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 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
891 Node* bs_res = bs->split_if_pre(this, n);
892 if (bs_res != NULL) {
893 return bs_res;
894 }
895 // Cloning these guys is unlikely to win
896 int n_op = n->Opcode();
897 if( n_op == Op_MergeMem ) return n;
898 if( n->is_Proj() ) return n;
899 // Do not clone-up CmpFXXX variations, as these are always
900 // followed by a CmpI
901 if( n->is_Cmp() ) return n;
902 // Attempt to use a conditional move instead of a phi/branch
903 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) {
904 Node *cmov = conditional_move( n );
905 if( cmov ) return cmov;
906 }
907 if( n->is_CFG() || n->is_LoadStore() )
908 return n;
909 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd
910 n_op == Op_Opaque2 ) {
911 if( !C->major_progress() ) // If chance of no more loop opts...
912 _igvn._worklist.push(n); // maybe we'll remove them
913 return n;
914 }
915
916 if( n->is_Con() ) return n; // No cloning for Con nodes
917
918 Node *n_ctrl = get_ctrl(n);
919 if( !n_ctrl ) return n; // Dead node
920
921 Node* res = try_move_store_before_loop(n, n_ctrl);
922 if (res != NULL) {
923 return n;
924 }
925
926 // Attempt to remix address expressions for loop invariants
927 Node *m = remix_address_expressions( n );
928 if( m ) return m;
929
930 if (n->is_ConstraintCast()) {
931 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this);
932 // ConstraintCastNode::dominating_cast() uses node control input to determine domination.
933 // Node control inputs don't necessarily agree with loop control info (due to
934 // transformations happened in between), thus additional dominance check is needed
935 // to keep loop info valid.
936 if (dom_cast != NULL && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) {
937 _igvn.replace_node(n, dom_cast);
938 return dom_cast;
939 }
940 }
941
942 // Determine if the Node has inputs from some local Phi.
943 // Returns the block to clone thru.
944 Node *n_blk = has_local_phi_input( n );
945 if( !n_blk ) return n;
946
947 // Do not clone the trip counter through on a CountedLoop
948 // (messes up the canonical shape).
949 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n;
950
951 // Check for having no control input; not pinned. Allow
952 // dominating control.
953 if (n->in(0)) {
954 Node *dom = idom(n_blk);
955 if (dom_lca(n->in(0), dom) != n->in(0)) {
956 return n;
957 }
958 }
959 // Policy: when is it profitable. You must get more wins than
960 // policy before it is considered profitable. Policy is usually 0,
961 // so 1 win is considered profitable. Big merges will require big
962 // cloning, so get a larger policy.
963 int policy = n_blk->req() >> 2;
964
965 // If the loop is a candidate for range check elimination,
966 // delay splitting through it's phi until a later loop optimization
967 if (n_blk->is_CountedLoop()) {
968 IdealLoopTree *lp = get_loop(n_blk);
969 if (lp && lp->_rce_candidate) {
970 return n;
971 }
972 }
973
974 // Use same limit as split_if_with_blocks_post
975 if( C->live_nodes() > 35000 ) return n; // Method too big
976
977 // Split 'n' through the merge point if it is profitable
978 Node *phi = split_thru_phi( n, n_blk, policy );
979 if (!phi) return n;
980
981 // Found a Phi to split thru!
982 // Replace 'n' with the new phi
983 _igvn.replace_node( n, phi );
984 // Moved a load around the loop, 'en-registering' something.
985 if (n_blk->is_Loop() && n->is_Load() &&
986 !phi->in(LoopNode::LoopBackControl)->is_Load())
987 C->set_major_progress();
988
989 return phi;
990 }
991
992 static bool merge_point_too_heavy(Compile* C, Node* region) {
993 // Bail out if the region and its phis have too many users.
994 int weight = 0;
995 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
996 weight += region->fast_out(i)->outcnt();
997 }
998 int nodes_left = C->max_node_limit() - C->live_nodes();
999 if (weight * 8 > nodes_left) {
1000 if (PrintOpto) {
1001 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight);
1002 }
1003 return true;
1004 } else {
1005 return false;
1006 }
1007 }
1008
1009 static bool merge_point_safe(Node* region) {
1010 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode
1011 // having a PhiNode input. This sidesteps the dangerous case where the split
1012 // ConvI2LNode may become TOP if the input Value() does not
1013 // overlap the ConvI2L range, leaving a node which may not dominate its
1014 // uses.
1015 // A better fix for this problem can be found in the BugTraq entry, but
1016 // expediency for Mantis demands this hack.
1017 // 6855164: If the merge point has a FastLockNode with a PhiNode input, we stop
1018 // split_if_with_blocks from splitting a block because we could not move around
1019 // the FastLockNode.
1020 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1021 Node* n = region->fast_out(i);
1022 if (n->is_Phi()) {
1023 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1024 Node* m = n->fast_out(j);
1025 if (m->is_FastLock())
1026 return false;
1027 #if INCLUDE_SHENANDOAHGC
1028 if (m->is_ShenandoahBarrier() && m->has_out_with(Op_FastLock)) {
1029 return false;
1030 }
1031 #endif
1032 #ifdef _LP64
1033 if (m->Opcode() == Op_ConvI2L)
1034 return false;
1035 if (m->is_CastII() && m->isa_CastII()->has_range_check()) {
1036 return false;
1037 }
1038 #endif
1039 }
1040 }
1041 }
1042 return true;
1043 }
1044
1045
1046 //------------------------------place_near_use---------------------------------
1047 // Place some computation next to use but not inside inner loops.
1048 // For inner loop uses move it to the preheader area.
1049 Node *PhaseIdealLoop::place_near_use(Node *useblock) const {
1050 IdealLoopTree *u_loop = get_loop( useblock );
1051 if (u_loop->_irreducible) {
1052 return useblock;
1053 }
1054 if (u_loop->_child) {
1055 if (useblock == u_loop->_head && u_loop->_head->is_OuterStripMinedLoop()) {
1056 return u_loop->_head->in(LoopNode::EntryControl);
1057 }
1058 return useblock;
1059 }
1060 return u_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl);
1061 }
1062
1063
1064 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) {
1065 if (!n->is_If() || n->is_CountedLoopEnd()) {
1066 return false;
1067 }
1068 if (!n->in(0)->is_Region()) {
1069 return false;
1070 }
1071 Node* region = n->in(0);
1072 Node* dom = idom(region);
1073 if (!dom->is_If() || dom->in(1) != n->in(1)) {
1074 return false;
1075 }
1076 IfNode* dom_if = dom->as_If();
1077 Node* proj_true = dom_if->proj_out(1);
1078 Node* proj_false = dom_if->proj_out(0);
1079
1080 for (uint i = 1; i < region->req(); i++) {
1081 if (is_dominator(proj_true, region->in(i))) {
1082 continue;
1083 }
1084 if (is_dominator(proj_false, region->in(i))) {
1085 continue;
1086 }
1087 return false;
1088 }
1089
1090 return true;
1091 }
1092
1093 bool PhaseIdealLoop::can_split_if(Node *n_ctrl) {
1094 if (C->live_nodes() > 35000) {
1095 return false; // Method too big
1096 }
1097
1098 // Do not do 'split-if' if irreducible loops are present.
1099 if (_has_irreducible_loops) {
1100 return false;
1101 }
1102
1103 if (merge_point_too_heavy(C, n_ctrl)) {
1104 return false;
1105 }
1106
1107 // Do not do 'split-if' if some paths are dead. First do dead code
1108 // elimination and then see if its still profitable.
1109 for (uint i = 1; i < n_ctrl->req(); i++) {
1110 if (n_ctrl->in(i) == C->top()) {
1111 return false;
1112 }
1113 }
1114
1115 // If trying to do a 'Split-If' at the loop head, it is only
1116 // profitable if the cmp folds up on BOTH paths. Otherwise we
1117 // risk peeling a loop forever.
1118
1119 // CNC - Disabled for now. Requires careful handling of loop
1120 // body selection for the cloned code. Also, make sure we check
1121 // for any input path not being in the same loop as n_ctrl. For
1122 // irreducible loops we cannot check for 'n_ctrl->is_Loop()'
1123 // because the alternative loop entry points won't be converted
1124 // into LoopNodes.
1125 IdealLoopTree *n_loop = get_loop(n_ctrl);
1126 for (uint j = 1; j < n_ctrl->req(); j++) {
1127 if (get_loop(n_ctrl->in(j)) != n_loop) {
1128 return false;
1129 }
1130 }
1131
1132 // Check for safety of the merge point.
1133 if (!merge_point_safe(n_ctrl)) {
1134 return false;
1135 }
1136
1137 return true;
1138 }
1139
1140 //------------------------------split_if_with_blocks_post----------------------
1141 // Do the real work in a non-recursive function. CFG hackery wants to be
1142 // in the post-order, so it can dirty the I-DOM info and not use the dirtied
1143 // info.
1144 void PhaseIdealLoop::split_if_with_blocks_post(Node *n, bool last_round) {
1145
1146 // Cloning Cmp through Phi's involves the split-if transform.
1147 // FastLock is not used by an If
1148 if (n->is_Cmp() && !n->is_FastLock() && !last_round) {
1149 Node *n_ctrl = get_ctrl(n);
1150 // Determine if the Node has inputs from some local Phi.
1151 // Returns the block to clone thru.
1152 Node *n_blk = has_local_phi_input(n);
1153 if (n_blk != n_ctrl) {
1154 return;
1155 }
1156
1157 if (!can_split_if(n_ctrl)) {
1158 return;
1159 }
1160
1161 if (n->outcnt() != 1) {
1162 return; // Multiple bool's from 1 compare?
1163 }
1164 Node *bol = n->unique_out();
1165 assert(bol->is_Bool(), "expect a bool here");
1166 if (bol->outcnt() != 1) {
1167 return;// Multiple branches from 1 compare?
1168 }
1169 Node *iff = bol->unique_out();
1170
1171 // Check some safety conditions
1172 if (iff->is_If()) { // Classic split-if?
1173 if (iff->in(0) != n_ctrl) {
1174 return; // Compare must be in same blk as if
1175 }
1176 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE
1177 // Can't split CMove with different control edge.
1178 if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) {
1179 return;
1180 }
1181 if (get_ctrl(iff->in(2)) == n_ctrl ||
1182 get_ctrl(iff->in(3)) == n_ctrl) {
1183 return; // Inputs not yet split-up
1184 }
1185 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) {
1186 return; // Loop-invar test gates loop-varying CMOVE
1187 }
1188 } else {
1189 return; // some other kind of node, such as an Allocate
1190 }
1191
1192 // When is split-if profitable? Every 'win' on means some control flow
1193 // goes dead, so it's almost always a win.
1194 int policy = 0;
1195 // Split compare 'n' through the merge point if it is profitable
1196 Node *phi = split_thru_phi( n, n_ctrl, policy);
1197 if (!phi) {
1198 return;
1199 }
1200
1201 // Found a Phi to split thru!
