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 //------------------------------conditional_move-------------------------------
497 // Attempt to replace a Phi with a conditional move. We have some pretty
498 // strict profitability requirements. All Phis at the merge point must
499 // be converted, so we can remove the control flow. We need to limit the
500 // number of c-moves to a small handful. All code that was in the side-arms
501 // of the CFG diamond is now speculatively executed. This code has to be
502 // "cheap enough". We are pretty much limited to CFG diamonds that merge
503 // 1 or 2 items with a total of 1 or 2 ops executed speculatively.
504 Node *PhaseIdealLoop::conditional_move( Node *region ) {
505
506 assert(region->is_Region(), "sanity check");
507 if (region->req() != 3) return NULL;
508
509 // Check for CFG diamond
510 Node *lp = region->in(1);
511 Node *rp = region->in(2);
512 if (!lp || !rp) return NULL;
513 Node *lp_c = lp->in(0);
514 if (lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If()) return NULL;
515 IfNode *iff = lp_c->as_If();
516
517 // Check for ops pinned in an arm of the diamond.
518 // Can't remove the control flow in this case
519 if (lp->outcnt() > 1) return NULL;
520 if (rp->outcnt() > 1) return NULL;
521
522 IdealLoopTree* r_loop = get_loop(region);
523 assert(r_loop == get_loop(iff), "sanity");
524 // Always convert to CMOVE if all results are used only outside this loop.
525 bool used_inside_loop = (r_loop == _ltree_root);
526
527 // Check profitability
528 int cost = 0;
529 int phis = 0;
530 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
531 Node *out = region->fast_out(i);
532 if (!out->is_Phi()) continue; // Ignore other control edges, etc
533 phis++;
534 PhiNode* phi = out->as_Phi();
535 BasicType bt = phi->type()->basic_type();
536 switch (bt) {
537 case T_DOUBLE:
538 case T_FLOAT:
539 if (C->use_cmove()) {
540 continue; //TODO: maybe we want to add some cost
541 }
542 cost += Matcher::float_cmove_cost(); // Could be very expensive
543 break;
544 case T_LONG: {
545 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's
546 }
547 case T_INT: // These all CMOV fine
548 case T_ADDRESS: { // (RawPtr)
549 cost++;
550 break;
551 }
552 case T_NARROWOOP: // Fall through
553 case T_OBJECT: { // Base oops are OK, but not derived oops
554 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr();
555 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a
556 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus
557 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we
558 // have a Phi for the base here that we convert to a CMOVE all is well
559 // and good. But if the base is dead, we'll not make a CMOVE. Later
560 // the allocator will have to produce a base by creating a CMOVE of the
561 // relevant bases. This puts the allocator in the business of
562 // manufacturing expensive instructions, generally a bad plan.
563 // Just Say No to Conditionally-Moved Derived Pointers.
564 if (tp && tp->offset() != 0)
565 return NULL;
566 cost++;
567 break;
568 }
569 default:
570 return NULL; // In particular, can't do memory or I/O
571 }
572 // Add in cost any speculative ops
573 for (uint j = 1; j < region->req(); j++) {
574 Node *proj = region->in(j);
575 Node *inp = phi->in(j);
576 if (get_ctrl(inp) == proj) { // Found local op
577 cost++;
578 // Check for a chain of dependent ops; these will all become
579 // speculative in a CMOV.
580 for (uint k = 1; k < inp->req(); k++)
581 if (get_ctrl(inp->in(k)) == proj)
582 cost += ConditionalMoveLimit; // Too much speculative goo
583 }
584 }
585 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode.
586 // This will likely Split-If, a higher-payoff operation.
587 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) {
588 Node* use = phi->fast_out(k);
589 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr())
590 cost += ConditionalMoveLimit;
591 // Is there a use inside the loop?
592 // Note: check only basic types since CMoveP is pinned.
593 if (!used_inside_loop && is_java_primitive(bt)) {
594 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use);
595 if (r_loop == u_loop || r_loop->is_member(u_loop)) {
596 used_inside_loop = true;
597 }
598 }
599 }
600 }//for
601 Node* bol = iff->in(1);
602 assert(bol->Opcode() == Op_Bool, "");
603 int cmp_op = bol->in(1)->Opcode();
604 // It is expensive to generate flags from a float compare.
605 // Avoid duplicated float compare.
606 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL;
607
608 float infrequent_prob = PROB_UNLIKELY_MAG(3);
609 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop.
610 if (used_inside_loop) {
611 if (cost >= ConditionalMoveLimit) return NULL; // Too much goo
612
613 // BlockLayoutByFrequency optimization moves infrequent branch
614 // from hot path. No point in CMOV'ing in such case (110 is used
615 // instead of 100 to take into account not exactness of float value).
616 if (BlockLayoutByFrequency) {
617 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f);
618 }
619 }
620 // Check for highly predictable branch. No point in CMOV'ing if
621 // we are going to predict accurately all the time.
622 if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) {
623 //keep going
624 } else if (iff->_prob < infrequent_prob ||
625 iff->_prob > (1.0f - infrequent_prob))
626 return NULL;
627
628 // --------------
629 // Now replace all Phis with CMOV's
630 Node *cmov_ctrl = iff->in(0);
631 uint flip = (lp->Opcode() == Op_IfTrue);
632 Node_List wq;
633 while (1) {
634 PhiNode* phi = NULL;
635 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
636 Node *out = region->fast_out(i);
637 if (out->is_Phi()) {
638 phi = out->as_Phi();
639 break;
640 }
641 }
642 if (phi == NULL) break;
643 if (PrintOpto && VerifyLoopOptimizations) { tty->print_cr("CMOV"); }
644 // Move speculative ops
645 wq.push(phi);
646 while (wq.size() > 0) {
647 Node *n = wq.pop();
648 for (uint j = 1; j < n->req(); j++) {
649 Node* m = n->in(j);
650 if (m != NULL && !is_dominator(get_ctrl(m), cmov_ctrl)) {
651 #ifndef PRODUCT
652 if (PrintOpto && VerifyLoopOptimizations) {
653 tty->print(" speculate: ");
654 m->dump();
655 }
656 #endif
657 set_ctrl(m, cmov_ctrl);
658 wq.push(m);
659 }
660 }
661 }
662 Node *cmov = CMoveNode::make(cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi));
663 register_new_node( cmov, cmov_ctrl );
664 _igvn.replace_node( phi, cmov );
665 #ifndef PRODUCT
666 if (TraceLoopOpts) {
667 tty->print("CMOV ");
668 r_loop->dump_head();
669 if (Verbose) {
670 bol->in(1)->dump(1);
671 cmov->dump(1);
672 }
673 }
674 if (VerifyLoopOptimizations) verify();
675 #endif
676 }
677
678 // The useless CFG diamond will fold up later; see the optimization in
679 // RegionNode::Ideal.
680 _igvn._worklist.push(region);
681
682 return iff->in(1);
683 }
684
685 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) {
686 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
687 Node* u = m->fast_out(i);
688 if (u->is_CFG()) {
689 if (u->Opcode() == Op_NeverBranch) {
690 u = ((NeverBranchNode*)u)->proj_out(0);
691 enqueue_cfg_uses(u, wq);
692 } else {
693 wq.push(u);
694 }
695 }
696 }
697 }
698
699 // Try moving a store out of a loop, right before the loop
700 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) {
701 // Store has to be first in the loop body
702 IdealLoopTree *n_loop = get_loop(n_ctrl);
703 if (n->is_Store() && n_loop != _ltree_root &&
704 n_loop->is_loop() && n_loop->_head->is_Loop() &&
705 n->in(0) != NULL) {
706 Node* address = n->in(MemNode::Address);
707 Node* value = n->in(MemNode::ValueIn);
708 Node* mem = n->in(MemNode::Memory);
709 IdealLoopTree* address_loop = get_loop(get_ctrl(address));
710 IdealLoopTree* value_loop = get_loop(get_ctrl(value));
711
712 // - address and value must be loop invariant
713 // - memory must be a memory Phi for the loop
714 // - Store must be the only store on this memory slice in the
715 // loop: if there's another store following this one then value
716 // written at iteration i by the second store could be overwritten
717 // at iteration i+n by the first store: it's not safe to move the
718 // first store out of the loop
719 // - nothing must observe the memory Phi: it guarantees no read
720 // before the store, we are also guaranteed the store post
721 // dominates the loop head (ignoring a possible early
722 // exit). Otherwise there would be extra Phi involved between the
723 // loop's Phi and the store.
724 // - there must be no early exit from the loop before the Store
725 // (such an exit most of the time would be an extra use of the
726 // memory Phi but sometimes is a bottom memory Phi that takes the
727 // store as input).
728
729 if (!n_loop->is_member(address_loop) &&
730 !n_loop->is_member(value_loop) &&
731 mem->is_Phi() && mem->in(0) == n_loop->_head &&
732 mem->outcnt() == 1 &&
733 mem->in(LoopNode::LoopBackControl) == n) {
734
735 assert(n_loop->_tail != NULL, "need a tail");
736 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop");
737
738 // Verify that there's no early exit of the loop before the store.
739 bool ctrl_ok = false;
740 {
741 // Follow control from loop head until n, we exit the loop or
742 // we reach the tail
743 ResourceMark rm;
744 Unique_Node_List wq;
745 wq.push(n_loop->_head);
746
747 for (uint next = 0; next < wq.size(); ++next) {
748 Node *m = wq.at(next);
749 if (m == n->in(0)) {
750 ctrl_ok = true;
751 continue;
752 }
753 assert(!has_ctrl(m), "should be CFG");
754 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) {
755 ctrl_ok = false;
756 break;
757 }
758 enqueue_cfg_uses(m, wq);
759 if (wq.size() > 10) {
760 ctrl_ok = false;
761 break;
762 }
763 }
764 }
765 if (ctrl_ok) {
766 // move the Store
767 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem);
768 _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl));
769 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl));
770 // Disconnect the phi now. An empty phi can confuse other
771 // optimizations in this pass of loop opts.
772 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl));
773 n_loop->_body.yank(mem);
774
775 set_ctrl_and_loop(n, n->in(0));
776
777 return n;
778 }
779 }
780 }
781 return NULL;
782 }
783
784 // Try moving a store out of a loop, right after the loop
785 void PhaseIdealLoop::try_move_store_after_loop(Node* n) {
786 if (n->is_Store() && n->in(0) != NULL) {
787 Node *n_ctrl = get_ctrl(n);
788 IdealLoopTree *n_loop = get_loop(n_ctrl);
789 // Store must be in a loop
790 if (n_loop != _ltree_root && !n_loop->_irreducible) {
791 Node* address = n->in(MemNode::Address);
792 Node* value = n->in(MemNode::ValueIn);
793 IdealLoopTree* address_loop = get_loop(get_ctrl(address));
794 // address must be loop invariant
795 if (!n_loop->is_member(address_loop)) {
796 // Store must be last on this memory slice in the loop and
797 // nothing in the loop must observe it
798 Node* phi = NULL;
799 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
800 Node* u = n->fast_out(i);
801 if (has_ctrl(u)) { // control use?
802 IdealLoopTree *u_loop = get_loop(get_ctrl(u));
803 if (!n_loop->is_member(u_loop)) {
804 continue;
805 }
806 if (u->is_Phi() && u->in(0) == n_loop->_head) {
807 assert(_igvn.type(u) == Type::MEMORY, "bad phi");
808 // multiple phis on the same slice are possible
809 if (phi != NULL) {
810 return;
811 }
812 phi = u;
813 continue;
814 }
815 }
816 return;
817 }
818 if (phi != NULL) {
819 // Nothing in the loop before the store (next iteration)
820 // must observe the stored value
821 bool mem_ok = true;
822 {
823 ResourceMark rm;
824 Unique_Node_List wq;
825 wq.push(phi);
826 for (uint next = 0; next < wq.size() && mem_ok; ++next) {
827 Node *m = wq.at(next);
828 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) {
829 Node* u = m->fast_out(i);
830 if (u->is_Store() || u->is_Phi()) {
831 if (u != n) {
832 wq.push(u);
833 mem_ok = (wq.size() <= 10);
834 }
835 } else {
836 mem_ok = false;
837 break;
838 }
839 }
840 }
841 }
842 if (mem_ok) {
843 // Move the store out of the loop if the LCA of all
844 // users (except for the phi) is outside the loop.
845 Node* hook = new Node(1);
846 _igvn.rehash_node_delayed(phi);
847 int count = phi->replace_edge(n, hook);
848 assert(count > 0, "inconsistent phi");
849
850 // Compute latest point this store can go
851 Node* lca = get_late_ctrl(n, get_ctrl(n));
852 if (n_loop->is_member(get_loop(lca))) {
853 // LCA is in the loop - bail out
854 _igvn.replace_node(hook, n);
855 return;
856 }
857 #ifdef ASSERT
858 if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) {
859 assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined");
860 n_loop->_head->as_Loop()->verify_strip_mined(1);
861 Node* outer = n_loop->_head->as_CountedLoop()->outer_loop();
862 IdealLoopTree* outer_loop = get_loop(outer);
863 assert(n_loop->_parent == outer_loop, "broken loop tree");
864 assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state");
865 }
866 #endif
867
868 // Move store out of the loop
869 _igvn.replace_node(hook, n->in(MemNode::Memory));
870 _igvn.replace_input_of(n, 0, lca);
871 set_ctrl_and_loop(n, lca);
872
873 // Disconnect the phi now. An empty phi can confuse other
874 // optimizations in this pass of loop opts..
