report false-positives on progress. I
1432 // tried setting control edges on the x's to force them to
1433 // not combine, but the matching gets worried when it tries
1434 // to fold a StoreP and an AddP together (as part of an
1435 // address expression) and the AddP and StoreP have
1436 // different controls.
1437 if (!x->is_Load() && !x->is_DecodeNarrowPtr()) _igvn._worklist.yank(x);
1438 }
1439 _igvn.remove_dead_node(n);
1440 }
1441 }
1442 }
1443 }
1444
1445 try_move_store_after_loop(n);
1446
1447 // Check for Opaque2's who's loop has disappeared - who's input is in the
1448 // same loop nest as their output. Remove 'em, they are no longer useful.
1449 if( n_op == Op_Opaque2 &&
1450 n->in(1) != NULL &&
1451 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) {
1452 _igvn.replace_node( n, n->in(1) );
1453 }
1454
1455 #if INCLUDE_ZGC
1456 if (UseZGC) {
1457 ZBarrierSetC2::loop_optimize_gc_barrier(this, n, last_round);
1458 }
1459 #endif
1460 }
1461
1462 //------------------------------split_if_with_blocks---------------------------
1463 // Check for aggressive application of 'split-if' optimization,
1464 // using basic block level info.
1465 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack, bool last_round) {
1466 Node* root = C->root();
1467 visited.set(root->_idx); // first, mark root as visited
1468 // Do pre-visit work for root
1469 Node* n = split_if_with_blocks_pre(root);
1470 uint cnt = n->outcnt();
1471 uint i = 0;
1472
1473 while (true) {
1474 // Visit all children
1475 if (i < cnt) {
1476 Node* use = n->raw_out(i);
1477 ++i;
1478 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
1479 // Now do pre-visit work for this use
1480 use = split_if_with_blocks_pre(use);
1481 nstack.push(n, i); // Save parent and next use's index.
1482 n = use; // Process all children of current use.
1483 cnt = use->outcnt();
1484 i = 0;
1485 }
1486 }
1487 else {
1488 // All of n's children have been processed, complete post-processing.
1489 if (cnt != 0 && !n->is_Con()) {
1490 assert(has_node(n), "no dead nodes");
1491 split_if_with_blocks_post(n, last_round);
1492 }
1493 if (must_throttle_split_if()) {
1494 nstack.clear();
1495 }
1496 if (nstack.is_empty()) {
1497 // Finished all nodes on stack.
1498 break;
1499 }
1500 // Get saved parent node and next use's index. Visit the rest of uses.
1501 n = nstack.node();
1502 cnt = n->outcnt();
1503 i = nstack.index();
1504 nstack.pop();
1505 }
1506 }
1507 }
1508
1509
1510 //=============================================================================
1511 //
1512 // C L O N E A L O O P B O D Y
1513 //
1514
1515 //------------------------------clone_iff--------------------------------------
1516 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1517 // "Nearly" because all Nodes have been cloned from the original in the loop,
1518 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1519 // through the Phi recursively, and return a Bool.
1520 Node* PhaseIdealLoop::clone_iff(PhiNode *phi, IdealLoopTree *loop) {
1521
1522 // Convert this Phi into a Phi merging Bools
1523 uint i;
1524 for (i = 1; i < phi->req(); i++) {
1525 Node *b = phi->in(i);
1526 if (b->is_Phi()) {
1527 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi(), loop));
1528 } else {
1529 assert(b->is_Bool() || b->Opcode() == Op_Opaque4, "");
1530 }
1531 }
1532
1533 Node* n = phi->in(1);
1534 Node* sample_opaque = NULL;
1535 Node *sample_bool = NULL;
1536 if (n->Opcode() == Op_Opaque4) {
1537 sample_opaque = n;
1538 sample_bool = n->in(1);
1539 assert(sample_bool->is_Bool(), "wrong type");
1540 } else {
1541 sample_bool = n;
1542 }
1543 Node *sample_cmp = sample_bool->in(1);
1544
1545 // Make Phis to merge the Cmp's inputs.
1546 PhiNode *phi1 = new PhiNode(phi->in(0), Type::TOP);
1547 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP);
1548 for (i = 1; i < phi->req(); i++) {
1549 Node *n1 = sample_opaque == NULL ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1);
1550 Node *n2 = sample_opaque == NULL ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2);
1551 phi1->set_req(i, n1);
1552 phi2->set_req(i, n2);
1553 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
1554 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
1555 }
1556 // See if these Phis have been made before.
1557 // Register with optimizer
1558 Node *hit1 = _igvn.hash_find_insert(phi1);
1559 if (hit1) { // Hit, toss just made Phi
1560 _igvn.remove_dead_node(phi1); // Remove new phi
1561 assert(hit1->is_Phi(), "" );
1562 phi1 = (PhiNode*)hit1; // Use existing phi
1563 } else { // Miss
1564 _igvn.register_new_node_with_optimizer(phi1);
1565 }
1566 Node *hit2 = _igvn.hash_find_insert(phi2);
1567 if (hit2) { // Hit, toss just made Phi
1568 _igvn.remove_dead_node(phi2); // Remove new phi
1569 assert(hit2->is_Phi(), "" );
1570 phi2 = (PhiNode*)hit2; // Use existing phi
1571 } else { // Miss
1572 _igvn.register_new_node_with_optimizer(phi2);
1573 }
1574 // Register Phis with loop/block info
1575 set_ctrl(phi1, phi->in(0));
1576 set_ctrl(phi2, phi->in(0));
1577 // Make a new Cmp
1578 Node *cmp = sample_cmp->clone();
1579 cmp->set_req(1, phi1);
1580 cmp->set_req(2, phi2);
1581 _igvn.register_new_node_with_optimizer(cmp);
1582 set_ctrl(cmp, phi->in(0));
1583
1584 // Make a new Bool
1585 Node *b = sample_bool->clone();
1586 b->set_req(1,cmp);
1587 _igvn.register_new_node_with_optimizer(b);
1588 set_ctrl(b, phi->in(0));
1589
1590 if (sample_opaque != NULL) {
1591 Node* opaque = sample_opaque->clone();
1592 opaque->set_req(1, b);
1593 _igvn.register_new_node_with_optimizer(opaque);
1594 set_ctrl(opaque, phi->in(0));
1595 return opaque;
1596 }
1597
1598 assert(b->is_Bool(), "");
1599 return b;
1600 }
1601
1602 //------------------------------clone_bool-------------------------------------
1603 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1604 // "Nearly" because all Nodes have been cloned from the original in the loop,
1605 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1606 // through the Phi recursively, and return a Bool.
1607 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) {
1608 uint i;
1609 // Convert this Phi into a Phi merging Bools
1610 for( i = 1; i < phi->req(); i++ ) {
1611 Node *b = phi->in(i);
1612 if( b->is_Phi() ) {
1613 _igvn.replace_input_of(phi, i, clone_bool( b->as_Phi(), loop ));
1614 } else {
1615 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
1616 }
1617 }
1618
1619 Node *sample_cmp = phi->in(1);
1620
1621 // Make Phis to merge the Cmp's inputs.
1622 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP );
1623 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP );
1624 for( uint j = 1; j < phi->req(); j++ ) {
1625 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
1626 Node *n1, *n2;
1627 if( cmp_top->is_Cmp() ) {
1628 n1 = cmp_top->in(1);
1629 n2 = cmp_top->in(2);
1630 } else {
1631 n1 = n2 = cmp_top;
1632 }
1633 phi1->set_req( j, n1 );
1634 phi2->set_req( j, n2 );
1635 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
1636 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
1637 }
1638
1639 // See if these Phis have been made before.
1640 // Register with optimizer
1641 Node *hit1 = _igvn.hash_find_insert(phi1);
1642 if( hit1 ) { // Hit, toss just made Phi
1643 _igvn.remove_dead_node(phi1); // Remove new phi
1644 assert( hit1->is_Phi(), "" );
1645 phi1 = (PhiNode*)hit1; // Use existing phi
1646 } else { // Miss
1647 _igvn.register_new_node_with_optimizer(phi1);
1648 }
1649 Node *hit2 = _igvn.hash_find_insert(phi2);
1650 if( hit2 ) { // Hit, toss just made Phi
1651 _igvn.remove_dead_node(phi2); // Remove new phi
1652 assert( hit2->is_Phi(), "" );
1653 phi2 = (PhiNode*)hit2; // Use existing phi
1654 } else { // Miss
1655 _igvn.register_new_node_with_optimizer(phi2);
1656 }
1657 // Register Phis with loop/block info
1658 set_ctrl(phi1, phi->in(0));
1659 set_ctrl(phi2, phi->in(0));
1660 // Make a new Cmp
1661 Node *cmp = sample_cmp->clone();
1662 cmp->set_req( 1, phi1 );
1663 cmp->set_req( 2, phi2 );
1664 _igvn.register_new_node_with_optimizer(cmp);
1665 set_ctrl(cmp, phi->in(0));
1666
1667 assert( cmp->is_Cmp(), "" );
1668 return (CmpNode*)cmp;
1669 }
1670
1671 //------------------------------sink_use---------------------------------------
1672 // If 'use' was in the loop-exit block, it now needs to be sunk
1673 // below the post-loop merge point.
