1 /*
2 * Copyright (c) 2005, 2021, 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 "compiler/compileLog.hpp"
27 #include "gc/shared/collectedHeap.inline.hpp"
28 #include "gc/shared/tlab_globals.hpp"
29 #include "libadt/vectset.hpp"
30 #include "memory/universe.hpp"
31 #include "opto/addnode.hpp"
32 #include "opto/arraycopynode.hpp"
33 #include "opto/callnode.hpp"
34 #include "opto/castnode.hpp"
35 #include "opto/cfgnode.hpp"
36 #include "opto/compile.hpp"
37 #include "opto/convertnode.hpp"
38 #include "opto/graphKit.hpp"
39 #include "opto/intrinsicnode.hpp"
40 #include "opto/locknode.hpp"
41 #include "opto/loopnode.hpp"
42 #include "opto/macro.hpp"
43 #include "opto/memnode.hpp"
44 #include "opto/narrowptrnode.hpp"
45 #include "opto/node.hpp"
46 #include "opto/opaquenode.hpp"
47 #include "opto/phaseX.hpp"
48 #include "opto/rootnode.hpp"
49 #include "opto/runtime.hpp"
50 #include "opto/subnode.hpp"
51 #include "opto/subtypenode.hpp"
52 #include "opto/type.hpp"
53 #include "prims/jvmtiExport.hpp"
54 #include "runtime/sharedRuntime.hpp"
55 #include "utilities/macros.hpp"
56 #include "utilities/powerOfTwo.hpp"
57 #if INCLUDE_G1GC
58 #include "gc/g1/g1ThreadLocalData.hpp"
59 #endif // INCLUDE_G1GC
60
61
62 //
63 // Replace any references to "oldref" in inputs to "use" with "newref".
64 // Returns the number of replacements made.
65 //
66 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
67 int nreplacements = 0;
68 uint req = use->req();
69 for (uint j = 0; j < use->len(); j++) {
70 Node *uin = use->in(j);
71 if (uin == oldref) {
72 if (j < req)
73 use->set_req(j, newref);
74 else
75 use->set_prec(j, newref);
76 nreplacements++;
77 } else if (j >= req && uin == NULL) {
78 break;
79 }
80 }
81 return nreplacements;
82 }
83
84 void PhaseMacroExpand::migrate_outs(Node *old, Node *target) {
85 assert(old != NULL, "sanity");
86 for (DUIterator_Fast imax, i = old->fast_outs(imax); i < imax; i++) {
87 Node* use = old->fast_out(i);
88 _igvn.rehash_node_delayed(use);
89 imax -= replace_input(use, old, target);
90 // back up iterator
91 --i;
92 }
93 assert(old->outcnt() == 0, "all uses must be deleted");
94 }
95
96 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
97 Node* cmp;
98 if (mask != 0) {
99 Node* and_node = transform_later(new AndXNode(word, MakeConX(mask)));
100 cmp = transform_later(new CmpXNode(and_node, MakeConX(bits)));
101 } else {
102 cmp = word;
103 }
104 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
105 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
106 transform_later(iff);
107
108 // Fast path taken.
109 Node *fast_taken = transform_later(new IfFalseNode(iff));
110
111 // Fast path not-taken, i.e. slow path
112 Node *slow_taken = transform_later(new IfTrueNode(iff));
113
114 if (return_fast_path) {
115 region->init_req(edge, slow_taken); // Capture slow-control
116 return fast_taken;
117 } else {
118 region->init_req(edge, fast_taken); // Capture fast-control
119 return slow_taken;
120 }
121 }
122
123 //--------------------copy_predefined_input_for_runtime_call--------------------
124 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
125 // Set fixed predefined input arguments
126 call->init_req( TypeFunc::Control, ctrl );
127 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
128 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
129 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
130 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
131 }
132
133 //------------------------------make_slow_call---------------------------------
134 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type,
135 address slow_call, const char* leaf_name, Node* slow_path,
136 Node* parm0, Node* parm1, Node* parm2) {
137
138 // Slow-path call
139 CallNode *call = leaf_name
140 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
141 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), TypeRawPtr::BOTTOM );
142
143 // Slow path call has no side-effects, uses few values
144 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
145 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0);
146 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1);
147 if (parm2 != NULL) call->init_req(TypeFunc::Parms+2, parm2);
148 call->copy_call_debug_info(&_igvn, oldcall);
149 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
150 _igvn.replace_node(oldcall, call);
151 transform_later(call);
152
153 return call;
154 }
155
156 void PhaseMacroExpand::eliminate_gc_barrier(Node* p2x) {
157 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
158 bs->eliminate_gc_barrier(this, p2x);
159 }
160
161 // Search for a memory operation for the specified memory slice.
162 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
163 Node *orig_mem = mem;
164 Node *alloc_mem = alloc->in(TypeFunc::Memory);
165 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
166 while (true) {
167 if (mem == alloc_mem || mem == start_mem ) {
168 return mem; // hit one of our sentinels
169 } else if (mem->is_MergeMem()) {
170 mem = mem->as_MergeMem()->memory_at(alias_idx);
171 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
172 Node *in = mem->in(0);
173 // we can safely skip over safepoints, calls, locks and membars because we
174 // already know that the object is safe to eliminate.
175 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
176 return in;
177 } else if (in->is_Call()) {
178 CallNode *call = in->as_Call();
179 if (call->may_modify(tinst, phase)) {
180 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape");
181 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) {
182 return in;
183 }
184 }
185 mem = in->in(TypeFunc::Memory);
186 } else if (in->is_MemBar()) {
187 ArrayCopyNode* ac = NULL;
188 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) {
189 if (ac != NULL) {
190 assert(ac->is_clonebasic(), "Only basic clone is a non escaping clone");
191 return ac;
192 }
193 }
194 mem = in->in(TypeFunc::Memory);
195 } else {
196 #ifdef ASSERT
197 in->dump();
198 mem->dump();
199 assert(false, "unexpected projection");
200 #endif
201 }
202 } else if (mem->is_Store()) {
203 const TypePtr* atype = mem->as_Store()->adr_type();
204 int adr_idx = phase->C->get_alias_index(atype);
205 if (adr_idx == alias_idx) {
206 assert(atype->isa_oopptr(), "address type must be oopptr");
207 int adr_offset = atype->offset();
208 uint adr_iid = atype->is_oopptr()->instance_id();
209 // Array elements references have the same alias_idx
210 // but different offset and different instance_id.
211 if (adr_offset == offset && adr_iid == alloc->_idx) {
212 return mem;
213 }
214 } else {
215 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
216 }
217 mem = mem->in(MemNode::Memory);
218 } else if (mem->is_ClearArray()) {
219 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
220 // Can not bypass initialization of the instance
221 // we are looking.
222 debug_only(intptr_t offset;)
223 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
224 InitializeNode* init = alloc->as_Allocate()->initialization();
225 // We are looking for stored value, return Initialize node
226 // or memory edge from Allocate node.
227 if (init != NULL) {
228 return init;
229 } else {
230 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
231 }
232 }
233 // Otherwise skip it (the call updated 'mem' value).
234 } else if (mem->Opcode() == Op_SCMemProj) {
235 mem = mem->in(0);
236 Node* adr = NULL;
237 if (mem->is_LoadStore()) {
238 adr = mem->in(MemNode::Address);
239 } else {
240 assert(mem->Opcode() == Op_EncodeISOArray ||
241 mem->Opcode() == Op_StrCompressedCopy, "sanity");
242 adr = mem->in(3); // Destination array
243 }
244 const TypePtr* atype = adr->bottom_type()->is_ptr();
245 int adr_idx = phase->C->get_alias_index(atype);
246 if (adr_idx == alias_idx) {
247 DEBUG_ONLY(mem->dump();)
248 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
249 return NULL;
250 }
251 mem = mem->in(MemNode::Memory);
252 } else if (mem->Opcode() == Op_StrInflatedCopy) {
253 Node* adr = mem->in(3); // Destination array
254 const TypePtr* atype = adr->bottom_type()->is_ptr();
255 int adr_idx = phase->C->get_alias_index(atype);
256 if (adr_idx == alias_idx) {
257 DEBUG_ONLY(mem->dump();)
258 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
259 return NULL;
260 }
261 mem = mem->in(MemNode::Memory);
262 } else {
263 return mem;
264 }
265 assert(mem != orig_mem, "dead memory loop");
266 }
267 }
268
269 // Generate loads from source of the arraycopy for fields of
270 // destination needed at a deoptimization point
271 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) {
272 BasicType bt = ft;
273 const Type *type = ftype;
274 if (ft == T_NARROWOOP) {
275 bt = T_OBJECT;
276 type = ftype->make_oopptr();
277 }
278 Node* res = NULL;
279 if (ac->is_clonebasic()) {
280 assert(ac->in(ArrayCopyNode::Src) != ac->in(ArrayCopyNode::Dest), "clone source equals destination");
281 Node* base = ac->in(ArrayCopyNode::Src);
282 Node* adr = _igvn.transform(new AddPNode(base, base, MakeConX(offset)));
283 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
284 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
285 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
286 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
287 } else {
288 if (ac->modifies(offset, offset, &_igvn, true)) {
289 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
290 uint shift = exact_log2(type2aelembytes(bt));
291 Node* src_pos = ac->in(ArrayCopyNode::SrcPos);
292 Node* dest_pos = ac->in(ArrayCopyNode::DestPos);
293 const TypeInt* src_pos_t = _igvn.type(src_pos)->is_int();
294 const TypeInt* dest_pos_t = _igvn.type(dest_pos)->is_int();
295
296 Node* adr = NULL;
297 const TypePtr* adr_type = NULL;
298 if (src_pos_t->is_con() && dest_pos_t->is_con()) {
299 intptr_t off = ((src_pos_t->get_con() - dest_pos_t->get_con()) << shift) + offset;
300 Node* base = ac->in(ArrayCopyNode::Src);
301 adr = _igvn.transform(new AddPNode(base, base, MakeConX(off)));
302 adr_type = _igvn.type(base)->is_ptr()->add_offset(off);
303 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
304 // Don't emit a new load from src if src == dst but try to get the value from memory instead
305 return value_from_mem(ac->in(TypeFunc::Memory), ctl, ft, ftype, adr_type->isa_oopptr(), alloc);
306 }
307 } else {
308 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
309 #ifdef _LP64
310 diff = _igvn.transform(new ConvI2LNode(diff));
311 #endif
312 diff = _igvn.transform(new LShiftXNode(diff, intcon(shift)));
313
314 Node* off = _igvn.transform(new AddXNode(MakeConX(offset), diff));
315 Node* base = ac->in(ArrayCopyNode::Src);
316 adr = _igvn.transform(new AddPNode(base, base, off));
317 adr_type = _igvn.type(base)->is_ptr()->add_offset(Type::OffsetBot);
318 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
319 // Non constant offset in the array: we can't statically
320 // determine the value
321 return NULL;
322 }
323 }
324 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
325 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
326 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
327 }
328 }
329 if (res != NULL) {
330 if (ftype->isa_narrowoop()) {
331 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
332 res = _igvn.transform(new EncodePNode(res, ftype));
333 }
334 return res;
335 }
336 return NULL;
337 }
338
339 //
340 // Given a Memory Phi, compute a value Phi containing the values from stores
341 // on the input paths.
342 // Note: this function is recursive, its depth is limited by the "level" argument
343 // Returns the computed Phi, or NULL if it cannot compute it.
344 Node *PhaseMacroExpand::value_from_mem_phi(Node *mem, BasicType ft, const Type *phi_type, const TypeOopPtr *adr_t, AllocateNode *alloc, Node_Stack *value_phis, int level) {
345 assert(mem->is_Phi(), "sanity");
346 int alias_idx = C->get_alias_index(adr_t);
347 int offset = adr_t->offset();
348 int instance_id = adr_t->instance_id();
349
350 // Check if an appropriate value phi already exists.
351 Node* region = mem->in(0);
352 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
353 Node* phi = region->fast_out(k);
354 if (phi->is_Phi() && phi != mem &&
355 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
356 return phi;
357 }
358 }
359 // Check if an appropriate new value phi already exists.