875 if (phi->in(LoopNode::LoopBackControl) == phi) {
876 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl));
877 n_loop->_body.yank(phi);
878 }
879 }
880 }
881 }
882 }
883 }
884 }
885
886 //------------------------------split_if_with_blocks_pre-----------------------
887 // Do the real work in a non-recursive function. Data nodes want to be
888 // cloned in the pre-order so they can feed each other nicely.
889 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
890 // Cloning these guys is unlikely to win
891 int n_op = n->Opcode();
892 if( n_op == Op_MergeMem ) return n;
893 if( n->is_Proj() ) return n;
894 // Do not clone-up CmpFXXX variations, as these are always
895 // followed by a CmpI
896 if( n->is_Cmp() ) return n;
897 // Attempt to use a conditional move instead of a phi/branch
898 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) {
899 Node *cmov = conditional_move( n );
900 if( cmov ) return cmov;
901 }
902 if( n->is_CFG() || n->is_LoadStore() )
903 return n;
904 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd
905 n_op == Op_Opaque2 ) {
906 if( !C->major_progress() ) // If chance of no more loop opts...
907 _igvn._worklist.push(n); // maybe we'll remove them
908 return n;
909 }
910
911 if( n->is_Con() ) return n; // No cloning for Con nodes
912
913 Node *n_ctrl = get_ctrl(n);
914 if( !n_ctrl ) return n; // Dead node
915
916 Node* res = try_move_store_before_loop(n, n_ctrl);
917 if (res != NULL) {
918 return n;
919 }
920
921 // Attempt to remix address expressions for loop invariants
922 Node *m = remix_address_expressions( n );
923 if( m ) return m;
924
925 if (n->is_ConstraintCast()) {
926 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this);
927 // ConstraintCastNode::dominating_cast() uses node control input to determine domination.
928 // Node control inputs don't necessarily agree with loop control info (due to
929 // transformations happened in between), thus additional dominance check is needed
930 // to keep loop info valid.
931 if (dom_cast != NULL && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) {
932 _igvn.replace_node(n, dom_cast);
933 return dom_cast;
934 }
935 }
936
937 // Determine if the Node has inputs from some local Phi.
938 // Returns the block to clone thru.
939 Node *n_blk = has_local_phi_input( n );
940 if( !n_blk ) return n;
941
942 // Do not clone the trip counter through on a CountedLoop
943 // (messes up the canonical shape).
944 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n;
945
946 // Check for having no control input; not pinned. Allow
947 // dominating control.
948 if (n->in(0)) {
949 Node *dom = idom(n_blk);
950 if (dom_lca(n->in(0), dom) != n->in(0)) {
951 return n;
952 }
953 }
954 // Policy: when is it profitable. You must get more wins than
955 // policy before it is considered profitable. Policy is usually 0,
956 // so 1 win is considered profitable. Big merges will require big
957 // cloning, so get a larger policy.
958 int policy = n_blk->req() >> 2;
959
960 // If the loop is a candidate for range check elimination,
961 // delay splitting through it's phi until a later loop optimization
962 if (n_blk->is_CountedLoop()) {
963 IdealLoopTree *lp = get_loop(n_blk);
964 if (lp && lp->_rce_candidate) {
965 return n;
966 }
967 }
968
969 // Use same limit as split_if_with_blocks_post
970 if( C->live_nodes() > 35000 ) return n; // Method too big
971
972 // Split 'n' through the merge point if it is profitable
973 Node *phi = split_thru_phi( n, n_blk, policy );
974 if (!phi) return n;
975
976 // Found a Phi to split thru!
977 // Replace 'n' with the new phi
978 _igvn.replace_node( n, phi );
979 // Moved a load around the loop, 'en-registering' something.
980 if (n_blk->is_Loop() && n->is_Load() &&
981 !phi->in(LoopNode::LoopBackControl)->is_Load())
982 C->set_major_progress();
983
984 return phi;
985 }
986
987 static bool merge_point_too_heavy(Compile* C, Node* region) {
988 // Bail out if the region and its phis have too many users.
989 int weight = 0;
990 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
991 weight += region->fast_out(i)->outcnt();
992 }
993 int nodes_left = C->max_node_limit() - C->live_nodes();
994 if (weight * 8 > nodes_left) {
995 if (PrintOpto) {
996 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight);
997 }
998 return true;
999 } else {
1000 return false;
1001 }
1002 }
1003
1004 static bool merge_point_safe(Node* region) {
1005 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode
1006 // having a PhiNode input. This sidesteps the dangerous case where the split
1007 // ConvI2LNode may become TOP if the input Value() does not
1008 // overlap the ConvI2L range, leaving a node which may not dominate its
1009 // uses.
1010 // A better fix for this problem can be found in the BugTraq entry, but
1011 // expediency for Mantis demands this hack.
1012 // 6855164: If the merge point has a FastLockNode with a PhiNode input, we stop
1013 // split_if_with_blocks from splitting a block because we could not move around
1014 // the FastLockNode.
1015 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1016 Node* n = region->fast_out(i);
1017 if (n->is_Phi()) {
1018 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1019 Node* m = n->fast_out(j);
1020 if (m->is_FastLock())
1021 return false;
1022 #ifdef _LP64
1023 if (m->Opcode() == Op_ConvI2L)
1024 return false;
1025 if (m->is_CastII() && m->isa_CastII()->has_range_check()) {
1026 return false;
1027 }
1028 #endif
1029 }
1030 }
1031 }
1032 return true;
1033 }
1034
1035
1036 //------------------------------place_near_use---------------------------------
1037 // Place some computation next to use but not inside inner loops.
1038 // For inner loop uses move it to the preheader area.
1039 Node *PhaseIdealLoop::place_near_use(Node *useblock) const {
1040 IdealLoopTree *u_loop = get_loop( useblock );
1041 if (u_loop->_irreducible) {
1042 return useblock;
1043 }
1044 if (u_loop->_child) {
1045 if (useblock == u_loop->_head && u_loop->_head->is_OuterStripMinedLoop()) {
1046 return u_loop->_head->in(LoopNode::EntryControl);
1047 }
1048 return useblock;
1049 }
1050 return u_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl);
1051 }
1052
1053
1054 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) {
1055 if (!n->is_If() || n->is_CountedLoopEnd()) {
1056 return false;
1057 }
1058 if (!n->in(0)->is_Region()) {
1059 return false;
1060 }
1061 Node* region = n->in(0);
1062 Node* dom = idom(region);
1063 if (!dom->is_If() || dom->in(1) != n->in(1)) {
1064 return false;
1065 }
1066 IfNode* dom_if = dom->as_If();
1067 Node* proj_true = dom_if->proj_out(1);
1068 Node* proj_false = dom_if->proj_out(0);
1069
1070 for (uint i = 1; i < region->req(); i++) {
1071 if (is_dominator(proj_true, region->in(i))) {
1072 continue;
1073 }
1074 if (is_dominator(proj_false, region->in(i))) {
1075 continue;
1076 }
1077 return false;
1078 }
1079
1080 return true;
1081 }
1082
1083 bool PhaseIdealLoop::can_split_if(Node *n_ctrl) {
1084 if (C->live_nodes() > 35000) {
1085 return false; // Method too big
1086 }
1087
1088 // Do not do 'split-if' if irreducible loops are present.
1089 if (_has_irreducible_loops) {
1090 return false;
1091 }
1092
1093 if (merge_point_too_heavy(C, n_ctrl)) {
1094 return false;
1095 }
1096
1097 // Do not do 'split-if' if some paths are dead. First do dead code
1098 // elimination and then see if its still profitable.
1099 for (uint i = 1; i < n_ctrl->req(); i++) {
1100 if (n_ctrl->in(i) == C->top()) {
1101 return false;
1102 }
1103 }
1104
1105 // If trying to do a 'Split-If' at the loop head, it is only
1106 // profitable if the cmp folds up on BOTH paths. Otherwise we
1107 // risk peeling a loop forever.
1108
1109 // CNC - Disabled for now. Requires careful handling of loop
1110 // body selection for the cloned code. Also, make sure we check
1111 // for any input path not being in the same loop as n_ctrl. For
1112 // irreducible loops we cannot check for 'n_ctrl->is_Loop()'
1113 // because the alternative loop entry points won't be converted
1114 // into LoopNodes.
1115 IdealLoopTree *n_loop = get_loop(n_ctrl);
1116 for (uint j = 1; j < n_ctrl->req(); j++) {
1117 if (get_loop(n_ctrl->in(j)) != n_loop) {
1118 return false;
1119 }
1120 }
1121
1122 // Check for safety of the merge point.
1123 if (!merge_point_safe(n_ctrl)) {
1124 return false;
1125 }
1126
1127 return true;
1128 }
1129
1130 //------------------------------split_if_with_blocks_post----------------------
1131 // Do the real work in a non-recursive function. CFG hackery wants to be
1132 // in the post-order, so it can dirty the I-DOM info and not use the dirtied
1133 // info.
1134 void PhaseIdealLoop::split_if_with_blocks_post(Node *n, bool last_round) {
1135
1136 // Cloning Cmp through Phi's involves the split-if transform.
1137 // FastLock is not used by an If
1138 if (n->is_Cmp() && !n->is_FastLock() && !last_round) {
1139 Node *n_ctrl = get_ctrl(n);
1140 // Determine if the Node has inputs from some local Phi.
1141 // Returns the block to clone thru.
1142 Node *n_blk = has_local_phi_input(n);
1143 if (n_blk != n_ctrl) {
1144 return;
1145 }
1146
1147 if (!can_split_if(n_ctrl)) {
1148 return;
1149 }
1150
1151 if (n->outcnt() != 1) {
1152 return; // Multiple bool's from 1 compare?
1153 }
1154 Node *bol = n->unique_out();
1155 assert(bol->is_Bool(), "expect a bool here");
1156 if (bol->outcnt() != 1) {
1157 return;// Multiple branches from 1 compare?
1158 }
1159 Node *iff = bol->unique_out();
1160
1161 // Check some safety conditions
1162 if (iff->is_If()) { // Classic split-if?
1163 if (iff->in(0) != n_ctrl) {
1164 return; // Compare must be in same blk as if
1165 }
1166 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE
1167 // Can't split CMove with different control edge.
1168 if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) {
1169 return;
1170 }
1171 if (get_ctrl(iff->in(2)) == n_ctrl ||
1172 get_ctrl(iff->in(3)) == n_ctrl) {
1173 return; // Inputs not yet split-up
1174 }
1175 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) {
1176 return; // Loop-invar test gates loop-varying CMOVE
1177 }
1178 } else {
1179 return; // some other kind of node, such as an Allocate
1180 }
1181
1182 // When is split-if profitable? Every 'win' on means some control flow
1183 // goes dead, so it's almost always a win.
1184 int policy = 0;
1185 // Split compare 'n' through the merge point if it is profitable
1186 Node *phi = split_thru_phi( n, n_ctrl, policy);
1187 if (!phi) {
1188 return;
1189 }
1190
1191 // Found a Phi to split thru!
1192 // Replace 'n' with the new phi
1193 _igvn.replace_node(n, phi);
1194
1195 // Now split the bool up thru the phi
1196 Node *bolphi = split_thru_phi(bol, n_ctrl, -1);
1197 guarantee(bolphi != NULL, "null boolean phi node");
1198
1199 _igvn.replace_node(bol, bolphi);
1200 assert(iff->in(1) == bolphi, "");
1201
1202 if (bolphi->Value(&_igvn)->singleton()) {
1203 return;
1204 }
1205
1206 // Conditional-move? Must split up now
1207 if (!iff->is_If()) {
1208 Node *cmovphi = split_thru_phi(iff, n_ctrl, -1);
1209 _igvn.replace_node(iff, cmovphi);
1210 return;
1211 }
1212
1213 // Now split the IF
1214 do_split_if(iff);
1215 return;
1216 }
1217
1218 // Two identical ifs back to back can be merged
1219 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) {
1220 Node *n_ctrl = n->in(0);
1221 PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1));
1222 IfNode* dom_if = idom(n_ctrl)->as_If();
1223 Node* proj_true = dom_if->proj_out(1);
1224 Node* proj_false = dom_if->proj_out(0);
1225 Node* con_true = _igvn.makecon(TypeInt::ONE);
1226 Node* con_false = _igvn.makecon(TypeInt::ZERO);
1227
1228 for (uint i = 1; i < n_ctrl->req(); i++) {
1229 if (is_dominator(proj_true, n_ctrl->in(i))) {
1230 bolphi->init_req(i, con_true);
1231 } else {
1232 assert(is_dominator(proj_false, n_ctrl->in(i)), "bad if");
1233 bolphi->init_req(i, con_false);
1234 }
1235 }
1236 register_new_node(bolphi, n_ctrl);
1237 _igvn.replace_input_of(n, 1, bolphi);
1238
1239 // Now split the IF
1240 do_split_if(n);
1241 return;
1242 }
1243
1244 // Check for an IF ready to split; one that has its
1245 // condition codes input coming from a Phi at the block start.