1202 // Replace 'n' with the new phi
1203 _igvn.replace_node(n, phi);
1204
1205 // Now split the bool up thru the phi
1206 Node *bolphi = split_thru_phi(bol, n_ctrl, -1);
1207 guarantee(bolphi != NULL, "null boolean phi node");
1208
1209 _igvn.replace_node(bol, bolphi);
1210 assert(iff->in(1) == bolphi, "");
1211
1212 if (bolphi->Value(&_igvn)->singleton()) {
1213 return;
1214 }
1215
1216 // Conditional-move? Must split up now
1217 if (!iff->is_If()) {
1218 Node *cmovphi = split_thru_phi(iff, n_ctrl, -1);
1219 _igvn.replace_node(iff, cmovphi);
1220 return;
1221 }
1222
1223 // Now split the IF
1224 do_split_if(iff);
1225 return;
1226 }
1227
1228 // Two identical ifs back to back can be merged
1229 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) {
1230 Node *n_ctrl = n->in(0);
1231 PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1));
1232 IfNode* dom_if = idom(n_ctrl)->as_If();
1233 Node* proj_true = dom_if->proj_out(1);
1234 Node* proj_false = dom_if->proj_out(0);
1235 Node* con_true = _igvn.makecon(TypeInt::ONE);
1236 Node* con_false = _igvn.makecon(TypeInt::ZERO);
1237
1238 for (uint i = 1; i < n_ctrl->req(); i++) {
1239 if (is_dominator(proj_true, n_ctrl->in(i))) {
1240 bolphi->init_req(i, con_true);
1241 } else {
1242 assert(is_dominator(proj_false, n_ctrl->in(i)), "bad if");
1243 bolphi->init_req(i, con_false);
1244 }
1245 }
1246 register_new_node(bolphi, n_ctrl);
1247 _igvn.replace_input_of(n, 1, bolphi);
1248
1249 // Now split the IF
1250 do_split_if(n);
1251 return;
1252 }
1253
1254 // Check for an IF ready to split; one that has its
1255 // condition codes input coming from a Phi at the block start.
1256 int n_op = n->Opcode();
1257
1258 // Check for an IF being dominated by another IF same test
1259 if (n_op == Op_If ||
1260 n_op == Op_RangeCheck) {
1261 Node *bol = n->in(1);
1262 uint max = bol->outcnt();
1263 // Check for same test used more than once?
1264 if (max > 1 && bol->is_Bool()) {
1265 // Search up IDOMs to see if this IF is dominated.
1266 Node *cutoff = get_ctrl(bol);
1267
1268 // Now search up IDOMs till cutoff, looking for a dominating test
1269 Node *prevdom = n;
1270 Node *dom = idom(prevdom);
1271 while (dom != cutoff) {
1272 if (dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom) {
1273 // Replace the dominated test with an obvious true or false.
1274 // Place it on the IGVN worklist for later cleanup.
1275 C->set_major_progress();
1276 dominated_by(prevdom, n, false, true);
1277 #ifndef PRODUCT
1278 if( VerifyLoopOptimizations ) verify();
1279 #endif
1280 return;
1281 }
1282 prevdom = dom;
1283 dom = idom(prevdom);
1284 }
1285 }
1286 }
1287
1288 // See if a shared loop-varying computation has no loop-varying uses.
1289 // Happens if something is only used for JVM state in uncommon trap exits,
1290 // like various versions of induction variable+offset. Clone the
1291 // computation per usage to allow it to sink out of the loop.
1292 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about)
1293 Node *n_ctrl = get_ctrl(n);
1294 IdealLoopTree *n_loop = get_loop(n_ctrl);
1295 if( n_loop != _ltree_root ) {
1296 DUIterator_Fast imax, i = n->fast_outs(imax);
1297 for (; i < imax; i++) {
1298 Node* u = n->fast_out(i);
1299 if( !has_ctrl(u) ) break; // Found control user
1300 IdealLoopTree *u_loop = get_loop(get_ctrl(u));
1301 if( u_loop == n_loop ) break; // Found loop-varying use
1302 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop
1303 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003
1304 }
1305 bool did_break = (i < imax); // Did we break out of the previous loop?
1306 if (!did_break && n->outcnt() > 1) { // All uses in outer loops!
1307 Node *late_load_ctrl = NULL;
1308 if (n->is_Load()) {
1309 // If n is a load, get and save the result from get_late_ctrl(),
1310 // to be later used in calculating the control for n's clones.
1311 clear_dom_lca_tags();
1312 late_load_ctrl = get_late_ctrl(n, n_ctrl);
1313 }
1314 // If n is a load, and the late control is the same as the current
1315 // control, then the cloning of n is a pointless exercise, because
1316 // GVN will ensure that we end up where we started.
1317 if (!n->is_Load() || late_load_ctrl != n_ctrl) {
1318 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1319 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) {
1320 Node *u = n->last_out(j); // Clone private computation per use
1321 _igvn.rehash_node_delayed(u);
1322 Node *x = n->clone(); // Clone computation
1323 Node *x_ctrl = NULL;
1324 if( u->is_Phi() ) {
1325 // Replace all uses of normal nodes. Replace Phi uses
1326 // individually, so the separate Nodes can sink down
1327 // different paths.
1328 uint k = 1;
1329 while( u->in(k) != n ) k++;
1330 u->set_req( k, x );
1331 // x goes next to Phi input path
1332 x_ctrl = u->in(0)->in(k);
1333 --j;
1334 } else { // Normal use
1335 // Replace all uses
1336 for( uint k = 0; k < u->req(); k++ ) {
1337 if( u->in(k) == n ) {
1338 u->set_req( k, x );
1339 --j;
1340 }
1341 }
1342 x_ctrl = get_ctrl(u);
1343 }
1344
1345 // Find control for 'x' next to use but not inside inner loops.
1346 // For inner loop uses get the preheader area.
1347 x_ctrl = place_near_use(x_ctrl);
1348
1349 if (bs->sink_node(this, n, x, x_ctrl, n_ctrl)) {
1350 continue;
1351 }
1352
1353 if (n->is_Load()) {
1354 // For loads, add a control edge to a CFG node outside of the loop
1355 // to force them to not combine and return back inside the loop
1356 // during GVN optimization (4641526).
1357 //
1358 // Because we are setting the actual control input, factor in
1359 // the result from get_late_ctrl() so we respect any
1360 // anti-dependences. (6233005).
1361 x_ctrl = dom_lca(late_load_ctrl, x_ctrl);
1362
1363 // Don't allow the control input to be a CFG splitting node.
1364 // Such nodes should only have ProjNodes as outs, e.g. IfNode
1365 // should only have IfTrueNode and IfFalseNode (4985384).
1366 x_ctrl = find_non_split_ctrl(x_ctrl);
1367 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone");
1368
1369 x->set_req(0, x_ctrl);
1370 }
1371 register_new_node(x, x_ctrl);
1372
1373 // Some institutional knowledge is needed here: 'x' is
1374 // yanked because if the optimizer runs GVN on it all the
1375 // cloned x's will common up and undo this optimization and
1376 // be forced back in the loop. This is annoying because it
1377 // makes +VerifyOpto report false-positives on progress. I
1378 // tried setting control edges on the x's to force them to
1379 // not combine, but the matching gets worried when it tries
1380 // to fold a StoreP and an AddP together (as part of an
1381 // address expression) and the AddP and StoreP have
1382 // different controls.
1383 if (!x->is_Load() && !x->is_DecodeNarrowPtr()) _igvn._worklist.yank(x);
1384 }
1385 _igvn.remove_dead_node(n);
1386 }
1387 }
1388 }
1389 }
1390
1391 try_move_store_after_loop(n);
1392
1393 // Check for Opaque2's who's loop has disappeared - who's input is in the
1394 // same loop nest as their output. Remove 'em, they are no longer useful.
1395 if( n_op == Op_Opaque2 &&
1396 n->in(1) != NULL &&
1397 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) {
1398 _igvn.replace_node( n, n->in(1) );
1399 }
1400
1401 #if INCLUDE_ZGC
1402 if (UseZGC) {
1403 ZBarrierSetC2::loop_optimize_gc_barrier(this, n, last_round);
1404 }
1405 #endif
1406 }
1407
1408 //------------------------------split_if_with_blocks---------------------------
1409 // Check for aggressive application of 'split-if' optimization,
1410 // using basic block level info.
1411 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack, bool last_round) {
1412 Node *n = C->root();
1413 visited.set(n->_idx); // first, mark node as visited
1414 // Do pre-visit work for root
1415 n = split_if_with_blocks_pre( n );
1416 uint cnt = n->outcnt();
1417 uint i = 0;
1418 while (true) {
1419 // Visit all children
1420 if (i < cnt) {
1421 Node* use = n->raw_out(i);
1422 ++i;
1423 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
1424 // Now do pre-visit work for this use
1425 use = split_if_with_blocks_pre( use );
1426 nstack.push(n, i); // Save parent and next use's index.
1427 n = use; // Process all children of current use.
1428 cnt = use->outcnt();
1429 i = 0;
1430 }
1431 }
1432 else {
1433 // All of n's children have been processed, complete post-processing.
1434 if (cnt != 0 && !n->is_Con()) {
1435 assert(has_node(n), "no dead nodes");
1436 split_if_with_blocks_post( n, last_round );
1437 }
1438 if (nstack.is_empty()) {
1439 // Finished all nodes on stack.
1440 break;
1441 }
1442 // Get saved parent node and next use's index. Visit the rest of uses.
1443 n = nstack.node();
1444 cnt = n->outcnt();
1445 i = nstack.index();
1446 nstack.pop();
1447 }
1448 }
1449 }
1450
1451
1452 //=============================================================================
1453 //
1454 // C L O N E A L O O P B O D Y
1455 //
1456
1457 //------------------------------clone_iff--------------------------------------
1458 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1459 // "Nearly" because all Nodes have been cloned from the original in the loop,
1460 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1461 // through the Phi recursively, and return a Bool.
1462 Node* PhaseIdealLoop::clone_iff(PhiNode *phi, IdealLoopTree *loop) {
1463
1464 // Convert this Phi into a Phi merging Bools
1465 uint i;
1466 for (i = 1; i < phi->req(); i++) {
1467 Node *b = phi->in(i);
1468 if (b->is_Phi()) {
1469 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi(), loop));
1470 } else {
1471 assert(b->is_Bool() || b->Opcode() == Op_Opaque4, "");
1472 }
1473 }
1474
1475 Node* n = phi->in(1);
1476 Node* sample_opaque = NULL;
1477 Node *sample_bool = NULL;
1478 if (n->Opcode() == Op_Opaque4) {
1479 sample_opaque = n;
1480 sample_bool = n->in(1);
1481 assert(sample_bool->is_Bool(), "wrong type");
1482 } else {
1483 sample_bool = n;
1484 }
1485 Node *sample_cmp = sample_bool->in(1);
1486
1487 // Make Phis to merge the Cmp's inputs.
1488 PhiNode *phi1 = new PhiNode(phi->in(0), Type::TOP);
1489 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP);
1490 for (i = 1; i < phi->req(); i++) {
1491 Node *n1 = sample_opaque == NULL ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1);
1492 Node *n2 = sample_opaque == NULL ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2);
1493 phi1->set_req(i, n1);
1494 phi2->set_req(i, n2);
1495 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
1496 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
1497 }
1498 // See if these Phis have been made before.