360 Node* new_phi = value_phis->find(mem->_idx);
361 if (new_phi != NULL)
362 return new_phi;
363
364 if (level <= 0) {
365 return NULL; // Give up: phi tree too deep
366 }
367 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
368 Node *alloc_mem = alloc->in(TypeFunc::Memory);
369
370 uint length = mem->req();
371 GrowableArray <Node *> values(length, length, NULL);
372
373 // create a new Phi for the value
374 PhiNode *phi = new PhiNode(mem->in(0), phi_type, NULL, mem->_idx, instance_id, alias_idx, offset);
375 transform_later(phi);
376 value_phis->push(phi, mem->_idx);
377
378 for (uint j = 1; j < length; j++) {
379 Node *in = mem->in(j);
380 if (in == NULL || in->is_top()) {
381 values.at_put(j, in);
382 } else {
383 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
384 if (val == start_mem || val == alloc_mem) {
385 // hit a sentinel, return appropriate 0 value
386 values.at_put(j, _igvn.zerocon(ft));
387 continue;
388 }
389 if (val->is_Initialize()) {
390 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
391 }
392 if (val == NULL) {
393 return NULL; // can't find a value on this path
394 }
395 if (val == mem) {
396 values.at_put(j, mem);
397 } else if (val->is_Store()) {
398 Node* n = val->in(MemNode::ValueIn);
399 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
400 n = bs->step_over_gc_barrier(n);
401 if (is_subword_type(ft)) {
402 n = Compile::narrow_value(ft, n, phi_type, &_igvn, true);
403 }
404 values.at_put(j, n);
405 } else if(val->is_Proj() && val->in(0) == alloc) {
406 values.at_put(j, _igvn.zerocon(ft));
407 } else if (val->is_Phi()) {
408 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
409 if (val == NULL) {
410 return NULL;
411 }
412 values.at_put(j, val);
413 } else if (val->Opcode() == Op_SCMemProj) {
414 assert(val->in(0)->is_LoadStore() ||
415 val->in(0)->Opcode() == Op_EncodeISOArray ||
416 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
417 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
418 return NULL;
419 } else if (val->is_ArrayCopy()) {
420 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc);
421 if (res == NULL) {
422 return NULL;
423 }
424 values.at_put(j, res);
425 } else {
426 DEBUG_ONLY( val->dump(); )
427 assert(false, "unknown node on this path");
428 return NULL; // unknown node on this path
429 }
430 }
431 }
432 // Set Phi's inputs
433 for (uint j = 1; j < length; j++) {
434 if (values.at(j) == mem) {
435 phi->init_req(j, phi);
436 } else {
437 phi->init_req(j, values.at(j));
438 }
439 }
440 return phi;
441 }
442
443 // Search the last value stored into the object's field.
444 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
445 assert(adr_t->is_known_instance_field(), "instance required");
446 int instance_id = adr_t->instance_id();
447 assert((uint)instance_id == alloc->_idx, "wrong allocation");
448
449 int alias_idx = C->get_alias_index(adr_t);
450 int offset = adr_t->offset();
451 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
452 Node *alloc_ctrl = alloc->in(TypeFunc::Control);
453 Node *alloc_mem = alloc->in(TypeFunc::Memory);
454 VectorSet visited;
455
456 bool done = sfpt_mem == alloc_mem;
457 Node *mem = sfpt_mem;
458 while (!done) {
459 if (visited.test_set(mem->_idx)) {
460 return NULL; // found a loop, give up
461 }
462 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
463 if (mem == start_mem || mem == alloc_mem) {
464 done = true; // hit a sentinel, return appropriate 0 value
465 } else if (mem->is_Initialize()) {
466 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
467 if (mem == NULL) {
468 done = true; // Something go wrong.
469 } else if (mem->is_Store()) {
470 const TypePtr* atype = mem->as_Store()->adr_type();
471 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
472 done = true;
473 }
474 } else if (mem->is_Store()) {
475 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
476 assert(atype != NULL, "address type must be oopptr");
477 assert(C->get_alias_index(atype) == alias_idx &&
478 atype->is_known_instance_field() && atype->offset() == offset &&
479 atype->instance_id() == instance_id, "store is correct memory slice");
480 done = true;
481 } else if (mem->is_Phi()) {
482 // try to find a phi's unique input
483 Node *unique_input = NULL;
484 Node *top = C->top();
485 for (uint i = 1; i < mem->req(); i++) {
486 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
487 if (n == NULL || n == top || n == mem) {
488 continue;
489 } else if (unique_input == NULL) {
490 unique_input = n;
491 } else if (unique_input != n) {
492 unique_input = top;
493 break;
494 }
495 }
496 if (unique_input != NULL && unique_input != top) {
497 mem = unique_input;
498 } else {
499 done = true;
500 }
501 } else if (mem->is_ArrayCopy()) {
502 done = true;
503 } else {
504 DEBUG_ONLY( mem->dump(); )
505 assert(false, "unexpected node");
506 }
507 }
508 if (mem != NULL) {
509 if (mem == start_mem || mem == alloc_mem) {
510 // hit a sentinel, return appropriate 0 value
511 return _igvn.zerocon(ft);
512 } else if (mem->is_Store()) {
513 Node* n = mem->in(MemNode::ValueIn);
514 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
515 n = bs->step_over_gc_barrier(n);
516 return n;
517 } else if (mem->is_Phi()) {
518 // attempt to produce a Phi reflecting the values on the input paths of the Phi
519 Node_Stack value_phis(8);
520 Node* phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
521 if (phi != NULL) {
522 return phi;
523 } else {
524 // Kill all new Phis
525 while(value_phis.is_nonempty()) {
526 Node* n = value_phis.node();
527 _igvn.replace_node(n, C->top());
528 value_phis.pop();
529 }
530 }
531 } else if (mem->is_ArrayCopy()) {
532 Node* ctl = mem->in(0);
533 Node* m = mem->in(TypeFunc::Memory);
534 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
535 // pin the loads in the uncommon trap path
536 ctl = sfpt_ctl;
537 m = sfpt_mem;
538 }
539 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc);
540 }
541 }
542 // Something go wrong.
543 return NULL;
544 }
545
546 // Check the possibility of scalar replacement.
547 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
548 // Scan the uses of the allocation to check for anything that would
549 // prevent us from eliminating it.
550 NOT_PRODUCT( const char* fail_eliminate = NULL; )
551 DEBUG_ONLY( Node* disq_node = NULL; )
552 bool can_eliminate = true;
553
554 Node* res = alloc->result_cast();
555 const TypeOopPtr* res_type = NULL;
556 if (res == NULL) {
557 // All users were eliminated.
558 } else if (!res->is_CheckCastPP()) {
559 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
560 can_eliminate = false;
561 } else {
562 res_type = _igvn.type(res)->isa_oopptr();
563 if (res_type == NULL) {
564 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
565 can_eliminate = false;
566 } else if (res_type->isa_aryptr()) {
567 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
568 if (length < 0) {
569 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
570 can_eliminate = false;
571 }
572 }
573 }
574
575 if (can_eliminate && res != NULL) {
576 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
577 for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
578 j < jmax && can_eliminate; j++) {
579 Node* use = res->fast_out(j);
580
581 if (use->is_AddP()) {
582 const TypePtr* addp_type = _igvn.type(use)->is_ptr();
583 int offset = addp_type->offset();
584
585 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
586 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";)
587 can_eliminate = false;
588 break;
589 }
590 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
591 k < kmax && can_eliminate; k++) {
592 Node* n = use->fast_out(k);
593 if (!n->is_Store() && n->Opcode() != Op_CastP2X && !bs->is_gc_pre_barrier_node(n)) {
594 DEBUG_ONLY(disq_node = n;)
595 if (n->is_Load() || n->is_LoadStore()) {
596 NOT_PRODUCT(fail_eliminate = "Field load";)
597 } else {
598 NOT_PRODUCT(fail_eliminate = "Not store field referrence";)
599 }
600 can_eliminate = false;
601 }
602 }
603 } else if (use->is_ArrayCopy() &&
604 (use->as_ArrayCopy()->is_clonebasic() ||
605 use->as_ArrayCopy()->is_arraycopy_validated() ||
606 use->as_ArrayCopy()->is_copyof_validated() ||
607 use->as_ArrayCopy()->is_copyofrange_validated()) &&
608 use->in(ArrayCopyNode::Dest) == res) {
609 // ok to eliminate
610 } else if (use->is_SafePoint()) {
611 SafePointNode* sfpt = use->as_SafePoint();
612 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
613 // Object is passed as argument.
614 DEBUG_ONLY(disq_node = use;)
615 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
616 can_eliminate = false;
617 }
618 Node* sfptMem = sfpt->memory();
619 if (sfptMem == NULL || sfptMem->is_top()) {
620 DEBUG_ONLY(disq_node = use;)
621 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
622 can_eliminate = false;
623 } else {
624 safepoints.append_if_missing(sfpt);
625 }
626 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
627 if (use->is_Phi()) {
628 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
629 NOT_PRODUCT(fail_eliminate = "Object is return value";)
630 } else {
631 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
632 }
633 DEBUG_ONLY(disq_node = use;)
634 } else {
635 if (use->Opcode() == Op_Return) {
636 NOT_PRODUCT(fail_eliminate = "Object is return value";)
637 }else {
638 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
639 }
640 DEBUG_ONLY(disq_node = use;)
641 }
642 can_eliminate = false;
643 }
644 }
645 }
646
647 #ifndef PRODUCT
648 if (PrintEliminateAllocations) {
649 if (can_eliminate) {
650 tty->print("Scalar ");
651 if (res == NULL)
652 alloc->dump();
653 else
654 res->dump();
655 } else if (alloc->_is_scalar_replaceable) {
656 tty->print("NotScalar (%s)", fail_eliminate);
657 if (res == NULL)
658 alloc->dump();
659 else
660 res->dump();
661 #ifdef ASSERT
662 if (disq_node != NULL) {
663 tty->print(" >>>> ");
664 disq_node->dump();
665 }
666 #endif /*ASSERT*/
667 }
668 }
669 #endif
670 return can_eliminate;
671 }
672
673 // Do scalar replacement.
674 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
675 GrowableArray <SafePointNode *> safepoints_done;
676
677 ciKlass* klass = NULL;
678 ciInstanceKlass* iklass = NULL;
679 int nfields = 0;
680 int array_base = 0;
681 int element_size = 0;
682 BasicType basic_elem_type = T_ILLEGAL;
683 ciType* elem_type = NULL;
684
685 Node* res = alloc->result_cast();
686 assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result");
687 const TypeOopPtr* res_type = NULL;
688 if (res != NULL) { // Could be NULL when there are no users
689 res_type = _igvn.type(res)->isa_oopptr();
690 }
691
692 if (res != NULL) {
693 klass = res_type->klass();
694 if (res_type->isa_instptr()) {
695 // find the fields of the class which will be needed for safepoint debug information
696 assert(klass->is_instance_klass(), "must be an instance klass.");
697 iklass = klass->as_instance_klass();
698 nfields = iklass->nof_nonstatic_fields();
699 } else {
700 // find the array's elements which will be needed for safepoint debug information
701 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
702 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass.");
703 elem_type = klass->as_array_klass()->element_type();
704 basic_elem_type = elem_type->basic_type();
705 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
706 element_size = type2aelembytes(basic_elem_type);
707 }
708 }
709 //
710 // Process the safepoint uses
711 //
712 while (safepoints.length() > 0) {
713 SafePointNode* sfpt = safepoints.pop();
714 Node* mem = sfpt->memory();
715 Node* ctl = sfpt->control();
716 assert(sfpt->jvms() != NULL, "missed JVMS");
717 // Fields of scalar objs are referenced only at the end
718 // of regular debuginfo at the last (youngest) JVMS.
719 // Record relative start index.
720 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
721 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type,
722 #ifdef ASSERT
723 alloc,
724 #endif
725 first_ind, nfields);
726 sobj->init_req(0, C->root());
727 transform_later(sobj);
728
729 // Scan object's fields adding an input to the safepoint for each field.
730 for (int j = 0; j < nfields; j++) {
731 intptr_t offset;
732 ciField* field = NULL;
733 if (iklass != NULL) {
734 field = iklass->nonstatic_field_at(j);
735 offset = field->offset();
736 elem_type = field->type();
737 basic_elem_type = field->layout_type();
738 } else {
739 offset = array_base + j * (intptr_t)element_size;
740 }
741
742 const Type *field_type;
743 // The next code is taken from Parse::do_get_xxx().
744 if (is_reference_type(basic_elem_type)) {
745 if (!elem_type->is_loaded()) {
746 field_type = TypeInstPtr::BOTTOM;
747 } else if (field != NULL && field->is_static_constant()) {
748 // This can happen if the constant oop is non-perm.
749 ciObject* con = field->constant_value().as_object();
750 // Do not "join" in the previous type; it doesn't add value,
751 // and may yield a vacuous result if the field is of interface type.
752 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
753 assert(field_type != NULL, "field singleton type must be consistent");
754 } else {
755 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
756 }
757 if (UseCompressedOops) {
758 field_type = field_type->make_narrowoop();
759 basic_elem_type = T_NARROWOOP;
760 }
761 } else {
762 field_type = Type::get_const_basic_type(basic_elem_type);
763 }
764
765 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr();
766
767 Node *field_val = value_from_mem(mem, ctl, basic_elem_type, field_type, field_addr_type, alloc);
768 if (field_val == NULL) {
769 // We weren't able to find a value for this field,
770 // give up on eliminating this allocation.
771
772 // Remove any extra entries we added to the safepoint.