1246 int n_op = n->Opcode();
1247
1248 // Check for an IF being dominated by another IF same test
1249 if (n_op == Op_If ||
1250 n_op == Op_RangeCheck) {
1251 Node *bol = n->in(1);
1252 uint max = bol->outcnt();
1253 // Check for same test used more than once?
1254 if (max > 1 && bol->is_Bool()) {
1255 // Search up IDOMs to see if this IF is dominated.
1256 Node *cutoff = get_ctrl(bol);
1257
1258 // Now search up IDOMs till cutoff, looking for a dominating test
1259 Node *prevdom = n;
1260 Node *dom = idom(prevdom);
1261 while (dom != cutoff) {
1262 if (dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom) {
1263 // Replace the dominated test with an obvious true or false.
1264 // Place it on the IGVN worklist for later cleanup.
1265 C->set_major_progress();
1266 dominated_by(prevdom, n, false, true);
1267 #ifndef PRODUCT
1268 if( VerifyLoopOptimizations ) verify();
1269 #endif
1270 return;
1271 }
1272 prevdom = dom;
1273 dom = idom(prevdom);
1274 }
1275 }
1276 }
1277
1278 // See if a shared loop-varying computation has no loop-varying uses.
1279 // Happens if something is only used for JVM state in uncommon trap exits,
1280 // like various versions of induction variable+offset. Clone the
1281 // computation per usage to allow it to sink out of the loop.
1282 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about)
1283 Node *n_ctrl = get_ctrl(n);
1284 IdealLoopTree *n_loop = get_loop(n_ctrl);
1285 if( n_loop != _ltree_root ) {
1286 DUIterator_Fast imax, i = n->fast_outs(imax);
1287 for (; i < imax; i++) {
1288 Node* u = n->fast_out(i);
1289 if( !has_ctrl(u) ) break; // Found control user
1290 IdealLoopTree *u_loop = get_loop(get_ctrl(u));
1291 if( u_loop == n_loop ) break; // Found loop-varying use
1292 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop
1293 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003
1294 }
1295 bool did_break = (i < imax); // Did we break out of the previous loop?
1296 if (!did_break && n->outcnt() > 1) { // All uses in outer loops!
1297 Node *late_load_ctrl = NULL;
1298 if (n->is_Load()) {
1299 // If n is a load, get and save the result from get_late_ctrl(),
1300 // to be later used in calculating the control for n's clones.
1301 clear_dom_lca_tags();
1302 late_load_ctrl = get_late_ctrl(n, n_ctrl);
1303 }
1304 // If n is a load, and the late control is the same as the current
1305 // control, then the cloning of n is a pointless exercise, because
1306 // GVN will ensure that we end up where we started.
1307 if (!n->is_Load() || late_load_ctrl != n_ctrl) {
1308 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) {
1309 Node *u = n->last_out(j); // Clone private computation per use
1310 _igvn.rehash_node_delayed(u);
1311 Node *x = n->clone(); // Clone computation
1312 Node *x_ctrl = NULL;
1313 if( u->is_Phi() ) {
1314 // Replace all uses of normal nodes. Replace Phi uses
1315 // individually, so the separate Nodes can sink down
1316 // different paths.
1317 uint k = 1;
1318 while( u->in(k) != n ) k++;
1319 u->set_req( k, x );
1320 // x goes next to Phi input path
1321 x_ctrl = u->in(0)->in(k);
1322 --j;
1323 } else { // Normal use
1324 // Replace all uses
1325 for( uint k = 0; k < u->req(); k++ ) {
1326 if( u->in(k) == n ) {
1327 u->set_req( k, x );
1328 --j;
1329 }
1330 }
1331 x_ctrl = get_ctrl(u);
1332 }
1333
1334 // Find control for 'x' next to use but not inside inner loops.
1335 // For inner loop uses get the preheader area.
1336 x_ctrl = place_near_use(x_ctrl);
1337
1338 if (n->is_Load()) {
1339 // For loads, add a control edge to a CFG node outside of the loop
1340 // to force them to not combine and return back inside the loop
1341 // during GVN optimization (4641526).
1342 //
1343 // Because we are setting the actual control input, factor in
1344 // the result from get_late_ctrl() so we respect any
1345 // anti-dependences. (6233005).
1346 x_ctrl = dom_lca(late_load_ctrl, x_ctrl);
1347
1348 // Don't allow the control input to be a CFG splitting node.
1349 // Such nodes should only have ProjNodes as outs, e.g. IfNode
1350 // should only have IfTrueNode and IfFalseNode (4985384).
1351 x_ctrl = find_non_split_ctrl(x_ctrl);
1352 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone");
1353
1354 x->set_req(0, x_ctrl);
1355 }
1356 register_new_node(x, x_ctrl);
1357
1358 // Some institutional knowledge is needed here: 'x' is
1359 // yanked because if the optimizer runs GVN on it all the
1360 // cloned x's will common up and undo this optimization and
1361 // be forced back in the loop. This is annoying because it
1362 // makes +VerifyOpto report false-positives on progress. I
1363 // tried setting control edges on the x's to force them to
1364 // not combine, but the matching gets worried when it tries
1365 // to fold a StoreP and an AddP together (as part of an
1366 // address expression) and the AddP and StoreP have
1367 // different controls.
1368 if (!x->is_Load() && !x->is_DecodeNarrowPtr()) _igvn._worklist.yank(x);
1369 }
1370 _igvn.remove_dead_node(n);
1371 }
1372 }
1373 }
1374 }
1375
1376 try_move_store_after_loop(n);
1377
1378 // Check for Opaque2's who's loop has disappeared - who's input is in the
1379 // same loop nest as their output. Remove 'em, they are no longer useful.
1380 if( n_op == Op_Opaque2 &&
1381 n->in(1) != NULL &&
1382 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) {
1383 _igvn.replace_node( n, n->in(1) );
1384 }
1385
1386 #if INCLUDE_ZGC
1387 if (UseZGC) {
1388 ZBarrierSetC2::loop_optimize_gc_barrier(this, n, last_round);
1389 }
1390 #endif
1391 }
1392
1393 //------------------------------split_if_with_blocks---------------------------
1394 // Check for aggressive application of 'split-if' optimization,
1395 // using basic block level info.
1396 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack, bool last_round) {
1397 Node *n = C->root();
1398 visited.set(n->_idx); // first, mark node as visited
1399 // Do pre-visit work for root
1400 n = split_if_with_blocks_pre( n );
1401 uint cnt = n->outcnt();
1402 uint i = 0;
1403 while (true) {
1404 // Visit all children
1405 if (i < cnt) {
1406 Node* use = n->raw_out(i);
1407 ++i;
1408 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
1409 // Now do pre-visit work for this use
1410 use = split_if_with_blocks_pre( use );
1411 nstack.push(n, i); // Save parent and next use's index.
1412 n = use; // Process all children of current use.
1413 cnt = use->outcnt();
1414 i = 0;
1415 }
1416 }
1417 else {
1418 // All of n's children have been processed, complete post-processing.
1419 if (cnt != 0 && !n->is_Con()) {
1420 assert(has_node(n), "no dead nodes");
1421 split_if_with_blocks_post( n, last_round );
1422 }
1423 if (nstack.is_empty()) {
1424 // Finished all nodes on stack.
1425 break;
1426 }
1427 // Get saved parent node and next use's index. Visit the rest of uses.
1428 n = nstack.node();
1429 cnt = n->outcnt();
1430 i = nstack.index();
1431 nstack.pop();
1432 }
1433 }
1434 }
1435
1436
1437 //=============================================================================
1438 //
1439 // C L O N E A L O O P B O D Y
1440 //
1441
1442 //------------------------------clone_iff--------------------------------------
1443 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1444 // "Nearly" because all Nodes have been cloned from the original in the loop,
1445 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1446 // through the Phi recursively, and return a Bool.
1447 Node* PhaseIdealLoop::clone_iff(PhiNode *phi, IdealLoopTree *loop) {
1448
1449 // Convert this Phi into a Phi merging Bools
1450 uint i;
1451 for (i = 1; i < phi->req(); i++) {
1452 Node *b = phi->in(i);
1453 if (b->is_Phi()) {
1454 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi(), loop));
1455 } else {
1456 assert(b->is_Bool() || b->Opcode() == Op_Opaque4, "");
1457 }
1458 }
1459
1460 Node* n = phi->in(1);
1461 Node* sample_opaque = NULL;
1462 Node *sample_bool = NULL;
1463 if (n->Opcode() == Op_Opaque4) {
1464 sample_opaque = n;
1465 sample_bool = n->in(1);
1466 assert(sample_bool->is_Bool(), "wrong type");
1467 } else {
1468 sample_bool = n;
1469 }
1470 Node *sample_cmp = sample_bool->in(1);
1471
1472 // Make Phis to merge the Cmp's inputs.
1473 PhiNode *phi1 = new PhiNode(phi->in(0), Type::TOP);
1474 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP);
1475 for (i = 1; i < phi->req(); i++) {
1476 Node *n1 = sample_opaque == NULL ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1);
1477 Node *n2 = sample_opaque == NULL ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2);
1478 phi1->set_req(i, n1);
1479 phi2->set_req(i, n2);
1480 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
1481 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
1482 }
1483 // See if these Phis have been made before.
1484 // Register with optimizer
1485 Node *hit1 = _igvn.hash_find_insert(phi1);
1486 if (hit1) { // Hit, toss just made Phi
1487 _igvn.remove_dead_node(phi1); // Remove new phi
1488 assert(hit1->is_Phi(), "" );
1489 phi1 = (PhiNode*)hit1; // Use existing phi
1490 } else { // Miss
1491 _igvn.register_new_node_with_optimizer(phi1);
1492 }
1493 Node *hit2 = _igvn.hash_find_insert(phi2);
1494 if (hit2) { // Hit, toss just made Phi
1495 _igvn.remove_dead_node(phi2); // Remove new phi
1496 assert(hit2->is_Phi(), "" );
1497 phi2 = (PhiNode*)hit2; // Use existing phi
1498 } else { // Miss
1499 _igvn.register_new_node_with_optimizer(phi2);
1500 }
1501 // Register Phis with loop/block info
1502 set_ctrl(phi1, phi->in(0));
1503 set_ctrl(phi2, phi->in(0));
1504 // Make a new Cmp
1505 Node *cmp = sample_cmp->clone();
1506 cmp->set_req(1, phi1);
1507 cmp->set_req(2, phi2);
1508 _igvn.register_new_node_with_optimizer(cmp);
1509 set_ctrl(cmp, phi->in(0));
1510
1511 // Make a new Bool
1512 Node *b = sample_bool->clone();
1513 b->set_req(1,cmp);
1514 _igvn.register_new_node_with_optimizer(b);
1515 set_ctrl(b, phi->in(0));
1516
1517 if (sample_opaque != NULL) {
1518 Node* opaque = sample_opaque->clone();
1519 opaque->set_req(1, b);
1520 _igvn.register_new_node_with_optimizer(opaque);
1521 set_ctrl(opaque, phi->in(0));
1522 return opaque;
1523 }
1524
1525 assert(b->is_Bool(), "");
1526 return b;
1527 }
1528
1529 //------------------------------clone_bool-------------------------------------
1530 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1531 // "Nearly" because all Nodes have been cloned from the original in the loop,
1532 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1533 // through the Phi recursively, and return a Bool.
1534 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) {
1535 uint i;
1536 // Convert this Phi into a Phi merging Bools
1537 for( i = 1; i < phi->req(); i++ ) {
1538 Node *b = phi->in(i);
1539 if( b->is_Phi() ) {
1540 _igvn.replace_input_of(phi, i, clone_bool( b->as_Phi(), loop ));
1541 } else {
1542 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
1543 }
1544 }
1545
1546 Node *sample_cmp = phi->in(1);
1547
1548 // Make Phis to merge the Cmp's inputs.
1549 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP );
1550 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP );
1551 for( uint j = 1; j < phi->req(); j++ ) {
1552 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
1553 Node *n1, *n2;
1554 if( cmp_top->is_Cmp() ) {
1555 n1 = cmp_top->in(1);
1556 n2 = cmp_top->in(2);
1557 } else {
1558 n1 = n2 = cmp_top;
1559 }
1560 phi1->set_req( j, n1 );
1561 phi2->set_req( j, n2 );
1562 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
1563 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
1564 }
1565
1566 // See if these Phis have been made before.
