1 /*
2 * Copyright (c) 2005, 2025, 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 "compiler/compileLog.hpp"
26 #include "gc/shared/collectedHeap.inline.hpp"
27 #include "gc/shared/tlab_globals.hpp"
28 #include "libadt/vectset.hpp"
29 #include "memory/universe.hpp"
30 #include "opto/addnode.hpp"
31 #include "opto/arraycopynode.hpp"
32 #include "opto/callnode.hpp"
33 #include "opto/castnode.hpp"
34 #include "opto/cfgnode.hpp"
35 #include "opto/compile.hpp"
36 #include "opto/convertnode.hpp"
37 #include "opto/graphKit.hpp"
38 #include "opto/intrinsicnode.hpp"
39 #include "opto/locknode.hpp"
40 #include "opto/loopnode.hpp"
41 #include "opto/macro.hpp"
42 #include "opto/memnode.hpp"
43 #include "opto/narrowptrnode.hpp"
44 #include "opto/node.hpp"
45 #include "opto/opaquenode.hpp"
46 #include "opto/phaseX.hpp"
47 #include "opto/rootnode.hpp"
48 #include "opto/runtime.hpp"
49 #include "opto/subnode.hpp"
50 #include "opto/subtypenode.hpp"
51 #include "opto/type.hpp"
52 #include "prims/jvmtiExport.hpp"
53 #include "runtime/continuation.hpp"
54 #include "runtime/sharedRuntime.hpp"
55 #include "utilities/macros.hpp"
56 #include "utilities/powerOfTwo.hpp"
57 #if INCLUDE_G1GC
58 #include "gc/g1/g1ThreadLocalData.hpp"
59 #endif // INCLUDE_G1GC
60
61
62 //
63 // Replace any references to "oldref" in inputs to "use" with "newref".
64 // Returns the number of replacements made.
65 //
66 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
67 int nreplacements = 0;
68 uint req = use->req();
69 for (uint j = 0; j < use->len(); j++) {
70 Node *uin = use->in(j);
71 if (uin == oldref) {
72 if (j < req)
73 use->set_req(j, newref);
74 else
75 use->set_prec(j, newref);
76 nreplacements++;
77 } else if (j >= req && uin == nullptr) {
78 break;
79 }
80 }
81 return nreplacements;
82 }
83
84 void PhaseMacroExpand::migrate_outs(Node *old, Node *target) {
85 assert(old != nullptr, "sanity");
86 for (DUIterator_Fast imax, i = old->fast_outs(imax); i < imax; i++) {
87 Node* use = old->fast_out(i);
88 _igvn.rehash_node_delayed(use);
89 imax -= replace_input(use, old, target);
90 // back up iterator
91 --i;
92 }
93 assert(old->outcnt() == 0, "all uses must be deleted");
94 }
95
96 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word) {
97 Node* cmp = word;
98 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
99 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
100 transform_later(iff);
101
102 // Fast path taken.
103 Node *fast_taken = transform_later(new IfFalseNode(iff));
104
105 // Fast path not-taken, i.e. slow path
106 Node *slow_taken = transform_later(new IfTrueNode(iff));
107
108 region->init_req(edge, fast_taken); // Capture fast-control
109 return slow_taken;
110 }
111
112 //--------------------copy_predefined_input_for_runtime_call--------------------
113 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
114 // Set fixed predefined input arguments
115 call->init_req( TypeFunc::Control, ctrl );
116 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
117 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
118 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
119 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
120 }
121
122 //------------------------------make_slow_call---------------------------------
123 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type,
124 address slow_call, const char* leaf_name, Node* slow_path,
125 Node* parm0, Node* parm1, Node* parm2) {
126
127 // Slow-path call
128 CallNode *call = leaf_name
129 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
130 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), TypeRawPtr::BOTTOM );
131
132 // Slow path call has no side-effects, uses few values
133 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
134 if (parm0 != nullptr) call->init_req(TypeFunc::Parms+0, parm0);
135 if (parm1 != nullptr) call->init_req(TypeFunc::Parms+1, parm1);
136 if (parm2 != nullptr) call->init_req(TypeFunc::Parms+2, parm2);
137 call->copy_call_debug_info(&_igvn, oldcall);
138 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
139 _igvn.replace_node(oldcall, call);
140 transform_later(call);
141
142 return call;
143 }
144
145 void PhaseMacroExpand::eliminate_gc_barrier(Node* p2x) {
146 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
147 bs->eliminate_gc_barrier(this, p2x);
148 #ifndef PRODUCT
149 if (PrintOptoStatistics) {
150 AtomicAccess::inc(&PhaseMacroExpand::_GC_barriers_removed_counter);
151 }
152 #endif
153 }
154
155 // Search for a memory operation for the specified memory slice.
156 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
157 Node *orig_mem = mem;
158 Node *alloc_mem = alloc->as_Allocate()->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
159 assert(alloc_mem != nullptr, "Allocation without a memory projection.");
160 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
161 while (true) {
162 if (mem == alloc_mem || mem == start_mem ) {
163 return mem; // hit one of our sentinels
164 } else if (mem->is_MergeMem()) {
165 mem = mem->as_MergeMem()->memory_at(alias_idx);
166 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
167 Node *in = mem->in(0);
168 // we can safely skip over safepoints, calls, locks and membars because we
169 // already know that the object is safe to eliminate.
170 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
171 return in;
172 } else if (in->is_Call()) {
173 CallNode *call = in->as_Call();
174 if (call->may_modify(tinst, phase)) {
175 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape");
176 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) {
177 return in;
178 }
179 }
180 mem = in->in(TypeFunc::Memory);
181 } else if (in->is_MemBar()) {
182 ArrayCopyNode* ac = nullptr;
183 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) {
184 if (ac != nullptr) {
185 assert(ac->is_clonebasic(), "Only basic clone is a non escaping clone");
186 return ac;
187 }
188 }
189 mem = in->in(TypeFunc::Memory);
190 } else {
191 #ifdef ASSERT
192 in->dump();
193 mem->dump();
194 assert(false, "unexpected projection");
195 #endif
196 }
197 } else if (mem->is_Store()) {
198 const TypePtr* atype = mem->as_Store()->adr_type();
199 int adr_idx = phase->C->get_alias_index(atype);
200 if (adr_idx == alias_idx) {
201 assert(atype->isa_oopptr(), "address type must be oopptr");
202 int adr_offset = atype->offset();
203 uint adr_iid = atype->is_oopptr()->instance_id();
204 // Array elements references have the same alias_idx
205 // but different offset and different instance_id.
206 if (adr_offset == offset && adr_iid == alloc->_idx) {
207 return mem;
208 }
209 } else {
210 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
211 }
212 mem = mem->in(MemNode::Memory);
213 } else if (mem->is_ClearArray()) {
214 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
215 // Can not bypass initialization of the instance
216 // we are looking.
217 DEBUG_ONLY(intptr_t offset;)
218 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
219 InitializeNode* init = alloc->as_Allocate()->initialization();
220 // We are looking for stored value, return Initialize node
221 // or memory edge from Allocate node.
222 if (init != nullptr) {
223 return init;
224 } else {
225 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
226 }
227 }
228 // Otherwise skip it (the call updated 'mem' value).
229 } else if (mem->Opcode() == Op_SCMemProj) {
230 mem = mem->in(0);
231 Node* adr = nullptr;
232 if (mem->is_LoadStore()) {
233 adr = mem->in(MemNode::Address);
234 } else {
235 assert(mem->Opcode() == Op_EncodeISOArray ||
236 mem->Opcode() == Op_StrCompressedCopy, "sanity");
237 adr = mem->in(3); // Destination array
238 }
239 const TypePtr* atype = adr->bottom_type()->is_ptr();
240 int adr_idx = phase->C->get_alias_index(atype);
241 if (adr_idx == alias_idx) {
242 DEBUG_ONLY(mem->dump();)
243 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
244 return nullptr;
245 }
246 mem = mem->in(MemNode::Memory);
247 } else if (mem->Opcode() == Op_StrInflatedCopy) {
248 Node* adr = mem->in(3); // Destination array
249 const TypePtr* atype = adr->bottom_type()->is_ptr();
250 int adr_idx = phase->C->get_alias_index(atype);
251 if (adr_idx == alias_idx) {
252 DEBUG_ONLY(mem->dump();)
253 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
254 return nullptr;
255 }
256 mem = mem->in(MemNode::Memory);
257 } else {
258 return mem;
259 }
260 assert(mem != orig_mem, "dead memory loop");
261 }
262 }
263
264 // Generate loads from source of the arraycopy for fields of
265 // destination needed at a deoptimization point
266 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) {
267 BasicType bt = ft;
268 const Type *type = ftype;
269 if (ft == T_NARROWOOP) {
270 bt = T_OBJECT;
271 type = ftype->make_oopptr();
272 }
273 Node* res = nullptr;
274 if (ac->is_clonebasic()) {
275 assert(ac->in(ArrayCopyNode::Src) != ac->in(ArrayCopyNode::Dest), "clone source equals destination");
276 Node* base = ac->in(ArrayCopyNode::Src);
277 Node* adr = _igvn.transform(new AddPNode(base, base, _igvn.MakeConX(offset)));
278 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
279 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
280 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
281 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
282 } else {
283 if (ac->modifies(offset, offset, &_igvn, true)) {
284 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
285 uint shift = exact_log2(type2aelembytes(bt));
286 Node* src_pos = ac->in(ArrayCopyNode::SrcPos);
287 Node* dest_pos = ac->in(ArrayCopyNode::DestPos);
288 const TypeInt* src_pos_t = _igvn.type(src_pos)->is_int();
289 const TypeInt* dest_pos_t = _igvn.type(dest_pos)->is_int();
290
291 Node* adr = nullptr;
292 const TypePtr* adr_type = nullptr;
293 if (src_pos_t->is_con() && dest_pos_t->is_con()) {
294 intptr_t off = ((src_pos_t->get_con() - dest_pos_t->get_con()) << shift) + offset;
295 Node* base = ac->in(ArrayCopyNode::Src);
296 adr = _igvn.transform(new AddPNode(base, base, _igvn.MakeConX(off)));
297 adr_type = _igvn.type(base)->is_ptr()->add_offset(off);
298 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
299 // Don't emit a new load from src if src == dst but try to get the value from memory instead
300 return value_from_mem(ac->in(TypeFunc::Memory), ctl, ft, ftype, adr_type->isa_oopptr(), alloc);
301 }
302 } else {
303 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
304 #ifdef _LP64
305 diff = _igvn.transform(new ConvI2LNode(diff));
306 #endif
307 diff = _igvn.transform(new LShiftXNode(diff, _igvn.intcon(shift)));
308
309 Node* off = _igvn.transform(new AddXNode(_igvn.MakeConX(offset), diff));
310 Node* base = ac->in(ArrayCopyNode::Src);
311 adr = _igvn.transform(new AddPNode(base, base, off));
312 adr_type = _igvn.type(base)->is_ptr()->add_offset(Type::OffsetBot);
313 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
314 // Non constant offset in the array: we can't statically
315 // determine the value
316 return nullptr;
317 }
318 }
319 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
320 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
321 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
322 }
323 }
324 if (res != nullptr) {
325 if (ftype->isa_narrowoop()) {
326 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
327 res = _igvn.transform(new EncodePNode(res, ftype));
328 }
329 return res;
330 }
331 return nullptr;
332 }
333
334 //
335 // Given a Memory Phi, compute a value Phi containing the values from stores
336 // on the input paths.
337 // Note: this function is recursive, its depth is limited by the "level" argument
338 // Returns the computed Phi, or null if it cannot compute it.
339 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) {
340 assert(mem->is_Phi(), "sanity");
341 int alias_idx = C->get_alias_index(adr_t);
342 int offset = adr_t->offset();
343 int instance_id = adr_t->instance_id();
344
345 // Check if an appropriate value phi already exists.
346 Node* region = mem->in(0);
347 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
348 Node* phi = region->fast_out(k);
349 if (phi->is_Phi() && phi != mem &&
350 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
351 return phi;
352 }
353 }
354 // Check if an appropriate new value phi already exists.
355 Node* new_phi = value_phis->find(mem->_idx);
356 if (new_phi != nullptr)
357 return new_phi;
358
359 if (level <= 0) {
360 return nullptr; // Give up: phi tree too deep
361 }
362 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
363 Node *alloc_mem = alloc->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
364 assert(alloc_mem != nullptr, "Allocation without a memory projection.");
365
366 uint length = mem->req();
367 GrowableArray <Node *> values(length, length, nullptr);
368
369 // create a new Phi for the value
370 PhiNode *phi = new PhiNode(mem->in(0), phi_type, nullptr, mem->_idx, instance_id, alias_idx, offset);
371 transform_later(phi);
372 value_phis->push(phi, mem->_idx);
373
374 for (uint j = 1; j < length; j++) {
375 Node *in = mem->in(j);
376 if (in == nullptr || in->is_top()) {
377 values.at_put(j, in);
378 } else {
379 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
380 if (val == start_mem || val == alloc_mem) {
381 // hit a sentinel, return appropriate 0 value
382 values.at_put(j, _igvn.zerocon(ft));
383 continue;
384 }
385 if (val->is_Initialize()) {
386 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
387 }
388 if (val == nullptr) {
389 return nullptr; // can't find a value on this path
390 }
391 if (val == mem) {
392 values.at_put(j, mem);
393 } else if (val->is_Store()) {
394 Node* n = val->in(MemNode::ValueIn);
395 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
396 n = bs->step_over_gc_barrier(n);
397 if (is_subword_type(ft)) {
398 n = Compile::narrow_value(ft, n, phi_type, &_igvn, true);
399 }
400 values.at_put(j, n);
401 } else if(val->is_Proj() && val->in(0) == alloc) {
402 values.at_put(j, _igvn.zerocon(ft));
403 } else if (val->is_Phi()) {
404 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
405 if (val == nullptr) {
406 return nullptr;
407 }
408 values.at_put(j, val);
409 } else if (val->Opcode() == Op_SCMemProj) {
410 assert(val->in(0)->is_LoadStore() ||
411 val->in(0)->Opcode() == Op_EncodeISOArray ||
412 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
413 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
414 return nullptr;
415 } else if (val->is_ArrayCopy()) {
416 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc);
417 if (res == nullptr) {
418 return nullptr;
419 }
420 values.at_put(j, res);
421 } else if (val->is_top()) {
422 // This indicates that this path into the phi is dead. Top will eventually also propagate into the Region.