773 uint last = sfpt->req() - 1;
774 for (int k = 0; k < j; k++) {
775 sfpt->del_req(last--);
776 }
777 _igvn._worklist.push(sfpt);
778 // rollback processed safepoints
779 while (safepoints_done.length() > 0) {
780 SafePointNode* sfpt_done = safepoints_done.pop();
781 // remove any extra entries we added to the safepoint
782 last = sfpt_done->req() - 1;
783 for (int k = 0; k < nfields; k++) {
784 sfpt_done->del_req(last--);
785 }
786 JVMState *jvms = sfpt_done->jvms();
787 jvms->set_endoff(sfpt_done->req());
788 // Now make a pass over the debug information replacing any references
789 // to SafePointScalarObjectNode with the allocated object.
790 int start = jvms->debug_start();
791 int end = jvms->debug_end();
792 for (int i = start; i < end; i++) {
793 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
794 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
795 if (scobj->first_index(jvms) == sfpt_done->req() &&
796 scobj->n_fields() == (uint)nfields) {
797 assert(scobj->alloc() == alloc, "sanity");
798 sfpt_done->set_req(i, res);
799 }
800 }
801 }
802 _igvn._worklist.push(sfpt_done);
803 }
804 #ifndef PRODUCT
805 if (PrintEliminateAllocations) {
806 if (field != NULL) {
807 tty->print("=== At SafePoint node %d can't find value of Field: ",
808 sfpt->_idx);
809 field->print();
810 int field_idx = C->get_alias_index(field_addr_type);
811 tty->print(" (alias_idx=%d)", field_idx);
812 } else { // Array's element
813 tty->print("=== At SafePoint node %d can't find value of array element [%d]",
814 sfpt->_idx, j);
815 }
816 tty->print(", which prevents elimination of: ");
817 if (res == NULL)
818 alloc->dump();
819 else
820 res->dump();
821 }
822 #endif
823 return false;
824 }
825 if (UseCompressedOops && field_type->isa_narrowoop()) {
826 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
827 // to be able scalar replace the allocation.
828 if (field_val->is_EncodeP()) {
829 field_val = field_val->in(1);
830 } else {
831 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
832 }
833 }
834 sfpt->add_req(field_val);
835 }
836 JVMState *jvms = sfpt->jvms();
837 jvms->set_endoff(sfpt->req());
838 // Now make a pass over the debug information replacing any references
839 // to the allocated object with "sobj"
840 int start = jvms->debug_start();
841 int end = jvms->debug_end();
842 sfpt->replace_edges_in_range(res, sobj, start, end, &_igvn);
843 _igvn._worklist.push(sfpt);
844 safepoints_done.append_if_missing(sfpt); // keep it for rollback
845 }
846 return true;
847 }
848
849 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
850 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
851 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
852 if (ctl_proj != NULL) {
853 igvn.replace_node(ctl_proj, n->in(0));
854 }
855 if (mem_proj != NULL) {
856 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
857 }
858 }
859
860 // Process users of eliminated allocation.
861 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) {
862 Node* res = alloc->result_cast();
863 if (res != NULL) {
864 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
865 Node *use = res->last_out(j);
866 uint oc1 = res->outcnt();
867
868 if (use->is_AddP()) {
869 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
870 Node *n = use->last_out(k);
871 uint oc2 = use->outcnt();
872 if (n->is_Store()) {
873 #ifdef ASSERT
874 // Verify that there is no dependent MemBarVolatile nodes,
875 // they should be removed during IGVN, see MemBarNode::Ideal().
876 for (DUIterator_Fast pmax, p = n->fast_outs(pmax);
877 p < pmax; p++) {
878 Node* mb = n->fast_out(p);
879 assert(mb->is_Initialize() || !mb->is_MemBar() ||
880 mb->req() <= MemBarNode::Precedent ||
881 mb->in(MemBarNode::Precedent) != n,
882 "MemBarVolatile should be eliminated for non-escaping object");
883 }
884 #endif
885 _igvn.replace_node(n, n->in(MemNode::Memory));
886 } else {
887 eliminate_gc_barrier(n);
888 }
889 k -= (oc2 - use->outcnt());
890 }
891 _igvn.remove_dead_node(use);
892 } else if (use->is_ArrayCopy()) {
893 // Disconnect ArrayCopy node
894 ArrayCopyNode* ac = use->as_ArrayCopy();
895 if (ac->is_clonebasic()) {
896 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
897 disconnect_projections(ac, _igvn);
898 assert(alloc->in(TypeFunc::Memory)->is_Proj() && alloc->in(TypeFunc::Memory)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
899 Node* membar_before = alloc->in(TypeFunc::Memory)->in(0);
900 disconnect_projections(membar_before->as_MemBar(), _igvn);
901 if (membar_after->is_MemBar()) {
902 disconnect_projections(membar_after->as_MemBar(), _igvn);
903 }
904 } else {
905 assert(ac->is_arraycopy_validated() ||
906 ac->is_copyof_validated() ||
907 ac->is_copyofrange_validated(), "unsupported");
908 CallProjections callprojs;
909 ac->extract_projections(&callprojs, true);
910
911 _igvn.replace_node(callprojs.fallthrough_ioproj, ac->in(TypeFunc::I_O));
912 _igvn.replace_node(callprojs.fallthrough_memproj, ac->in(TypeFunc::Memory));
913 _igvn.replace_node(callprojs.fallthrough_catchproj, ac->in(TypeFunc::Control));
914
915 // Set control to top. IGVN will remove the remaining projections
916 ac->set_req(0, top());
917 ac->replace_edge(res, top(), &_igvn);
918
919 // Disconnect src right away: it can help find new
920 // opportunities for allocation elimination
921 Node* src = ac->in(ArrayCopyNode::Src);
922 ac->replace_edge(src, top(), &_igvn);
923 // src can be top at this point if src and dest of the
924 // arraycopy were the same
925 if (src->outcnt() == 0 && !src->is_top()) {
926 _igvn.remove_dead_node(src);
927 }
928 }
929 _igvn._worklist.push(ac);
930 } else {
931 eliminate_gc_barrier(use);
932 }
933 j -= (oc1 - res->outcnt());
934 }
935 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
936 _igvn.remove_dead_node(res);
937 }
938
939 //
940 // Process other users of allocation's projections
941 //
942 if (_callprojs.resproj != NULL && _callprojs.resproj->outcnt() != 0) {
943 // First disconnect stores captured by Initialize node.
944 // If Initialize node is eliminated first in the following code,
945 // it will kill such stores and DUIterator_Last will assert.
946 for (DUIterator_Fast jmax, j = _callprojs.resproj->fast_outs(jmax); j < jmax; j++) {
947 Node* use = _callprojs.resproj->fast_out(j);
948 if (use->is_AddP()) {
949 // raw memory addresses used only by the initialization
950 _igvn.replace_node(use, C->top());
951 --j; --jmax;
952 }
953 }
954 for (DUIterator_Last jmin, j = _callprojs.resproj->last_outs(jmin); j >= jmin; ) {
955 Node* use = _callprojs.resproj->last_out(j);
956 uint oc1 = _callprojs.resproj->outcnt();
957 if (use->is_Initialize()) {
958 // Eliminate Initialize node.
959 InitializeNode *init = use->as_Initialize();
960 assert(init->outcnt() <= 2, "only a control and memory projection expected");
961 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
962 if (ctrl_proj != NULL) {
963 _igvn.replace_node(ctrl_proj, init->in(TypeFunc::Control));
964 #ifdef ASSERT
965 // If the InitializeNode has no memory out, it will die, and tmp will become NULL
966 Node* tmp = init->in(TypeFunc::Control);
967 assert(tmp == NULL || tmp == _callprojs.fallthrough_catchproj, "allocation control projection");
968 #endif
969 }
970 Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory);
971 if (mem_proj != NULL) {
972 Node *mem = init->in(TypeFunc::Memory);
973 #ifdef ASSERT
974 if (mem->is_MergeMem()) {
975 assert(mem->in(TypeFunc::Memory) == _callprojs.fallthrough_memproj, "allocation memory projection");
976 } else {
977 assert(mem == _callprojs.fallthrough_memproj, "allocation memory projection");
978 }
979 #endif
980 _igvn.replace_node(mem_proj, mem);
981 }
982 } else {
983 assert(false, "only Initialize or AddP expected");
984 }
985 j -= (oc1 - _callprojs.resproj->outcnt());
986 }
987 }
988 if (_callprojs.fallthrough_catchproj != NULL) {
989 _igvn.replace_node(_callprojs.fallthrough_catchproj, alloc->in(TypeFunc::Control));
990 }
991 if (_callprojs.fallthrough_memproj != NULL) {
992 _igvn.replace_node(_callprojs.fallthrough_memproj, alloc->in(TypeFunc::Memory));
993 }
994 if (_callprojs.catchall_memproj != NULL) {
995 _igvn.replace_node(_callprojs.catchall_memproj, C->top());
996 }
997 if (_callprojs.fallthrough_ioproj != NULL) {
998 _igvn.replace_node(_callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
999 }
1000 if (_callprojs.catchall_ioproj != NULL) {
1001 _igvn.replace_node(_callprojs.catchall_ioproj, C->top());
1002 }
1003 if (_callprojs.catchall_catchproj != NULL) {
1004 _igvn.replace_node(_callprojs.catchall_catchproj, C->top());
1005 }
1006 }
1007
1008 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1009 // If reallocation fails during deoptimization we'll pop all
1010 // interpreter frames for this compiled frame and that won't play
1011 // nice with JVMTI popframe.
1012 // We avoid this issue by eager reallocation when the popframe request
1013 // is received.
1014 if (!EliminateAllocations || !alloc->_is_non_escaping) {
1015 return false;
1016 }
1017 Node* klass = alloc->in(AllocateNode::KlassNode);
1018 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1019 Node* res = alloc->result_cast();
1020 // Eliminate boxing allocations which are not used
1021 // regardless scalar replacable status.
1022 bool boxing_alloc = C->eliminate_boxing() &&
1023 tklass->klass()->is_instance_klass() &&
1024 tklass->klass()->as_instance_klass()->is_box_klass();
1025 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) {
1026 return false;
1027 }
1028
1029 alloc->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1030
1031 GrowableArray <SafePointNode *> safepoints;
1032 if (!can_eliminate_allocation(alloc, safepoints)) {
1033 return false;
1034 }
1035
1036 if (!alloc->_is_scalar_replaceable) {
1037 assert(res == NULL, "sanity");
1038 // We can only eliminate allocation if all debug info references
1039 // are already replaced with SafePointScalarObject because
1040 // we can't search for a fields value without instance_id.
1041 if (safepoints.length() > 0) {
1042 return false;
1043 }
1044 }
1045
1046 if (!scalar_replacement(alloc, safepoints)) {
1047 return false;
1048 }
1049
1050 CompileLog* log = C->log();
1051 if (log != NULL) {
1052 log->head("eliminate_allocation type='%d'",
1053 log->identify(tklass->klass()));
1054 JVMState* p = alloc->jvms();
1055 while (p != NULL) {
1056 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1057 p = p->caller();
1058 }
1059 log->tail("eliminate_allocation");
1060 }
1061
1062 process_users_of_allocation(alloc);
1063
1064 #ifndef PRODUCT
1065 if (PrintEliminateAllocations) {
1066 if (alloc->is_AllocateArray())
1067 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1068 else
1069 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1070 }
1071 #endif
1072
1073 return true;
1074 }
1075
1076 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1077 // EA should remove all uses of non-escaping boxing node.
1078 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != NULL) {
1079 return false;
1080 }
1081
1082 assert(boxing->result_cast() == NULL, "unexpected boxing node result");
1083
1084 boxing->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1085
1086 const TypeTuple* r = boxing->tf()->range();
1087 assert(r->cnt() > TypeFunc::Parms, "sanity");
1088 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1089 assert(t != NULL, "sanity");
1090
1091 CompileLog* log = C->log();
1092 if (log != NULL) {
1093 log->head("eliminate_boxing type='%d'",
1094 log->identify(t->klass()));
1095 JVMState* p = boxing->jvms();
1096 while (p != NULL) {
1097 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1098 p = p->caller();
1099 }
1100 log->tail("eliminate_boxing");
1101 }
1102
1103 process_users_of_allocation(boxing);
1104
1105 #ifndef PRODUCT
1106 if (PrintEliminateAllocations) {
1107 tty->print("++++ Eliminated: %d ", boxing->_idx);
1108 boxing->method()->print_short_name(tty);
1109 tty->cr();
1110 }
1111 #endif
1112
1113 return true;
1114 }
1115
1116 //---------------------------set_eden_pointers-------------------------
1117 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) {
1118 if (UseTLAB) { // Private allocation: load from TLS
1119 Node* thread = transform_later(new ThreadLocalNode());
1120 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset());
1121 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset());
1122 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset);
1123 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset);
1124 } else { // Shared allocation: load from globals
1125 CollectedHeap* ch = Universe::heap();
1126 address top_adr = (address)ch->top_addr();
1127 address end_adr = (address)ch->end_addr();
1128 eden_top_adr = makecon(TypeRawPtr::make(top_adr));
1129 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr);
1130 }
1131 }
1132
1133
1134 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1135 Node* adr = basic_plus_adr(base, offset);
1136 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1137 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered);
1138 transform_later(value);
1139 return value;
1140 }
1141
1142
1143 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1144 Node* adr = basic_plus_adr(base, offset);
1145 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered);
1146 transform_later(mem);
1147 return mem;
1148 }
1149
1150 //=============================================================================
1151 //
1152 // A L L O C A T I O N
1153 //
1154 // Allocation attempts to be fast in the case of frequent small objects.
