1 /*
2 * Copyright (c) 2005, 2023, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "ci/ciFlatArrayKlass.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "gc/shared/collectedHeap.inline.hpp"
29 #include "gc/shared/tlab_globals.hpp"
30 #include "libadt/vectset.hpp"
31 #include "memory/universe.hpp"
32 #include "opto/addnode.hpp"
33 #include "opto/arraycopynode.hpp"
34 #include "opto/callnode.hpp"
35 #include "opto/castnode.hpp"
36 #include "opto/cfgnode.hpp"
37 #include "opto/compile.hpp"
38 #include "opto/convertnode.hpp"
39 #include "opto/graphKit.hpp"
40 #include "opto/inlinetypenode.hpp"
41 #include "opto/intrinsicnode.hpp"
42 #include "opto/locknode.hpp"
43 #include "opto/loopnode.hpp"
44 #include "opto/macro.hpp"
45 #include "opto/memnode.hpp"
46 #include "opto/narrowptrnode.hpp"
47 #include "opto/node.hpp"
48 #include "opto/opaquenode.hpp"
49 #include "opto/phaseX.hpp"
50 #include "opto/rootnode.hpp"
51 #include "opto/runtime.hpp"
52 #include "opto/subnode.hpp"
53 #include "opto/subtypenode.hpp"
54 #include "opto/type.hpp"
55 #include "prims/jvmtiExport.hpp"
56 #include "runtime/continuation.hpp"
57 #include "runtime/sharedRuntime.hpp"
58 #include "runtime/stubRoutines.hpp"
59 #include "utilities/macros.hpp"
60 #include "utilities/powerOfTwo.hpp"
61 #if INCLUDE_G1GC
62 #include "gc/g1/g1ThreadLocalData.hpp"
63 #endif // INCLUDE_G1GC
64
65
66 //
67 // Replace any references to "oldref" in inputs to "use" with "newref".
68 // Returns the number of replacements made.
69 //
70 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
71 int nreplacements = 0;
72 uint req = use->req();
73 for (uint j = 0; j < use->len(); j++) {
74 Node *uin = use->in(j);
75 if (uin == oldref) {
76 if (j < req)
77 use->set_req(j, newref);
78 else
79 use->set_prec(j, newref);
80 nreplacements++;
81 } else if (j >= req && uin == nullptr) {
82 break;
83 }
84 }
85 return nreplacements;
86 }
87
88 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
89 Node* cmp;
90 if (mask != 0) {
91 Node* and_node = transform_later(new AndXNode(word, MakeConX(mask)));
92 cmp = transform_later(new CmpXNode(and_node, MakeConX(bits)));
93 } else {
94 cmp = word;
95 }
96 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
97 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
98 transform_later(iff);
99
100 // Fast path taken.
101 Node *fast_taken = transform_later(new IfFalseNode(iff));
102
103 // Fast path not-taken, i.e. slow path
104 Node *slow_taken = transform_later(new IfTrueNode(iff));
105
106 if (return_fast_path) {
107 region->init_req(edge, slow_taken); // Capture slow-control
108 return fast_taken;
109 } else {
110 region->init_req(edge, fast_taken); // Capture fast-control
111 return slow_taken;
112 }
113 }
114
115 //--------------------copy_predefined_input_for_runtime_call--------------------
116 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
117 // Set fixed predefined input arguments
118 call->init_req( TypeFunc::Control, ctrl );
119 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
120 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
121 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
122 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
123 }
124
125 //------------------------------make_slow_call---------------------------------
126 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type,
127 address slow_call, const char* leaf_name, Node* slow_path,
128 Node* parm0, Node* parm1, Node* parm2) {
129
130 // Slow-path call
131 CallNode *call = leaf_name
132 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
133 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), TypeRawPtr::BOTTOM );
134
135 // Slow path call has no side-effects, uses few values
136 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
137 if (parm0 != nullptr) call->init_req(TypeFunc::Parms+0, parm0);
138 if (parm1 != nullptr) call->init_req(TypeFunc::Parms+1, parm1);
139 if (parm2 != nullptr) call->init_req(TypeFunc::Parms+2, parm2);
140 call->copy_call_debug_info(&_igvn, oldcall);
141 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
142 _igvn.replace_node(oldcall, call);
143 transform_later(call);
144
145 return call;
146 }
147
148 void PhaseMacroExpand::eliminate_gc_barrier(Node* p2x) {
149 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
150 bs->eliminate_gc_barrier(&_igvn, p2x);
151 #ifndef PRODUCT
152 if (PrintOptoStatistics) {
153 Atomic::inc(&PhaseMacroExpand::_GC_barriers_removed_counter);
154 }
155 #endif
156 }
157
158 // Search for a memory operation for the specified memory slice.
159 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
160 Node *orig_mem = mem;
161 Node *alloc_mem = alloc->in(TypeFunc::Memory);
162 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
163 while (true) {
164 if (mem == alloc_mem || mem == start_mem ) {
165 return mem; // hit one of our sentinels
166 } else if (mem->is_MergeMem()) {
167 mem = mem->as_MergeMem()->memory_at(alias_idx);
168 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
169 Node *in = mem->in(0);
170 // we can safely skip over safepoints, calls, locks and membars because we
171 // already know that the object is safe to eliminate.
172 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
173 return in;
174 } else if (in->is_Call()) {
175 CallNode *call = in->as_Call();
176 if (call->may_modify(tinst, phase)) {
177 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape");
178 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) {
179 return in;
180 }
181 }
182 mem = in->in(TypeFunc::Memory);
183 } else if (in->is_MemBar()) {
184 ArrayCopyNode* ac = nullptr;
185 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) {
186 if (ac != nullptr) {
187 assert(ac->is_clonebasic(), "Only basic clone is a non escaping clone");
188 return ac;
189 }
190 }
191 mem = in->in(TypeFunc::Memory);
192 } else {
193 #ifdef ASSERT
194 in->dump();
195 mem->dump();
196 assert(false, "unexpected projection");
197 #endif
198 }
199 } else if (mem->is_Store()) {
200 const TypePtr* atype = mem->as_Store()->adr_type();
201 int adr_idx = phase->C->get_alias_index(atype);
202 if (adr_idx == alias_idx) {
203 assert(atype->isa_oopptr(), "address type must be oopptr");
204 int adr_offset = atype->flat_offset();
205 uint adr_iid = atype->is_oopptr()->instance_id();
206 // Array elements references have the same alias_idx
207 // but different offset and different instance_id.
208 if (adr_offset == offset && adr_iid == alloc->_idx) {
209 return mem;
210 }
211 } else {
212 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
213 }
214 mem = mem->in(MemNode::Memory);
215 } else if (mem->is_ClearArray()) {
216 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
217 // Can not bypass initialization of the instance
218 // we are looking.
219 debug_only(intptr_t offset;)
220 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
221 InitializeNode* init = alloc->as_Allocate()->initialization();
222 // We are looking for stored value, return Initialize node
223 // or memory edge from Allocate node.
224 if (init != nullptr) {
225 return init;
226 } else {
227 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
228 }
229 }
230 // Otherwise skip it (the call updated 'mem' value).
231 } else if (mem->Opcode() == Op_SCMemProj) {
232 mem = mem->in(0);
233 Node* adr = nullptr;
234 if (mem->is_LoadStore()) {
235 adr = mem->in(MemNode::Address);
236 } else {
237 assert(mem->Opcode() == Op_EncodeISOArray ||
238 mem->Opcode() == Op_StrCompressedCopy, "sanity");
239 adr = mem->in(3); // Destination array
240 }
241 const TypePtr* atype = adr->bottom_type()->is_ptr();
242 int adr_idx = phase->C->get_alias_index(atype);
243 if (adr_idx == alias_idx) {
244 DEBUG_ONLY(mem->dump();)
245 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
246 return nullptr;
247 }
248 mem = mem->in(MemNode::Memory);
249 } else if (mem->Opcode() == Op_StrInflatedCopy) {
250 Node* adr = mem->in(3); // Destination array
251 const TypePtr* atype = adr->bottom_type()->is_ptr();
252 int adr_idx = phase->C->get_alias_index(atype);
253 if (adr_idx == alias_idx) {
254 DEBUG_ONLY(mem->dump();)
255 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
256 return nullptr;
257 }
258 mem = mem->in(MemNode::Memory);
259 } else {
260 return mem;
261 }
262 assert(mem != orig_mem, "dead memory loop");
263 }
264 }
265
266 // Generate loads from source of the arraycopy for fields of
267 // destination needed at a deoptimization point
268 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) {
269 BasicType bt = ft;
270 const Type *type = ftype;
271 if (ft == T_NARROWOOP) {
272 bt = T_OBJECT;
273 type = ftype->make_oopptr();
274 }
275 Node* res = nullptr;
276 if (ac->is_clonebasic()) {
277 assert(ac->in(ArrayCopyNode::Src) != ac->in(ArrayCopyNode::Dest), "clone source equals destination");
278 Node* base = ac->in(ArrayCopyNode::Src);
279 Node* adr = _igvn.transform(new AddPNode(base, base, _igvn.MakeConX(offset)));
280 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
281 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
282 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
283 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
284 } else {
285 if (ac->modifies(offset, offset, &_igvn, true)) {
286 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
287 uint shift = exact_log2(type2aelembytes(bt));
288 Node* src_pos = ac->in(ArrayCopyNode::SrcPos);
289 Node* dest_pos = ac->in(ArrayCopyNode::DestPos);
290 const TypeInt* src_pos_t = _igvn.type(src_pos)->is_int();
291 const TypeInt* dest_pos_t = _igvn.type(dest_pos)->is_int();
292
293 Node* adr = nullptr;
294 Node* base = ac->in(ArrayCopyNode::Src);
295 const TypeAryPtr* adr_type = _igvn.type(base)->is_aryptr();
296 if (adr_type->is_flat()) {
297 shift = adr_type->flat_log_elem_size();
298 }
299 if (src_pos_t->is_con() && dest_pos_t->is_con()) {
300 intptr_t off = ((src_pos_t->get_con() - dest_pos_t->get_con()) << shift) + offset;
301 adr = _igvn.transform(new AddPNode(base, base, _igvn.MakeConX(off)));
302 adr_type = _igvn.type(adr)->is_aryptr();
303 assert(adr_type == _igvn.type(base)->is_aryptr()->add_field_offset_and_offset(off), "incorrect address type");
304 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
305 // Don't emit a new load from src if src == dst but try to get the value from memory instead
306 return value_from_mem(ac->in(TypeFunc::Memory), ctl, ft, ftype, adr_type, alloc);
307 }
308 } else {
309 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
310 // Non constant offset in the array: we can't statically
311 // determine the value
312 return nullptr;
313 }
314 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
315 #ifdef _LP64
316 diff = _igvn.transform(new ConvI2LNode(diff));
317 #endif
318 diff = _igvn.transform(new LShiftXNode(diff, _igvn.intcon(shift)));
319
320 Node* off = _igvn.transform(new AddXNode(_igvn.MakeConX(offset), diff));
321 adr = _igvn.transform(new AddPNode(base, base, off));
322 // In the case of a flat inline type array, each field has its
323 // own slice so we need to extract the field being accessed from
324 // the address computation
325 adr_type = adr_type->add_field_offset_and_offset(offset)->add_offset(Type::OffsetBot)->is_aryptr();
326 adr = _igvn.transform(new CastPPNode(adr, adr_type));
327 }
328 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
329 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
330 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
331 }
332 }
333 if (res != nullptr) {
334 if (ftype->isa_narrowoop()) {
335 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
336 assert(res->isa_DecodeN(), "should be narrow oop");
337 res = _igvn.transform(new EncodePNode(res, ftype));
338 }
339 return res;
340 }
341 return nullptr;
342 }
343
344 //
345 // Given a Memory Phi, compute a value Phi containing the values from stores
346 // on the input paths.
347 // Note: this function is recursive, its depth is limited by the "level" argument
348 // Returns the computed Phi, or null if it cannot compute it.