423 // IGVN will clean this up later.
424 values.at_put(j, val);
425 } else {
426 DEBUG_ONLY( val->dump(); )
427 assert(false, "unknown node on this path");
428 return nullptr; // unknown node on this path
429 }
430 }
431 }
432 // Set Phi's inputs
433 for (uint j = 1; j < length; j++) {
434 if (values.at(j) == mem) {
435 phi->init_req(j, phi);
436 } else {
437 phi->init_req(j, values.at(j));
438 }
439 }
440 return phi;
441 }
442
443 // Search the last value stored into the object's field.
444 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
445 assert(adr_t->is_known_instance_field(), "instance required");
446 int instance_id = adr_t->instance_id();
447 assert((uint)instance_id == alloc->_idx, "wrong allocation");
448
449 int alias_idx = C->get_alias_index(adr_t);
450 int offset = adr_t->offset();
451 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
452 Node *alloc_ctrl = alloc->in(TypeFunc::Control);
453 Node *alloc_mem = alloc->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
454 assert(alloc_mem != nullptr, "Allocation without a memory projection.");
455 VectorSet visited;
456
457 bool done = sfpt_mem == alloc_mem;
458 Node *mem = sfpt_mem;
459 while (!done) {
460 if (visited.test_set(mem->_idx)) {
461 return nullptr; // found a loop, give up
462 }
463 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
464 if (mem == start_mem || mem == alloc_mem) {
465 done = true; // hit a sentinel, return appropriate 0 value
466 } else if (mem->is_Initialize()) {
467 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
468 if (mem == nullptr) {
469 done = true; // Something go wrong.
470 } else if (mem->is_Store()) {
471 const TypePtr* atype = mem->as_Store()->adr_type();
472 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
473 done = true;
474 }
475 } else if (mem->is_Store()) {
476 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
477 assert(atype != nullptr, "address type must be oopptr");
478 assert(C->get_alias_index(atype) == alias_idx &&
479 atype->is_known_instance_field() && atype->offset() == offset &&
480 atype->instance_id() == instance_id, "store is correct memory slice");
481 done = true;
482 } else if (mem->is_Phi()) {
483 // try to find a phi's unique input
484 Node *unique_input = nullptr;
485 Node *top = C->top();
486 for (uint i = 1; i < mem->req(); i++) {
487 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
488 if (n == nullptr || n == top || n == mem) {
489 continue;
490 } else if (unique_input == nullptr) {
491 unique_input = n;
492 } else if (unique_input != n) {
493 unique_input = top;
494 break;
495 }
496 }
497 if (unique_input != nullptr && unique_input != top) {
498 mem = unique_input;
499 } else {
500 done = true;
501 }
502 } else if (mem->is_ArrayCopy()) {
503 done = true;
504 } else if (mem->is_top()) {
505 // The slice is on a dead path. Returning nullptr would lead to elimination
506 // bailout, but we want to prevent that. Just forwarding the top is also legal,
507 // and IGVN can just clean things up, and remove whatever receives top.
508 return mem;
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()) {
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->klass_is_exact()) {
574 NOT_PRODUCT(fail_eliminate = "Not an exact type.";)
575 can_eliminate = false;
576 } else if (res_type->isa_aryptr()) {
577 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
578 if (length < 0) {
579 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
580 can_eliminate = false;
581 }
582 }
583 }
584
585 if (can_eliminate && res != nullptr) {
586 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
587 for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
588 j < jmax && can_eliminate; j++) {
589 Node* use = res->fast_out(j);
590
591 if (use->is_AddP()) {
592 const TypePtr* addp_type = igvn->type(use)->is_ptr();
593 int offset = addp_type->offset();
594
595 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
596 NOT_PRODUCT(fail_eliminate = "Undefined field reference";)
597 can_eliminate = false;
598 break;
599 }
600 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
601 k < kmax && can_eliminate; k++) {
602 Node* n = use->fast_out(k);
603 if (n->is_Mem() && n->as_Mem()->is_mismatched_access()) {
604 DEBUG_ONLY(disq_node = n);
605 NOT_PRODUCT(fail_eliminate = "Mismatched access");
606 can_eliminate = false;
607 }
608 if (!n->is_Store() && n->Opcode() != Op_CastP2X && !bs->is_gc_pre_barrier_node(n) && !reduce_merge_precheck) {
609 DEBUG_ONLY(disq_node = n;)
610 if (n->is_Load() || n->is_LoadStore()) {
611 NOT_PRODUCT(fail_eliminate = "Field load";)
612 } else {
613 NOT_PRODUCT(fail_eliminate = "Not store field reference";)
614 }
615 can_eliminate = false;
616 }
617 }
618 } else if (use->is_ArrayCopy() &&
619 (use->as_ArrayCopy()->is_clonebasic() ||
620 use->as_ArrayCopy()->is_arraycopy_validated() ||
621 use->as_ArrayCopy()->is_copyof_validated() ||
622 use->as_ArrayCopy()->is_copyofrange_validated()) &&
623 use->in(ArrayCopyNode::Dest) == res) {
624 // ok to eliminate
625 } else if (use->is_SafePoint()) {
626 SafePointNode* sfpt = use->as_SafePoint();
627 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
628 // Object is passed as argument.
629 DEBUG_ONLY(disq_node = use;)
630 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
631 can_eliminate = false;
632 }
633 Node* sfptMem = sfpt->memory();
634 if (sfptMem == nullptr || sfptMem->is_top()) {
635 DEBUG_ONLY(disq_node = use;)
636 NOT_PRODUCT(fail_eliminate = "null or TOP memory";)
637 can_eliminate = false;
638 } else if (!reduce_merge_precheck) {
639 safepoints->append_if_missing(sfpt);
640 }
641 } else if (reduce_merge_precheck &&
642 (use->is_Phi() || use->is_EncodeP() ||
643 use->Opcode() == Op_MemBarRelease ||
644 (UseStoreStoreForCtor && use->Opcode() == Op_MemBarStoreStore))) {
645 // Nothing to do
646 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
647 if (use->is_Phi()) {
648 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
649 NOT_PRODUCT(fail_eliminate = "Object is return value";)
650 } else {
651 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
652 }
653 DEBUG_ONLY(disq_node = use;)
654 } else {
655 if (use->Opcode() == Op_Return) {
656 NOT_PRODUCT(fail_eliminate = "Object is return value";)
657 } else {
658 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
659 }
660 DEBUG_ONLY(disq_node = use;)
661 }
662 can_eliminate = false;
663 }
664 }
665 }
666
667 #ifndef PRODUCT
668 if (PrintEliminateAllocations && safepoints != nullptr) {
669 if (can_eliminate) {
670 tty->print("Scalar ");
671 if (res == nullptr)
672 alloc->dump();
673 else
674 res->dump();
675 } else if (alloc->_is_scalar_replaceable) {
676 tty->print("NotScalar (%s)", fail_eliminate);
677 if (res == nullptr)
678 alloc->dump();
679 else
680 res->dump();
681 #ifdef ASSERT
682 if (disq_node != nullptr) {
683 tty->print(" >>>> ");
684 disq_node->dump();
685 }
686 #endif /*ASSERT*/
687 }
688 }
689
690 if (TraceReduceAllocationMerges && !can_eliminate && reduce_merge_precheck) {
691 tty->print_cr("\tCan't eliminate allocation because '%s': ", fail_eliminate != nullptr ? fail_eliminate : "");
692 DEBUG_ONLY(if (disq_node != nullptr) disq_node->dump();)
693 }
694 #endif
695 return can_eliminate;
696 }
697
698 void PhaseMacroExpand::undo_previous_scalarizations(GrowableArray <SafePointNode *> safepoints_done, AllocateNode* alloc) {
699 Node* res = alloc->result_cast();
700 int nfields = 0;
701 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
702
703 if (res != nullptr) {
704 const TypeOopPtr* res_type = _igvn.type(res)->isa_oopptr();
705
706 if (res_type->isa_instptr()) {
707 // find the fields of the class which will be needed for safepoint debug information
708 ciInstanceKlass* iklass = res_type->is_instptr()->instance_klass();
709 nfields = iklass->nof_nonstatic_fields();
710 } else {
711 // find the array's elements which will be needed for safepoint debug information
712 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
713 assert(nfields >= 0, "must be an array klass.");
714 }
715 }
716
717 // rollback processed safepoints
718 while (safepoints_done.length() > 0) {
719 SafePointNode* sfpt_done = safepoints_done.pop();
720 // remove any extra entries we added to the safepoint
721 uint last = sfpt_done->req() - 1;
722 for (int k = 0; k < nfields; k++) {
723 sfpt_done->del_req(last--);
724 }
725 JVMState *jvms = sfpt_done->jvms();
726 jvms->set_endoff(sfpt_done->req());
727 // Now make a pass over the debug information replacing any references
728 // to SafePointScalarObjectNode with the allocated object.
729 int start = jvms->debug_start();
730 int end = jvms->debug_end();
731 for (int i = start; i < end; i++) {
732 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
733 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
734 if (scobj->first_index(jvms) == sfpt_done->req() &&
735 scobj->n_fields() == (uint)nfields) {
736 assert(scobj->alloc() == alloc, "sanity");
737 sfpt_done->set_req(i, res);
738 }
739 }
740 }
741 _igvn._worklist.push(sfpt_done);
742 }
743 }
744
745 #ifdef ASSERT
746 // Verify if a value can be written into a field.
747 void verify_type_compatability(const Type* value_type, const Type* field_type) {
748 BasicType value_bt = value_type->basic_type();
749 BasicType field_bt = field_type->basic_type();
750
751 // Primitive types must match.
752 if (is_java_primitive(value_bt) && value_bt == field_bt) { return; }
753
754 // I have been struggling to make a similar assert for non-primitive
755 // types. I we can add one in the future. For now, I just let them
756 // pass without checks.
757 // In particular, I was struggling with a value that came from a call,
758 // and had only a non-null check CastPP. There was also a checkcast
759 // in the graph to verify the interface, but the corresponding
760 // CheckCastPP result was not updated in the stack slot, and so
761 // we ended up using the CastPP. That means that the field knows
762 // that it should get an oop from an interface, but the value lost
763 // that information, and so it is not a subtype.
764 // There may be other issues, feel free to investigate further!
765 if (!is_java_primitive(value_bt)) { return; }
766
767 tty->print_cr("value not compatible for field: %s vs %s",
768 type2name(value_bt),
769 type2name(field_bt));
770 tty->print("value_type: ");
771 value_type->dump();
772 tty->cr();
773 tty->print("field_type: ");
774 field_type->dump();
775 tty->cr();
776 assert(false, "value_type does not fit field_type");
777 }
778 #endif
779
780 SafePointScalarObjectNode* PhaseMacroExpand::create_scalarized_object_description(AllocateNode *alloc, SafePointNode* sfpt) {
781 // Fields of scalar objs are referenced only at the end
782 // of regular debuginfo at the last (youngest) JVMS.
783 // Record relative start index.
784 ciInstanceKlass* iklass = nullptr;
785 BasicType basic_elem_type = T_ILLEGAL;
786 const Type* field_type = nullptr;
787 const TypeOopPtr* res_type = nullptr;
788 int nfields = 0;
789 int array_base = 0;
790 int element_size = 0;
791 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
792 Node* res = alloc->result_cast();
793
794 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
795 assert(sfpt->jvms() != nullptr, "missed JVMS");
796
797 if (res != nullptr) { // Could be null when there are no users
798 res_type = _igvn.type(res)->isa_oopptr();
799
800 if (res_type->isa_instptr()) {
801 // find the fields of the class which will be needed for safepoint debug information
802 iklass = res_type->is_instptr()->instance_klass();
803 nfields = iklass->nof_nonstatic_fields();
804 } else {
805 // find the array's elements which will be needed for safepoint debug information
806 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
807 assert(nfields >= 0, "must be an array klass.");
808 basic_elem_type = res_type->is_aryptr()->elem()->array_element_basic_type();
809 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
810 element_size = type2aelembytes(basic_elem_type);
811 field_type = res_type->is_aryptr()->elem();
812 }
813 }
814
815 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type, alloc, first_ind, sfpt->jvms()->depth(), nfields);
816 sobj->init_req(0, C->root());
817 transform_later(sobj);
818
819 // Scan object's fields adding an input to the safepoint for each field.