1674 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) {
1675 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) {
1676 set_ctrl(use, post_loop);
1677 for (DUIterator j = use->outs(); use->has_out(j); j++)
1678 sink_use(use->out(j), post_loop);
1679 }
1680 }
1681
1682 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new,
1683 IdealLoopTree* loop, IdealLoopTree* outer_loop,
1684 Node_List*& split_if_set, Node_List*& split_bool_set,
1685 Node_List*& split_cex_set, Node_List& worklist,
1686 uint new_counter, CloneLoopMode mode) {
1687 Node* nnn = old_new[old->_idx];
1688 // Copy uses to a worklist, so I can munge the def-use info
1689 // with impunity.
1690 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
1691 worklist.push(old->fast_out(j));
1692
1693 while( worklist.size() ) {
1694 Node *use = worklist.pop();
1695 if (!has_node(use)) continue; // Ignore dead nodes
1696 if (use->in(0) == C->top()) continue;
1697 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
1698 // Check for data-use outside of loop - at least one of OLD or USE
1699 // must not be a CFG node.
1700 #ifdef ASSERT
1701 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == NULL) {
1702 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint();
1703 assert(mode == ControlAroundStripMined && use == sfpt, "missed a node");
1704 }
1705 #endif
1706 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) {
1707
1708 // If the Data use is an IF, that means we have an IF outside of the
1709 // loop that is switching on a condition that is set inside of the
1710 // loop. Happens if people set a loop-exit flag; then test the flag
1711 // in the loop to break the loop, then test is again outside of the
1712 // loop to determine which way the loop exited.
1713 // Loop predicate If node connects to Bool node through Opaque1 node.
1714 if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use) || use->Opcode() == Op_Opaque4) {
1715 // Since this code is highly unlikely, we lazily build the worklist
1716 // of such Nodes to go split.
1717 if (!split_if_set) {
1718 ResourceArea *area = Thread::current()->resource_area();
1719 split_if_set = new Node_List(area);
1720 }
1721 split_if_set->push(use);
1722 }
1723 if (use->is_Bool()) {
1724 if (!split_bool_set) {
1725 ResourceArea *area = Thread::current()->resource_area();
1726 split_bool_set = new Node_List(area);
1727 }
1728 split_bool_set->push(use);
1729 }
1730 if (use->Opcode() == Op_CreateEx) {
1731 if (!split_cex_set) {
1732 ResourceArea *area = Thread::current()->resource_area();
1733 split_cex_set = new Node_List(area);
1734 }
1735 split_cex_set->push(use);
1736 }
1737
1738
1739 // Get "block" use is in
1740 uint idx = 0;
1741 while( use->in(idx) != old ) idx++;
1742 Node *prev = use->is_CFG() ? use : get_ctrl(use);
1743 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" );
1744 Node *cfg = prev->_idx >= new_counter
1745 ? prev->in(2)
1746 : idom(prev);
1747 if( use->is_Phi() ) // Phi use is in prior block
1748 cfg = prev->in(idx); // NOT in block of Phi itself
1749 if (cfg->is_top()) { // Use is dead?
1750 _igvn.replace_input_of(use, idx, C->top());
1751 continue;
1752 }
1753
1754 // If use is referenced through control edge... (idx == 0)
1755 if (mode == IgnoreStripMined && idx == 0) {
1756 LoopNode *head = loop->_head->as_Loop();
1757 if (head->is_strip_mined() && is_dominator(head->outer_loop_exit(), prev)) {
1758 // That node is outside the inner loop, leave it outside the
1759 // outer loop as well to not confuse verification code.
1760 assert(!loop->_parent->is_member(use_loop), "should be out of the outer loop");
1761 _igvn.replace_input_of(use, 0, head->outer_loop_exit());
1762 continue;
1763 }
1764 }
1765
1766 while(!outer_loop->is_member(get_loop(cfg))) {
1767 prev = cfg;
1768 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg);
1769 }
1770 // If the use occurs after merging several exits from the loop, then
1771 // old value must have dominated all those exits. Since the same old
1772 // value was used on all those exits we did not need a Phi at this
1773 // merge point. NOW we do need a Phi here. Each loop exit value
1774 // is now merged with the peeled body exit; each exit gets its own
1775 // private Phi and those Phis need to be merged here.
1776 Node *phi;
1777 if( prev->is_Region() ) {
1778 if( idx == 0 ) { // Updating control edge?
1779 phi = prev; // Just use existing control
1780 } else { // Else need a new Phi
1781 phi = PhiNode::make( prev, old );
1782 // Now recursively fix up the new uses of old!
1783 for( uint i = 1; i < prev->req(); i++ ) {
1784 worklist.push(phi); // Onto worklist once for each 'old' input
1785 }
1786 }
1787 } else {
1788 // Get new RegionNode merging old and new loop exits
1789 prev = old_new[prev->_idx];
1790 assert( prev, "just made this in step 7" );
1791 if( idx == 0) { // Updating control edge?
1792 phi = prev; // Just use existing control
1793 } else { // Else need a new Phi
1794 // Make a new Phi merging data values properly
1795 phi = PhiNode::make( prev, old );
1796 phi->set_req( 1, nnn );
1797 }
1798 }
1799 // If inserting a new Phi, check for prior hits
1800 if( idx != 0 ) {
1801 Node *hit = _igvn.hash_find_insert(phi);
1802 if( hit == NULL ) {
1803 _igvn.register_new_node_with_optimizer(phi); // Register new phi
1804 } else { // or
1805 // Remove the new phi from the graph and use the hit
1806 _igvn.remove_dead_node(phi);
1807 phi = hit; // Use existing phi
1808 }
1809 set_ctrl(phi, prev);
1810 }
1811 // Make 'use' use the Phi instead of the old loop body exit value
1812 _igvn.replace_input_of(use, idx, phi);
1813 if( use->_idx >= new_counter ) { // If updating new phis
1814 // Not needed for correctness, but prevents a weak assert
1815 // in AddPNode from tripping (when we end up with different
1816 // base & derived Phis that will become the same after
1817 // IGVN does CSE).
1818 Node *hit = _igvn.hash_find_insert(use);
1819 if( hit ) // Go ahead and re-hash for hits.
1820 _igvn.replace_node( use, hit );
1821 }
1822
1823 // If 'use' was in the loop-exit block, it now needs to be sunk
1824 // below the post-loop merge point.
1825 sink_use( use, prev );
1826 }
1827 }
1828 }
1829
1830 static void clone_outer_loop_helper(Node* n, const IdealLoopTree *loop, const IdealLoopTree* outer_loop,
1831 const Node_List &old_new, Unique_Node_List& wq, PhaseIdealLoop* phase,
1832 bool check_old_new) {
1833 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1834 Node* u = n->fast_out(j);
1835 assert(check_old_new || old_new[u->_idx] == NULL, "shouldn't have been cloned");
1836 if (!u->is_CFG() && (!check_old_new || old_new[u->_idx] == NULL)) {
1837 Node* c = phase->get_ctrl(u);
1838 IdealLoopTree* u_loop = phase->get_loop(c);
1839 assert(!loop->is_member(u_loop), "can be in outer loop or out of both loops only");
1840 if (outer_loop->is_member(u_loop)) {
1841 wq.push(u);
1842 }
1843 }
1844 }
1845 }
1846
1847 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
1848 IdealLoopTree* outer_loop, int dd, Node_List &old_new,
1849 Node_List& extra_data_nodes) {
1850 if (head->is_strip_mined() && mode != IgnoreStripMined) {
1851 CountedLoopNode* cl = head->as_CountedLoop();
1852 Node* l = cl->outer_loop();
1853 Node* tail = cl->outer_loop_tail();
1854 IfNode* le = cl->outer_loop_end();
1855 Node* sfpt = cl->outer_safepoint();
1856 CountedLoopEndNode* cle = cl->loopexit();
1857 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop();
1858 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null();
1859 Node* cle_out = cle->proj_out(false);
1860
1861 Node* new_sfpt = NULL;
1862 Node* new_cle_out = cle_out->clone();
1863 old_new.map(cle_out->_idx, new_cle_out);
1864 if (mode == CloneIncludesStripMined) {
1865 // clone outer loop body
1866 Node* new_l = l->clone();
1867 Node* new_tail = tail->clone();
1868 IfNode* new_le = le->clone()->as_If();
1869 new_sfpt = sfpt->clone();
1870
1871 set_loop(new_l, outer_loop->_parent);
1872 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd);
1873 set_loop(new_cle_out, outer_loop->_parent);
1874 set_idom(new_cle_out, new_cle, dd);
1875 set_loop(new_sfpt, outer_loop->_parent);
1876 set_idom(new_sfpt, new_cle_out, dd);
1877 set_loop(new_le, outer_loop->_parent);
1878 set_idom(new_le, new_sfpt, dd);
1879 set_loop(new_tail, outer_loop->_parent);
1880 set_idom(new_tail, new_le, dd);
1881 set_idom(new_cl, new_l, dd);
1882
1883 old_new.map(l->_idx, new_l);
1884 old_new.map(tail->_idx, new_tail);
1885 old_new.map(le->_idx, new_le);
1886 old_new.map(sfpt->_idx, new_sfpt);
1887
1888 new_l->set_req(LoopNode::LoopBackControl, new_tail);
1889 new_l->set_req(0, new_l);
1890 new_tail->set_req(0, new_le);
1891 new_le->set_req(0, new_sfpt);
1892 new_sfpt->set_req(0, new_cle_out);
1893 new_cle_out->set_req(0, new_cle);
1894 new_cl->set_req(LoopNode::EntryControl, new_l);
1895
1896 _igvn.register_new_node_with_optimizer(new_l);
1897 _igvn.register_new_node_with_optimizer(new_tail);
1898 _igvn.register_new_node_with_optimizer(new_le);
1899 } else {
1900 Node *newhead = old_new[loop->_head->_idx];
1901 newhead->as_Loop()->clear_strip_mined();
1902 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl));
1903 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
1904 }
1905 // Look at data node that were assigned a control in the outer
1906 // loop: they are kept in the outer loop by the safepoint so start
1907 // from the safepoint node's inputs.