1567 // Register with optimizer
1568 Node *hit1 = _igvn.hash_find_insert(phi1);
1569 if( hit1 ) { // Hit, toss just made Phi
1570 _igvn.remove_dead_node(phi1); // Remove new phi
1571 assert( hit1->is_Phi(), "" );
1572 phi1 = (PhiNode*)hit1; // Use existing phi
1573 } else { // Miss
1574 _igvn.register_new_node_with_optimizer(phi1);
1575 }
1576 Node *hit2 = _igvn.hash_find_insert(phi2);
1577 if( hit2 ) { // Hit, toss just made Phi
1578 _igvn.remove_dead_node(phi2); // Remove new phi
1579 assert( hit2->is_Phi(), "" );
1580 phi2 = (PhiNode*)hit2; // Use existing phi
1581 } else { // Miss
1582 _igvn.register_new_node_with_optimizer(phi2);
1583 }
1584 // Register Phis with loop/block info
1585 set_ctrl(phi1, phi->in(0));
1586 set_ctrl(phi2, phi->in(0));
1587 // Make a new Cmp
1588 Node *cmp = sample_cmp->clone();
1589 cmp->set_req( 1, phi1 );
1590 cmp->set_req( 2, phi2 );
1591 _igvn.register_new_node_with_optimizer(cmp);
1592 set_ctrl(cmp, phi->in(0));
1593
1594 assert( cmp->is_Cmp(), "" );
1595 return (CmpNode*)cmp;
1596 }
1597
1598 //------------------------------sink_use---------------------------------------
1599 // If 'use' was in the loop-exit block, it now needs to be sunk
1600 // below the post-loop merge point.
1601 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) {
1602 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) {
1603 set_ctrl(use, post_loop);
1604 for (DUIterator j = use->outs(); use->has_out(j); j++)
1605 sink_use(use->out(j), post_loop);
1606 }
1607 }
1608
1609 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new,
1610 IdealLoopTree* loop, IdealLoopTree* outer_loop,
1611 Node_List*& split_if_set, Node_List*& split_bool_set,
1612 Node_List*& split_cex_set, Node_List& worklist,
1613 uint new_counter, CloneLoopMode mode) {
1614 Node* nnn = old_new[old->_idx];
1615 // Copy uses to a worklist, so I can munge the def-use info
1616 // with impunity.
1617 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
1618 worklist.push(old->fast_out(j));
1619
1620 while( worklist.size() ) {
1621 Node *use = worklist.pop();
1622 if (!has_node(use)) continue; // Ignore dead nodes
1623 if (use->in(0) == C->top()) continue;
1624 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
1625 // Check for data-use outside of loop - at least one of OLD or USE
1626 // must not be a CFG node.
1627 #ifdef ASSERT
1628 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == NULL) {
1629 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint();
1630 assert(mode == ControlAroundStripMined && use == sfpt, "missed a node");
1631 }
1632 #endif
1633 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) {
1634
1635 // If the Data use is an IF, that means we have an IF outside of the
1636 // loop that is switching on a condition that is set inside of the
1637 // loop. Happens if people set a loop-exit flag; then test the flag
1638 // in the loop to break the loop, then test is again outside of the
1639 // loop to determine which way the loop exited.
1640 // Loop predicate If node connects to Bool node through Opaque1 node.
1641 if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use) || use->Opcode() == Op_Opaque4) {
1642 // Since this code is highly unlikely, we lazily build the worklist
1643 // of such Nodes to go split.
1644 if (!split_if_set) {
1645 ResourceArea *area = Thread::current()->resource_area();
1646 split_if_set = new Node_List(area);
1647 }
1648 split_if_set->push(use);
1649 }
1650 if (use->is_Bool()) {
1651 if (!split_bool_set) {
1652 ResourceArea *area = Thread::current()->resource_area();
1653 split_bool_set = new Node_List(area);
1654 }
1655 split_bool_set->push(use);
1656 }
1657 if (use->Opcode() == Op_CreateEx) {
1658 if (!split_cex_set) {
1659 ResourceArea *area = Thread::current()->resource_area();
1660 split_cex_set = new Node_List(area);
1661 }
1662 split_cex_set->push(use);
1663 }
1664
1665
1666 // Get "block" use is in
1667 uint idx = 0;
1668 while( use->in(idx) != old ) idx++;
1669 Node *prev = use->is_CFG() ? use : get_ctrl(use);
1670 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" );
1671 Node *cfg = prev->_idx >= new_counter
1672 ? prev->in(2)
1673 : idom(prev);
1674 if( use->is_Phi() ) // Phi use is in prior block
1675 cfg = prev->in(idx); // NOT in block of Phi itself
1676 if (cfg->is_top()) { // Use is dead?
1677 _igvn.replace_input_of(use, idx, C->top());
1678 continue;
1679 }
1680
1681 // If use is referenced through control edge... (idx == 0)
1682 if (mode == IgnoreStripMined && idx == 0) {
1683 LoopNode *head = loop->_head->as_Loop();
1684 if (head->is_strip_mined() && is_dominator(head->outer_loop_exit(), prev)) {
1685 // That node is outside the inner loop, leave it outside the
1686 // outer loop as well to not confuse verification code.
1687 assert(!loop->_parent->is_member(use_loop), "should be out of the outer loop");
1688 _igvn.replace_input_of(use, 0, head->outer_loop_exit());
1689 continue;
1690 }
1691 }
1692
1693 while(!outer_loop->is_member(get_loop(cfg))) {
1694 prev = cfg;
1695 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg);
1696 }
1697 // If the use occurs after merging several exits from the loop, then
1698 // old value must have dominated all those exits. Since the same old
1699 // value was used on all those exits we did not need a Phi at this
1700 // merge point. NOW we do need a Phi here. Each loop exit value
1701 // is now merged with the peeled body exit; each exit gets its own
1702 // private Phi and those Phis need to be merged here.
1703 Node *phi;
1704 if( prev->is_Region() ) {
1705 if( idx == 0 ) { // Updating control edge?
1706 phi = prev; // Just use existing control
1707 } else { // Else need a new Phi
1708 phi = PhiNode::make( prev, old );
1709 // Now recursively fix up the new uses of old!
1710 for( uint i = 1; i < prev->req(); i++ ) {
1711 worklist.push(phi); // Onto worklist once for each 'old' input
1712 }
1713 }
1714 } else {
1715 // Get new RegionNode merging old and new loop exits
1716 prev = old_new[prev->_idx];
1717 assert( prev, "just made this in step 7" );
1718 if( idx == 0) { // Updating control edge?
1719 phi = prev; // Just use existing control
1720 } else { // Else need a new Phi
1721 // Make a new Phi merging data values properly
1722 phi = PhiNode::make( prev, old );
1723 phi->set_req( 1, nnn );
1724 }
1725 }
1726 // If inserting a new Phi, check for prior hits
1727 if( idx != 0 ) {
1728 Node *hit = _igvn.hash_find_insert(phi);
1729 if( hit == NULL ) {
1730 _igvn.register_new_node_with_optimizer(phi); // Register new phi
1731 } else { // or
1732 // Remove the new phi from the graph and use the hit
1733 _igvn.remove_dead_node(phi);
1734 phi = hit; // Use existing phi
1735 }
1736 set_ctrl(phi, prev);
1737 }
1738 // Make 'use' use the Phi instead of the old loop body exit value
1739 _igvn.replace_input_of(use, idx, phi);
1740 if( use->_idx >= new_counter ) { // If updating new phis
1741 // Not needed for correctness, but prevents a weak assert
1742 // in AddPNode from tripping (when we end up with different
1743 // base & derived Phis that will become the same after
1744 // IGVN does CSE).
1745 Node *hit = _igvn.hash_find_insert(use);
1746 if( hit ) // Go ahead and re-hash for hits.
1747 _igvn.replace_node( use, hit );
1748 }
1749
1750 // If 'use' was in the loop-exit block, it now needs to be sunk
1751 // below the post-loop merge point.
1752 sink_use( use, prev );
1753 }
1754 }
1755 }
1756
1757 static void clone_outer_loop_helper(Node* n, const IdealLoopTree *loop, const IdealLoopTree* outer_loop,
1758 const Node_List &old_new, Unique_Node_List& wq, PhaseIdealLoop* phase,
1759 bool check_old_new) {
1760 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1761 Node* u = n->fast_out(j);
1762 assert(check_old_new || old_new[u->_idx] == NULL, "shouldn't have been cloned");
1763 if (!u->is_CFG() && (!check_old_new || old_new[u->_idx] == NULL)) {
1764 Node* c = phase->get_ctrl(u);
1765 IdealLoopTree* u_loop = phase->get_loop(c);
1766 assert(!loop->is_member(u_loop), "can be in outer loop or out of both loops only");
1767 if (outer_loop->is_member(u_loop)) {
1768 wq.push(u);
1769 }
1770 }
1771 }
1772 }
1773
1774 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
1775 IdealLoopTree* outer_loop, int dd, Node_List &old_new,
1776 Node_List& extra_data_nodes) {
1777 if (head->is_strip_mined() && mode != IgnoreStripMined) {
1778 CountedLoopNode* cl = head->as_CountedLoop();
1779 Node* l = cl->outer_loop();
1780 Node* tail = cl->outer_loop_tail();
1781 IfNode* le = cl->outer_loop_end();
1782 Node* sfpt = cl->outer_safepoint();
1783 CountedLoopEndNode* cle = cl->loopexit();
1784 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop();
1785 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null();
1786 Node* cle_out = cle->proj_out(false);
1787
1788 Node* new_sfpt = NULL;
1789 Node* new_cle_out = cle_out->clone();
1790 old_new.map(cle_out->_idx, new_cle_out);
1791 if (mode == CloneIncludesStripMined) {
1792 // clone outer loop body
1793 Node* new_l = l->clone();
1794 Node* new_tail = tail->clone();
1795 IfNode* new_le = le->clone()->as_If();
1796 new_sfpt = sfpt->clone();
1797
1798 set_loop(new_l, outer_loop->_parent);
1799 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd);
1800 set_loop(new_cle_out, outer_loop->_parent);
1801 set_idom(new_cle_out, new_cle, dd);
1802 set_loop(new_sfpt, outer_loop->_parent);
1803 set_idom(new_sfpt, new_cle_out, dd);
1804 set_loop(new_le, outer_loop->_parent);
1805 set_idom(new_le, new_sfpt, dd);
1806 set_loop(new_tail, outer_loop->_parent);
1807 set_idom(new_tail, new_le, dd);
1808 set_idom(new_cl, new_l, dd);
1809
1810 old_new.map(l->_idx, new_l);
1811 old_new.map(tail->_idx, new_tail);
1812 old_new.map(le->_idx, new_le);
1813 old_new.map(sfpt->_idx, new_sfpt);
1814
1815 new_l->set_req(LoopNode::LoopBackControl, new_tail);
1816 new_l->set_req(0, new_l);
1817 new_tail->set_req(0, new_le);
1818 new_le->set_req(0, new_sfpt);
1819 new_sfpt->set_req(0, new_cle_out);
1820 new_cle_out->set_req(0, new_cle);
1821 new_cl->set_req(LoopNode::EntryControl, new_l);
1822
1823 _igvn.register_new_node_with_optimizer(new_l);
1824 _igvn.register_new_node_with_optimizer(new_tail);
1825 _igvn.register_new_node_with_optimizer(new_le);
1826 } else {
1827 Node *newhead = old_new[loop->_head->_idx];
1828 newhead->as_Loop()->clear_strip_mined();
1829 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl));
1830 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
1831 }
1832 // Look at data node that were assigned a control in the outer
1833 // loop: they are kept in the outer loop by the safepoint so start
1834 // from the safepoint node's inputs.
1835 IdealLoopTree* outer_loop = get_loop(l);
1836 Node_Stack stack(2);
1837 stack.push(sfpt, 1);
1838 uint new_counter = C->unique();
1839 while (stack.size() > 0) {
1840 Node* n = stack.node();
1841 uint i = stack.index();
1842 while (i < n->req() &&
1843 (n->in(i) == NULL ||
1844 !has_ctrl(n->in(i)) ||
1845 get_loop(get_ctrl(n->in(i))) != outer_loop ||
1846 (old_new[n->in(i)->_idx] != NULL && old_new[n->in(i)->_idx]->_idx >= new_counter))) {
1847 i++;
1848 }
1849 if (i < n->req()) {
1850 stack.set_index(i+1);
1851 stack.push(n->in(i), 0);
1852 } else {
1853 assert(old_new[n->_idx] == NULL || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet");
1854 Node* m = n == sfpt ? new_sfpt : n->clone();
1855 if (m != NULL) {
1856 for (uint i = 0; i < n->req(); i++) {
1857 if (m->in(i) != NULL && old_new[m->in(i)->_idx] != NULL) {
1858 m->set_req(i, old_new[m->in(i)->_idx]);
1859 }
1860 }
1861 } else {
1862 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?");
1863 }
1864 if (n != sfpt) {
1865 extra_data_nodes.push(n);
1866 _igvn.register_new_node_with_optimizer(m);
1867 assert(get_ctrl(n) == cle_out, "what other control?");
1868 set_ctrl(m, new_cle_out);
1869 old_new.map(n->_idx, m);
1870 }
1871 stack.pop();
1872 }
1873 }
1874 if (mode == CloneIncludesStripMined) {
1875 _igvn.register_new_node_with_optimizer(new_sfpt);
1876 _igvn.register_new_node_with_optimizer(new_cle_out);
1877 }
1878 // Some other transformation may have pessimistically assign some
1879 // data nodes to the outer loop. Set their control so they are out
1880 // of the outer loop.