1499 // Register with optimizer
1500 Node *hit1 = _igvn.hash_find_insert(phi1);
1501 if (hit1) { // Hit, toss just made Phi
1502 _igvn.remove_dead_node(phi1); // Remove new phi
1503 assert(hit1->is_Phi(), "" );
1504 phi1 = (PhiNode*)hit1; // Use existing phi
1505 } else { // Miss
1506 _igvn.register_new_node_with_optimizer(phi1);
1507 }
1508 Node *hit2 = _igvn.hash_find_insert(phi2);
1509 if (hit2) { // Hit, toss just made Phi
1510 _igvn.remove_dead_node(phi2); // Remove new phi
1511 assert(hit2->is_Phi(), "" );
1512 phi2 = (PhiNode*)hit2; // Use existing phi
1513 } else { // Miss
1514 _igvn.register_new_node_with_optimizer(phi2);
1515 }
1516 // Register Phis with loop/block info
1517 set_ctrl(phi1, phi->in(0));
1518 set_ctrl(phi2, phi->in(0));
1519 // Make a new Cmp
1520 Node *cmp = sample_cmp->clone();
1521 cmp->set_req(1, phi1);
1522 cmp->set_req(2, phi2);
1523 _igvn.register_new_node_with_optimizer(cmp);
1524 set_ctrl(cmp, phi->in(0));
1525
1526 // Make a new Bool
1527 Node *b = sample_bool->clone();
1528 b->set_req(1,cmp);
1529 _igvn.register_new_node_with_optimizer(b);
1530 set_ctrl(b, phi->in(0));
1531
1532 if (sample_opaque != NULL) {
1533 Node* opaque = sample_opaque->clone();
1534 opaque->set_req(1, b);
1535 _igvn.register_new_node_with_optimizer(opaque);
1536 set_ctrl(opaque, phi->in(0));
1537 return opaque;
1538 }
1539
1540 assert(b->is_Bool(), "");
1541 return b;
1542 }
1543
1544 //------------------------------clone_bool-------------------------------------
1545 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1546 // "Nearly" because all Nodes have been cloned from the original in the loop,
1547 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1548 // through the Phi recursively, and return a Bool.
1549 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) {
1550 uint i;
1551 // Convert this Phi into a Phi merging Bools
1552 for( i = 1; i < phi->req(); i++ ) {
1553 Node *b = phi->in(i);
1554 if( b->is_Phi() ) {
1555 _igvn.replace_input_of(phi, i, clone_bool( b->as_Phi(), loop ));
1556 } else {
1557 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
1558 }
1559 }
1560
1561 Node *sample_cmp = phi->in(1);
1562
1563 // Make Phis to merge the Cmp's inputs.
1564 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP );
1565 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP );
1566 for( uint j = 1; j < phi->req(); j++ ) {
1567 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
1568 Node *n1, *n2;
1569 if( cmp_top->is_Cmp() ) {
1570 n1 = cmp_top->in(1);
1571 n2 = cmp_top->in(2);
1572 } else {
1573 n1 = n2 = cmp_top;
1574 }
1575 phi1->set_req( j, n1 );
1576 phi2->set_req( j, n2 );
1577 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
1578 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
1579 }
1580
1581 // See if these Phis have been made before.
1582 // Register with optimizer
1583 Node *hit1 = _igvn.hash_find_insert(phi1);
1584 if( hit1 ) { // Hit, toss just made Phi
1585 _igvn.remove_dead_node(phi1); // Remove new phi
1586 assert( hit1->is_Phi(), "" );
1587 phi1 = (PhiNode*)hit1; // Use existing phi
1588 } else { // Miss
1589 _igvn.register_new_node_with_optimizer(phi1);
1590 }
1591 Node *hit2 = _igvn.hash_find_insert(phi2);
1592 if( hit2 ) { // Hit, toss just made Phi
1593 _igvn.remove_dead_node(phi2); // Remove new phi
1594 assert( hit2->is_Phi(), "" );
1595 phi2 = (PhiNode*)hit2; // Use existing phi
1596 } else { // Miss
1597 _igvn.register_new_node_with_optimizer(phi2);
1598 }
1599 // Register Phis with loop/block info
1600 set_ctrl(phi1, phi->in(0));
1601 set_ctrl(phi2, phi->in(0));
1602 // Make a new Cmp
1603 Node *cmp = sample_cmp->clone();
1604 cmp->set_req( 1, phi1 );
1605 cmp->set_req( 2, phi2 );
1606 _igvn.register_new_node_with_optimizer(cmp);
1607 set_ctrl(cmp, phi->in(0));
1608
1609 assert( cmp->is_Cmp(), "" );
1610 return (CmpNode*)cmp;
1611 }
1612
1613 //------------------------------sink_use---------------------------------------
1614 // If 'use' was in the loop-exit block, it now needs to be sunk
1615 // below the post-loop merge point.
1616 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) {
1617 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) {
1618 set_ctrl(use, post_loop);
1619 for (DUIterator j = use->outs(); use->has_out(j); j++)
1620 sink_use(use->out(j), post_loop);
1621 }
1622 }
1623
1624 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new,
1625 IdealLoopTree* loop, IdealLoopTree* outer_loop,
1626 Node_List*& split_if_set, Node_List*& split_bool_set,
1627 Node_List*& split_cex_set, Node_List& worklist,
1628 uint new_counter, CloneLoopMode mode) {
1629 Node* nnn = old_new[old->_idx];
1630 // Copy uses to a worklist, so I can munge the def-use info
1631 // with impunity.
1632 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
1633 worklist.push(old->fast_out(j));
1634
1635 while( worklist.size() ) {
1636 Node *use = worklist.pop();
1637 if (!has_node(use)) continue; // Ignore dead nodes
1638 if (use->in(0) == C->top()) continue;
1639 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
1640 // Check for data-use outside of loop - at least one of OLD or USE
1641 // must not be a CFG node.
1642 #ifdef ASSERT
1643 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == NULL) {
1644 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint();
1645 assert(mode == ControlAroundStripMined && use == sfpt, "missed a node");
1646 }
1647 #endif
1648 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) {
1649
1650 // If the Data use is an IF, that means we have an IF outside of the
1651 // loop that is switching on a condition that is set inside of the
1652 // loop. Happens if people set a loop-exit flag; then test the flag
1653 // in the loop to break the loop, then test is again outside of the
1654 // loop to determine which way the loop exited.
1655 // Loop predicate If node connects to Bool node through Opaque1 node.
1656 if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use) || use->Opcode() == Op_Opaque4) {
1657 // Since this code is highly unlikely, we lazily build the worklist
1658 // of such Nodes to go split.
1659 if (!split_if_set) {
1660 ResourceArea *area = Thread::current()->resource_area();
1661 split_if_set = new Node_List(area);
1662 }
1663 split_if_set->push(use);
1664 }
1665 if (use->is_Bool()) {
1666 if (!split_bool_set) {
1667 ResourceArea *area = Thread::current()->resource_area();
1668 split_bool_set = new Node_List(area);
1669 }
1670 split_bool_set->push(use);
1671 }
1672 if (use->Opcode() == Op_CreateEx) {
1673 if (!split_cex_set) {
1674 ResourceArea *area = Thread::current()->resource_area();
1675 split_cex_set = new Node_List(area);
1676 }
1677 split_cex_set->push(use);
1678 }
1679
1680
1681 // Get "block" use is in
1682 uint idx = 0;
1683 while( use->in(idx) != old ) idx++;
1684 Node *prev = use->is_CFG() ? use : get_ctrl(use);
1685 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" );
1686 Node *cfg = prev->_idx >= new_counter
1687 ? prev->in(2)
1688 : idom(prev);
1689 if( use->is_Phi() ) // Phi use is in prior block
1690 cfg = prev->in(idx); // NOT in block of Phi itself
1691 if (cfg->is_top()) { // Use is dead?
1692 _igvn.replace_input_of(use, idx, C->top());
1693 continue;
1694 }
1695
1696 // If use is referenced through control edge... (idx == 0)
1697 if (mode == IgnoreStripMined && idx == 0) {
1698 LoopNode *head = loop->_head->as_Loop();
1699 if (head->is_strip_mined() && is_dominator(head->outer_loop_exit(), prev)) {
1700 // That node is outside the inner loop, leave it outside the
1701 // outer loop as well to not confuse verification code.
1702 assert(!loop->_parent->is_member(use_loop), "should be out of the outer loop");
1703 _igvn.replace_input_of(use, 0, head->outer_loop_exit());
1704 continue;
1705 }
1706 }
1707
1708 while(!outer_loop->is_member(get_loop(cfg))) {
1709 prev = cfg;
1710 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg);
1711 }
1712 // If the use occurs after merging several exits from the loop, then
1713 // old value must have dominated all those exits. Since the same old
1714 // value was used on all those exits we did not need a Phi at this
1715 // merge point. NOW we do need a Phi here. Each loop exit value
1716 // is now merged with the peeled body exit; each exit gets its own
1717 // private Phi and those Phis need to be merged here.
1718 Node *phi;
1719 if( prev->is_Region() ) {
1720 if( idx == 0 ) { // Updating control edge?
1721 phi = prev; // Just use existing control
1722 } else { // Else need a new Phi
1723 phi = PhiNode::make( prev, old );
1724 // Now recursively fix up the new uses of old!
1725 for( uint i = 1; i < prev->req(); i++ ) {
1726 worklist.push(phi); // Onto worklist once for each 'old' input
1727 }
1728 }
1729 } else {
1730 // Get new RegionNode merging old and new loop exits
1731 prev = old_new[prev->_idx];
1732 assert( prev, "just made this in step 7" );
1733 if( idx == 0) { // Updating control edge?
1734 phi = prev; // Just use existing control
1735 } else { // Else need a new Phi
1736 // Make a new Phi merging data values properly
1737 phi = PhiNode::make( prev, old );
1738 phi->set_req( 1, nnn );
1739 }
1740 }
1741 // If inserting a new Phi, check for prior hits
1742 if( idx != 0 ) {
1743 Node *hit = _igvn.hash_find_insert(phi);
1744 if( hit == NULL ) {
1745 _igvn.register_new_node_with_optimizer(phi); // Register new phi
1746 } else { // or
1747 // Remove the new phi from the graph and use the hit
1748 _igvn.remove_dead_node(phi);
1749 phi = hit; // Use existing phi
1750 }
1751 set_ctrl(phi, prev);
1752 }
1753 // Make 'use' use the Phi instead of the old loop body exit value
1754 _igvn.replace_input_of(use, idx, phi);
1755 if( use->_idx >= new_counter ) { // If updating new phis
1756 // Not needed for correctness, but prevents a weak assert
1757 // in AddPNode from tripping (when we end up with different
1758 // base & derived Phis that will become the same after
1759 // IGVN does CSE).
1760 Node *hit = _igvn.hash_find_insert(use);
1761 if( hit ) // Go ahead and re-hash for hits.
1762 _igvn.replace_node( use, hit );
1763 }
1764
1765 // If 'use' was in the loop-exit block, it now needs to be sunk
1766 // below the post-loop merge point.