1908 IdealLoopTree* outer_loop = get_loop(l);
1909 Node_Stack stack(2);
1910 stack.push(sfpt, 1);
1911 uint new_counter = C->unique();
1912 while (stack.size() > 0) {
1913 Node* n = stack.node();
1914 uint i = stack.index();
1915 while (i < n->req() &&
1916 (n->in(i) == NULL ||
1917 !has_ctrl(n->in(i)) ||
1918 get_loop(get_ctrl(n->in(i))) != outer_loop ||
1919 (old_new[n->in(i)->_idx] != NULL && old_new[n->in(i)->_idx]->_idx >= new_counter))) {
1920 i++;
1921 }
1922 if (i < n->req()) {
1923 stack.set_index(i+1);
1924 stack.push(n->in(i), 0);
1925 } else {
1926 assert(old_new[n->_idx] == NULL || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet");
1927 Node* m = n == sfpt ? new_sfpt : n->clone();
1928 if (m != NULL) {
1929 for (uint i = 0; i < n->req(); i++) {
1930 if (m->in(i) != NULL && old_new[m->in(i)->_idx] != NULL) {
1931 m->set_req(i, old_new[m->in(i)->_idx]);
1932 }
1933 }
1934 } else {
1935 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?");
1936 }
1937 if (n != sfpt) {
1938 extra_data_nodes.push(n);
1939 _igvn.register_new_node_with_optimizer(m);
1940 assert(get_ctrl(n) == cle_out, "what other control?");
1941 set_ctrl(m, new_cle_out);
1942 old_new.map(n->_idx, m);
1943 }
1944 stack.pop();
1945 }
1946 }
1947 if (mode == CloneIncludesStripMined) {
1948 _igvn.register_new_node_with_optimizer(new_sfpt);
1949 _igvn.register_new_node_with_optimizer(new_cle_out);
1950 }
1951 // Some other transformation may have pessimistically assign some
1952 // data nodes to the outer loop. Set their control so they are out
1953 // of the outer loop.
1954 ResourceMark rm;
1955 Unique_Node_List wq;
1956 for (uint i = 0; i < extra_data_nodes.size(); i++) {
1957 Node* old = extra_data_nodes.at(i);
1958 clone_outer_loop_helper(old, loop, outer_loop, old_new, wq, this, true);
1959 }
1960 Node* new_ctrl = cl->outer_loop_exit();
1961 assert(get_loop(new_ctrl) != outer_loop, "must be out of the loop nest");
1962 for (uint i = 0; i < wq.size(); i++) {
1963 Node* n = wq.at(i);
1964 set_ctrl(n, new_ctrl);
1965 clone_outer_loop_helper(n, loop, outer_loop, old_new, wq, this, false);
1966 }
1967 } else {
1968 Node *newhead = old_new[loop->_head->_idx];
1969 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
1970 }
1971 }
1972
1973 //------------------------------clone_loop-------------------------------------
1974 //
1975 // C L O N E A L O O P B O D Y
1976 //
1977 // This is the basic building block of the loop optimizations. It clones an
1978 // entire loop body. It makes an old_new loop body mapping; with this mapping
1979 // you can find the new-loop equivalent to an old-loop node. All new-loop
1980 // nodes are exactly equal to their old-loop counterparts, all edges are the
1981 // same. All exits from the old-loop now have a RegionNode that merges the
1982 // equivalent new-loop path. This is true even for the normal "loop-exit"
1983 // condition. All uses of loop-invariant old-loop values now come from (one
1984 // or more) Phis that merge their new-loop equivalents.
1985 //
1986 // This operation leaves the graph in an illegal state: there are two valid
1987 // control edges coming from the loop pre-header to both loop bodies. I'll
1988 // definitely have to hack the graph after running this transform.
1989 //
1990 // From this building block I will further edit edges to perform loop peeling
1991 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
1992 //
1993 // Parameter side_by_size_idom:
1994 // When side_by_size_idom is NULL, the dominator tree is constructed for
1995 // the clone loop to dominate the original. Used in construction of
1996 // pre-main-post loop sequence.
1997 // When nonnull, the clone and original are side-by-side, both are
1998 // dominated by the side_by_side_idom node. Used in construction of
1999 // unswitched loops.
2000 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
2001 CloneLoopMode mode, Node* side_by_side_idom) {
2002
2003 LoopNode* head = loop->_head->as_Loop();
2004 head->verify_strip_mined(1);
2005
2006 if (C->do_vector_loop() && PrintOpto) {
2007 const char* mname = C->method()->name()->as_quoted_ascii();
2008 if (mname != NULL) {
2009 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname);
2010 }
2011 }
2012
2013 CloneMap& cm = C->clone_map();
2014 Dict* dict = cm.dict();
2015 if (C->do_vector_loop()) {
2016 cm.set_clone_idx(cm.max_gen()+1);
2017 #ifndef PRODUCT
2018 if (PrintOpto) {
2019 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx());
2020 loop->dump_head();
2021 }
2022 #endif
2023 }
2024
2025 // Step 1: Clone the loop body. Make the old->new mapping.
2026 uint i;
2027 for( i = 0; i < loop->_body.size(); i++ ) {
2028 Node *old = loop->_body.at(i);
2029 Node *nnn = old->clone();
2030 old_new.map( old->_idx, nnn );
2031 if (C->do_vector_loop()) {
2032 cm.verify_insert_and_clone(old, nnn, cm.clone_idx());
2033 }
2034 _igvn.register_new_node_with_optimizer(nnn);
2035 }
2036
2037 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop;
2038
2039 // Step 2: Fix the edges in the new body. If the old input is outside the
2040 // loop use it. If the old input is INside the loop, use the corresponding
2041 // new node instead.
2042 for( i = 0; i < loop->_body.size(); i++ ) {
2043 Node *old = loop->_body.at(i);
2044 Node *nnn = old_new[old->_idx];
2045 // Fix CFG/Loop controlling the new node
2046 if (has_ctrl(old)) {
2047 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
2048 } else {
2049 set_loop(nnn, outer_loop->_parent);
2050 if (old->outcnt() > 0) {
2051 set_idom( nnn, old_new[idom(old)->_idx], dd );
2052 }
2053 }
2054 // Correct edges to the new node
2055 for( uint j = 0; j < nnn->req(); j++ ) {
2056 Node *n = nnn->in(j);
2057 if( n ) {
2058 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n );
2059 if( loop->is_member( old_in_loop ) )
2060 nnn->set_req(j, old_new[n->_idx]);
2061 }
2062 }
2063 _igvn.hash_find_insert(nnn);
2064 }
2065
2066 ResourceArea *area = Thread::current()->resource_area();
2067 Node_List extra_data_nodes(area); // data nodes in the outer strip mined loop
2068 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes);
2069
2070 // Step 3: Now fix control uses. Loop varying control uses have already
2071 // been fixed up (as part of all input edges in Step 2). Loop invariant
2072 // control uses must be either an IfFalse or an IfTrue. Make a merge
2073 // point to merge the old and new IfFalse/IfTrue nodes; make the use
2074 // refer to this.
2075 Node_List worklist(area);
2076 uint new_counter = C->unique();
2077 for( i = 0; i < loop->_body.size(); i++ ) {
2078 Node* old = loop->_body.at(i);
2079 if( !old->is_CFG() ) continue;
2080
2081 // Copy uses to a worklist, so I can munge the def-use info
2082 // with impunity.
2083 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
2084 worklist.push(old->fast_out(j));
2085
2086 while( worklist.size() ) { // Visit all uses
2087 Node *use = worklist.pop();
2088 if (!has_node(use)) continue; // Ignore dead nodes
2089 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
2090 if( !loop->is_member( use_loop ) && use->is_CFG() ) {
2091 // Both OLD and USE are CFG nodes here.
2092 assert( use->is_Proj(), "" );
2093 Node* nnn = old_new[old->_idx];
2094
2095 Node* newuse = NULL;
2096 if (head->is_strip_mined() && mode != IgnoreStripMined) {
2097 CountedLoopNode* cl = head->as_CountedLoop();
2098 CountedLoopEndNode* cle = cl->loopexit();
2099 Node* cle_out = cle->proj_out_or_null(false);
2100 if (use == cle_out) {
2101 IfNode* le = cl->outer_loop_end();
2102 use = le->proj_out(false);
2103 use_loop = get_loop(use);
2104 if (mode == CloneIncludesStripMined) {
2105 nnn = old_new[le->_idx];
2106 } else {
2107 newuse = old_new[cle_out->_idx];
2108 }
2109 }
2110 }
2111 if (newuse == NULL) {
2112 newuse = use->clone();
2113 }
2114
2115 // Clone the loop exit control projection
2116 if (C->do_vector_loop()) {
2117 cm.verify_insert_and_clone(use, newuse, cm.clone_idx());
2118 }
2119 newuse->set_req(0,nnn);
2120 _igvn.register_new_node_with_optimizer(newuse);
2121 set_loop(newuse, use_loop);
2122 set_idom(newuse, nnn, dom_depth(nnn) + 1 );
2123
2124 // We need a Region to merge the exit from the peeled body and the
2125 // exit from the old loop body.