349 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) {
350 assert(mem->is_Phi(), "sanity");
351 int alias_idx = C->get_alias_index(adr_t);
352 int offset = adr_t->flat_offset();
353 int instance_id = adr_t->instance_id();
354
355 // Check if an appropriate value phi already exists.
356 Node* region = mem->in(0);
357 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
358 Node* phi = region->fast_out(k);
359 if (phi->is_Phi() && phi != mem &&
360 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
361 return phi;
362 }
363 }
364 // Check if an appropriate new value phi already exists.
365 Node* new_phi = value_phis->find(mem->_idx);
366 if (new_phi != nullptr)
367 return new_phi;
368
369 if (level <= 0) {
370 return nullptr; // Give up: phi tree too deep
371 }
372 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
373 Node *alloc_mem = alloc->in(TypeFunc::Memory);
374
375 uint length = mem->req();
376 GrowableArray <Node *> values(length, length, nullptr);
377
378 // create a new Phi for the value
379 PhiNode *phi = new PhiNode(mem->in(0), phi_type, nullptr, mem->_idx, instance_id, alias_idx, offset);
380 transform_later(phi);
381 value_phis->push(phi, mem->_idx);
382
383 for (uint j = 1; j < length; j++) {
384 Node *in = mem->in(j);
385 if (in == nullptr || in->is_top()) {
386 values.at_put(j, in);
387 } else {
388 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
389 if (val == start_mem || val == alloc_mem) {
390 // hit a sentinel, return appropriate 0 value
391 Node* default_value = alloc->in(AllocateNode::DefaultValue);
392 if (default_value != nullptr) {
393 values.at_put(j, default_value);
394 } else {
395 assert(alloc->in(AllocateNode::RawDefaultValue) == nullptr, "default value may not be null");
396 values.at_put(j, _igvn.zerocon(ft));
397 }
398 continue;
399 }
400 if (val->is_Initialize()) {
401 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
402 }
403 if (val == nullptr) {
404 return nullptr; // can't find a value on this path
405 }
406 if (val == mem) {
407 values.at_put(j, mem);
408 } else if (val->is_Store()) {
409 Node* n = val->in(MemNode::ValueIn);
410 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
411 n = bs->step_over_gc_barrier(n);
412 if (is_subword_type(ft)) {
413 n = Compile::narrow_value(ft, n, phi_type, &_igvn, true);
414 }
415 values.at_put(j, n);
416 } else if(val->is_Proj() && val->in(0) == alloc) {
417 Node* default_value = alloc->in(AllocateNode::DefaultValue);
418 if (default_value != nullptr) {
419 values.at_put(j, default_value);
420 } else {
421 assert(alloc->in(AllocateNode::RawDefaultValue) == nullptr, "default value may not be null");
422 values.at_put(j, _igvn.zerocon(ft));
423 }
424 } else if (val->is_Phi()) {
425 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
426 if (val == nullptr) {
427 return nullptr;
428 }
429 values.at_put(j, val);
430 } else if (val->Opcode() == Op_SCMemProj) {
431 assert(val->in(0)->is_LoadStore() ||
432 val->in(0)->Opcode() == Op_EncodeISOArray ||
433 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
434 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
435 return nullptr;
436 } else if (val->is_ArrayCopy()) {
437 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc);
438 if (res == nullptr) {
439 return nullptr;
440 }
441 values.at_put(j, res);
442 } else {
443 DEBUG_ONLY( val->dump(); )
444 assert(false, "unknown node on this path");
445 return nullptr; // unknown node on this path
446 }
447 }
448 }
449 // Set Phi's inputs
450 for (uint j = 1; j < length; j++) {
451 if (values.at(j) == mem) {
452 phi->init_req(j, phi);
453 } else {
454 phi->init_req(j, values.at(j));
455 }
456 }
457 return phi;
458 }
459
460 // Search the last value stored into the object's field.
461 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
462 assert(adr_t->is_known_instance_field(), "instance required");
463 int instance_id = adr_t->instance_id();
464 assert((uint)instance_id == alloc->_idx, "wrong allocation");
465
466 int alias_idx = C->get_alias_index(adr_t);
467 int offset = adr_t->flat_offset();
468 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
469 Node *alloc_mem = alloc->in(TypeFunc::Memory);
470 VectorSet visited;
471
472 bool done = sfpt_mem == alloc_mem;
473 Node *mem = sfpt_mem;
474 while (!done) {
475 if (visited.test_set(mem->_idx)) {
476 return nullptr; // found a loop, give up
477 }
478 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
479 if (mem == start_mem || mem == alloc_mem) {
480 done = true; // hit a sentinel, return appropriate 0 value
481 } else if (mem->is_Initialize()) {
482 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
483 if (mem == nullptr) {
484 done = true; // Something went wrong.
485 } else if (mem->is_Store()) {
486 const TypePtr* atype = mem->as_Store()->adr_type();
487 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
488 done = true;
489 }
490 } else if (mem->is_Store()) {
491 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
492 assert(atype != nullptr, "address type must be oopptr");
493 assert(C->get_alias_index(atype) == alias_idx &&
494 atype->is_known_instance_field() && atype->flat_offset() == offset &&
495 atype->instance_id() == instance_id, "store is correct memory slice");
496 done = true;
497 } else if (mem->is_Phi()) {
498 // try to find a phi's unique input
499 Node *unique_input = nullptr;
500 Node *top = C->top();
501 for (uint i = 1; i < mem->req(); i++) {
502 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
503 if (n == nullptr || n == top || n == mem) {
504 continue;
505 } else if (unique_input == nullptr) {
506 unique_input = n;
507 } else if (unique_input != n) {
508 unique_input = top;
509 break;
510 }
511 }
512 if (unique_input != nullptr && unique_input != top) {
513 mem = unique_input;
514 } else {
515 done = true;
516 }
517 } else if (mem->is_ArrayCopy()) {
518 done = true;
519 } else {
520 DEBUG_ONLY( mem->dump(); )
521 assert(false, "unexpected node");
522 }
523 }
524 if (mem != nullptr) {
525 if (mem == start_mem || mem == alloc_mem) {
526 // hit a sentinel, return appropriate 0 value
527 Node* default_value = alloc->in(AllocateNode::DefaultValue);
528 if (default_value != nullptr) {
529 return default_value;
530 }
531 assert(alloc->in(AllocateNode::RawDefaultValue) == nullptr, "default value may not be null");
532 return _igvn.zerocon(ft);
533 } else if (mem->is_Store()) {
534 Node* n = mem->in(MemNode::ValueIn);
535 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
536 n = bs->step_over_gc_barrier(n);
537 return n;
538 } else if (mem->is_Phi()) {
539 // attempt to produce a Phi reflecting the values on the input paths of the Phi
540 Node_Stack value_phis(8);
541 Node* phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
542 if (phi != nullptr) {
543 return phi;
544 } else {
545 // Kill all new Phis
546 while(value_phis.is_nonempty()) {
547 Node* n = value_phis.node();
548 _igvn.replace_node(n, C->top());
549 value_phis.pop();
550 }
551 }
552 } else if (mem->is_ArrayCopy()) {
553 Node* ctl = mem->in(0);
554 Node* m = mem->in(TypeFunc::Memory);
555 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
556 // pin the loads in the uncommon trap path
557 ctl = sfpt_ctl;
558 m = sfpt_mem;
559 }
560 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc);
561 }
562 }
563 // Something went wrong.
564 return nullptr;
565 }
566
567 // Search the last value stored into the inline type's fields.
568 Node* PhaseMacroExpand::inline_type_from_mem(Node* mem, Node* ctl, ciInlineKlass* vk, const TypeAryPtr* adr_type, int offset, AllocateNode* alloc) {
569 // Subtract the offset of the first field to account for the missing oop header
570 offset -= vk->first_field_offset();
571 // Create a new InlineTypeNode and retrieve the field values from memory
572 InlineTypeNode* vt = InlineTypeNode::make_uninitialized(_igvn, vk);
573 transform_later(vt);
574 for (int i = 0; i < vk->nof_declared_nonstatic_fields(); ++i) {
575 ciType* field_type = vt->field_type(i);
576 int field_offset = offset + vt->field_offset(i);
577 Node* value = nullptr;
578 if (vt->field_is_flat(i)) {
579 value = inline_type_from_mem(mem, ctl, field_type->as_inline_klass(), adr_type, field_offset, alloc);
580 } else {
581 const Type* ft = Type::get_const_type(field_type);
582 BasicType bt = type2field[field_type->basic_type()];
583 if (UseCompressedOops && !is_java_primitive(bt)) {
584 ft = ft->make_narrowoop();
585 bt = T_NARROWOOP;
586 }
587 // Each inline type field has its own memory slice
588 adr_type = adr_type->with_field_offset(field_offset);
589 value = value_from_mem(mem, ctl, bt, ft, adr_type, alloc);
590 if (value != nullptr && ft->isa_narrowoop()) {
591 assert(UseCompressedOops, "unexpected narrow oop");
592 if (value->is_EncodeP()) {
593 value = value->in(1);
594 } else {
595 value = transform_later(new DecodeNNode(value, value->get_ptr_type()));
596 }
597 }
598 }
599 if (value != nullptr) {
600 vt->set_field_value(i, value);
601 } else {
602 // We might have reached the TrackedInitializationLimit
603 return nullptr;
604 }
605 }
606 return vt;
607 }
608
609 // Check the possibility of scalar replacement.
610 bool PhaseMacroExpand::can_eliminate_allocation(PhaseIterGVN* igvn, AllocateNode *alloc, GrowableArray <SafePointNode *>* safepoints) {
611 // Scan the uses of the allocation to check for anything that would
612 // prevent us from eliminating it.
613 NOT_PRODUCT( const char* fail_eliminate = nullptr; )
614 DEBUG_ONLY( Node* disq_node = nullptr; )
615 bool can_eliminate = true;
616 bool reduce_merge_precheck = (safepoints == nullptr);
617
618 Unique_Node_List worklist;
619 Node* res = alloc->result_cast();
620 const TypeOopPtr* res_type = nullptr;
621 if (res == nullptr) {
622 // All users were eliminated.
623 } else if (!res->is_CheckCastPP()) {
624 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
625 can_eliminate = false;
626 } else {
627 worklist.push(res);
628 res_type = igvn->type(res)->isa_oopptr();
629 if (res_type == nullptr) {
630 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
631 can_eliminate = false;
632 } else if (res_type->isa_aryptr()) {
633 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
634 if (length < 0) {
635 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
636 can_eliminate = false;
637 }
638 }
639 }
640
641 while (can_eliminate && worklist.size() > 0) {
642 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
643 res = worklist.pop();
644 for (DUIterator_Fast jmax, j = res->fast_outs(jmax); j < jmax && can_eliminate; j++) {
645 Node* use = res->fast_out(j);
646
647 if (use->is_AddP()) {
648 const TypePtr* addp_type = igvn->type(use)->is_ptr();
649 int offset = addp_type->offset();
650
651 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
652 NOT_PRODUCT(fail_eliminate = "Undefined field reference";)
653 can_eliminate = false;
654 break;
655 }
656 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
657 k < kmax && can_eliminate; k++) {
658 Node* n = use->fast_out(k);
659 if (!n->is_Store() && n->Opcode() != Op_CastP2X && !bs->is_gc_pre_barrier_node(n)) {
660 DEBUG_ONLY(disq_node = n;)
661 if (n->is_Load() || n->is_LoadStore()) {
662 NOT_PRODUCT(fail_eliminate = "Field load";)
663 } else {
664 NOT_PRODUCT(fail_eliminate = "Not store field reference";)
665 }
666 can_eliminate = false;
667 }
668 }
669 } else if (use->is_ArrayCopy() &&
670 (use->as_ArrayCopy()->is_clonebasic() ||
671 use->as_ArrayCopy()->is_arraycopy_validated() ||
672 use->as_ArrayCopy()->is_copyof_validated() ||
673 use->as_ArrayCopy()->is_copyofrange_validated()) &&
674 use->in(ArrayCopyNode::Dest) == res) {
675 // ok to eliminate
676 } else if (use->is_SafePoint()) {
677 SafePointNode* sfpt = use->as_SafePoint();
678 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
679 // Object is passed as argument.