1881 ResourceMark rm;
1882 Unique_Node_List wq;
1883 for (uint i = 0; i < extra_data_nodes.size(); i++) {
1884 Node* old = extra_data_nodes.at(i);
1885 clone_outer_loop_helper(old, loop, outer_loop, old_new, wq, this, true);
1886 }
1887 Node* new_ctrl = cl->outer_loop_exit();
1888 assert(get_loop(new_ctrl) != outer_loop, "must be out of the loop nest");
1889 for (uint i = 0; i < wq.size(); i++) {
1890 Node* n = wq.at(i);
1891 set_ctrl(n, new_ctrl);
1892 clone_outer_loop_helper(n, loop, outer_loop, old_new, wq, this, false);
1893 }
1894 } else {
1895 Node *newhead = old_new[loop->_head->_idx];
1896 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
1897 }
1898 }
1899
1900 //------------------------------clone_loop-------------------------------------
1901 //
1902 // C L O N E A L O O P B O D Y
1903 //
1904 // This is the basic building block of the loop optimizations. It clones an
1905 // entire loop body. It makes an old_new loop body mapping; with this mapping
1906 // you can find the new-loop equivalent to an old-loop node. All new-loop
1907 // nodes are exactly equal to their old-loop counterparts, all edges are the
1908 // same. All exits from the old-loop now have a RegionNode that merges the
1909 // equivalent new-loop path. This is true even for the normal "loop-exit"
1910 // condition. All uses of loop-invariant old-loop values now come from (one
1911 // or more) Phis that merge their new-loop equivalents.
1912 //
1913 // This operation leaves the graph in an illegal state: there are two valid
1914 // control edges coming from the loop pre-header to both loop bodies. I'll
1915 // definitely have to hack the graph after running this transform.
1916 //
1917 // From this building block I will further edit edges to perform loop peeling
1918 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
1919 //
1920 // Parameter side_by_size_idom:
1921 // When side_by_size_idom is NULL, the dominator tree is constructed for
1922 // the clone loop to dominate the original. Used in construction of
1923 // pre-main-post loop sequence.
1924 // When nonnull, the clone and original are side-by-side, both are
1925 // dominated by the side_by_side_idom node. Used in construction of
1926 // unswitched loops.
1927 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
1928 CloneLoopMode mode, Node* side_by_side_idom) {
1929
1930 LoopNode* head = loop->_head->as_Loop();
1931 head->verify_strip_mined(1);
1932
1933 if (C->do_vector_loop() && PrintOpto) {
1934 const char* mname = C->method()->name()->as_quoted_ascii();
1935 if (mname != NULL) {
1936 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname);
1937 }
1938 }
1939
1940 CloneMap& cm = C->clone_map();
1941 Dict* dict = cm.dict();
1942 if (C->do_vector_loop()) {
1943 cm.set_clone_idx(cm.max_gen()+1);
1944 #ifndef PRODUCT
1945 if (PrintOpto) {
1946 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx());
1947 loop->dump_head();
1948 }
1949 #endif
1950 }
1951
1952 // Step 1: Clone the loop body. Make the old->new mapping.
1953 uint i;
1954 for( i = 0; i < loop->_body.size(); i++ ) {
1955 Node *old = loop->_body.at(i);
1956 Node *nnn = old->clone();
1957 old_new.map( old->_idx, nnn );
1958 if (C->do_vector_loop()) {
1959 cm.verify_insert_and_clone(old, nnn, cm.clone_idx());
1960 }
1961 _igvn.register_new_node_with_optimizer(nnn);
1962 }
1963
1964 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop;
1965
1966 // Step 2: Fix the edges in the new body. If the old input is outside the
1967 // loop use it. If the old input is INside the loop, use the corresponding
1968 // new node instead.
1969 for( i = 0; i < loop->_body.size(); i++ ) {
1970 Node *old = loop->_body.at(i);
1971 Node *nnn = old_new[old->_idx];
1972 // Fix CFG/Loop controlling the new node
1973 if (has_ctrl(old)) {
1974 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
1975 } else {
1976 set_loop(nnn, outer_loop->_parent);
1977 if (old->outcnt() > 0) {
1978 set_idom( nnn, old_new[idom(old)->_idx], dd );
1979 }
1980 }
1981 // Correct edges to the new node
1982 for( uint j = 0; j < nnn->req(); j++ ) {
1983 Node *n = nnn->in(j);
1984 if( n ) {
1985 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n );
1986 if( loop->is_member( old_in_loop ) )
1987 nnn->set_req(j, old_new[n->_idx]);
1988 }
1989 }
1990 _igvn.hash_find_insert(nnn);
1991 }
1992
1993 ResourceArea *area = Thread::current()->resource_area();
1994 Node_List extra_data_nodes(area); // data nodes in the outer strip mined loop
1995 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes);
1996
1997 // Step 3: Now fix control uses. Loop varying control uses have already
1998 // been fixed up (as part of all input edges in Step 2). Loop invariant
1999 // control uses must be either an IfFalse or an IfTrue. Make a merge
2000 // point to merge the old and new IfFalse/IfTrue nodes; make the use
2001 // refer to this.
2002 Node_List worklist(area);
2003 uint new_counter = C->unique();
2004 for( i = 0; i < loop->_body.size(); i++ ) {
2005 Node* old = loop->_body.at(i);
2006 if( !old->is_CFG() ) continue;
2007
2008 // Copy uses to a worklist, so I can munge the def-use info
2009 // with impunity.
2010 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
2011 worklist.push(old->fast_out(j));
2012
2013 while( worklist.size() ) { // Visit all uses
2014 Node *use = worklist.pop();
2015 if (!has_node(use)) continue; // Ignore dead nodes
2016 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
2017 if( !loop->is_member( use_loop ) && use->is_CFG() ) {
2018 // Both OLD and USE are CFG nodes here.
2019 assert( use->is_Proj(), "" );
2020 Node* nnn = old_new[old->_idx];
2021
2022 Node* newuse = NULL;
2023 if (head->is_strip_mined() && mode != IgnoreStripMined) {
2024 CountedLoopNode* cl = head->as_CountedLoop();
2025 CountedLoopEndNode* cle = cl->loopexit();
2026 Node* cle_out = cle->proj_out_or_null(false);
2027 if (use == cle_out) {
2028 IfNode* le = cl->outer_loop_end();
2029 use = le->proj_out(false);
2030 use_loop = get_loop(use);
2031 if (mode == CloneIncludesStripMined) {
2032 nnn = old_new[le->_idx];
2033 } else {
2034 newuse = old_new[cle_out->_idx];
2035 }
2036 }
2037 }
2038 if (newuse == NULL) {
2039 newuse = use->clone();
2040 }
2041
2042 // Clone the loop exit control projection
2043 if (C->do_vector_loop()) {
2044 cm.verify_insert_and_clone(use, newuse, cm.clone_idx());
2045 }
2046 newuse->set_req(0,nnn);
2047 _igvn.register_new_node_with_optimizer(newuse);
2048 set_loop(newuse, use_loop);
2049 set_idom(newuse, nnn, dom_depth(nnn) + 1 );
2050
2051 // We need a Region to merge the exit from the peeled body and the
2052 // exit from the old loop body.
2053 RegionNode *r = new RegionNode(3);
2054 // Map the old use to the new merge point
2055 old_new.map( use->_idx, r );
2056 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use));
2057 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" );
2058
2059 // The original user of 'use' uses 'r' instead.
2060 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
2061 Node* useuse = use->last_out(l);
2062 _igvn.rehash_node_delayed(useuse);
2063 uint uses_found = 0;
2064 if( useuse->in(0) == use ) {
2065 useuse->set_req(0, r);
2066 uses_found++;
2067 if( useuse->is_CFG() ) {
2068 assert( dom_depth(useuse) > dd_r, "" );
2069 set_idom(useuse, r, dom_depth(useuse));
2070 }
2071 }
2072 for( uint k = 1; k < useuse->req(); k++ ) {
2073 if( useuse->in(k) == use ) {
2074 useuse->set_req(k, r);
2075 uses_found++;
2076 if (useuse->is_Loop() && k == LoopNode::EntryControl) {
2077 assert(dom_depth(useuse) > dd_r , "");
2078 set_idom(useuse, r, dom_depth(useuse));
2079 }
2080 }
2081 }
2082 l -= uses_found; // we deleted 1 or more copies of this edge
2083 }
2084
2085 // Now finish up 'r'
2086 r->set_req( 1, newuse );
2087 r->set_req( 2, use );
2088 _igvn.register_new_node_with_optimizer(r);
2089 set_loop(r, use_loop);
2090 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r);
2091 } // End of if a loop-exit test
2092 }
2093 }
2094
2095 // Step 4: If loop-invariant use is not control, it must be dominated by a
2096 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region
2097 // there if needed. Make a Phi there merging old and new used values.
2098 Node_List *split_if_set = NULL;
2099 Node_List *split_bool_set = NULL;
2100 Node_List *split_cex_set = NULL;
2101 for( i = 0; i < loop->_body.size(); i++ ) {
2102 Node* old = loop->_body.at(i);
2103 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2104 split_bool_set, split_cex_set, worklist, new_counter,
2105 mode);
2106 }
2107
2108 for (i = 0; i < extra_data_nodes.size(); i++) {
2109 Node* old = extra_data_nodes.at(i);
2110 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2111 split_bool_set, split_cex_set, worklist, new_counter,
2112 mode);
2113 }
2114
2115 // Check for IFs that need splitting/cloning. Happens if an IF outside of
2116 // the loop uses a condition set in the loop. The original IF probably
2117 // takes control from one or more OLD Regions (which in turn get from NEW
2118 // Regions). In any case, there will be a set of Phis for each merge point
2119 // from the IF up to where the original BOOL def exists the loop.
2120 if (split_if_set) {
2121 while (split_if_set->size()) {
2122 Node *iff = split_if_set->pop();
2123 if (iff->in(1)->is_Phi()) {
2124 Node *b = clone_iff(iff->in(1)->as_Phi(), loop);
2125 _igvn.replace_input_of(iff, 1, b);
2126 }
2127 }
2128 }
2129 if (split_bool_set) {
2130 while (split_bool_set->size()) {
2131 Node *b = split_bool_set->pop();
2132 Node *phi = b->in(1);
2133 assert(phi->is_Phi(), "");
2134 CmpNode *cmp = clone_bool((PhiNode*)phi, loop);
2135 _igvn.replace_input_of(b, 1, cmp);
2136 }
2137 }
2138 if (split_cex_set) {
2139 while (split_cex_set->size()) {
2140 Node *b = split_cex_set->pop();
2141 assert(b->in(0)->is_Region(), "");
2142 assert(b->in(1)->is_Phi(), "");
2143 assert(b->in(0)->in(0) == b->in(1)->in(0), "");
2144 split_up(b, b->in(0), NULL);
2145 }
2146 }
2147
2148 }
2149
2150
2151 //---------------------- stride_of_possible_iv -------------------------------------
2152 // Looks for an iff/bool/comp with one operand of the compare
2153 // being a cycle involving an add and a phi,
2154 // with an optional truncation (left-shift followed by a right-shift)
2155 // of the add. Returns zero if not an iv.
2156 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
2157 Node* trunc1 = NULL;
2158 Node* trunc2 = NULL;
2159 const TypeInt* ttype = NULL;
2160 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) {
2161 return 0;
2162 }
2163 BoolNode* bl = iff->in(1)->as_Bool();
2164 Node* cmp = bl->in(1);
2165 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) {
2166 return 0;
2167 }
2168 // Must have an invariant operand
2169 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) {
2170 return 0;
2171 }
2172 Node* add2 = NULL;
2173 Node* cmp1 = cmp->in(1);
2174 if (cmp1->is_Phi()) {
2175 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
2176 Node* phi = cmp1;
2177 for (uint i = 1; i < phi->req(); i++) {
2178 Node* in = phi->in(i);
2179 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in,
2180 &trunc1, &trunc2, &ttype);
2181 if (add && add->in(1) == phi) {
2182 add2 = add->in(2);
2183 break;
2184 }
2185 }
2186 } else {
2187 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
2188 Node* addtrunc = cmp1;
2189 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc,
2190 &trunc1, &trunc2, &ttype);
2191 if (add && add->in(1)->is_Phi()) {
2192 Node* phi = add->in(1);
2193 for (uint i = 1; i < phi->req(); i++) {
2194 if (phi->in(i) == addtrunc) {
2195 add2 = add->in(2);
2196 break;
2197 }
2198 }
2199 }
2200 }
2201 if (add2 != NULL) {
2202 const TypeInt* add2t = _igvn.type(add2)->is_int();
2203 if (add2t->is_con()) {
2204 return add2t->get_con();
2205 }
2206 }
2207 return 0;
2208 }
2209
2210
2211 //---------------------- stay_in_loop -------------------------------------
2212 // Return the (unique) control output node that's in the loop (if it exists.)
