1 /*
2 * Copyright (c) 1999, 2019, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "gc/shared/barrierSet.hpp"
27 #include "gc/shared/c2/barrierSetC2.hpp"
28 #include "memory/allocation.inline.hpp"
29 #include "memory/resourceArea.hpp"
30 #include "opto/addnode.hpp"
31 #include "opto/callnode.hpp"
32 #include "opto/castnode.hpp"
33 #include "opto/connode.hpp"
34 #include "opto/castnode.hpp"
35 #include "opto/divnode.hpp"
36 #include "opto/loopnode.hpp"
37 #include "opto/matcher.hpp"
38 #include "opto/mulnode.hpp"
39 #include "opto/movenode.hpp"
40 #include "opto/opaquenode.hpp"
41 #include "opto/rootnode.hpp"
42 #include "opto/subnode.hpp"
43 #include "utilities/macros.hpp"
44 #if INCLUDE_ZGC
45 #include "gc/z/c2/zBarrierSetC2.hpp"
46 #endif
47
48 //=============================================================================
49 //------------------------------split_thru_phi---------------------------------
50 // Split Node 'n' through merge point if there is enough win.
51 Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) {
52 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) {
53 // ConvI2L may have type information on it which is unsafe to push up
54 // so disable this for now
55 return NULL;
56 }
57
58 // Splitting range check CastIIs through a loop induction Phi can
59 // cause new Phis to be created that are left unrelated to the loop
60 // induction Phi and prevent optimizations (vectorization)
61 if (n->Opcode() == Op_CastII && n->as_CastII()->has_range_check() &&
62 region->is_CountedLoop() && n->in(1) == region->as_CountedLoop()->phi()) {
63 return NULL;
64 }
65
66 int wins = 0;
67 assert(!n->is_CFG(), "");
68 assert(region->is_Region(), "");
69
70 const Type* type = n->bottom_type();
71 const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr();
72 Node *phi;
73 if (t_oop != NULL && t_oop->is_known_instance_field()) {
74 int iid = t_oop->instance_id();
75 int index = C->get_alias_index(t_oop);
76 int offset = t_oop->offset();
77 phi = new PhiNode(region, type, NULL, iid, index, offset);
78 } else {
79 phi = PhiNode::make_blank(region, n);
80 }
81 uint old_unique = C->unique();
82 for (uint i = 1; i < region->req(); i++) {
83 Node *x;
84 Node* the_clone = NULL;
85 if (region->in(i) == C->top()) {
86 x = C->top(); // Dead path? Use a dead data op
87 } else {
88 x = n->clone(); // Else clone up the data op
89 the_clone = x; // Remember for possible deletion.
90 // Alter data node to use pre-phi inputs
91 if (n->in(0) == region)
92 x->set_req( 0, region->in(i) );
93 for (uint j = 1; j < n->req(); j++) {
94 Node *in = n->in(j);
95 if (in->is_Phi() && in->in(0) == region)
96 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone
97 }
98 }
99 // Check for a 'win' on some paths
100 const Type *t = x->Value(&_igvn);
101
102 bool singleton = t->singleton();
103
104 // A TOP singleton indicates that there are no possible values incoming
105 // along a particular edge. In most cases, this is OK, and the Phi will
106 // be eliminated later in an Ideal call. However, we can't allow this to
107 // happen if the singleton occurs on loop entry, as the elimination of
108 // the PhiNode may cause the resulting node to migrate back to a previous
109 // loop iteration.
110 if (singleton && t == Type::TOP) {
111 // Is_Loop() == false does not confirm the absence of a loop (e.g., an
112 // irreducible loop may not be indicated by an affirmative is_Loop());
113 // therefore, the only top we can split thru a phi is on a backedge of
114 // a loop.
115 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
116 }
117
118 if (singleton) {
119 wins++;
120 x = ((PhaseGVN&)_igvn).makecon(t);
121 } else {
122 // We now call Identity to try to simplify the cloned node.
123 // Note that some Identity methods call phase->type(this).
124 // Make sure that the type array is big enough for
125 // our new node, even though we may throw the node away.
126 // (Note: This tweaking with igvn only works because x is a new node.)
127 _igvn.set_type(x, t);
128 // If x is a TypeNode, capture any more-precise type permanently into Node
129 // otherwise it will be not updated during igvn->transform since
130 // igvn->type(x) is set to x->Value() already.
131 x->raise_bottom_type(t);
132 Node *y = _igvn.apply_identity(x);
133 if (y != x) {
134 wins++;
135 x = y;
136 } else {
137 y = _igvn.hash_find(x);
138 if (y) {
139 wins++;
140 x = y;
141 } else {
142 // Else x is a new node we are keeping
143 // We do not need register_new_node_with_optimizer
144 // because set_type has already been called.
145 _igvn._worklist.push(x);
146 }
147 }
148 }
149 if (x != the_clone && the_clone != NULL)
150 _igvn.remove_dead_node(the_clone);
151 phi->set_req( i, x );
152 }
153 // Too few wins?
154 if (wins <= policy) {
155 _igvn.remove_dead_node(phi);
156 return NULL;
157 }
158
159 // Record Phi
160 register_new_node( phi, region );
161
162 for (uint i2 = 1; i2 < phi->req(); i2++) {
163 Node *x = phi->in(i2);
164 // If we commoned up the cloned 'x' with another existing Node,
165 // the existing Node picks up a new use. We need to make the
166 // existing Node occur higher up so it dominates its uses.
167 Node *old_ctrl;
168 IdealLoopTree *old_loop;
169
170 if (x->is_Con()) {
171 // Constant's control is always root.
172 set_ctrl(x, C->root());
173 continue;
174 }
175 // The occasional new node
176 if (x->_idx >= old_unique) { // Found a new, unplaced node?
177 old_ctrl = NULL;
178 old_loop = NULL; // Not in any prior loop
179 } else {
180 old_ctrl = get_ctrl(x);
181 old_loop = get_loop(old_ctrl); // Get prior loop
182 }
183 // New late point must dominate new use
184 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2));
185 if (new_ctrl == old_ctrl) // Nothing is changed
186 continue;
187
188 IdealLoopTree *new_loop = get_loop(new_ctrl);
189
190 // Don't move x into a loop if its uses are
191 // outside of loop. Otherwise x will be cloned
192 // for each use outside of this loop.
193 IdealLoopTree *use_loop = get_loop(region);
194 if (!new_loop->is_member(use_loop) &&
195 (old_loop == NULL || !new_loop->is_member(old_loop))) {
196 // Take early control, later control will be recalculated
197 // during next iteration of loop optimizations.
198 new_ctrl = get_early_ctrl(x);
199 new_loop = get_loop(new_ctrl);
200 }
201 // Set new location
202 set_ctrl(x, new_ctrl);
203 // If changing loop bodies, see if we need to collect into new body
204 if (old_loop != new_loop) {
205 if (old_loop && !old_loop->_child)
206 old_loop->_body.yank(x);
207 if (!new_loop->_child)
208 new_loop->_body.push(x); // Collect body info
209 }
210 }
211
212 return phi;
213 }
214
215 //------------------------------dominated_by------------------------------------
216 // Replace the dominated test with an obvious true or false. Place it on the
217 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the
218 // live path up to the dominating control.
219 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff, bool flip, bool exclude_loop_predicate ) {
220 if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); }
221
222 // prevdom is the dominating projection of the dominating test.
223 assert( iff->is_If(), "" );
224 assert(iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd || iff->Opcode() == Op_RangeCheck, "Check this code when new subtype is added");
225 int pop = prevdom->Opcode();
226 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" );
227 if (flip) {
228 if (pop == Op_IfTrue)
229 pop = Op_IfFalse;
230 else
231 pop = Op_IfTrue;
232 }
233 // 'con' is set to true or false to kill the dominated test.
234 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO);
235 set_ctrl(con, C->root()); // Constant gets a new use
236 // Hack the dominated test
237 _igvn.replace_input_of(iff, 1, con);
238
239 // If I dont have a reachable TRUE and FALSE path following the IfNode then
240 // I can assume this path reaches an infinite loop. In this case it's not
241 // important to optimize the data Nodes - either the whole compilation will
242 // be tossed or this path (and all data Nodes) will go dead.
243 if (iff->outcnt() != 2) return;
244
245 // Make control-dependent data Nodes on the live path (path that will remain
246 // once the dominated IF is removed) become control-dependent on the
247 // dominating projection.
248 Node* dp = iff->as_If()->proj_out_or_null(pop == Op_IfTrue);
249
250 // Loop predicates may have depending checks which should not
251 // be skipped. For example, range check predicate has two checks
252 // for lower and upper bounds.
253 if (dp == NULL)
254 return;
255
256 ProjNode* dp_proj = dp->as_Proj();
257 ProjNode* unc_proj = iff->as_If()->proj_out(1 - dp_proj->_con)->as_Proj();
258 if (exclude_loop_predicate &&
259 (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) != NULL ||
260 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_profile_predicate) != NULL ||
261 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check) != NULL)) {
262 // If this is a range check (IfNode::is_range_check), do not
263 // reorder because Compile::allow_range_check_smearing might have
264 // changed the check.
265 return; // Let IGVN transformation change control dependence.
266 }
267
268 IdealLoopTree *old_loop = get_loop(dp);
269
270 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
271 Node* cd = dp->fast_out(i); // Control-dependent node
272 if (cd->depends_only_on_test()) {
273 assert(cd->in(0) == dp, "");
274 _igvn.replace_input_of(cd, 0, prevdom);
275 set_early_ctrl(cd);
276 IdealLoopTree *new_loop = get_loop(get_ctrl(cd));
277 if (old_loop != new_loop) {
278 if (!old_loop->_child) old_loop->_body.yank(cd);
279 if (!new_loop->_child) new_loop->_body.push(cd);
280 }
281 --i;
282 --imax;
283 }
284 }
285 }
286
287 //------------------------------has_local_phi_input----------------------------
288 // Return TRUE if 'n' has Phi inputs from its local block and no other
289 // block-local inputs (all non-local-phi inputs come from earlier blocks)
290 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) {
291 Node *n_ctrl = get_ctrl(n);
292 // See if some inputs come from a Phi in this block, or from before
293 // this block.
294 uint i;
295 for( i = 1; i < n->req(); i++ ) {
296 Node *phi = n->in(i);
297 if( phi->is_Phi() && phi->in(0) == n_ctrl )
298 break;
299 }
300 if( i >= n->req() )
301 return NULL; // No Phi inputs; nowhere to clone thru
302
303 // Check for inputs created between 'n' and the Phi input. These
304 // must split as well; they have already been given the chance
305 // (courtesy of a post-order visit) and since they did not we must
306 // recover the 'cost' of splitting them by being very profitable
307 // when splitting 'n'. Since this is unlikely we simply give up.
308 for( i = 1; i < n->req(); i++ ) {
309 Node *m = n->in(i);
310 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) {
311 // We allow the special case of AddP's with no local inputs.
312 // This allows us to split-up address expressions.
313 if (m->is_AddP() &&
314 get_ctrl(m->in(2)) != n_ctrl &&
315 get_ctrl(m->in(3)) != n_ctrl) {
316 // Move the AddP up to dominating point
317 Node* c = find_non_split_ctrl(idom(n_ctrl));
318 if (c->is_OuterStripMinedLoop()) {
319 c->as_Loop()->verify_strip_mined(1);
320 c = c->in(LoopNode::EntryControl);
321 }
322 set_ctrl_and_loop(m, c);
323 continue;
324 }
325 return NULL;
326 }
327 assert(m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control");
328 }
329
330 return n_ctrl;
331 }
332
333 //------------------------------remix_address_expressions----------------------
334 // Rework addressing expressions to get the most loop-invariant stuff
335 // moved out. We'd like to do all associative operators, but it's especially
336 // important (common) to do address expressions.
337 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) {
338 if (!has_ctrl(n)) return NULL;
339 Node *n_ctrl = get_ctrl(n);
340 IdealLoopTree *n_loop = get_loop(n_ctrl);
341
342 // See if 'n' mixes loop-varying and loop-invariant inputs and
343 // itself is loop-varying.
344
345 // Only interested in binary ops (and AddP)
346 if( n->req() < 3 || n->req() > 4 ) return NULL;
347
348 Node *n1_ctrl = get_ctrl(n->in( 1));
349 Node *n2_ctrl = get_ctrl(n->in( 2));
350 Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3));
351 IdealLoopTree *n1_loop = get_loop( n1_ctrl );
352 IdealLoopTree *n2_loop = get_loop( n2_ctrl );
353 IdealLoopTree *n3_loop = get_loop( n3_ctrl );
354
355 // Does one of my inputs spin in a tighter loop than self?
356 if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) ||
357 (n_loop->is_member( n2_loop ) && n_loop != n2_loop) ||
358 (n_loop->is_member( n3_loop ) && n_loop != n3_loop) )
359 return NULL; // Leave well enough alone
360
361 // Is at least one of my inputs loop-invariant?
362 if( n1_loop == n_loop &&
363 n2_loop == n_loop &&
364 n3_loop == n_loop )
365 return NULL; // No loop-invariant inputs
366
367
368 int n_op = n->Opcode();
369
370 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2).
371 if( n_op == Op_LShiftI ) {
372 // Scale is loop invariant
373 Node *scale = n->in(2);
374 Node *scale_ctrl = get_ctrl(scale);
375 IdealLoopTree *scale_loop = get_loop(scale_ctrl );
376 if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) )
377 return NULL;
378 const TypeInt *scale_t = scale->bottom_type()->isa_int();
379 if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 )
380 return NULL; // Dont bother with byte/short masking
381 // Add must vary with loop (else shift would be loop-invariant)
382 Node *add = n->in(1);
383 Node *add_ctrl = get_ctrl(add);
384 IdealLoopTree *add_loop = get_loop(add_ctrl);
385 //assert( n_loop == add_loop, "" );
386 if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops
387
388 // Convert I-V into I+ (0-V); same for V-I
389 if( add->Opcode() == Op_SubI &&
390 _igvn.type( add->in(1) ) != TypeInt::ZERO ) {
391 Node *zero = _igvn.intcon(0);
392 set_ctrl(zero, C->root());
393 Node *neg = new SubINode( _igvn.intcon(0), add->in(2) );
394 register_new_node( neg, get_ctrl(add->in(2) ) );
395 add = new AddINode( add->in(1), neg );
396 register_new_node( add, add_ctrl );
397 }
398 if( add->Opcode() != Op_AddI ) return NULL;
399 // See if one add input is loop invariant
400 Node *add_var = add->in(1);
401 Node *add_var_ctrl = get_ctrl(add_var);
402 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
403 Node *add_invar = add->in(2);
404 Node *add_invar_ctrl = get_ctrl(add_invar);
405 IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl );
406 if( add_var_loop == n_loop ) {
407 } else if( add_invar_loop == n_loop ) {
408 // Swap to find the invariant part
409 add_invar = add_var;
410 add_invar_ctrl = add_var_ctrl;
411 add_invar_loop = add_var_loop;
412 add_var = add->in(2);
413 Node *add_var_ctrl = get_ctrl(add_var);
414 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
415 } else // Else neither input is loop invariant
416 return NULL;
417 if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) )
418 return NULL; // No invariant part of the add?
419
420 // Yes! Reshape address expression!
421 Node *inv_scale = new LShiftINode( add_invar, scale );
422 Node *inv_scale_ctrl =
423 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ?