820 for (int j = 0; j < nfields; j++) {
821 intptr_t offset;
822 ciField* field = nullptr;
823 if (iklass != nullptr) {
824 field = iklass->nonstatic_field_at(j);
825 offset = field->offset_in_bytes();
826 ciType* elem_type = field->type();
827 basic_elem_type = field->layout_type();
828
829 // The next code is taken from Parse::do_get_xxx().
830 if (is_reference_type(basic_elem_type)) {
831 if (!elem_type->is_loaded()) {
832 field_type = TypeInstPtr::BOTTOM;
833 } else if (field != nullptr && field->is_static_constant()) {
834 ciObject* con = field->constant_value().as_object();
835 // Do not "join" in the previous type; it doesn't add value,
836 // and may yield a vacuous result if the field is of interface type.
837 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
838 assert(field_type != nullptr, "field singleton type must be consistent");
839 } else {
840 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
841 }
842 if (UseCompressedOops) {
843 field_type = field_type->make_narrowoop();
844 basic_elem_type = T_NARROWOOP;
845 }
846 } else {
847 field_type = Type::get_const_basic_type(basic_elem_type);
848 }
849 } else {
850 offset = array_base + j * (intptr_t)element_size;
851 }
852
853 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr();
854
855 Node *field_val = value_from_mem(sfpt->memory(), sfpt->control(), basic_elem_type, field_type, field_addr_type, alloc);
856
857 // We weren't able to find a value for this field,
858 // give up on eliminating this allocation.
859 if (field_val == nullptr) {
860 uint last = sfpt->req() - 1;
861 for (int k = 0; k < j; k++) {
862 sfpt->del_req(last--);
863 }
864 _igvn._worklist.push(sfpt);
865
866 #ifndef PRODUCT
867 if (PrintEliminateAllocations) {
868 if (field != nullptr) {
869 tty->print("=== At SafePoint node %d can't find value of field: ", sfpt->_idx);
870 field->print();
871 int field_idx = C->get_alias_index(field_addr_type);
872 tty->print(" (alias_idx=%d)", field_idx);
873 } else { // Array's element
874 tty->print("=== At SafePoint node %d can't find value of array element [%d]", sfpt->_idx, j);
875 }
876 tty->print(", which prevents elimination of: ");
877 if (res == nullptr)
878 alloc->dump();
879 else
880 res->dump();
881 }
882 #endif
883
884 return nullptr;
885 }
886
887 if (UseCompressedOops && field_type->isa_narrowoop()) {
888 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
889 // to be able scalar replace the allocation.
890 if (field_val->is_EncodeP()) {
891 field_val = field_val->in(1);
892 } else {
893 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
894 }
895 }
896 DEBUG_ONLY(verify_type_compatability(field_val->bottom_type(), field_type);)
897 sfpt->add_req(field_val);
898 }
899
900 sfpt->jvms()->set_endoff(sfpt->req());
901
902 return sobj;
903 }
904
905 // Do scalar replacement.
906 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
907 GrowableArray <SafePointNode *> safepoints_done;
908 Node* res = alloc->result_cast();
909 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
910
911 // Process the safepoint uses
912 while (safepoints.length() > 0) {
913 SafePointNode* sfpt = safepoints.pop();
914 SafePointScalarObjectNode* sobj = create_scalarized_object_description(alloc, sfpt);
915
916 if (sobj == nullptr) {
917 undo_previous_scalarizations(safepoints_done, alloc);
918 return false;
919 }
920
921 // Now make a pass over the debug information replacing any references
922 // to the allocated object with "sobj"
923 JVMState *jvms = sfpt->jvms();
924 sfpt->replace_edges_in_range(res, sobj, jvms->debug_start(), jvms->debug_end(), &_igvn);
925 _igvn._worklist.push(sfpt);
926
927 // keep it for rollback
928 safepoints_done.append_if_missing(sfpt);
929 }
930
931 return true;
932 }
933
934 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
935 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
936 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
937 if (ctl_proj != nullptr) {
938 igvn.replace_node(ctl_proj, n->in(0));
939 }
940 if (mem_proj != nullptr) {
941 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
942 }
943 }
944
945 // Process users of eliminated allocation.
946 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) {
947 Node* res = alloc->result_cast();
948 if (res != nullptr) {
949 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
950 Node *use = res->last_out(j);
951 uint oc1 = res->outcnt();
952
953 if (use->is_AddP()) {
954 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
955 Node *n = use->last_out(k);
956 uint oc2 = use->outcnt();
957 if (n->is_Store()) {
958 #ifdef ASSERT
959 // Verify that there is no dependent MemBarVolatile nodes,
960 // they should be removed during IGVN, see MemBarNode::Ideal().
961 for (DUIterator_Fast pmax, p = n->fast_outs(pmax);
962 p < pmax; p++) {
963 Node* mb = n->fast_out(p);
964 assert(mb->is_Initialize() || !mb->is_MemBar() ||
965 mb->req() <= MemBarNode::Precedent ||
966 mb->in(MemBarNode::Precedent) != n,
967 "MemBarVolatile should be eliminated for non-escaping object");
968 }
969 #endif
970 _igvn.replace_node(n, n->in(MemNode::Memory));
971 } else {
972 eliminate_gc_barrier(n);
973 }
974 k -= (oc2 - use->outcnt());
975 }
976 _igvn.remove_dead_node(use);
977 } else if (use->is_ArrayCopy()) {
978 // Disconnect ArrayCopy node
979 ArrayCopyNode* ac = use->as_ArrayCopy();
980 if (ac->is_clonebasic()) {
981 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
982 disconnect_projections(ac, _igvn);
983 assert(alloc->in(TypeFunc::Memory)->is_Proj() && alloc->in(TypeFunc::Memory)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
984 Node* membar_before = alloc->in(TypeFunc::Memory)->in(0);
985 disconnect_projections(membar_before->as_MemBar(), _igvn);
986 if (membar_after->is_MemBar()) {
987 disconnect_projections(membar_after->as_MemBar(), _igvn);
988 }
989 } else {
990 assert(ac->is_arraycopy_validated() ||
991 ac->is_copyof_validated() ||
992 ac->is_copyofrange_validated(), "unsupported");
993 CallProjections callprojs;
994 ac->extract_projections(&callprojs, true);
995
996 _igvn.replace_node(callprojs.fallthrough_ioproj, ac->in(TypeFunc::I_O));
997 _igvn.replace_node(callprojs.fallthrough_memproj, ac->in(TypeFunc::Memory));
998 _igvn.replace_node(callprojs.fallthrough_catchproj, ac->in(TypeFunc::Control));
999
1000 // Set control to top. IGVN will remove the remaining projections
1001 ac->set_req(0, top());
1002 ac->replace_edge(res, top(), &_igvn);
1003
1004 // Disconnect src right away: it can help find new
1005 // opportunities for allocation elimination
1006 Node* src = ac->in(ArrayCopyNode::Src);
1007 ac->replace_edge(src, top(), &_igvn);
1008 // src can be top at this point if src and dest of the
1009 // arraycopy were the same
1010 if (src->outcnt() == 0 && !src->is_top()) {
1011 _igvn.remove_dead_node(src);
1012 }
1013 }
1014 _igvn._worklist.push(ac);
1015 } else {
1016 eliminate_gc_barrier(use);
1017 }
1018 j -= (oc1 - res->outcnt());
1019 }
1020 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
1021 _igvn.remove_dead_node(res);
1022 }
1023
1024 //
1025 // Process other users of allocation's projections
1026 //
1027 if (_callprojs.resproj != nullptr && _callprojs.resproj->outcnt() != 0) {
1028 // First disconnect stores captured by Initialize node.
1029 // If Initialize node is eliminated first in the following code,
1030 // it will kill such stores and DUIterator_Last will assert.
1031 for (DUIterator_Fast jmax, j = _callprojs.resproj->fast_outs(jmax); j < jmax; j++) {
1032 Node* use = _callprojs.resproj->fast_out(j);
1033 if (use->is_AddP()) {
1034 // raw memory addresses used only by the initialization
1035 _igvn.replace_node(use, C->top());
1036 --j; --jmax;
1037 }
1038 }
1039 for (DUIterator_Last jmin, j = _callprojs.resproj->last_outs(jmin); j >= jmin; ) {
1040 Node* use = _callprojs.resproj->last_out(j);
1041 uint oc1 = _callprojs.resproj->outcnt();
1042 if (use->is_Initialize()) {
1043 // Eliminate Initialize node.
1044 InitializeNode *init = use->as_Initialize();
1045 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
1046 if (ctrl_proj != nullptr) {
1047 _igvn.replace_node(ctrl_proj, init->in(TypeFunc::Control));
1048 #ifdef ASSERT
1049 // If the InitializeNode has no memory out, it will die, and tmp will become null
1050 Node* tmp = init->in(TypeFunc::Control);
1051 assert(tmp == nullptr || tmp == _callprojs.fallthrough_catchproj, "allocation control projection");
1052 #endif
1053 }
1054 Node* mem = init->in(TypeFunc::Memory);
1055 #ifdef ASSERT
1056 if (init->number_of_projs(TypeFunc::Memory) > 0) {
1057 if (mem->is_MergeMem()) {
1058 assert(mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw) == _callprojs.fallthrough_memproj, "allocation memory projection");
1059 } else {
1060 assert(mem == _callprojs.fallthrough_memproj, "allocation memory projection");
1061 }
1062 }
1063 #endif
1064 init->replace_mem_projs_by(mem, &_igvn);
1065 assert(init->outcnt() == 0, "should only have had a control and some memory projections, and we removed them");
1066 } else {
1067 assert(false, "only Initialize or AddP expected");
1068 }
1069 j -= (oc1 - _callprojs.resproj->outcnt());
1070 }
1071 }
1072 if (_callprojs.fallthrough_catchproj != nullptr) {
1073 _igvn.replace_node(_callprojs.fallthrough_catchproj, alloc->in(TypeFunc::Control));
1074 }
1075 if (_callprojs.fallthrough_memproj != nullptr) {
1076 _igvn.replace_node(_callprojs.fallthrough_memproj, alloc->in(TypeFunc::Memory));
1077 }
1078 if (_callprojs.catchall_memproj != nullptr) {
1079 _igvn.replace_node(_callprojs.catchall_memproj, C->top());
1080 }
1081 if (_callprojs.fallthrough_ioproj != nullptr) {
1082 _igvn.replace_node(_callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
1083 }
1084 if (_callprojs.catchall_ioproj != nullptr) {
1085 _igvn.replace_node(_callprojs.catchall_ioproj, C->top());
1086 }
1087 if (_callprojs.catchall_catchproj != nullptr) {
1088 _igvn.replace_node(_callprojs.catchall_catchproj, C->top());
1089 }
1090 }
1091
1092 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1093 // If reallocation fails during deoptimization we'll pop all
1094 // interpreter frames for this compiled frame and that won't play
1095 // nice with JVMTI popframe.
1096 // We avoid this issue by eager reallocation when the popframe request
1097 // is received.
1098 if (!EliminateAllocations || !alloc->_is_non_escaping) {
1099 return false;
1100 }
1101 Node* klass = alloc->in(AllocateNode::KlassNode);
1102 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1103 Node* res = alloc->result_cast();
1104 // Eliminate boxing allocations which are not used
1105 // regardless scalar replaceable status.
1106 bool boxing_alloc = C->eliminate_boxing() &&
1107 tklass->isa_instklassptr() &&
1108 tklass->is_instklassptr()->instance_klass()->is_box_klass();
1109 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != nullptr))) {
1110 return false;
1111 }
1112
1113 alloc->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1114
1115 GrowableArray <SafePointNode *> safepoints;
1116 if (!can_eliminate_allocation(&_igvn, alloc, &safepoints)) {
1117 return false;
1118 }
1119
1120 if (!alloc->_is_scalar_replaceable) {
1121 assert(res == nullptr, "sanity");
1122 // We can only eliminate allocation if all debug info references
1123 // are already replaced with SafePointScalarObject because
1124 // we can't search for a fields value without instance_id.
1125 if (safepoints.length() > 0) {
1126 return false;
1127 }
1128 }
1129
1130 if (!scalar_replacement(alloc, safepoints)) {
1131 return false;
1132 }
1133
1134 CompileLog* log = C->log();
1135 if (log != nullptr) {
1136 log->head("eliminate_allocation type='%d'",
1137 log->identify(tklass->exact_klass()));
1138 JVMState* p = alloc->jvms();
1139 while (p != nullptr) {
1140 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1141 p = p->caller();
1142 }
1143 log->tail("eliminate_allocation");
1144 }
1145
1146 process_users_of_allocation(alloc);
1147
1148 #ifndef PRODUCT
1149 if (PrintEliminateAllocations) {
1150 if (alloc->is_AllocateArray())
1151 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1152 else
1153 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1154 }
1155 #endif
1156
1157 return true;
1158 }
1159
1160 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1161 // EA should remove all uses of non-escaping boxing node.
