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