1767 sink_use( use, prev );
1768 }
1769 }
1770 }
1771
1772 static void clone_outer_loop_helper(Node* n, const IdealLoopTree *loop, const IdealLoopTree* outer_loop,
1773 const Node_List &old_new, Unique_Node_List& wq, PhaseIdealLoop* phase,
1774 bool check_old_new) {
1775 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1776 Node* u = n->fast_out(j);
1777 assert(check_old_new || old_new[u->_idx] == NULL, "shouldn't have been cloned");
1778 if (!u->is_CFG() && (!check_old_new || old_new[u->_idx] == NULL)) {
1779 Node* c = phase->get_ctrl(u);
1780 IdealLoopTree* u_loop = phase->get_loop(c);
1781 assert(!loop->is_member(u_loop), "can be in outer loop or out of both loops only");
1782 if (outer_loop->is_member(u_loop)) {
1783 wq.push(u);
1784 }
1785 }
1786 }
1787 }
1788
1789 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
1790 IdealLoopTree* outer_loop, int dd, Node_List &old_new,
1791 Node_List& extra_data_nodes) {
1792 if (head->is_strip_mined() && mode != IgnoreStripMined) {
1793 CountedLoopNode* cl = head->as_CountedLoop();
1794 Node* l = cl->outer_loop();
1795 Node* tail = cl->outer_loop_tail();
1796 IfNode* le = cl->outer_loop_end();
1797 Node* sfpt = cl->outer_safepoint();
1798 CountedLoopEndNode* cle = cl->loopexit();
1799 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop();
1800 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null();
1801 Node* cle_out = cle->proj_out(false);
1802
1803 Node* new_sfpt = NULL;
1804 Node* new_cle_out = cle_out->clone();
1805 old_new.map(cle_out->_idx, new_cle_out);
1806 if (mode == CloneIncludesStripMined) {
1807 // clone outer loop body
1808 Node* new_l = l->clone();
1809 Node* new_tail = tail->clone();
1810 IfNode* new_le = le->clone()->as_If();
1811 new_sfpt = sfpt->clone();
1812
1813 set_loop(new_l, outer_loop->_parent);
1814 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd);
1815 set_loop(new_cle_out, outer_loop->_parent);
1816 set_idom(new_cle_out, new_cle, dd);
1817 set_loop(new_sfpt, outer_loop->_parent);
1818 set_idom(new_sfpt, new_cle_out, dd);
1819 set_loop(new_le, outer_loop->_parent);
1820 set_idom(new_le, new_sfpt, dd);
1821 set_loop(new_tail, outer_loop->_parent);
1822 set_idom(new_tail, new_le, dd);
1823 set_idom(new_cl, new_l, dd);
1824
1825 old_new.map(l->_idx, new_l);
1826 old_new.map(tail->_idx, new_tail);
1827 old_new.map(le->_idx, new_le);
1828 old_new.map(sfpt->_idx, new_sfpt);
1829
1830 new_l->set_req(LoopNode::LoopBackControl, new_tail);
1831 new_l->set_req(0, new_l);
1832 new_tail->set_req(0, new_le);
1833 new_le->set_req(0, new_sfpt);
1834 new_sfpt->set_req(0, new_cle_out);
1835 new_cle_out->set_req(0, new_cle);
1836 new_cl->set_req(LoopNode::EntryControl, new_l);
1837
1838 _igvn.register_new_node_with_optimizer(new_l);
1839 _igvn.register_new_node_with_optimizer(new_tail);
1840 _igvn.register_new_node_with_optimizer(new_le);
1841 } else {
1842 Node *newhead = old_new[loop->_head->_idx];
1843 newhead->as_Loop()->clear_strip_mined();
1844 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl));
1845 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
1846 }
1847 // Look at data node that were assigned a control in the outer
1848 // loop: they are kept in the outer loop by the safepoint so start
1849 // from the safepoint node's inputs.
1850 IdealLoopTree* outer_loop = get_loop(l);
1851 Node_Stack stack(2);
1852 stack.push(sfpt, 1);
1853 uint new_counter = C->unique();
1854 while (stack.size() > 0) {
1855 Node* n = stack.node();
1856 uint i = stack.index();
1857 while (i < n->req() &&
1858 (n->in(i) == NULL ||
1859 !has_ctrl(n->in(i)) ||
1860 get_loop(get_ctrl(n->in(i))) != outer_loop ||
1861 (old_new[n->in(i)->_idx] != NULL && old_new[n->in(i)->_idx]->_idx >= new_counter))) {
1862 i++;
1863 }
1864 if (i < n->req()) {
1865 stack.set_index(i+1);
1866 stack.push(n->in(i), 0);
1867 } else {
1868 assert(old_new[n->_idx] == NULL || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet");
1869 Node* m = n == sfpt ? new_sfpt : n->clone();
1870 if (m != NULL) {
1871 for (uint i = 0; i < n->req(); i++) {
1872 if (m->in(i) != NULL && old_new[m->in(i)->_idx] != NULL) {
1873 m->set_req(i, old_new[m->in(i)->_idx]);
1874 }
1875 }
1876 } else {
1877 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?");
1878 }
1879 if (n != sfpt) {
1880 extra_data_nodes.push(n);
1881 _igvn.register_new_node_with_optimizer(m);
1882 assert(get_ctrl(n) == cle_out, "what other control?");
1883 set_ctrl(m, new_cle_out);
1884 old_new.map(n->_idx, m);
1885 }
1886 stack.pop();
1887 }
1888 }
1889 if (mode == CloneIncludesStripMined) {
1890 _igvn.register_new_node_with_optimizer(new_sfpt);
1891 _igvn.register_new_node_with_optimizer(new_cle_out);
1892 }
1893 // Some other transformation may have pessimistically assign some
1894 // data nodes to the outer loop. Set their control so they are out
1895 // of the outer loop.
1896 ResourceMark rm;
1897 Unique_Node_List wq;
1898 for (uint i = 0; i < extra_data_nodes.size(); i++) {
1899 Node* old = extra_data_nodes.at(i);
1900 clone_outer_loop_helper(old, loop, outer_loop, old_new, wq, this, true);
1901 }
1902 Node* new_ctrl = cl->outer_loop_exit();
1903 assert(get_loop(new_ctrl) != outer_loop, "must be out of the loop nest");
1904 for (uint i = 0; i < wq.size(); i++) {
1905 Node* n = wq.at(i);
1906 set_ctrl(n, new_ctrl);
1907 clone_outer_loop_helper(n, loop, outer_loop, old_new, wq, this, false);
1908 }
1909 } else {
1910 Node *newhead = old_new[loop->_head->_idx];
1911 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
1912 }
1913 }
1914
1915 //------------------------------clone_loop-------------------------------------
1916 //
1917 // C L O N E A L O O P B O D Y
1918 //
1919 // This is the basic building block of the loop optimizations. It clones an
1920 // entire loop body. It makes an old_new loop body mapping; with this mapping
1921 // you can find the new-loop equivalent to an old-loop node. All new-loop
1922 // nodes are exactly equal to their old-loop counterparts, all edges are the
1923 // same. All exits from the old-loop now have a RegionNode that merges the
1924 // equivalent new-loop path. This is true even for the normal "loop-exit"
1925 // condition. All uses of loop-invariant old-loop values now come from (one
1926 // or more) Phis that merge their new-loop equivalents.
1927 //
1928 // This operation leaves the graph in an illegal state: there are two valid
1929 // control edges coming from the loop pre-header to both loop bodies. I'll
1930 // definitely have to hack the graph after running this transform.
1931 //
1932 // From this building block I will further edit edges to perform loop peeling
1933 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
1934 //
1935 // Parameter side_by_size_idom:
1936 // When side_by_size_idom is NULL, the dominator tree is constructed for
1937 // the clone loop to dominate the original. Used in construction of
1938 // pre-main-post loop sequence.
1939 // When nonnull, the clone and original are side-by-side, both are
1940 // dominated by the side_by_side_idom node. Used in construction of
1941 // unswitched loops.
1942 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
1943 CloneLoopMode mode, Node* side_by_side_idom) {
1944
1945 LoopNode* head = loop->_head->as_Loop();
1946 head->verify_strip_mined(1);
1947
1948 if (C->do_vector_loop() && PrintOpto) {
1949 const char* mname = C->method()->name()->as_quoted_ascii();
1950 if (mname != NULL) {
1951 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname);
1952 }
1953 }
1954
1955 CloneMap& cm = C->clone_map();
1956 Dict* dict = cm.dict();
1957 if (C->do_vector_loop()) {
1958 cm.set_clone_idx(cm.max_gen()+1);
1959 #ifndef PRODUCT
1960 if (PrintOpto) {
1961 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx());
1962 loop->dump_head();
1963 }
1964 #endif
1965 }
1966
1967 // Step 1: Clone the loop body. Make the old->new mapping.
1968 uint i;
1969 for( i = 0; i < loop->_body.size(); i++ ) {
1970 Node *old = loop->_body.at(i);
1971 Node *nnn = old->clone();
1972 old_new.map( old->_idx, nnn );
1973 if (C->do_vector_loop()) {
1974 cm.verify_insert_and_clone(old, nnn, cm.clone_idx());
1975 }
1976 _igvn.register_new_node_with_optimizer(nnn);
1977 }
1978
1979 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop;
1980
1981 // Step 2: Fix the edges in the new body. If the old input is outside the
1982 // loop use it. If the old input is INside the loop, use the corresponding
1983 // new node instead.
1984 for( i = 0; i < loop->_body.size(); i++ ) {
1985 Node *old = loop->_body.at(i);
1986 Node *nnn = old_new[old->_idx];
1987 // Fix CFG/Loop controlling the new node
1988 if (has_ctrl(old)) {
1989 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
1990 } else {
1991 set_loop(nnn, outer_loop->_parent);
1992 if (old->outcnt() > 0) {
1993 set_idom( nnn, old_new[idom(old)->_idx], dd );
1994 }
1995 }
1996 // Correct edges to the new node
1997 for( uint j = 0; j < nnn->req(); j++ ) {
1998 Node *n = nnn->in(j);
1999 if( n ) {
2000 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n );
2001 if( loop->is_member( old_in_loop ) )
2002 nnn->set_req(j, old_new[n->_idx]);
2003 }
2004 }
2005 _igvn.hash_find_insert(nnn);
2006 }
2007
2008 ResourceArea *area = Thread::current()->resource_area();
2009 Node_List extra_data_nodes(area); // data nodes in the outer strip mined loop
2010 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes);
2011
2012 // Step 3: Now fix control uses. Loop varying control uses have already
2013 // been fixed up (as part of all input edges in Step 2). Loop invariant
2014 // control uses must be either an IfFalse or an IfTrue. Make a merge
2015 // point to merge the old and new IfFalse/IfTrue nodes; make the use
2016 // refer to this.
2017 Node_List worklist(area);
2018 uint new_counter = C->unique();
2019 for( i = 0; i < loop->_body.size(); i++ ) {
2020 Node* old = loop->_body.at(i);
2021 if( !old->is_CFG() ) continue;
2022
2023 // Copy uses to a worklist, so I can munge the def-use info
2024 // with impunity.
2025 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
2026 worklist.push(old->fast_out(j));
2027
2028 while( worklist.size() ) { // Visit all uses
2029 Node *use = worklist.pop();
2030 if (!has_node(use)) continue; // Ignore dead nodes
2031 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
2032 if( !loop->is_member( use_loop ) && use->is_CFG() ) {
2033 // Both OLD and USE are CFG nodes here.
2034 assert( use->is_Proj(), "" );
2035 Node* nnn = old_new[old->_idx];
2036
2037 Node* newuse = NULL;
2038 if (head->is_strip_mined() && mode != IgnoreStripMined) {
2039 CountedLoopNode* cl = head->as_CountedLoop();
2040 CountedLoopEndNode* cle = cl->loopexit();
2041 Node* cle_out = cle->proj_out_or_null(false);
2042 if (use == cle_out) {
2043 IfNode* le = cl->outer_loop_end();
2044 use = le->proj_out(false);
2045 use_loop = get_loop(use);
2046 if (mode == CloneIncludesStripMined) {
2047 nnn = old_new[le->_idx];
2048 } else {
2049 newuse = old_new[cle_out->_idx];
2050 }
2051 }
2052 }
2053 if (newuse == NULL) {
2054 newuse = use->clone();
2055 }
2056
2057 // Clone the loop exit control projection
2058 if (C->do_vector_loop()) {
2059 cm.verify_insert_and_clone(use, newuse, cm.clone_idx());
2060 }
2061 newuse->set_req(0,nnn);
2062 _igvn.register_new_node_with_optimizer(newuse);
2063 set_loop(newuse, use_loop);
2064 set_idom(newuse, nnn, dom_depth(nnn) + 1 );
2065
2066 // We need a Region to merge the exit from the peeled body and the
2067 // exit from the old loop body.
