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