424 add_invar_ctrl : scale_ctrl;
425 register_new_node( inv_scale, inv_scale_ctrl );
426 Node *var_scale = new LShiftINode( add_var, scale );
427 register_new_node( var_scale, n_ctrl );
428 Node *var_add = new AddINode( var_scale, inv_scale );
429 register_new_node( var_add, n_ctrl );
430 _igvn.replace_node( n, var_add );
431 return var_add;
432 }
433
434 // Replace (I+V) with (V+I)
435 if( n_op == Op_AddI ||
436 n_op == Op_AddL ||
437 n_op == Op_AddF ||
438 n_op == Op_AddD ||
439 n_op == Op_MulI ||
440 n_op == Op_MulL ||
441 n_op == Op_MulF ||
442 n_op == Op_MulD ) {
443 if( n2_loop == n_loop ) {
444 assert( n1_loop != n_loop, "" );
445 n->swap_edges(1, 2);
446 }
447 }
448
449 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V),
450 // but not if I2 is a constant.
451 if( n_op == Op_AddP ) {
452 if( n2_loop == n_loop && n3_loop != n_loop ) {
453 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) {
454 Node *n22_ctrl = get_ctrl(n->in(2)->in(2));
455 Node *n23_ctrl = get_ctrl(n->in(2)->in(3));
456 IdealLoopTree *n22loop = get_loop( n22_ctrl );
457 IdealLoopTree *n23_loop = get_loop( n23_ctrl );
458 if( n22loop != n_loop && n22loop->is_member(n_loop) &&
459 n23_loop == n_loop ) {
460 Node *add1 = new AddPNode( n->in(1), n->in(2)->in(2), n->in(3) );
461 // Stuff new AddP in the loop preheader
462 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) );
463 Node *add2 = new AddPNode( n->in(1), add1, n->in(2)->in(3) );
464 register_new_node( add2, n_ctrl );
465 _igvn.replace_node( n, add2 );
466 return add2;
467 }
468 }
469 }
470
471 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V)
472 if (n2_loop != n_loop && n3_loop == n_loop) {
473 if (n->in(3)->Opcode() == Op_AddX) {
474 Node *V = n->in(3)->in(1);
475 Node *I = n->in(3)->in(2);
476 if (is_member(n_loop,get_ctrl(V))) {
477 } else {
478 Node *tmp = V; V = I; I = tmp;
479 }
480 if (!is_member(n_loop,get_ctrl(I))) {
481 Node *add1 = new AddPNode(n->in(1), n->in(2), I);
482 // Stuff new AddP in the loop preheader
483 register_new_node(add1, n_loop->_head->in(LoopNode::EntryControl));
484 Node *add2 = new AddPNode(n->in(1), add1, V);
485 register_new_node(add2, n_ctrl);
486 _igvn.replace_node(n, add2);
487 return add2;
488 }
489 }
490 }
491 }
492
493 return NULL;
494 }
495
496 // Optimize ((in1[2*i] * in2[2*i]) + (in1[2*i+1] * in2[2*i+1]))
497 Node *PhaseIdealLoop::convert_add_to_muladd(Node* n) {
498 assert(n->Opcode() == Op_AddI, "sanity");
499 Node * nn = NULL;
500 Node * in1 = n->in(1);
501 Node * in2 = n->in(2);
502 if (in1->Opcode() == Op_MulI && in2->Opcode() == Op_MulI) {
503 IdealLoopTree* loop_n = get_loop(get_ctrl(n));
504 if (loop_n->_head->as_Loop()->is_valid_counted_loop() &&
505 Matcher::match_rule_supported(Op_MulAddS2I) &&
506 Matcher::match_rule_supported(Op_MulAddVS2VI)) {
507 Node* mul_in1 = in1->in(1);
508 Node* mul_in2 = in1->in(2);
509 Node* mul_in3 = in2->in(1);
510 Node* mul_in4 = in2->in(2);
511 if (mul_in1->Opcode() == Op_LoadS &&
512 mul_in2->Opcode() == Op_LoadS &&
513 mul_in3->Opcode() == Op_LoadS &&
514 mul_in4->Opcode() == Op_LoadS) {
515 IdealLoopTree* loop1 = get_loop(get_ctrl(mul_in1));
516 IdealLoopTree* loop2 = get_loop(get_ctrl(mul_in2));
517 IdealLoopTree* loop3 = get_loop(get_ctrl(mul_in3));
518 IdealLoopTree* loop4 = get_loop(get_ctrl(mul_in4));
519 IdealLoopTree* loop5 = get_loop(get_ctrl(in1));
520 IdealLoopTree* loop6 = get_loop(get_ctrl(in2));
521 // All nodes should be in the same counted loop.
522 if (loop_n == loop1 && loop_n == loop2 && loop_n == loop3 &&
523 loop_n == loop4 && loop_n == loop5 && loop_n == loop6) {
524 Node* adr1 = mul_in1->in(MemNode::Address);
525 Node* adr2 = mul_in2->in(MemNode::Address);
526 Node* adr3 = mul_in3->in(MemNode::Address);
527 Node* adr4 = mul_in4->in(MemNode::Address);
528 if (adr1->is_AddP() && adr2->is_AddP() && adr3->is_AddP() && adr4->is_AddP()) {
529 if ((adr1->in(AddPNode::Base) == adr3->in(AddPNode::Base)) &&
530 (adr2->in(AddPNode::Base) == adr4->in(AddPNode::Base))) {
531 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in3, mul_in4);
532 register_new_node(nn, get_ctrl(n));
533 _igvn.replace_node(n, nn);
534 return nn;
535 } else if ((adr1->in(AddPNode::Base) == adr4->in(AddPNode::Base)) &&
536 (adr2->in(AddPNode::Base) == adr3->in(AddPNode::Base))) {
537 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in4, mul_in3);
538 register_new_node(nn, get_ctrl(n));
539 _igvn.replace_node(n, nn);
540 return nn;
541 }
542 }
543 }
544 }
545 }
546 }
547 return nn;
548 }
549
550 //------------------------------conditional_move-------------------------------
551 // Attempt to replace a Phi with a conditional move. We have some pretty
552 // strict profitability requirements. All Phis at the merge point must
553 // be converted, so we can remove the control flow. We need to limit the
554 // number of c-moves to a small handful. All code that was in the side-arms
555 // of the CFG diamond is now speculatively executed. This code has to be
556 // "cheap enough". We are pretty much limited to CFG diamonds that merge
557 // 1 or 2 items with a total of 1 or 2 ops executed speculatively.
558 Node *PhaseIdealLoop::conditional_move( Node *region ) {
559
560 assert(region->is_Region(), "sanity check");
561 if (region->req() != 3) return NULL;
562
563 // Check for CFG diamond
564 Node *lp = region->in(1);
565 Node *rp = region->in(2);
566 if (!lp || !rp) return NULL;
567 Node *lp_c = lp->in(0);
568 if (lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If()) return NULL;
569 IfNode *iff = lp_c->as_If();
570
571 // Check for ops pinned in an arm of the diamond.
572 // Can't remove the control flow in this case
573 if (lp->outcnt() > 1) return NULL;
574 if (rp->outcnt() > 1) return NULL;
575
576 IdealLoopTree* r_loop = get_loop(region);
577 assert(r_loop == get_loop(iff), "sanity");
578 // Always convert to CMOVE if all results are used only outside this loop.
579 bool used_inside_loop = (r_loop == _ltree_root);
580
581 // Check profitability
582 int cost = 0;
583 int phis = 0;
584 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
585 Node *out = region->fast_out(i);
586 if (!out->is_Phi()) continue; // Ignore other control edges, etc
587 phis++;
588 PhiNode* phi = out->as_Phi();
589 BasicType bt = phi->type()->basic_type();
590 switch (bt) {
591 case T_DOUBLE:
592 case T_FLOAT:
593 if (C->use_cmove()) {
594 continue; //TODO: maybe we want to add some cost
595 }
596 cost += Matcher::float_cmove_cost(); // Could be very expensive
597 break;
598 case T_LONG: {
599 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's
600 }
601 case T_INT: // These all CMOV fine
602 case T_ADDRESS: { // (RawPtr)
603 cost++;
604 break;
605 }
606 case T_NARROWOOP: // Fall through
607 case T_OBJECT: { // Base oops are OK, but not derived oops
608 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr();
609 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a
610 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus
611 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we
612 // have a Phi for the base here that we convert to a CMOVE all is well
613 // and good. But if the base is dead, we'll not make a CMOVE. Later
614 // the allocator will have to produce a base by creating a CMOVE of the
615 // relevant bases. This puts the allocator in the business of
616 // manufacturing expensive instructions, generally a bad plan.
617 // Just Say No to Conditionally-Moved Derived Pointers.
618 if (tp && tp->offset() != 0)
619 return NULL;
620 cost++;
621 break;
622 }
623 default:
624 return NULL; // In particular, can't do memory or I/O
625 }
626 // Add in cost any speculative ops
627 for (uint j = 1; j < region->req(); j++) {
628 Node *proj = region->in(j);
629 Node *inp = phi->in(j);
630 if (get_ctrl(inp) == proj) { // Found local op
631 cost++;
632 // Check for a chain of dependent ops; these will all become
633 // speculative in a CMOV.
634 for (uint k = 1; k < inp->req(); k++)
635 if (get_ctrl(inp->in(k)) == proj)
636 cost += ConditionalMoveLimit; // Too much speculative goo
637 }
638 }
639 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode.
640 // This will likely Split-If, a higher-payoff operation.
641 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) {
642 Node* use = phi->fast_out(k);
643 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr())
644 cost += ConditionalMoveLimit;
645 // Is there a use inside the loop?
646 // Note: check only basic types since CMoveP is pinned.
647 if (!used_inside_loop && is_java_primitive(bt)) {
648 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use);
649 if (r_loop == u_loop || r_loop->is_member(u_loop)) {
650 used_inside_loop = true;
651 }
652 }
653 }
654 }//for
655 Node* bol = iff->in(1);
656 assert(bol->Opcode() == Op_Bool, "");
657 int cmp_op = bol->in(1)->Opcode();
658 // It is expensive to generate flags from a float compare.
659 // Avoid duplicated float compare.
660 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL;
661
662 float infrequent_prob = PROB_UNLIKELY_MAG(3);
663 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop.
664 if (used_inside_loop) {
665 if (cost >= ConditionalMoveLimit) return NULL; // Too much goo
666
667 // BlockLayoutByFrequency optimization moves infrequent branch
668 // from hot path. No point in CMOV'ing in such case (110 is used
669 // instead of 100 to take into account not exactness of float value).
670 if (BlockLayoutByFrequency) {
671 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f);
672 }
673 }
674 // Check for highly predictable branch. No point in CMOV'ing if
675 // we are going to predict accurately all the time.
676 if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) {
677 //keep going
678 } else if (iff->_prob < infrequent_prob ||
679 iff->_prob > (1.0f - infrequent_prob))
680 return NULL;
681
682 // --------------
683 // Now replace all Phis with CMOV's
684 Node *cmov_ctrl = iff->in(0);
685 uint flip = (lp->Opcode() == Op_IfTrue);
686 Node_List wq;
687 while (1) {
688 PhiNode* phi = NULL;
689 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
690 Node *out = region->fast_out(i);
691 if (out->is_Phi()) {
692 phi = out->as_Phi();
693 break;
694 }
695 }
696 if (phi == NULL) break;
697 if (PrintOpto && VerifyLoopOptimizations) { tty->print_cr("CMOV"); }
698 // Move speculative ops
699 wq.push(phi);
700 while (wq.size() > 0) {
701 Node *n = wq.pop();
702 for (uint j = 1; j < n->req(); j++) {
703 Node* m = n->in(j);
704 if (m != NULL && !is_dominator(get_ctrl(m), cmov_ctrl)) {
705 #ifndef PRODUCT
706 if (PrintOpto && VerifyLoopOptimizations) {
707 tty->print(" speculate: ");
708 m->dump();
709 }
710 #endif
711 set_ctrl(m, cmov_ctrl);
712 wq.push(m);
713 }
714 }
715 }
716 Node *cmov = CMoveNode::make(cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi));
717 register_new_node( cmov, cmov_ctrl );
718 _igvn.replace_node( phi, cmov );
719 #ifndef PRODUCT
720 if (TraceLoopOpts) {
721 tty->print("CMOV ");
722 r_loop->dump_head();
723 if (Verbose) {
724 bol->in(1)->dump(1);
725 cmov->dump(1);
726 }
727 }
728 if (VerifyLoopOptimizations) verify();
729 #endif
730 }
731
732 // The useless CFG diamond will fold up later; see the optimization in
733 // RegionNode::Ideal.
734 _igvn._worklist.push(region);
735
736 return iff->in(1);
737 }
738
739 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) {
740 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
741 Node* u = m->fast_out(i);
742 if (u->is_CFG()) {
743 if (u->Opcode() == Op_NeverBranch) {
744 u = ((NeverBranchNode*)u)->proj_out(0);
745 enqueue_cfg_uses(u, wq);
746 } else {
747 wq.push(u);
748 }
749 }
750 }
751 }
752
753 // Try moving a store out of a loop, right before the loop
754 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) {
755 // Store has to be first in the loop body
756 IdealLoopTree *n_loop = get_loop(n_ctrl);
757 if (n->is_Store() && n_loop != _ltree_root &&
758 n_loop->is_loop() && n_loop->_head->is_Loop() &&
759 n->in(0) != NULL) {
760 Node* address = n->in(MemNode::Address);
761 Node* value = n->in(MemNode::ValueIn);
762 Node* mem = n->in(MemNode::Memory);
763 IdealLoopTree* address_loop = get_loop(get_ctrl(address));
764 IdealLoopTree* value_loop = get_loop(get_ctrl(value));
765
766 // - address and value must be loop invariant
767 // - memory must be a memory Phi for the loop
768 // - Store must be the only store on this memory slice in the
769 // loop: if there's another store following this one then value
770 // written at iteration i by the second store could be overwritten
771 // at iteration i+n by the first store: it's not safe to move the
772 // first store out of the loop
773 // - nothing must observe the memory Phi: it guarantees no read
774 // before the store, we are also guaranteed the store post
775 // dominates the loop head (ignoring a possible early
776 // exit). Otherwise there would be extra Phi involved between the
777 // loop's Phi and the store.
778 // - there must be no early exit from the loop before the Store
779 // (such an exit most of the time would be an extra use of the
780 // memory Phi but sometimes is a bottom memory Phi that takes the
781 // store as input).
782
783 if (!n_loop->is_member(address_loop) &&
784 !n_loop->is_member(value_loop) &&
785 mem->is_Phi() && mem->in(0) == n_loop->_head &&
786 mem->outcnt() == 1 &&
787 mem->in(LoopNode::LoopBackControl) == n) {
788
789 assert(n_loop->_tail != NULL, "need a tail");
790 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop");
791
792 // Verify that there's no early exit of the loop before the store.
793 bool ctrl_ok = false;
794 {
795 // Follow control from loop head until n, we exit the loop or
796 // we reach the tail
797 ResourceMark rm;
798 Unique_Node_List wq;
799 wq.push(n_loop->_head);
800
801 for (uint next = 0; next < wq.size(); ++next) {
802 Node *m = wq.at(next);
803 if (m == n->in(0)) {
804 ctrl_ok = true;
805 continue;
806 }
807 assert(!has_ctrl(m), "should be CFG");
808 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) {
809 ctrl_ok = false;
810 break;
811 }
812 enqueue_cfg_uses(m, wq);
813 if (wq.size() > 10) {
814 ctrl_ok = false;
815 break;
816 }
817 }
818 }
819 if (ctrl_ok) {
820 // move the Store
821 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem);
822 _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl));
823 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl));
824 // Disconnect the phi now. An empty phi can confuse other
825 // optimizations in this pass of loop opts.