2068 RegionNode *r = new RegionNode(3);
2069 // Map the old use to the new merge point
2070 old_new.map( use->_idx, r );
2071 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use));
2072 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" );
2073
2074 // The original user of 'use' uses 'r' instead.
2075 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
2076 Node* useuse = use->last_out(l);
2077 _igvn.rehash_node_delayed(useuse);
2078 uint uses_found = 0;
2079 if( useuse->in(0) == use ) {
2080 useuse->set_req(0, r);
2081 uses_found++;
2082 if( useuse->is_CFG() ) {
2083 assert( dom_depth(useuse) > dd_r, "" );
2084 set_idom(useuse, r, dom_depth(useuse));
2085 }
2086 }
2087 for( uint k = 1; k < useuse->req(); k++ ) {
2088 if( useuse->in(k) == use ) {
2089 useuse->set_req(k, r);
2090 uses_found++;
2091 if (useuse->is_Loop() && k == LoopNode::EntryControl) {
2092 assert(dom_depth(useuse) > dd_r , "");
2093 set_idom(useuse, r, dom_depth(useuse));
2094 }
2095 }
2096 }
2097 l -= uses_found; // we deleted 1 or more copies of this edge
2098 }
2099
2100 // Now finish up 'r'
2101 r->set_req( 1, newuse );
2102 r->set_req( 2, use );
2103 _igvn.register_new_node_with_optimizer(r);
2104 set_loop(r, use_loop);
2105 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r);
2106 } // End of if a loop-exit test
2107 }
2108 }
2109
2110 // Step 4: If loop-invariant use is not control, it must be dominated by a
2111 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region
2112 // there if needed. Make a Phi there merging old and new used values.
2113 Node_List *split_if_set = NULL;
2114 Node_List *split_bool_set = NULL;
2115 Node_List *split_cex_set = NULL;
2116 for( i = 0; i < loop->_body.size(); i++ ) {
2117 Node* old = loop->_body.at(i);
2118 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2119 split_bool_set, split_cex_set, worklist, new_counter,
2120 mode);
2121 }
2122
2123 for (i = 0; i < extra_data_nodes.size(); i++) {
2124 Node* old = extra_data_nodes.at(i);
2125 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2126 split_bool_set, split_cex_set, worklist, new_counter,
2127 mode);
2128 }
2129
2130 // Check for IFs that need splitting/cloning. Happens if an IF outside of
2131 // the loop uses a condition set in the loop. The original IF probably
2132 // takes control from one or more OLD Regions (which in turn get from NEW
2133 // Regions). In any case, there will be a set of Phis for each merge point
2134 // from the IF up to where the original BOOL def exists the loop.
2135 if (split_if_set) {
2136 while (split_if_set->size()) {
2137 Node *iff = split_if_set->pop();
2138 if (iff->in(1)->is_Phi()) {
2139 Node *b = clone_iff(iff->in(1)->as_Phi(), loop);
2140 _igvn.replace_input_of(iff, 1, b);
2141 }
2142 }
2143 }
2144 if (split_bool_set) {
2145 while (split_bool_set->size()) {
2146 Node *b = split_bool_set->pop();
2147 Node *phi = b->in(1);
2148 assert(phi->is_Phi(), "");
2149 CmpNode *cmp = clone_bool((PhiNode*)phi, loop);
2150 _igvn.replace_input_of(b, 1, cmp);
2151 }
2152 }
2153 if (split_cex_set) {
2154 while (split_cex_set->size()) {
2155 Node *b = split_cex_set->pop();
2156 assert(b->in(0)->is_Region(), "");
2157 assert(b->in(1)->is_Phi(), "");
2158 assert(b->in(0)->in(0) == b->in(1)->in(0), "");
2159 split_up(b, b->in(0), NULL);
2160 }
2161 }
2162
2163 }
2164
2165
2166 //---------------------- stride_of_possible_iv -------------------------------------
2167 // Looks for an iff/bool/comp with one operand of the compare
2168 // being a cycle involving an add and a phi,
2169 // with an optional truncation (left-shift followed by a right-shift)
2170 // of the add. Returns zero if not an iv.
2171 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
2172 Node* trunc1 = NULL;
2173 Node* trunc2 = NULL;
2174 const TypeInt* ttype = NULL;
2175 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) {
2176 return 0;
2177 }
2178 BoolNode* bl = iff->in(1)->as_Bool();
2179 Node* cmp = bl->in(1);
2180 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) {
2181 return 0;
2182 }
2183 // Must have an invariant operand
2184 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) {
2185 return 0;
2186 }
2187 Node* add2 = NULL;
2188 Node* cmp1 = cmp->in(1);
2189 if (cmp1->is_Phi()) {
2190 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
2191 Node* phi = cmp1;
2192 for (uint i = 1; i < phi->req(); i++) {
2193 Node* in = phi->in(i);
2194 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in,
2195 &trunc1, &trunc2, &ttype);
2196 if (add && add->in(1) == phi) {
2197 add2 = add->in(2);
2198 break;
2199 }
2200 }
2201 } else {
2202 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
2203 Node* addtrunc = cmp1;
2204 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc,
2205 &trunc1, &trunc2, &ttype);
2206 if (add && add->in(1)->is_Phi()) {
2207 Node* phi = add->in(1);
2208 for (uint i = 1; i < phi->req(); i++) {
2209 if (phi->in(i) == addtrunc) {
2210 add2 = add->in(2);
2211 break;
2212 }
2213 }
2214 }
2215 }
2216 if (add2 != NULL) {
2217 const TypeInt* add2t = _igvn.type(add2)->is_int();
2218 if (add2t->is_con()) {
2219 return add2t->get_con();
2220 }
2221 }
2222 return 0;
2223 }
2224
2225
2226 //---------------------- stay_in_loop -------------------------------------
2227 // Return the (unique) control output node that's in the loop (if it exists.)
2228 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
2229 Node* unique = NULL;
2230 if (!n) return NULL;
2231 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2232 Node* use = n->fast_out(i);
2233 if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
2234 if (unique != NULL) {
2235 return NULL;
2236 }
2237 unique = use;
2238 }
2239 }
2240 return unique;
2241 }
2242
2243 //------------------------------ register_node -------------------------------------
2244 // Utility to register node "n" with PhaseIdealLoop
2245 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) {
2246 _igvn.register_new_node_with_optimizer(n);
2247 loop->_body.push(n);
2248 if (n->is_CFG()) {
2249 set_loop(n, loop);
2250 set_idom(n, pred, ddepth);
2251 } else {
2252 set_ctrl(n, pred);
2253 }
2254 }
2255
2256 //------------------------------ proj_clone -------------------------------------
2257 // Utility to create an if-projection
2258 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
2259 ProjNode* c = p->clone()->as_Proj();
2260 c->set_req(0, iff);
2261 return c;
2262 }
2263
2264 //------------------------------ short_circuit_if -------------------------------------
2265 // Force the iff control output to be the live_proj
2266 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
2267 guarantee(live_proj != NULL, "null projection");
2268 int proj_con = live_proj->_con;
2269 assert(proj_con == 0 || proj_con == 1, "false or true projection");
2270 Node *con = _igvn.intcon(proj_con);
2271 set_ctrl(con, C->root());
2272 if (iff) {
2273 iff->set_req(1, con);
2274 }
2275 return con;
2276 }
2277
2278 //------------------------------ insert_if_before_proj -------------------------------------
2279 // Insert a new if before an if projection (* - new node)
2280 //
2281 // before
2282 // if(test)
2283 // / \
2284 // v v
2285 // other-proj proj (arg)
2286 //
2287 // after
2288 // if(test)
2289 // / \
2290 // / v
2291 // | * proj-clone
2292 // v |
2293 // other-proj v
2294 // * new_if(relop(cmp[IU](left,right)))
2295 // / \
2296 // v v
2297 // * new-proj proj
2298 // (returned)
2299 //
2300 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) {
2301 IfNode* iff = proj->in(0)->as_If();
2302 IdealLoopTree *loop = get_loop(proj);
2303 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2304 int ddepth = dom_depth(proj);
2305
2306 _igvn.rehash_node_delayed(iff);
2307 _igvn.rehash_node_delayed(proj);
2308
2309 proj->set_req(0, NULL); // temporary disconnect
2310 ProjNode* proj2 = proj_clone(proj, iff);
2311 register_node(proj2, loop, iff, ddepth);
2312
2313 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right);
2314 register_node(cmp, loop, proj2, ddepth);
2315
2316 BoolNode* bol = new BoolNode(cmp, relop);
2317 register_node(bol, loop, proj2, ddepth);
2318
2319 int opcode = iff->Opcode();
2320 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode");
2321 IfNode* new_if = (opcode == Op_If) ? new IfNode(proj2, bol, iff->_prob, iff->_fcnt):
2322 new RangeCheckNode(proj2, bol, iff->_prob, iff->_fcnt);
2323 register_node(new_if, loop, proj2, ddepth);
2324
2325 proj->set_req(0, new_if); // reattach
2326 set_idom(proj, new_if, ddepth);
2327
2328 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
2329 guarantee(new_exit != NULL, "null exit node");
2330 register_node(new_exit, get_loop(other_proj), new_if, ddepth);
2331
2332 return new_exit;
2333 }
2334
2335 //------------------------------ insert_region_before_proj -------------------------------------
2336 // Insert a region before an if projection (* - new node)
2337 //
2338 // before
2339 // if(test)
2340 // / |
2341 // v |
2342 // proj v
2343 // other-proj
2344 //
2345 // after
2346 // if(test)
2347 // / |
2348 // v |
2349 // * proj-clone v
2350 // | other-proj
2351 // v
2352 // * new-region
2353 // |
2354 // v
2355 // * dum_if
2356 // / \
2357 // v \
2358 // * dum-proj v
2359 // proj
2360 //
2361 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
2362 IfNode* iff = proj->in(0)->as_If();
2363 IdealLoopTree *loop = get_loop(proj);
2364 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2365 int ddepth = dom_depth(proj);
2366
2367 _igvn.rehash_node_delayed(iff);
2368 _igvn.rehash_node_delayed(proj);
2369
2370 proj->set_req(0, NULL); // temporary disconnect
2371 ProjNode* proj2 = proj_clone(proj, iff);
2372 register_node(proj2, loop, iff, ddepth);
2373
2374 RegionNode* reg = new RegionNode(2);
2375 reg->set_req(1, proj2);
2376 register_node(reg, loop, iff, ddepth);
2377
2378 IfNode* dum_if = new IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt);
2379 register_node(dum_if, loop, reg, ddepth);
2380
2381 proj->set_req(0, dum_if); // reattach
2382 set_idom(proj, dum_if, ddepth);
2383
2384 ProjNode* dum_proj = proj_clone(other_proj, dum_if);
2385 register_node(dum_proj, loop, dum_if, ddepth);
2386
2387 return reg;
2388 }
2389
2390 //------------------------------ insert_cmpi_loop_exit -------------------------------------
2391 // Clone a signed compare loop exit from an unsigned compare and
2392 // insert it before the unsigned cmp on the stay-in-loop path.