680 DEBUG_ONLY(disq_node = use;)
681 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
682 can_eliminate = false;
683 }
684 Node* sfptMem = sfpt->memory();
685 if (sfptMem == nullptr || sfptMem->is_top()) {
686 DEBUG_ONLY(disq_node = use;)
687 NOT_PRODUCT(fail_eliminate = "null or TOP memory";)
688 can_eliminate = false;
689 } else if (!reduce_merge_precheck) {
690 safepoints->append_if_missing(sfpt);
691 }
692 } else if (use->is_InlineType() && use->as_InlineType()->get_oop() == res) {
693 // Look at uses
694 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); k < kmax; k++) {
695 Node* u = use->fast_out(k);
696 if (u->is_InlineType()) {
697 // Use in flat field can be eliminated
698 InlineTypeNode* vt = u->as_InlineType();
699 for (uint i = 0; i < vt->field_count(); ++i) {
700 if (vt->field_value(i) == use && !vt->field_is_flat(i)) {
701 can_eliminate = false; // Use in non-flat field
702 break;
703 }
704 }
705 } else {
706 // Add other uses to the worklist to process individually
707 worklist.push(u);
708 }
709 }
710 } else if (use->Opcode() == Op_StoreX && use->in(MemNode::Address) == res) {
711 // Store to mark word of inline type larval buffer
712 assert(res_type->is_inlinetypeptr(), "Unexpected store to mark word");
713 } else if (reduce_merge_precheck && (use->is_Phi() || use->is_EncodeP() || use->Opcode() == Op_MemBarRelease)) {
714 // Nothing to do
715 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
716 if (use->is_Phi()) {
717 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
718 NOT_PRODUCT(fail_eliminate = "Object is return value";)
719 } else {
720 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
721 }
722 DEBUG_ONLY(disq_node = use;)
723 } else {
724 if (use->Opcode() == Op_Return) {
725 NOT_PRODUCT(fail_eliminate = "Object is return value";)
726 } else {
727 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
728 }
729 DEBUG_ONLY(disq_node = use;)
730 }
731 can_eliminate = false;
732 } else {
733 assert(use->Opcode() == Op_CastP2X, "should be");
734 assert(!use->has_out_with(Op_OrL), "should have been removed because oop is never null");
735 }
736 }
737 }
738
739 #ifndef PRODUCT
740 if (PrintEliminateAllocations && safepoints != nullptr) {
741 if (can_eliminate) {
742 tty->print("Scalar ");
743 if (res == nullptr)
744 alloc->dump();
745 else
746 res->dump();
747 } else {
748 tty->print("NotScalar (%s)", fail_eliminate);
749 if (res == nullptr)
750 alloc->dump();
751 else
752 res->dump();
753 #ifdef ASSERT
754 if (disq_node != nullptr) {
755 tty->print(" >>>> ");
756 disq_node->dump();
757 }
758 #endif /*ASSERT*/
759 }
760 }
761 #endif
762 return can_eliminate;
763 }
764
765 void PhaseMacroExpand::undo_previous_scalarizations(GrowableArray <SafePointNode *> safepoints_done, AllocateNode* alloc) {
766 Node* res = alloc->result_cast();
767 int nfields = 0;
768 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
769
770 if (res != nullptr) {
771 const TypeOopPtr* res_type = _igvn.type(res)->isa_oopptr();
772
773 if (res_type->isa_instptr()) {
774 // find the fields of the class which will be needed for safepoint debug information
775 ciInstanceKlass* iklass = res_type->is_instptr()->instance_klass();
776 nfields = iklass->nof_nonstatic_fields();
777 } else {
778 // find the array's elements which will be needed for safepoint debug information
779 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
780 assert(nfields >= 0, "must be an array klass.");
781 }
782 }
783
784 // rollback processed safepoints
785 while (safepoints_done.length() > 0) {
786 SafePointNode* sfpt_done = safepoints_done.pop();
787 // remove any extra entries we added to the safepoint
788 uint last = sfpt_done->req() - 1;
789 for (int k = 0; k < nfields; k++) {
790 sfpt_done->del_req(last--);
791 }
792 JVMState *jvms = sfpt_done->jvms();
793 jvms->set_endoff(sfpt_done->req());
794 // Now make a pass over the debug information replacing any references
795 // to SafePointScalarObjectNode with the allocated object.
796 int start = jvms->debug_start();
797 int end = jvms->debug_end();
798 for (int i = start; i < end; i++) {
799 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
800 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
801 if (scobj->first_index(jvms) == sfpt_done->req() &&
802 scobj->n_fields() == (uint)nfields) {
803 assert(scobj->alloc() == alloc, "sanity");
804 sfpt_done->set_req(i, res);
805 }
806 }
807 }
808 _igvn._worklist.push(sfpt_done);
809 }
810 }
811
812 SafePointScalarObjectNode* PhaseMacroExpand::create_scalarized_object_description(AllocateNode *alloc, SafePointNode* sfpt,
813 Unique_Node_List* value_worklist) {
814 // Fields of scalar objs are referenced only at the end
815 // of regular debuginfo at the last (youngest) JVMS.
816 // Record relative start index.
817 ciInstanceKlass* iklass = nullptr;
818 BasicType basic_elem_type = T_ILLEGAL;
819 const Type* field_type = nullptr;
820 const TypeOopPtr* res_type = nullptr;
821 int nfields = 0;
822 int array_base = 0;
823 int element_size = 0;
824 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
825 Node* res = alloc->result_cast();
826
827 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
828 assert(sfpt->jvms() != nullptr, "missed JVMS");
829
830 if (res != nullptr) { // Could be null when there are no users
831 res_type = _igvn.type(res)->isa_oopptr();
832
833 if (res_type->isa_instptr()) {
834 // find the fields of the class which will be needed for safepoint debug information
835 iklass = res_type->is_instptr()->instance_klass();
836 nfields = iklass->nof_nonstatic_fields();
837 } else {
838 // find the array's elements which will be needed for safepoint debug information
839 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
840 assert(nfields >= 0, "must be an array klass.");
841 basic_elem_type = res_type->is_aryptr()->elem()->array_element_basic_type();
842 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
843 element_size = type2aelembytes(basic_elem_type);
844 field_type = res_type->is_aryptr()->elem();
845 if (res_type->is_flat()) {
846 // Flat inline type array
847 element_size = res_type->is_aryptr()->flat_elem_size();
848 }
849 }
850 }
851
852 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type, alloc, first_ind, nfields);
853 sobj->init_req(0, C->root());
854 transform_later(sobj);
855
856 // Scan object's fields adding an input to the safepoint for each field.
857 for (int j = 0; j < nfields; j++) {
858 intptr_t offset;
859 ciField* field = nullptr;
860 if (iklass != nullptr) {
861 field = iklass->nonstatic_field_at(j);
862 offset = field->offset_in_bytes();
863 ciType* elem_type = field->type();
864 basic_elem_type = field->layout_type();
865 assert(!field->is_flat(), "flat inline type fields should not have safepoint uses");
866
867 // The next code is taken from Parse::do_get_xxx().
868 if (is_reference_type(basic_elem_type)) {
869 if (!elem_type->is_loaded()) {
870 field_type = TypeInstPtr::BOTTOM;
871 } else if (field != nullptr && field->is_static_constant()) {
872 ciObject* con = field->constant_value().as_object();
873 // Do not "join" in the previous type; it doesn't add value,
874 // and may yield a vacuous result if the field is of interface type.
875 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
876 assert(field_type != nullptr, "field singleton type must be consistent");
877 } else {
878 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
879 }
880 if (UseCompressedOops) {
881 field_type = field_type->make_narrowoop();
882 basic_elem_type = T_NARROWOOP;
883 }
884 } else {
885 field_type = Type::get_const_basic_type(basic_elem_type);
886 }
887 } else {
888 offset = array_base + j * (intptr_t)element_size;
889 }
890
891 Node* field_val = nullptr;
892 const TypeOopPtr* field_addr_type = res_type->add_offset(offset)->isa_oopptr();
893 if (res_type->is_flat()) {
894 ciInlineKlass* vk = res_type->is_aryptr()->elem()->inline_klass();
895 assert(vk->flat_array(), "must be flat");
896 field_val = inline_type_from_mem(sfpt->memory(), sfpt->control(), vk, field_addr_type->isa_aryptr(), 0, alloc);
897 } else {
898 field_val = value_from_mem(sfpt->memory(), sfpt->control(), basic_elem_type, field_type, field_addr_type, alloc);
899 }
900
901 // We weren't able to find a value for this field,
902 // give up on eliminating this allocation.
903 if (field_val == nullptr) {
904 uint last = sfpt->req() - 1;
905 for (int k = 0; k < j; k++) {
906 sfpt->del_req(last--);
907 }
908 _igvn._worklist.push(sfpt);
909
910 #ifndef PRODUCT
911 if (PrintEliminateAllocations) {
912 if (field != nullptr) {
913 tty->print("=== At SafePoint node %d can't find value of field: ", sfpt->_idx);
914 field->print();
915 int field_idx = C->get_alias_index(field_addr_type);
916 tty->print(" (alias_idx=%d)", field_idx);
917 } else { // Array's element
918 tty->print("=== At SafePoint node %d can't find value of array element [%d]", sfpt->_idx, j);
919 }
920 tty->print(", which prevents elimination of: ");
921 if (res == nullptr)
922 alloc->dump();
923 else
924 res->dump();
925 }
926 #endif
927
928 return nullptr;
929 }
930
931 if (UseCompressedOops && field_type->isa_narrowoop()) {
932 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
933 // to be able scalar replace the allocation.
934 if (field_val->is_EncodeP()) {
935 field_val = field_val->in(1);
936 } else if (!field_val->is_InlineType()) {
937 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
938 }
939 }
940 if (field_val->is_InlineType()) {
941 // Keep track of inline types to scalarize them later
942 value_worklist->push(field_val);
943 }
944 sfpt->add_req(field_val);
945 }
946
947 sfpt->jvms()->set_endoff(sfpt->req());
948
949 return sobj;
950 }
951
952 // Do scalar replacement.
953 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
954 GrowableArray <SafePointNode *> safepoints_done;
955 Node* res = alloc->result_cast();
956 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
957 const TypeOopPtr* res_type = nullptr;
958 if (res != nullptr) { // Could be null when there are no users
959 res_type = _igvn.type(res)->isa_oopptr();
960 }
961
962 // Process the safepoint uses
963 assert(safepoints.length() == 0 || !res_type->is_inlinetypeptr(), "Inline type allocations should not have safepoint uses");
964 Unique_Node_List value_worklist;
965 while (safepoints.length() > 0) {
966 SafePointNode* sfpt = safepoints.pop();
967 SafePointScalarObjectNode* sobj = create_scalarized_object_description(alloc, sfpt, &value_worklist);
968
969 if (sobj == nullptr) {
970 undo_previous_scalarizations(safepoints_done, alloc);
971 return false;
972 }
973
974 // Now make a pass over the debug information replacing any references
975 // to the allocated object with "sobj"
976 JVMState *jvms = sfpt->jvms();
977 sfpt->replace_edges_in_range(res, sobj, jvms->debug_start(), jvms->debug_end(), &_igvn);
978 _igvn._worklist.push(sfpt);
979
980 // keep it for rollback
981 safepoints_done.append_if_missing(sfpt);
982 }
983 // Scalarize inline types that were added to the safepoint.
984 // Don't allow linking a constant oop (if available) for flat array elements
985 // because Deoptimization::reassign_flat_array_elements needs field values.