1155 // It breaks down like this:
1156 //
1157 // 1) Size in doublewords is computed. This is a constant for objects and
1158 // variable for most arrays. Doubleword units are used to avoid size
1159 // overflow of huge doubleword arrays. We need doublewords in the end for
1160 // rounding.
1161 //
1162 // 2) Size is checked for being 'too large'. Too-large allocations will go
1163 // the slow path into the VM. The slow path can throw any required
1164 // exceptions, and does all the special checks for very large arrays. The
1165 // size test can constant-fold away for objects. For objects with
1166 // finalizers it constant-folds the otherway: you always go slow with
1167 // finalizers.
1168 //
1169 // 3) If NOT using TLABs, this is the contended loop-back point.
1170 // Load-Locked the heap top. If using TLABs normal-load the heap top.
1171 //
1172 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
1173 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
1174 // "size*8" we always enter the VM, where "largish" is a constant picked small
1175 // enough that there's always space between the eden max and 4Gig (old space is
1176 // there so it's quite large) and large enough that the cost of entering the VM
1177 // is dwarfed by the cost to initialize the space.
1178 //
1179 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1180 // down. If contended, repeat at step 3. If using TLABs normal-store
1181 // adjusted heap top back down; there is no contention.
1182 //
1183 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
1184 // fields.
1185 //
1186 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1187 // oop flavor.
1188 //
1189 //=============================================================================
1190 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1191 // Allocations bigger than this always go the slow route.
1192 // This value must be small enough that allocation attempts that need to
1193 // trigger exceptions go the slow route. Also, it must be small enough so
1194 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1195 //=============================================================================j//
1196 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1197 // The allocator will coalesce int->oop copies away. See comment in
1198 // coalesce.cpp about how this works. It depends critically on the exact
1199 // code shape produced here, so if you are changing this code shape
1200 // make sure the GC info for the heap-top is correct in and around the
1201 // slow-path call.
1202 //
1203
1204 void PhaseMacroExpand::expand_allocate_common(
1205 AllocateNode* alloc, // allocation node to be expanded
1206 Node* length, // array length for an array allocation
1207 const TypeFunc* slow_call_type, // Type of slow call
1208 address slow_call_address, // Address of slow call
1209 Node* valid_length_test // whether length is valid or not
1210 )
1211 {
1212 Node* ctrl = alloc->in(TypeFunc::Control);
1213 Node* mem = alloc->in(TypeFunc::Memory);
1214 Node* i_o = alloc->in(TypeFunc::I_O);
1215 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1216 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1217 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1218 assert(ctrl != NULL, "must have control");
1219
1220 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1221 // they will not be used if "always_slow" is set
1222 enum { slow_result_path = 1, fast_result_path = 2 };
1223 Node *result_region = NULL;
1224 Node *result_phi_rawmem = NULL;
1225 Node *result_phi_rawoop = NULL;
1226 Node *result_phi_i_o = NULL;
1227
1228 // The initial slow comparison is a size check, the comparison
1229 // we want to do is a BoolTest::gt
1230 bool expand_fast_path = true;
1231 int tv = _igvn.find_int_con(initial_slow_test, -1);
1232 if (tv >= 0) {
1233 // InitialTest has constant result
1234 // 0 - can fit in TLAB
1235 // 1 - always too big or negative
1236 assert(tv <= 1, "0 or 1 if a constant");
1237 expand_fast_path = (tv == 0);
1238 initial_slow_test = NULL;
1239 } else {
1240 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1241 }
1242
1243 if (C->env()->dtrace_alloc_probes() ||
1244 (!UseTLAB && !Universe::heap()->supports_inline_contig_alloc())) {
1245 // Force slow-path allocation
1246 expand_fast_path = false;
1247 initial_slow_test = NULL;
1248 }
1249
1250 bool allocation_has_use = (alloc->result_cast() != NULL);
1251 if (!allocation_has_use) {
1252 InitializeNode* init = alloc->initialization();
1253 if (init != NULL) {
1254 init->remove(&_igvn);
1255 }
1256 if (expand_fast_path && (initial_slow_test == NULL)) {
1257 // Remove allocation node and return.
1258 // Size is a non-negative constant -> no initial check needed -> directly to fast path.
1259 // Also, no usages -> empty fast path -> no fall out to slow path -> nothing left.
1260 #ifndef PRODUCT
1261 if (PrintEliminateAllocations) {
1262 tty->print("NotUsed ");
1263 Node* res = alloc->proj_out_or_null(TypeFunc::Parms);
1264 if (res != NULL) {
1265 res->dump();
1266 } else {
1267 alloc->dump();
1268 }
1269 }
1270 #endif
1271 yank_alloc_node(alloc);
1272 return;
1273 }
1274 }
1275
1276 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1277 Node *slow_region = NULL;
1278 Node *toobig_false = ctrl;
1279
1280 // generate the initial test if necessary
1281 if (initial_slow_test != NULL ) {
1282 assert (expand_fast_path, "Only need test if there is a fast path");
1283 slow_region = new RegionNode(3);
1284
1285 // Now make the initial failure test. Usually a too-big test but
1286 // might be a TRUE for finalizers or a fancy class check for
1287 // newInstance0.
1288 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1289 transform_later(toobig_iff);
1290 // Plug the failing-too-big test into the slow-path region
1291 Node *toobig_true = new IfTrueNode( toobig_iff );
1292 transform_later(toobig_true);
1293 slow_region ->init_req( too_big_or_final_path, toobig_true );
1294 toobig_false = new IfFalseNode( toobig_iff );
1295 transform_later(toobig_false);
1296 } else {
1297 // No initial test, just fall into next case
1298 assert(allocation_has_use || !expand_fast_path, "Should already have been handled");
1299 toobig_false = ctrl;
1300 debug_only(slow_region = NodeSentinel);
1301 }
1302
1303 // If we are here there are several possibilities
1304 // - expand_fast_path is false - then only a slow path is expanded. That's it.
1305 // no_initial_check means a constant allocation.
1306 // - If check always evaluates to false -> expand_fast_path is false (see above)
1307 // - If check always evaluates to true -> directly into fast path (but may bailout to slowpath)
1308 // if !allocation_has_use the fast path is empty
1309 // if !allocation_has_use && no_initial_check
1310 // - Then there are no fastpath that can fall out to slowpath -> no allocation code at all.
1311 // removed by yank_alloc_node above.
1312
1313 Node *slow_mem = mem; // save the current memory state for slow path
1314 // generate the fast allocation code unless we know that the initial test will always go slow
1315 if (expand_fast_path) {
1316 // Fast path modifies only raw memory.
1317 if (mem->is_MergeMem()) {
1318 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1319 }
1320
1321 // allocate the Region and Phi nodes for the result
1322 result_region = new RegionNode(3);
1323 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1324 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1325
1326 // Grab regular I/O before optional prefetch may change it.
1327 // Slow-path does no I/O so just set it to the original I/O.
1328 result_phi_i_o->init_req(slow_result_path, i_o);
1329
1330 // Name successful fast-path variables
1331 Node* fast_oop_ctrl;
1332 Node* fast_oop_rawmem;
1333 if (allocation_has_use) {
1334 Node* needgc_ctrl = NULL;
1335 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1336
1337 intx prefetch_lines = length != NULL ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1338 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1339 Node* fast_oop = bs->obj_allocate(this, mem, toobig_false, size_in_bytes, i_o, needgc_ctrl,
1340 fast_oop_ctrl, fast_oop_rawmem,
1341 prefetch_lines);
1342
1343 if (initial_slow_test != NULL) {
1344 // This completes all paths into the slow merge point
1345 slow_region->init_req(need_gc_path, needgc_ctrl);
1346 transform_later(slow_region);
1347 } else {
1348 // No initial slow path needed!
1349 // Just fall from the need-GC path straight into the VM call.
1350 slow_region = needgc_ctrl;
1351 }
1352
1353 InitializeNode* init = alloc->initialization();
1354 fast_oop_rawmem = initialize_object(alloc,
1355 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1356 klass_node, length, size_in_bytes);
1357 expand_initialize_membar(alloc, init, fast_oop_ctrl, fast_oop_rawmem);
1358 expand_dtrace_alloc_probe(alloc, fast_oop, fast_oop_ctrl, fast_oop_rawmem);
1359
1360 result_phi_rawoop->init_req(fast_result_path, fast_oop);
1361 } else {
1362 assert (initial_slow_test != NULL, "sanity");
1363 fast_oop_ctrl = toobig_false;
1364 fast_oop_rawmem = mem;
1365 transform_later(slow_region);
1366 }
1367
1368 // Plug in the successful fast-path into the result merge point
1369 result_region ->init_req(fast_result_path, fast_oop_ctrl);
1370 result_phi_i_o ->init_req(fast_result_path, i_o);
1371 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1372 } else {
1373 slow_region = ctrl;
1374 result_phi_i_o = i_o; // Rename it to use in the following code.
1375 }
1376
1377 // Generate slow-path call
1378 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1379 OptoRuntime::stub_name(slow_call_address),
1380 TypePtr::BOTTOM);
1381 call->init_req(TypeFunc::Control, slow_region);
1382 call->init_req(TypeFunc::I_O, top()); // does no i/o
1383 call->init_req(TypeFunc::Memory, slow_mem); // may gc ptrs
1384 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1385 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1386
1387 call->init_req(TypeFunc::Parms+0, klass_node);
1388 if (length != NULL) {
1389 call->init_req(TypeFunc::Parms+1, length);
1390 }
1391
1392 // Copy debug information and adjust JVMState information, then replace
1393 // allocate node with the call
1394 call->copy_call_debug_info(&_igvn, alloc);
1395 // For array allocations, copy the valid length check to the call node so Compile::final_graph_reshaping() can verify
1396 // that the call has the expected number of CatchProj nodes (in case the allocation always fails and the fallthrough
1397 // path dies).
1398 if (valid_length_test != NULL) {
1399 call->add_req(valid_length_test);
1400 }
1401 if (expand_fast_path) {
1402 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1403 } else {
1404 // Hook i_o projection to avoid its elimination during allocation
1405 // replacement (when only a slow call is generated).
1406 call->set_req(TypeFunc::I_O, result_phi_i_o);
1407 }
1408 _igvn.replace_node(alloc, call);
1409 transform_later(call);
1410
1411 // Identify the output projections from the allocate node and
1412 // adjust any references to them.
1413 // The control and io projections look like:
1414 //
1415 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1416 // Allocate Catch
1417 // ^---Proj(io) <-------+ ^---CatchProj(io)
1418 //
1419 // We are interested in the CatchProj nodes.
1420 //
1421 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1422
1423 // An allocate node has separate memory projections for the uses on
1424 // the control and i_o paths. Replace the control memory projection with
1425 // result_phi_rawmem (unless we are only generating a slow call when
1426 // both memory projections are combined)
1427 if (expand_fast_path && _callprojs.fallthrough_memproj != NULL) {
1428 migrate_outs(_callprojs.fallthrough_memproj, result_phi_rawmem);
1429 }
1430 // Now change uses of catchall_memproj to use fallthrough_memproj and delete
1431 // catchall_memproj so we end up with a call that has only 1 memory projection.
1432 if (_callprojs.catchall_memproj != NULL ) {
1433 if (_callprojs.fallthrough_memproj == NULL) {
1434 _callprojs.fallthrough_memproj = new ProjNode(call, TypeFunc::Memory);
1435 transform_later(_callprojs.fallthrough_memproj);
1436 }
1437 migrate_outs(_callprojs.catchall_memproj, _callprojs.fallthrough_memproj);
1438 _igvn.remove_dead_node(_callprojs.catchall_memproj);
1439 }
1440
1441 // An allocate node has separate i_o projections for the uses on the control
1442 // and i_o paths. Always replace the control i_o projection with result i_o
1443 // otherwise incoming i_o become dead when only a slow call is generated
1444 // (it is different from memory projections where both projections are
1445 // combined in such case).
1446 if (_callprojs.fallthrough_ioproj != NULL) {
1447 migrate_outs(_callprojs.fallthrough_ioproj, result_phi_i_o);
1448 }
1449 // Now change uses of catchall_ioproj to use fallthrough_ioproj and delete
1450 // catchall_ioproj so we end up with a call that has only 1 i_o projection.