826 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl));
827 n_loop->_body.yank(mem);
828
829 set_ctrl_and_loop(n, n->in(0));
830
831 return n;
832 }
833 }
834 }
835 return NULL;
836 }
837
838 // Try moving a store out of a loop, right after the loop
839 void PhaseIdealLoop::try_move_store_after_loop(Node* n) {
840 if (n->is_Store() && n->in(0) != NULL) {
841 Node *n_ctrl = get_ctrl(n);
842 IdealLoopTree *n_loop = get_loop(n_ctrl);
843 // Store must be in a loop
844 if (n_loop != _ltree_root && !n_loop->_irreducible) {
845 Node* address = n->in(MemNode::Address);
846 Node* value = n->in(MemNode::ValueIn);
847 IdealLoopTree* address_loop = get_loop(get_ctrl(address));
848 // address must be loop invariant
849 if (!n_loop->is_member(address_loop)) {
850 // Store must be last on this memory slice in the loop and
851 // nothing in the loop must observe it
852 Node* phi = NULL;
853 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
854 Node* u = n->fast_out(i);
855 if (has_ctrl(u)) { // control use?
856 IdealLoopTree *u_loop = get_loop(get_ctrl(u));
857 if (!n_loop->is_member(u_loop)) {
858 continue;
859 }
860 if (u->is_Phi() && u->in(0) == n_loop->_head) {
861 assert(_igvn.type(u) == Type::MEMORY, "bad phi");
862 // multiple phis on the same slice are possible
863 if (phi != NULL) {
864 return;
865 }
866 phi = u;
867 continue;
868 }
869 }
870 return;
871 }
872 if (phi != NULL) {
873 // Nothing in the loop before the store (next iteration)
874 // must observe the stored value
875 bool mem_ok = true;
876 {
877 ResourceMark rm;
878 Unique_Node_List wq;
879 wq.push(phi);
880 for (uint next = 0; next < wq.size() && mem_ok; ++next) {
881 Node *m = wq.at(next);
882 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) {
883 Node* u = m->fast_out(i);
884 if (u->is_Store() || u->is_Phi()) {
885 if (u != n) {
886 wq.push(u);
887 mem_ok = (wq.size() <= 10);
888 }
889 } else {
890 mem_ok = false;
891 break;
892 }
893 }
894 }
895 }
896 if (mem_ok) {
897 // Move the store out of the loop if the LCA of all
898 // users (except for the phi) is outside the loop.
899 Node* hook = new Node(1);
900 _igvn.rehash_node_delayed(phi);
901 int count = phi->replace_edge(n, hook);
902 assert(count > 0, "inconsistent phi");
903
904 // Compute latest point this store can go
905 Node* lca = get_late_ctrl(n, get_ctrl(n));
906 if (n_loop->is_member(get_loop(lca))) {
907 // LCA is in the loop - bail out
908 _igvn.replace_node(hook, n);
909 return;
910 }
911 #ifdef ASSERT
912 if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) {
913 assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined");
914 n_loop->_head->as_Loop()->verify_strip_mined(1);
915 Node* outer = n_loop->_head->as_CountedLoop()->outer_loop();
916 IdealLoopTree* outer_loop = get_loop(outer);
917 assert(n_loop->_parent == outer_loop, "broken loop tree");
918 assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state");
919 }
920 #endif
921
922 // Move store out of the loop
923 _igvn.replace_node(hook, n->in(MemNode::Memory));
924 _igvn.replace_input_of(n, 0, lca);
925 set_ctrl_and_loop(n, lca);
926
927 // Disconnect the phi now. An empty phi can confuse other
928 // optimizations in this pass of loop opts..
929 if (phi->in(LoopNode::LoopBackControl) == phi) {
930 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl));
931 n_loop->_body.yank(phi);
932 }
933 }
934 }
935 }
936 }
937 }
938 }
939
940 //------------------------------split_if_with_blocks_pre-----------------------
941 // Do the real work in a non-recursive function. Data nodes want to be
942 // cloned in the pre-order so they can feed each other nicely.
943 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
944 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
945 Node* bs_res = bs->split_if_pre(this, n);
946 if (bs_res != NULL) {
947 return bs_res;
948 }
949 // Cloning these guys is unlikely to win
950 int n_op = n->Opcode();
951 if( n_op == Op_MergeMem ) return n;
952 if( n->is_Proj() ) return n;
953 // Do not clone-up CmpFXXX variations, as these are always
954 // followed by a CmpI
955 if( n->is_Cmp() ) return n;
956 // Attempt to use a conditional move instead of a phi/branch
957 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) {
958 Node *cmov = conditional_move( n );
959 if( cmov ) return cmov;
960 }
961 if( n->is_CFG() || n->is_LoadStore() )
962 return n;
963 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd
964 n_op == Op_Opaque2 ) {
965 if( !C->major_progress() ) // If chance of no more loop opts...
966 _igvn._worklist.push(n); // maybe we'll remove them
967 return n;
968 }
969
970 if( n->is_Con() ) return n; // No cloning for Con nodes
971
972 Node *n_ctrl = get_ctrl(n);
973 if( !n_ctrl ) return n; // Dead node
974
975 Node* res = try_move_store_before_loop(n, n_ctrl);
976 if (res != NULL) {
977 return n;
978 }
979
980 // Attempt to remix address expressions for loop invariants
981 Node *m = remix_address_expressions( n );
982 if( m ) return m;
983
984 if (n_op == Op_AddI) {
985 Node *nn = convert_add_to_muladd( n );
986 if ( nn ) return nn;
987 }
988
989 if (n->is_ConstraintCast()) {
990 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this);
991 // ConstraintCastNode::dominating_cast() uses node control input to determine domination.
992 // Node control inputs don't necessarily agree with loop control info (due to
993 // transformations happened in between), thus additional dominance check is needed
994 // to keep loop info valid.
995 if (dom_cast != NULL && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) {
996 _igvn.replace_node(n, dom_cast);
997 return dom_cast;
998 }
999 }
1000
1001 // Determine if the Node has inputs from some local Phi.
1002 // Returns the block to clone thru.
1003 Node *n_blk = has_local_phi_input( n );
1004 if( !n_blk ) return n;
1005
1006 // Do not clone the trip counter through on a CountedLoop
1007 // (messes up the canonical shape).
1008 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n;
1009
1010 // Check for having no control input; not pinned. Allow
1011 // dominating control.
1012 if (n->in(0)) {
1013 Node *dom = idom(n_blk);
1014 if (dom_lca(n->in(0), dom) != n->in(0)) {
1015 return n;
1016 }
1017 }
1018 // Policy: when is it profitable. You must get more wins than
1019 // policy before it is considered profitable. Policy is usually 0,
1020 // so 1 win is considered profitable. Big merges will require big
1021 // cloning, so get a larger policy.
1022 int policy = n_blk->req() >> 2;
1023
1024 // If the loop is a candidate for range check elimination,
1025 // delay splitting through it's phi until a later loop optimization
1026 if (n_blk->is_CountedLoop()) {
1027 IdealLoopTree *lp = get_loop(n_blk);
1028 if (lp && lp->_rce_candidate) {
1029 return n;
1030 }
1031 }
1032
1033 if (must_throttle_split_if()) return n;
1034
1035 // Split 'n' through the merge point if it is profitable
1036 Node *phi = split_thru_phi( n, n_blk, policy );
1037 if (!phi) return n;
1038
1039 // Found a Phi to split thru!
1040 // Replace 'n' with the new phi
1041 _igvn.replace_node( n, phi );
1042 // Moved a load around the loop, 'en-registering' something.
1043 if (n_blk->is_Loop() && n->is_Load() &&
1044 !phi->in(LoopNode::LoopBackControl)->is_Load())
1045 C->set_major_progress();
1046
1047 return phi;
1048 }
1049
1050 static bool merge_point_too_heavy(Compile* C, Node* region) {
1051 // Bail out if the region and its phis have too many users.
1052 int weight = 0;
1053 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1054 weight += region->fast_out(i)->outcnt();
1055 }
1056 int nodes_left = C->max_node_limit() - C->live_nodes();
1057 if (weight * 8 > nodes_left) {
1058 if (PrintOpto) {
1059 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight);
1060 }
1061 return true;
1062 } else {
1063 return false;
1064 }
1065 }
1066
1067 static bool merge_point_safe(Node* region) {
1068 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode
1069 // having a PhiNode input. This sidesteps the dangerous case where the split
1070 // ConvI2LNode may become TOP if the input Value() does not
1071 // overlap the ConvI2L range, leaving a node which may not dominate its
1072 // uses.
1073 // A better fix for this problem can be found in the BugTraq entry, but
1074 // expediency for Mantis demands this hack.
1075 // 6855164: If the merge point has a FastLockNode with a PhiNode input, we stop
1076 // split_if_with_blocks from splitting a block because we could not move around
1077 // the FastLockNode.
1078 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1079 Node* n = region->fast_out(i);
1080 if (n->is_Phi()) {
1081 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1082 Node* m = n->fast_out(j);
1083 if (m->is_FastLock())
1084 return false;
1085 #ifdef _LP64
1086 if (m->Opcode() == Op_ConvI2L)
1087 return false;
1088 if (m->is_CastII() && m->isa_CastII()->has_range_check()) {
1089 return false;
1090 }
1091 #endif
1092 }
1093 }
1094 }
1095 return true;
1096 }
1097
1098
1099 //------------------------------place_near_use---------------------------------
1100 // Place some computation next to use but not inside inner loops.
1101 // For inner loop uses move it to the preheader area.
1102 Node *PhaseIdealLoop::place_near_use(Node *useblock) const {
1103 IdealLoopTree *u_loop = get_loop( useblock );
1104 if (u_loop->_irreducible) {
1105 return useblock;
1106 }
1107 if (u_loop->_child) {
1108 if (useblock == u_loop->_head && u_loop->_head->is_OuterStripMinedLoop()) {
1109 return u_loop->_head->in(LoopNode::EntryControl);
1110 }
1111 return useblock;
1112 }
1113 return u_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl);
1114 }
1115
1116
1117 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) {
1118 if (!n->is_If() || n->is_CountedLoopEnd()) {
1119 return false;
1120 }
1121 if (!n->in(0)->is_Region()) {
1122 return false;
1123 }
1124 Node* region = n->in(0);
1125 Node* dom = idom(region);
1126 if (!dom->is_If() || dom->in(1) != n->in(1)) {
1127 return false;
1128 }
1129 IfNode* dom_if = dom->as_If();
1130 Node* proj_true = dom_if->proj_out(1);
1131 Node* proj_false = dom_if->proj_out(0);
1132
1133 for (uint i = 1; i < region->req(); i++) {
1134 if (is_dominator(proj_true, region->in(i))) {
1135 continue;
1136 }
1137 if (is_dominator(proj_false, region->in(i))) {
1138 continue;
1139 }
1140 return false;
1141 }
1142
1143 return true;
1144 }
1145
1146
1147 bool PhaseIdealLoop::can_split_if(Node* n_ctrl) {
1148 if (must_throttle_split_if()) {
1149 return false;
1150 }
1151
1152 // Do not do 'split-if' if irreducible loops are present.
1153 if (_has_irreducible_loops) {
1154 return false;
1155 }
1156
1157 if (merge_point_too_heavy(C, n_ctrl)) {
1158 return false;
1159 }
1160
1161 // Do not do 'split-if' if some paths are dead. First do dead code
1162 // elimination and then see if its still profitable.
1163 for (uint i = 1; i < n_ctrl->req(); i++) {
1164 if (n_ctrl->in(i) == C->top()) {
1165 return false;
1166 }
1167 }
1168
1169 // If trying to do a 'Split-If' at the loop head, it is only
1170 // profitable if the cmp folds up on BOTH paths. Otherwise we
1171 // risk peeling a loop forever.
1172
1173 // CNC - Disabled for now. Requires careful handling of loop
1174 // body selection for the cloned code. Also, make sure we check
1175 // for any input path not being in the same loop as n_ctrl. For
1176 // irreducible loops we cannot check for 'n_ctrl->is_Loop()'
1177 // because the alternative loop entry points won't be converted
1178 // into LoopNodes.
1179 IdealLoopTree *n_loop = get_loop(n_ctrl);
1180 for (uint j = 1; j < n_ctrl->req(); j++) {
1181 if (get_loop(n_ctrl->in(j)) != n_loop) {
1182 return false;
1183 }
1184 }
1185
1186 // Check for safety of the merge point.
1187 if (!merge_point_safe(n_ctrl)) {
1188 return false;
1189 }
1190
1191 return true;
1192 }
1193
1194 //------------------------------split_if_with_blocks_post----------------------
1195 // Do the real work in a non-recursive function. CFG hackery wants to be
1196 // in the post-order, so it can dirty the I-DOM info and not use the dirtied
1197 // info.
1198 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) {
1199
1200 // Cloning Cmp through Phi's involves the split-if transform.
1201 // FastLock is not used by an If
1202 if (n->is_Cmp() && !n->is_FastLock()) {
1203 Node *n_ctrl = get_ctrl(n);
1204 // Determine if the Node has inputs from some local Phi.
1205 // Returns the block to clone thru.
1206 Node *n_blk = has_local_phi_input(n);
1207 if (n_blk != n_ctrl) {
1208 return;
1209 }
1210
1211 if (!can_split_if(n_ctrl)) {
1212 return;
1213 }
1214
1215 if (n->outcnt() != 1) {
1216 return; // Multiple bool's from 1 compare?
1217 }
1218 Node *bol = n->unique_out();
1219 assert(bol->is_Bool(), "expect a bool here");
1220 if (bol->outcnt() != 1) {
1221 return;// Multiple branches from 1 compare?
1222 }
1223 Node *iff = bol->unique_out();
1224
1225 // Check some safety conditions
1226 if (iff->is_If()) { // Classic split-if?
1227 if (iff->in(0) != n_ctrl) {
1228 return; // Compare must be in same blk as if
1229 }
1230 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE
1231 // Can't split CMove with different control edge.
1232 if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) {
1233 return;
1234 }
1235 if (get_ctrl(iff->in(2)) == n_ctrl ||
1236 get_ctrl(iff->in(3)) == n_ctrl) {
1237 return; // Inputs not yet split-up
1238 }
1239 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) {
1240 return; // Loop-invar test gates loop-varying CMOVE
1241 }
1242 } else {
1243 return; // some other kind of node, such as an Allocate
1244 }
1245
1246 // When is split-if profitable? Every 'win' on means some control flow
1247 // goes dead, so it's almost always a win.
1248 int policy = 0;
1249 // Split compare 'n' through the merge point if it is profitable
1250 Node *phi = split_thru_phi( n, n_ctrl, policy);
1251 if (!phi) {
1252 return;
1253 }
1254
1255 // Found a Phi to split thru!
1256 // Replace 'n' with the new phi
1257 _igvn.replace_node(n, phi);
1258
1259 // Now split the bool up thru the phi
1260 Node *bolphi = split_thru_phi(bol, n_ctrl, -1);
1261 guarantee(bolphi != NULL, "null boolean phi node");
1262
1263 _igvn.replace_node(bol, bolphi);
1264 assert(iff->in(1) == bolphi, "");
1265
1266 if (bolphi->Value(&_igvn)->singleton()) {
1267 return;
1268 }
1269
1270 // Conditional-move? Must split up now
1271 if (!iff->is_If()) {
1272 Node *cmovphi = split_thru_phi(iff, n_ctrl, -1);
1273 _igvn.replace_node(iff, cmovphi);
1274 return;
1275 }
1276
1277 // Now split the IF
1278 do_split_if(iff);
1279 return;
1280 }
1281
1282 // Two identical ifs back to back can be merged
1283 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) {
1284 Node *n_ctrl = n->in(0);
1285 PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1));
1286 IfNode* dom_if = idom(n_ctrl)->as_If();
1287 Node* proj_true = dom_if->proj_out(1);
1288 Node* proj_false = dom_if->proj_out(0);
1289 Node* con_true = _igvn.makecon(TypeInt::ONE);
1290 Node* con_false = _igvn.makecon(TypeInt::ZERO);
1291
1292 for (uint i = 1; i < n_ctrl->req(); i++) {
1293 if (is_dominator(proj_true, n_ctrl->in(i))) {
1294 bolphi->init_req(i, con_true);
1295 } else {
1296 assert(is_dominator(proj_false, n_ctrl->in(i)), "bad if");
1297 bolphi->init_req(i, con_false);
1298 }
1299 }
1300 register_new_node(bolphi, n_ctrl);
1301 _igvn.replace_input_of(n, 1, bolphi);
1302
1303 // Now split the IF
1304 do_split_if(n);
1305 return;
1306 }
1307
1308 // Check for an IF ready to split; one that has its
1309 // condition codes input coming from a Phi at the block start.