986 bool allow_oop = (res_type != nullptr) && !res_type->is_flat();
987 for (uint i = 0; i < value_worklist.size(); ++i) {
988 InlineTypeNode* vt = value_worklist.at(i)->as_InlineType();
989 vt->make_scalar_in_safepoints(&_igvn, allow_oop);
990 }
991 return true;
992 }
993
994 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
995 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
996 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
997 if (ctl_proj != nullptr) {
998 igvn.replace_node(ctl_proj, n->in(0));
999 }
1000 if (mem_proj != nullptr) {
1001 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
1002 }
1003 }
1004
1005 // Process users of eliminated allocation.
1006 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc, bool inline_alloc) {
1007 Unique_Node_List worklist;
1008 Node* res = alloc->result_cast();
1009 if (res != nullptr) {
1010 worklist.push(res);
1011 }
1012 while (worklist.size() > 0) {
1013 res = worklist.pop();
1014 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
1015 Node *use = res->last_out(j);
1016 uint oc1 = res->outcnt();
1017
1018 if (use->is_AddP()) {
1019 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
1020 Node *n = use->last_out(k);
1021 uint oc2 = use->outcnt();
1022 if (n->is_Store()) {
1023 for (DUIterator_Fast pmax, p = n->fast_outs(pmax); p < pmax; p++) {
1024 MemBarNode* mb = n->fast_out(p)->isa_MemBar();
1025 if (mb != nullptr && mb->req() <= MemBarNode::Precedent && mb->in(MemBarNode::Precedent) == n) {
1026 // MemBarVolatiles should have been removed by MemBarNode::Ideal() for non-inline allocations
1027 assert(inline_alloc, "MemBarVolatile should be eliminated for non-escaping object");
1028 mb->remove(&_igvn);
1029 }
1030 }
1031 _igvn.replace_node(n, n->in(MemNode::Memory));
1032 } else {
1033 eliminate_gc_barrier(n);
1034 }
1035 k -= (oc2 - use->outcnt());
1036 }
1037 _igvn.remove_dead_node(use);
1038 } else if (use->is_ArrayCopy()) {
1039 // Disconnect ArrayCopy node
1040 ArrayCopyNode* ac = use->as_ArrayCopy();
1041 if (ac->is_clonebasic()) {
1042 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
1043 disconnect_projections(ac, _igvn);
1044 assert(alloc->in(TypeFunc::Memory)->is_Proj() && alloc->in(TypeFunc::Memory)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
1045 Node* membar_before = alloc->in(TypeFunc::Memory)->in(0);
1046 disconnect_projections(membar_before->as_MemBar(), _igvn);
1047 if (membar_after->is_MemBar()) {
1048 disconnect_projections(membar_after->as_MemBar(), _igvn);
1049 }
1050 } else {
1051 assert(ac->is_arraycopy_validated() ||
1052 ac->is_copyof_validated() ||
1053 ac->is_copyofrange_validated(), "unsupported");
1054 CallProjections* callprojs = ac->extract_projections(true);
1055
1056 _igvn.replace_node(callprojs->fallthrough_ioproj, ac->in(TypeFunc::I_O));
1057 _igvn.replace_node(callprojs->fallthrough_memproj, ac->in(TypeFunc::Memory));
1058 _igvn.replace_node(callprojs->fallthrough_catchproj, ac->in(TypeFunc::Control));
1059
1060 // Set control to top. IGVN will remove the remaining projections
1061 ac->set_req(0, top());
1062 ac->replace_edge(res, top(), &_igvn);
1063
1064 // Disconnect src right away: it can help find new
1065 // opportunities for allocation elimination
1066 Node* src = ac->in(ArrayCopyNode::Src);
1067 ac->replace_edge(src, top(), &_igvn);
1068 // src can be top at this point if src and dest of the
1069 // arraycopy were the same
1070 if (src->outcnt() == 0 && !src->is_top()) {
1071 _igvn.remove_dead_node(src);
1072 }
1073 }
1074 _igvn._worklist.push(ac);
1075 } else if (use->is_InlineType()) {
1076 assert(use->as_InlineType()->get_oop() == res, "unexpected inline type ptr use");
1077 // Cut off oop input and remove known instance id from type
1078 _igvn.rehash_node_delayed(use);
1079 use->as_InlineType()->set_oop(_igvn.zerocon(T_PRIMITIVE_OBJECT));
1080 const TypeOopPtr* toop = _igvn.type(use)->is_oopptr()->cast_to_instance_id(TypeOopPtr::InstanceBot);
1081 _igvn.set_type(use, toop);
1082 use->as_InlineType()->set_type(toop);
1083 // Process users
1084 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); k < kmax; k++) {
1085 Node* u = use->fast_out(k);
1086 if (!u->is_InlineType()) {
1087 worklist.push(u);
1088 }
1089 }
1090 } else if (use->Opcode() == Op_StoreX && use->in(MemNode::Address) == res) {
1091 // Store to mark word of inline type larval buffer
1092 assert(inline_alloc, "Unexpected store to mark word");
1093 _igvn.replace_node(use, use->in(MemNode::Memory));
1094 } else {
1095 eliminate_gc_barrier(use);
1096 }
1097 j -= (oc1 - res->outcnt());
1098 }
1099 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
1100 _igvn.remove_dead_node(res);
1101 }
1102
1103 //
1104 // Process other users of allocation's projections
1105 //
1106 if (_callprojs->resproj[0] != nullptr && _callprojs->resproj[0]->outcnt() != 0) {
1107 // First disconnect stores captured by Initialize node.
1108 // If Initialize node is eliminated first in the following code,
1109 // it will kill such stores and DUIterator_Last will assert.
1110 for (DUIterator_Fast jmax, j = _callprojs->resproj[0]->fast_outs(jmax); j < jmax; j++) {
1111 Node* use = _callprojs->resproj[0]->fast_out(j);
1112 if (use->is_AddP()) {
1113 // raw memory addresses used only by the initialization
1114 _igvn.replace_node(use, C->top());
1115 --j; --jmax;
1116 }
1117 }
1118 for (DUIterator_Last jmin, j = _callprojs->resproj[0]->last_outs(jmin); j >= jmin; ) {
1119 Node* use = _callprojs->resproj[0]->last_out(j);
1120 uint oc1 = _callprojs->resproj[0]->outcnt();
1121 if (use->is_Initialize()) {
1122 // Eliminate Initialize node.
1123 InitializeNode *init = use->as_Initialize();
1124 assert(init->outcnt() <= 2, "only a control and memory projection expected");
1125 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
1126 if (ctrl_proj != nullptr) {
1127 _igvn.replace_node(ctrl_proj, init->in(TypeFunc::Control));
1128 #ifdef ASSERT
1129 // If the InitializeNode has no memory out, it will die, and tmp will become null
1130 Node* tmp = init->in(TypeFunc::Control);
1131 assert(tmp == nullptr || tmp == _callprojs->fallthrough_catchproj, "allocation control projection");
1132 #endif
1133 }
1134 Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory);
1135 if (mem_proj != nullptr) {
1136 Node *mem = init->in(TypeFunc::Memory);
1137 #ifdef ASSERT
1138 if (mem->is_MergeMem()) {
1139 assert(mem->in(TypeFunc::Memory) == _callprojs->fallthrough_memproj, "allocation memory projection");
1140 } else {
1141 assert(mem == _callprojs->fallthrough_memproj, "allocation memory projection");
1142 }
1143 #endif
1144 _igvn.replace_node(mem_proj, mem);
1145 }
1146 } else if (use->Opcode() == Op_MemBarStoreStore) {
1147 // Inline type buffer allocations are followed by a membar
1148 assert(inline_alloc, "Unexpected MemBarStoreStore");
1149 use->as_MemBar()->remove(&_igvn);
1150 } else {
1151 assert(false, "only Initialize or AddP expected");
1152 }
1153 j -= (oc1 - _callprojs->resproj[0]->outcnt());
1154 }
1155 }
1156 if (_callprojs->fallthrough_catchproj != nullptr) {
1157 _igvn.replace_node(_callprojs->fallthrough_catchproj, alloc->in(TypeFunc::Control));
1158 }
1159 if (_callprojs->fallthrough_memproj != nullptr) {
1160 _igvn.replace_node(_callprojs->fallthrough_memproj, alloc->in(TypeFunc::Memory));
1161 }
1162 if (_callprojs->catchall_memproj != nullptr) {
1163 _igvn.replace_node(_callprojs->catchall_memproj, C->top());
1164 }
1165 if (_callprojs->fallthrough_ioproj != nullptr) {
1166 _igvn.replace_node(_callprojs->fallthrough_ioproj, alloc->in(TypeFunc::I_O));
1167 }
1168 if (_callprojs->catchall_ioproj != nullptr) {
1169 _igvn.replace_node(_callprojs->catchall_ioproj, C->top());
1170 }
1171 if (_callprojs->catchall_catchproj != nullptr) {
1172 _igvn.replace_node(_callprojs->catchall_catchproj, C->top());
1173 }
1174 }
1175
1176 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1177 // If reallocation fails during deoptimization we'll pop all
1178 // interpreter frames for this compiled frame and that won't play
1179 // nice with JVMTI popframe.
1180 // We avoid this issue by eager reallocation when the popframe request
1181 // is received.
1182 if (!EliminateAllocations) {
1183 return false;
1184 }
1185 Node* klass = alloc->in(AllocateNode::KlassNode);
1186 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1187
1188 // Attempt to eliminate inline type buffer allocations
1189 // regardless of usage and escape/replaceable status.
1190 bool inline_alloc = tklass->isa_instklassptr() &&
1191 tklass->is_instklassptr()->instance_klass()->is_inlinetype();
1192 if (!alloc->_is_non_escaping && !inline_alloc) {
1193 return false;
1194 }
1195 // Eliminate boxing allocations which are not used
1196 // regardless scalar replaceable status.
1197 Node* res = alloc->result_cast();
1198 bool boxing_alloc = (res == nullptr) && C->eliminate_boxing() &&
1199 tklass->isa_instklassptr() &&
1200 tklass->is_instklassptr()->instance_klass()->is_box_klass();
1201 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != nullptr))) {
1202 return false;
1203 }
1204
1205 _callprojs = alloc->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1206
1207 GrowableArray <SafePointNode *> safepoints;
1208 if (!can_eliminate_allocation(&_igvn, alloc, &safepoints)) {
1209 return false;
1210 }
1211
1212 if (!alloc->_is_scalar_replaceable) {
1213 assert(res == nullptr || inline_alloc, "sanity");
1214 // We can only eliminate allocation if all debug info references
1215 // are already replaced with SafePointScalarObject because
1216 // we can't search for a fields value without instance_id.
1217 if (safepoints.length() > 0) {
1218 assert(!inline_alloc, "Inline type allocations should not have safepoint uses");
1219 return false;
1220 }
1221 }
1222
1223 if (!scalar_replacement(alloc, safepoints)) {
1224 return false;
1225 }
1226
1227 CompileLog* log = C->log();
1228 if (log != nullptr) {
1229 log->head("eliminate_allocation type='%d'",
1230 log->identify(tklass->exact_klass()));
1231 JVMState* p = alloc->jvms();
1232 while (p != nullptr) {
1233 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1234 p = p->caller();
1235 }
1236 log->tail("eliminate_allocation");
1237 }
1238
1239 process_users_of_allocation(alloc, inline_alloc);
1240
1241 #ifndef PRODUCT
1242 if (PrintEliminateAllocations) {
1243 if (alloc->is_AllocateArray())
1244 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1245 else
1246 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1247 }
1248 #endif
1249
1250 return true;
1251 }
1252
1253 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1254 // EA should remove all uses of non-escaping boxing node.