2126 RegionNode *r = new RegionNode(3);
2127 // Map the old use to the new merge point
2128 old_new.map( use->_idx, r );
2129 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use));
2130 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" );
2131
2132 // The original user of 'use' uses 'r' instead.
2133 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
2134 Node* useuse = use->last_out(l);
2135 _igvn.rehash_node_delayed(useuse);
2136 uint uses_found = 0;
2137 if( useuse->in(0) == use ) {
2138 useuse->set_req(0, r);
2139 uses_found++;
2140 if( useuse->is_CFG() ) {
2141 assert( dom_depth(useuse) > dd_r, "" );
2142 set_idom(useuse, r, dom_depth(useuse));
2143 }
2144 }
2145 for( uint k = 1; k < useuse->req(); k++ ) {
2146 if( useuse->in(k) == use ) {
2147 useuse->set_req(k, r);
2148 uses_found++;
2149 if (useuse->is_Loop() && k == LoopNode::EntryControl) {
2150 assert(dom_depth(useuse) > dd_r , "");
2151 set_idom(useuse, r, dom_depth(useuse));
2152 }
2153 }
2154 }
2155 l -= uses_found; // we deleted 1 or more copies of this edge
2156 }
2157
2158 // Now finish up 'r'
2159 r->set_req( 1, newuse );
2160 r->set_req( 2, use );
2161 _igvn.register_new_node_with_optimizer(r);
2162 set_loop(r, use_loop);
2163 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r);
2164 } // End of if a loop-exit test
2165 }
2166 }
2167
2168 // Step 4: If loop-invariant use is not control, it must be dominated by a
2169 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region
2170 // there if needed. Make a Phi there merging old and new used values.
2171 Node_List *split_if_set = NULL;
2172 Node_List *split_bool_set = NULL;
2173 Node_List *split_cex_set = NULL;
2174 for( i = 0; i < loop->_body.size(); i++ ) {
2175 Node* old = loop->_body.at(i);
2176 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2177 split_bool_set, split_cex_set, worklist, new_counter,
2178 mode);
2179 }
2180
2181 for (i = 0; i < extra_data_nodes.size(); i++) {
2182 Node* old = extra_data_nodes.at(i);
2183 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2184 split_bool_set, split_cex_set, worklist, new_counter,
2185 mode);
2186 }
2187
2188 // Check for IFs that need splitting/cloning. Happens if an IF outside of
2189 // the loop uses a condition set in the loop. The original IF probably
2190 // takes control from one or more OLD Regions (which in turn get from NEW
2191 // Regions). In any case, there will be a set of Phis for each merge point
2192 // from the IF up to where the original BOOL def exists the loop.
2193 if (split_if_set) {
2194 while (split_if_set->size()) {
2195 Node *iff = split_if_set->pop();
2196 if (iff->in(1)->is_Phi()) {
2197 Node *b = clone_iff(iff->in(1)->as_Phi(), loop);
2198 _igvn.replace_input_of(iff, 1, b);
2199 }
2200 }
2201 }
2202 if (split_bool_set) {
2203 while (split_bool_set->size()) {
2204 Node *b = split_bool_set->pop();
2205 Node *phi = b->in(1);
2206 assert(phi->is_Phi(), "");
2207 CmpNode *cmp = clone_bool((PhiNode*)phi, loop);
2208 _igvn.replace_input_of(b, 1, cmp);
2209 }
2210 }
2211 if (split_cex_set) {
2212 while (split_cex_set->size()) {
2213 Node *b = split_cex_set->pop();
2214 assert(b->in(0)->is_Region(), "");
2215 assert(b->in(1)->is_Phi(), "");
2216 assert(b->in(0)->in(0) == b->in(1)->in(0), "");
2217 split_up(b, b->in(0), NULL);
2218 }
2219 }
2220
2221 }
2222
2223
2224 //---------------------- stride_of_possible_iv -------------------------------------
2225 // Looks for an iff/bool/comp with one operand of the compare
2226 // being a cycle involving an add and a phi,
2227 // with an optional truncation (left-shift followed by a right-shift)
2228 // of the add. Returns zero if not an iv.
2229 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
2230 Node* trunc1 = NULL;
2231 Node* trunc2 = NULL;
2232 const TypeInt* ttype = NULL;
2233 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) {
2234 return 0;
2235 }
2236 BoolNode* bl = iff->in(1)->as_Bool();
2237 Node* cmp = bl->in(1);
2238 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) {
2239 return 0;
2240 }
2241 // Must have an invariant operand
2242 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) {
2243 return 0;
2244 }
2245 Node* add2 = NULL;
2246 Node* cmp1 = cmp->in(1);
2247 if (cmp1->is_Phi()) {
2248 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
2249 Node* phi = cmp1;
2250 for (uint i = 1; i < phi->req(); i++) {
2251 Node* in = phi->in(i);
2252 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in,
2253 &trunc1, &trunc2, &ttype);
2254 if (add && add->in(1) == phi) {
2255 add2 = add->in(2);
2256 break;
2257 }
2258 }
2259 } else {
2260 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
2261 Node* addtrunc = cmp1;
2262 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc,
2263 &trunc1, &trunc2, &ttype);
2264 if (add && add->in(1)->is_Phi()) {
2265 Node* phi = add->in(1);
2266 for (uint i = 1; i < phi->req(); i++) {
2267 if (phi->in(i) == addtrunc) {
2268 add2 = add->in(2);
2269 break;
2270 }
2271 }
2272 }
2273 }
2274 if (add2 != NULL) {
2275 const TypeInt* add2t = _igvn.type(add2)->is_int();
2276 if (add2t->is_con()) {
2277 return add2t->get_con();
2278 }
2279 }
2280 return 0;
2281 }
2282
2283
2284 //---------------------- stay_in_loop -------------------------------------
2285 // Return the (unique) control output node that's in the loop (if it exists.)
2286 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
2287 Node* unique = NULL;
2288 if (!n) return NULL;
2289 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2290 Node* use = n->fast_out(i);
2291 if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
2292 if (unique != NULL) {
2293 return NULL;
2294 }
2295 unique = use;
2296 }
2297 }
2298 return unique;
2299 }
2300
2301 //------------------------------ register_node -------------------------------------
2302 // Utility to register node "n" with PhaseIdealLoop
2303 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) {
2304 _igvn.register_new_node_with_optimizer(n);
2305 loop->_body.push(n);
2306 if (n->is_CFG()) {
2307 set_loop(n, loop);
2308 set_idom(n, pred, ddepth);
2309 } else {
2310 set_ctrl(n, pred);
2311 }
2312 }
2313
2314 //------------------------------ proj_clone -------------------------------------
2315 // Utility to create an if-projection
2316 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
2317 ProjNode* c = p->clone()->as_Proj();
2318 c->set_req(0, iff);
2319 return c;
2320 }
2321
2322 //------------------------------ short_circuit_if -------------------------------------
2323 // Force the iff control output to be the live_proj
2324 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
2325 guarantee(live_proj != NULL, "null projection");
2326 int proj_con = live_proj->_con;
2327 assert(proj_con == 0 || proj_con == 1, "false or true projection");
2328 Node *con = _igvn.intcon(proj_con);
2329 set_ctrl(con, C->root());
2330 if (iff) {
2331 iff->set_req(1, con);
2332 }
2333 return con;
2334 }
2335
2336 //------------------------------ insert_if_before_proj -------------------------------------
2337 // Insert a new if before an if projection (* - new node)
2338 //
2339 // before
2340 // if(test)
2341 // / \
2342 // v v
2343 // other-proj proj (arg)
2344 //
2345 // after
2346 // if(test)
2347 // / \
2348 // / v
2349 // | * proj-clone
2350 // v |
2351 // other-proj v
2352 // * new_if(relop(cmp[IU](left,right)))
2353 // / \
2354 // v v
2355 // * new-proj proj
2356 // (returned)
2357 //
2358 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) {
2359 IfNode* iff = proj->in(0)->as_If();
2360 IdealLoopTree *loop = get_loop(proj);
2361 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2362 int ddepth = dom_depth(proj);
2363
2364 _igvn.rehash_node_delayed(iff);
2365 _igvn.rehash_node_delayed(proj);
2366
2367 proj->set_req(0, NULL); // temporary disconnect
2368 ProjNode* proj2 = proj_clone(proj, iff);
2369 register_node(proj2, loop, iff, ddepth);
2370
2371 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right);
2372 register_node(cmp, loop, proj2, ddepth);
2373
2374 BoolNode* bol = new BoolNode(cmp, relop);
2375 register_node(bol, loop, proj2, ddepth);
2376
2377 int opcode = iff->Opcode();
2378 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode");
2379 IfNode* new_if = (opcode == Op_If) ? new IfNode(proj2, bol, iff->_prob, iff->_fcnt):
2380 new RangeCheckNode(proj2, bol, iff->_prob, iff->_fcnt);
2381 register_node(new_if, loop, proj2, ddepth);
2382
2383 proj->set_req(0, new_if); // reattach
2384 set_idom(proj, new_if, ddepth);
2385
2386 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
2387 guarantee(new_exit != NULL, "null exit node");
2388 register_node(new_exit, get_loop(other_proj), new_if, ddepth);
2389
2390 return new_exit;
2391 }
2392
2393 //------------------------------ insert_region_before_proj -------------------------------------
2394 // Insert a region before an if projection (* - new node)
2395 //
2396 // before
2397 // if(test)
2398 // / |
2399 // v |
2400 // proj v
2401 // other-proj
2402 //
2403 // after
2404 // if(test)
2405 // / |
2406 // v |
2407 // * proj-clone v
2408 // | other-proj
2409 // v
2410 // * new-region
2411 // |
2412 // v
2413 // * dum_if
2414 // / \
2415 // v \
2416 // * dum-proj v
2417 // proj
2418 //
2419 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
2420 IfNode* iff = proj->in(0)->as_If();
2421 IdealLoopTree *loop = get_loop(proj);
2422 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2423 int ddepth = dom_depth(proj);
2424
2425 _igvn.rehash_node_delayed(iff);
2426 _igvn.rehash_node_delayed(proj);
2427
2428 proj->set_req(0, NULL); // temporary disconnect
2429 ProjNode* proj2 = proj_clone(proj, iff);
2430 register_node(proj2, loop, iff, ddepth);
2431
2432 RegionNode* reg = new RegionNode(2);
2433 reg->set_req(1, proj2);
2434 register_node(reg, loop, iff, ddepth);
2435
2436 IfNode* dum_if = new IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt);
2437 register_node(dum_if, loop, reg, ddepth);
2438
2439 proj->set_req(0, dum_if); // reattach
2440 set_idom(proj, dum_if, ddepth);
2441
2442 ProjNode* dum_proj = proj_clone(other_proj, dum_if);
2443 register_node(dum_proj, loop, dum_if, ddepth);
2444
2445 return reg;
2446 }
2447
2448 //------------------------------ insert_cmpi_loop_exit -------------------------------------
2449 // Clone a signed compare loop exit from an unsigned compare and
2450 // insert it before the unsigned cmp on the stay-in-loop path.