1162 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != nullptr) {
1163 return false;
1164 }
1165
1166 assert(boxing->result_cast() == nullptr, "unexpected boxing node result");
1167
1168 boxing->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1169
1170 const TypeTuple* r = boxing->tf()->range();
1171 assert(r->cnt() > TypeFunc::Parms, "sanity");
1172 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1173 assert(t != nullptr, "sanity");
1174
1175 CompileLog* log = C->log();
1176 if (log != nullptr) {
1177 log->head("eliminate_boxing type='%d'",
1178 log->identify(t->instance_klass()));
1179 JVMState* p = boxing->jvms();
1180 while (p != nullptr) {
1181 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1182 p = p->caller();
1183 }
1184 log->tail("eliminate_boxing");
1185 }
1186
1187 process_users_of_allocation(boxing);
1188
1189 #ifndef PRODUCT
1190 if (PrintEliminateAllocations) {
1191 tty->print("++++ Eliminated: %d ", boxing->_idx);
1192 boxing->method()->print_short_name(tty);
1193 tty->cr();
1194 }
1195 #endif
1196
1197 return true;
1198 }
1199
1200
1201 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1202 Node* adr = basic_plus_adr(base, offset);
1203 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1204 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered);
1205 transform_later(value);
1206 return value;
1207 }
1208
1209
1210 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1211 Node* adr = basic_plus_adr(base, offset);
1212 mem = StoreNode::make(_igvn, ctl, mem, adr, nullptr, value, bt, MemNode::unordered);
1213 transform_later(mem);
1214 return mem;
1215 }
1216
1217 //=============================================================================
1218 //
1219 // A L L O C A T I O N
1220 //
1221 // Allocation attempts to be fast in the case of frequent small objects.
1222 // It breaks down like this:
1223 //
1224 // 1) Size in doublewords is computed. This is a constant for objects and
1225 // variable for most arrays. Doubleword units are used to avoid size
1226 // overflow of huge doubleword arrays. We need doublewords in the end for
1227 // rounding.
1228 //
1229 // 2) Size is checked for being 'too large'. Too-large allocations will go
1230 // the slow path into the VM. The slow path can throw any required
1231 // exceptions, and does all the special checks for very large arrays. The
1232 // size test can constant-fold away for objects. For objects with
1233 // finalizers it constant-folds the otherway: you always go slow with
1234 // finalizers.
1235 //
1236 // 3) If NOT using TLABs, this is the contended loop-back point.
1237 // Load-Locked the heap top. If using TLABs normal-load the heap top.
1238 //
1239 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
1240 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
1241 // "size*8" we always enter the VM, where "largish" is a constant picked small
1242 // enough that there's always space between the eden max and 4Gig (old space is
1243 // there so it's quite large) and large enough that the cost of entering the VM
1244 // is dwarfed by the cost to initialize the space.
1245 //
1246 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1247 // down. If contended, repeat at step 3. If using TLABs normal-store
1248 // adjusted heap top back down; there is no contention.
1249 //
1250 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
1251 // fields.
1252 //
1253 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1254 // oop flavor.
1255 //
1256 //=============================================================================
1257 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1258 // Allocations bigger than this always go the slow route.
1259 // This value must be small enough that allocation attempts that need to
1260 // trigger exceptions go the slow route. Also, it must be small enough so
1261 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1262 //=============================================================================j//
1263 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1264 // The allocator will coalesce int->oop copies away. See comment in
1265 // coalesce.cpp about how this works. It depends critically on the exact
1266 // code shape produced here, so if you are changing this code shape
1267 // make sure the GC info for the heap-top is correct in and around the
1268 // slow-path call.
1269 //
1270
1271 void PhaseMacroExpand::expand_allocate_common(
1272 AllocateNode* alloc, // allocation node to be expanded
1273 Node* length, // array length for an array allocation
1274 const TypeFunc* slow_call_type, // Type of slow call
1275 address slow_call_address, // Address of slow call
1276 Node* valid_length_test // whether length is valid or not
1277 )
1278 {
1279 Node* ctrl = alloc->in(TypeFunc::Control);
1280 Node* mem = alloc->in(TypeFunc::Memory);
1281 Node* i_o = alloc->in(TypeFunc::I_O);
1282 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1283 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1284 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1285 assert(ctrl != nullptr, "must have control");
1286
1287 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1288 // they will not be used if "always_slow" is set
1289 enum { slow_result_path = 1, fast_result_path = 2 };
1290 Node *result_region = nullptr;
1291 Node *result_phi_rawmem = nullptr;
1292 Node *result_phi_rawoop = nullptr;
1293 Node *result_phi_i_o = nullptr;
1294
1295 // The initial slow comparison is a size check, the comparison
1296 // we want to do is a BoolTest::gt
1297 bool expand_fast_path = true;
1298 int tv = _igvn.find_int_con(initial_slow_test, -1);
1299 if (tv >= 0) {
1300 // InitialTest has constant result
1301 // 0 - can fit in TLAB
1302 // 1 - always too big or negative
1303 assert(tv <= 1, "0 or 1 if a constant");
1304 expand_fast_path = (tv == 0);
1305 initial_slow_test = nullptr;
1306 } else {
1307 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1308 }
1309
1310 if (!UseTLAB) {
1311 // Force slow-path allocation
1312 expand_fast_path = false;
1313 initial_slow_test = nullptr;
1314 }
1315
1316 bool allocation_has_use = (alloc->result_cast() != nullptr);
1317 if (!allocation_has_use) {
1318 InitializeNode* init = alloc->initialization();
1319 if (init != nullptr) {
1320 init->remove(&_igvn);
1321 }
1322 if (expand_fast_path && (initial_slow_test == nullptr)) {
1323 // Remove allocation node and return.
1324 // Size is a non-negative constant -> no initial check needed -> directly to fast path.
1325 // Also, no usages -> empty fast path -> no fall out to slow path -> nothing left.
1326 #ifndef PRODUCT
1327 if (PrintEliminateAllocations) {
1328 tty->print("NotUsed ");
1329 Node* res = alloc->proj_out_or_null(TypeFunc::Parms);
1330 if (res != nullptr) {
1331 res->dump();
1332 } else {
1333 alloc->dump();
1334 }
1335 }
1336 #endif
1337 yank_alloc_node(alloc);
1338 return;
1339 }
1340 }
1341
1342 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1343 Node *slow_region = nullptr;
1344 Node *toobig_false = ctrl;
1345
1346 // generate the initial test if necessary
1347 if (initial_slow_test != nullptr ) {
1348 assert (expand_fast_path, "Only need test if there is a fast path");
1349 slow_region = new RegionNode(3);
1350
1351 // Now make the initial failure test. Usually a too-big test but
1352 // might be a TRUE for finalizers.
1353 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1354 transform_later(toobig_iff);
1355 // Plug the failing-too-big test into the slow-path region
1356 Node *toobig_true = new IfTrueNode( toobig_iff );
1357 transform_later(toobig_true);
1358 slow_region ->init_req( too_big_or_final_path, toobig_true );
1359 toobig_false = new IfFalseNode( toobig_iff );
1360 transform_later(toobig_false);
1361 } else {
1362 // No initial test, just fall into next case
1363 assert(allocation_has_use || !expand_fast_path, "Should already have been handled");
1364 toobig_false = ctrl;
1365 DEBUG_ONLY(slow_region = NodeSentinel);
1366 }
1367
1368 // If we are here there are several possibilities
1369 // - expand_fast_path is false - then only a slow path is expanded. That's it.
1370 // no_initial_check means a constant allocation.
1371 // - If check always evaluates to false -> expand_fast_path is false (see above)
1372 // - If check always evaluates to true -> directly into fast path (but may bailout to slowpath)
1373 // if !allocation_has_use the fast path is empty
1374 // if !allocation_has_use && no_initial_check
1375 // - Then there are no fastpath that can fall out to slowpath -> no allocation code at all.
1376 // removed by yank_alloc_node above.
1377
1378 Node *slow_mem = mem; // save the current memory state for slow path
1379 // generate the fast allocation code unless we know that the initial test will always go slow
1380 if (expand_fast_path) {
1381 // Fast path modifies only raw memory.
1382 if (mem->is_MergeMem()) {
1383 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1384 }
1385
1386 // allocate the Region and Phi nodes for the result
1387 result_region = new RegionNode(3);
1388 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1389 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1390
1391 // Grab regular I/O before optional prefetch may change it.
1392 // Slow-path does no I/O so just set it to the original I/O.
1393 result_phi_i_o->init_req(slow_result_path, i_o);
1394
1395 // Name successful fast-path variables
1396 Node* fast_oop_ctrl;
1397 Node* fast_oop_rawmem;
1398 if (allocation_has_use) {
1399 Node* needgc_ctrl = nullptr;
1400 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1401
1402 intx prefetch_lines = length != nullptr ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1403 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1404 Node* fast_oop = bs->obj_allocate(this, mem, toobig_false, size_in_bytes, i_o, needgc_ctrl,
1405 fast_oop_ctrl, fast_oop_rawmem,
1406 prefetch_lines);
1407
1408 if (initial_slow_test != nullptr) {
1409 // This completes all paths into the slow merge point
1410 slow_region->init_req(need_gc_path, needgc_ctrl);
1411 transform_later(slow_region);
1412 } else {
1413 // No initial slow path needed!
1414 // Just fall from the need-GC path straight into the VM call.
1415 slow_region = needgc_ctrl;
1416 }
1417
1418 InitializeNode* init = alloc->initialization();
1419 fast_oop_rawmem = initialize_object(alloc,
1420 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1421 klass_node, length, size_in_bytes);
1422 expand_initialize_membar(alloc, init, fast_oop_ctrl, fast_oop_rawmem);
1423 expand_dtrace_alloc_probe(alloc, fast_oop, fast_oop_ctrl, fast_oop_rawmem);
1424
1425 result_phi_rawoop->init_req(fast_result_path, fast_oop);
1426 } else {
1427 assert (initial_slow_test != nullptr, "sanity");
1428 fast_oop_ctrl = toobig_false;
1429 fast_oop_rawmem = mem;
1430 transform_later(slow_region);
1431 }
1432
1433 // Plug in the successful fast-path into the result merge point
1434 result_region ->init_req(fast_result_path, fast_oop_ctrl);
1435 result_phi_i_o ->init_req(fast_result_path, i_o);
1436 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1437 } else {
1438 slow_region = ctrl;
1439 result_phi_i_o = i_o; // Rename it to use in the following code.
1440 }
1441
1442 // Generate slow-path call
1443 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1444 OptoRuntime::stub_name(slow_call_address),
1445 TypePtr::BOTTOM);
1446 call->init_req(TypeFunc::Control, slow_region);
1447 call->init_req(TypeFunc::I_O, top()); // does no i/o
1448 call->init_req(TypeFunc::Memory, slow_mem); // may gc ptrs
1449 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1450 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1451
1452 call->init_req(TypeFunc::Parms+0, klass_node);
1453 if (length != nullptr) {
1454 call->init_req(TypeFunc::Parms+1, length);
1455 }
1456
1457 // Copy debug information and adjust JVMState information, then replace
1458 // allocate node with the call
1459 call->copy_call_debug_info(&_igvn, alloc);
1460 // For array allocations, copy the valid length check to the call node so Compile::final_graph_reshaping() can verify
1461 // that the call has the expected number of CatchProj nodes (in case the allocation always fails and the fallthrough
1462 // path dies).
1463 if (valid_length_test != nullptr) {
1464 call->add_req(valid_length_test);
1465 }
1466 if (expand_fast_path) {
1467 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1468 } else {
1469 // Hook i_o projection to avoid its elimination during allocation
1470 // replacement (when only a slow call is generated).
1471 call->set_req(TypeFunc::I_O, result_phi_i_o);
1472 }
1473 _igvn.replace_node(alloc, call);
1474 transform_later(call);
1475
1476 // Identify the output projections from the allocate node and
1477 // adjust any references to them.
1478 // The control and io projections look like:
1479 //
1480 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1481 // Allocate Catch
1482 // ^---Proj(io) <-------+ ^---CatchProj(io)
1483 //
1484 // We are interested in the CatchProj nodes.
1485 //
1486 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1487
1488 // An allocate node has separate memory projections for the uses on
1489 // the control and i_o paths. Replace the control memory projection with
1490 // result_phi_rawmem (unless we are only generating a slow call when
1491 // both memory projections are combined)
1492 if (expand_fast_path && _callprojs.fallthrough_memproj != nullptr) {
1493 migrate_outs(_callprojs.fallthrough_memproj, result_phi_rawmem);
1494 }
1495 // Now change uses of catchall_memproj to use fallthrough_memproj and delete
1496 // catchall_memproj so we end up with a call that has only 1 memory projection.
1497 if (_callprojs.catchall_memproj != nullptr ) {
1498 if (_callprojs.fallthrough_memproj == nullptr) {
1499 _callprojs.fallthrough_memproj = new ProjNode(call, TypeFunc::Memory);
1500 transform_later(_callprojs.fallthrough_memproj);
1501 }
1502 migrate_outs(_callprojs.catchall_memproj, _callprojs.fallthrough_memproj);
1503 _igvn.remove_dead_node(_callprojs.catchall_memproj);
1504 }
1505
1506 // An allocate node has separate i_o projections for the uses on the control
1507 // and i_o paths. Always replace the control i_o projection with result i_o
1508 // otherwise incoming i_o become dead when only a slow call is generated
1509 // (it is different from memory projections where both projections are
1510 // combined in such case).