1451 if (_callprojs.catchall_ioproj != NULL ) {
1452 if (_callprojs.fallthrough_ioproj == NULL) {
1453 _callprojs.fallthrough_ioproj = new ProjNode(call, TypeFunc::I_O);
1454 transform_later(_callprojs.fallthrough_ioproj);
1455 }
1456 migrate_outs(_callprojs.catchall_ioproj, _callprojs.fallthrough_ioproj);
1457 _igvn.remove_dead_node(_callprojs.catchall_ioproj);
1458 }
1459
1460 // if we generated only a slow call, we are done
1461 if (!expand_fast_path) {
1462 // Now we can unhook i_o.
1463 if (result_phi_i_o->outcnt() > 1) {
1464 call->set_req(TypeFunc::I_O, top());
1465 } else {
1466 assert(result_phi_i_o->unique_ctrl_out() == call, "sanity");
1467 // Case of new array with negative size known during compilation.
1468 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1469 // following code since call to runtime will throw exception.
1470 // As result there will be no users of i_o after the call.
1471 // Leave i_o attached to this call to avoid problems in preceding graph.
1472 }
1473 return;
1474 }
1475
1476 if (_callprojs.fallthrough_catchproj != NULL) {
1477 ctrl = _callprojs.fallthrough_catchproj->clone();
1478 transform_later(ctrl);
1479 _igvn.replace_node(_callprojs.fallthrough_catchproj, result_region);
1480 } else {
1481 ctrl = top();
1482 }
1483 Node *slow_result;
1484 if (_callprojs.resproj == NULL) {
1485 // no uses of the allocation result
1486 slow_result = top();
1487 } else {
1488 slow_result = _callprojs.resproj->clone();
1489 transform_later(slow_result);
1490 _igvn.replace_node(_callprojs.resproj, result_phi_rawoop);
1491 }
1492
1493 // Plug slow-path into result merge point
1494 result_region->init_req( slow_result_path, ctrl);
1495 transform_later(result_region);
1496 if (allocation_has_use) {
1497 result_phi_rawoop->init_req(slow_result_path, slow_result);
1498 transform_later(result_phi_rawoop);
1499 }
1500 result_phi_rawmem->init_req(slow_result_path, _callprojs.fallthrough_memproj);
1501 transform_later(result_phi_rawmem);
1502 transform_later(result_phi_i_o);
1503 // This completes all paths into the result merge point
1504 }
1505
1506 // Remove alloc node that has no uses.
1507 void PhaseMacroExpand::yank_alloc_node(AllocateNode* alloc) {
1508 Node* ctrl = alloc->in(TypeFunc::Control);
1509 Node* mem = alloc->in(TypeFunc::Memory);
1510 Node* i_o = alloc->in(TypeFunc::I_O);
1511
1512 alloc->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1513 if (_callprojs.resproj != NULL) {
1514 for (DUIterator_Fast imax, i = _callprojs.resproj->fast_outs(imax); i < imax; i++) {
1515 Node* use = _callprojs.resproj->fast_out(i);
1516 use->isa_MemBar()->remove(&_igvn);
1517 --imax;
1518 --i; // back up iterator
1519 }
1520 assert(_callprojs.resproj->outcnt() == 0, "all uses must be deleted");
1521 _igvn.remove_dead_node(_callprojs.resproj);
1522 }
1523 if (_callprojs.fallthrough_catchproj != NULL) {
1524 migrate_outs(_callprojs.fallthrough_catchproj, ctrl);
1525 _igvn.remove_dead_node(_callprojs.fallthrough_catchproj);
1526 }
1527 if (_callprojs.catchall_catchproj != NULL) {
1528 _igvn.rehash_node_delayed(_callprojs.catchall_catchproj);
1529 _callprojs.catchall_catchproj->set_req(0, top());
1530 }
1531 if (_callprojs.fallthrough_proj != NULL) {
1532 Node* catchnode = _callprojs.fallthrough_proj->unique_ctrl_out();
1533 _igvn.remove_dead_node(catchnode);
1534 _igvn.remove_dead_node(_callprojs.fallthrough_proj);
1535 }
1536 if (_callprojs.fallthrough_memproj != NULL) {
1537 migrate_outs(_callprojs.fallthrough_memproj, mem);
1538 _igvn.remove_dead_node(_callprojs.fallthrough_memproj);
1539 }
1540 if (_callprojs.fallthrough_ioproj != NULL) {
1541 migrate_outs(_callprojs.fallthrough_ioproj, i_o);
1542 _igvn.remove_dead_node(_callprojs.fallthrough_ioproj);
1543 }
1544 if (_callprojs.catchall_memproj != NULL) {
1545 _igvn.rehash_node_delayed(_callprojs.catchall_memproj);
1546 _callprojs.catchall_memproj->set_req(0, top());
1547 }
1548 if (_callprojs.catchall_ioproj != NULL) {
1549 _igvn.rehash_node_delayed(_callprojs.catchall_ioproj);
1550 _callprojs.catchall_ioproj->set_req(0, top());
1551 }
1552 #ifndef PRODUCT
1553 if (PrintEliminateAllocations) {
1554 if (alloc->is_AllocateArray()) {
1555 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1556 } else {
1557 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1558 }
1559 }
1560 #endif
1561 _igvn.remove_dead_node(alloc);
1562 }
1563
1564 void PhaseMacroExpand::expand_initialize_membar(AllocateNode* alloc, InitializeNode* init,
1565 Node*& fast_oop_ctrl, Node*& fast_oop_rawmem) {
1566 // If initialization is performed by an array copy, any required
1567 // MemBarStoreStore was already added. If the object does not
1568 // escape no need for a MemBarStoreStore. If the object does not
1569 // escape in its initializer and memory barrier (MemBarStoreStore or
1570 // stronger) is already added at exit of initializer, also no need
1571 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore
1572 // so that stores that initialize this object can't be reordered
1573 // with a subsequent store that makes this object accessible by
1574 // other threads.
1575 // Other threads include java threads and JVM internal threads
1576 // (for example concurrent GC threads). Current concurrent GC
1577 // implementation: G1 will not scan newly created object,
1578 // so it's safe to skip storestore barrier when allocation does
1579 // not escape.
1580 if (!alloc->does_not_escape_thread() &&
1581 !alloc->is_allocation_MemBar_redundant() &&
1582 (init == NULL || !init->is_complete_with_arraycopy())) {
1583 if (init == NULL || init->req() < InitializeNode::RawStores) {
1584 // No InitializeNode or no stores captured by zeroing
1585 // elimination. Simply add the MemBarStoreStore after object
1586 // initialization.
1587 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1588 transform_later(mb);
1589
1590 mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
1591 mb->init_req(TypeFunc::Control, fast_oop_ctrl);
1592 fast_oop_ctrl = new ProjNode(mb, TypeFunc::Control);
1593 transform_later(fast_oop_ctrl);
1594 fast_oop_rawmem = new ProjNode(mb, TypeFunc::Memory);
1595 transform_later(fast_oop_rawmem);
1596 } else {
1597 // Add the MemBarStoreStore after the InitializeNode so that
1598 // all stores performing the initialization that were moved
1599 // before the InitializeNode happen before the storestore
1600 // barrier.
1601
1602 Node* init_ctrl = init->proj_out_or_null(TypeFunc::Control);
1603 Node* init_mem = init->proj_out_or_null(TypeFunc::Memory);
1604
1605 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1606 transform_later(mb);
1607
1608 Node* ctrl = new ProjNode(init, TypeFunc::Control);
1609 transform_later(ctrl);
1610 Node* mem = new ProjNode(init, TypeFunc::Memory);
1611 transform_later(mem);
1612
1613 // The MemBarStoreStore depends on control and memory coming
1614 // from the InitializeNode
1615 mb->init_req(TypeFunc::Memory, mem);
1616 mb->init_req(TypeFunc::Control, ctrl);
1617
1618 ctrl = new ProjNode(mb, TypeFunc::Control);
1619 transform_later(ctrl);
1620 mem = new ProjNode(mb, TypeFunc::Memory);
1621 transform_later(mem);
1622
1623 // All nodes that depended on the InitializeNode for control
1624 // and memory must now depend on the MemBarNode that itself
1625 // depends on the InitializeNode
1626 if (init_ctrl != NULL) {
1627 _igvn.replace_node(init_ctrl, ctrl);
1628 }
1629 if (init_mem != NULL) {
1630 _igvn.replace_node(init_mem, mem);
1631 }
1632 }
1633 }
1634 }
1635
1636 void PhaseMacroExpand::expand_dtrace_alloc_probe(AllocateNode* alloc, Node* oop,
1637 Node*& ctrl, Node*& rawmem) {
1638 if (C->env()->dtrace_extended_probes()) {
1639 // Slow-path call
1640 int size = TypeFunc::Parms + 2;
1641 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1642 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base),
1643 "dtrace_object_alloc",
1644 TypeRawPtr::BOTTOM);
1645
1646 // Get base of thread-local storage area
1647 Node* thread = new ThreadLocalNode();
1648 transform_later(thread);
1649
1650 call->init_req(TypeFunc::Parms + 0, thread);
1651 call->init_req(TypeFunc::Parms + 1, oop);
1652 call->init_req(TypeFunc::Control, ctrl);
1653 call->init_req(TypeFunc::I_O , top()); // does no i/o
1654 call->init_req(TypeFunc::Memory , ctrl);
1655 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1656 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1657 transform_later(call);
1658 ctrl = new ProjNode(call, TypeFunc::Control);
1659 transform_later(ctrl);
1660 rawmem = new ProjNode(call, TypeFunc::Memory);
1661 transform_later(rawmem);
1662 }
1663 }
1664
1665 // Helper for PhaseMacroExpand::expand_allocate_common.
1666 // Initializes the newly-allocated storage.
1667 Node*
1668 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1669 Node* control, Node* rawmem, Node* object,
1670 Node* klass_node, Node* length,
1671 Node* size_in_bytes) {
1672 InitializeNode* init = alloc->initialization();
1673 // Store the klass & mark bits
1674 Node* mark_node = alloc->make_ideal_mark(&_igvn, object, control, rawmem);
1675 if (!mark_node->is_Con()) {
1676 transform_later(mark_node);
1677 }
1678 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, TypeX_X->basic_type());
1679
1680 if (!UseCompactObjectHeaders) {
1681 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1682 }
1683
1684 int header_size = alloc->minimum_header_size(); // conservatively small
1685
1686 // Array length
1687 if (length != NULL) { // Arrays need length field
1688 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1689 // conservatively small header size:
1690 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1691 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1692 if (k->is_array_klass()) // we know the exact header size in most cases:
1693 header_size = Klass::layout_helper_header_size(k->layout_helper());
1694 }
1695
1696 // Clear the object body, if necessary.
1697 if (init == NULL) {
1698 // The init has somehow disappeared; be cautious and clear everything.
1699 //
1700 // This can happen if a node is allocated but an uncommon trap occurs
1701 // immediately. In this case, the Initialize gets associated with the
1702 // trap, and may be placed in a different (outer) loop, if the Allocate
1703 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1704 // there can be two Allocates to one Initialize. The answer in all these
1705 // edge cases is safety first. It is always safe to clear immediately
1706 // within an Allocate, and then (maybe or maybe not) clear some more later.
1707 if (!(UseTLAB && ZeroTLAB)) {
1708 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1709 header_size, size_in_bytes,
1710 &_igvn);
1711 }
1712 } else {
1713 if (!init->is_complete()) {
1714 // Try to win by zeroing only what the init does not store.
1715 // We can also try to do some peephole optimizations,
1716 // such as combining some adjacent subword stores.
1717 rawmem = init->complete_stores(control, rawmem, object,
1718 header_size, size_in_bytes, &_igvn);
1719 }
1720 // We have no more use for this link, since the AllocateNode goes away:
1721 init->set_req(InitializeNode::RawAddress, top());
1722 // (If we keep the link, it just confuses the register allocator,
1723 // who thinks he sees a real use of the address by the membar.)
1724 }
1725
1726 return rawmem;
1727 }
1728
1729 // Generate prefetch instructions for next allocations.
1730 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1731 Node*& contended_phi_rawmem,
1732 Node* old_eden_top, Node* new_eden_top,
1733 intx lines) {
1734 enum { fall_in_path = 1, pf_path = 2 };
1735 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1736 // Generate prefetch allocation with watermark check.
1737 // As an allocation hits the watermark, we will prefetch starting
1738 // at a "distance" away from watermark.