2451 // All new nodes inserted in the dominator tree between the original
2452 // if and it's projections. The original if test is replaced with
2453 // a constant to force the stay-in-loop path.
2454 //
2455 // This is done to make sure that the original if and it's projections
2456 // still dominate the same set of control nodes, that the ctrl() relation
2457 // from data nodes to them is preserved, and that their loop nesting is
2458 // preserved.
2459 //
2460 // before
2461 // if(i <u limit) unsigned compare loop exit
2462 // / |
2463 // v v
2464 // exit-proj stay-in-loop-proj
2465 //
2466 // after
2467 // if(stay-in-loop-const) original if
2468 // / |
2469 // / v
2470 // / if(i < limit) new signed test
2471 // / / |
2472 // / / v
2473 // / / if(i <u limit) new cloned unsigned test
2474 // / / / |
2475 // v v v |
2476 // region |
2477 // | |
2478 // dum-if |
2479 // / | |
2480 // ether | |
2481 // v v
2482 // exit-proj stay-in-loop-proj
2483 //
2484 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) {
2485 const bool Signed = true;
2486 const bool Unsigned = false;
2487
2488 BoolNode* bol = if_cmpu->in(1)->as_Bool();
2489 if (bol->_test._test != BoolTest::lt) return NULL;
2490 CmpNode* cmpu = bol->in(1)->as_Cmp();
2491 if (cmpu->Opcode() != Op_CmpU) return NULL;
2492 int stride = stride_of_possible_iv(if_cmpu);
2493 if (stride == 0) return NULL;
2494
2495 Node* lp_proj = stay_in_loop(if_cmpu, loop);
2496 guarantee(lp_proj != NULL, "null loop node");
2497
2498 ProjNode* lp_continue = lp_proj->as_Proj();
2499 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
2500
2501 Node* limit = NULL;
2502 if (stride > 0) {
2503 limit = cmpu->in(2);
2504 } else {
2505 limit = _igvn.makecon(TypeInt::ZERO);
2506 set_ctrl(limit, C->root());
2507 }
2508 // Create a new region on the exit path
2509 RegionNode* reg = insert_region_before_proj(lp_exit);
2510 guarantee(reg != NULL, "null region node");
2511
2512 // Clone the if-cmpu-true-false using a signed compare
2513 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
2514 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue);
2515 reg->add_req(cmpi_exit);
2516
2517 // Clone the if-cmpu-true-false
2518 BoolTest::mask rel_u = bol->_test._test;
2519 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
2520 reg->add_req(cmpu_exit);
2521
2522 // Force original if to stay in loop.
2523 short_circuit_if(if_cmpu, lp_continue);
2524
2525 return cmpi_exit->in(0)->as_If();
2526 }
2527
2528 //------------------------------ remove_cmpi_loop_exit -------------------------------------
2529 // Remove a previously inserted signed compare loop exit.
2530 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
2531 Node* lp_proj = stay_in_loop(if_cmp, loop);
2532 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
2533 stay_in_loop(lp_proj, loop)->is_If() &&
2534 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
2535 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
2536 set_ctrl(con, C->root());
2537 if_cmp->set_req(1, con);
2538 }
2539
2540 //------------------------------ scheduled_nodelist -------------------------------------
2541 // Create a post order schedule of nodes that are in the
2542 // "member" set. The list is returned in "sched".
2543 // The first node in "sched" is the loop head, followed by
2544 // nodes which have no inputs in the "member" set, and then
2545 // followed by the nodes that have an immediate input dependence
2546 // on a node in "sched".
2547 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
2548
2549 assert(member.test(loop->_head->_idx), "loop head must be in member set");
2550 Arena *a = Thread::current()->resource_area();
2551 VectorSet visited(a);
2552 Node_Stack nstack(a, loop->_body.size());
2553
2554 Node* n = loop->_head; // top of stack is cached in "n"
2555 uint idx = 0;
2556 visited.set(n->_idx);
2557
2558 // Initially push all with no inputs from within member set
2559 for(uint i = 0; i < loop->_body.size(); i++ ) {
2560 Node *elt = loop->_body.at(i);
2561 if (member.test(elt->_idx)) {
2562 bool found = false;
2563 for (uint j = 0; j < elt->req(); j++) {
2564 Node* def = elt->in(j);
2565 if (def && member.test(def->_idx) && def != elt) {
2566 found = true;
2567 break;
2568 }
2569 }
2570 if (!found && elt != loop->_head) {
2571 nstack.push(n, idx);
2572 n = elt;
2573 assert(!visited.test(n->_idx), "not seen yet");
2574 visited.set(n->_idx);
2575 }
2576 }
2577 }
2578
2579 // traverse out's that are in the member set
2580 while (true) {
2581 if (idx < n->outcnt()) {
2582 Node* use = n->raw_out(idx);
2583 idx++;
2584 if (!visited.test_set(use->_idx)) {
2585 if (member.test(use->_idx)) {
2586 nstack.push(n, idx);
2587 n = use;
2588 idx = 0;
2589 }
2590 }
2591 } else {
2592 // All outputs processed
2593 sched.push(n);
2594 if (nstack.is_empty()) break;
2595 n = nstack.node();
2596 idx = nstack.index();
2597 nstack.pop();
2598 }
2599 }
2600 }
2601
2602
2603 //------------------------------ has_use_in_set -------------------------------------
2604 // Has a use in the vector set
2605 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
2606 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2607 Node* use = n->fast_out(j);
2608 if (vset.test(use->_idx)) {
2609 return true;
2610 }
2611 }
2612 return false;
2613 }
2614
2615
2616 //------------------------------ has_use_internal_to_set -------------------------------------
2617 // Has use internal to the vector set (ie. not in a phi at the loop head)
2618 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
2619 Node* head = loop->_head;
2620 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2621 Node* use = n->fast_out(j);
2622 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
2623 return true;
2624 }
2625 }
2626 return false;
2627 }
2628
2629
2630 //------------------------------ clone_for_use_outside_loop -------------------------------------
2631 // clone "n" for uses that are outside of loop
2632 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
2633 int cloned = 0;
2634 assert(worklist.size() == 0, "should be empty");
2635 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2636 Node* use = n->fast_out(j);
2637 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
2638 worklist.push(use);
2639 }
2640 }
2641 while( worklist.size() ) {
2642 Node *use = worklist.pop();
2643 if (!has_node(use) || use->in(0) == C->top()) continue;
2644 uint j;
2645 for (j = 0; j < use->req(); j++) {
2646 if (use->in(j) == n) break;
2647 }
2648 assert(j < use->req(), "must be there");
2649
2650 // clone "n" and insert it between the inputs of "n" and the use outside the loop
2651 Node* n_clone = n->clone();
2652 _igvn.replace_input_of(use, j, n_clone);
2653 cloned++;
2654 Node* use_c;
2655 if (!use->is_Phi()) {
2656 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
2657 } else {
2658 // Use in a phi is considered a use in the associated predecessor block
2659 use_c = use->in(0)->in(j);
2660 }
2661 set_ctrl(n_clone, use_c);
2662 assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
2663 get_loop(use_c)->_body.push(n_clone);
2664 _igvn.register_new_node_with_optimizer(n_clone);
2665 #if !defined(PRODUCT)
2666 if (TracePartialPeeling) {
2667 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
2668 }
2669 #endif
2670 }
2671 return cloned;
2672 }
2673
2674
2675 //------------------------------ clone_for_special_use_inside_loop -------------------------------------
2676 // clone "n" for special uses that are in the not_peeled region.