1511 if (_callprojs.fallthrough_ioproj != nullptr) {
1512 migrate_outs(_callprojs.fallthrough_ioproj, result_phi_i_o);
1513 }
1514 // Now change uses of catchall_ioproj to use fallthrough_ioproj and delete
1515 // catchall_ioproj so we end up with a call that has only 1 i_o projection.
1516 if (_callprojs.catchall_ioproj != nullptr ) {
1517 if (_callprojs.fallthrough_ioproj == nullptr) {
1518 _callprojs.fallthrough_ioproj = new ProjNode(call, TypeFunc::I_O);
1519 transform_later(_callprojs.fallthrough_ioproj);
1520 }
1521 migrate_outs(_callprojs.catchall_ioproj, _callprojs.fallthrough_ioproj);
1522 _igvn.remove_dead_node(_callprojs.catchall_ioproj);
1523 }
1524
1525 // if we generated only a slow call, we are done
1526 if (!expand_fast_path) {
1527 // Now we can unhook i_o.
1528 if (result_phi_i_o->outcnt() > 1) {
1529 call->set_req(TypeFunc::I_O, top());
1530 } else {
1531 assert(result_phi_i_o->unique_ctrl_out() == call, "sanity");
1532 // Case of new array with negative size known during compilation.
1533 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1534 // following code since call to runtime will throw exception.
1535 // As result there will be no users of i_o after the call.
1536 // Leave i_o attached to this call to avoid problems in preceding graph.
1537 }
1538 return;
1539 }
1540
1541 if (_callprojs.fallthrough_catchproj != nullptr) {
1542 ctrl = _callprojs.fallthrough_catchproj->clone();
1543 transform_later(ctrl);
1544 _igvn.replace_node(_callprojs.fallthrough_catchproj, result_region);
1545 } else {
1546 ctrl = top();
1547 }
1548 Node *slow_result;
1549 if (_callprojs.resproj == nullptr) {
1550 // no uses of the allocation result
1551 slow_result = top();
1552 } else {
1553 slow_result = _callprojs.resproj->clone();
1554 transform_later(slow_result);
1555 _igvn.replace_node(_callprojs.resproj, result_phi_rawoop);
1556 }
1557
1558 // Plug slow-path into result merge point
1559 result_region->init_req( slow_result_path, ctrl);
1560 transform_later(result_region);
1561 if (allocation_has_use) {
1562 result_phi_rawoop->init_req(slow_result_path, slow_result);
1563 transform_later(result_phi_rawoop);
1564 }
1565 result_phi_rawmem->init_req(slow_result_path, _callprojs.fallthrough_memproj);
1566 transform_later(result_phi_rawmem);
1567 transform_later(result_phi_i_o);
1568 // This completes all paths into the result merge point
1569 }
1570
1571 // Remove alloc node that has no uses.
1572 void PhaseMacroExpand::yank_alloc_node(AllocateNode* alloc) {
1573 Node* ctrl = alloc->in(TypeFunc::Control);
1574 Node* mem = alloc->in(TypeFunc::Memory);
1575 Node* i_o = alloc->in(TypeFunc::I_O);
1576
1577 alloc->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1578 if (_callprojs.resproj != nullptr) {
1579 for (DUIterator_Fast imax, i = _callprojs.resproj->fast_outs(imax); i < imax; i++) {
1580 Node* use = _callprojs.resproj->fast_out(i);
1581 use->isa_MemBar()->remove(&_igvn);
1582 --imax;
1583 --i; // back up iterator
1584 }
1585 assert(_callprojs.resproj->outcnt() == 0, "all uses must be deleted");
1586 _igvn.remove_dead_node(_callprojs.resproj);
1587 }
1588 if (_callprojs.fallthrough_catchproj != nullptr) {
1589 migrate_outs(_callprojs.fallthrough_catchproj, ctrl);
1590 _igvn.remove_dead_node(_callprojs.fallthrough_catchproj);
1591 }
1592 if (_callprojs.catchall_catchproj != nullptr) {
1593 _igvn.rehash_node_delayed(_callprojs.catchall_catchproj);
1594 _callprojs.catchall_catchproj->set_req(0, top());
1595 }
1596 if (_callprojs.fallthrough_proj != nullptr) {
1597 Node* catchnode = _callprojs.fallthrough_proj->unique_ctrl_out();
1598 _igvn.remove_dead_node(catchnode);
1599 _igvn.remove_dead_node(_callprojs.fallthrough_proj);
1600 }
1601 if (_callprojs.fallthrough_memproj != nullptr) {
1602 migrate_outs(_callprojs.fallthrough_memproj, mem);
1603 _igvn.remove_dead_node(_callprojs.fallthrough_memproj);
1604 }
1605 if (_callprojs.fallthrough_ioproj != nullptr) {
1606 migrate_outs(_callprojs.fallthrough_ioproj, i_o);
1607 _igvn.remove_dead_node(_callprojs.fallthrough_ioproj);
1608 }
1609 if (_callprojs.catchall_memproj != nullptr) {
1610 _igvn.rehash_node_delayed(_callprojs.catchall_memproj);
1611 _callprojs.catchall_memproj->set_req(0, top());
1612 }
1613 if (_callprojs.catchall_ioproj != nullptr) {
1614 _igvn.rehash_node_delayed(_callprojs.catchall_ioproj);
1615 _callprojs.catchall_ioproj->set_req(0, top());
1616 }
1617 #ifndef PRODUCT
1618 if (PrintEliminateAllocations) {
1619 if (alloc->is_AllocateArray()) {
1620 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1621 } else {
1622 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1623 }
1624 }
1625 #endif
1626 _igvn.remove_dead_node(alloc);
1627 }
1628
1629 void PhaseMacroExpand::expand_initialize_membar(AllocateNode* alloc, InitializeNode* init,
1630 Node*& fast_oop_ctrl, Node*& fast_oop_rawmem) {
1631 // If initialization is performed by an array copy, any required
1632 // MemBarStoreStore was already added. If the object does not
1633 // escape no need for a MemBarStoreStore. If the object does not
1634 // escape in its initializer and memory barrier (MemBarStoreStore or
1635 // stronger) is already added at exit of initializer, also no need
1636 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore
1637 // so that stores that initialize this object can't be reordered
1638 // with a subsequent store that makes this object accessible by
1639 // other threads.
1640 // Other threads include java threads and JVM internal threads
1641 // (for example concurrent GC threads). Current concurrent GC
1642 // implementation: G1 will not scan newly created object,
1643 // so it's safe to skip storestore barrier when allocation does
1644 // not escape.
1645 if (!alloc->does_not_escape_thread() &&
1646 !alloc->is_allocation_MemBar_redundant() &&
1647 (init == nullptr || !init->is_complete_with_arraycopy())) {
1648 if (init == nullptr || init->req() < InitializeNode::RawStores) {
1649 // No InitializeNode or no stores captured by zeroing
1650 // elimination. Simply add the MemBarStoreStore after object
1651 // initialization.
1652 // What we want is to prevent the compiler and the CPU from re-ordering the stores that initialize this object
1653 // with subsequent stores to any slice. As a consequence, this MemBar should capture the entire memory state at
1654 // this point in the IR and produce a new memory state that should cover all slices. However, the Initialize node
1655 // only captures/produces a partial memory state making it complicated to insert such a MemBar. Because
1656 // re-ordering by the compiler can't happen by construction (a later Store that publishes the just allocated
1657 // object reference is indirectly control dependent on the Initialize node), preventing reordering by the CPU is
1658 // sufficient. For that a MemBar on the raw memory slice is good enough.
1659 // If init is null, this allocation does have an InitializeNode but this logic can't locate it (see comment in
1660 // PhaseMacroExpand::initialize_object()).
1661 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxRaw);
1662 transform_later(mb);
1663
1664 mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
1665 mb->init_req(TypeFunc::Control, fast_oop_ctrl);
1666 fast_oop_ctrl = new ProjNode(mb, TypeFunc::Control);
1667 transform_later(fast_oop_ctrl);
1668 fast_oop_rawmem = new ProjNode(mb, TypeFunc::Memory);
1669 transform_later(fast_oop_rawmem);
1670 } else {
1671 // Add the MemBarStoreStore after the InitializeNode so that
1672 // all stores performing the initialization that were moved
1673 // before the InitializeNode happen before the storestore
1674 // barrier.
1675
1676 Node* init_ctrl = init->proj_out_or_null(TypeFunc::Control);
1677
1678 // See comment above that explains why a raw memory MemBar is good enough.
1679 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxRaw);
1680 transform_later(mb);
1681
1682 Node* ctrl = new ProjNode(init, TypeFunc::Control);
1683 transform_later(ctrl);
1684 Node* old_raw_mem_proj = nullptr;
1685 auto find_raw_mem = [&](ProjNode* proj) {
1686 if (C->get_alias_index(proj->adr_type()) == Compile::AliasIdxRaw) {
1687 assert(old_raw_mem_proj == nullptr, "only one expected");
1688 old_raw_mem_proj = proj;
1689 }
1690 };
1691 init->for_each_proj(find_raw_mem, TypeFunc::Memory);
1692 assert(old_raw_mem_proj != nullptr, "should have found raw mem Proj");
1693 Node* raw_mem_proj = new ProjNode(init, TypeFunc::Memory);
1694 transform_later(raw_mem_proj);
1695
1696 // The MemBarStoreStore depends on control and memory coming
1697 // from the InitializeNode
1698 mb->init_req(TypeFunc::Memory, raw_mem_proj);
1699 mb->init_req(TypeFunc::Control, ctrl);
1700
1701 ctrl = new ProjNode(mb, TypeFunc::Control);
1702 transform_later(ctrl);
1703 Node* mem = new ProjNode(mb, TypeFunc::Memory);
1704 transform_later(mem);
1705
1706 // All nodes that depended on the InitializeNode for control
1707 // and memory must now depend on the MemBarNode that itself
1708 // depends on the InitializeNode
1709 if (init_ctrl != nullptr) {
1710 _igvn.replace_node(init_ctrl, ctrl);
1711 }
1712 _igvn.replace_node(old_raw_mem_proj, mem);
1713 }
1714 }
1715 }
1716
1717 void PhaseMacroExpand::expand_dtrace_alloc_probe(AllocateNode* alloc, Node* oop,
1718 Node*& ctrl, Node*& rawmem) {
1719 if (C->env()->dtrace_alloc_probes()) {
1720 // Slow-path call
1721 int size = TypeFunc::Parms + 2;
1722 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1723 CAST_FROM_FN_PTR(address,
1724 static_cast<int (*)(JavaThread*, oopDesc*)>(SharedRuntime::dtrace_object_alloc)),
1725 "dtrace_object_alloc",
1726 TypeRawPtr::BOTTOM);
1727
1728 // Get base of thread-local storage area
1729 Node* thread = new ThreadLocalNode();
1730 transform_later(thread);
1731
1732 call->init_req(TypeFunc::Parms + 0, thread);
1733 call->init_req(TypeFunc::Parms + 1, oop);
1734 call->init_req(TypeFunc::Control, ctrl);
1735 call->init_req(TypeFunc::I_O , top()); // does no i/o
1736 call->init_req(TypeFunc::Memory , rawmem);
1737 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1738 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1739 transform_later(call);
1740 ctrl = new ProjNode(call, TypeFunc::Control);
1741 transform_later(ctrl);
1742 rawmem = new ProjNode(call, TypeFunc::Memory);
1743 transform_later(rawmem);
1744 }
1745 }
1746
1747 // Helper for PhaseMacroExpand::expand_allocate_common.
1748 // Initializes the newly-allocated storage.
1749 Node*
1750 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1751 Node* control, Node* rawmem, Node* object,
1752 Node* klass_node, Node* length,
1753 Node* size_in_bytes) {
1754 InitializeNode* init = alloc->initialization();
1755 // Store the klass & mark bits
1756 Node* mark_node = alloc->make_ideal_mark(&_igvn, object, control, rawmem);
1757 if (!mark_node->is_Con()) {
1758 transform_later(mark_node);
1759 }
1760 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, TypeX_X->basic_type());
1761
1762 if (!UseCompactObjectHeaders) {
1763 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1764 }
1765 int header_size = alloc->minimum_header_size(); // conservatively small
1766
1767 // Array length
1768 if (length != nullptr) { // Arrays need length field
1769 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1770 // conservatively small header size:
1771 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1772 if (_igvn.type(klass_node)->isa_aryklassptr()) { // we know the exact header size in most cases:
1773 BasicType elem = _igvn.type(klass_node)->is_klassptr()->as_instance_type()->isa_aryptr()->elem()->array_element_basic_type();
1774 if (is_reference_type(elem, true)) {
1775 elem = T_OBJECT;
1776 }
1777 header_size = Klass::layout_helper_header_size(Klass::array_layout_helper(elem));
1778 }
1779 }
1780
1781 // Clear the object body, if necessary.
1782 if (init == nullptr) {
1783 // The init has somehow disappeared; be cautious and clear everything.
1784 //
1785 // This can happen if a node is allocated but an uncommon trap occurs
1786 // immediately. In this case, the Initialize gets associated with the
1787 // trap, and may be placed in a different (outer) loop, if the Allocate
1788 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1789 // there can be two Allocates to one Initialize. The answer in all these
1790 // edge cases is safety first. It is always safe to clear immediately
1791 // within an Allocate, and then (maybe or maybe not) clear some more later.