2393 // All new nodes inserted in the dominator tree between the original
2394 // if and it's projections. The original if test is replaced with
2395 // a constant to force the stay-in-loop path.
2396 //
2397 // This is done to make sure that the original if and it's projections
2398 // still dominate the same set of control nodes, that the ctrl() relation
2399 // from data nodes to them is preserved, and that their loop nesting is
2400 // preserved.
2401 //
2402 // before
2403 // if(i <u limit) unsigned compare loop exit
2404 // / |
2405 // v v
2406 // exit-proj stay-in-loop-proj
2407 //
2408 // after
2409 // if(stay-in-loop-const) original if
2410 // / |
2411 // / v
2412 // / if(i < limit) new signed test
2413 // / / |
2414 // / / v
2415 // / / if(i <u limit) new cloned unsigned test
2416 // / / / |
2417 // v v v |
2418 // region |
2419 // | |
2420 // dum-if |
2421 // / | |
2422 // ether | |
2423 // v v
2424 // exit-proj stay-in-loop-proj
2425 //
2426 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) {
2427 const bool Signed = true;
2428 const bool Unsigned = false;
2429
2430 BoolNode* bol = if_cmpu->in(1)->as_Bool();
2431 if (bol->_test._test != BoolTest::lt) return NULL;
2432 CmpNode* cmpu = bol->in(1)->as_Cmp();
2433 if (cmpu->Opcode() != Op_CmpU) return NULL;
2434 int stride = stride_of_possible_iv(if_cmpu);
2435 if (stride == 0) return NULL;
2436
2437 Node* lp_proj = stay_in_loop(if_cmpu, loop);
2438 guarantee(lp_proj != NULL, "null loop node");
2439
2440 ProjNode* lp_continue = lp_proj->as_Proj();
2441 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
2442
2443 Node* limit = NULL;
2444 if (stride > 0) {
2445 limit = cmpu->in(2);
2446 } else {
2447 limit = _igvn.makecon(TypeInt::ZERO);
2448 set_ctrl(limit, C->root());
2449 }
2450 // Create a new region on the exit path
2451 RegionNode* reg = insert_region_before_proj(lp_exit);
2452 guarantee(reg != NULL, "null region node");
2453
2454 // Clone the if-cmpu-true-false using a signed compare
2455 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
2456 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue);
2457 reg->add_req(cmpi_exit);
2458
2459 // Clone the if-cmpu-true-false
2460 BoolTest::mask rel_u = bol->_test._test;
2461 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
2462 reg->add_req(cmpu_exit);
2463
2464 // Force original if to stay in loop.
2465 short_circuit_if(if_cmpu, lp_continue);
2466
2467 return cmpi_exit->in(0)->as_If();
2468 }
2469
2470 //------------------------------ remove_cmpi_loop_exit -------------------------------------
2471 // Remove a previously inserted signed compare loop exit.
2472 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
2473 Node* lp_proj = stay_in_loop(if_cmp, loop);
2474 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
2475 stay_in_loop(lp_proj, loop)->is_If() &&
2476 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
2477 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
2478 set_ctrl(con, C->root());
2479 if_cmp->set_req(1, con);
2480 }
2481
2482 //------------------------------ scheduled_nodelist -------------------------------------
2483 // Create a post order schedule of nodes that are in the
2484 // "member" set. The list is returned in "sched".
2485 // The first node in "sched" is the loop head, followed by
2486 // nodes which have no inputs in the "member" set, and then
2487 // followed by the nodes that have an immediate input dependence
2488 // on a node in "sched".
2489 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
2490
2491 assert(member.test(loop->_head->_idx), "loop head must be in member set");
2492 Arena *a = Thread::current()->resource_area();
2493 VectorSet visited(a);
2494 Node_Stack nstack(a, loop->_body.size());
2495
2496 Node* n = loop->_head; // top of stack is cached in "n"
2497 uint idx = 0;
2498 visited.set(n->_idx);
2499
2500 // Initially push all with no inputs from within member set
2501 for(uint i = 0; i < loop->_body.size(); i++ ) {
2502 Node *elt = loop->_body.at(i);
2503 if (member.test(elt->_idx)) {
2504 bool found = false;
2505 for (uint j = 0; j < elt->req(); j++) {
2506 Node* def = elt->in(j);
2507 if (def && member.test(def->_idx) && def != elt) {
2508 found = true;
2509 break;
2510 }
2511 }
2512 if (!found && elt != loop->_head) {
2513 nstack.push(n, idx);
2514 n = elt;
2515 assert(!visited.test(n->_idx), "not seen yet");
2516 visited.set(n->_idx);
2517 }
2518 }
2519 }
2520
2521 // traverse out's that are in the member set
2522 while (true) {
2523 if (idx < n->outcnt()) {
2524 Node* use = n->raw_out(idx);
2525 idx++;
2526 if (!visited.test_set(use->_idx)) {
2527 if (member.test(use->_idx)) {
2528 nstack.push(n, idx);
2529 n = use;
2530 idx = 0;
2531 }
2532 }
2533 } else {
2534 // All outputs processed
2535 sched.push(n);
2536 if (nstack.is_empty()) break;
2537 n = nstack.node();
2538 idx = nstack.index();
2539 nstack.pop();
2540 }
2541 }
2542 }
2543
2544
2545 //------------------------------ has_use_in_set -------------------------------------
2546 // Has a use in the vector set
2547 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
2548 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2549 Node* use = n->fast_out(j);
2550 if (vset.test(use->_idx)) {
2551 return true;
2552 }
2553 }
2554 return false;
2555 }
2556
2557
2558 //------------------------------ has_use_internal_to_set -------------------------------------
2559 // Has use internal to the vector set (ie. not in a phi at the loop head)
2560 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
2561 Node* head = loop->_head;
2562 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2563 Node* use = n->fast_out(j);
2564 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
2565 return true;
2566 }
2567 }
2568 return false;
2569 }
2570
2571
2572 //------------------------------ clone_for_use_outside_loop -------------------------------------
2573 // clone "n" for uses that are outside of loop
2574 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
2575 int cloned = 0;
2576 assert(worklist.size() == 0, "should be empty");
2577 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2578 Node* use = n->fast_out(j);
2579 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
2580 worklist.push(use);
2581 }
2582 }
2583 while( worklist.size() ) {
2584 Node *use = worklist.pop();
2585 if (!has_node(use) || use->in(0) == C->top()) continue;
2586 uint j;
2587 for (j = 0; j < use->req(); j++) {
2588 if (use->in(j) == n) break;
2589 }
2590 assert(j < use->req(), "must be there");
2591
2592 // clone "n" and insert it between the inputs of "n" and the use outside the loop
2593 Node* n_clone = n->clone();
2594 _igvn.replace_input_of(use, j, n_clone);
2595 cloned++;
2596 Node* use_c;
2597 if (!use->is_Phi()) {
2598 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
2599 } else {
2600 // Use in a phi is considered a use in the associated predecessor block
2601 use_c = use->in(0)->in(j);
2602 }
2603 set_ctrl(n_clone, use_c);
2604 assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
2605 get_loop(use_c)->_body.push(n_clone);
2606 _igvn.register_new_node_with_optimizer(n_clone);
2607 #if !defined(PRODUCT)
2608 if (TracePartialPeeling) {
2609 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
2610 }
2611 #endif
2612 }
2613 return cloned;
2614 }
2615
2616
2617 //------------------------------ clone_for_special_use_inside_loop -------------------------------------
2618 // clone "n" for special uses that are in the not_peeled region.
2619 // If these def-uses occur in separate blocks, the code generator
2620 // marks the method as not compilable. For example, if a "BoolNode"
2621 // is in a different basic block than the "IfNode" that uses it, then
2622 // the compilation is aborted in the code generator.
2623 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
2624 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
2625 if (n->is_Phi() || n->is_Load()) {
2626 return;
2627 }
2628 assert(worklist.size() == 0, "should be empty");
2629 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2630 Node* use = n->fast_out(j);
2631 if ( not_peel.test(use->_idx) &&
2632 (use->is_If() || use->is_CMove() || use->is_Bool()) &&
2633 use->in(1) == n) {
2634 worklist.push(use);
2635 }
2636 }
2637 if (worklist.size() > 0) {
2638 // clone "n" and insert it between inputs of "n" and the use
2639 Node* n_clone = n->clone();
2640 loop->_body.push(n_clone);
2641 _igvn.register_new_node_with_optimizer(n_clone);
2642 set_ctrl(n_clone, get_ctrl(n));
2643 sink_list.push(n_clone);
2644 not_peel <<= n_clone->_idx; // add n_clone to not_peel set.
2645 #if !defined(PRODUCT)
2646 if (TracePartialPeeling) {
2647 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
2648 }
2649 #endif
2650 while( worklist.size() ) {
2651 Node *use = worklist.pop();
2652 _igvn.rehash_node_delayed(use);
2653 for (uint j = 1; j < use->req(); j++) {
2654 if (use->in(j) == n) {
2655 use->set_req(j, n_clone);
2656 }
2657 }
2658 }
2659 }
2660 }
2661
2662
2663 //------------------------------ insert_phi_for_loop -------------------------------------
2664 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
2665 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
2666 Node *phi = PhiNode::make(lp, back_edge_val);
2667 phi->set_req(LoopNode::EntryControl, lp_entry_val);
2668 // Use existing phi if it already exists
2669 Node *hit = _igvn.hash_find_insert(phi);
2670 if( hit == NULL ) {
2671 _igvn.register_new_node_with_optimizer(phi);
2672 set_ctrl(phi, lp);
2673 } else {
2674 // Remove the new phi from the graph and use the hit
2675 _igvn.remove_dead_node(phi);
2676 phi = hit;
2677 }
2678 _igvn.replace_input_of(use, idx, phi);
2679 }
2680
2681 #ifdef ASSERT
2682 //------------------------------ is_valid_loop_partition -------------------------------------
2683 // Validate the loop partition sets: peel and not_peel
2684 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
2685 VectorSet& not_peel ) {
2686 uint i;
2687 // Check that peel_list entries are in the peel set
2688 for (i = 0; i < peel_list.size(); i++) {
2689 if (!peel.test(peel_list.at(i)->_idx)) {
2690 return false;
2691 }
2692 }
2693 // Check at loop members are in one of peel set or not_peel set
2694 for (i = 0; i < loop->_body.size(); i++ ) {
2695 Node *def = loop->_body.at(i);
2696 uint di = def->_idx;
2697 // Check that peel set elements are in peel_list
2698 if (peel.test(di)) {
2699 if (not_peel.test(di)) {
2700 return false;
2701 }
2702 // Must be in peel_list also
2703 bool found = false;
2704 for (uint j = 0; j < peel_list.size(); j++) {
2705 if (peel_list.at(j)->_idx == di) {
2706 found = true;
2707 break;
2708 }
2709 }
2710 if (!found) {
2711 return false;
2712 }
2713 } else if (not_peel.test(di)) {
2714 if (peel.test(di)) {
2715 return false;
2716 }
2717 } else {
2718 return false;
2719 }
2720 }
2721 return true;
2722 }
2723
2724 //------------------------------ is_valid_clone_loop_exit_use -------------------------------------
2725 // Ensure a use outside of loop is of the right form
2726 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
2727 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2728 return (use->is_Phi() &&
2729 use_c->is_Region() && use_c->req() == 3 &&
2730 (use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
2731 use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
2732 use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
2733 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
2734 }
2735
2736 //------------------------------ is_valid_clone_loop_form -------------------------------------
2737 // Ensure that all uses outside of loop are of the right form
2738 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
2739 uint orig_exit_idx, uint clone_exit_idx) {
2740 uint len = peel_list.size();
2741 for (uint i = 0; i < len; i++) {
2742 Node *def = peel_list.at(i);
2743
2744 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2745 Node *use = def->fast_out(j);
2746 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2747 if (!loop->is_member(get_loop(use_c))) {
2748 // use is not in the loop, check for correct structure
2749 if (use->in(0) == def) {
2750 // Okay
2751 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
2752 return false;
2753 }
2754 }
2755 }
2756 }
2757 return true;
2758 }
2759 #endif
2760
2761 //------------------------------ partial_peel -------------------------------------
2762 // Partially peel (aka loop rotation) the top portion of a loop (called
2763 // the peel section below) by cloning it and placing one copy just before
2764 // the new loop head and the other copy at the bottom of the new loop.