2677 // If these def-uses occur in separate blocks, the code generator
2678 // marks the method as not compilable. For example, if a "BoolNode"
2679 // is in a different basic block than the "IfNode" that uses it, then
2680 // the compilation is aborted in the code generator.
2681 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
2682 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
2683 if (n->is_Phi() || n->is_Load()) {
2684 return;
2685 }
2686 assert(worklist.size() == 0, "should be empty");
2687 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2688 Node* use = n->fast_out(j);
2689 if ( not_peel.test(use->_idx) &&
2690 (use->is_If() || use->is_CMove() || use->is_Bool()) &&
2691 use->in(1) == n) {
2692 worklist.push(use);
2693 }
2694 }
2695 if (worklist.size() > 0) {
2696 // clone "n" and insert it between inputs of "n" and the use
2697 Node* n_clone = n->clone();
2698 loop->_body.push(n_clone);
2699 _igvn.register_new_node_with_optimizer(n_clone);
2700 set_ctrl(n_clone, get_ctrl(n));
2701 sink_list.push(n_clone);
2702 not_peel <<= n_clone->_idx; // add n_clone to not_peel set.
2703 #if !defined(PRODUCT)
2704 if (TracePartialPeeling) {
2705 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
2706 }
2707 #endif
2708 while( worklist.size() ) {
2709 Node *use = worklist.pop();
2710 _igvn.rehash_node_delayed(use);
2711 for (uint j = 1; j < use->req(); j++) {
2712 if (use->in(j) == n) {
2713 use->set_req(j, n_clone);
2714 }
2715 }
2716 }
2717 }
2718 }
2719
2720
2721 //------------------------------ insert_phi_for_loop -------------------------------------
2722 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
2723 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
2724 Node *phi = PhiNode::make(lp, back_edge_val);
2725 phi->set_req(LoopNode::EntryControl, lp_entry_val);
2726 // Use existing phi if it already exists
2727 Node *hit = _igvn.hash_find_insert(phi);
2728 if( hit == NULL ) {
2729 _igvn.register_new_node_with_optimizer(phi);
2730 set_ctrl(phi, lp);
2731 } else {
2732 // Remove the new phi from the graph and use the hit
2733 _igvn.remove_dead_node(phi);
2734 phi = hit;
2735 }
2736 _igvn.replace_input_of(use, idx, phi);
2737 }
2738
2739 #ifdef ASSERT
2740 //------------------------------ is_valid_loop_partition -------------------------------------
2741 // Validate the loop partition sets: peel and not_peel
2742 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
2743 VectorSet& not_peel ) {
2744 uint i;
2745 // Check that peel_list entries are in the peel set
2746 for (i = 0; i < peel_list.size(); i++) {
2747 if (!peel.test(peel_list.at(i)->_idx)) {
2748 return false;
2749 }
2750 }
2751 // Check at loop members are in one of peel set or not_peel set
2752 for (i = 0; i < loop->_body.size(); i++ ) {
2753 Node *def = loop->_body.at(i);
2754 uint di = def->_idx;
2755 // Check that peel set elements are in peel_list
2756 if (peel.test(di)) {
2757 if (not_peel.test(di)) {
2758 return false;
2759 }
2760 // Must be in peel_list also
2761 bool found = false;
2762 for (uint j = 0; j < peel_list.size(); j++) {
2763 if (peel_list.at(j)->_idx == di) {
2764 found = true;
2765 break;
2766 }
2767 }
2768 if (!found) {
2769 return false;
2770 }
2771 } else if (not_peel.test(di)) {
2772 if (peel.test(di)) {
2773 return false;
2774 }
2775 } else {
2776 return false;
2777 }
2778 }
2779 return true;
2780 }
2781
2782 //------------------------------ is_valid_clone_loop_exit_use -------------------------------------
2783 // Ensure a use outside of loop is of the right form
2784 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
2785 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2786 return (use->is_Phi() &&
2787 use_c->is_Region() && use_c->req() == 3 &&
2788 (use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
2789 use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
2790 use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
2791 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
2792 }
2793
2794 //------------------------------ is_valid_clone_loop_form -------------------------------------
2795 // Ensure that all uses outside of loop are of the right form
2796 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
2797 uint orig_exit_idx, uint clone_exit_idx) {
2798 uint len = peel_list.size();
2799 for (uint i = 0; i < len; i++) {
2800 Node *def = peel_list.at(i);
2801
2802 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2803 Node *use = def->fast_out(j);
2804 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2805 if (!loop->is_member(get_loop(use_c))) {
2806 // use is not in the loop, check for correct structure
2807 if (use->in(0) == def) {
2808 // Okay
2809 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
2810 return false;
2811 }
2812 }
2813 }
2814 }
2815 return true;
2816 }
2817 #endif
2818
2819 //------------------------------ partial_peel -------------------------------------
2820 // Partially peel (aka loop rotation) the top portion of a loop (called
2821 // the peel section below) by cloning it and placing one copy just before
2822 // the new loop head and the other copy at the bottom of the new loop.
2823 //
2824 // before after where it came from
2825 //
2826 // stmt1 stmt1
2827 // loop: stmt2 clone
2828 // stmt2 if condA goto exitA clone
2829 // if condA goto exitA new_loop: new
2830 // stmt3 stmt3 clone
2831 // if !condB goto loop if condB goto exitB clone
2832 // exitB: stmt2 orig
2833 // stmt4 if !condA goto new_loop orig
2834 // exitA: goto exitA
2835 // exitB:
2836 // stmt4
2837 // exitA:
2838 //
2839 // Step 1: find the cut point: an exit test on probable
2840 // induction variable.
2841 // Step 2: schedule (with cloning) operations in the peel
2842 // section that can be executed after the cut into
2843 // the section that is not peeled. This may need
2844 // to clone operations into exit blocks. For
2845 // instance, a reference to A[i] in the not-peel
2846 // section and a reference to B[i] in an exit block
2847 // may cause a left-shift of i by 2 to be placed
2848 // in the peel block. This step will clone the left
2849 // shift into the exit block and sink the left shift
2850 // from the peel to the not-peel section.
2851 // Step 3: clone the loop, retarget the control, and insert
2852 // phis for values that are live across the new loop
2853 // head. This is very dependent on the graph structure
2854 // from clone_loop. It creates region nodes for
2855 // exit control and associated phi nodes for values
2856 // flow out of the loop through that exit. The region
2857 // node is dominated by the clone's control projection.
2858 // So the clone's peel section is placed before the
2859 // new loop head, and the clone's not-peel section is
2860 // forms the top part of the new loop. The original
2861 // peel section forms the tail of the new loop.
2862 // Step 4: update the dominator tree and recompute the
2863 // dominator depth.