1255 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != nullptr) {
1256 return false;
1257 }
1258
1259 assert(boxing->result_cast() == nullptr, "unexpected boxing node result");
1260
1261 _callprojs = boxing->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1262
1263 const TypeTuple* r = boxing->tf()->range_sig();
1264 assert(r->cnt() > TypeFunc::Parms, "sanity");
1265 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1266 assert(t != nullptr, "sanity");
1267
1268 CompileLog* log = C->log();
1269 if (log != nullptr) {
1270 log->head("eliminate_boxing type='%d'",
1271 log->identify(t->instance_klass()));
1272 JVMState* p = boxing->jvms();
1273 while (p != nullptr) {
1274 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1275 p = p->caller();
1276 }
1277 log->tail("eliminate_boxing");
1278 }
1279
1280 process_users_of_allocation(boxing);
1281
1282 #ifndef PRODUCT
1283 if (PrintEliminateAllocations) {
1284 tty->print("++++ Eliminated: %d ", boxing->_idx);
1285 boxing->method()->print_short_name(tty);
1286 tty->cr();
1287 }
1288 #endif
1289
1290 return true;
1291 }
1292
1293
1294 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1295 Node* adr = basic_plus_adr(base, offset);
1296 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1297 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered);
1298 transform_later(value);
1299 return value;
1300 }
1301
1302
1303 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1304 Node* adr = basic_plus_adr(base, offset);
1305 mem = StoreNode::make(_igvn, ctl, mem, adr, nullptr, value, bt, MemNode::unordered);
1306 transform_later(mem);
1307 return mem;
1308 }
1309
1310 //=============================================================================
1311 //
1312 // A L L O C A T I O N
1313 //
1314 // Allocation attempts to be fast in the case of frequent small objects.
1315 // It breaks down like this:
1316 //
1317 // 1) Size in doublewords is computed. This is a constant for objects and
1318 // variable for most arrays. Doubleword units are used to avoid size
1319 // overflow of huge doubleword arrays. We need doublewords in the end for
1320 // rounding.
1321 //
1322 // 2) Size is checked for being 'too large'. Too-large allocations will go
1323 // the slow path into the VM. The slow path can throw any required
1324 // exceptions, and does all the special checks for very large arrays. The
1325 // size test can constant-fold away for objects. For objects with
1326 // finalizers it constant-folds the otherway: you always go slow with
1327 // finalizers.
1328 //
1329 // 3) If NOT using TLABs, this is the contended loop-back point.
1330 // Load-Locked the heap top. If using TLABs normal-load the heap top.
1331 //
1332 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
1333 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
1334 // "size*8" we always enter the VM, where "largish" is a constant picked small
1335 // enough that there's always space between the eden max and 4Gig (old space is
1336 // there so it's quite large) and large enough that the cost of entering the VM
1337 // is dwarfed by the cost to initialize the space.
1338 //
1339 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1340 // down. If contended, repeat at step 3. If using TLABs normal-store
1341 // adjusted heap top back down; there is no contention.
1342 //
1343 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
1344 // fields.
1345 //
1346 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1347 // oop flavor.
1348 //
1349 //=============================================================================
1350 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1351 // Allocations bigger than this always go the slow route.
1352 // This value must be small enough that allocation attempts that need to
1353 // trigger exceptions go the slow route. Also, it must be small enough so
1354 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1355 //=============================================================================j//
1356 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1357 // The allocator will coalesce int->oop copies away. See comment in
1358 // coalesce.cpp about how this works. It depends critically on the exact
1359 // code shape produced here, so if you are changing this code shape
1360 // make sure the GC info for the heap-top is correct in and around the
1361 // slow-path call.
1362 //
1363
1364 void PhaseMacroExpand::expand_allocate_common(
1365 AllocateNode* alloc, // allocation node to be expanded
1366 Node* length, // array length for an array allocation
1367 const TypeFunc* slow_call_type, // Type of slow call
1368 address slow_call_address, // Address of slow call
1369 Node* valid_length_test // whether length is valid or not
1370 )
1371 {
1372 Node* ctrl = alloc->in(TypeFunc::Control);
1373 Node* mem = alloc->in(TypeFunc::Memory);
1374 Node* i_o = alloc->in(TypeFunc::I_O);
1375 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1376 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1377 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1378 assert(ctrl != nullptr, "must have control");
1379
1380 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1381 // they will not be used if "always_slow" is set
1382 enum { slow_result_path = 1, fast_result_path = 2 };
1383 Node *result_region = nullptr;
1384 Node *result_phi_rawmem = nullptr;
1385 Node *result_phi_rawoop = nullptr;
1386 Node *result_phi_i_o = nullptr;
1387
1388 // The initial slow comparison is a size check, the comparison
1389 // we want to do is a BoolTest::gt
1390 bool expand_fast_path = true;
1391 int tv = _igvn.find_int_con(initial_slow_test, -1);
1392 if (tv >= 0) {
1393 // InitialTest has constant result
1394 // 0 - can fit in TLAB
1395 // 1 - always too big or negative
1396 assert(tv <= 1, "0 or 1 if a constant");
1397 expand_fast_path = (tv == 0);
1398 initial_slow_test = nullptr;
1399 } else {
1400 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1401 }
1402
1403 if (!UseTLAB) {
1404 // Force slow-path allocation
1405 expand_fast_path = false;
1406 initial_slow_test = nullptr;
1407 }
1408
1409 bool allocation_has_use = (alloc->result_cast() != nullptr);
1410 if (!allocation_has_use) {
1411 InitializeNode* init = alloc->initialization();
1412 if (init != nullptr) {
1413 init->remove(&_igvn);
1414 }
1415 if (expand_fast_path && (initial_slow_test == nullptr)) {
1416 // Remove allocation node and return.
1417 // Size is a non-negative constant -> no initial check needed -> directly to fast path.
1418 // Also, no usages -> empty fast path -> no fall out to slow path -> nothing left.
1419 #ifndef PRODUCT
1420 if (PrintEliminateAllocations) {
1421 tty->print("NotUsed ");
1422 Node* res = alloc->proj_out_or_null(TypeFunc::Parms);
1423 if (res != nullptr) {
1424 res->dump();
1425 } else {
1426 alloc->dump();
1427 }
1428 }
1429 #endif
1430 yank_alloc_node(alloc);
1431 return;
1432 }
1433 }
1434
1435 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1436 Node *slow_region = nullptr;
1437 Node *toobig_false = ctrl;
1438
1439 // generate the initial test if necessary
1440 if (initial_slow_test != nullptr ) {
1441 assert (expand_fast_path, "Only need test if there is a fast path");
1442 slow_region = new RegionNode(3);
1443
1444 // Now make the initial failure test. Usually a too-big test but
1445 // might be a TRUE for finalizers or a fancy class check for
1446 // newInstance0.
1447 IfNode* toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1448 transform_later(toobig_iff);
1449 // Plug the failing-too-big test into the slow-path region
1450 Node* toobig_true = new IfTrueNode(toobig_iff);
1451 transform_later(toobig_true);
1452 slow_region ->init_req( too_big_or_final_path, toobig_true );
1453 toobig_false = new IfFalseNode(toobig_iff);
1454 transform_later(toobig_false);
1455 } else {
1456 // No initial test, just fall into next case
1457 assert(allocation_has_use || !expand_fast_path, "Should already have been handled");
1458 toobig_false = ctrl;
1459 debug_only(slow_region = NodeSentinel);
1460 }
1461
1462 // If we are here there are several possibilities
1463 // - expand_fast_path is false - then only a slow path is expanded. That's it.
1464 // no_initial_check means a constant allocation.
1465 // - If check always evaluates to false -> expand_fast_path is false (see above)
1466 // - If check always evaluates to true -> directly into fast path (but may bailout to slowpath)
1467 // if !allocation_has_use the fast path is empty
1468 // if !allocation_has_use && no_initial_check
1469 // - Then there are no fastpath that can fall out to slowpath -> no allocation code at all.
1470 // removed by yank_alloc_node above.
1471
1472 Node *slow_mem = mem; // save the current memory state for slow path
1473 // generate the fast allocation code unless we know that the initial test will always go slow
1474 if (expand_fast_path) {
1475 // Fast path modifies only raw memory.
1476 if (mem->is_MergeMem()) {
1477 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1478 }
1479
1480 // allocate the Region and Phi nodes for the result
1481 result_region = new RegionNode(3);
1482 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1483 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1484
1485 // Grab regular I/O before optional prefetch may change it.
1486 // Slow-path does no I/O so just set it to the original I/O.
1487 result_phi_i_o->init_req(slow_result_path, i_o);
1488
1489 // Name successful fast-path variables
1490 Node* fast_oop_ctrl;
1491 Node* fast_oop_rawmem;
1492
1493 if (allocation_has_use) {
1494 Node* needgc_ctrl = nullptr;
1495 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1496
1497 intx prefetch_lines = length != nullptr ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1498 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1499 Node* fast_oop = bs->obj_allocate(this, mem, toobig_false, size_in_bytes, i_o, needgc_ctrl,
1500 fast_oop_ctrl, fast_oop_rawmem,
1501 prefetch_lines);
1502
1503 if (initial_slow_test != nullptr) {
1504 // This completes all paths into the slow merge point
1505 slow_region->init_req(need_gc_path, needgc_ctrl);
1506 transform_later(slow_region);
1507 } else {
1508 // No initial slow path needed!
1509 // Just fall from the need-GC path straight into the VM call.
1510 slow_region = needgc_ctrl;
1511 }
1512
1513 InitializeNode* init = alloc->initialization();
1514 fast_oop_rawmem = initialize_object(alloc,
1515 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1516 klass_node, length, size_in_bytes);
1517 expand_initialize_membar(alloc, init, fast_oop_ctrl, fast_oop_rawmem);
1518 expand_dtrace_alloc_probe(alloc, fast_oop, fast_oop_ctrl, fast_oop_rawmem);
1519
1520 result_phi_rawoop->init_req(fast_result_path, fast_oop);
1521 } else {
1522 assert (initial_slow_test != nullptr, "sanity");
1523 fast_oop_ctrl = toobig_false;
1524 fast_oop_rawmem = mem;
1525 transform_later(slow_region);
1526 }
1527
1528 // Plug in the successful fast-path into the result merge point
1529 result_region ->init_req(fast_result_path, fast_oop_ctrl);
1530 result_phi_i_o ->init_req(fast_result_path, i_o);
1531 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1532 } else {
1533 slow_region = ctrl;
1534 result_phi_i_o = i_o; // Rename it to use in the following code.
1535 }
1536
1537 // Generate slow-path call
1538 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1539 OptoRuntime::stub_name(slow_call_address),
1540 TypePtr::BOTTOM);
1541 call->init_req(TypeFunc::Control, slow_region);
1542 call->init_req(TypeFunc::I_O, top()); // does no i/o
1543 call->init_req(TypeFunc::Memory, slow_mem); // may gc ptrs
1544 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1545 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1546
1547 call->init_req(TypeFunc::Parms+0, klass_node);
1548 if (length != nullptr) {
1549 call->init_req(TypeFunc::Parms+1, length);
1550 } else {
1551 // Let the runtime know if this is a larval allocation
1552 call->init_req(TypeFunc::Parms+1, _igvn.intcon(alloc->_larval));
1553 }
1554
1555 // Copy debug information and adjust JVMState information, then replace
1556 // allocate node with the call
1557 call->copy_call_debug_info(&_igvn, alloc);
1558 // For array allocations, copy the valid length check to the call node so Compile::final_graph_reshaping() can verify
1559 // that the call has the expected number of CatchProj nodes (in case the allocation always fails and the fallthrough
1560 // path dies).
1561 if (valid_length_test != nullptr) {
1562 call->add_req(valid_length_test);
1563 }
1564 if (expand_fast_path) {
1565 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1566 } else {
1567 // Hook i_o projection to avoid its elimination during allocation
1568 // replacement (when only a slow call is generated).
1569 call->set_req(TypeFunc::I_O, result_phi_i_o);
1570 }
1571 _igvn.replace_node(alloc, call);
1572 transform_later(call);
1573
1574 // Identify the output projections from the allocate node and
1575 // adjust any references to them.