2765 //
2766 // before after where it came from
2767 //
2768 // stmt1 stmt1
2769 // loop: stmt2 clone
2770 // stmt2 if condA goto exitA clone
2771 // if condA goto exitA new_loop: new
2772 // stmt3 stmt3 clone
2773 // if !condB goto loop if condB goto exitB clone
2774 // exitB: stmt2 orig
2775 // stmt4 if !condA goto new_loop orig
2776 // exitA: goto exitA
2777 // exitB:
2778 // stmt4
2779 // exitA:
2780 //
2781 // Step 1: find the cut point: an exit test on probable
2782 // induction variable.
2783 // Step 2: schedule (with cloning) operations in the peel
2784 // section that can be executed after the cut into
2785 // the section that is not peeled. This may need
2786 // to clone operations into exit blocks. For
2787 // instance, a reference to A[i] in the not-peel
2788 // section and a reference to B[i] in an exit block
2789 // may cause a left-shift of i by 2 to be placed
2790 // in the peel block. This step will clone the left
2791 // shift into the exit block and sink the left shift
2792 // from the peel to the not-peel section.
2793 // Step 3: clone the loop, retarget the control, and insert
2794 // phis for values that are live across the new loop
2795 // head. This is very dependent on the graph structure
2796 // from clone_loop. It creates region nodes for
2797 // exit control and associated phi nodes for values
2798 // flow out of the loop through that exit. The region
2799 // node is dominated by the clone's control projection.
2800 // So the clone's peel section is placed before the
2801 // new loop head, and the clone's not-peel section is
2802 // forms the top part of the new loop. The original
2803 // peel section forms the tail of the new loop.
2804 // Step 4: update the dominator tree and recompute the
2805 // dominator depth.
2806 //
2807 // orig
2808 //
2809 // stmt1
2810 // |
2811 // v
2812 // loop predicate
2813 // |
2814 // v
2815 // loop<----+
2816 // | |
2817 // stmt2 |
2818 // | |
2819 // v |
2820 // ifA |
2821 // / | |
2822 // v v |
2823 // false true ^ <-- last_peel
2824 // / | |
2825 // / ===|==cut |
2826 // / stmt3 | <-- first_not_peel
2827 // / | |
2828 // | v |
2829 // v ifB |
2830 // exitA: / \ |
2831 // / \ |
2832 // v v |
2833 // false true |
2834 // / \ |
2835 // / ----+
2836 // |
2837 // v
2838 // exitB:
2839 // stmt4
2840 //
2841 //
2842 // after clone loop
2843 //
2844 // stmt1
2845 // |
2846 // v
2847 // loop predicate
2848 // / \
2849 // clone / \ orig
2850 // / \
2851 // / \
2852 // v v
2853 // +---->loop loop<----+
2854 // | | | |
2855 // | stmt2 stmt2 |
2856 // | | | |
2857 // | v v |
2858 // | ifA ifA |
2859 // | | \ / | |
2860 // | v v v v |
2861 // ^ true false false true ^ <-- last_peel
2862 // | | ^ \ / | |
2863 // | cut==|== \ \ / ===|==cut |
2864 // | stmt3 \ \ / stmt3 | <-- first_not_peel
2865 // | | dom | | | |
2866 // | v \ 1v v2 v |
2867 // | ifB regionA ifB |
2868 // | / \ | / \ |
2869 // | / \ v / \ |
2870 // | v v exitA: v v |
2871 // | true false false true |
2872 // | / ^ \ / \ |
2873 // +---- \ \ / ----+
2874 // dom \ /
2875 // \ 1v v2
2876 // regionB
2877 // |
2878 // v
2879 // exitB:
2880 // stmt4
2881 //
2882 //
2883 // after partial peel
2884 //
2885 // stmt1
2886 // |
2887 // v
2888 // loop predicate
2889 // /
2890 // clone / orig
2891 // / TOP
2892 // / \
2893 // v v
2894 // TOP->loop loop----+
2895 // | | |
2896 // stmt2 stmt2 |
2897 // | | |
2898 // v v |
2899 // ifA ifA |
2900 // | \ / | |
2901 // v v v v |
2902 // true false false true | <-- last_peel
2903 // | ^ \ / +------|---+
2904 // +->newloop \ \ / === ==cut | |
2905 // | stmt3 \ \ / TOP | |
2906 // | | dom | | stmt3 | | <-- first_not_peel
2907 // | v \ 1v v2 v | |
2908 // | ifB regionA ifB ^ v
2909 // | / \ | / \ | |
2910 // | / \ v / \ | |
2911 // | v v exitA: v v | |
2912 // | true false false true | |
2913 // | / ^ \ / \ | |
2914 // | | \ \ / v | |
2915 // | | dom \ / TOP | |
2916 // | | \ 1v v2 | |
2917 // ^ v regionB | |
2918 // | | | | |
2919 // | | v ^ v
2920 // | | exitB: | |
2921 // | | stmt4 | |
2922 // | +------------>-----------------+ |
2923 // | |
2924 // +-----------------<---------------------+
2925 //
2926 //
2927 // final graph
2928 //
2929 // stmt1
2930 // |
2931 // v
2932 // loop predicate
2933 // |
2934 // v
2935 // stmt2 clone
2936 // |
2937 // v
2938 // ........> ifA clone
2939 // : / |
2940 // dom / |
2941 // : v v
2942 // : false true
2943 // : | |
2944 // : | v
2945 // : | newloop<-----+
2946 // : | | |
2947 // : | stmt3 clone |
2948 // : | | |
2949 // : | v |
2950 // : | ifB |
2951 // : | / \ |
2952 // : | v v |
2953 // : | false true |
2954 // : | | | |
2955 // : | v stmt2 |
2956 // : | exitB: | |
2957 // : | stmt4 v |
2958 // : | ifA orig |
2959 // : | / \ |
2960 // : | / \ |
2961 // : | v v |
2962 // : | false true |
2963 // : | / \ |
2964 // : v v -----+
2965 // RegionA
2966 // |
2967 // v
2968 // exitA
2969 //
2970 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
2971
2972 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only");
2973 if (!loop->_head->is_Loop()) {
2974 return false; }
2975
2976 LoopNode *head = loop->_head->as_Loop();
2977
2978 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
2979 return false;
2980 }
2981
2982 // Check for complex exit control
2983 for(uint ii = 0; ii < loop->_body.size(); ii++ ) {
2984 Node *n = loop->_body.at(ii);
2985 int opc = n->Opcode();
2986 if (n->is_Call() ||
2987 opc == Op_Catch ||
2988 opc == Op_CatchProj ||
2989 opc == Op_Jump ||
2990 opc == Op_JumpProj) {
2991 #if !defined(PRODUCT)
2992 if (TracePartialPeeling) {
2993 tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
2994 }
2995 #endif
2996 return false;
2997 }
2998 }
2999
3000 int dd = dom_depth(head);
3001
3002 // Step 1: find cut point
3003
3004 // Walk up dominators to loop head looking for first loop exit
3005 // which is executed on every path thru loop.
3006 IfNode *peel_if = NULL;
3007 IfNode *peel_if_cmpu = NULL;
3008
3009 Node *iff = loop->tail();
3010 while( iff != head ) {
3011 if( iff->is_If() ) {
3012 Node *ctrl = get_ctrl(iff->in(1));
3013 if (ctrl->is_top()) return false; // Dead test on live IF.
3014 // If loop-varying exit-test, check for induction variable
3015 if( loop->is_member(get_loop(ctrl)) &&
3016 loop->is_loop_exit(iff) &&
3017 is_possible_iv_test(iff)) {
3018 Node* cmp = iff->in(1)->in(1);
3019 if (cmp->Opcode() == Op_CmpI) {
3020 peel_if = iff->as_If();
3021 } else {
3022 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
3023 peel_if_cmpu = iff->as_If();
3024 }
3025 }
3026 }
3027 iff = idom(iff);
3028 }
3029 // Prefer signed compare over unsigned compare.
3030 IfNode* new_peel_if = NULL;
3031 if (peel_if == NULL) {
3032 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) {
3033 return false; // No peel point found
3034 }
3035 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
3036 if (new_peel_if == NULL) {
3037 return false; // No peel point found
3038 }
3039 peel_if = new_peel_if;
3040 }
3041 Node* last_peel = stay_in_loop(peel_if, loop);
3042 Node* first_not_peeled = stay_in_loop(last_peel, loop);
3043 if (first_not_peeled == NULL || first_not_peeled == head) {
3044 return false;
3045 }
3046
3047 #if !defined(PRODUCT)
3048 if (TraceLoopOpts) {
3049 tty->print("PartialPeel ");
3050 loop->dump_head();
3051 }
3052
3053 if (TracePartialPeeling) {
3054 tty->print_cr("before partial peel one iteration");
3055 Node_List wl;
3056 Node* t = head->in(2);
3057 while (true) {
3058 wl.push(t);
3059 if (t == head) break;
3060 t = idom(t);
3061 }
3062 while (wl.size() > 0) {
3063 Node* tt = wl.pop();
3064 tt->dump();
3065 if (tt == last_peel) tty->print_cr("-- cut --");
3066 }
3067 }
3068 #endif
3069 ResourceArea *area = Thread::current()->resource_area();
3070 VectorSet peel(area);
3071 VectorSet not_peel(area);
3072 Node_List peel_list(area);
3073 Node_List worklist(area);
3074 Node_List sink_list(area);
3075
3076 // Set of cfg nodes to peel are those that are executable from
3077 // the head through last_peel.
3078 assert(worklist.size() == 0, "should be empty");
3079 worklist.push(head);
3080 peel.set(head->_idx);
3081 while (worklist.size() > 0) {
3082 Node *n = worklist.pop();
3083 if (n != last_peel) {
3084 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3085 Node* use = n->fast_out(j);
3086 if (use->is_CFG() &&
3087 loop->is_member(get_loop(use)) &&
3088 !peel.test_set(use->_idx)) {
3089 worklist.push(use);
3090 }
3091 }
3092 }
3093 }
3094
3095 // Set of non-cfg nodes to peel are those that are control
3096 // dependent on the cfg nodes.
3097 uint i;
3098 for(i = 0; i < loop->_body.size(); i++ ) {
3099 Node *n = loop->_body.at(i);
3100 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
3101 if (peel.test(n_c->_idx)) {
3102 peel.set(n->_idx);
3103 } else {
3104 not_peel.set(n->_idx);
3105 }
3106 }
3107
3108 // Step 2: move operations from the peeled section down into the
3109 // not-peeled section
3110
3111 // Get a post order schedule of nodes in the peel region
3112 // Result in right-most operand.