1310 int n_op = n->Opcode();
1311
1312 // Check for an IF being dominated by another IF same test
1313 if (n_op == Op_If ||
1314 n_op == Op_RangeCheck) {
1315 Node *bol = n->in(1);
1316 uint max = bol->outcnt();
1317 // Check for same test used more than once?
1318 if (max > 1 && bol->is_Bool()) {
1319 // Search up IDOMs to see if this IF is dominated.
1320 Node *cutoff = get_ctrl(bol);
1321
1322 // Now search up IDOMs till cutoff, looking for a dominating test
1323 Node *prevdom = n;
1324 Node *dom = idom(prevdom);
1325 while (dom != cutoff) {
1326 if (dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom) {
1327 // Replace the dominated test with an obvious true or false.
1328 // Place it on the IGVN worklist for later cleanup.
1329 C->set_major_progress();
1330 dominated_by(prevdom, n, false, true);
1331 #ifndef PRODUCT
1332 if( VerifyLoopOptimizations ) verify();
1333 #endif
1334 return;
1335 }
1336 prevdom = dom;
1337 dom = idom(prevdom);
1338 }
1339 }
1340 }
1341
1342 // See if a shared loop-varying computation has no loop-varying uses.
1343 // Happens if something is only used for JVM state in uncommon trap exits,
1344 // like various versions of induction variable+offset. Clone the
1345 // computation per usage to allow it to sink out of the loop.
1346 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about)
1347 Node *n_ctrl = get_ctrl(n);
1348 IdealLoopTree *n_loop = get_loop(n_ctrl);
1349 if( n_loop != _ltree_root ) {
1350 DUIterator_Fast imax, i = n->fast_outs(imax);
1351 for (; i < imax; i++) {
1352 Node* u = n->fast_out(i);
1353 if( !has_ctrl(u) ) break; // Found control user
1354 IdealLoopTree *u_loop = get_loop(get_ctrl(u));
1355 if( u_loop == n_loop ) break; // Found loop-varying use
1356 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop
1357 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003
1358 }
1359 bool did_break = (i < imax); // Did we break out of the previous loop?
1360 if (!did_break && n->outcnt() > 1) { // All uses in outer loops!
1361 Node *late_load_ctrl = NULL;
1362 if (n->is_Load()) {
1363 // If n is a load, get and save the result from get_late_ctrl(),
1364 // to be later used in calculating the control for n's clones.
1365 clear_dom_lca_tags();
1366 late_load_ctrl = get_late_ctrl(n, n_ctrl);
1367 }
1368 // If n is a load, and the late control is the same as the current
1369 // control, then the cloning of n is a pointless exercise, because
1370 // GVN will ensure that we end up where we started.
1371 if (!n->is_Load() || late_load_ctrl != n_ctrl) {
1372 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1373 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) {
1374 Node *u = n->last_out(j); // Clone private computation per use
1375 _igvn.rehash_node_delayed(u);
1376 Node *x = n->clone(); // Clone computation
1377 Node *x_ctrl = NULL;
1378 if( u->is_Phi() ) {
1379 // Replace all uses of normal nodes. Replace Phi uses
1380 // individually, so the separate Nodes can sink down
1381 // different paths.
1382 uint k = 1;
1383 while( u->in(k) != n ) k++;
1384 u->set_req( k, x );
1385 // x goes next to Phi input path
1386 x_ctrl = u->in(0)->in(k);
1387 --j;
1388 } else { // Normal use
1389 // Replace all uses
1390 for( uint k = 0; k < u->req(); k++ ) {
1391 if( u->in(k) == n ) {
1392 u->set_req( k, x );
1393 --j;
1394 }
1395 }
1396 x_ctrl = get_ctrl(u);
1397 }
1398
1399 // Find control for 'x' next to use but not inside inner loops.
1400 // For inner loop uses get the preheader area.
1401 x_ctrl = place_near_use(x_ctrl);
1402
1403 if (bs->sink_node(this, n, x, x_ctrl, n_ctrl)) {
1404 continue;
1405 }
1406
1407 if (n->is_Load()) {
1408 // For loads, add a control edge to a CFG node outside of the loop
1409 // to force them to not combine and return back inside the loop
1410 // during GVN optimization (4641526).
1411 //
1412 // Because we are setting the actual control input, factor in
1413 // the result from get_late_ctrl() so we respect any
1414 // anti-dependences. (6233005).
1415 x_ctrl = dom_lca(late_load_ctrl, x_ctrl);
1416
1417 // Don't allow the control input to be a CFG splitting node.
1418 // Such nodes should only have ProjNodes as outs, e.g. IfNode
1419 // should only have IfTrueNode and IfFalseNode (4985384).
1420 x_ctrl = find_non_split_ctrl(x_ctrl);
1421 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone");
1422
1423 x->set_req(0, x_ctrl);
1424 }
1425 register_new_node(x, x_ctrl);
1426
1427 // Some institutional knowledge is needed here: 'x' is
1428 // yanked because if the optimizer runs GVN on it all the
1429 // cloned x's will common up and undo this optimization and
1430 // be forced back in the loop.
1431 // I tried setting control edges on the x's to force them to
1432 // not combine, but the matching gets worried when it tries
1433 // to fold a StoreP and an AddP together (as part of an
1434 // address expression) and the AddP and StoreP have
1435 // different controls.
1436 if (!x->is_Load() && !x->is_DecodeNarrowPtr()) _igvn._worklist.yank(x);
1437 }
1438 _igvn.remove_dead_node(n);
1439 }
1440 }
1441 }
1442 }
1443
1444 try_move_store_after_loop(n);
1445
1446 // Check for Opaque2's who's loop has disappeared - who's input is in the
1447 // same loop nest as their output. Remove 'em, they are no longer useful.
1448 if( n_op == Op_Opaque2 &&
1449 n->in(1) != NULL &&
1450 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) {
1451 _igvn.replace_node( n, n->in(1) );
1452 }
1453 }
1454
1455 //------------------------------split_if_with_blocks---------------------------
1456 // Check for aggressive application of 'split-if' optimization,
1457 // using basic block level info.
1458 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack) {
1459 Node* root = C->root();
1460 visited.set(root->_idx); // first, mark root as visited
1461 // Do pre-visit work for root
1462 Node* n = split_if_with_blocks_pre(root);
1463 uint cnt = n->outcnt();
1464 uint i = 0;
1465
1466 while (true) {
1467 // Visit all children
1468 if (i < cnt) {
1469 Node* use = n->raw_out(i);
1470 ++i;
1471 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
1472 // Now do pre-visit work for this use
1473 use = split_if_with_blocks_pre(use);
1474 nstack.push(n, i); // Save parent and next use's index.
1475 n = use; // Process all children of current use.
1476 cnt = use->outcnt();
1477 i = 0;
1478 }
1479 }
1480 else {
1481 // All of n's children have been processed, complete post-processing.
1482 if (cnt != 0 && !n->is_Con()) {
1483 assert(has_node(n), "no dead nodes");
1484 split_if_with_blocks_post(n);
1485 }
1486 if (must_throttle_split_if()) {
1487 nstack.clear();
1488 }
1489 if (nstack.is_empty()) {
1490 // Finished all nodes on stack.
1491 break;
1492 }
1493 // Get saved parent node and next use's index. Visit the rest of uses.
1494 n = nstack.node();
1495 cnt = n->outcnt();
1496 i = nstack.index();
1497 nstack.pop();
1498 }
1499 }
1500 }
1501
1502
1503 //=============================================================================
1504 //
1505 // C L O N E A L O O P B O D Y
1506 //
1507
1508 //------------------------------clone_iff--------------------------------------
1509 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1510 // "Nearly" because all Nodes have been cloned from the original in the loop,
1511 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1512 // through the Phi recursively, and return a Bool.
1513 Node* PhaseIdealLoop::clone_iff(PhiNode *phi, IdealLoopTree *loop) {
1514
1515 // Convert this Phi into a Phi merging Bools
1516 uint i;
1517 for (i = 1; i < phi->req(); i++) {
1518 Node *b = phi->in(i);
1519 if (b->is_Phi()) {
1520 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi(), loop));
1521 } else {
1522 assert(b->is_Bool() || b->Opcode() == Op_Opaque4, "");
1523 }
1524 }
1525
1526 Node* n = phi->in(1);
1527 Node* sample_opaque = NULL;
1528 Node *sample_bool = NULL;
1529 if (n->Opcode() == Op_Opaque4) {
1530 sample_opaque = n;
1531 sample_bool = n->in(1);
1532 assert(sample_bool->is_Bool(), "wrong type");
1533 } else {
1534 sample_bool = n;
1535 }
1536 Node *sample_cmp = sample_bool->in(1);
1537
1538 // Make Phis to merge the Cmp's inputs.
1539 PhiNode *phi1 = new PhiNode(phi->in(0), Type::TOP);
1540 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP);
1541 for (i = 1; i < phi->req(); i++) {
1542 Node *n1 = sample_opaque == NULL ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1);
1543 Node *n2 = sample_opaque == NULL ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2);
1544 phi1->set_req(i, n1);
1545 phi2->set_req(i, n2);
1546 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
1547 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
1548 }
1549 // See if these Phis have been made before.
1550 // Register with optimizer
1551 Node *hit1 = _igvn.hash_find_insert(phi1);
1552 if (hit1) { // Hit, toss just made Phi
1553 _igvn.remove_dead_node(phi1); // Remove new phi
1554 assert(hit1->is_Phi(), "" );
1555 phi1 = (PhiNode*)hit1; // Use existing phi
1556 } else { // Miss
1557 _igvn.register_new_node_with_optimizer(phi1);
1558 }
1559 Node *hit2 = _igvn.hash_find_insert(phi2);
1560 if (hit2) { // Hit, toss just made Phi
1561 _igvn.remove_dead_node(phi2); // Remove new phi
1562 assert(hit2->is_Phi(), "" );
1563 phi2 = (PhiNode*)hit2; // Use existing phi
1564 } else { // Miss
1565 _igvn.register_new_node_with_optimizer(phi2);
1566 }
1567 // Register Phis with loop/block info
1568 set_ctrl(phi1, phi->in(0));
1569 set_ctrl(phi2, phi->in(0));
1570 // Make a new Cmp
1571 Node *cmp = sample_cmp->clone();
1572 cmp->set_req(1, phi1);
1573 cmp->set_req(2, phi2);
1574 _igvn.register_new_node_with_optimizer(cmp);
1575 set_ctrl(cmp, phi->in(0));
1576
1577 // Make a new Bool
1578 Node *b = sample_bool->clone();
1579 b->set_req(1,cmp);
1580 _igvn.register_new_node_with_optimizer(b);
1581 set_ctrl(b, phi->in(0));
1582
1583 if (sample_opaque != NULL) {
1584 Node* opaque = sample_opaque->clone();
1585 opaque->set_req(1, b);
1586 _igvn.register_new_node_with_optimizer(opaque);
1587 set_ctrl(opaque, phi->in(0));
1588 return opaque;
1589 }
1590
1591 assert(b->is_Bool(), "");
1592 return b;
1593 }
1594
1595 //------------------------------clone_bool-------------------------------------
1596 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1597 // "Nearly" because all Nodes have been cloned from the original in the loop,
1598 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1599 // through the Phi recursively, and return a Bool.
1600 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) {
1601 uint i;
1602 // Convert this Phi into a Phi merging Bools
1603 for( i = 1; i < phi->req(); i++ ) {
1604 Node *b = phi->in(i);
1605 if( b->is_Phi() ) {
1606 _igvn.replace_input_of(phi, i, clone_bool( b->as_Phi(), loop ));
1607 } else {
1608 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
1609 }
1610 }
1611
1612 Node *sample_cmp = phi->in(1);
1613
1614 // Make Phis to merge the Cmp's inputs.
1615 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP );
1616 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP );
1617 for( uint j = 1; j < phi->req(); j++ ) {
1618 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
1619 Node *n1, *n2;
1620 if( cmp_top->is_Cmp() ) {
1621 n1 = cmp_top->in(1);
1622 n2 = cmp_top->in(2);
1623 } else {
1624 n1 = n2 = cmp_top;
1625 }
1626 phi1->set_req( j, n1 );
1627 phi2->set_req( j, n2 );
1628 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
1629 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
1630 }
1631
1632 // See if these Phis have been made before.
1633 // Register with optimizer
1634 Node *hit1 = _igvn.hash_find_insert(phi1);
1635 if( hit1 ) { // Hit, toss just made Phi
1636 _igvn.remove_dead_node(phi1); // Remove new phi
1637 assert( hit1->is_Phi(), "" );
1638 phi1 = (PhiNode*)hit1; // Use existing phi
1639 } else { // Miss
1640 _igvn.register_new_node_with_optimizer(phi1);
1641 }
1642 Node *hit2 = _igvn.hash_find_insert(phi2);
1643 if( hit2 ) { // Hit, toss just made Phi
1644 _igvn.remove_dead_node(phi2); // Remove new phi
1645 assert( hit2->is_Phi(), "" );
1646 phi2 = (PhiNode*)hit2; // Use existing phi
1647 } else { // Miss
1648 _igvn.register_new_node_with_optimizer(phi2);
1649 }
1650 // Register Phis with loop/block info
1651 set_ctrl(phi1, phi->in(0));
1652 set_ctrl(phi2, phi->in(0));
1653 // Make a new Cmp
1654 Node *cmp = sample_cmp->clone();
1655 cmp->set_req( 1, phi1 );
1656 cmp->set_req( 2, phi2 );
1657 _igvn.register_new_node_with_optimizer(cmp);
1658 set_ctrl(cmp, phi->in(0));
1659
1660 assert( cmp->is_Cmp(), "" );
1661 return (CmpNode*)cmp;
1662 }
1663
1664 //------------------------------sink_use---------------------------------------
1665 // If 'use' was in the loop-exit block, it now needs to be sunk
1666 // below the post-loop merge point.
1667 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) {
1668 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) {
1669 set_ctrl(use, post_loop);
1670 for (DUIterator j = use->outs(); use->has_out(j); j++)
1671 sink_use(use->out(j), post_loop);
1672 }
1673 }
1674
1675 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new,
1676 IdealLoopTree* loop, IdealLoopTree* outer_loop,
1677 Node_List*& split_if_set, Node_List*& split_bool_set,
1678 Node_List*& split_cex_set, Node_List& worklist,
1679 uint new_counter, CloneLoopMode mode) {
1680 Node* nnn = old_new[old->_idx];
1681 // Copy uses to a worklist, so I can munge the def-use info
1682 // with impunity.
1683 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
1684 worklist.push(old->fast_out(j));
1685
1686 while( worklist.size() ) {
1687 Node *use = worklist.pop();
1688 if (!has_node(use)) continue; // Ignore dead nodes
1689 if (use->in(0) == C->top()) continue;
1690 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
1691 // Check for data-use outside of loop - at least one of OLD or USE
1692 // must not be a CFG node.