1576 // The control and io projections look like:
1577 //
1578 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1579 // Allocate Catch
1580 // ^---Proj(io) <-------+ ^---CatchProj(io)
1581 //
1582 // We are interested in the CatchProj nodes.
1583 //
1584 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1585
1586 // An allocate node has separate memory projections for the uses on
1587 // the control and i_o paths. Replace the control memory projection with
1588 // result_phi_rawmem (unless we are only generating a slow call when
1589 // both memory projections are combined)
1590 if (expand_fast_path && _callprojs->fallthrough_memproj != nullptr) {
1591 _igvn.replace_in_uses(_callprojs->fallthrough_memproj, result_phi_rawmem);
1592 }
1593 // Now change uses of catchall_memproj to use fallthrough_memproj and delete
1594 // catchall_memproj so we end up with a call that has only 1 memory projection.
1595 if (_callprojs->catchall_memproj != nullptr) {
1596 if (_callprojs->fallthrough_memproj == nullptr) {
1597 _callprojs->fallthrough_memproj = new ProjNode(call, TypeFunc::Memory);
1598 transform_later(_callprojs->fallthrough_memproj);
1599 }
1600 _igvn.replace_in_uses(_callprojs->catchall_memproj, _callprojs->fallthrough_memproj);
1601 _igvn.remove_dead_node(_callprojs->catchall_memproj);
1602 }
1603
1604 // An allocate node has separate i_o projections for the uses on the control
1605 // and i_o paths. Always replace the control i_o projection with result i_o
1606 // otherwise incoming i_o become dead when only a slow call is generated
1607 // (it is different from memory projections where both projections are
1608 // combined in such case).
1609 if (_callprojs->fallthrough_ioproj != nullptr) {
1610 _igvn.replace_in_uses(_callprojs->fallthrough_ioproj, result_phi_i_o);
1611 }
1612 // Now change uses of catchall_ioproj to use fallthrough_ioproj and delete
1613 // catchall_ioproj so we end up with a call that has only 1 i_o projection.
1614 if (_callprojs->catchall_ioproj != nullptr) {
1615 if (_callprojs->fallthrough_ioproj == nullptr) {
1616 _callprojs->fallthrough_ioproj = new ProjNode(call, TypeFunc::I_O);
1617 transform_later(_callprojs->fallthrough_ioproj);
1618 }
1619 _igvn.replace_in_uses(_callprojs->catchall_ioproj, _callprojs->fallthrough_ioproj);
1620 _igvn.remove_dead_node(_callprojs->catchall_ioproj);
1621 }
1622
1623 // if we generated only a slow call, we are done
1624 if (!expand_fast_path) {
1625 // Now we can unhook i_o.
1626 if (result_phi_i_o->outcnt() > 1) {
1627 call->set_req(TypeFunc::I_O, top());
1628 } else {
1629 assert(result_phi_i_o->unique_ctrl_out() == call, "sanity");
1630 // Case of new array with negative size known during compilation.
1631 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1632 // following code since call to runtime will throw exception.
1633 // As result there will be no users of i_o after the call.
1634 // Leave i_o attached to this call to avoid problems in preceding graph.
1635 }
1636 return;
1637 }
1638
1639 if (_callprojs->fallthrough_catchproj != nullptr) {
1640 ctrl = _callprojs->fallthrough_catchproj->clone();
1641 transform_later(ctrl);
1642 _igvn.replace_node(_callprojs->fallthrough_catchproj, result_region);
1643 } else {
1644 ctrl = top();
1645 }
1646 Node *slow_result;
1647 if (_callprojs->resproj[0] == nullptr) {
1648 // no uses of the allocation result
1649 slow_result = top();
1650 } else {
1651 slow_result = _callprojs->resproj[0]->clone();
1652 transform_later(slow_result);
1653 _igvn.replace_node(_callprojs->resproj[0], result_phi_rawoop);
1654 }
1655
1656 // Plug slow-path into result merge point
1657 result_region->init_req( slow_result_path, ctrl);
1658 transform_later(result_region);
1659 if (allocation_has_use) {
1660 result_phi_rawoop->init_req(slow_result_path, slow_result);
1661 transform_later(result_phi_rawoop);
1662 }
1663 result_phi_rawmem->init_req(slow_result_path, _callprojs->fallthrough_memproj);
1664 transform_later(result_phi_rawmem);
1665 transform_later(result_phi_i_o);
1666 // This completes all paths into the result merge point
1667 }
1668
1669 // Remove alloc node that has no uses.
1670 void PhaseMacroExpand::yank_alloc_node(AllocateNode* alloc) {
1671 Node* ctrl = alloc->in(TypeFunc::Control);
1672 Node* mem = alloc->in(TypeFunc::Memory);
1673 Node* i_o = alloc->in(TypeFunc::I_O);
1674
1675 _callprojs = alloc->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1676 if (_callprojs->resproj[0] != nullptr) {
1677 for (DUIterator_Fast imax, i = _callprojs->resproj[0]->fast_outs(imax); i < imax; i++) {
1678 Node* use = _callprojs->resproj[0]->fast_out(i);
1679 use->isa_MemBar()->remove(&_igvn);
1680 --imax;
1681 --i; // back up iterator
1682 }
1683 assert(_callprojs->resproj[0]->outcnt() == 0, "all uses must be deleted");
1684 _igvn.remove_dead_node(_callprojs->resproj[0]);
1685 }
1686 if (_callprojs->fallthrough_catchproj != nullptr) {
1687 _igvn.replace_in_uses(_callprojs->fallthrough_catchproj, ctrl);
1688 _igvn.remove_dead_node(_callprojs->fallthrough_catchproj);
1689 }
1690 if (_callprojs->catchall_catchproj != nullptr) {
1691 _igvn.rehash_node_delayed(_callprojs->catchall_catchproj);
1692 _callprojs->catchall_catchproj->set_req(0, top());
1693 }
1694 if (_callprojs->fallthrough_proj != nullptr) {
1695 Node* catchnode = _callprojs->fallthrough_proj->unique_ctrl_out();
1696 _igvn.remove_dead_node(catchnode);
1697 _igvn.remove_dead_node(_callprojs->fallthrough_proj);
1698 }
1699 if (_callprojs->fallthrough_memproj != nullptr) {
1700 _igvn.replace_in_uses(_callprojs->fallthrough_memproj, mem);
1701 _igvn.remove_dead_node(_callprojs->fallthrough_memproj);
1702 }
1703 if (_callprojs->fallthrough_ioproj != nullptr) {
1704 _igvn.replace_in_uses(_callprojs->fallthrough_ioproj, i_o);
1705 _igvn.remove_dead_node(_callprojs->fallthrough_ioproj);
1706 }
1707 if (_callprojs->catchall_memproj != nullptr) {
1708 _igvn.rehash_node_delayed(_callprojs->catchall_memproj);
1709 _callprojs->catchall_memproj->set_req(0, top());
1710 }
1711 if (_callprojs->catchall_ioproj != nullptr) {
1712 _igvn.rehash_node_delayed(_callprojs->catchall_ioproj);
1713 _callprojs->catchall_ioproj->set_req(0, top());
1714 }
1715 #ifndef PRODUCT
1716 if (PrintEliminateAllocations) {
1717 if (alloc->is_AllocateArray()) {
1718 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1719 } else {
1720 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1721 }
1722 }
1723 #endif
1724 _igvn.remove_dead_node(alloc);
1725 }
1726
1727 void PhaseMacroExpand::expand_initialize_membar(AllocateNode* alloc, InitializeNode* init,
1728 Node*& fast_oop_ctrl, Node*& fast_oop_rawmem) {
1729 // If initialization is performed by an array copy, any required
1730 // MemBarStoreStore was already added. If the object does not
1731 // escape no need for a MemBarStoreStore. If the object does not
1732 // escape in its initializer and memory barrier (MemBarStoreStore or
1733 // stronger) is already added at exit of initializer, also no need
1734 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore
1735 // so that stores that initialize this object can't be reordered
1736 // with a subsequent store that makes this object accessible by
1737 // other threads.
1738 // Other threads include java threads and JVM internal threads
1739 // (for example concurrent GC threads). Current concurrent GC
1740 // implementation: G1 will not scan newly created object,
1741 // so it's safe to skip storestore barrier when allocation does
1742 // not escape.
1743 if (!alloc->does_not_escape_thread() &&
1744 !alloc->is_allocation_MemBar_redundant() &&
1745 (init == nullptr || !init->is_complete_with_arraycopy())) {
1746 if (init == nullptr || init->req() < InitializeNode::RawStores) {
1747 // No InitializeNode or no stores captured by zeroing
1748 // elimination. Simply add the MemBarStoreStore after object
1749 // initialization.
1750 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1751 transform_later(mb);
1752
1753 mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
1754 mb->init_req(TypeFunc::Control, fast_oop_ctrl);
1755 fast_oop_ctrl = new ProjNode(mb, TypeFunc::Control);
1756 transform_later(fast_oop_ctrl);
1757 fast_oop_rawmem = new ProjNode(mb, TypeFunc::Memory);
1758 transform_later(fast_oop_rawmem);
1759 } else {
1760 // Add the MemBarStoreStore after the InitializeNode so that
1761 // all stores performing the initialization that were moved
1762 // before the InitializeNode happen before the storestore
1763 // barrier.
1764
1765 Node* init_ctrl = init->proj_out_or_null(TypeFunc::Control);
1766 Node* init_mem = init->proj_out_or_null(TypeFunc::Memory);
1767
1768 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1769 transform_later(mb);
1770
1771 Node* ctrl = new ProjNode(init, TypeFunc::Control);
1772 transform_later(ctrl);
1773 Node* mem = new ProjNode(init, TypeFunc::Memory);
1774 transform_later(mem);
1775
1776 // The MemBarStoreStore depends on control and memory coming
1777 // from the InitializeNode
1778 mb->init_req(TypeFunc::Memory, mem);
1779 mb->init_req(TypeFunc::Control, ctrl);
1780
1781 ctrl = new ProjNode(mb, TypeFunc::Control);
1782 transform_later(ctrl);
1783 mem = new ProjNode(mb, TypeFunc::Memory);
1784 transform_later(mem);
1785
1786 // All nodes that depended on the InitializeNode for control
1787 // and memory must now depend on the MemBarNode that itself
1788 // depends on the InitializeNode
1789 if (init_ctrl != nullptr) {
1790 _igvn.replace_node(init_ctrl, ctrl);
1791 }
1792 if (init_mem != nullptr) {
1793 _igvn.replace_node(init_mem, mem);
1794 }
1795 }
1796 }
1797 }
1798
1799 void PhaseMacroExpand::expand_dtrace_alloc_probe(AllocateNode* alloc, Node* oop,
1800 Node*& ctrl, Node*& rawmem) {
1801 if (C->env()->dtrace_alloc_probes()) {
1802 // Slow-path call
1803 int size = TypeFunc::Parms + 2;
1804 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1805 CAST_FROM_FN_PTR(address,
1806 static_cast<int (*)(JavaThread*, oopDesc*)>(SharedRuntime::dtrace_object_alloc)),
1807 "dtrace_object_alloc",
1808 TypeRawPtr::BOTTOM);
1809
1810 // Get base of thread-local storage area
1811 Node* thread = new ThreadLocalNode();
1812 transform_later(thread);
1813
1814 call->init_req(TypeFunc::Parms + 0, thread);
1815 call->init_req(TypeFunc::Parms + 1, oop);
1816 call->init_req(TypeFunc::Control, ctrl);
1817 call->init_req(TypeFunc::I_O , top()); // does no i/o
1818 call->init_req(TypeFunc::Memory , rawmem);
1819 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1820 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1821 transform_later(call);
1822 ctrl = new ProjNode(call, TypeFunc::Control);
1823 transform_later(ctrl);
1824 rawmem = new ProjNode(call, TypeFunc::Memory);
1825 transform_later(rawmem);
1826 }
1827 }
1828
1829 // Helper for PhaseMacroExpand::expand_allocate_common.
1830 // Initializes the newly-allocated storage.