1792 if (!(UseTLAB && ZeroTLAB)) {
1793 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1794 header_size, size_in_bytes,
1795 &_igvn);
1796 }
1797 } else {
1798 if (!init->is_complete()) {
1799 // Try to win by zeroing only what the init does not store.
1800 // We can also try to do some peephole optimizations,
1801 // such as combining some adjacent subword stores.
1802 rawmem = init->complete_stores(control, rawmem, object,
1803 header_size, size_in_bytes, &_igvn);
1804 }
1805 // We have no more use for this link, since the AllocateNode goes away:
1806 init->set_req(InitializeNode::RawAddress, top());
1807 // (If we keep the link, it just confuses the register allocator,
1808 // who thinks he sees a real use of the address by the membar.)
1809 }
1810
1811 return rawmem;
1812 }
1813
1814 // Generate prefetch instructions for next allocations.
1815 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1816 Node*& contended_phi_rawmem,
1817 Node* old_eden_top, Node* new_eden_top,
1818 intx lines) {
1819 enum { fall_in_path = 1, pf_path = 2 };
1820 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1821 // Generate prefetch allocation with watermark check.
1822 // As an allocation hits the watermark, we will prefetch starting
1823 // at a "distance" away from watermark.
1824
1825 Node *pf_region = new RegionNode(3);
1826 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY,
1827 TypeRawPtr::BOTTOM );
1828 // I/O is used for Prefetch
1829 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO );
1830
1831 Node *thread = new ThreadLocalNode();
1832 transform_later(thread);
1833
1834 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread,
1835 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1836 transform_later(eden_pf_adr);
1837
1838 Node *old_pf_wm = new LoadPNode(needgc_false,
1839 contended_phi_rawmem, eden_pf_adr,
1840 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM,
1841 MemNode::unordered);
1842 transform_later(old_pf_wm);
1843
1844 // check against new_eden_top
1845 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm );
1846 transform_later(need_pf_cmp);
1847 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge );
1848 transform_later(need_pf_bol);
1849 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol,
1850 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1851 transform_later(need_pf_iff);
1852
1853 // true node, add prefetchdistance
1854 Node *need_pf_true = new IfTrueNode( need_pf_iff );
1855 transform_later(need_pf_true);
1856
1857 Node *need_pf_false = new IfFalseNode( need_pf_iff );
1858 transform_later(need_pf_false);
1859
1860 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm,
1861 _igvn.MakeConX(AllocatePrefetchDistance) );
1862 transform_later(new_pf_wmt );
1863 new_pf_wmt->set_req(0, need_pf_true);
1864
1865 Node *store_new_wmt = new StorePNode(need_pf_true,
1866 contended_phi_rawmem, eden_pf_adr,
1867 TypeRawPtr::BOTTOM, new_pf_wmt,
1868 MemNode::unordered);
1869 transform_later(store_new_wmt);
1870
1871 // adding prefetches
1872 pf_phi_abio->init_req( fall_in_path, i_o );
1873
1874 Node *prefetch_adr;
1875 Node *prefetch;
1876 uint step_size = AllocatePrefetchStepSize;
1877 uint distance = 0;
1878
1879 for ( intx i = 0; i < lines; i++ ) {
1880 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt,
1881 _igvn.MakeConX(distance) );
1882 transform_later(prefetch_adr);
1883 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1884 transform_later(prefetch);
1885 distance += step_size;
1886 i_o = prefetch;
1887 }
1888 pf_phi_abio->set_req( pf_path, i_o );
1889
1890 pf_region->init_req( fall_in_path, need_pf_false );
1891 pf_region->init_req( pf_path, need_pf_true );
1892
1893 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
1894 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
1895
1896 transform_later(pf_region);
1897 transform_later(pf_phi_rawmem);
1898 transform_later(pf_phi_abio);
1899
1900 needgc_false = pf_region;
1901 contended_phi_rawmem = pf_phi_rawmem;
1902 i_o = pf_phi_abio;
1903 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) {
1904 // Insert a prefetch instruction for each allocation.
1905 // This code is used to generate 1 prefetch instruction per cache line.
1906
1907 // Generate several prefetch instructions.
1908 uint step_size = AllocatePrefetchStepSize;
1909 uint distance = AllocatePrefetchDistance;
1910
1911 // Next cache address.
1912 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top,
1913 _igvn.MakeConX(step_size + distance));
1914 transform_later(cache_adr);
1915 cache_adr = new CastP2XNode(needgc_false, cache_adr);
1916 transform_later(cache_adr);
1917 // Address is aligned to execute prefetch to the beginning of cache line size.
1918 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
1919 cache_adr = new AndXNode(cache_adr, mask);
1920 transform_later(cache_adr);
1921 cache_adr = new CastX2PNode(cache_adr);
1922 transform_later(cache_adr);
1923
1924 // Prefetch
1925 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr );
1926 prefetch->set_req(0, needgc_false);
1927 transform_later(prefetch);
1928 contended_phi_rawmem = prefetch;
1929 Node *prefetch_adr;
1930 distance = step_size;
1931 for ( intx i = 1; i < lines; i++ ) {
1932 prefetch_adr = new AddPNode( cache_adr, cache_adr,
1933 _igvn.MakeConX(distance) );
1934 transform_later(prefetch_adr);
1935 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr );
1936 transform_later(prefetch);
1937 distance += step_size;
1938 contended_phi_rawmem = prefetch;
1939 }
1940 } else if( AllocatePrefetchStyle > 0 ) {
1941 // Insert a prefetch for each allocation only on the fast-path
1942 Node *prefetch_adr;
1943 Node *prefetch;
1944 // Generate several prefetch instructions.
1945 uint step_size = AllocatePrefetchStepSize;
1946 uint distance = AllocatePrefetchDistance;
1947 for ( intx i = 0; i < lines; i++ ) {
1948 prefetch_adr = new AddPNode( old_eden_top, new_eden_top,
1949 _igvn.MakeConX(distance) );
1950 transform_later(prefetch_adr);
1951 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1952 // Do not let it float too high, since if eden_top == eden_end,
1953 // both might be null.
1954 if( i == 0 ) { // Set control for first prefetch, next follows it
1955 prefetch->init_req(0, needgc_false);
1956 }
1957 transform_later(prefetch);
1958 distance += step_size;
1959 i_o = prefetch;
1960 }
1961 }
1962 return i_o;
1963 }
1964
1965
1966 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
1967 expand_allocate_common(alloc, nullptr,
1968 OptoRuntime::new_instance_Type(),
1969 OptoRuntime::new_instance_Java(), nullptr);
1970 }
1971
1972 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
1973 Node* length = alloc->in(AllocateNode::ALength);
1974 Node* valid_length_test = alloc->in(AllocateNode::ValidLengthTest);
1975 InitializeNode* init = alloc->initialization();
1976 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1977 const TypeAryKlassPtr* ary_klass_t = _igvn.type(klass_node)->isa_aryklassptr();
1978 address slow_call_address; // Address of slow call
1979 if (init != nullptr && init->is_complete_with_arraycopy() &&
1980 ary_klass_t && ary_klass_t->elem()->isa_klassptr() == nullptr) {
1981 // Don't zero type array during slow allocation in VM since
1982 // it will be initialized later by arraycopy in compiled code.
1983 slow_call_address = OptoRuntime::new_array_nozero_Java();
1984 } else {
1985 slow_call_address = OptoRuntime::new_array_Java();
1986 }
1987 expand_allocate_common(alloc, length,
1988 OptoRuntime::new_array_Type(),
1989 slow_call_address, valid_length_test);
1990 }
1991
1992 //-------------------mark_eliminated_box----------------------------------
1993 //
1994 // During EA obj may point to several objects but after few ideal graph
1995 // transformations (CCP) it may point to only one non escaping object
1996 // (but still using phi), corresponding locks and unlocks will be marked
1997 // for elimination. Later obj could be replaced with a new node (new phi)
1998 // and which does not have escape information. And later after some graph
1999 // reshape other locks and unlocks (which were not marked for elimination
2000 // before) are connected to this new obj (phi) but they still will not be
2001 // marked for elimination since new obj has no escape information.
2002 // Mark all associated (same box and obj) lock and unlock nodes for
2003 // elimination if some of them marked already.
2004 void PhaseMacroExpand::mark_eliminated_box(Node* box, Node* obj) {
2005 BoxLockNode* oldbox = box->as_BoxLock();
2006 if (oldbox->is_eliminated()) {
2007 return; // This BoxLock node was processed already.
2008 }
2009 assert(!oldbox->is_unbalanced(), "this should not be called for unbalanced region");
2010 // New implementation (EliminateNestedLocks) has separate BoxLock
2011 // node for each locked region so mark all associated locks/unlocks as
2012 // eliminated even if different objects are referenced in one locked region
2013 // (for example, OSR compilation of nested loop inside locked scope).
2014 if (EliminateNestedLocks ||
2015 oldbox->as_BoxLock()->is_simple_lock_region(nullptr, obj, nullptr)) {
2016 // Box is used only in one lock region. Mark this box as eliminated.
2017 oldbox->set_local(); // This verifies correct state of BoxLock
2018 _igvn.hash_delete(oldbox);
2019 oldbox->set_eliminated(); // This changes box's hash value
2020 _igvn.hash_insert(oldbox);
2021
2022 for (uint i = 0; i < oldbox->outcnt(); i++) {
2023 Node* u = oldbox->raw_out(i);
2024 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
2025 AbstractLockNode* alock = u->as_AbstractLock();
2026 // Check lock's box since box could be referenced by Lock's debug info.
2027 if (alock->box_node() == oldbox) {
2028 // Mark eliminated all related locks and unlocks.
2029 #ifdef ASSERT
2030 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4");
2031 #endif
2032 alock->set_non_esc_obj();
2033 }
2034 }
2035 }
2036 return;
2037 }
2038
2039 // Create new "eliminated" BoxLock node and use it in monitor debug info
2040 // instead of oldbox for the same object.
2041 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
2042
2043 // Note: BoxLock node is marked eliminated only here and it is used
2044 // to indicate that all associated lock and unlock nodes are marked
2045 // for elimination.
2046 newbox->set_local(); // This verifies correct state of BoxLock
2047 newbox->set_eliminated();
2048 transform_later(newbox);
2049
2050 // Replace old box node with new box for all users of the same object.
2051 for (uint i = 0; i < oldbox->outcnt();) {
2052 bool next_edge = true;
2053
2054 Node* u = oldbox->raw_out(i);
2055 if (u->is_AbstractLock()) {
2056 AbstractLockNode* alock = u->as_AbstractLock();
2057 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
2058 // Replace Box and mark eliminated all related locks and unlocks.
2059 #ifdef ASSERT
2060 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5");
2061 #endif
2062 alock->set_non_esc_obj();
2063 _igvn.rehash_node_delayed(alock);
2064 alock->set_box_node(newbox);
2065 next_edge = false;
2066 }
2067 }
2068 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
2069 FastLockNode* flock = u->as_FastLock();
2070 assert(flock->box_node() == oldbox, "sanity");
2071 _igvn.rehash_node_delayed(flock);
2072 flock->set_box_node(newbox);
2073 next_edge = false;
2074 }
2075
2076 // Replace old box in monitor debug info.
2077 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
2078 SafePointNode* sfn = u->as_SafePoint();
2079 JVMState* youngest_jvms = sfn->jvms();
2080 int max_depth = youngest_jvms->depth();
2081 for (int depth = 1; depth <= max_depth; depth++) {
2082 JVMState* jvms = youngest_jvms->of_depth(depth);
2083 int num_mon = jvms->nof_monitors();
2084 // Loop over monitors
2085 for (int idx = 0; idx < num_mon; idx++) {
2086 Node* obj_node = sfn->monitor_obj(jvms, idx);
2087 Node* box_node = sfn->monitor_box(jvms, idx);
2088 if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
2089 int j = jvms->monitor_box_offset(idx);
2090 _igvn.replace_input_of(u, j, newbox);
2091 next_edge = false;
2092 }
2093 }
2094 }
2095 }
2096 if (next_edge) i++;
2097 }
2098 }
2099
2100 //-----------------------mark_eliminated_locking_nodes-----------------------
2101 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
2102 if (!alock->is_balanced()) {
2103 return; // Can't do any more elimination for this locking region
2104 }
2105 if (EliminateNestedLocks) {
2106 if (alock->is_nested()) {
2107 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
2108 return;
2109 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
2110 // Only Lock node has JVMState needed here.
2111 // Not that preceding claim is documented anywhere else.
2112 if (alock->jvms() != nullptr) {
2113 if (alock->as_Lock()->is_nested_lock_region()) {
2114 // Mark eliminated related nested locks and unlocks.
2115 Node* obj = alock->obj_node();
2116 BoxLockNode* box_node = alock->box_node()->as_BoxLock();
2117 assert(!box_node->is_eliminated(), "should not be marked yet");
2118 // Note: BoxLock node is marked eliminated only here
2119 // and it is used to indicate that all associated lock
2120 // and unlock nodes are marked for elimination.
2121 box_node->set_eliminated(); // Box's hash is always NO_HASH here
2122 for (uint i = 0; i < box_node->outcnt(); i++) {
2123 Node* u = box_node->raw_out(i);
2124 if (u->is_AbstractLock()) {
2125 alock = u->as_AbstractLock();
2126 if (alock->box_node() == box_node) {
2127 // Verify that this Box is referenced only by related locks.