1693 #ifdef ASSERT
1694 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == NULL) {
1695 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint();
1696 assert(mode == ControlAroundStripMined && use == sfpt, "missed a node");
1697 }
1698 #endif
1699 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) {
1700
1701 // If the Data use is an IF, that means we have an IF outside of the
1702 // loop that is switching on a condition that is set inside of the
1703 // loop. Happens if people set a loop-exit flag; then test the flag
1704 // in the loop to break the loop, then test is again outside of the
1705 // loop to determine which way the loop exited.
1706 // Loop predicate If node connects to Bool node through Opaque1 node.
1707 if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use) || use->Opcode() == Op_Opaque4) {
1708 // Since this code is highly unlikely, we lazily build the worklist
1709 // of such Nodes to go split.
1710 if (!split_if_set) {
1711 ResourceArea *area = Thread::current()->resource_area();
1712 split_if_set = new Node_List(area);
1713 }
1714 split_if_set->push(use);
1715 }
1716 if (use->is_Bool()) {
1717 if (!split_bool_set) {
1718 ResourceArea *area = Thread::current()->resource_area();
1719 split_bool_set = new Node_List(area);
1720 }
1721 split_bool_set->push(use);
1722 }
1723 if (use->Opcode() == Op_CreateEx) {
1724 if (!split_cex_set) {
1725 ResourceArea *area = Thread::current()->resource_area();
1726 split_cex_set = new Node_List(area);
1727 }
1728 split_cex_set->push(use);
1729 }
1730
1731
1732 // Get "block" use is in
1733 uint idx = 0;
1734 while( use->in(idx) != old ) idx++;
1735 Node *prev = use->is_CFG() ? use : get_ctrl(use);
1736 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" );
1737 Node *cfg = prev->_idx >= new_counter
1738 ? prev->in(2)
1739 : idom(prev);
1740 if( use->is_Phi() ) // Phi use is in prior block
1741 cfg = prev->in(idx); // NOT in block of Phi itself
1742 if (cfg->is_top()) { // Use is dead?
1743 _igvn.replace_input_of(use, idx, C->top());
1744 continue;
1745 }
1746
1747 // If use is referenced through control edge... (idx == 0)
1748 if (mode == IgnoreStripMined && idx == 0) {
1749 LoopNode *head = loop->_head->as_Loop();
1750 if (head->is_strip_mined() && is_dominator(head->outer_loop_exit(), prev)) {
1751 // That node is outside the inner loop, leave it outside the
1752 // outer loop as well to not confuse verification code.
1753 assert(!loop->_parent->is_member(use_loop), "should be out of the outer loop");
1754 _igvn.replace_input_of(use, 0, head->outer_loop_exit());
1755 continue;
1756 }
1757 }
1758
1759 while(!outer_loop->is_member(get_loop(cfg))) {
1760 prev = cfg;
1761 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg);
1762 }
1763 // If the use occurs after merging several exits from the loop, then
1764 // old value must have dominated all those exits. Since the same old
1765 // value was used on all those exits we did not need a Phi at this
1766 // merge point. NOW we do need a Phi here. Each loop exit value
1767 // is now merged with the peeled body exit; each exit gets its own
1768 // private Phi and those Phis need to be merged here.
1769 Node *phi;
1770 if( prev->is_Region() ) {
1771 if( idx == 0 ) { // Updating control edge?
1772 phi = prev; // Just use existing control
1773 } else { // Else need a new Phi
1774 phi = PhiNode::make( prev, old );
1775 // Now recursively fix up the new uses of old!
1776 for( uint i = 1; i < prev->req(); i++ ) {
1777 worklist.push(phi); // Onto worklist once for each 'old' input
1778 }
1779 }
1780 } else {
1781 // Get new RegionNode merging old and new loop exits
1782 prev = old_new[prev->_idx];
1783 assert( prev, "just made this in step 7" );
1784 if( idx == 0) { // Updating control edge?
1785 phi = prev; // Just use existing control
1786 } else { // Else need a new Phi
1787 // Make a new Phi merging data values properly
1788 phi = PhiNode::make( prev, old );
1789 phi->set_req( 1, nnn );
1790 }
1791 }
1792 // If inserting a new Phi, check for prior hits
1793 if( idx != 0 ) {
1794 Node *hit = _igvn.hash_find_insert(phi);
1795 if( hit == NULL ) {
1796 _igvn.register_new_node_with_optimizer(phi); // Register new phi
1797 } else { // or
1798 // Remove the new phi from the graph and use the hit
1799 _igvn.remove_dead_node(phi);
1800 phi = hit; // Use existing phi
1801 }
1802 set_ctrl(phi, prev);
1803 }
1804 // Make 'use' use the Phi instead of the old loop body exit value
1805 _igvn.replace_input_of(use, idx, phi);
1806 if( use->_idx >= new_counter ) { // If updating new phis
1807 // Not needed for correctness, but prevents a weak assert
1808 // in AddPNode from tripping (when we end up with different
1809 // base & derived Phis that will become the same after
1810 // IGVN does CSE).
1811 Node *hit = _igvn.hash_find_insert(use);
1812 if( hit ) // Go ahead and re-hash for hits.
1813 _igvn.replace_node( use, hit );
1814 }
1815
1816 // If 'use' was in the loop-exit block, it now needs to be sunk
1817 // below the post-loop merge point.
1818 sink_use( use, prev );
1819 }
1820 }
1821 }
1822
1823 static void clone_outer_loop_helper(Node* n, const IdealLoopTree *loop, const IdealLoopTree* outer_loop,
1824 const Node_List &old_new, Unique_Node_List& wq, PhaseIdealLoop* phase,
1825 bool check_old_new) {
1826 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1827 Node* u = n->fast_out(j);
1828 assert(check_old_new || old_new[u->_idx] == NULL, "shouldn't have been cloned");
1829 if (!u->is_CFG() && (!check_old_new || old_new[u->_idx] == NULL)) {
1830 Node* c = phase->get_ctrl(u);
1831 IdealLoopTree* u_loop = phase->get_loop(c);
1832 assert(!loop->is_member(u_loop), "can be in outer loop or out of both loops only");
1833 if (outer_loop->is_member(u_loop)) {
1834 wq.push(u);
1835 }
1836 }
1837 }
1838 }
1839
1840 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
1841 IdealLoopTree* outer_loop, int dd, Node_List &old_new,
1842 Node_List& extra_data_nodes) {
1843 if (head->is_strip_mined() && mode != IgnoreStripMined) {
1844 CountedLoopNode* cl = head->as_CountedLoop();
1845 Node* l = cl->outer_loop();
1846 Node* tail = cl->outer_loop_tail();
1847 IfNode* le = cl->outer_loop_end();
1848 Node* sfpt = cl->outer_safepoint();
1849 CountedLoopEndNode* cle = cl->loopexit();
1850 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop();
1851 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null();
1852 Node* cle_out = cle->proj_out(false);
1853
1854 Node* new_sfpt = NULL;
1855 Node* new_cle_out = cle_out->clone();
1856 old_new.map(cle_out->_idx, new_cle_out);
1857 if (mode == CloneIncludesStripMined) {
1858 // clone outer loop body
1859 Node* new_l = l->clone();
1860 Node* new_tail = tail->clone();
1861 IfNode* new_le = le->clone()->as_If();
1862 new_sfpt = sfpt->clone();
1863
1864 set_loop(new_l, outer_loop->_parent);
1865 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd);
1866 set_loop(new_cle_out, outer_loop->_parent);
1867 set_idom(new_cle_out, new_cle, dd);
1868 set_loop(new_sfpt, outer_loop->_parent);
1869 set_idom(new_sfpt, new_cle_out, dd);
1870 set_loop(new_le, outer_loop->_parent);
1871 set_idom(new_le, new_sfpt, dd);
1872 set_loop(new_tail, outer_loop->_parent);
1873 set_idom(new_tail, new_le, dd);
1874 set_idom(new_cl, new_l, dd);
1875
1876 old_new.map(l->_idx, new_l);
1877 old_new.map(tail->_idx, new_tail);
1878 old_new.map(le->_idx, new_le);
1879 old_new.map(sfpt->_idx, new_sfpt);
1880
1881 new_l->set_req(LoopNode::LoopBackControl, new_tail);
1882 new_l->set_req(0, new_l);
1883 new_tail->set_req(0, new_le);
1884 new_le->set_req(0, new_sfpt);
1885 new_sfpt->set_req(0, new_cle_out);
1886 new_cle_out->set_req(0, new_cle);
1887 new_cl->set_req(LoopNode::EntryControl, new_l);
1888
1889 _igvn.register_new_node_with_optimizer(new_l);
1890 _igvn.register_new_node_with_optimizer(new_tail);
1891 _igvn.register_new_node_with_optimizer(new_le);
1892 } else {
1893 Node *newhead = old_new[loop->_head->_idx];
1894 newhead->as_Loop()->clear_strip_mined();
1895 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl));
1896 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
1897 }
1898 // Look at data node that were assigned a control in the outer
1899 // loop: they are kept in the outer loop by the safepoint so start
1900 // from the safepoint node's inputs.
1901 IdealLoopTree* outer_loop = get_loop(l);
1902 Node_Stack stack(2);
1903 stack.push(sfpt, 1);
1904 uint new_counter = C->unique();
1905 while (stack.size() > 0) {
1906 Node* n = stack.node();
1907 uint i = stack.index();
1908 while (i < n->req() &&
1909 (n->in(i) == NULL ||
1910 !has_ctrl(n->in(i)) ||
1911 get_loop(get_ctrl(n->in(i))) != outer_loop ||
1912 (old_new[n->in(i)->_idx] != NULL && old_new[n->in(i)->_idx]->_idx >= new_counter))) {
1913 i++;
1914 }
1915 if (i < n->req()) {
1916 stack.set_index(i+1);
1917 stack.push(n->in(i), 0);
1918 } else {
1919 assert(old_new[n->_idx] == NULL || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet");
1920 Node* m = n == sfpt ? new_sfpt : n->clone();
1921 if (m != NULL) {
1922 for (uint i = 0; i < n->req(); i++) {
1923 if (m->in(i) != NULL && old_new[m->in(i)->_idx] != NULL) {
1924 m->set_req(i, old_new[m->in(i)->_idx]);
1925 }
1926 }
1927 } else {
1928 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?");
1929 }
1930 if (n != sfpt) {
1931 extra_data_nodes.push(n);
1932 _igvn.register_new_node_with_optimizer(m);
1933 assert(get_ctrl(n) == cle_out, "what other control?");
1934 set_ctrl(m, new_cle_out);
1935 old_new.map(n->_idx, m);
1936 }
1937 stack.pop();
1938 }
1939 }
1940 if (mode == CloneIncludesStripMined) {
1941 _igvn.register_new_node_with_optimizer(new_sfpt);
1942 _igvn.register_new_node_with_optimizer(new_cle_out);
1943 }
1944 // Some other transformation may have pessimistically assign some
1945 // data nodes to the outer loop. Set their control so they are out
1946 // of the outer loop.
1947 ResourceMark rm;
1948 Unique_Node_List wq;
1949 for (uint i = 0; i < extra_data_nodes.size(); i++) {
1950 Node* old = extra_data_nodes.at(i);
1951 clone_outer_loop_helper(old, loop, outer_loop, old_new, wq, this, true);
1952 }
1953 Node* new_ctrl = cl->outer_loop_exit();
1954 assert(get_loop(new_ctrl) != outer_loop, "must be out of the loop nest");
1955 for (uint i = 0; i < wq.size(); i++) {
1956 Node* n = wq.at(i);
1957 set_ctrl(n, new_ctrl);
1958 clone_outer_loop_helper(n, loop, outer_loop, old_new, wq, this, false);
1959 }
1960 } else {
1961 Node *newhead = old_new[loop->_head->_idx];
1962 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
1963 }
1964 }
1965
1966 //------------------------------clone_loop-------------------------------------
1967 //
1968 // C L O N E A L O O P B O D Y
1969 //
1970 // This is the basic building block of the loop optimizations. It clones an
1971 // entire loop body. It makes an old_new loop body mapping; with this mapping
1972 // you can find the new-loop equivalent to an old-loop node. All new-loop
1973 // nodes are exactly equal to their old-loop counterparts, all edges are the
1974 // same. All exits from the old-loop now have a RegionNode that merges the
1975 // equivalent new-loop path. This is true even for the normal "loop-exit"
1976 // condition. All uses of loop-invariant old-loop values now come from (one
1977 // or more) Phis that merge their new-loop equivalents.
1978 //
1979 // This operation leaves the graph in an illegal state: there are two valid
1980 // control edges coming from the loop pre-header to both loop bodies. I'll
1981 // definitely have to hack the graph after running this transform.
1982 //
1983 // From this building block I will further edit edges to perform loop peeling
1984 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
1985 //
1986 // Parameter side_by_size_idom:
1987 // When side_by_size_idom is NULL, the dominator tree is constructed for
1988 // the clone loop to dominate the original. Used in construction of
1989 // pre-main-post loop sequence.
1990 // When nonnull, the clone and original are side-by-side, both are
1991 // dominated by the side_by_side_idom node. Used in construction of
1992 // unswitched loops.
1993 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
1994 CloneLoopMode mode, Node* side_by_side_idom) {
1995
1996 LoopNode* head = loop->_head->as_Loop();
1997 head->verify_strip_mined(1);
1998
1999 if (C->do_vector_loop() && PrintOpto) {
2000 const char* mname = C->method()->name()->as_quoted_ascii();
2001 if (mname != NULL) {
2002 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname);
2003 }
2004 }
2005
2006 CloneMap& cm = C->clone_map();
2007 Dict* dict = cm.dict();
2008 if (C->do_vector_loop()) {
2009 cm.set_clone_idx(cm.max_gen()+1);
2010 #ifndef PRODUCT
2011 if (PrintOpto) {
2012 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx());
2013 loop->dump_head();
2014 }
2015 #endif
2016 }
2017
2018 // Step 1: Clone the loop body. Make the old->new mapping.
2019 uint i;
2020 for( i = 0; i < loop->_body.size(); i++ ) {
2021 Node *old = loop->_body.at(i);
2022 Node *nnn = old->clone();
2023 old_new.map( old->_idx, nnn );
2024 if (C->do_vector_loop()) {
2025 cm.verify_insert_and_clone(old, nnn, cm.clone_idx());
2026 }
2027 _igvn.register_new_node_with_optimizer(nnn);
2028 }
2029
2030 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop;
2031
2032 // Step 2: Fix the edges in the new body. If the old input is outside the
2033 // loop use it. If the old input is INside the loop, use the corresponding
2034 // new node instead.
2035 for( i = 0; i < loop->_body.size(); i++ ) {
2036 Node *old = loop->_body.at(i);
2037 Node *nnn = old_new[old->_idx];
2038 // Fix CFG/Loop controlling the new node
2039 if (has_ctrl(old)) {
2040 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
2041 } else {
2042 set_loop(nnn, outer_loop->_parent);
2043 if (old->outcnt() > 0) {
2044 set_idom( nnn, old_new[idom(old)->_idx], dd );
2045 }
2046 }
2047 // Correct edges to the new node
2048 for( uint j = 0; j < nnn->req(); j++ ) {
2049 Node *n = nnn->in(j);
2050 if( n ) {
2051 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n );
2052 if( loop->is_member( old_in_loop ) )
2053 nnn->set_req(j, old_new[n->_idx]);
2054 }
2055 }
2056 _igvn.hash_find_insert(nnn);
2057 }
2058
2059 ResourceArea *area = Thread::current()->resource_area();
2060 Node_List extra_data_nodes(area); // data nodes in the outer strip mined loop
2061 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes);
2062
2063 // Step 3: Now fix control uses. Loop varying control uses have already
2064 // been fixed up (as part of all input edges in Step 2). Loop invariant
2065 // control uses must be either an IfFalse or an IfTrue. Make a merge
2066 // point to merge the old and new IfFalse/IfTrue nodes; make the use
2067 // refer to this.