1739
1740 Node *pf_region = new RegionNode(3);
1741 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY,
1742 TypeRawPtr::BOTTOM );
1743 // I/O is used for Prefetch
1744 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO );
1745
1746 Node *thread = new ThreadLocalNode();
1747 transform_later(thread);
1748
1749 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread,
1750 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1751 transform_later(eden_pf_adr);
1752
1753 Node *old_pf_wm = new LoadPNode(needgc_false,
1754 contended_phi_rawmem, eden_pf_adr,
1755 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM,
1756 MemNode::unordered);
1757 transform_later(old_pf_wm);
1758
1759 // check against new_eden_top
1760 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm );
1761 transform_later(need_pf_cmp);
1762 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge );
1763 transform_later(need_pf_bol);
1764 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol,
1765 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1766 transform_later(need_pf_iff);
1767
1768 // true node, add prefetchdistance
1769 Node *need_pf_true = new IfTrueNode( need_pf_iff );
1770 transform_later(need_pf_true);
1771
1772 Node *need_pf_false = new IfFalseNode( need_pf_iff );
1773 transform_later(need_pf_false);
1774
1775 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm,
1776 _igvn.MakeConX(AllocatePrefetchDistance) );
1777 transform_later(new_pf_wmt );
1778 new_pf_wmt->set_req(0, need_pf_true);
1779
1780 Node *store_new_wmt = new StorePNode(need_pf_true,
1781 contended_phi_rawmem, eden_pf_adr,
1782 TypeRawPtr::BOTTOM, new_pf_wmt,
1783 MemNode::unordered);
1784 transform_later(store_new_wmt);
1785
1786 // adding prefetches
1787 pf_phi_abio->init_req( fall_in_path, i_o );
1788
1789 Node *prefetch_adr;
1790 Node *prefetch;
1791 uint step_size = AllocatePrefetchStepSize;
1792 uint distance = 0;
1793
1794 for ( intx i = 0; i < lines; i++ ) {
1795 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt,
1796 _igvn.MakeConX(distance) );
1797 transform_later(prefetch_adr);
1798 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1799 transform_later(prefetch);
1800 distance += step_size;
1801 i_o = prefetch;
1802 }
1803 pf_phi_abio->set_req( pf_path, i_o );
1804
1805 pf_region->init_req( fall_in_path, need_pf_false );
1806 pf_region->init_req( pf_path, need_pf_true );
1807
1808 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
1809 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
1810
1811 transform_later(pf_region);
1812 transform_later(pf_phi_rawmem);
1813 transform_later(pf_phi_abio);
1814
1815 needgc_false = pf_region;
1816 contended_phi_rawmem = pf_phi_rawmem;
1817 i_o = pf_phi_abio;
1818 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) {
1819 // Insert a prefetch instruction for each allocation.
1820 // This code is used to generate 1 prefetch instruction per cache line.
1821
1822 // Generate several prefetch instructions.
1823 uint step_size = AllocatePrefetchStepSize;
1824 uint distance = AllocatePrefetchDistance;
1825
1826 // Next cache address.
1827 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top,
1828 _igvn.MakeConX(step_size + distance));
1829 transform_later(cache_adr);
1830 cache_adr = new CastP2XNode(needgc_false, cache_adr);
1831 transform_later(cache_adr);
1832 // Address is aligned to execute prefetch to the beginning of cache line size
1833 // (it is important when BIS instruction is used on SPARC as prefetch).
1834 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
1835 cache_adr = new AndXNode(cache_adr, mask);
1836 transform_later(cache_adr);
1837 cache_adr = new CastX2PNode(cache_adr);
1838 transform_later(cache_adr);
1839
1840 // Prefetch
1841 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr );
1842 prefetch->set_req(0, needgc_false);
1843 transform_later(prefetch);
1844 contended_phi_rawmem = prefetch;
1845 Node *prefetch_adr;
1846 distance = step_size;
1847 for ( intx i = 1; i < lines; i++ ) {
1848 prefetch_adr = new AddPNode( cache_adr, cache_adr,
1849 _igvn.MakeConX(distance) );
1850 transform_later(prefetch_adr);
1851 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr );
1852 transform_later(prefetch);
1853 distance += step_size;
1854 contended_phi_rawmem = prefetch;
1855 }
1856 } else if( AllocatePrefetchStyle > 0 ) {
1857 // Insert a prefetch for each allocation only on the fast-path
1858 Node *prefetch_adr;
1859 Node *prefetch;
1860 // Generate several prefetch instructions.
1861 uint step_size = AllocatePrefetchStepSize;
1862 uint distance = AllocatePrefetchDistance;
1863 for ( intx i = 0; i < lines; i++ ) {
1864 prefetch_adr = new AddPNode( old_eden_top, new_eden_top,
1865 _igvn.MakeConX(distance) );
1866 transform_later(prefetch_adr);
1867 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1868 // Do not let it float too high, since if eden_top == eden_end,
1869 // both might be null.
1870 if( i == 0 ) { // Set control for first prefetch, next follows it
1871 prefetch->init_req(0, needgc_false);
1872 }
1873 transform_later(prefetch);
1874 distance += step_size;
1875 i_o = prefetch;
1876 }
1877 }
1878 return i_o;
1879 }
1880
1881
1882 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
1883 expand_allocate_common(alloc, NULL,
1884 OptoRuntime::new_instance_Type(),
1885 OptoRuntime::new_instance_Java(), NULL);
1886 }
1887
1888 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
1889 Node* length = alloc->in(AllocateNode::ALength);
1890 Node* valid_length_test = alloc->in(AllocateNode::ValidLengthTest);
1891 InitializeNode* init = alloc->initialization();
1892 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1893 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1894 address slow_call_address; // Address of slow call
1895 if (init != NULL && init->is_complete_with_arraycopy() &&
1896 k->is_type_array_klass()) {
1897 // Don't zero type array during slow allocation in VM since
1898 // it will be initialized later by arraycopy in compiled code.
1899 slow_call_address = OptoRuntime::new_array_nozero_Java();
1900 } else {
1901 slow_call_address = OptoRuntime::new_array_Java();
1902 }
1903 expand_allocate_common(alloc, length,
1904 OptoRuntime::new_array_Type(),
1905 slow_call_address, valid_length_test);
1906 }
1907
1908 //-------------------mark_eliminated_box----------------------------------
1909 //
1910 // During EA obj may point to several objects but after few ideal graph
1911 // transformations (CCP) it may point to only one non escaping object
1912 // (but still using phi), corresponding locks and unlocks will be marked
1913 // for elimination. Later obj could be replaced with a new node (new phi)
1914 // and which does not have escape information. And later after some graph
1915 // reshape other locks and unlocks (which were not marked for elimination
1916 // before) are connected to this new obj (phi) but they still will not be
1917 // marked for elimination since new obj has no escape information.
1918 // Mark all associated (same box and obj) lock and unlock nodes for
1919 // elimination if some of them marked already.
1920 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) {
1921 if (oldbox->as_BoxLock()->is_eliminated()) {
1922 return; // This BoxLock node was processed already.
1923 }
1924 // New implementation (EliminateNestedLocks) has separate BoxLock
1925 // node for each locked region so mark all associated locks/unlocks as
1926 // eliminated even if different objects are referenced in one locked region
1927 // (for example, OSR compilation of nested loop inside locked scope).
1928 if (EliminateNestedLocks ||
1929 oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj, NULL)) {
1930 // Box is used only in one lock region. Mark this box as eliminated.
1931 _igvn.hash_delete(oldbox);
1932 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value
1933 _igvn.hash_insert(oldbox);
1934
1935 for (uint i = 0; i < oldbox->outcnt(); i++) {
1936 Node* u = oldbox->raw_out(i);
1937 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
1938 AbstractLockNode* alock = u->as_AbstractLock();
1939 // Check lock's box since box could be referenced by Lock's debug info.
1940 if (alock->box_node() == oldbox) {
1941 // Mark eliminated all related locks and unlocks.
1942 #ifdef ASSERT
1943 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4");
1944 #endif
1945 alock->set_non_esc_obj();
1946 }
1947 }
1948 }
1949 return;
1950 }
1951
1952 // Create new "eliminated" BoxLock node and use it in monitor debug info
1953 // instead of oldbox for the same object.
1954 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
1955
1956 // Note: BoxLock node is marked eliminated only here and it is used
1957 // to indicate that all associated lock and unlock nodes are marked
1958 // for elimination.
1959 newbox->set_eliminated();
1960 transform_later(newbox);
1961
1962 // Replace old box node with new box for all users of the same object.
1963 for (uint i = 0; i < oldbox->outcnt();) {
1964 bool next_edge = true;
1965
1966 Node* u = oldbox->raw_out(i);
1967 if (u->is_AbstractLock()) {
1968 AbstractLockNode* alock = u->as_AbstractLock();
1969 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
1970 // Replace Box and mark eliminated all related locks and unlocks.
1971 #ifdef ASSERT
1972 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5");
1973 #endif
1974 alock->set_non_esc_obj();
1975 _igvn.rehash_node_delayed(alock);
1976 alock->set_box_node(newbox);
1977 next_edge = false;
1978 }
1979 }
1980 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
1981 FastLockNode* flock = u->as_FastLock();
1982 assert(flock->box_node() == oldbox, "sanity");
1983 _igvn.rehash_node_delayed(flock);
1984 flock->set_box_node(newbox);
1985 next_edge = false;
1986 }
1987
1988 // Replace old box in monitor debug info.
1989 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
1990 SafePointNode* sfn = u->as_SafePoint();
1991 JVMState* youngest_jvms = sfn->jvms();
1992 int max_depth = youngest_jvms->depth();
1993 for (int depth = 1; depth <= max_depth; depth++) {
1994 JVMState* jvms = youngest_jvms->of_depth(depth);
1995 int num_mon = jvms->nof_monitors();
1996 // Loop over monitors
1997 for (int idx = 0; idx < num_mon; idx++) {
1998 Node* obj_node = sfn->monitor_obj(jvms, idx);
1999 Node* box_node = sfn->monitor_box(jvms, idx);
2000 if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
2001 int j = jvms->monitor_box_offset(idx);
2002 _igvn.replace_input_of(u, j, newbox);
2003 next_edge = false;
2004 }
2005 }
2006 }
2007 }
2008 if (next_edge) i++;
2009 }
2010 }
2011
2012 //-----------------------mark_eliminated_locking_nodes-----------------------
2013 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
2014 if (EliminateNestedLocks) {
2015 if (alock->is_nested()) {
2016 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
2017 return;
2018 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
2019 // Only Lock node has JVMState needed here.
2020 // Not that preceding claim is documented anywhere else.
2021 if (alock->jvms() != NULL) {
2022 if (alock->as_Lock()->is_nested_lock_region()) {
2023 // Mark eliminated related nested locks and unlocks.
2024 Node* obj = alock->obj_node();
2025 BoxLockNode* box_node = alock->box_node()->as_BoxLock();
2026 assert(!box_node->is_eliminated(), "should not be marked yet");
2027 // Note: BoxLock node is marked eliminated only here
2028 // and it is used to indicate that all associated lock
2029 // and unlock nodes are marked for elimination.
2030 box_node->set_eliminated(); // Box's hash is always NO_HASH here
2031 for (uint i = 0; i < box_node->outcnt(); i++) {
2032 Node* u = box_node->raw_out(i);
2033 if (u->is_AbstractLock()) {
2034 alock = u->as_AbstractLock();
2035 if (alock->box_node() == box_node) {
2036 // Verify that this Box is referenced only by related locks.
2037 assert(alock->obj_node()->eqv_uncast(obj), "");
2038 // Mark all related locks and unlocks.
2039 #ifdef ASSERT
2040 alock->log_lock_optimization(C, "eliminate_lock_set_nested");
2041 #endif
2042 alock->set_nested();
2043 }
2044 }
2045 }
2046 } else {
2047 #ifdef ASSERT
2048 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region");
2049 if (C->log() != NULL)
2050 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output
2051 #endif
2052 }
2053 }
2054 return;
2055 }
2056 // Process locks for non escaping object
2057 assert(alock->is_non_esc_obj(), "");
2058 } // EliminateNestedLocks
2059
2060 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2061 // Look for all locks of this object and mark them and
2062 // corresponding BoxLock nodes as eliminated.
2063 Node* obj = alock->obj_node();
2064 for (uint j = 0; j < obj->outcnt(); j++) {
2065 Node* o = obj->raw_out(j);
2066 if (o->is_AbstractLock() &&
2067 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2068 alock = o->as_AbstractLock();
2069 Node* box = alock->box_node();
2070 // Replace old box node with new eliminated box for all users
2071 // of the same object and mark related locks as eliminated.
2072 mark_eliminated_box(box, obj);
2073 }
2074 }
2075 }
2076 }
2077
2078 // we have determined that this lock/unlock can be eliminated, we simply
2079 // eliminate the node without expanding it.
2080 //
2081 // Note: The membar's associated with the lock/unlock are currently not
2082 // eliminated. This should be investigated as a future enhancement.
2083 //
2084 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2085
2086 if (!alock->is_eliminated()) {
2087 return false;
2088 }
2089 #ifdef ASSERT
2090 if (!alock->is_coarsened()) {
2091 // Check that new "eliminated" BoxLock node is created.
2092 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2093 assert(oldbox->is_eliminated(), "should be done already");
2094 }
2095 #endif
2096
2097 alock->log_lock_optimization(C, "eliminate_lock");
2098
2099 #ifndef PRODUCT
2100 if (PrintEliminateLocks) {
2101 tty->print_cr("++++ Eliminated: %d %s '%s'", alock->_idx, (alock->is_Lock() ? "Lock" : "Unlock"), alock->kind_as_string());
2102 }
2103 #endif
2104
2105 Node* mem = alock->in(TypeFunc::Memory);
2106 Node* ctrl = alock->in(TypeFunc::Control);
2107 guarantee(ctrl != NULL, "missing control projection, cannot replace_node() with NULL");
2108
2109 alock->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
2110 // There are 2 projections from the lock. The lock node will
2111 // be deleted when its last use is subsumed below.