2128 assert(alock->obj_node()->eqv_uncast(obj), "");
2129 // Mark all related locks and unlocks.
2130 #ifdef ASSERT
2131 alock->log_lock_optimization(C, "eliminate_lock_set_nested");
2132 #endif
2133 alock->set_nested();
2134 }
2135 }
2136 }
2137 } else {
2138 #ifdef ASSERT
2139 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region");
2140 if (C->log() != nullptr)
2141 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output
2142 #endif
2143 }
2144 }
2145 return;
2146 }
2147 // Process locks for non escaping object
2148 assert(alock->is_non_esc_obj(), "");
2149 } // EliminateNestedLocks
2150
2151 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2152 // Look for all locks of this object and mark them and
2153 // corresponding BoxLock nodes as eliminated.
2154 Node* obj = alock->obj_node();
2155 for (uint j = 0; j < obj->outcnt(); j++) {
2156 Node* o = obj->raw_out(j);
2157 if (o->is_AbstractLock() &&
2158 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2159 alock = o->as_AbstractLock();
2160 Node* box = alock->box_node();
2161 // Replace old box node with new eliminated box for all users
2162 // of the same object and mark related locks as eliminated.
2163 mark_eliminated_box(box, obj);
2164 }
2165 }
2166 }
2167 }
2168
2169 // we have determined that this lock/unlock can be eliminated, we simply
2170 // eliminate the node without expanding it.
2171 //
2172 // Note: The membar's associated with the lock/unlock are currently not
2173 // eliminated. This should be investigated as a future enhancement.
2174 //
2175 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2176
2177 if (!alock->is_eliminated()) {
2178 return false;
2179 }
2180 #ifdef ASSERT
2181 if (!alock->is_coarsened()) {
2182 // Check that new "eliminated" BoxLock node is created.
2183 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2184 assert(oldbox->is_eliminated(), "should be done already");
2185 }
2186 #endif
2187
2188 alock->log_lock_optimization(C, "eliminate_lock");
2189
2190 #ifndef PRODUCT
2191 if (PrintEliminateLocks) {
2192 tty->print_cr("++++ Eliminated: %d %s '%s'", alock->_idx, (alock->is_Lock() ? "Lock" : "Unlock"), alock->kind_as_string());
2193 }
2194 #endif
2195
2196 Node* mem = alock->in(TypeFunc::Memory);
2197 Node* ctrl = alock->in(TypeFunc::Control);
2198 guarantee(ctrl != nullptr, "missing control projection, cannot replace_node() with null");
2199
2200 alock->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
2201 // There are 2 projections from the lock. The lock node will
2202 // be deleted when its last use is subsumed below.
2203 assert(alock->outcnt() == 2 &&
2204 _callprojs.fallthrough_proj != nullptr &&
2205 _callprojs.fallthrough_memproj != nullptr,
2206 "Unexpected projections from Lock/Unlock");
2207
2208 Node* fallthroughproj = _callprojs.fallthrough_proj;
2209 Node* memproj_fallthrough = _callprojs.fallthrough_memproj;
2210
2211 // The memory projection from a lock/unlock is RawMem
2212 // The input to a Lock is merged memory, so extract its RawMem input
2213 // (unless the MergeMem has been optimized away.)
2214 if (alock->is_Lock()) {
2215 // Search for MemBarAcquireLock node and delete it also.
2216 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2217 assert(membar != nullptr && membar->Opcode() == Op_MemBarAcquireLock, "");
2218 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2219 Node* memproj = membar->proj_out(TypeFunc::Memory);
2220 _igvn.replace_node(ctrlproj, fallthroughproj);
2221 _igvn.replace_node(memproj, memproj_fallthrough);
2222
2223 // Delete FastLock node also if this Lock node is unique user
2224 // (a loop peeling may clone a Lock node).
2225 Node* flock = alock->as_Lock()->fastlock_node();
2226 if (flock->outcnt() == 1) {
2227 assert(flock->unique_out() == alock, "sanity");
2228 _igvn.replace_node(flock, top());
2229 }
2230 }
2231
2232 // Search for MemBarReleaseLock node and delete it also.
2233 if (alock->is_Unlock() && ctrl->is_Proj() && ctrl->in(0)->is_MemBar()) {
2234 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2235 assert(membar->Opcode() == Op_MemBarReleaseLock &&
2236 mem->is_Proj() && membar == mem->in(0), "");
2237 _igvn.replace_node(fallthroughproj, ctrl);
2238 _igvn.replace_node(memproj_fallthrough, mem);
2239 fallthroughproj = ctrl;
2240 memproj_fallthrough = mem;
2241 ctrl = membar->in(TypeFunc::Control);
2242 mem = membar->in(TypeFunc::Memory);
2243 }
2244
2245 _igvn.replace_node(fallthroughproj, ctrl);
2246 _igvn.replace_node(memproj_fallthrough, mem);
2247 return true;
2248 }
2249
2250
2251 //------------------------------expand_lock_node----------------------
2252 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2253
2254 Node* ctrl = lock->in(TypeFunc::Control);
2255 Node* mem = lock->in(TypeFunc::Memory);
2256 Node* obj = lock->obj_node();
2257 Node* box = lock->box_node();
2258 Node* flock = lock->fastlock_node();
2259
2260 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2261
2262 // Make the merge point
2263 Node *region;
2264 Node *mem_phi;
2265 Node *slow_path;
2266
2267 region = new RegionNode(3);
2268 // create a Phi for the memory state
2269 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2270
2271 // Optimize test; set region slot 2
2272 slow_path = opt_bits_test(ctrl, region, 2, flock);
2273 mem_phi->init_req(2, mem);
2274
2275 // Make slow path call
2276 CallNode* call = make_slow_call(lock, OptoRuntime::complete_monitor_enter_Type(),
2277 OptoRuntime::complete_monitor_locking_Java(), nullptr, slow_path,
2278 obj, box, nullptr);
2279
2280 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
2281
2282 // Slow path can only throw asynchronous exceptions, which are always
2283 // de-opted. So the compiler thinks the slow-call can never throw an
2284 // exception. If it DOES throw an exception we would need the debug
2285 // info removed first (since if it throws there is no monitor).
2286 assert(_callprojs.fallthrough_ioproj == nullptr && _callprojs.catchall_ioproj == nullptr &&
2287 _callprojs.catchall_memproj == nullptr && _callprojs.catchall_catchproj == nullptr, "Unexpected projection from Lock");
2288
2289 // Capture slow path
2290 // disconnect fall-through projection from call and create a new one
2291 // hook up users of fall-through projection to region
2292 Node *slow_ctrl = _callprojs.fallthrough_proj->clone();
2293 transform_later(slow_ctrl);
2294 _igvn.hash_delete(_callprojs.fallthrough_proj);
2295 _callprojs.fallthrough_proj->disconnect_inputs(C);
2296 region->init_req(1, slow_ctrl);
2297 // region inputs are now complete
2298 transform_later(region);
2299 _igvn.replace_node(_callprojs.fallthrough_proj, region);
2300
2301 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2302
2303 mem_phi->init_req(1, memproj);
2304
2305 transform_later(mem_phi);
2306
2307 _igvn.replace_node(_callprojs.fallthrough_memproj, mem_phi);
2308 }
2309
2310 //------------------------------expand_unlock_node----------------------
2311 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2312
2313 Node* ctrl = unlock->in(TypeFunc::Control);
2314 Node* mem = unlock->in(TypeFunc::Memory);
2315 Node* obj = unlock->obj_node();
2316 Node* box = unlock->box_node();
2317
2318 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2319
2320 // No need for a null check on unlock
2321
2322 // Make the merge point
2323 Node *region;
2324 Node *mem_phi;
2325
2326 region = new RegionNode(3);
2327 // create a Phi for the memory state
2328 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2329
2330 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2331 funlock = transform_later( funlock )->as_FastUnlock();
2332 // Optimize test; set region slot 2
2333 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock);
2334 Node *thread = transform_later(new ThreadLocalNode());
2335
2336 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2337 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2338 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2339
2340 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
2341 assert(_callprojs.fallthrough_ioproj == nullptr && _callprojs.catchall_ioproj == nullptr &&
2342 _callprojs.catchall_memproj == nullptr && _callprojs.catchall_catchproj == nullptr, "Unexpected projection from Lock");
2343
2344 // No exceptions for unlocking
2345 // Capture slow path
2346 // disconnect fall-through projection from call and create a new one
2347 // hook up users of fall-through projection to region
2348 Node *slow_ctrl = _callprojs.fallthrough_proj->clone();
2349 transform_later(slow_ctrl);
2350 _igvn.hash_delete(_callprojs.fallthrough_proj);
2351 _callprojs.fallthrough_proj->disconnect_inputs(C);
2352 region->init_req(1, slow_ctrl);
2353 // region inputs are now complete
2354 transform_later(region);
2355 _igvn.replace_node(_callprojs.fallthrough_proj, region);
2356
2357 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2358 mem_phi->init_req(1, memproj );
2359 mem_phi->init_req(2, mem);
2360 transform_later(mem_phi);
2361
2362 _igvn.replace_node(_callprojs.fallthrough_memproj, mem_phi);
2363 }
2364
2365 void PhaseMacroExpand::expand_subtypecheck_node(SubTypeCheckNode *check) {
2366 assert(check->in(SubTypeCheckNode::Control) == nullptr, "should be pinned");
2367 Node* bol = check->unique_out();
2368 Node* obj_or_subklass = check->in(SubTypeCheckNode::ObjOrSubKlass);
2369 Node* superklass = check->in(SubTypeCheckNode::SuperKlass);
2370 assert(bol->is_Bool() && bol->as_Bool()->_test._test == BoolTest::ne, "unexpected bool node");
2371
2372 for (DUIterator_Last imin, i = bol->last_outs(imin); i >= imin; --i) {
2373 Node* iff = bol->last_out(i);
2374 assert(iff->is_If(), "where's the if?");
2375
2376 if (iff->in(0)->is_top()) {
2377 _igvn.replace_input_of(iff, 1, C->top());
2378 continue;
2379 }
2380
2381 Node* iftrue = iff->as_If()->proj_out(1);
2382 Node* iffalse = iff->as_If()->proj_out(0);
2383 Node* ctrl = iff->in(0);
2384
2385 Node* subklass = nullptr;
2386 if (_igvn.type(obj_or_subklass)->isa_klassptr()) {
2387 subklass = obj_or_subklass;
2388 } else {
2389 Node* k_adr = basic_plus_adr(obj_or_subklass, oopDesc::klass_offset_in_bytes());
2390 subklass = _igvn.transform(LoadKlassNode::make(_igvn, C->immutable_memory(), k_adr, TypeInstPtr::KLASS));
2391 }
2392
2393 Node* not_subtype_ctrl = Phase::gen_subtype_check(subklass, superklass, &ctrl, nullptr, _igvn, check->method(), check->bci());
2394
2395 _igvn.replace_input_of(iff, 0, C->top());
2396 _igvn.replace_node(iftrue, not_subtype_ctrl);
2397 _igvn.replace_node(iffalse, ctrl);
2398 }
2399 _igvn.replace_node(check, C->top());
2400 }
2401
2402 // Perform refining of strip mined loop nodes in the macro nodes list.
2403 void PhaseMacroExpand::refine_strip_mined_loop_macro_nodes() {
2404 for (int i = C->macro_count(); i > 0; i--) {
2405 Node* n = C->macro_node(i - 1);
2406 if (n->is_OuterStripMinedLoop()) {
2407 n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn);
2408 }
2409 }
2410 }
2411
2412 //---------------------------eliminate_macro_nodes----------------------
2413 // Eliminate scalar replaced allocations and associated locks.
2414 void PhaseMacroExpand::eliminate_macro_nodes() {
2415 if (C->macro_count() == 0)
2416 return;
2417
2418 if (StressMacroElimination) {
2419 C->shuffle_macro_nodes();
2420 }
2421 NOT_PRODUCT(int membar_before = count_MemBar(C);)
2422
2423 // Before elimination may re-mark (change to Nested or NonEscObj)
2424 // all associated (same box and obj) lock and unlock nodes.
2425 int cnt = C->macro_count();
2426 for (int i=0; i < cnt; i++) {
2427 Node *n = C->macro_node(i);
2428 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2429 mark_eliminated_locking_nodes(n->as_AbstractLock());
2430 }
2431 }
2432 // Re-marking may break consistency of Coarsened locks.
2433 if (!C->coarsened_locks_consistent()) {
2434 return; // recompile without Coarsened locks if broken
2435 } else {
2436 // After coarsened locks are eliminated locking regions
2437 // become unbalanced. We should not execute any more
2438 // locks elimination optimizations on them.