1831 Node* PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1832 Node* control, Node* rawmem, Node* object,
1833 Node* klass_node, Node* length,
1834 Node* size_in_bytes) {
1835 InitializeNode* init = alloc->initialization();
1836 // Store the klass & mark bits
1837 Node* mark_node = alloc->make_ideal_mark(&_igvn, control, rawmem);
1838 if (!mark_node->is_Con()) {
1839 transform_later(mark_node);
1840 }
1841 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, TypeX_X->basic_type());
1842
1843 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1844 int header_size = alloc->minimum_header_size(); // conservatively small
1845
1846 // Array length
1847 if (length != nullptr) { // Arrays need length field
1848 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1849 // conservatively small header size:
1850 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1851 if (_igvn.type(klass_node)->isa_aryklassptr()) { // we know the exact header size in most cases:
1852 BasicType elem = _igvn.type(klass_node)->is_klassptr()->as_instance_type()->isa_aryptr()->elem()->array_element_basic_type();
1853 if (is_reference_type(elem, true)) {
1854 elem = T_OBJECT;
1855 }
1856 header_size = Klass::layout_helper_header_size(Klass::array_layout_helper(elem));
1857 }
1858 }
1859
1860 // Clear the object body, if necessary.
1861 if (init == nullptr) {
1862 // The init has somehow disappeared; be cautious and clear everything.
1863 //
1864 // This can happen if a node is allocated but an uncommon trap occurs
1865 // immediately. In this case, the Initialize gets associated with the
1866 // trap, and may be placed in a different (outer) loop, if the Allocate
1867 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1868 // there can be two Allocates to one Initialize. The answer in all these
1869 // edge cases is safety first. It is always safe to clear immediately
1870 // within an Allocate, and then (maybe or maybe not) clear some more later.
1871 if (!(UseTLAB && ZeroTLAB)) {
1872 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1873 alloc->in(AllocateNode::DefaultValue),
1874 alloc->in(AllocateNode::RawDefaultValue),
1875 header_size, size_in_bytes,
1876 &_igvn);
1877 }
1878 } else {
1879 if (!init->is_complete()) {
1880 // Try to win by zeroing only what the init does not store.
1881 // We can also try to do some peephole optimizations,
1882 // such as combining some adjacent subword stores.
1883 rawmem = init->complete_stores(control, rawmem, object,
1884 header_size, size_in_bytes, &_igvn);
1885 }
1886 // We have no more use for this link, since the AllocateNode goes away:
1887 init->set_req(InitializeNode::RawAddress, top());
1888 // (If we keep the link, it just confuses the register allocator,
1889 // who thinks he sees a real use of the address by the membar.)
1890 }
1891
1892 return rawmem;
1893 }
1894
1895 // Generate prefetch instructions for next allocations.
1896 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1897 Node*& contended_phi_rawmem,
1898 Node* old_eden_top, Node* new_eden_top,
1899 intx lines) {
1900 enum { fall_in_path = 1, pf_path = 2 };
1901 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1902 // Generate prefetch allocation with watermark check.
1903 // As an allocation hits the watermark, we will prefetch starting
1904 // at a "distance" away from watermark.
1905
1906 Node *pf_region = new RegionNode(3);
1907 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY,
1908 TypeRawPtr::BOTTOM );
1909 // I/O is used for Prefetch
1910 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO );
1911
1912 Node *thread = new ThreadLocalNode();
1913 transform_later(thread);
1914
1915 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread,
1916 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1917 transform_later(eden_pf_adr);
1918
1919 Node *old_pf_wm = new LoadPNode(needgc_false,
1920 contended_phi_rawmem, eden_pf_adr,
1921 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM,
1922 MemNode::unordered);
1923 transform_later(old_pf_wm);
1924
1925 // check against new_eden_top
1926 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm );
1927 transform_later(need_pf_cmp);
1928 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge );
1929 transform_later(need_pf_bol);
1930 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol,
1931 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1932 transform_later(need_pf_iff);
1933
1934 // true node, add prefetchdistance
1935 Node *need_pf_true = new IfTrueNode( need_pf_iff );
1936 transform_later(need_pf_true);
1937
1938 Node *need_pf_false = new IfFalseNode( need_pf_iff );
1939 transform_later(need_pf_false);
1940
1941 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm,
1942 _igvn.MakeConX(AllocatePrefetchDistance) );
1943 transform_later(new_pf_wmt );
1944 new_pf_wmt->set_req(0, need_pf_true);
1945
1946 Node *store_new_wmt = new StorePNode(need_pf_true,
1947 contended_phi_rawmem, eden_pf_adr,
1948 TypeRawPtr::BOTTOM, new_pf_wmt,
1949 MemNode::unordered);
1950 transform_later(store_new_wmt);
1951
1952 // adding prefetches
1953 pf_phi_abio->init_req( fall_in_path, i_o );
1954
1955 Node *prefetch_adr;
1956 Node *prefetch;
1957 uint step_size = AllocatePrefetchStepSize;
1958 uint distance = 0;
1959
1960 for ( intx i = 0; i < lines; i++ ) {
1961 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt,
1962 _igvn.MakeConX(distance) );
1963 transform_later(prefetch_adr);
1964 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1965 transform_later(prefetch);
1966 distance += step_size;
1967 i_o = prefetch;
1968 }
1969 pf_phi_abio->set_req( pf_path, i_o );
1970
1971 pf_region->init_req( fall_in_path, need_pf_false );
1972 pf_region->init_req( pf_path, need_pf_true );
1973
1974 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
1975 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
1976
1977 transform_later(pf_region);
1978 transform_later(pf_phi_rawmem);
1979 transform_later(pf_phi_abio);
1980
1981 needgc_false = pf_region;
1982 contended_phi_rawmem = pf_phi_rawmem;
1983 i_o = pf_phi_abio;
1984 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) {
1985 // Insert a prefetch instruction for each allocation.
1986 // This code is used to generate 1 prefetch instruction per cache line.
1987
1988 // Generate several prefetch instructions.
1989 uint step_size = AllocatePrefetchStepSize;
1990 uint distance = AllocatePrefetchDistance;
1991
1992 // Next cache address.
1993 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top,
1994 _igvn.MakeConX(step_size + distance));
1995 transform_later(cache_adr);
1996 cache_adr = new CastP2XNode(needgc_false, cache_adr);
1997 transform_later(cache_adr);
1998 // Address is aligned to execute prefetch to the beginning of cache line size
1999 // (it is important when BIS instruction is used on SPARC as prefetch).
2000 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
2001 cache_adr = new AndXNode(cache_adr, mask);
2002 transform_later(cache_adr);
2003 cache_adr = new CastX2PNode(cache_adr);
2004 transform_later(cache_adr);
2005
2006 // Prefetch
2007 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr );
2008 prefetch->set_req(0, needgc_false);
2009 transform_later(prefetch);
2010 contended_phi_rawmem = prefetch;
2011 Node *prefetch_adr;
2012 distance = step_size;
2013 for ( intx i = 1; i < lines; i++ ) {
2014 prefetch_adr = new AddPNode( cache_adr, cache_adr,
2015 _igvn.MakeConX(distance) );
2016 transform_later(prefetch_adr);
2017 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr );
2018 transform_later(prefetch);
2019 distance += step_size;
2020 contended_phi_rawmem = prefetch;
2021 }
2022 } else if( AllocatePrefetchStyle > 0 ) {
2023 // Insert a prefetch for each allocation only on the fast-path
2024 Node *prefetch_adr;
2025 Node *prefetch;
2026 // Generate several prefetch instructions.
2027 uint step_size = AllocatePrefetchStepSize;
2028 uint distance = AllocatePrefetchDistance;
2029 for ( intx i = 0; i < lines; i++ ) {
2030 prefetch_adr = new AddPNode( old_eden_top, new_eden_top,
2031 _igvn.MakeConX(distance) );
2032 transform_later(prefetch_adr);
2033 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
2034 // Do not let it float too high, since if eden_top == eden_end,
2035 // both might be null.
2036 if( i == 0 ) { // Set control for first prefetch, next follows it
2037 prefetch->init_req(0, needgc_false);
2038 }
2039 transform_later(prefetch);
2040 distance += step_size;
2041 i_o = prefetch;
2042 }
2043 }
2044 return i_o;
2045 }
2046
2047
2048 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
2049 expand_allocate_common(alloc, nullptr,
2050 OptoRuntime::new_instance_Type(),
2051 OptoRuntime::new_instance_Java(), nullptr);
2052 }
2053
2054 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
2055 Node* length = alloc->in(AllocateNode::ALength);
2056 Node* valid_length_test = alloc->in(AllocateNode::ValidLengthTest);
2057 InitializeNode* init = alloc->initialization();
2058 Node* klass_node = alloc->in(AllocateNode::KlassNode);
2059 const TypeAryKlassPtr* ary_klass_t = _igvn.type(klass_node)->isa_aryklassptr();
2060 address slow_call_address; // Address of slow call
2061 if (init != nullptr && init->is_complete_with_arraycopy() &&
2062 ary_klass_t && ary_klass_t->elem()->isa_klassptr() == nullptr) {
2063 // Don't zero type array during slow allocation in VM since
2064 // it will be initialized later by arraycopy in compiled code.
2065 slow_call_address = OptoRuntime::new_array_nozero_Java();
2066 } else {
2067 slow_call_address = OptoRuntime::new_array_Java();
2068 }
2069 expand_allocate_common(alloc, length,
2070 OptoRuntime::new_array_Type(),
2071 slow_call_address, valid_length_test);
2072 }
2073
2074 //-------------------mark_eliminated_box----------------------------------
2075 //
2076 // During EA obj may point to several objects but after few ideal graph
2077 // transformations (CCP) it may point to only one non escaping object
2078 // (but still using phi), corresponding locks and unlocks will be marked
2079 // for elimination. Later obj could be replaced with a new node (new phi)
2080 // and which does not have escape information. And later after some graph
2081 // reshape other locks and unlocks (which were not marked for elimination
2082 // before) are connected to this new obj (phi) but they still will not be
2083 // marked for elimination since new obj has no escape information.
2084 // Mark all associated (same box and obj) lock and unlock nodes for
2085 // elimination if some of them marked already.
2086 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) {
2087 if (oldbox->as_BoxLock()->is_eliminated()) {
2088 return; // This BoxLock node was processed already.
2089 }
2090 // New implementation (EliminateNestedLocks) has separate BoxLock
2091 // node for each locked region so mark all associated locks/unlocks as
2092 // eliminated even if different objects are referenced in one locked region
2093 // (for example, OSR compilation of nested loop inside locked scope).
2094 if (EliminateNestedLocks ||
2095 oldbox->as_BoxLock()->is_simple_lock_region(nullptr, obj, nullptr)) {
2096 // Box is used only in one lock region. Mark this box as eliminated.
2097 _igvn.hash_delete(oldbox);
2098 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value
2099 _igvn.hash_insert(oldbox);
2100
2101 for (uint i = 0; i < oldbox->outcnt(); i++) {
2102 Node* u = oldbox->raw_out(i);
2103 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
2104 AbstractLockNode* alock = u->as_AbstractLock();
2105 // Check lock's box since box could be referenced by Lock's debug info.
2106 if (alock->box_node() == oldbox) {
2107 // Mark eliminated all related locks and unlocks.
2108 #ifdef ASSERT
2109 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4");
2110 #endif
2111 alock->set_non_esc_obj();
2112 }
2113 }
2114 }
2115 return;
2116 }
2117
2118 // Create new "eliminated" BoxLock node and use it in monitor debug info
2119 // instead of oldbox for the same object.
2120 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
2121
2122 // Note: BoxLock node is marked eliminated only here and it is used
2123 // to indicate that all associated lock and unlock nodes are marked
2124 // for elimination.
2125 newbox->set_eliminated();
2126 transform_later(newbox);
2127
2128 // Replace old box node with new box for all users of the same object.