2112 assert(alock->outcnt() == 2 &&
2113 _callprojs.fallthrough_proj != NULL &&
2114 _callprojs.fallthrough_memproj != NULL,
2115 "Unexpected projections from Lock/Unlock");
2116
2117 Node* fallthroughproj = _callprojs.fallthrough_proj;
2118 Node* memproj_fallthrough = _callprojs.fallthrough_memproj;
2119
2120 // The memory projection from a lock/unlock is RawMem
2121 // The input to a Lock is merged memory, so extract its RawMem input
2122 // (unless the MergeMem has been optimized away.)
2123 if (alock->is_Lock()) {
2124 // Seach for MemBarAcquireLock node and delete it also.
2125 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2126 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, "");
2127 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2128 Node* memproj = membar->proj_out(TypeFunc::Memory);
2129 _igvn.replace_node(ctrlproj, fallthroughproj);
2130 _igvn.replace_node(memproj, memproj_fallthrough);
2131
2132 // Delete FastLock node also if this Lock node is unique user
2133 // (a loop peeling may clone a Lock node).
2134 Node* flock = alock->as_Lock()->fastlock_node();
2135 if (flock->outcnt() == 1) {
2136 assert(flock->unique_out() == alock, "sanity");
2137 _igvn.replace_node(flock, top());
2138 }
2139 }
2140
2141 // Seach for MemBarReleaseLock node and delete it also.
2142 if (alock->is_Unlock() && ctrl->is_Proj() && ctrl->in(0)->is_MemBar()) {
2143 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2144 assert(membar->Opcode() == Op_MemBarReleaseLock &&
2145 mem->is_Proj() && membar == mem->in(0), "");
2146 _igvn.replace_node(fallthroughproj, ctrl);
2147 _igvn.replace_node(memproj_fallthrough, mem);
2148 fallthroughproj = ctrl;
2149 memproj_fallthrough = mem;
2150 ctrl = membar->in(TypeFunc::Control);
2151 mem = membar->in(TypeFunc::Memory);
2152 }
2153
2154 _igvn.replace_node(fallthroughproj, ctrl);
2155 _igvn.replace_node(memproj_fallthrough, mem);
2156 return true;
2157 }
2158
2159
2160 //------------------------------expand_lock_node----------------------
2161 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2162
2163 Node* ctrl = lock->in(TypeFunc::Control);
2164 Node* mem = lock->in(TypeFunc::Memory);
2165 Node* obj = lock->obj_node();
2166 Node* box = lock->box_node();
2167 Node* flock = lock->fastlock_node();
2168
2169 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2170
2171 // Make the merge point
2172 Node *region;
2173 Node *mem_phi;
2174 Node *slow_path;
2175
2176 if (UseOptoBiasInlining) {
2177 /*
2178 * See the full description in MacroAssembler::biased_locking_enter().
2179 *
2180 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) {
2181 * // The object is biased.
2182 * proto_node = klass->prototype_header;
2183 * o_node = thread | proto_node;
2184 * x_node = o_node ^ mark_word;
2185 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ?
2186 * // Done.
2187 * } else {
2188 * if( (x_node & biased_lock_mask) != 0 ) {
2189 * // The klass's prototype header is no longer biased.
2190 * cas(&mark_word, mark_word, proto_node)
2191 * goto cas_lock;
2192 * } else {
2193 * // The klass's prototype header is still biased.
2194 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch?
2195 * old = mark_word;
2196 * new = o_node;
2197 * } else {
2198 * // Different thread or anonymous biased.
2199 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask);
2200 * new = thread | old;
2201 * }
2202 * // Try to rebias.
2203 * if( cas(&mark_word, old, new) == 0 ) {
2204 * // Done.
2205 * } else {
2206 * goto slow_path; // Failed.
2207 * }
2208 * }
2209 * }
2210 * } else {
2211 * // The object is not biased.
2212 * cas_lock:
2213 * if( FastLock(obj) == 0 ) {
2214 * // Done.
2215 * } else {
2216 * slow_path:
2217 * OptoRuntime::complete_monitor_locking_Java(obj);
2218 * }
2219 * }
2220 */
2221
2222 region = new RegionNode(5);
2223 // create a Phi for the memory state
2224 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2225
2226 Node* fast_lock_region = new RegionNode(3);
2227 Node* fast_lock_mem_phi = new PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM);
2228
2229 // First, check mark word for the biased lock pattern.
2230 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2231
2232 // Get fast path - mark word has the biased lock pattern.
2233 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node,
2234 markWord::biased_lock_mask_in_place,
2235 markWord::biased_lock_pattern, true);
2236 // fast_lock_region->in(1) is set to slow path.
2237 fast_lock_mem_phi->init_req(1, mem);
2238
2239 // Now check that the lock is biased to the current thread and has
2240 // the same epoch and bias as Klass::_prototype_header.
2241
2242 // Special-case a fresh allocation to avoid building nodes:
2243 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn);
2244 if (klass_node == NULL) {
2245 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
2246 klass_node = transform_later(LoadKlassNode::make(_igvn, NULL, mem, k_adr, _igvn.type(k_adr)->is_ptr()));
2247 #ifdef _LP64
2248 if (UseCompressedClassPointers && klass_node->is_DecodeNKlass()) {
2249 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity");
2250 klass_node->in(1)->init_req(0, ctrl);
2251 } else
2252 #endif
2253 klass_node->init_req(0, ctrl);
2254 }
2255 Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type());
2256
2257 Node* thread = transform_later(new ThreadLocalNode());
2258 Node* cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2259 Node* o_node = transform_later(new OrXNode(cast_thread, proto_node));
2260 Node* x_node = transform_later(new XorXNode(o_node, mark_node));
2261
2262 // Get slow path - mark word does NOT match the value.
2263 STATIC_ASSERT(markWord::age_mask_in_place <= INT_MAX);
2264 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node,
2265 (~(int)markWord::age_mask_in_place), 0);
2266 // region->in(3) is set to fast path - the object is biased to the current thread.
2267 mem_phi->init_req(3, mem);
2268
2269
2270 // Mark word does NOT match the value (thread | Klass::_prototype_header).
2271
2272
2273 // First, check biased pattern.
2274 // Get fast path - _prototype_header has the same biased lock pattern.
2275 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node,
2276 markWord::biased_lock_mask_in_place, 0, true);
2277
2278 not_biased_ctrl = fast_lock_region->in(2); // Slow path
2279 // fast_lock_region->in(2) - the prototype header is no longer biased
2280 // and we have to revoke the bias on this object.
2281 // We are going to try to reset the mark of this object to the prototype
2282 // value and fall through to the CAS-based locking scheme.
2283 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
2284 Node* cas = new StoreXConditionalNode(not_biased_ctrl, mem, adr,
2285 proto_node, mark_node);
2286 transform_later(cas);
2287 Node* proj = transform_later(new SCMemProjNode(cas));
2288 fast_lock_mem_phi->init_req(2, proj);
2289
2290
2291 // Second, check epoch bits.
2292 Node* rebiased_region = new RegionNode(3);
2293 Node* old_phi = new PhiNode( rebiased_region, TypeX_X);
2294 Node* new_phi = new PhiNode( rebiased_region, TypeX_X);
2295
2296 // Get slow path - mark word does NOT match epoch bits.
2297 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node,
2298 markWord::epoch_mask_in_place, 0);
2299 // The epoch of the current bias is not valid, attempt to rebias the object
2300 // toward the current thread.
2301 rebiased_region->init_req(2, epoch_ctrl);
2302 old_phi->init_req(2, mark_node);
2303 new_phi->init_req(2, o_node);
2304
2305 // rebiased_region->in(1) is set to fast path.
2306 // The epoch of the current bias is still valid but we know
2307 // nothing about the owner; it might be set or it might be clear.
2308 Node* cmask = MakeConX(markWord::biased_lock_mask_in_place |
2309 markWord::age_mask_in_place |
2310 markWord::epoch_mask_in_place);
2311 Node* old = transform_later(new AndXNode(mark_node, cmask));
2312 cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2313 Node* new_mark = transform_later(new OrXNode(cast_thread, old));
2314 old_phi->init_req(1, old);
2315 new_phi->init_req(1, new_mark);
2316
2317 transform_later(rebiased_region);
2318 transform_later(old_phi);
2319 transform_later(new_phi);
2320
2321 // Try to acquire the bias of the object using an atomic operation.
2322 // If this fails we will go in to the runtime to revoke the object's bias.
2323 cas = new StoreXConditionalNode(rebiased_region, mem, adr, new_phi, old_phi);
2324 transform_later(cas);
2325 proj = transform_later(new SCMemProjNode(cas));
2326
2327 // Get slow path - Failed to CAS.
2328 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0);
2329 mem_phi->init_req(4, proj);
2330 // region->in(4) is set to fast path - the object is rebiased to the current thread.
2331
2332 // Failed to CAS.
2333 slow_path = new RegionNode(3);
2334 Node *slow_mem = new PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM);
2335
2336 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control
2337 slow_mem->init_req(1, proj);
2338
2339 // Call CAS-based locking scheme (FastLock node).
2340
2341 transform_later(fast_lock_region);
2342 transform_later(fast_lock_mem_phi);
2343
2344 // Get slow path - FastLock failed to lock the object.
2345 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0);
2346 mem_phi->init_req(2, fast_lock_mem_phi);
2347 // region->in(2) is set to fast path - the object is locked to the current thread.
2348
2349 slow_path->init_req(2, ctrl); // Capture slow-control
2350 slow_mem->init_req(2, fast_lock_mem_phi);
2351
2352 transform_later(slow_path);
2353 transform_later(slow_mem);
2354 // Reset lock's memory edge.
2355 lock->set_req(TypeFunc::Memory, slow_mem);
2356
2357 } else {
2358 region = new RegionNode(3);
2359 // create a Phi for the memory state
2360 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2361
2362 // Optimize test; set region slot 2
2363 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2364 mem_phi->init_req(2, mem);
2365 }
2366
2367 // Make slow path call
2368 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(),
2369 OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path,
2370 obj, box, NULL);
2371
2372 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
2373
2374 // Slow path can only throw asynchronous exceptions, which are always
2375 // de-opted. So the compiler thinks the slow-call can never throw an
2376 // exception. If it DOES throw an exception we would need the debug
2377 // info removed first (since if it throws there is no monitor).
2378 assert(_callprojs.fallthrough_ioproj == NULL && _callprojs.catchall_ioproj == NULL &&
2379 _callprojs.catchall_memproj == NULL && _callprojs.catchall_catchproj == NULL, "Unexpected projection from Lock");
2380
2381 // Capture slow path
2382 // disconnect fall-through projection from call and create a new one
2383 // hook up users of fall-through projection to region
2384 Node *slow_ctrl = _callprojs.fallthrough_proj->clone();
2385 transform_later(slow_ctrl);
2386 _igvn.hash_delete(_callprojs.fallthrough_proj);
2387 _callprojs.fallthrough_proj->disconnect_inputs(C);
2388 region->init_req(1, slow_ctrl);
2389 // region inputs are now complete
2390 transform_later(region);
2391 _igvn.replace_node(_callprojs.fallthrough_proj, region);
2392
2393 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2394 mem_phi->init_req(1, memproj );
2395 transform_later(mem_phi);
2396 _igvn.replace_node(_callprojs.fallthrough_memproj, mem_phi);
2397 }
2398
2399 //------------------------------expand_unlock_node----------------------
2400 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2401
2402 Node* ctrl = unlock->in(TypeFunc::Control);
2403 Node* mem = unlock->in(TypeFunc::Memory);
2404 Node* obj = unlock->obj_node();
2405 Node* box = unlock->box_node();
2406
2407 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2408
2409 // No need for a null check on unlock
2410
2411 // Make the merge point
2412 Node *region;
2413 Node *mem_phi;
2414
2415 if (UseOptoBiasInlining) {
2416 // Check for biased locking unlock case, which is a no-op.
2417 // See the full description in MacroAssembler::biased_locking_exit().