2068 Node_List worklist(area);
2069 uint new_counter = C->unique();
2070 for( i = 0; i < loop->_body.size(); i++ ) {
2071 Node* old = loop->_body.at(i);
2072 if( !old->is_CFG() ) continue;
2073
2074 // Copy uses to a worklist, so I can munge the def-use info
2075 // with impunity.
2076 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
2077 worklist.push(old->fast_out(j));
2078
2079 while( worklist.size() ) { // Visit all uses
2080 Node *use = worklist.pop();
2081 if (!has_node(use)) continue; // Ignore dead nodes
2082 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
2083 if( !loop->is_member( use_loop ) && use->is_CFG() ) {
2084 // Both OLD and USE are CFG nodes here.
2085 assert( use->is_Proj(), "" );
2086 Node* nnn = old_new[old->_idx];
2087
2088 Node* newuse = NULL;
2089 if (head->is_strip_mined() && mode != IgnoreStripMined) {
2090 CountedLoopNode* cl = head->as_CountedLoop();
2091 CountedLoopEndNode* cle = cl->loopexit();
2092 Node* cle_out = cle->proj_out_or_null(false);
2093 if (use == cle_out) {
2094 IfNode* le = cl->outer_loop_end();
2095 use = le->proj_out(false);
2096 use_loop = get_loop(use);
2097 if (mode == CloneIncludesStripMined) {
2098 nnn = old_new[le->_idx];
2099 } else {
2100 newuse = old_new[cle_out->_idx];
2101 }
2102 }
2103 }
2104 if (newuse == NULL) {
2105 newuse = use->clone();
2106 }
2107
2108 // Clone the loop exit control projection
2109 if (C->do_vector_loop()) {
2110 cm.verify_insert_and_clone(use, newuse, cm.clone_idx());
2111 }
2112 newuse->set_req(0,nnn);
2113 _igvn.register_new_node_with_optimizer(newuse);
2114 set_loop(newuse, use_loop);
2115 set_idom(newuse, nnn, dom_depth(nnn) + 1 );
2116
2117 // We need a Region to merge the exit from the peeled body and the
2118 // exit from the old loop body.
2119 RegionNode *r = new RegionNode(3);
2120 // Map the old use to the new merge point
2121 old_new.map( use->_idx, r );
2122 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use));
2123 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" );
2124
2125 // The original user of 'use' uses 'r' instead.
2126 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
2127 Node* useuse = use->last_out(l);
2128 _igvn.rehash_node_delayed(useuse);
2129 uint uses_found = 0;
2130 if( useuse->in(0) == use ) {
2131 useuse->set_req(0, r);
2132 uses_found++;
2133 if( useuse->is_CFG() ) {
2134 assert( dom_depth(useuse) > dd_r, "" );
2135 set_idom(useuse, r, dom_depth(useuse));
2136 }
2137 }
2138 for( uint k = 1; k < useuse->req(); k++ ) {
2139 if( useuse->in(k) == use ) {
2140 useuse->set_req(k, r);
2141 uses_found++;
2142 if (useuse->is_Loop() && k == LoopNode::EntryControl) {
2143 assert(dom_depth(useuse) > dd_r , "");
2144 set_idom(useuse, r, dom_depth(useuse));
2145 }
2146 }
2147 }
2148 l -= uses_found; // we deleted 1 or more copies of this edge
2149 }
2150
2151 // Now finish up 'r'
2152 r->set_req( 1, newuse );
2153 r->set_req( 2, use );
2154 _igvn.register_new_node_with_optimizer(r);
2155 set_loop(r, use_loop);
2156 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r);
2157 } // End of if a loop-exit test
2158 }
2159 }
2160
2161 // Step 4: If loop-invariant use is not control, it must be dominated by a
2162 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region
2163 // there if needed. Make a Phi there merging old and new used values.
2164 Node_List *split_if_set = NULL;
2165 Node_List *split_bool_set = NULL;
2166 Node_List *split_cex_set = NULL;
2167 for( i = 0; i < loop->_body.size(); i++ ) {
2168 Node* old = loop->_body.at(i);
2169 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2170 split_bool_set, split_cex_set, worklist, new_counter,
2171 mode);
2172 }
2173
2174 for (i = 0; i < extra_data_nodes.size(); i++) {
2175 Node* old = extra_data_nodes.at(i);
2176 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2177 split_bool_set, split_cex_set, worklist, new_counter,
2178 mode);
2179 }
2180
2181 // Check for IFs that need splitting/cloning. Happens if an IF outside of
2182 // the loop uses a condition set in the loop. The original IF probably
2183 // takes control from one or more OLD Regions (which in turn get from NEW
2184 // Regions). In any case, there will be a set of Phis for each merge point
2185 // from the IF up to where the original BOOL def exists the loop.
2186 if (split_if_set) {
2187 while (split_if_set->size()) {
2188 Node *iff = split_if_set->pop();
2189 if (iff->in(1)->is_Phi()) {
2190 Node *b = clone_iff(iff->in(1)->as_Phi(), loop);
2191 _igvn.replace_input_of(iff, 1, b);
2192 }
2193 }
2194 }
2195 if (split_bool_set) {
2196 while (split_bool_set->size()) {
2197 Node *b = split_bool_set->pop();
2198 Node *phi = b->in(1);
2199 assert(phi->is_Phi(), "");
2200 CmpNode *cmp = clone_bool((PhiNode*)phi, loop);
2201 _igvn.replace_input_of(b, 1, cmp);
2202 }
2203 }
2204 if (split_cex_set) {
2205 while (split_cex_set->size()) {
2206 Node *b = split_cex_set->pop();
2207 assert(b->in(0)->is_Region(), "");
2208 assert(b->in(1)->is_Phi(), "");
2209 assert(b->in(0)->in(0) == b->in(1)->in(0), "");
2210 split_up(b, b->in(0), NULL);
2211 }
2212 }
2213
2214 }
2215
2216
2217 //---------------------- stride_of_possible_iv -------------------------------------
2218 // Looks for an iff/bool/comp with one operand of the compare
2219 // being a cycle involving an add and a phi,
2220 // with an optional truncation (left-shift followed by a right-shift)
2221 // of the add. Returns zero if not an iv.
2222 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
2223 Node* trunc1 = NULL;
2224 Node* trunc2 = NULL;
2225 const TypeInt* ttype = NULL;
2226 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) {
2227 return 0;
2228 }
2229 BoolNode* bl = iff->in(1)->as_Bool();
2230 Node* cmp = bl->in(1);
2231 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) {
2232 return 0;
2233 }
2234 // Must have an invariant operand
2235 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) {
2236 return 0;
2237 }
2238 Node* add2 = NULL;
2239 Node* cmp1 = cmp->in(1);
2240 if (cmp1->is_Phi()) {
2241 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
2242 Node* phi = cmp1;
2243 for (uint i = 1; i < phi->req(); i++) {
2244 Node* in = phi->in(i);
2245 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in,
2246 &trunc1, &trunc2, &ttype);
2247 if (add && add->in(1) == phi) {
2248 add2 = add->in(2);
2249 break;
2250 }
2251 }
2252 } else {
2253 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
2254 Node* addtrunc = cmp1;
2255 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc,
2256 &trunc1, &trunc2, &ttype);
2257 if (add && add->in(1)->is_Phi()) {
2258 Node* phi = add->in(1);
2259 for (uint i = 1; i < phi->req(); i++) {
2260 if (phi->in(i) == addtrunc) {
2261 add2 = add->in(2);
2262 break;
2263 }
2264 }
2265 }
2266 }
2267 if (add2 != NULL) {
2268 const TypeInt* add2t = _igvn.type(add2)->is_int();
2269 if (add2t->is_con()) {
2270 return add2t->get_con();
2271 }
2272 }
2273 return 0;
2274 }
2275
2276
2277 //---------------------- stay_in_loop -------------------------------------
2278 // Return the (unique) control output node that's in the loop (if it exists.)
2279 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
2280 Node* unique = NULL;
2281 if (!n) return NULL;
2282 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2283 Node* use = n->fast_out(i);
2284 if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
2285 if (unique != NULL) {
2286 return NULL;
2287 }
2288 unique = use;
2289 }
2290 }
2291 return unique;
2292 }
2293
2294 //------------------------------ register_node -------------------------------------
2295 // Utility to register node "n" with PhaseIdealLoop
2296 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) {
2297 _igvn.register_new_node_with_optimizer(n);
2298 loop->_body.push(n);
2299 if (n->is_CFG()) {
2300 set_loop(n, loop);
2301 set_idom(n, pred, ddepth);
2302 } else {
2303 set_ctrl(n, pred);
2304 }
2305 }
2306
2307 //------------------------------ proj_clone -------------------------------------
2308 // Utility to create an if-projection
2309 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
2310 ProjNode* c = p->clone()->as_Proj();
2311 c->set_req(0, iff);
2312 return c;
2313 }
2314
2315 //------------------------------ short_circuit_if -------------------------------------
2316 // Force the iff control output to be the live_proj
2317 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
2318 guarantee(live_proj != NULL, "null projection");
2319 int proj_con = live_proj->_con;
2320 assert(proj_con == 0 || proj_con == 1, "false or true projection");
2321 Node *con = _igvn.intcon(proj_con);
2322 set_ctrl(con, C->root());
2323 if (iff) {
2324 iff->set_req(1, con);
2325 }
2326 return con;
2327 }
2328
2329 //------------------------------ insert_if_before_proj -------------------------------------
2330 // Insert a new if before an if projection (* - new node)
2331 //
2332 // before
2333 // if(test)
2334 // / \
2335 // v v
2336 // other-proj proj (arg)
2337 //
2338 // after
2339 // if(test)
2340 // / \
2341 // / v
2342 // | * proj-clone
2343 // v |
2344 // other-proj v
2345 // * new_if(relop(cmp[IU](left,right)))
2346 // / \
2347 // v v
2348 // * new-proj proj
2349 // (returned)
2350 //
2351 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, 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 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right);
2365 register_node(cmp, loop, proj2, ddepth);
2366
2367 BoolNode* bol = new BoolNode(cmp, relop);
2368 register_node(bol, loop, proj2, ddepth);
2369
2370 int opcode = iff->Opcode();
2371 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode");
2372 IfNode* new_if = (opcode == Op_If) ? new IfNode(proj2, bol, iff->_prob, iff->_fcnt):
2373 new RangeCheckNode(proj2, bol, iff->_prob, iff->_fcnt);
2374 register_node(new_if, loop, proj2, ddepth);
2375
2376 proj->set_req(0, new_if); // reattach
2377 set_idom(proj, new_if, ddepth);
2378
2379 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
2380 guarantee(new_exit != NULL, "null exit node");
2381 register_node(new_exit, get_loop(other_proj), new_if, ddepth);
2382
2383 return new_exit;
2384 }
2385
2386 //------------------------------ insert_region_before_proj -------------------------------------
2387 // Insert a region before an if projection (* - new node)
2388 //
2389 // before
2390 // if(test)
2391 // / |
2392 // v |
2393 // proj v
2394 // other-proj
2395 //
2396 // after
2397 // if(test)
2398 // / |
2399 // v |
2400 // * proj-clone v
2401 // | other-proj
2402 // v
2403 // * new-region
2404 // |
2405 // v
2406 // * dum_if
2407 // / \
2408 // v \
2409 // * dum-proj v
2410 // proj
2411 //
2412 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
2413 IfNode* iff = proj->in(0)->as_If();
2414 IdealLoopTree *loop = get_loop(proj);
2415 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2416 int ddepth = dom_depth(proj);
2417
2418 _igvn.rehash_node_delayed(iff);
2419 _igvn.rehash_node_delayed(proj);
2420
2421 proj->set_req(0, NULL); // temporary disconnect
2422 ProjNode* proj2 = proj_clone(proj, iff);
2423 register_node(proj2, loop, iff, ddepth);
2424
2425 RegionNode* reg = new RegionNode(2);
2426 reg->set_req(1, proj2);
2427 register_node(reg, loop, iff, ddepth);
2428
2429 IfNode* dum_if = new IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt);
2430 register_node(dum_if, loop, reg, ddepth);
2431
2432 proj->set_req(0, dum_if); // reattach
2433 set_idom(proj, dum_if, ddepth);
2434
2435 ProjNode* dum_proj = proj_clone(other_proj, dum_if);
2436 register_node(dum_proj, loop, dum_if, ddepth);
2437
2438 return reg;
2439 }
2440
2441 //------------------------------ insert_cmpi_loop_exit -------------------------------------
2442 // Clone a signed compare loop exit from an unsigned compare and
2443 // insert it before the unsigned cmp on the stay-in-loop path.
2444 // All new nodes inserted in the dominator tree between the original
2445 // if and it's projections. The original if test is replaced with
2446 // a constant to force the stay-in-loop path.
2447 //
2448 // This is done to make sure that the original if and it's projections
2449 // still dominate the same set of control nodes, that the ctrl() relation
2450 // from data nodes to them is preserved, and that their loop nesting is
2451 // preserved.
2452 //
2453 // before
2454 // if(i <u limit) unsigned compare loop exit
2455 // / |
2456 // v v
2457 // exit-proj stay-in-loop-proj
2458 //
2459 // after
2460 // if(stay-in-loop-const) original if
2461 // / |
2462 // / v
2463 // / if(i < limit) new signed test
2464 // / / |
2465 // / / v
2466 // / / if(i <u limit) new cloned unsigned test
2467 // / / / |
2468 // v v v |
2469 // region |
2470 // | |
2471 // dum-if |
2472 // / | |
2473 // ether | |
2474 // v v
2475 // exit-proj stay-in-loop-proj
2476 //
2477 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) {
2478 const bool Signed = true;
2479 const bool Unsigned = false;
2480
2481 BoolNode* bol = if_cmpu->in(1)->as_Bool();
2482 if (bol->_test._test != BoolTest::lt) return NULL;
2483 CmpNode* cmpu = bol->in(1)->as_Cmp();
2484 if (cmpu->Opcode() != Op_CmpU) return NULL;
2485 int stride = stride_of_possible_iv(if_cmpu);
2486 if (stride == 0) return NULL;
2487
2488 Node* lp_proj = stay_in_loop(if_cmpu, loop);
2489 guarantee(lp_proj != NULL, "null loop node");
2490
2491 ProjNode* lp_continue = lp_proj->as_Proj();
2492 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
2493
2494 Node* limit = NULL;
2495 if (stride > 0) {
2496 limit = cmpu->in(2);
2497 } else {
2498 limit = _igvn.makecon(TypeInt::ZERO);
2499 set_ctrl(limit, C->root());
2500 }
2501 // Create a new region on the exit path
2502 RegionNode* reg = insert_region_before_proj(lp_exit);
2503 guarantee(reg != NULL, "null region node");
2504
2505 // Clone the if-cmpu-true-false using a signed compare
2506 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
2507 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue);
2508 reg->add_req(cmpi_exit);
2509
2510 // Clone the if-cmpu-true-false
2511 BoolTest::mask rel_u = bol->_test._test;
2512 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
2513 reg->add_req(cmpu_exit);
2514
2515 // Force original if to stay in loop.
2516 short_circuit_if(if_cmpu, lp_continue);
2517
2518 return cmpi_exit->in(0)->as_If();
2519 }
2520
2521 //------------------------------ remove_cmpi_loop_exit -------------------------------------
2522 // Remove a previously inserted signed compare loop exit.