2864 //
2865 // orig
2866 //
2867 // stmt1
2868 // |
2869 // v
2870 // loop predicate
2871 // |
2872 // v
2873 // loop<----+
2874 // | |
2875 // stmt2 |
2876 // | |
2877 // v |
2878 // ifA |
2879 // / | |
2880 // v v |
2881 // false true ^ <-- last_peel
2882 // / | |
2883 // / ===|==cut |
2884 // / stmt3 | <-- first_not_peel
2885 // / | |
2886 // | v |
2887 // v ifB |
2888 // exitA: / \ |
2889 // / \ |
2890 // v v |
2891 // false true |
2892 // / \ |
2893 // / ----+
2894 // |
2895 // v
2896 // exitB:
2897 // stmt4
2898 //
2899 //
2900 // after clone loop
2901 //
2902 // stmt1
2903 // |
2904 // v
2905 // loop predicate
2906 // / \
2907 // clone / \ orig
2908 // / \
2909 // / \
2910 // v v
2911 // +---->loop loop<----+
2912 // | | | |
2913 // | stmt2 stmt2 |
2914 // | | | |
2915 // | v v |
2916 // | ifA ifA |
2917 // | | \ / | |
2918 // | v v v v |
2919 // ^ true false false true ^ <-- last_peel
2920 // | | ^ \ / | |
2921 // | cut==|== \ \ / ===|==cut |
2922 // | stmt3 \ \ / stmt3 | <-- first_not_peel
2923 // | | dom | | | |
2924 // | v \ 1v v2 v |
2925 // | ifB regionA ifB |
2926 // | / \ | / \ |
2927 // | / \ v / \ |
2928 // | v v exitA: v v |
2929 // | true false false true |
2930 // | / ^ \ / \ |
2931 // +---- \ \ / ----+
2932 // dom \ /
2933 // \ 1v v2
2934 // regionB
2935 // |
2936 // v
2937 // exitB:
2938 // stmt4
2939 //
2940 //
2941 // after partial peel
2942 //
2943 // stmt1
2944 // |
2945 // v
2946 // loop predicate
2947 // /
2948 // clone / orig
2949 // / TOP
2950 // / \
2951 // v v
2952 // TOP->loop loop----+
2953 // | | |
2954 // stmt2 stmt2 |
2955 // | | |
2956 // v v |
2957 // ifA ifA |
2958 // | \ / | |
2959 // v v v v |
2960 // true false false true | <-- last_peel
2961 // | ^ \ / +------|---+
2962 // +->newloop \ \ / === ==cut | |
2963 // | stmt3 \ \ / TOP | |
2964 // | | dom | | stmt3 | | <-- first_not_peel
2965 // | v \ 1v v2 v | |
2966 // | ifB regionA ifB ^ v
2967 // | / \ | / \ | |
2968 // | / \ v / \ | |
2969 // | v v exitA: v v | |
2970 // | true false false true | |
2971 // | / ^ \ / \ | |
2972 // | | \ \ / v | |
2973 // | | dom \ / TOP | |
2974 // | | \ 1v v2 | |
2975 // ^ v regionB | |
2976 // | | | | |
2977 // | | v ^ v
2978 // | | exitB: | |
2979 // | | stmt4 | |
2980 // | +------------>-----------------+ |
2981 // | |
2982 // +-----------------<---------------------+
2983 //
2984 //
2985 // final graph
2986 //
2987 // stmt1
2988 // |
2989 // v
2990 // loop predicate
2991 // |
2992 // v
2993 // stmt2 clone
2994 // |
2995 // v
2996 // ........> ifA clone
2997 // : / |
2998 // dom / |
2999 // : v v
3000 // : false true
3001 // : | |
3002 // : | v
3003 // : | newloop<-----+
3004 // : | | |
3005 // : | stmt3 clone |
3006 // : | | |
3007 // : | v |
3008 // : | ifB |
3009 // : | / \ |
3010 // : | v v |
3011 // : | false true |
3012 // : | | | |
3013 // : | v stmt2 |
3014 // : | exitB: | |
3015 // : | stmt4 v |
3016 // : | ifA orig |
3017 // : | / \ |
3018 // : | / \ |
3019 // : | v v |
3020 // : | false true |
3021 // : | / \ |
3022 // : v v -----+
3023 // RegionA
3024 // |
3025 // v
3026 // exitA
3027 //
3028 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
3029
3030 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only");
3031 if (!loop->_head->is_Loop()) {
3032 return false; }
3033
3034 LoopNode *head = loop->_head->as_Loop();
3035
3036 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
3037 return false;
3038 }
3039
3040 // Check for complex exit control
3041 for(uint ii = 0; ii < loop->_body.size(); ii++ ) {
3042 Node *n = loop->_body.at(ii);
3043 int opc = n->Opcode();
3044 if (n->is_Call() ||
3045 opc == Op_Catch ||
3046 opc == Op_CatchProj ||
3047 opc == Op_Jump ||
3048 opc == Op_JumpProj) {
3049 #if !defined(PRODUCT)
3050 if (TracePartialPeeling) {
3051 tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
3052 }
3053 #endif
3054 return false;
3055 }
3056 }
3057
3058 int dd = dom_depth(head);
3059
3060 // Step 1: find cut point
3061
3062 // Walk up dominators to loop head looking for first loop exit
3063 // which is executed on every path thru loop.
3064 IfNode *peel_if = NULL;
3065 IfNode *peel_if_cmpu = NULL;
3066
3067 Node *iff = loop->tail();
3068 while( iff != head ) {
3069 if( iff->is_If() ) {
3070 Node *ctrl = get_ctrl(iff->in(1));
3071 if (ctrl->is_top()) return false; // Dead test on live IF.
3072 // If loop-varying exit-test, check for induction variable
3073 if( loop->is_member(get_loop(ctrl)) &&
3074 loop->is_loop_exit(iff) &&
3075 is_possible_iv_test(iff)) {
3076 Node* cmp = iff->in(1)->in(1);
3077 if (cmp->Opcode() == Op_CmpI) {
3078 peel_if = iff->as_If();
3079 } else {
3080 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
3081 peel_if_cmpu = iff->as_If();
3082 }
3083 }
3084 }
3085 iff = idom(iff);
3086 }
3087 // Prefer signed compare over unsigned compare.
3088 IfNode* new_peel_if = NULL;
3089 if (peel_if == NULL) {
3090 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) {
3091 return false; // No peel point found
3092 }
3093 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
3094 if (new_peel_if == NULL) {
3095 return false; // No peel point found
3096 }
3097 peel_if = new_peel_if;
3098 }
3099 Node* last_peel = stay_in_loop(peel_if, loop);
3100 Node* first_not_peeled = stay_in_loop(last_peel, loop);
3101 if (first_not_peeled == NULL || first_not_peeled == head) {
3102 return false;
3103 }
3104
3105 #if !defined(PRODUCT)
3106 if (TraceLoopOpts) {
3107 tty->print("PartialPeel ");
3108 loop->dump_head();
3109 }
3110
3111 if (TracePartialPeeling) {
3112 tty->print_cr("before partial peel one iteration");
3113 Node_List wl;
3114 Node* t = head->in(2);
3115 while (true) {
3116 wl.push(t);
3117 if (t == head) break;
3118 t = idom(t);
3119 }
3120 while (wl.size() > 0) {
3121 Node* tt = wl.pop();
3122 tt->dump();
3123 if (tt == last_peel) tty->print_cr("-- cut --");
3124 }
3125 }
3126 #endif
3127 ResourceArea *area = Thread::current()->resource_area();
3128 VectorSet peel(area);
3129 VectorSet not_peel(area);
3130 Node_List peel_list(area);
3131 Node_List worklist(area);
3132 Node_List sink_list(area);
3133
3134 // Set of cfg nodes to peel are those that are executable from
3135 // the head through last_peel.
3136 assert(worklist.size() == 0, "should be empty");
3137 worklist.push(head);
3138 peel.set(head->_idx);
3139 while (worklist.size() > 0) {
3140 Node *n = worklist.pop();
3141 if (n != last_peel) {
3142 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3143 Node* use = n->fast_out(j);
3144 if (use->is_CFG() &&
3145 loop->is_member(get_loop(use)) &&
3146 !peel.test_set(use->_idx)) {
3147 worklist.push(use);
3148 }
3149 }
3150 }
3151 }
3152
3153 // Set of non-cfg nodes to peel are those that are control
3154 // dependent on the cfg nodes.
3155 uint i;
3156 for(i = 0; i < loop->_body.size(); i++ ) {
3157 Node *n = loop->_body.at(i);
3158 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
3159 if (peel.test(n_c->_idx)) {
3160 peel.set(n->_idx);
3161 } else {
3162 not_peel.set(n->_idx);
3163 }
3164 }
3165
3166 // Step 2: move operations from the peeled section down into the
3167 // not-peeled section
3168
3169 // Get a post order schedule of nodes in the peel region
3170 // Result in right-most operand.
3171 scheduled_nodelist(loop, peel, peel_list );
3172
3173 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
3174
3175 // For future check for too many new phis
3176 uint old_phi_cnt = 0;
3177 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
3178 Node* use = head->fast_out(j);
3179 if (use->is_Phi()) old_phi_cnt++;
3180 }
3181
3182 #if !defined(PRODUCT)
3183 if (TracePartialPeeling) {
3184 tty->print_cr("\npeeled list");
3185 }
3186 #endif
3187
3188 // Evacuate nodes in peel region into the not_peeled region if possible
3189 uint new_phi_cnt = 0;
3190 uint cloned_for_outside_use = 0;
3191 for (i = 0; i < peel_list.size();) {
3192 Node* n = peel_list.at(i);
3193 #if !defined(PRODUCT)
3194 if (TracePartialPeeling) n->dump();
3195 #endif
3196 bool incr = true;
3197 if ( !n->is_CFG() ) {
3198
3199 if ( has_use_in_set(n, not_peel) ) {
3200
3201 // If not used internal to the peeled region,
3202 // move "n" from peeled to not_peeled region.
3203
3204 if ( !has_use_internal_to_set(n, peel, loop) ) {
3205
3206 // if not pinned and not a load (which maybe anti-dependent on a store)
3207 // and not a CMove (Matcher expects only bool->cmove).
3208 if (n->in(0) == NULL && !n->is_Load() && !n->is_CMove()) {
3209 cloned_for_outside_use += clone_for_use_outside_loop( loop, n, worklist );
3210 sink_list.push(n);
3211 peel >>= n->_idx; // delete n from peel set.
3212 not_peel <<= n->_idx; // add n to not_peel set.
3213 peel_list.remove(i);
3214 incr = false;
3215 #if !defined(PRODUCT)
3216 if (TracePartialPeeling) {
3217 tty->print_cr("sink to not_peeled region: %d newbb: %d",
3218 n->_idx, get_ctrl(n)->_idx);
3219 }
3220 #endif
3221 }
3222 } else {
3223 // Otherwise check for special def-use cases that span
3224 // the peel/not_peel boundary such as bool->if
3225 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist );
3226 new_phi_cnt++;
3227 }
3228 }
3229 }
3230 if (incr) i++;
3231 }
3232
3233 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) {
3234 #if !defined(PRODUCT)
3235 if (TracePartialPeeling) {
3236 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c",
3237 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F');
3238 }
3239 #endif
3240 if (new_peel_if != NULL) {
3241 remove_cmpi_loop_exit(new_peel_if, loop);
3242 }
3243 // Inhibit more partial peeling on this loop
3244 assert(!head->is_partial_peel_loop(), "not partial peeled");
3245 head->mark_partial_peel_failed();
3246 if (cloned_for_outside_use > 0) {
3247 // Terminate this round of loop opts because
3248 // the graph outside this loop was changed.