2213 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
2214 Node* unique = NULL;
2215 if (!n) return NULL;
2216 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2217 Node* use = n->fast_out(i);
2218 if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
2219 if (unique != NULL) {
2220 return NULL;
2221 }
2222 unique = use;
2223 }
2224 }
2225 return unique;
2226 }
2227
2228 //------------------------------ register_node -------------------------------------
2229 // Utility to register node "n" with PhaseIdealLoop
2230 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) {
2231 _igvn.register_new_node_with_optimizer(n);
2232 loop->_body.push(n);
2233 if (n->is_CFG()) {
2234 set_loop(n, loop);
2235 set_idom(n, pred, ddepth);
2236 } else {
2237 set_ctrl(n, pred);
2238 }
2239 }
2240
2241 //------------------------------ proj_clone -------------------------------------
2242 // Utility to create an if-projection
2243 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
2244 ProjNode* c = p->clone()->as_Proj();
2245 c->set_req(0, iff);
2246 return c;
2247 }
2248
2249 //------------------------------ short_circuit_if -------------------------------------
2250 // Force the iff control output to be the live_proj
2251 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
2252 guarantee(live_proj != NULL, "null projection");
2253 int proj_con = live_proj->_con;
2254 assert(proj_con == 0 || proj_con == 1, "false or true projection");
2255 Node *con = _igvn.intcon(proj_con);
2256 set_ctrl(con, C->root());
2257 if (iff) {
2258 iff->set_req(1, con);
2259 }
2260 return con;
2261 }
2262
2263 //------------------------------ insert_if_before_proj -------------------------------------
2264 // Insert a new if before an if projection (* - new node)
2265 //
2266 // before
2267 // if(test)
2268 // / \
2269 // v v
2270 // other-proj proj (arg)
2271 //
2272 // after
2273 // if(test)
2274 // / \
2275 // / v
2276 // | * proj-clone
2277 // v |
2278 // other-proj v
2279 // * new_if(relop(cmp[IU](left,right)))
2280 // / \
2281 // v v
2282 // * new-proj proj
2283 // (returned)
2284 //
2285 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) {
2286 IfNode* iff = proj->in(0)->as_If();
2287 IdealLoopTree *loop = get_loop(proj);
2288 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2289 int ddepth = dom_depth(proj);
2290
2291 _igvn.rehash_node_delayed(iff);
2292 _igvn.rehash_node_delayed(proj);
2293
2294 proj->set_req(0, NULL); // temporary disconnect
2295 ProjNode* proj2 = proj_clone(proj, iff);
2296 register_node(proj2, loop, iff, ddepth);
2297
2298 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right);
2299 register_node(cmp, loop, proj2, ddepth);
2300
2301 BoolNode* bol = new BoolNode(cmp, relop);
2302 register_node(bol, loop, proj2, ddepth);
2303
2304 int opcode = iff->Opcode();
2305 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode");
2306 IfNode* new_if = (opcode == Op_If) ? new IfNode(proj2, bol, iff->_prob, iff->_fcnt):
2307 new RangeCheckNode(proj2, bol, iff->_prob, iff->_fcnt);
2308 register_node(new_if, loop, proj2, ddepth);
2309
2310 proj->set_req(0, new_if); // reattach
2311 set_idom(proj, new_if, ddepth);
2312
2313 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
2314 guarantee(new_exit != NULL, "null exit node");
2315 register_node(new_exit, get_loop(other_proj), new_if, ddepth);
2316
2317 return new_exit;
2318 }
2319
2320 //------------------------------ insert_region_before_proj -------------------------------------
2321 // Insert a region before an if projection (* - new node)
2322 //
2323 // before
2324 // if(test)
2325 // / |
2326 // v |
2327 // proj v
2328 // other-proj
2329 //
2330 // after
2331 // if(test)
2332 // / |
2333 // v |
2334 // * proj-clone v
2335 // | other-proj
2336 // v
2337 // * new-region
2338 // |
2339 // v
2340 // * dum_if
2341 // / \
2342 // v \
2343 // * dum-proj v
2344 // proj
2345 //
2346 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
2347 IfNode* iff = proj->in(0)->as_If();
2348 IdealLoopTree *loop = get_loop(proj);
2349 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2350 int ddepth = dom_depth(proj);
2351
2352 _igvn.rehash_node_delayed(iff);
2353 _igvn.rehash_node_delayed(proj);
2354
2355 proj->set_req(0, NULL); // temporary disconnect
2356 ProjNode* proj2 = proj_clone(proj, iff);
2357 register_node(proj2, loop, iff, ddepth);
2358
2359 RegionNode* reg = new RegionNode(2);
2360 reg->set_req(1, proj2);
2361 register_node(reg, loop, iff, ddepth);
2362
2363 IfNode* dum_if = new IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt);
2364 register_node(dum_if, loop, reg, ddepth);
2365
2366 proj->set_req(0, dum_if); // reattach
2367 set_idom(proj, dum_if, ddepth);
2368
2369 ProjNode* dum_proj = proj_clone(other_proj, dum_if);
2370 register_node(dum_proj, loop, dum_if, ddepth);
2371
2372 return reg;
2373 }
2374
2375 //------------------------------ insert_cmpi_loop_exit -------------------------------------
2376 // Clone a signed compare loop exit from an unsigned compare and
2377 // insert it before the unsigned cmp on the stay-in-loop path.
2378 // All new nodes inserted in the dominator tree between the original
2379 // if and it's projections. The original if test is replaced with
2380 // a constant to force the stay-in-loop path.
2381 //
2382 // This is done to make sure that the original if and it's projections
2383 // still dominate the same set of control nodes, that the ctrl() relation
2384 // from data nodes to them is preserved, and that their loop nesting is
2385 // preserved.
2386 //
2387 // before
2388 // if(i <u limit) unsigned compare loop exit
2389 // / |
2390 // v v
2391 // exit-proj stay-in-loop-proj
2392 //
2393 // after
2394 // if(stay-in-loop-const) original if
2395 // / |
2396 // / v
2397 // / if(i < limit) new signed test
2398 // / / |
2399 // / / v
2400 // / / if(i <u limit) new cloned unsigned test
2401 // / / / |
2402 // v v v |
2403 // region |
2404 // | |
2405 // dum-if |
2406 // / | |
2407 // ether | |
2408 // v v
2409 // exit-proj stay-in-loop-proj
2410 //
2411 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) {
2412 const bool Signed = true;
2413 const bool Unsigned = false;
2414
2415 BoolNode* bol = if_cmpu->in(1)->as_Bool();
2416 if (bol->_test._test != BoolTest::lt) return NULL;
2417 CmpNode* cmpu = bol->in(1)->as_Cmp();
2418 if (cmpu->Opcode() != Op_CmpU) return NULL;
2419 int stride = stride_of_possible_iv(if_cmpu);
2420 if (stride == 0) return NULL;
2421
2422 Node* lp_proj = stay_in_loop(if_cmpu, loop);
2423 guarantee(lp_proj != NULL, "null loop node");
2424
2425 ProjNode* lp_continue = lp_proj->as_Proj();
2426 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
2427
2428 Node* limit = NULL;
2429 if (stride > 0) {
2430 limit = cmpu->in(2);
2431 } else {
2432 limit = _igvn.makecon(TypeInt::ZERO);
2433 set_ctrl(limit, C->root());
2434 }
2435 // Create a new region on the exit path
2436 RegionNode* reg = insert_region_before_proj(lp_exit);
2437 guarantee(reg != NULL, "null region node");
2438
2439 // Clone the if-cmpu-true-false using a signed compare
2440 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
2441 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue);
2442 reg->add_req(cmpi_exit);
2443
2444 // Clone the if-cmpu-true-false
2445 BoolTest::mask rel_u = bol->_test._test;
2446 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
2447 reg->add_req(cmpu_exit);
2448
2449 // Force original if to stay in loop.
2450 short_circuit_if(if_cmpu, lp_continue);
2451
2452 return cmpi_exit->in(0)->as_If();
2453 }
2454
2455 //------------------------------ remove_cmpi_loop_exit -------------------------------------
2456 // Remove a previously inserted signed compare loop exit.
2457 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
2458 Node* lp_proj = stay_in_loop(if_cmp, loop);
2459 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
2460 stay_in_loop(lp_proj, loop)->is_If() &&
2461 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
2462 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
2463 set_ctrl(con, C->root());
2464 if_cmp->set_req(1, con);
2465 }
2466
2467 //------------------------------ scheduled_nodelist -------------------------------------
2468 // Create a post order schedule of nodes that are in the
2469 // "member" set. The list is returned in "sched".
2470 // The first node in "sched" is the loop head, followed by
2471 // nodes which have no inputs in the "member" set, and then
2472 // followed by the nodes that have an immediate input dependence
2473 // on a node in "sched".
2474 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
2475
2476 assert(member.test(loop->_head->_idx), "loop head must be in member set");
2477 Arena *a = Thread::current()->resource_area();
2478 VectorSet visited(a);
2479 Node_Stack nstack(a, loop->_body.size());
2480
2481 Node* n = loop->_head; // top of stack is cached in "n"
2482 uint idx = 0;
2483 visited.set(n->_idx);
2484
2485 // Initially push all with no inputs from within member set
2486 for(uint i = 0; i < loop->_body.size(); i++ ) {
2487 Node *elt = loop->_body.at(i);
2488 if (member.test(elt->_idx)) {
2489 bool found = false;
2490 for (uint j = 0; j < elt->req(); j++) {
2491 Node* def = elt->in(j);
2492 if (def && member.test(def->_idx) && def != elt) {
2493 found = true;
2494 break;
2495 }
2496 }
2497 if (!found && elt != loop->_head) {
2498 nstack.push(n, idx);
2499 n = elt;
2500 assert(!visited.test(n->_idx), "not seen yet");
2501 visited.set(n->_idx);
2502 }
2503 }
2504 }
2505
2506 // traverse out's that are in the member set
2507 while (true) {
2508 if (idx < n->outcnt()) {
2509 Node* use = n->raw_out(idx);
2510 idx++;
2511 if (!visited.test_set(use->_idx)) {
2512 if (member.test(use->_idx)) {
2513 nstack.push(n, idx);
2514 n = use;
2515 idx = 0;
2516 }
2517 }
2518 } else {
2519 // All outputs processed
2520 sched.push(n);
2521 if (nstack.is_empty()) break;
2522 n = nstack.node();
2523 idx = nstack.index();
2524 nstack.pop();
2525 }
2526 }
2527 }
2528
2529
2530 //------------------------------ has_use_in_set -------------------------------------
2531 // Has a use in the vector set
2532 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
2533 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2534 Node* use = n->fast_out(j);
2535 if (vset.test(use->_idx)) {
2536 return true;
2537 }
2538 }
2539 return false;
2540 }
2541
2542
2543 //------------------------------ has_use_internal_to_set -------------------------------------
2544 // Has use internal to the vector set (ie. not in a phi at the loop head)
2545 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
2546 Node* head = loop->_head;
2547 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2548 Node* use = n->fast_out(j);
2549 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
2550 return true;
2551 }
2552 }
2553 return false;
2554 }
2555
2556
2557 //------------------------------ clone_for_use_outside_loop -------------------------------------
2558 // clone "n" for uses that are outside of loop
2559 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
2560 int cloned = 0;
2561 assert(worklist.size() == 0, "should be empty");
2562 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2563 Node* use = n->fast_out(j);
2564 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
2565 worklist.push(use);
2566 }
2567 }
2568 while( worklist.size() ) {
2569 Node *use = worklist.pop();
2570 if (!has_node(use) || use->in(0) == C->top()) continue;
2571 uint j;
2572 for (j = 0; j < use->req(); j++) {
2573 if (use->in(j) == n) break;
2574 }
2575 assert(j < use->req(), "must be there");
2576
2577 // clone "n" and insert it between the inputs of "n" and the use outside the loop
2578 Node* n_clone = n->clone();
2579 _igvn.replace_input_of(use, j, n_clone);
2580 cloned++;
2581 Node* use_c;
2582 if (!use->is_Phi()) {
2583 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
2584 } else {
2585 // Use in a phi is considered a use in the associated predecessor block
2586 use_c = use->in(0)->in(j);
2587 }
2588 set_ctrl(n_clone, use_c);
2589 assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
2590 get_loop(use_c)->_body.push(n_clone);
2591 _igvn.register_new_node_with_optimizer(n_clone);
2592 #if !defined(PRODUCT)
2593 if (TracePartialPeeling) {
2594 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
2595 }
2596 #endif
2597 }
2598 return cloned;
2599 }
2600
2601
2602 //------------------------------ clone_for_special_use_inside_loop -------------------------------------
2603 // clone "n" for special uses that are in the not_peeled region.
2604 // If these def-uses occur in separate blocks, the code generator
2605 // marks the method as not compilable. For example, if a "BoolNode"
2606 // is in a different basic block than the "IfNode" that uses it, then
2607 // the compilation is aborted in the code generator.
2608 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
2609 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
2610 if (n->is_Phi() || n->is_Load()) {
2611 return;
2612 }
2613 assert(worklist.size() == 0, "should be empty");
2614 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2615 Node* use = n->fast_out(j);
2616 if ( not_peel.test(use->_idx) &&
2617 (use->is_If() || use->is_CMove() || use->is_Bool()) &&
2618 use->in(1) == n) {
2619 worklist.push(use);
2620 }
2621 }
2622 if (worklist.size() > 0) {
2623 // clone "n" and insert it between inputs of "n" and the use
2624 Node* n_clone = n->clone();
2625 loop->_body.push(n_clone);
2626 _igvn.register_new_node_with_optimizer(n_clone);
2627 set_ctrl(n_clone, get_ctrl(n));
2628 sink_list.push(n_clone);
2629 not_peel <<= n_clone->_idx; // add n_clone to not_peel set.