2523 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
2524 Node* lp_proj = stay_in_loop(if_cmp, loop);
2525 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
2526 stay_in_loop(lp_proj, loop)->is_If() &&
2527 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
2528 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
2529 set_ctrl(con, C->root());
2530 if_cmp->set_req(1, con);
2531 }
2532
2533 //------------------------------ scheduled_nodelist -------------------------------------
2534 // Create a post order schedule of nodes that are in the
2535 // "member" set. The list is returned in "sched".
2536 // The first node in "sched" is the loop head, followed by
2537 // nodes which have no inputs in the "member" set, and then
2538 // followed by the nodes that have an immediate input dependence
2539 // on a node in "sched".
2540 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
2541
2542 assert(member.test(loop->_head->_idx), "loop head must be in member set");
2543 Arena *a = Thread::current()->resource_area();
2544 VectorSet visited(a);
2545 Node_Stack nstack(a, loop->_body.size());
2546
2547 Node* n = loop->_head; // top of stack is cached in "n"
2548 uint idx = 0;
2549 visited.set(n->_idx);
2550
2551 // Initially push all with no inputs from within member set
2552 for(uint i = 0; i < loop->_body.size(); i++ ) {
2553 Node *elt = loop->_body.at(i);
2554 if (member.test(elt->_idx)) {
2555 bool found = false;
2556 for (uint j = 0; j < elt->req(); j++) {
2557 Node* def = elt->in(j);
2558 if (def && member.test(def->_idx) && def != elt) {
2559 found = true;
2560 break;
2561 }
2562 }
2563 if (!found && elt != loop->_head) {
2564 nstack.push(n, idx);
2565 n = elt;
2566 assert(!visited.test(n->_idx), "not seen yet");
2567 visited.set(n->_idx);
2568 }
2569 }
2570 }
2571
2572 // traverse out's that are in the member set
2573 while (true) {
2574 if (idx < n->outcnt()) {
2575 Node* use = n->raw_out(idx);
2576 idx++;
2577 if (!visited.test_set(use->_idx)) {
2578 if (member.test(use->_idx)) {
2579 nstack.push(n, idx);
2580 n = use;
2581 idx = 0;
2582 }
2583 }
2584 } else {
2585 // All outputs processed
2586 sched.push(n);
2587 if (nstack.is_empty()) break;
2588 n = nstack.node();
2589 idx = nstack.index();
2590 nstack.pop();
2591 }
2592 }
2593 }
2594
2595
2596 //------------------------------ has_use_in_set -------------------------------------
2597 // Has a use in the vector set
2598 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
2599 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2600 Node* use = n->fast_out(j);
2601 if (vset.test(use->_idx)) {
2602 return true;
2603 }
2604 }
2605 return false;
2606 }
2607
2608
2609 //------------------------------ has_use_internal_to_set -------------------------------------
2610 // Has use internal to the vector set (ie. not in a phi at the loop head)
2611 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
2612 Node* head = loop->_head;
2613 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2614 Node* use = n->fast_out(j);
2615 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
2616 return true;
2617 }
2618 }
2619 return false;
2620 }
2621
2622
2623 //------------------------------ clone_for_use_outside_loop -------------------------------------
2624 // clone "n" for uses that are outside of loop
2625 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
2626 int cloned = 0;
2627 assert(worklist.size() == 0, "should be empty");
2628 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2629 Node* use = n->fast_out(j);
2630 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
2631 worklist.push(use);
2632 }
2633 }
2634 while( worklist.size() ) {
2635 Node *use = worklist.pop();
2636 if (!has_node(use) || use->in(0) == C->top()) continue;
2637 uint j;
2638 for (j = 0; j < use->req(); j++) {
2639 if (use->in(j) == n) break;
2640 }
2641 assert(j < use->req(), "must be there");
2642
2643 // clone "n" and insert it between the inputs of "n" and the use outside the loop
2644 Node* n_clone = n->clone();
2645 _igvn.replace_input_of(use, j, n_clone);
2646 cloned++;
2647 Node* use_c;
2648 if (!use->is_Phi()) {
2649 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
2650 } else {
2651 // Use in a phi is considered a use in the associated predecessor block
2652 use_c = use->in(0)->in(j);
2653 }
2654 set_ctrl(n_clone, use_c);
2655 assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
2656 get_loop(use_c)->_body.push(n_clone);
2657 _igvn.register_new_node_with_optimizer(n_clone);
2658 #ifndef PRODUCT
2659 if (TracePartialPeeling) {
2660 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
2661 }
2662 #endif
2663 }
2664 return cloned;
2665 }
2666
2667
2668 //------------------------------ clone_for_special_use_inside_loop -------------------------------------
2669 // clone "n" for special uses that are in the not_peeled region.
2670 // If these def-uses occur in separate blocks, the code generator
2671 // marks the method as not compilable. For example, if a "BoolNode"
2672 // is in a different basic block than the "IfNode" that uses it, then
2673 // the compilation is aborted in the code generator.
2674 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
2675 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
2676 if (n->is_Phi() || n->is_Load()) {
2677 return;
2678 }
2679 assert(worklist.size() == 0, "should be empty");
2680 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2681 Node* use = n->fast_out(j);
2682 if ( not_peel.test(use->_idx) &&
2683 (use->is_If() || use->is_CMove() || use->is_Bool()) &&
2684 use->in(1) == n) {
2685 worklist.push(use);
2686 }
2687 }
2688 if (worklist.size() > 0) {
2689 // clone "n" and insert it between inputs of "n" and the use
2690 Node* n_clone = n->clone();
2691 loop->_body.push(n_clone);
2692 _igvn.register_new_node_with_optimizer(n_clone);
2693 set_ctrl(n_clone, get_ctrl(n));
2694 sink_list.push(n_clone);
2695 not_peel <<= n_clone->_idx; // add n_clone to not_peel set.
2696 #ifndef PRODUCT
2697 if (TracePartialPeeling) {
2698 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
2699 }
2700 #endif
2701 while( worklist.size() ) {
2702 Node *use = worklist.pop();
2703 _igvn.rehash_node_delayed(use);
2704 for (uint j = 1; j < use->req(); j++) {
2705 if (use->in(j) == n) {
2706 use->set_req(j, n_clone);
2707 }
2708 }
2709 }
2710 }
2711 }
2712
2713
2714 //------------------------------ insert_phi_for_loop -------------------------------------
2715 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
2716 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
2717 Node *phi = PhiNode::make(lp, back_edge_val);
2718 phi->set_req(LoopNode::EntryControl, lp_entry_val);
2719 // Use existing phi if it already exists
2720 Node *hit = _igvn.hash_find_insert(phi);
2721 if( hit == NULL ) {
2722 _igvn.register_new_node_with_optimizer(phi);
2723 set_ctrl(phi, lp);
2724 } else {
2725 // Remove the new phi from the graph and use the hit
2726 _igvn.remove_dead_node(phi);
2727 phi = hit;
2728 }
2729 _igvn.replace_input_of(use, idx, phi);
2730 }
2731
2732 #ifdef ASSERT
2733 //------------------------------ is_valid_loop_partition -------------------------------------
2734 // Validate the loop partition sets: peel and not_peel
2735 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
2736 VectorSet& not_peel ) {
2737 uint i;
2738 // Check that peel_list entries are in the peel set
2739 for (i = 0; i < peel_list.size(); i++) {
2740 if (!peel.test(peel_list.at(i)->_idx)) {
2741 return false;
2742 }
2743 }
2744 // Check at loop members are in one of peel set or not_peel set
2745 for (i = 0; i < loop->_body.size(); i++ ) {
2746 Node *def = loop->_body.at(i);
2747 uint di = def->_idx;
2748 // Check that peel set elements are in peel_list
2749 if (peel.test(di)) {
2750 if (not_peel.test(di)) {
2751 return false;
2752 }
2753 // Must be in peel_list also
2754 bool found = false;
2755 for (uint j = 0; j < peel_list.size(); j++) {
2756 if (peel_list.at(j)->_idx == di) {
2757 found = true;
2758 break;
2759 }
2760 }
2761 if (!found) {
2762 return false;
2763 }
2764 } else if (not_peel.test(di)) {
2765 if (peel.test(di)) {
2766 return false;
2767 }
2768 } else {
2769 return false;
2770 }
2771 }
2772 return true;
2773 }
2774
2775 //------------------------------ is_valid_clone_loop_exit_use -------------------------------------
2776 // Ensure a use outside of loop is of the right form
2777 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
2778 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2779 return (use->is_Phi() &&
2780 use_c->is_Region() && use_c->req() == 3 &&
2781 (use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
2782 use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
2783 use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
2784 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
2785 }
2786
2787 //------------------------------ is_valid_clone_loop_form -------------------------------------
2788 // Ensure that all uses outside of loop are of the right form
2789 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
2790 uint orig_exit_idx, uint clone_exit_idx) {
2791 uint len = peel_list.size();
2792 for (uint i = 0; i < len; i++) {
2793 Node *def = peel_list.at(i);
2794
2795 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2796 Node *use = def->fast_out(j);
2797 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2798 if (!loop->is_member(get_loop(use_c))) {
2799 // use is not in the loop, check for correct structure
2800 if (use->in(0) == def) {
2801 // Okay
2802 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
2803 return false;
2804 }
2805 }
2806 }
2807 }
2808 return true;
2809 }
2810 #endif
2811
2812 //------------------------------ partial_peel -------------------------------------
2813 // Partially peel (aka loop rotation) the top portion of a loop (called
2814 // the peel section below) by cloning it and placing one copy just before
2815 // the new loop head and the other copy at the bottom of the new loop.
2816 //
2817 // before after where it came from
2818 //
2819 // stmt1 stmt1
2820 // loop: stmt2 clone
2821 // stmt2 if condA goto exitA clone
2822 // if condA goto exitA new_loop: new
2823 // stmt3 stmt3 clone
2824 // if !condB goto loop if condB goto exitB clone
2825 // exitB: stmt2 orig
2826 // stmt4 if !condA goto new_loop orig
2827 // exitA: goto exitA
2828 // exitB:
2829 // stmt4
2830 // exitA:
2831 //
2832 // Step 1: find the cut point: an exit test on probable
2833 // induction variable.
2834 // Step 2: schedule (with cloning) operations in the peel
2835 // section that can be executed after the cut into
2836 // the section that is not peeled. This may need
2837 // to clone operations into exit blocks. For
2838 // instance, a reference to A[i] in the not-peel
2839 // section and a reference to B[i] in an exit block
2840 // may cause a left-shift of i by 2 to be placed
2841 // in the peel block. This step will clone the left
2842 // shift into the exit block and sink the left shift
2843 // from the peel to the not-peel section.
2844 // Step 3: clone the loop, retarget the control, and insert
2845 // phis for values that are live across the new loop
2846 // head. This is very dependent on the graph structure
2847 // from clone_loop. It creates region nodes for
2848 // exit control and associated phi nodes for values
2849 // flow out of the loop through that exit. The region
2850 // node is dominated by the clone's control projection.
2851 // So the clone's peel section is placed before the
2852 // new loop head, and the clone's not-peel section is
2853 // forms the top part of the new loop. The original
2854 // peel section forms the tail of the new loop.
2855 // Step 4: update the dominator tree and recompute the
2856 // dominator depth.
2857 //
2858 // orig
2859 //
2860 // stmt1
2861 // |
2862 // v
2863 // loop predicate
2864 // |
2865 // v
2866 // loop<----+
2867 // | |
2868 // stmt2 |
2869 // | |
2870 // v |
2871 // ifA |
2872 // / | |
2873 // v v |
2874 // false true ^ <-- last_peel
2875 // / | |
2876 // / ===|==cut |
2877 // / stmt3 | <-- first_not_peel
2878 // / | |
2879 // | v |
2880 // v ifB |
2881 // exitA: / \ |
2882 // / \ |
2883 // v v |
2884 // false true |
2885 // / \ |
2886 // / ----+
2887 // |
2888 // v
2889 // exitB:
2890 // stmt4
2891 //
2892 //
2893 // after clone loop
2894 //
2895 // stmt1
2896 // |
2897 // v
2898 // loop predicate
2899 // / \
2900 // clone / \ orig
2901 // / \
2902 // / \
2903 // v v
2904 // +---->loop loop<----+
2905 // | | | |
2906 // | stmt2 stmt2 |
2907 // | | | |
2908 // | v v |
2909 // | ifA ifA |
2910 // | | \ / | |
2911 // | v v v v |
2912 // ^ true false false true ^ <-- last_peel
2913 // | | ^ \ / | |
2914 // | cut==|== \ \ / ===|==cut |
2915 // | stmt3 \ \ / stmt3 | <-- first_not_peel
2916 // | | dom | | | |
2917 // | v \ 1v v2 v |
2918 // | ifB regionA ifB |
2919 // | / \ | / \ |
2920 // | / \ v / \ |
2921 // | v v exitA: v v |
2922 // | true false false true |
2923 // | / ^ \ / \ |
2924 // +---- \ \ / ----+
2925 // dom \ /
2926 // \ 1v v2
2927 // regionB
2928 // |
2929 // v
2930 // exitB:
2931 // stmt4
2932 //
2933 //
2934 // after partial peel
2935 //
2936 // stmt1
2937 // |
2938 // v
2939 // loop predicate
2940 // /
2941 // clone / orig
2942 // / TOP
2943 // / \
2944 // v v
2945 // TOP->loop loop----+
2946 // | | |
2947 // stmt2 stmt2 |
2948 // | | |
2949 // v v |
2950 // ifA ifA |
2951 // | \ / | |
2952 // v v v v |
2953 // true false false true | <-- last_peel
2954 // | ^ \ / +------|---+
2955 // +->newloop \ \ / === ==cut | |
2956 // | stmt3 \ \ / TOP | |
2957 // | | dom | | stmt3 | | <-- first_not_peel
2958 // | v \ 1v v2 v | |
2959 // | ifB regionA ifB ^ v
2960 // | / \ | / \ | |
2961 // | / \ v / \ | |
2962 // | v v exitA: v v | |
2963 // | true false false true | |
2964 // | / ^ \ / \ | |
2965 // | | \ \ / v | |
2966 // | | dom \ / TOP | |
2967 // | | \ 1v v2 | |
2968 // ^ v regionB | |
2969 // | | | | |
2970 // | | v ^ v
2971 // | | exitB: | |
2972 // | | stmt4 | |
2973 // | +------------>-----------------+ |
2974 // | |
2975 // +-----------------<---------------------+
2976 //
2977 //
2978 // final graph
2979 //
2980 // stmt1
2981 // |
2982 // v
2983 // loop predicate
2984 // |
2985 // v
2986 // stmt2 clone
2987 // |
2988 // v
2989 // ........> ifA clone
2990 // : / |
2991 // dom / |
2992 // : v v
2993 // : false true
2994 // : | |
2995 // : | v
2996 // : | newloop<-----+
2997 // : | | |
2998 // : | stmt3 clone |
2999 // : | | |
3000 // : | v |
3001 // : | ifB |
3002 // : | / \ |
3003 // : | v v |
3004 // : | false true |
3005 // : | | | |
3006 // : | v stmt2 |
3007 // : | exitB: | |
3008 // : | stmt4 v |
3009 // : | ifA orig |
3010 // : | / \ |
3011 // : | / \ |
3012 // : | v v |
3013 // : | false true |
3014 // : | / \ |
3015 // : v v -----+
3016 // RegionA
3017 // |
3018 // v
3019 // exitA
3020 //
3021 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
3022
3023 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only");
3024 if (!loop->_head->is_Loop()) {
3025 return false;
3026 }
3027 LoopNode *head = loop->_head->as_Loop();
3028
3029 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
3030 return false;
3031 }
3032
3033 // Check for complex exit control
3034 for (uint ii = 0; ii < loop->_body.size(); ii++) {
3035 Node *n = loop->_body.at(ii);
3036 int opc = n->Opcode();
3037 if (n->is_Call() ||
3038 opc == Op_Catch ||
3039 opc == Op_CatchProj ||
3040 opc == Op_Jump ||
3041 opc == Op_JumpProj) {
3042 #ifndef PRODUCT
3043 if (TracePartialPeeling) {
3044 tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
3045 }
3046 #endif
3047 return false;
3048 }
3049 }
3050
3051 int dd = dom_depth(head);
3052
3053 // Step 1: find cut point
3054
3055 // Walk up dominators to loop head looking for first loop exit
3056 // which is executed on every path thru loop.