3249 C->set_major_progress();
3250 return true;
3251 }
3252 return false;
3253 }
3254
3255 // Step 3: clone loop, retarget control, and insert new phis
3256
3257 // Create new loop head for new phis and to hang
3258 // the nodes being moved (sinked) from the peel region.
3259 LoopNode* new_head = new LoopNode(last_peel, last_peel);
3260 new_head->set_unswitch_count(head->unswitch_count()); // Preserve
3261 _igvn.register_new_node_with_optimizer(new_head);
3262 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
3263 _igvn.replace_input_of(first_not_peeled, 0, new_head);
3264 set_loop(new_head, loop);
3265 loop->_body.push(new_head);
3266 not_peel.set(new_head->_idx);
3267 set_idom(new_head, last_peel, dom_depth(first_not_peeled));
3268 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
3269
3270 while (sink_list.size() > 0) {
3271 Node* n = sink_list.pop();
3272 set_ctrl(n, new_head);
3273 }
3274
3275 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
3276
3277 clone_loop(loop, old_new, dd, IgnoreStripMined);
3278
3279 const uint clone_exit_idx = 1;
3280 const uint orig_exit_idx = 2;
3281 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop");
3282
3283 Node* head_clone = old_new[head->_idx];
3284 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop();
3285 Node* orig_tail_clone = head_clone->in(2);
3286
3287 // Add phi if "def" node is in peel set and "use" is not
3288
3289 for(i = 0; i < peel_list.size(); i++ ) {
3290 Node *def = peel_list.at(i);
3291 if (!def->is_CFG()) {
3292 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
3293 Node *use = def->fast_out(j);
3294 if (has_node(use) && use->in(0) != C->top() &&
3295 (!peel.test(use->_idx) ||
3296 (use->is_Phi() && use->in(0) == head)) ) {
3297 worklist.push(use);
3298 }
3299 }
3300 while( worklist.size() ) {
3301 Node *use = worklist.pop();
3302 for (uint j = 1; j < use->req(); j++) {
3303 Node* n = use->in(j);
3304 if (n == def) {
3305
3306 // "def" is in peel set, "use" is not in peel set
3307 // or "use" is in the entry boundary (a phi) of the peel set
3308
3309 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
3310
3311 if ( loop->is_member(get_loop( use_c )) ) {
3312 // use is in loop
3313 if (old_new[use->_idx] != NULL) { // null for dead code
3314 Node* use_clone = old_new[use->_idx];
3315 _igvn.replace_input_of(use, j, C->top());
3316 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
3317 }
3318 } else {
3319 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
3320 // use is not in the loop, check if the live range includes the cut
3321 Node* lp_if = use_c->in(orig_exit_idx)->in(0);
3322 if (not_peel.test(lp_if->_idx)) {
3323 assert(j == orig_exit_idx, "use from original loop");
3324 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
3325 }
3326 }
3327 }
3328 }
3329 }
3330 }
3331 }
3332
3333 // Step 3b: retarget control
3334
3335 // Redirect control to the new loop head if a cloned node in
3336 // the not_peeled region has control that points into the peeled region.
3337 // This necessary because the cloned peeled region will be outside
3338 // the loop.
3339 // from to
3340 // cloned-peeled <---+
3341 // new_head_clone: | <--+
3342 // cloned-not_peeled in(0) in(0)
3343 // orig-peeled
3344
3345 for(i = 0; i < loop->_body.size(); i++ ) {
3346 Node *n = loop->_body.at(i);
3347 if (!n->is_CFG() && n->in(0) != NULL &&
3348 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
3349 Node* n_clone = old_new[n->_idx];
3350 _igvn.replace_input_of(n_clone, 0, new_head_clone);
3351 }
3352 }
3353
3354 // Backedge of the surviving new_head (the clone) is original last_peel
3355 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel);
3356
3357 // Cut first node in original not_peel set
3358 _igvn.rehash_node_delayed(new_head); // Multiple edge updates:
3359 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of
3360 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls
3361
3362 // Copy head_clone back-branch info to original head
3363 // and remove original head's loop entry and
3364 // clone head's back-branch
3365 _igvn.rehash_node_delayed(head); // Multiple edge updates
3366 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl));
3367 head->set_req(LoopNode::LoopBackControl, C->top());
3368 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top());
3369
3370 // Similarly modify the phis
3371 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
3372 Node* use = head->fast_out(k);
3373 if (use->is_Phi() && use->outcnt() > 0) {
3374 Node* use_clone = old_new[use->_idx];
3375 _igvn.rehash_node_delayed(use); // Multiple edge updates
3376 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl));
3377 use->set_req(LoopNode::LoopBackControl, C->top());
3378 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top());
3379 }
3380 }
3381
3382 // Step 4: update dominator tree and dominator depth
3383
3384 set_idom(head, orig_tail_clone, dd);
3385 recompute_dom_depth();
3386
3387 // Inhibit more partial peeling on this loop
3388 new_head_clone->set_partial_peel_loop();
3389 C->set_major_progress();
3390 loop->record_for_igvn();
3391
3392 #if !defined(PRODUCT)
3393 if (TracePartialPeeling) {
3394 tty->print_cr("\nafter partial peel one iteration");
3395 Node_List wl(area);
3396 Node* t = last_peel;
3397 while (true) {
3398 wl.push(t);
3399 if (t == head_clone) break;
3400 t = idom(t);
3401 }
3402 while (wl.size() > 0) {
3403 Node* tt = wl.pop();
3404 if (tt == head) tty->print_cr("orig head");
3405 else if (tt == new_head_clone) tty->print_cr("new head");
3406 else if (tt == head_clone) tty->print_cr("clone head");
3407 tt->dump();
3408 }
3409 }
3410 #endif
3411 return true;
3412 }
3413
3414 //------------------------------reorg_offsets----------------------------------
3415 // Reorganize offset computations to lower register pressure. Mostly
3416 // prevent loop-fallout uses of the pre-incremented trip counter (which are
3417 // then alive with the post-incremented trip counter forcing an extra
3418 // register move)
3419 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) {
3420 // Perform it only for canonical counted loops.
3421 // Loop's shape could be messed up by iteration_split_impl.
3422 if (!loop->_head->is_CountedLoop())
3423 return;
3424 if (!loop->_head->as_Loop()->is_valid_counted_loop())
3425 return;
3426
3427 CountedLoopNode *cl = loop->_head->as_CountedLoop();
3428 CountedLoopEndNode *cle = cl->loopexit();
3429 Node *exit = cle->proj_out(false);
3430 Node *phi = cl->phi();
3431
3432 // Check for the special case of folks using the pre-incremented
3433 // trip-counter on the fall-out path (forces the pre-incremented
3434 // and post-incremented trip counter to be live at the same time).
3435 // Fix this by adjusting to use the post-increment trip counter.
3436
3437 bool progress = true;
3438 while (progress) {
3439 progress = false;
3440 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
3441 Node* use = phi->fast_out(i); // User of trip-counter
3442 if (!has_ctrl(use)) continue;
3443 Node *u_ctrl = get_ctrl(use);
3444 if (use->is_Phi()) {
3445 u_ctrl = NULL;
3446 for (uint j = 1; j < use->req(); j++)
3447 if (use->in(j) == phi)
3448 u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j));
3449 }
3450 IdealLoopTree *u_loop = get_loop(u_ctrl);
3451 // Look for loop-invariant use
3452 if (u_loop == loop) continue;
3453 if (loop->is_member(u_loop)) continue;
3454 // Check that use is live out the bottom. Assuming the trip-counter
3455 // update is right at the bottom, uses of of the loop middle are ok.
3456 if (dom_lca(exit, u_ctrl) != exit) continue;
3457 // Hit! Refactor use to use the post-incremented tripcounter.
3458 // Compute a post-increment tripcounter.
3459 Node* c = exit;
3460 if (cl->is_strip_mined()) {
3461 IdealLoopTree* outer_loop = get_loop(cl->outer_loop());
3462 if (!outer_loop->is_member(u_loop)) {
3463 c = cl->outer_loop_exit();
3464 }
3465 }
3466 Node *opaq = new Opaque2Node(C, cle->incr());
3467 register_new_node(opaq, c);
3468 Node *neg_stride = _igvn.intcon(-cle->stride_con());
3469 set_ctrl(neg_stride, C->root());
3470 Node *post = new AddINode(opaq, neg_stride);
3471 register_new_node(post, c);
3472 _igvn.rehash_node_delayed(use);
3473 for (uint j = 1; j < use->req(); j++) {
3474 if (use->in(j) == phi)
3475 use->set_req(j, post);
3476 }
3477 // Since DU info changed, rerun loop
3478 progress = true;
3479 break;
3480 }
3481 }
3482
3483 }