2418 region = new RegionNode(4);
2419 // create a Phi for the memory state
2420 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2421 mem_phi->init_req(3, mem);
2422
2423 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2424 ctrl = opt_bits_test(ctrl, region, 3, mark_node,
2425 markWord::biased_lock_mask_in_place,
2426 markWord::biased_lock_pattern);
2427 } else {
2428 region = new RegionNode(3);
2429 // create a Phi for the memory state
2430 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2431 }
2432
2433 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2434 funlock = transform_later( funlock )->as_FastUnlock();
2435 // Optimize test; set region slot 2
2436 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2437 Node *thread = transform_later(new ThreadLocalNode());
2438
2439 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2440 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2441 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2442
2443 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
2444 assert(_callprojs.fallthrough_ioproj == NULL && _callprojs.catchall_ioproj == NULL &&
2445 _callprojs.catchall_memproj == NULL && _callprojs.catchall_catchproj == NULL, "Unexpected projection from Lock");
2446
2447 // No exceptions for unlocking
2448 // Capture slow path
2449 // disconnect fall-through projection from call and create a new one
2450 // hook up users of fall-through projection to region
2451 Node *slow_ctrl = _callprojs.fallthrough_proj->clone();
2452 transform_later(slow_ctrl);
2453 _igvn.hash_delete(_callprojs.fallthrough_proj);
2454 _callprojs.fallthrough_proj->disconnect_inputs(C);
2455 region->init_req(1, slow_ctrl);
2456 // region inputs are now complete
2457 transform_later(region);
2458 _igvn.replace_node(_callprojs.fallthrough_proj, region);
2459
2460 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2461 mem_phi->init_req(1, memproj );
2462 mem_phi->init_req(2, mem);
2463 transform_later(mem_phi);
2464 _igvn.replace_node(_callprojs.fallthrough_memproj, mem_phi);
2465 }
2466
2467 void PhaseMacroExpand::expand_subtypecheck_node(SubTypeCheckNode *check) {
2468 assert(check->in(SubTypeCheckNode::Control) == NULL, "should be pinned");
2469 Node* bol = check->unique_out();
2470 Node* obj_or_subklass = check->in(SubTypeCheckNode::ObjOrSubKlass);
2471 Node* superklass = check->in(SubTypeCheckNode::SuperKlass);
2472 assert(bol->is_Bool() && bol->as_Bool()->_test._test == BoolTest::ne, "unexpected bool node");
2473
2474 for (DUIterator_Last imin, i = bol->last_outs(imin); i >= imin; --i) {
2475 Node* iff = bol->last_out(i);
2476 assert(iff->is_If(), "where's the if?");
2477
2478 if (iff->in(0)->is_top()) {
2479 _igvn.replace_input_of(iff, 1, C->top());
2480 continue;
2481 }
2482
2483 Node* iftrue = iff->as_If()->proj_out(1);
2484 Node* iffalse = iff->as_If()->proj_out(0);
2485 Node* ctrl = iff->in(0);
2486
2487 Node* subklass = NULL;
2488 if (_igvn.type(obj_or_subklass)->isa_klassptr()) {
2489 subklass = obj_or_subklass;
2490 } else {
2491 Node* k_adr = basic_plus_adr(obj_or_subklass, oopDesc::klass_offset_in_bytes());
2492 subklass = _igvn.transform(LoadKlassNode::make(_igvn, NULL, C->immutable_memory(), k_adr, TypeInstPtr::KLASS));
2493 }
2494
2495 Node* not_subtype_ctrl = Phase::gen_subtype_check(subklass, superklass, &ctrl, NULL, _igvn);
2496
2497 _igvn.replace_input_of(iff, 0, C->top());
2498 _igvn.replace_node(iftrue, not_subtype_ctrl);
2499 _igvn.replace_node(iffalse, ctrl);
2500 }
2501 _igvn.replace_node(check, C->top());
2502 }
2503
2504 //---------------------------eliminate_macro_nodes----------------------
2505 // Eliminate scalar replaced allocations and associated locks.
2506 void PhaseMacroExpand::eliminate_macro_nodes() {
2507 if (C->macro_count() == 0)
2508 return;
2509
2510 // Before elimination may re-mark (change to Nested or NonEscObj)
2511 // all associated (same box and obj) lock and unlock nodes.
2512 int cnt = C->macro_count();
2513 for (int i=0; i < cnt; i++) {
2514 Node *n = C->macro_node(i);
2515 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2516 mark_eliminated_locking_nodes(n->as_AbstractLock());
2517 }
2518 }
2519 // Re-marking may break consistency of Coarsened locks.
2520 if (!C->coarsened_locks_consistent()) {
2521 return; // recompile without Coarsened locks if broken
2522 }
2523
2524 // First, attempt to eliminate locks
2525 bool progress = true;
2526 while (progress) {
2527 progress = false;
2528 for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once
2529 Node* n = C->macro_node(i - 1);
2530 bool success = false;
2531 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2532 if (n->is_AbstractLock()) {
2533 success = eliminate_locking_node(n->as_AbstractLock());
2534 }
2535 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2536 progress = progress || success;
2537 }
2538 }
2539 // Next, attempt to eliminate allocations
2540 _has_locks = false;
2541 progress = true;
2542 while (progress) {
2543 progress = false;
2544 for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once
2545 Node* n = C->macro_node(i - 1);
2546 bool success = false;
2547 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2548 switch (n->class_id()) {
2549 case Node::Class_Allocate:
2550 case Node::Class_AllocateArray:
2551 success = eliminate_allocate_node(n->as_Allocate());
2552 break;
2553 case Node::Class_CallStaticJava:
2554 success = eliminate_boxing_node(n->as_CallStaticJava());
2555 break;
2556 case Node::Class_Lock:
2557 case Node::Class_Unlock:
2558 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2559 _has_locks = true;
2560 break;
2561 case Node::Class_ArrayCopy:
2562 break;
2563 case Node::Class_OuterStripMinedLoop:
2564 break;
2565 case Node::Class_SubTypeCheck:
2566 break;
2567 case Node::Class_Opaque1:
2568 break;
2569 default:
2570 assert(n->Opcode() == Op_LoopLimit ||
2571 n->Opcode() == Op_Opaque2 ||
2572 n->Opcode() == Op_Opaque3 ||
2573 n->Opcode() == Op_Opaque4 ||
2574 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n),
2575 "unknown node type in macro list");
2576 }
2577 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2578 progress = progress || success;
2579 }
2580 }
2581 }
2582
2583 //------------------------------expand_macro_nodes----------------------
2584 // Returns true if a failure occurred.
2585 bool PhaseMacroExpand::expand_macro_nodes() {
2586 // Last attempt to eliminate macro nodes.
2587 eliminate_macro_nodes();
2588 if (C->failing()) return true;
2589
2590 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2591 bool progress = true;
2592 while (progress) {
2593 progress = false;
2594 for (int i = C->macro_count(); i > 0; i--) {
2595 Node* n = C->macro_node(i-1);
2596 bool success = false;
2597 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2598 if (n->Opcode() == Op_LoopLimit) {
2599 // Remove it from macro list and put on IGVN worklist to optimize.
2600 C->remove_macro_node(n);
2601 _igvn._worklist.push(n);
2602 success = true;
2603 } else if (n->Opcode() == Op_CallStaticJava) {
2604 // Remove it from macro list and put on IGVN worklist to optimize.
2605 C->remove_macro_node(n);
2606 _igvn._worklist.push(n);
2607 success = true;
2608 } else if (n->is_Opaque1() || n->Opcode() == Op_Opaque2) {
2609 _igvn.replace_node(n, n->in(1));
2610 success = true;
2611 #if INCLUDE_RTM_OPT
2612 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) {
2613 assert(C->profile_rtm(), "should be used only in rtm deoptimization code");
2614 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), "");
2615 Node* cmp = n->unique_out();
2616 #ifdef ASSERT
2617 // Validate graph.
2618 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), "");
2619 BoolNode* bol = cmp->unique_out()->as_Bool();
2620 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() &&
2621 (bol->_test._test == BoolTest::ne), "");
2622 IfNode* ifn = bol->unique_out()->as_If();
2623 assert((ifn->outcnt() == 2) &&
2624 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != NULL, "");
2625 #endif
2626 Node* repl = n->in(1);
2627 if (!_has_locks) {
2628 // Remove RTM state check if there are no locks in the code.
2629 // Replace input to compare the same value.
2630 repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1);
2631 }
2632 _igvn.replace_node(n, repl);
2633 success = true;
2634 #endif
2635 } else if (n->Opcode() == Op_Opaque4) {
2636 // With Opaque4 nodes, the expectation is that the test of input 1
2637 // is always equal to the constant value of input 2. So we can
2638 // remove the Opaque4 and replace it by input 2. In debug builds,
2639 // leave the non constant test in instead to sanity check that it
2640 // never fails (if it does, that subgraph was constructed so, at
2641 // runtime, a Halt node is executed).
2642 #ifdef ASSERT
2643 _igvn.replace_node(n, n->in(1));
2644 #else
2645 _igvn.replace_node(n, n->in(2));
2646 #endif
2647 success = true;
2648 } else if (n->Opcode() == Op_OuterStripMinedLoop) {
2649 n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn);
2650 C->remove_macro_node(n);
2651 success = true;
2652 }
2653 assert(!success || (C->macro_count() == (old_macro_count - 1)), "elimination must have deleted one node from macro list");
2654 progress = progress || success;
2655 }
2656 }
2657
2658 // Clean up the graph so we're less likely to hit the maximum node
2659 // limit
2660 _igvn.set_delay_transform(false);
2661 _igvn.optimize();
2662 if (C->failing()) return true;
2663 _igvn.set_delay_transform(true);
2664
2665
2666 // Because we run IGVN after each expansion, some macro nodes may go
2667 // dead and be removed from the list as we iterate over it. Move
2668 // Allocate nodes (processed in a second pass) at the beginning of
2669 // the list and then iterate from the last element of the list until
2670 // an Allocate node is seen. This is robust to random deletion in
2671 // the list due to nodes going dead.
2672 C->sort_macro_nodes();
2673
2674 // expand arraycopy "macro" nodes first
2675 // For ReduceBulkZeroing, we must first process all arraycopy nodes
2676 // before the allocate nodes are expanded.
2677 while (C->macro_count() > 0) {
2678 int macro_count = C->macro_count();
2679 Node * n = C->macro_node(macro_count-1);
2680 assert(n->is_macro(), "only macro nodes expected here");
2681 if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) {
2682 // node is unreachable, so don't try to expand it
2683 C->remove_macro_node(n);
2684 continue;
2685 }
2686 if (n->is_Allocate()) {
2687 break;
2688 }
2689 // Make sure expansion will not cause node limit to be exceeded.
2690 // Worst case is a macro node gets expanded into about 200 nodes.
2691 // Allow 50% more for optimization.
2692 if (C->check_node_count(300, "out of nodes before macro expansion")) {
2693 return true;
2694 }
2695
2696 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2697 switch (n->class_id()) {
2698 case Node::Class_Lock:
2699 expand_lock_node(n->as_Lock());
2700 break;
2701 case Node::Class_Unlock:
2702 expand_unlock_node(n->as_Unlock());
2703 break;
2704 case Node::Class_ArrayCopy:
2705 expand_arraycopy_node(n->as_ArrayCopy());
2706 break;
2707 case Node::Class_SubTypeCheck:
2708 expand_subtypecheck_node(n->as_SubTypeCheck());
2709 break;
2710 default:
2711 assert(false, "unknown node type in macro list");
2712 }
2713 assert(C->macro_count() == (old_macro_count - 1), "expansion must have deleted one node from macro list");
2714 if (C->failing()) return true;
2715
2716 // Clean up the graph so we're less likely to hit the maximum node
2717 // limit
2718 _igvn.set_delay_transform(false);
2719 _igvn.optimize();
2720 if (C->failing()) return true;
2721 _igvn.set_delay_transform(true);
2722 }
2723
2724 // All nodes except Allocate nodes are expanded now. There could be
2725 // new optimization opportunities (such as folding newly created
2726 // load from a just allocated object). Run IGVN.
2727
2728 // expand "macro" nodes
2729 // nodes are removed from the macro list as they are processed
2730 while (C->macro_count() > 0) {
2731 int macro_count = C->macro_count();
2732 Node * n = C->macro_node(macro_count-1);
2733 assert(n->is_macro(), "only macro nodes expected here");
2734 if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) {
2735 // node is unreachable, so don't try to expand it
2736 C->remove_macro_node(n);
2737 continue;
2738 }
2739 // Make sure expansion will not cause node limit to be exceeded.
2740 // Worst case is a macro node gets expanded into about 200 nodes.
2741 // Allow 50% more for optimization.
2742 if (C->check_node_count(300, "out of nodes before macro expansion")) {
2743 return true;
2744 }
2745 switch (n->class_id()) {
2746 case Node::Class_Allocate:
2747 expand_allocate(n->as_Allocate());
2748 break;
2749 case Node::Class_AllocateArray:
2750 expand_allocate_array(n->as_AllocateArray());
2751 break;
2752 default:
2753 assert(false, "unknown node type in macro list");
2754 }
2755 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2756 if (C->failing()) return true;
2757
2758 // Clean up the graph so we're less likely to hit the maximum node
2759 // limit
2760 _igvn.set_delay_transform(false);
2761 _igvn.optimize();
2762 if (C->failing()) return true;
2763 _igvn.set_delay_transform(true);
2764 }
2765
2766 _igvn.set_delay_transform(false);
2767 return false;
2768 }