2630 #if !defined(PRODUCT)
2631 if (TracePartialPeeling) {
2632 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
2633 }
2634 #endif
2635 while( worklist.size() ) {
2636 Node *use = worklist.pop();
2637 _igvn.rehash_node_delayed(use);
2638 for (uint j = 1; j < use->req(); j++) {
2639 if (use->in(j) == n) {
2640 use->set_req(j, n_clone);
2641 }
2642 }
2643 }
2644 }
2645 }
2646
2647
2648 //------------------------------ insert_phi_for_loop -------------------------------------
2649 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
2650 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
2651 Node *phi = PhiNode::make(lp, back_edge_val);
2652 phi->set_req(LoopNode::EntryControl, lp_entry_val);
2653 // Use existing phi if it already exists
2654 Node *hit = _igvn.hash_find_insert(phi);
2655 if( hit == NULL ) {
2656 _igvn.register_new_node_with_optimizer(phi);
2657 set_ctrl(phi, lp);
2658 } else {
2659 // Remove the new phi from the graph and use the hit
2660 _igvn.remove_dead_node(phi);
2661 phi = hit;
2662 }
2663 _igvn.replace_input_of(use, idx, phi);
2664 }
2665
2666 #ifdef ASSERT
2667 //------------------------------ is_valid_loop_partition -------------------------------------
2668 // Validate the loop partition sets: peel and not_peel
2669 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
2670 VectorSet& not_peel ) {
2671 uint i;
2672 // Check that peel_list entries are in the peel set
2673 for (i = 0; i < peel_list.size(); i++) {
2674 if (!peel.test(peel_list.at(i)->_idx)) {
2675 return false;
2676 }
2677 }
2678 // Check at loop members are in one of peel set or not_peel set
2679 for (i = 0; i < loop->_body.size(); i++ ) {
2680 Node *def = loop->_body.at(i);
2681 uint di = def->_idx;
2682 // Check that peel set elements are in peel_list
2683 if (peel.test(di)) {
2684 if (not_peel.test(di)) {
2685 return false;
2686 }
2687 // Must be in peel_list also
2688 bool found = false;
2689 for (uint j = 0; j < peel_list.size(); j++) {
2690 if (peel_list.at(j)->_idx == di) {
2691 found = true;
2692 break;
2693 }
2694 }
2695 if (!found) {
2696 return false;
2697 }
2698 } else if (not_peel.test(di)) {
2699 if (peel.test(di)) {
2700 return false;
2701 }
2702 } else {
2703 return false;
2704 }
2705 }
2706 return true;
2707 }
2708
2709 //------------------------------ is_valid_clone_loop_exit_use -------------------------------------
2710 // Ensure a use outside of loop is of the right form
2711 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
2712 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2713 return (use->is_Phi() &&
2714 use_c->is_Region() && use_c->req() == 3 &&
2715 (use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
2716 use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
2717 use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
2718 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
2719 }
2720
2721 //------------------------------ is_valid_clone_loop_form -------------------------------------
2722 // Ensure that all uses outside of loop are of the right form
2723 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
2724 uint orig_exit_idx, uint clone_exit_idx) {
2725 uint len = peel_list.size();
2726 for (uint i = 0; i < len; i++) {
2727 Node *def = peel_list.at(i);
2728
2729 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2730 Node *use = def->fast_out(j);
2731 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2732 if (!loop->is_member(get_loop(use_c))) {
2733 // use is not in the loop, check for correct structure
2734 if (use->in(0) == def) {
2735 // Okay
2736 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
2737 return false;
2738 }
2739 }
2740 }
2741 }
2742 return true;
2743 }
2744 #endif
2745
2746 //------------------------------ partial_peel -------------------------------------
2747 // Partially peel (aka loop rotation) the top portion of a loop (called
2748 // the peel section below) by cloning it and placing one copy just before
2749 // the new loop head and the other copy at the bottom of the new loop.
2750 //
2751 // before after where it came from
2752 //
2753 // stmt1 stmt1
2754 // loop: stmt2 clone
2755 // stmt2 if condA goto exitA clone
2756 // if condA goto exitA new_loop: new
2757 // stmt3 stmt3 clone
2758 // if !condB goto loop if condB goto exitB clone
2759 // exitB: stmt2 orig
2760 // stmt4 if !condA goto new_loop orig
2761 // exitA: goto exitA
2762 // exitB:
2763 // stmt4
2764 // exitA:
2765 //
2766 // Step 1: find the cut point: an exit test on probable
2767 // induction variable.
2768 // Step 2: schedule (with cloning) operations in the peel
2769 // section that can be executed after the cut into
2770 // the section that is not peeled. This may need
2771 // to clone operations into exit blocks. For
2772 // instance, a reference to A[i] in the not-peel
2773 // section and a reference to B[i] in an exit block
2774 // may cause a left-shift of i by 2 to be placed
2775 // in the peel block. This step will clone the left
2776 // shift into the exit block and sink the left shift
2777 // from the peel to the not-peel section.
2778 // Step 3: clone the loop, retarget the control, and insert
2779 // phis for values that are live across the new loop
2780 // head. This is very dependent on the graph structure
2781 // from clone_loop. It creates region nodes for
2782 // exit control and associated phi nodes for values
2783 // flow out of the loop through that exit. The region
2784 // node is dominated by the clone's control projection.
2785 // So the clone's peel section is placed before the
2786 // new loop head, and the clone's not-peel section is
2787 // forms the top part of the new loop. The original
2788 // peel section forms the tail of the new loop.
2789 // Step 4: update the dominator tree and recompute the
2790 // dominator depth.
2791 //
2792 // orig
2793 //
2794 // stmt1
2795 // |
2796 // v
2797 // loop predicate
2798 // |
2799 // v
2800 // loop<----+
2801 // | |
2802 // stmt2 |
2803 // | |
2804 // v |
2805 // ifA |
2806 // / | |
2807 // v v |
2808 // false true ^ <-- last_peel
2809 // / | |
2810 // / ===|==cut |
2811 // / stmt3 | <-- first_not_peel
2812 // / | |
2813 // | v |
2814 // v ifB |
2815 // exitA: / \ |
2816 // / \ |
2817 // v v |
2818 // false true |
2819 // / \ |
2820 // / ----+
2821 // |
2822 // v
2823 // exitB:
2824 // stmt4
2825 //
2826 //
2827 // after clone loop
2828 //
2829 // stmt1
2830 // |
2831 // v
2832 // loop predicate
2833 // / \
2834 // clone / \ orig
2835 // / \
2836 // / \
2837 // v v
2838 // +---->loop loop<----+
2839 // | | | |
2840 // | stmt2 stmt2 |
2841 // | | | |
2842 // | v v |
2843 // | ifA ifA |
2844 // | | \ / | |
2845 // | v v v v |
2846 // ^ true false false true ^ <-- last_peel
2847 // | | ^ \ / | |
2848 // | cut==|== \ \ / ===|==cut |
2849 // | stmt3 \ \ / stmt3 | <-- first_not_peel
2850 // | | dom | | | |
2851 // | v \ 1v v2 v |
2852 // | ifB regionA ifB |
2853 // | / \ | / \ |
2854 // | / \ v / \ |
2855 // | v v exitA: v v |
2856 // | true false false true |
2857 // | / ^ \ / \ |
2858 // +---- \ \ / ----+
2859 // dom \ /
2860 // \ 1v v2
2861 // regionB
2862 // |
2863 // v
2864 // exitB:
2865 // stmt4
2866 //
2867 //
2868 // after partial peel
2869 //
2870 // stmt1
2871 // |
2872 // v
2873 // loop predicate
2874 // /
2875 // clone / orig
2876 // / TOP
2877 // / \
2878 // v v
2879 // TOP->loop loop----+
2880 // | | |
2881 // stmt2 stmt2 |
2882 // | | |
2883 // v v |
2884 // ifA ifA |
2885 // | \ / | |
2886 // v v v v |
2887 // true false false true | <-- last_peel
2888 // | ^ \ / +------|---+
2889 // +->newloop \ \ / === ==cut | |
2890 // | stmt3 \ \ / TOP | |
2891 // | | dom | | stmt3 | | <-- first_not_peel
2892 // | v \ 1v v2 v | |
2893 // | ifB regionA ifB ^ v
2894 // | / \ | / \ | |
2895 // | / \ v / \ | |
2896 // | v v exitA: v v | |
2897 // | true false false true | |
2898 // | / ^ \ / \ | |
2899 // | | \ \ / v | |
2900 // | | dom \ / TOP | |
2901 // | | \ 1v v2 | |
2902 // ^ v regionB | |
2903 // | | | | |
2904 // | | v ^ v
2905 // | | exitB: | |
2906 // | | stmt4 | |
2907 // | +------------>-----------------+ |
2908 // | |
2909 // +-----------------<---------------------+
2910 //
2911 //
2912 // final graph
2913 //
2914 // stmt1
2915 // |
2916 // v
2917 // loop predicate
2918 // |
2919 // v
2920 // stmt2 clone
2921 // |
2922 // v
2923 // ........> ifA clone
2924 // : / |
2925 // dom / |
2926 // : v v
2927 // : false true
2928 // : | |
2929 // : | v
2930 // : | newloop<-----+
2931 // : | | |
2932 // : | stmt3 clone |
2933 // : | | |
2934 // : | v |
2935 // : | ifB |
2936 // : | / \ |
2937 // : | v v |
2938 // : | false true |
2939 // : | | | |
2940 // : | v stmt2 |
2941 // : | exitB: | |
2942 // : | stmt4 v |
2943 // : | ifA orig |
2944 // : | / \ |
2945 // : | / \ |
2946 // : | v v |
2947 // : | false true |
2948 // : | / \ |
2949 // : v v -----+
2950 // RegionA
2951 // |
2952 // v
2953 // exitA
2954 //
2955 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
2956
2957 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only");
2958 if (!loop->_head->is_Loop()) {
2959 return false; }
2960
2961 LoopNode *head = loop->_head->as_Loop();
2962
2963 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
2964 return false;
2965 }
2966
2967 // Check for complex exit control
2968 for(uint ii = 0; ii < loop->_body.size(); ii++ ) {
2969 Node *n = loop->_body.at(ii);
2970 int opc = n->Opcode();
2971 if (n->is_Call() ||
2972 opc == Op_Catch ||
2973 opc == Op_CatchProj ||
2974 opc == Op_Jump ||
2975 opc == Op_JumpProj) {
2976 #if !defined(PRODUCT)
2977 if (TracePartialPeeling) {
2978 tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
2979 }
2980 #endif
2981 return false;
2982 }
2983 }
2984
2985 int dd = dom_depth(head);
2986
2987 // Step 1: find cut point
2988
2989 // Walk up dominators to loop head looking for first loop exit
2990 // which is executed on every path thru loop.
2991 IfNode *peel_if = NULL;
2992 IfNode *peel_if_cmpu = NULL;
2993
2994 Node *iff = loop->tail();
2995 while( iff != head ) {
2996 if( iff->is_If() ) {
2997 Node *ctrl = get_ctrl(iff->in(1));
2998 if (ctrl->is_top()) return false; // Dead test on live IF.
2999 // If loop-varying exit-test, check for induction variable
3000 if( loop->is_member(get_loop(ctrl)) &&
3001 loop->is_loop_exit(iff) &&
3002 is_possible_iv_test(iff)) {
3003 Node* cmp = iff->in(1)->in(1);
3004 if (cmp->Opcode() == Op_CmpI) {
3005 peel_if = iff->as_If();
3006 } else {
3007 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
3008 peel_if_cmpu = iff->as_If();
3009 }
3010 }
3011 }
3012 iff = idom(iff);
3013 }
3014 // Prefer signed compare over unsigned compare.
3015 IfNode* new_peel_if = NULL;
3016 if (peel_if == NULL) {
3017 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) {
3018 return false; // No peel point found
3019 }
3020 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
3021 if (new_peel_if == NULL) {
3022 return false; // No peel point found
3023 }
3024 peel_if = new_peel_if;
3025 }
3026 Node* last_peel = stay_in_loop(peel_if, loop);
3027 Node* first_not_peeled = stay_in_loop(last_peel, loop);
3028 if (first_not_peeled == NULL || first_not_peeled == head) {
3029 return false;
3030 }
3031
3032 #if !defined(PRODUCT)
3033 if (TraceLoopOpts) {
3034 tty->print("PartialPeel ");
3035 loop->dump_head();
3036 }
3037
3038 if (TracePartialPeeling) {
3039 tty->print_cr("before partial peel one iteration");
3040 Node_List wl;
3041 Node* t = head->in(2);
3042 while (true) {
3043 wl.push(t);
3044 if (t == head) break;
3045 t = idom(t);
3046 }
3047 while (wl.size() > 0) {
3048 Node* tt = wl.pop();
3049 tt->dump();
3050 if (tt == last_peel) tty->print_cr("-- cut --");
3051 }
3052 }
3053 #endif
3054 ResourceArea *area = Thread::current()->resource_area();
3055 VectorSet peel(area);
3056 VectorSet not_peel(area);
3057 Node_List peel_list(area);
3058 Node_List worklist(area);
3059 Node_List sink_list(area);
3060
3061 // Set of cfg nodes to peel are those that are executable from
3062 // the head through last_peel.