2439 C->mark_unbalanced_boxes();
2440 }
2441
2442 // First, attempt to eliminate locks
2443 bool progress = true;
2444 while (progress) {
2445 progress = false;
2446 for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once
2447 Node* n = C->macro_node(i - 1);
2448 bool success = false;
2449 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2450 if (n->is_AbstractLock()) {
2451 success = eliminate_locking_node(n->as_AbstractLock());
2452 #ifndef PRODUCT
2453 if (success && PrintOptoStatistics) {
2454 AtomicAccess::inc(&PhaseMacroExpand::_monitor_objects_removed_counter);
2455 }
2456 #endif
2457 }
2458 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2459 progress = progress || success;
2460 if (success) {
2461 C->print_method(PHASE_AFTER_MACRO_ELIMINATION_STEP, 5, n);
2462 }
2463 }
2464 }
2465 // Next, attempt to eliminate allocations
2466 progress = true;
2467 while (progress) {
2468 progress = false;
2469 for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once
2470 Node* n = C->macro_node(i - 1);
2471 bool success = false;
2472 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2473 switch (n->class_id()) {
2474 case Node::Class_Allocate:
2475 case Node::Class_AllocateArray:
2476 success = eliminate_allocate_node(n->as_Allocate());
2477 #ifndef PRODUCT
2478 if (success && PrintOptoStatistics) {
2479 AtomicAccess::inc(&PhaseMacroExpand::_objs_scalar_replaced_counter);
2480 }
2481 #endif
2482 break;
2483 case Node::Class_CallStaticJava:
2484 success = eliminate_boxing_node(n->as_CallStaticJava());
2485 break;
2486 case Node::Class_Lock:
2487 case Node::Class_Unlock:
2488 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2489 break;
2490 case Node::Class_ArrayCopy:
2491 break;
2492 case Node::Class_OuterStripMinedLoop:
2493 break;
2494 case Node::Class_SubTypeCheck:
2495 break;
2496 case Node::Class_Opaque1:
2497 break;
2498 default:
2499 assert(n->Opcode() == Op_LoopLimit ||
2500 n->Opcode() == Op_ModD ||
2501 n->Opcode() == Op_ModF ||
2502 n->is_OpaqueNotNull() ||
2503 n->is_OpaqueInitializedAssertionPredicate() ||
2504 n->Opcode() == Op_MaxL ||
2505 n->Opcode() == Op_MinL ||
2506 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n),
2507 "unknown node type in macro list");
2508 }
2509 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2510 progress = progress || success;
2511 if (success) {
2512 C->print_method(PHASE_AFTER_MACRO_ELIMINATION_STEP, 5, n);
2513 }
2514 }
2515 }
2516 #ifndef PRODUCT
2517 if (PrintOptoStatistics) {
2518 int membar_after = count_MemBar(C);
2519 AtomicAccess::add(&PhaseMacroExpand::_memory_barriers_removed_counter, membar_before - membar_after);
2520 }
2521 #endif
2522 }
2523
2524 void PhaseMacroExpand::eliminate_opaque_looplimit_macro_nodes() {
2525 if (C->macro_count() == 0) {
2526 return;
2527 }
2528 refine_strip_mined_loop_macro_nodes();
2529 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2530 bool progress = true;
2531 while (progress) {
2532 progress = false;
2533 for (int i = C->macro_count(); i > 0; i--) {
2534 Node* n = C->macro_node(i-1);
2535 bool success = false;
2536 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2537 if (n->Opcode() == Op_LoopLimit) {
2538 // Remove it from macro list and put on IGVN worklist to optimize.
2539 C->remove_macro_node(n);
2540 _igvn._worklist.push(n);
2541 success = true;
2542 } else if (n->Opcode() == Op_CallStaticJava) {
2543 // Remove it from macro list and put on IGVN worklist to optimize.
2544 C->remove_macro_node(n);
2545 _igvn._worklist.push(n);
2546 success = true;
2547 } else if (n->is_Opaque1()) {
2548 _igvn.replace_node(n, n->in(1));
2549 success = true;
2550 } else if (n->is_OpaqueNotNull()) {
2551 // Tests with OpaqueNotNull nodes are implicitly known to be true. Replace the node with true. In debug builds,
2552 // we leave the test in the graph to have an additional sanity check at runtime. If the test fails (i.e. a bug),
2553 // we will execute a Halt node.
2554 #ifdef ASSERT
2555 _igvn.replace_node(n, n->in(1));
2556 #else
2557 _igvn.replace_node(n, _igvn.intcon(1));
2558 #endif
2559 success = true;
2560 } else if (n->is_OpaqueInitializedAssertionPredicate()) {
2561 // Initialized Assertion Predicates must always evaluate to true. Therefore, we get rid of them in product
2562 // builds as they are useless. In debug builds we keep them as additional verification code. Even though
2563 // loop opts are already over, we want to keep Initialized Assertion Predicates alive as long as possible to
2564 // enable folding of dead control paths within which cast nodes become top after due to impossible types -
2565 // even after loop opts are over. Therefore, we delay the removal of these opaque nodes until now.
2566 #ifdef ASSERT
2567 _igvn.replace_node(n, n->in(1));
2568 #else
2569 _igvn.replace_node(n, _igvn.intcon(1));
2570 #endif // ASSERT
2571 } else if (n->Opcode() == Op_OuterStripMinedLoop) {
2572 C->remove_macro_node(n);
2573 success = true;
2574 } else if (n->Opcode() == Op_MaxL) {
2575 // Since MaxL and MinL are not implemented in the backend, we expand them to
2576 // a CMoveL construct now. At least until here, the type could be computed
2577 // precisely. CMoveL is not so smart, but we can give it at least the best
2578 // type we know abouot n now.
2579 Node* repl = MaxNode::signed_max(n->in(1), n->in(2), _igvn.type(n), _igvn);
2580 _igvn.replace_node(n, repl);
2581 success = true;
2582 } else if (n->Opcode() == Op_MinL) {
2583 Node* repl = MaxNode::signed_min(n->in(1), n->in(2), _igvn.type(n), _igvn);
2584 _igvn.replace_node(n, repl);
2585 success = true;
2586 }
2587 assert(!success || (C->macro_count() == (old_macro_count - 1)), "elimination must have deleted one node from macro list");
2588 progress = progress || success;
2589 if (success) {
2590 C->print_method(PHASE_AFTER_MACRO_ELIMINATION_STEP, 5, n);
2591 }
2592 }
2593 }
2594 }
2595
2596 //------------------------------expand_macro_nodes----------------------
2597 // Returns true if a failure occurred.
2598 bool PhaseMacroExpand::expand_macro_nodes() {
2599 if (StressMacroExpansion) {
2600 C->shuffle_macro_nodes();
2601 }
2602
2603 // Clean up the graph so we're less likely to hit the maximum node
2604 // limit
2605 _igvn.set_delay_transform(false);
2606 _igvn.optimize();
2607 if (C->failing()) return true;
2608 _igvn.set_delay_transform(true);
2609
2610
2611 // Because we run IGVN after each expansion, some macro nodes may go
2612 // dead and be removed from the list as we iterate over it. Move
2613 // Allocate nodes (processed in a second pass) at the beginning of
2614 // the list and then iterate from the last element of the list until
2615 // an Allocate node is seen. This is robust to random deletion in
2616 // the list due to nodes going dead.
2617 C->sort_macro_nodes();
2618
2619 // expand arraycopy "macro" nodes first
2620 // For ReduceBulkZeroing, we must first process all arraycopy nodes
2621 // before the allocate nodes are expanded.
2622 while (C->macro_count() > 0) {
2623 int macro_count = C->macro_count();
2624 Node * n = C->macro_node(macro_count-1);
2625 assert(n->is_macro(), "only macro nodes expected here");
2626 if (_igvn.type(n) == Type::TOP || (n->in(0) != nullptr && n->in(0)->is_top())) {
2627 // node is unreachable, so don't try to expand it
2628 C->remove_macro_node(n);
2629 continue;
2630 }
2631 if (n->is_Allocate()) {
2632 break;
2633 }
2634 // Make sure expansion will not cause node limit to be exceeded.
2635 // Worst case is a macro node gets expanded into about 200 nodes.
2636 // Allow 50% more for optimization.
2637 if (C->check_node_count(300, "out of nodes before macro expansion")) {
2638 return true;
2639 }
2640
2641 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2642 switch (n->class_id()) {
2643 case Node::Class_Lock:
2644 expand_lock_node(n->as_Lock());
2645 break;
2646 case Node::Class_Unlock:
2647 expand_unlock_node(n->as_Unlock());
2648 break;
2649 case Node::Class_ArrayCopy:
2650 expand_arraycopy_node(n->as_ArrayCopy());
2651 break;
2652 case Node::Class_SubTypeCheck:
2653 expand_subtypecheck_node(n->as_SubTypeCheck());
2654 break;
2655 default:
2656 switch (n->Opcode()) {
2657 case Op_ModD:
2658 case Op_ModF: {
2659 CallNode* mod_macro = n->as_Call();
2660 CallNode* call = new CallLeafPureNode(mod_macro->tf(), mod_macro->entry_point(),
2661 mod_macro->_name, TypeRawPtr::BOTTOM);
2662 call->init_req(TypeFunc::Control, mod_macro->in(TypeFunc::Control));
2663 call->init_req(TypeFunc::I_O, C->top());
2664 call->init_req(TypeFunc::Memory, C->top());
2665 call->init_req(TypeFunc::ReturnAdr, C->top());
2666 call->init_req(TypeFunc::FramePtr, C->top());
2667 for (unsigned int i = 0; i < mod_macro->tf()->domain()->cnt() - TypeFunc::Parms; i++) {
2668 call->init_req(TypeFunc::Parms + i, mod_macro->in(TypeFunc::Parms + i));
2669 }
2670 _igvn.replace_node(mod_macro, call);
2671 transform_later(call);
2672 break;
2673 }
2674 default:
2675 assert(false, "unknown node type in macro list");
2676 }
2677 }
2678 assert(C->macro_count() == (old_macro_count - 1), "expansion must have deleted one node from macro list");
2679 if (C->failing()) return true;
2680 C->print_method(PHASE_AFTER_MACRO_EXPANSION_STEP, 5, n);
2681
2682 // Clean up the graph so we're less likely to hit the maximum node
2683 // limit
2684 _igvn.set_delay_transform(false);
2685 _igvn.optimize();
2686 if (C->failing()) return true;
2687 _igvn.set_delay_transform(true);
2688 }
2689
2690 // All nodes except Allocate nodes are expanded now. There could be
2691 // new optimization opportunities (such as folding newly created
2692 // load from a just allocated object). Run IGVN.
2693
2694 // expand "macro" nodes
2695 // nodes are removed from the macro list as they are processed
2696 while (C->macro_count() > 0) {
2697 int macro_count = C->macro_count();
2698 Node * n = C->macro_node(macro_count-1);
2699 assert(n->is_macro(), "only macro nodes expected here");
2700 if (_igvn.type(n) == Type::TOP || (n->in(0) != nullptr && n->in(0)->is_top())) {
2701 // node is unreachable, so don't try to expand it
2702 C->remove_macro_node(n);
2703 continue;
2704 }
2705 // Make sure expansion will not cause node limit to be exceeded.
2706 // Worst case is a macro node gets expanded into about 200 nodes.
2707 // Allow 50% more for optimization.
2708 if (C->check_node_count(300, "out of nodes before macro expansion")) {
2709 return true;
2710 }
2711 switch (n->class_id()) {
2712 case Node::Class_Allocate:
2713 expand_allocate(n->as_Allocate());
2714 break;
2715 case Node::Class_AllocateArray:
2716 expand_allocate_array(n->as_AllocateArray());
2717 break;
2718 default:
2719 assert(false, "unknown node type in macro list");
2720 }
2721 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2722 if (C->failing()) return true;
2723 C->print_method(PHASE_AFTER_MACRO_EXPANSION_STEP, 5, n);
2724
2725 // Clean up the graph so we're less likely to hit the maximum node
2726 // limit
2727 _igvn.set_delay_transform(false);
2728 _igvn.optimize();
2729 if (C->failing()) return true;
2730 _igvn.set_delay_transform(true);
2731 }
2732
2733 _igvn.set_delay_transform(false);
2734 return false;
2735 }
2736
2737 #ifndef PRODUCT
2738 int PhaseMacroExpand::_objs_scalar_replaced_counter = 0;
2739 int PhaseMacroExpand::_monitor_objects_removed_counter = 0;
2740 int PhaseMacroExpand::_GC_barriers_removed_counter = 0;
2741 int PhaseMacroExpand::_memory_barriers_removed_counter = 0;
2742
2743 void PhaseMacroExpand::print_statistics() {
2744 tty->print("Objects scalar replaced = %d, ", AtomicAccess::load(&_objs_scalar_replaced_counter));
2745 tty->print("Monitor objects removed = %d, ", AtomicAccess::load(&_monitor_objects_removed_counter));
2746 tty->print("GC barriers removed = %d, ", AtomicAccess::load(&_GC_barriers_removed_counter));
2747 tty->print_cr("Memory barriers removed = %d", AtomicAccess::load(&_memory_barriers_removed_counter));
2748 }
2749
2750 int PhaseMacroExpand::count_MemBar(Compile *C) {
2751 if (!PrintOptoStatistics) {
2752 return 0;
2753 }
2754 Unique_Node_List ideal_nodes;
2755 int total = 0;
2756 ideal_nodes.map(C->live_nodes(), nullptr);
2757 ideal_nodes.push(C->root());
2758 for (uint next = 0; next < ideal_nodes.size(); ++next) {
2759 Node* n = ideal_nodes.at(next);
2760 if (n->is_MemBar()) {
2761 total++;
2762 }
2763 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2764 Node* m = n->fast_out(i);
2765 ideal_nodes.push(m);
2766 }
2767 }
2768 return total;
2769 }
2770 #endif