3113 scheduled_nodelist(loop, peel, peel_list );
3114
3115 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
3116
3117 // For future check for too many new phis
3118 uint old_phi_cnt = 0;
3119 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
3120 Node* use = head->fast_out(j);
3121 if (use->is_Phi()) old_phi_cnt++;
3122 }
3123
3124 #if !defined(PRODUCT)
3125 if (TracePartialPeeling) {
3126 tty->print_cr("\npeeled list");
3127 }
3128 #endif
3129
3130 // Evacuate nodes in peel region into the not_peeled region if possible
3131 uint new_phi_cnt = 0;
3132 uint cloned_for_outside_use = 0;
3133 for (i = 0; i < peel_list.size();) {
3134 Node* n = peel_list.at(i);
3135 #if !defined(PRODUCT)
3136 if (TracePartialPeeling) n->dump();
3137 #endif
3138 bool incr = true;
3139 if ( !n->is_CFG() ) {
3140
3141 if ( has_use_in_set(n, not_peel) ) {
3142
3143 // If not used internal to the peeled region,
3144 // move "n" from peeled to not_peeled region.
3145
3146 if ( !has_use_internal_to_set(n, peel, loop) ) {
3147
3148 // if not pinned and not a load (which maybe anti-dependent on a store)
3149 // and not a CMove (Matcher expects only bool->cmove).
3150 if (n->in(0) == NULL && !n->is_Load() && !n->is_CMove() && n->Opcode() != Op_ShenandoahWBMemProj) {
3151 cloned_for_outside_use += clone_for_use_outside_loop( loop, n, worklist );
3152 sink_list.push(n);
3153 peel >>= n->_idx; // delete n from peel set.
3154 not_peel <<= n->_idx; // add n to not_peel set.
3155 peel_list.remove(i);
3156 incr = false;
3157 #if !defined(PRODUCT)
3158 if (TracePartialPeeling) {
3159 tty->print_cr("sink to not_peeled region: %d newbb: %d",
3160 n->_idx, get_ctrl(n)->_idx);
3161 }
3162 #endif
3163 }
3164 } else {
3165 // Otherwise check for special def-use cases that span
3166 // the peel/not_peel boundary such as bool->if
3167 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist );
3168 new_phi_cnt++;
3169 }
3170 }
3171 }
3172 if (incr) i++;
3173 }
3174
3175 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) {
3176 #if !defined(PRODUCT)
3177 if (TracePartialPeeling) {
3178 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c",
3179 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F');
3180 }
3181 #endif
3182 if (new_peel_if != NULL) {
3183 remove_cmpi_loop_exit(new_peel_if, loop);
3184 }
3185 // Inhibit more partial peeling on this loop
3186 assert(!head->is_partial_peel_loop(), "not partial peeled");
3187 head->mark_partial_peel_failed();
3188 if (cloned_for_outside_use > 0) {
3189 // Terminate this round of loop opts because
3190 // the graph outside this loop was changed.
3191 C->set_major_progress();
3192 return true;
3193 }
3194 return false;
3195 }
3196
3197 // Step 3: clone loop, retarget control, and insert new phis
3198
3199 // Create new loop head for new phis and to hang
3200 // the nodes being moved (sinked) from the peel region.
3201 LoopNode* new_head = new LoopNode(last_peel, last_peel);
3202 new_head->set_unswitch_count(head->unswitch_count()); // Preserve
3203 _igvn.register_new_node_with_optimizer(new_head);
3204 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
3205 _igvn.replace_input_of(first_not_peeled, 0, new_head);
3206 set_loop(new_head, loop);
3207 loop->_body.push(new_head);
3208 not_peel.set(new_head->_idx);
3209 set_idom(new_head, last_peel, dom_depth(first_not_peeled));
3210 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
3211
3212 while (sink_list.size() > 0) {
3213 Node* n = sink_list.pop();
3214 set_ctrl(n, new_head);
3215 }
3216
3217 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
3218
3219 clone_loop(loop, old_new, dd, IgnoreStripMined);
3220
3221 const uint clone_exit_idx = 1;
3222 const uint orig_exit_idx = 2;
3223 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop");
3224
3225 Node* head_clone = old_new[head->_idx];
3226 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop();
3227 Node* orig_tail_clone = head_clone->in(2);
3228
3229 // Add phi if "def" node is in peel set and "use" is not
3230
3231 for(i = 0; i < peel_list.size(); i++ ) {
3232 Node *def = peel_list.at(i);
3233 if (!def->is_CFG()) {
3234 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
3235 Node *use = def->fast_out(j);
3236 if (has_node(use) && use->in(0) != C->top() &&
3237 (!peel.test(use->_idx) ||
3238 (use->is_Phi() && use->in(0) == head)) ) {
3239 worklist.push(use);
3240 }
3241 }
3242 while( worklist.size() ) {
3243 Node *use = worklist.pop();
3244 for (uint j = 1; j < use->req(); j++) {
3245 Node* n = use->in(j);
3246 if (n == def) {
3247
3248 // "def" is in peel set, "use" is not in peel set
3249 // or "use" is in the entry boundary (a phi) of the peel set
3250
3251 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
3252
3253 if ( loop->is_member(get_loop( use_c )) ) {
3254 // use is in loop
3255 if (old_new[use->_idx] != NULL) { // null for dead code
3256 Node* use_clone = old_new[use->_idx];
3257 _igvn.replace_input_of(use, j, C->top());
3258 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
3259 }
3260 } else {
3261 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
3262 // use is not in the loop, check if the live range includes the cut
3263 Node* lp_if = use_c->in(orig_exit_idx)->in(0);
3264 if (not_peel.test(lp_if->_idx)) {
3265 assert(j == orig_exit_idx, "use from original loop");
3266 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
3267 }
3268 }
3269 }
3270 }
3271 }
3272 }
3273 }
3274
3275 // Step 3b: retarget control
3276
3277 // Redirect control to the new loop head if a cloned node in
3278 // the not_peeled region has control that points into the peeled region.
3279 // This necessary because the cloned peeled region will be outside
3280 // the loop.
3281 // from to
3282 // cloned-peeled <---+
3283 // new_head_clone: | <--+
3284 // cloned-not_peeled in(0) in(0)
3285 // orig-peeled
3286
3287 for(i = 0; i < loop->_body.size(); i++ ) {
3288 Node *n = loop->_body.at(i);
3289 if (!n->is_CFG() && n->in(0) != NULL &&
3290 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
3291 Node* n_clone = old_new[n->_idx];
3292 _igvn.replace_input_of(n_clone, 0, new_head_clone);
3293 }
3294 }
3295
3296 // Backedge of the surviving new_head (the clone) is original last_peel
3297 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel);
3298
3299 // Cut first node in original not_peel set
3300 _igvn.rehash_node_delayed(new_head); // Multiple edge updates:
3301 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of
3302 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls
3303
3304 // Copy head_clone back-branch info to original head
3305 // and remove original head's loop entry and
3306 // clone head's back-branch
3307 _igvn.rehash_node_delayed(head); // Multiple edge updates
3308 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl));
3309 head->set_req(LoopNode::LoopBackControl, C->top());
3310 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top());
3311
3312 // Similarly modify the phis
3313 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
3314 Node* use = head->fast_out(k);
3315 if (use->is_Phi() && use->outcnt() > 0) {
3316 Node* use_clone = old_new[use->_idx];
3317 _igvn.rehash_node_delayed(use); // Multiple edge updates
3318 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl));
3319 use->set_req(LoopNode::LoopBackControl, C->top());
3320 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top());
3321 }
3322 }
3323
3324 // Step 4: update dominator tree and dominator depth
3325
3326 set_idom(head, orig_tail_clone, dd);
3327 recompute_dom_depth();
3328
3329 // Inhibit more partial peeling on this loop
3330 new_head_clone->set_partial_peel_loop();
3331 C->set_major_progress();
3332 loop->record_for_igvn();
3333
3334 #if !defined(PRODUCT)
3335 if (TracePartialPeeling) {
3336 tty->print_cr("\nafter partial peel one iteration");
3337 Node_List wl(area);
3338 Node* t = last_peel;
3339 while (true) {
3340 wl.push(t);
3341 if (t == head_clone) break;
3342 t = idom(t);
3343 }
3344 while (wl.size() > 0) {
3345 Node* tt = wl.pop();
3346 if (tt == head) tty->print_cr("orig head");
3347 else if (tt == new_head_clone) tty->print_cr("new head");
3348 else if (tt == head_clone) tty->print_cr("clone head");
3349 tt->dump();
3350 }
3351 }
3352 #endif
3353 return true;
3354 }
3355
3356 //------------------------------reorg_offsets----------------------------------
3357 // Reorganize offset computations to lower register pressure. Mostly
3358 // prevent loop-fallout uses of the pre-incremented trip counter (which are
3359 // then alive with the post-incremented trip counter forcing an extra
3360 // register move)
3361 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) {
3362 // Perform it only for canonical counted loops.
3363 // Loop's shape could be messed up by iteration_split_impl.
3364 if (!loop->_head->is_CountedLoop())
3365 return;
3366 if (!loop->_head->as_Loop()->is_valid_counted_loop())
3367 return;
3368
3369 CountedLoopNode *cl = loop->_head->as_CountedLoop();
3370 CountedLoopEndNode *cle = cl->loopexit();
3371 Node *exit = cle->proj_out(false);
3372 Node *phi = cl->phi();
3373
3374 // Check for the special case of folks using the pre-incremented
3375 // trip-counter on the fall-out path (forces the pre-incremented
3376 // and post-incremented trip counter to be live at the same time).
3377 // Fix this by adjusting to use the post-increment trip counter.
3378
3379 bool progress = true;
3380 while (progress) {
3381 progress = false;
3382 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
3383 Node* use = phi->fast_out(i); // User of trip-counter
3384 if (!has_ctrl(use)) continue;
3385 Node *u_ctrl = get_ctrl(use);
3386 if (use->is_Phi()) {
3387 u_ctrl = NULL;
3388 for (uint j = 1; j < use->req(); j++)
3389 if (use->in(j) == phi)
3390 u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j));
3391 }
3392 IdealLoopTree *u_loop = get_loop(u_ctrl);
3393 // Look for loop-invariant use
3394 if (u_loop == loop) continue;
3395 if (loop->is_member(u_loop)) continue;
3396 // Check that use is live out the bottom. Assuming the trip-counter
3397 // update is right at the bottom, uses of of the loop middle are ok.
3398 if (dom_lca(exit, u_ctrl) != exit) continue;
3399 // Hit! Refactor use to use the post-incremented tripcounter.
3400 // Compute a post-increment tripcounter.
3401 Node* c = exit;
3402 if (cl->is_strip_mined()) {
3403 IdealLoopTree* outer_loop = get_loop(cl->outer_loop());
3404 if (!outer_loop->is_member(u_loop)) {
3405 c = cl->outer_loop_exit();
3406 }
3407 }
3408 Node *opaq = new Opaque2Node(C, cle->incr());
3409 register_new_node(opaq, c);
3410 Node *neg_stride = _igvn.intcon(-cle->stride_con());
3411 set_ctrl(neg_stride, C->root());
3412 Node *post = new AddINode(opaq, neg_stride);
3413 register_new_node(post, c);
3414 _igvn.rehash_node_delayed(use);
3415 for (uint j = 1; j < use->req(); j++) {
3416 if (use->in(j) == phi)
3417 use->set_req(j, post);
3418 }
3419 // Since DU info changed, rerun loop
3420 progress = true;
3421 break;
3422 }
3423 }
3424
3425 }