3063 assert(worklist.size() == 0, "should be empty");
3064 worklist.push(head);
3065 peel.set(head->_idx);
3066 while (worklist.size() > 0) {
3067 Node *n = worklist.pop();
3068 if (n != last_peel) {
3069 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3070 Node* use = n->fast_out(j);
3071 if (use->is_CFG() &&
3072 loop->is_member(get_loop(use)) &&
3073 !peel.test_set(use->_idx)) {
3074 worklist.push(use);
3075 }
3076 }
3077 }
3078 }
3079
3080 // Set of non-cfg nodes to peel are those that are control
3081 // dependent on the cfg nodes.
3082 uint i;
3083 for(i = 0; i < loop->_body.size(); i++ ) {
3084 Node *n = loop->_body.at(i);
3085 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
3086 if (peel.test(n_c->_idx)) {
3087 peel.set(n->_idx);
3088 } else {
3089 not_peel.set(n->_idx);
3090 }
3091 }
3092
3093 // Step 2: move operations from the peeled section down into the
3094 // not-peeled section
3095
3096 // Get a post order schedule of nodes in the peel region
3097 // Result in right-most operand.
3098 scheduled_nodelist(loop, peel, peel_list );
3099
3100 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
3101
3102 // For future check for too many new phis
3103 uint old_phi_cnt = 0;
3104 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
3105 Node* use = head->fast_out(j);
3106 if (use->is_Phi()) old_phi_cnt++;
3107 }
3108
3109 #if !defined(PRODUCT)
3110 if (TracePartialPeeling) {
3111 tty->print_cr("\npeeled list");
3112 }
3113 #endif
3114
3115 // Evacuate nodes in peel region into the not_peeled region if possible
3116 uint new_phi_cnt = 0;
3117 uint cloned_for_outside_use = 0;
3118 for (i = 0; i < peel_list.size();) {
3119 Node* n = peel_list.at(i);
3120 #if !defined(PRODUCT)
3121 if (TracePartialPeeling) n->dump();
3122 #endif
3123 bool incr = true;
3124 if ( !n->is_CFG() ) {
3125
3126 if ( has_use_in_set(n, not_peel) ) {
3127
3128 // If not used internal to the peeled region,
3129 // move "n" from peeled to not_peeled region.
3130
3131 if ( !has_use_internal_to_set(n, peel, loop) ) {
3132
3133 // if not pinned and not a load (which maybe anti-dependent on a store)
3134 // and not a CMove (Matcher expects only bool->cmove).
3135 if ( n->in(0) == NULL && !n->is_Load() && !n->is_CMove() ) {
3136 cloned_for_outside_use += clone_for_use_outside_loop( loop, n, worklist );
3137 sink_list.push(n);
3138 peel >>= n->_idx; // delete n from peel set.
3139 not_peel <<= n->_idx; // add n to not_peel set.
3140 peel_list.remove(i);
3141 incr = false;
3142 #if !defined(PRODUCT)
3143 if (TracePartialPeeling) {
3144 tty->print_cr("sink to not_peeled region: %d newbb: %d",
3145 n->_idx, get_ctrl(n)->_idx);
3146 }
3147 #endif
3148 }
3149 } else {
3150 // Otherwise check for special def-use cases that span
3151 // the peel/not_peel boundary such as bool->if
3152 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist );
3153 new_phi_cnt++;
3154 }
3155 }
3156 }
3157 if (incr) i++;
3158 }
3159
3160 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) {
3161 #if !defined(PRODUCT)
3162 if (TracePartialPeeling) {
3163 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c",
3164 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F');
3165 }
3166 #endif
3167 if (new_peel_if != NULL) {
3168 remove_cmpi_loop_exit(new_peel_if, loop);
3169 }
3170 // Inhibit more partial peeling on this loop
3171 assert(!head->is_partial_peel_loop(), "not partial peeled");
3172 head->mark_partial_peel_failed();
3173 if (cloned_for_outside_use > 0) {
3174 // Terminate this round of loop opts because
3175 // the graph outside this loop was changed.
3176 C->set_major_progress();
3177 return true;
3178 }
3179 return false;
3180 }
3181
3182 // Step 3: clone loop, retarget control, and insert new phis
3183
3184 // Create new loop head for new phis and to hang
3185 // the nodes being moved (sinked) from the peel region.
3186 LoopNode* new_head = new LoopNode(last_peel, last_peel);
3187 new_head->set_unswitch_count(head->unswitch_count()); // Preserve
3188 _igvn.register_new_node_with_optimizer(new_head);
3189 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
3190 _igvn.replace_input_of(first_not_peeled, 0, new_head);
3191 set_loop(new_head, loop);
3192 loop->_body.push(new_head);
3193 not_peel.set(new_head->_idx);
3194 set_idom(new_head, last_peel, dom_depth(first_not_peeled));
3195 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
3196
3197 while (sink_list.size() > 0) {
3198 Node* n = sink_list.pop();
3199 set_ctrl(n, new_head);
3200 }
3201
3202 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
3203
3204 clone_loop(loop, old_new, dd, IgnoreStripMined);
3205
3206 const uint clone_exit_idx = 1;
3207 const uint orig_exit_idx = 2;
3208 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop");
3209
3210 Node* head_clone = old_new[head->_idx];
3211 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop();
3212 Node* orig_tail_clone = head_clone->in(2);
3213
3214 // Add phi if "def" node is in peel set and "use" is not
3215
3216 for(i = 0; i < peel_list.size(); i++ ) {
3217 Node *def = peel_list.at(i);
3218 if (!def->is_CFG()) {
3219 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
3220 Node *use = def->fast_out(j);
3221 if (has_node(use) && use->in(0) != C->top() &&
3222 (!peel.test(use->_idx) ||
3223 (use->is_Phi() && use->in(0) == head)) ) {
3224 worklist.push(use);
3225 }
3226 }
3227 while( worklist.size() ) {
3228 Node *use = worklist.pop();
3229 for (uint j = 1; j < use->req(); j++) {
3230 Node* n = use->in(j);
3231 if (n == def) {
3232
3233 // "def" is in peel set, "use" is not in peel set
3234 // or "use" is in the entry boundary (a phi) of the peel set
3235
3236 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
3237
3238 if ( loop->is_member(get_loop( use_c )) ) {
3239 // use is in loop
3240 if (old_new[use->_idx] != NULL) { // null for dead code
3241 Node* use_clone = old_new[use->_idx];
3242 _igvn.replace_input_of(use, j, C->top());
3243 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
3244 }
3245 } else {
3246 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
3247 // use is not in the loop, check if the live range includes the cut
3248 Node* lp_if = use_c->in(orig_exit_idx)->in(0);
3249 if (not_peel.test(lp_if->_idx)) {
3250 assert(j == orig_exit_idx, "use from original loop");
3251 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
3252 }
3253 }
3254 }
3255 }
3256 }
3257 }
3258 }
3259
3260 // Step 3b: retarget control
3261
3262 // Redirect control to the new loop head if a cloned node in
3263 // the not_peeled region has control that points into the peeled region.
3264 // This necessary because the cloned peeled region will be outside
3265 // the loop.
3266 // from to
3267 // cloned-peeled <---+
3268 // new_head_clone: | <--+
3269 // cloned-not_peeled in(0) in(0)
3270 // orig-peeled
3271
3272 for(i = 0; i < loop->_body.size(); i++ ) {
3273 Node *n = loop->_body.at(i);
3274 if (!n->is_CFG() && n->in(0) != NULL &&
3275 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
3276 Node* n_clone = old_new[n->_idx];
3277 _igvn.replace_input_of(n_clone, 0, new_head_clone);
3278 }
3279 }
3280
3281 // Backedge of the surviving new_head (the clone) is original last_peel
3282 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel);
3283
3284 // Cut first node in original not_peel set
3285 _igvn.rehash_node_delayed(new_head); // Multiple edge updates:
3286 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of
3287 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls
3288
3289 // Copy head_clone back-branch info to original head
3290 // and remove original head's loop entry and
3291 // clone head's back-branch
3292 _igvn.rehash_node_delayed(head); // Multiple edge updates
3293 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl));
3294 head->set_req(LoopNode::LoopBackControl, C->top());
3295 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top());
3296
3297 // Similarly modify the phis
3298 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
3299 Node* use = head->fast_out(k);
3300 if (use->is_Phi() && use->outcnt() > 0) {
3301 Node* use_clone = old_new[use->_idx];
3302 _igvn.rehash_node_delayed(use); // Multiple edge updates
3303 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl));
3304 use->set_req(LoopNode::LoopBackControl, C->top());
3305 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top());
3306 }
3307 }
3308
3309 // Step 4: update dominator tree and dominator depth
3310
3311 set_idom(head, orig_tail_clone, dd);
3312 recompute_dom_depth();
3313
3314 // Inhibit more partial peeling on this loop
3315 new_head_clone->set_partial_peel_loop();
3316 C->set_major_progress();
3317 loop->record_for_igvn();
3318
3319 #if !defined(PRODUCT)
3320 if (TracePartialPeeling) {
3321 tty->print_cr("\nafter partial peel one iteration");
3322 Node_List wl(area);
3323 Node* t = last_peel;
3324 while (true) {
3325 wl.push(t);
3326 if (t == head_clone) break;
3327 t = idom(t);
3328 }
3329 while (wl.size() > 0) {
3330 Node* tt = wl.pop();
3331 if (tt == head) tty->print_cr("orig head");
3332 else if (tt == new_head_clone) tty->print_cr("new head");
3333 else if (tt == head_clone) tty->print_cr("clone head");
3334 tt->dump();
3335 }
3336 }
3337 #endif
3338 return true;
3339 }
3340
3341 //------------------------------reorg_offsets----------------------------------
3342 // Reorganize offset computations to lower register pressure. Mostly
3343 // prevent loop-fallout uses of the pre-incremented trip counter (which are
3344 // then alive with the post-incremented trip counter forcing an extra
3345 // register move)
3346 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) {
3347 // Perform it only for canonical counted loops.
3348 // Loop's shape could be messed up by iteration_split_impl.
3349 if (!loop->_head->is_CountedLoop())
3350 return;
3351 if (!loop->_head->as_Loop()->is_valid_counted_loop())
3352 return;
3353
3354 CountedLoopNode *cl = loop->_head->as_CountedLoop();
3355 CountedLoopEndNode *cle = cl->loopexit();
3356 Node *exit = cle->proj_out(false);
3357 Node *phi = cl->phi();
3358
3359 // Check for the special case of folks using the pre-incremented
3360 // trip-counter on the fall-out path (forces the pre-incremented
3361 // and post-incremented trip counter to be live at the same time).
3362 // Fix this by adjusting to use the post-increment trip counter.
3363
3364 bool progress = true;
3365 while (progress) {
3366 progress = false;
3367 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
3368 Node* use = phi->fast_out(i); // User of trip-counter
3369 if (!has_ctrl(use)) continue;
3370 Node *u_ctrl = get_ctrl(use);
3371 if (use->is_Phi()) {
3372 u_ctrl = NULL;
3373 for (uint j = 1; j < use->req(); j++)
3374 if (use->in(j) == phi)
3375 u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j));
3376 }
3377 IdealLoopTree *u_loop = get_loop(u_ctrl);
3378 // Look for loop-invariant use
3379 if (u_loop == loop) continue;
3380 if (loop->is_member(u_loop)) continue;
3381 // Check that use is live out the bottom. Assuming the trip-counter
3382 // update is right at the bottom, uses of of the loop middle are ok.
3383 if (dom_lca(exit, u_ctrl) != exit) continue;
3384 // Hit! Refactor use to use the post-incremented tripcounter.
3385 // Compute a post-increment tripcounter.
3386 Node* c = exit;
3387 if (cl->is_strip_mined()) {
3388 IdealLoopTree* outer_loop = get_loop(cl->outer_loop());
3389 if (!outer_loop->is_member(u_loop)) {
3390 c = cl->outer_loop_exit();
3391 }
3392 }
3393 Node *opaq = new Opaque2Node(C, cle->incr());
3394 register_new_node(opaq, c);
3395 Node *neg_stride = _igvn.intcon(-cle->stride_con());
3396 set_ctrl(neg_stride, C->root());
3397 Node *post = new AddINode(opaq, neg_stride);
3398 register_new_node(post, c);
3399 _igvn.rehash_node_delayed(use);
3400 for (uint j = 1; j < use->req(); j++) {
3401 if (use->in(j) == phi)
3402 use->set_req(j, post);
3403 }
3404 // Since DU info changed, rerun loop
3405 progress = true;
3406 break;
3407 }
3408 }
3409
3410 }