3057 IfNode *peel_if = NULL;
3058 IfNode *peel_if_cmpu = NULL;
3059
3060 Node *iff = loop->tail();
3061 while (iff != head) {
3062 if (iff->is_If()) {
3063 Node *ctrl = get_ctrl(iff->in(1));
3064 if (ctrl->is_top()) return false; // Dead test on live IF.
3065 // If loop-varying exit-test, check for induction variable
3066 if (loop->is_member(get_loop(ctrl)) &&
3067 loop->is_loop_exit(iff) &&
3068 is_possible_iv_test(iff)) {
3069 Node* cmp = iff->in(1)->in(1);
3070 if (cmp->Opcode() == Op_CmpI) {
3071 peel_if = iff->as_If();
3072 } else {
3073 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
3074 peel_if_cmpu = iff->as_If();
3075 }
3076 }
3077 }
3078 iff = idom(iff);
3079 }
3080
3081 // Prefer signed compare over unsigned compare.
3082 IfNode* new_peel_if = NULL;
3083 if (peel_if == NULL) {
3084 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) {
3085 return false; // No peel point found
3086 }
3087 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
3088 if (new_peel_if == NULL) {
3089 return false; // No peel point found
3090 }
3091 peel_if = new_peel_if;
3092 }
3093 Node* last_peel = stay_in_loop(peel_if, loop);
3094 Node* first_not_peeled = stay_in_loop(last_peel, loop);
3095 if (first_not_peeled == NULL || first_not_peeled == head) {
3096 return false;
3097 }
3098
3099 #ifndef PRODUCT
3100 if (TraceLoopOpts) {
3101 tty->print("PartialPeel ");
3102 loop->dump_head();
3103 }
3104
3105 if (TracePartialPeeling) {
3106 tty->print_cr("before partial peel one iteration");
3107 Node_List wl;
3108 Node* t = head->in(2);
3109 while (true) {
3110 wl.push(t);
3111 if (t == head) break;
3112 t = idom(t);
3113 }
3114 while (wl.size() > 0) {
3115 Node* tt = wl.pop();
3116 tt->dump();
3117 if (tt == last_peel) tty->print_cr("-- cut --");
3118 }
3119 }
3120 #endif
3121 ResourceArea *area = Thread::current()->resource_area();
3122 VectorSet peel(area);
3123 VectorSet not_peel(area);
3124 Node_List peel_list(area);
3125 Node_List worklist(area);
3126 Node_List sink_list(area);
3127
3128 if (!may_require_nodes(loop->est_loop_clone_sz(2))) {
3129 return false;
3130 }
3131
3132 // Set of cfg nodes to peel are those that are executable from
3133 // the head through last_peel.
3134 assert(worklist.size() == 0, "should be empty");
3135 worklist.push(head);
3136 peel.set(head->_idx);
3137 while (worklist.size() > 0) {
3138 Node *n = worklist.pop();
3139 if (n != last_peel) {
3140 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3141 Node* use = n->fast_out(j);
3142 if (use->is_CFG() &&
3143 loop->is_member(get_loop(use)) &&
3144 !peel.test_set(use->_idx)) {
3145 worklist.push(use);
3146 }
3147 }
3148 }
3149 }
3150
3151 // Set of non-cfg nodes to peel are those that are control
3152 // dependent on the cfg nodes.
3153 uint i;
3154 for(i = 0; i < loop->_body.size(); i++ ) {
3155 Node *n = loop->_body.at(i);
3156 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
3157 if (peel.test(n_c->_idx)) {
3158 peel.set(n->_idx);
3159 } else {
3160 not_peel.set(n->_idx);
3161 }
3162 }
3163
3164 // Step 2: move operations from the peeled section down into the
3165 // not-peeled section
3166
3167 // Get a post order schedule of nodes in the peel region
3168 // Result in right-most operand.
3169 scheduled_nodelist(loop, peel, peel_list );
3170
3171 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
3172
3173 // For future check for too many new phis
3174 uint old_phi_cnt = 0;
3175 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
3176 Node* use = head->fast_out(j);
3177 if (use->is_Phi()) old_phi_cnt++;
3178 }
3179
3180 #ifndef PRODUCT
3181 if (TracePartialPeeling) {
3182 tty->print_cr("\npeeled list");
3183 }
3184 #endif
3185
3186 // Evacuate nodes in peel region into the not_peeled region if possible
3187 uint new_phi_cnt = 0;
3188 uint cloned_for_outside_use = 0;
3189 for (i = 0; i < peel_list.size();) {
3190 Node* n = peel_list.at(i);
3191 #ifndef PRODUCT
3192 if (TracePartialPeeling) n->dump();
3193 #endif
3194 bool incr = true;
3195 if ( !n->is_CFG() ) {
3196
3197 if ( has_use_in_set(n, not_peel) ) {
3198
3199 // If not used internal to the peeled region,
3200 // move "n" from peeled to not_peeled region.
3201
3202 if ( !has_use_internal_to_set(n, peel, loop) ) {
3203
3204 // if not pinned and not a load (which maybe anti-dependent on a store)
3205 // and not a CMove (Matcher expects only bool->cmove).
3206 if (n->in(0) == NULL && !n->is_Load() && !n->is_CMove()) {
3207 cloned_for_outside_use += clone_for_use_outside_loop( loop, n, worklist );
3208 sink_list.push(n);
3209 peel >>= n->_idx; // delete n from peel set.
3210 not_peel <<= n->_idx; // add n to not_peel set.
3211 peel_list.remove(i);
3212 incr = false;
3213 #ifndef PRODUCT
3214 if (TracePartialPeeling) {
3215 tty->print_cr("sink to not_peeled region: %d newbb: %d",
3216 n->_idx, get_ctrl(n)->_idx);
3217 }
3218 #endif
3219 }
3220 } else {
3221 // Otherwise check for special def-use cases that span
3222 // the peel/not_peel boundary such as bool->if
3223 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist );
3224 new_phi_cnt++;
3225 }
3226 }
3227 }
3228 if (incr) i++;
3229 }
3230
3231 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) {
3232 #ifndef PRODUCT
3233 if (TracePartialPeeling) {
3234 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c",
3235 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F');
3236 }
3237 #endif
3238 if (new_peel_if != NULL) {
3239 remove_cmpi_loop_exit(new_peel_if, loop);
3240 }
3241 // Inhibit more partial peeling on this loop
3242 assert(!head->is_partial_peel_loop(), "not partial peeled");
3243 head->mark_partial_peel_failed();
3244 if (cloned_for_outside_use > 0) {
3245 // Terminate this round of loop opts because
3246 // the graph outside this loop was changed.
3247 C->set_major_progress();
3248 return true;
3249 }
3250 return false;
3251 }
3252
3253 // Step 3: clone loop, retarget control, and insert new phis
3254
3255 // Create new loop head for new phis and to hang
3256 // the nodes being moved (sinked) from the peel region.
3257 LoopNode* new_head = new LoopNode(last_peel, last_peel);
3258 new_head->set_unswitch_count(head->unswitch_count()); // Preserve
3259 _igvn.register_new_node_with_optimizer(new_head);
3260 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
3261 _igvn.replace_input_of(first_not_peeled, 0, new_head);
3262 set_loop(new_head, loop);
3263 loop->_body.push(new_head);
3264 not_peel.set(new_head->_idx);
3265 set_idom(new_head, last_peel, dom_depth(first_not_peeled));
3266 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
3267
3268 while (sink_list.size() > 0) {
3269 Node* n = sink_list.pop();
3270 set_ctrl(n, new_head);
3271 }
3272
3273 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
3274
3275 clone_loop(loop, old_new, dd, IgnoreStripMined);
3276
3277 const uint clone_exit_idx = 1;
3278 const uint orig_exit_idx = 2;
3279 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop");
3280
3281 Node* head_clone = old_new[head->_idx];
3282 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop();
3283 Node* orig_tail_clone = head_clone->in(2);
3284
3285 // Add phi if "def" node is in peel set and "use" is not
3286
3287 for(i = 0; i < peel_list.size(); i++ ) {
3288 Node *def = peel_list.at(i);
3289 if (!def->is_CFG()) {
3290 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
3291 Node *use = def->fast_out(j);
3292 if (has_node(use) && use->in(0) != C->top() &&
3293 (!peel.test(use->_idx) ||
3294 (use->is_Phi() && use->in(0) == head)) ) {
3295 worklist.push(use);
3296 }
3297 }
3298 while( worklist.size() ) {
3299 Node *use = worklist.pop();
3300 for (uint j = 1; j < use->req(); j++) {
3301 Node* n = use->in(j);
3302 if (n == def) {
3303
3304 // "def" is in peel set, "use" is not in peel set
3305 // or "use" is in the entry boundary (a phi) of the peel set
3306
3307 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
3308
3309 if ( loop->is_member(get_loop( use_c )) ) {
3310 // use is in loop
3311 if (old_new[use->_idx] != NULL) { // null for dead code
3312 Node* use_clone = old_new[use->_idx];
3313 _igvn.replace_input_of(use, j, C->top());
3314 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
3315 }
3316 } else {
3317 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
3318 // use is not in the loop, check if the live range includes the cut
3319 Node* lp_if = use_c->in(orig_exit_idx)->in(0);
3320 if (not_peel.test(lp_if->_idx)) {
3321 assert(j == orig_exit_idx, "use from original loop");
3322 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
3323 }
3324 }
3325 }
3326 }
3327 }
3328 }
3329 }
3330
3331 // Step 3b: retarget control
3332
3333 // Redirect control to the new loop head if a cloned node in
3334 // the not_peeled region has control that points into the peeled region.
3335 // This necessary because the cloned peeled region will be outside
3336 // the loop.
3337 // from to
3338 // cloned-peeled <---+
3339 // new_head_clone: | <--+
3340 // cloned-not_peeled in(0) in(0)
3341 // orig-peeled
3342
3343 for(i = 0; i < loop->_body.size(); i++ ) {
3344 Node *n = loop->_body.at(i);
3345 if (!n->is_CFG() && n->in(0) != NULL &&
3346 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
3347 Node* n_clone = old_new[n->_idx];
3348 _igvn.replace_input_of(n_clone, 0, new_head_clone);
3349 }
3350 }
3351
3352 // Backedge of the surviving new_head (the clone) is original last_peel
3353 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel);
3354
3355 // Cut first node in original not_peel set
3356 _igvn.rehash_node_delayed(new_head); // Multiple edge updates:
3357 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of
3358 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls
3359
3360 // Copy head_clone back-branch info to original head
3361 // and remove original head's loop entry and
3362 // clone head's back-branch
3363 _igvn.rehash_node_delayed(head); // Multiple edge updates
3364 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl));
3365 head->set_req(LoopNode::LoopBackControl, C->top());
3366 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top());
3367
3368 // Similarly modify the phis
3369 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
3370 Node* use = head->fast_out(k);
3371 if (use->is_Phi() && use->outcnt() > 0) {
3372 Node* use_clone = old_new[use->_idx];
3373 _igvn.rehash_node_delayed(use); // Multiple edge updates
3374 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl));
3375 use->set_req(LoopNode::LoopBackControl, C->top());
3376 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top());
3377 }
3378 }
3379
3380 // Step 4: update dominator tree and dominator depth
3381
3382 set_idom(head, orig_tail_clone, dd);
3383 recompute_dom_depth();
3384
3385 // Inhibit more partial peeling on this loop
3386 new_head_clone->set_partial_peel_loop();
3387 C->set_major_progress();
3388 loop->record_for_igvn();
3389
3390 #ifndef PRODUCT
3391 if (TracePartialPeeling) {
3392 tty->print_cr("\nafter partial peel one iteration");
3393 Node_List wl(area);
3394 Node* t = last_peel;
3395 while (true) {
3396 wl.push(t);
3397 if (t == head_clone) break;
3398 t = idom(t);
3399 }
3400 while (wl.size() > 0) {
3401 Node* tt = wl.pop();
3402 if (tt == head) tty->print_cr("orig head");
3403 else if (tt == new_head_clone) tty->print_cr("new head");
3404 else if (tt == head_clone) tty->print_cr("clone head");
3405 tt->dump();
3406 }
3407 }
3408 #endif
3409 return true;
3410 }
3411
3412 //------------------------------reorg_offsets----------------------------------
3413 // Reorganize offset computations to lower register pressure. Mostly
3414 // prevent loop-fallout uses of the pre-incremented trip counter (which are
3415 // then alive with the post-incremented trip counter forcing an extra
3416 // register move)
3417 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) {
3418 // Perform it only for canonical counted loops.
3419 // Loop's shape could be messed up by iteration_split_impl.
3420 if (!loop->_head->is_CountedLoop())
3421 return;
3422 if (!loop->_head->as_Loop()->is_valid_counted_loop())
3423 return;
3424
3425 CountedLoopNode *cl = loop->_head->as_CountedLoop();
3426 CountedLoopEndNode *cle = cl->loopexit();
3427 Node *exit = cle->proj_out(false);
3428 Node *phi = cl->phi();
3429
3430 // Check for the special case when using the pre-incremented trip-counter on
3431 // the fall-out path (forces the pre-incremented and post-incremented trip
3432 // counter to be live at the same time). Fix this by adjusting to use the
3433 // post-increment trip counter.
3434
3435 bool progress = true;
3436 while (progress) {
3437 progress = false;
3438 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
3439 Node* use = phi->fast_out(i); // User of trip-counter
3440 if (!has_ctrl(use)) continue;
3441 Node *u_ctrl = get_ctrl(use);
3442 if (use->is_Phi()) {
3443 u_ctrl = NULL;
3444 for (uint j = 1; j < use->req(); j++)
3445 if (use->in(j) == phi)
3446 u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j));
3447 }
3448 IdealLoopTree *u_loop = get_loop(u_ctrl);
3449 // Look for loop-invariant use
3450 if (u_loop == loop) continue;
3451 if (loop->is_member(u_loop)) continue;
3452 // Check that use is live out the bottom. Assuming the trip-counter
3453 // update is right at the bottom, uses of of the loop middle are ok.
3454 if (dom_lca(exit, u_ctrl) != exit) continue;
3455 // Hit! Refactor use to use the post-incremented tripcounter.
3456 // Compute a post-increment tripcounter.
3457 Node* c = exit;
3458 if (cl->is_strip_mined()) {
3459 IdealLoopTree* outer_loop = get_loop(cl->outer_loop());
3460 if (!outer_loop->is_member(u_loop)) {
3461 c = cl->outer_loop_exit();
3462 }
3463 }
3464 Node *opaq = new Opaque2Node(C, cle->incr());
3465 register_new_node(opaq, c);
3466 Node *neg_stride = _igvn.intcon(-cle->stride_con());
3467 set_ctrl(neg_stride, C->root());
3468 Node *post = new AddINode(opaq, neg_stride);
3469 register_new_node(post, c);
3470 _igvn.rehash_node_delayed(use);
3471 for (uint j = 1; j < use->req(); j++) {
3472 if (use->in(j) == phi)
3473 use->set_req(j, post);
3474 }
3475 // Since DU info changed, rerun loop
3476 progress = true;
3477 break;
3478 }
3479 }
3480
3481 }