2129 for (uint i = 0; i < oldbox->outcnt();) {
2130 bool next_edge = true;
2131
2132 Node* u = oldbox->raw_out(i);
2133 if (u->is_AbstractLock()) {
2134 AbstractLockNode* alock = u->as_AbstractLock();
2135 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
2136 // Replace Box and mark eliminated all related locks and unlocks.
2137 #ifdef ASSERT
2138 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5");
2139 #endif
2140 alock->set_non_esc_obj();
2141 _igvn.rehash_node_delayed(alock);
2142 alock->set_box_node(newbox);
2143 next_edge = false;
2144 }
2145 }
2146 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
2147 FastLockNode* flock = u->as_FastLock();
2148 assert(flock->box_node() == oldbox, "sanity");
2149 _igvn.rehash_node_delayed(flock);
2150 flock->set_box_node(newbox);
2151 next_edge = false;
2152 }
2153
2154 // Replace old box in monitor debug info.
2155 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
2156 SafePointNode* sfn = u->as_SafePoint();
2157 JVMState* youngest_jvms = sfn->jvms();
2158 int max_depth = youngest_jvms->depth();
2159 for (int depth = 1; depth <= max_depth; depth++) {
2160 JVMState* jvms = youngest_jvms->of_depth(depth);
2161 int num_mon = jvms->nof_monitors();
2162 // Loop over monitors
2163 for (int idx = 0; idx < num_mon; idx++) {
2164 Node* obj_node = sfn->monitor_obj(jvms, idx);
2165 Node* box_node = sfn->monitor_box(jvms, idx);
2166 if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
2167 int j = jvms->monitor_box_offset(idx);
2168 _igvn.replace_input_of(u, j, newbox);
2169 next_edge = false;
2170 }
2171 }
2172 }
2173 }
2174 if (next_edge) i++;
2175 }
2176 }
2177
2178 //-----------------------mark_eliminated_locking_nodes-----------------------
2179 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
2180 if (EliminateNestedLocks) {
2181 if (alock->is_nested()) {
2182 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
2183 return;
2184 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
2185 // Only Lock node has JVMState needed here.
2186 // Not that preceding claim is documented anywhere else.
2187 if (alock->jvms() != nullptr) {
2188 if (alock->as_Lock()->is_nested_lock_region()) {
2189 // Mark eliminated related nested locks and unlocks.
2190 Node* obj = alock->obj_node();
2191 BoxLockNode* box_node = alock->box_node()->as_BoxLock();
2192 assert(!box_node->is_eliminated(), "should not be marked yet");
2193 // Note: BoxLock node is marked eliminated only here
2194 // and it is used to indicate that all associated lock
2195 // and unlock nodes are marked for elimination.
2196 box_node->set_eliminated(); // Box's hash is always NO_HASH here
2197 for (uint i = 0; i < box_node->outcnt(); i++) {
2198 Node* u = box_node->raw_out(i);
2199 if (u->is_AbstractLock()) {
2200 alock = u->as_AbstractLock();
2201 if (alock->box_node() == box_node) {
2202 // Verify that this Box is referenced only by related locks.
2203 assert(alock->obj_node()->eqv_uncast(obj), "");
2204 // Mark all related locks and unlocks.
2205 #ifdef ASSERT
2206 alock->log_lock_optimization(C, "eliminate_lock_set_nested");
2207 #endif
2208 alock->set_nested();
2209 }
2210 }
2211 }
2212 } else {
2213 #ifdef ASSERT
2214 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region");
2215 if (C->log() != nullptr)
2216 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output
2217 #endif
2218 }
2219 }
2220 return;
2221 }
2222 // Process locks for non escaping object
2223 assert(alock->is_non_esc_obj(), "");
2224 } // EliminateNestedLocks
2225
2226 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2227 // Look for all locks of this object and mark them and
2228 // corresponding BoxLock nodes as eliminated.
2229 Node* obj = alock->obj_node();
2230 for (uint j = 0; j < obj->outcnt(); j++) {
2231 Node* o = obj->raw_out(j);
2232 if (o->is_AbstractLock() &&
2233 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2234 alock = o->as_AbstractLock();
2235 Node* box = alock->box_node();
2236 // Replace old box node with new eliminated box for all users
2237 // of the same object and mark related locks as eliminated.
2238 mark_eliminated_box(box, obj);
2239 }
2240 }
2241 }
2242 }
2243
2244 void PhaseMacroExpand::inline_type_guard(Node** ctrl, LockNode* lock) {
2245 Node* obj = lock->obj_node();
2246 const TypePtr* obj_type = _igvn.type(obj)->make_ptr();
2247 if (!obj_type->can_be_inline_type()) {
2248 return;
2249 }
2250 Node* mark = make_load(*ctrl, lock->memory(), obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2251 Node* value_mask = _igvn.MakeConX(markWord::inline_type_pattern);
2252 Node* is_value = _igvn.transform(new AndXNode(mark, value_mask));
2253 Node* cmp = _igvn.transform(new CmpXNode(is_value, value_mask));
2254 Node* bol = _igvn.transform(new BoolNode(cmp, BoolTest::eq));
2255 Node* unc_ctrl = generate_slow_guard(ctrl, bol, nullptr);
2256
2257 int trap_request = Deoptimization::make_trap_request(Deoptimization::Reason_class_check, Deoptimization::Action_none);
2258 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2259 const TypePtr* no_memory_effects = nullptr;
2260 CallNode* unc = new CallStaticJavaNode(OptoRuntime::uncommon_trap_Type(), call_addr, "uncommon_trap",
2261 no_memory_effects);
2262 unc->init_req(TypeFunc::Control, unc_ctrl);
2263 unc->init_req(TypeFunc::I_O, lock->i_o());
2264 unc->init_req(TypeFunc::Memory, lock->memory());
2265 unc->init_req(TypeFunc::FramePtr, lock->in(TypeFunc::FramePtr));
2266 unc->init_req(TypeFunc::ReturnAdr, lock->in(TypeFunc::ReturnAdr));
2267 unc->init_req(TypeFunc::Parms+0, _igvn.intcon(trap_request));
2268 unc->set_cnt(PROB_UNLIKELY_MAG(4));
2269 unc->copy_call_debug_info(&_igvn, lock);
2270
2271 assert(unc->peek_monitor_box() == lock->box_node(), "wrong monitor");
2272 assert((obj_type->is_inlinetypeptr() && unc->peek_monitor_obj()->is_SafePointScalarObject()) ||
2273 (obj->is_InlineType() && obj->in(1) == unc->peek_monitor_obj()) ||
2274 (obj == unc->peek_monitor_obj()), "wrong monitor");
2275
2276 // pop monitor and push obj back on stack: we trap before the monitorenter
2277 unc->pop_monitor();
2278 unc->grow_stack(unc->jvms(), 1);
2279 unc->set_stack(unc->jvms(), unc->jvms()->stk_size()-1, obj);
2280 _igvn.register_new_node_with_optimizer(unc);
2281
2282 unc_ctrl = _igvn.transform(new ProjNode(unc, TypeFunc::Control));
2283 Node* halt = _igvn.transform(new HaltNode(unc_ctrl, lock->in(TypeFunc::FramePtr), "monitor enter on inline type"));
2284 _igvn.add_input_to(C->root(), halt);
2285 }
2286
2287 // we have determined that this lock/unlock can be eliminated, we simply
2288 // eliminate the node without expanding it.
2289 //
2290 // Note: The membar's associated with the lock/unlock are currently not
2291 // eliminated. This should be investigated as a future enhancement.
2292 //
2293 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2294
2295 if (!alock->is_eliminated()) {
2296 return false;
2297 }
2298 #ifdef ASSERT
2299 if (!alock->is_coarsened()) {
2300 // Check that new "eliminated" BoxLock node is created.
2301 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2302 assert(oldbox->is_eliminated(), "should be done already");
2303 }
2304 #endif
2305
2306 alock->log_lock_optimization(C, "eliminate_lock");
2307
2308 #ifndef PRODUCT
2309 if (PrintEliminateLocks) {
2310 tty->print_cr("++++ Eliminated: %d %s '%s'", alock->_idx, (alock->is_Lock() ? "Lock" : "Unlock"), alock->kind_as_string());
2311 }
2312 #endif
2313
2314 Node* mem = alock->in(TypeFunc::Memory);
2315 Node* ctrl = alock->in(TypeFunc::Control);
2316 guarantee(ctrl != nullptr, "missing control projection, cannot replace_node() with null");
2317
2318 _callprojs = alock->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2319 // There are 2 projections from the lock. The lock node will
2320 // be deleted when its last use is subsumed below.
2321 assert(alock->outcnt() == 2 &&
2322 _callprojs->fallthrough_proj != nullptr &&
2323 _callprojs->fallthrough_memproj != nullptr,
2324 "Unexpected projections from Lock/Unlock");
2325
2326 Node* fallthroughproj = _callprojs->fallthrough_proj;
2327 Node* memproj_fallthrough = _callprojs->fallthrough_memproj;
2328
2329 // The memory projection from a lock/unlock is RawMem
2330 // The input to a Lock is merged memory, so extract its RawMem input
2331 // (unless the MergeMem has been optimized away.)
2332 if (alock->is_Lock()) {
2333 // Deoptimize and re-execute if object is an inline type
2334 inline_type_guard(&ctrl, alock->as_Lock());
2335
2336 // Search for MemBarAcquireLock node and delete it also.
2337 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2338 assert(membar != nullptr && membar->Opcode() == Op_MemBarAcquireLock, "");
2339 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2340 Node* memproj = membar->proj_out(TypeFunc::Memory);
2341 _igvn.replace_node(ctrlproj, fallthroughproj);
2342 _igvn.replace_node(memproj, memproj_fallthrough);
2343
2344 // Delete FastLock node also if this Lock node is unique user
2345 // (a loop peeling may clone a Lock node).
2346 Node* flock = alock->as_Lock()->fastlock_node();
2347 if (flock->outcnt() == 1) {
2348 assert(flock->unique_out() == alock, "sanity");
2349 _igvn.replace_node(flock, top());
2350 }
2351 }
2352
2353 // Search for MemBarReleaseLock node and delete it also.
2354 if (alock->is_Unlock() && ctrl->is_Proj() && ctrl->in(0)->is_MemBar()) {
2355 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2356 assert(membar->Opcode() == Op_MemBarReleaseLock &&
2357 mem->is_Proj() && membar == mem->in(0), "");
2358 _igvn.replace_node(fallthroughproj, ctrl);
2359 _igvn.replace_node(memproj_fallthrough, mem);
2360 fallthroughproj = ctrl;
2361 memproj_fallthrough = mem;
2362 ctrl = membar->in(TypeFunc::Control);
2363 mem = membar->in(TypeFunc::Memory);
2364 }
2365
2366 _igvn.replace_node(fallthroughproj, ctrl);
2367 _igvn.replace_node(memproj_fallthrough, mem);
2368 return true;
2369 }
2370
2371
2372 //------------------------------expand_lock_node----------------------
2373 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2374
2375 Node* ctrl = lock->in(TypeFunc::Control);
2376 Node* mem = lock->in(TypeFunc::Memory);
2377 Node* obj = lock->obj_node();
2378 Node* box = lock->box_node();
2379 Node* flock = lock->fastlock_node();
2380
2381 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2382
2383 // Make the merge point
2384 Node *region;
2385 Node *mem_phi;
2386 Node *slow_path;
2387
2388 region = new RegionNode(3);
2389 // create a Phi for the memory state
2390 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2391
2392 // Optimize test; set region slot 2
2393 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2394 mem_phi->init_req(2, mem);
2395
2396 // Deoptimize and re-execute if object is an inline type
2397 inline_type_guard(&slow_path, lock);
2398
2399 // Make slow path call
2400 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(),
2401 OptoRuntime::complete_monitor_locking_Java(), nullptr, slow_path,
2402 obj, box, nullptr);
2403
2404 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2405
2406 // Slow path can only throw asynchronous exceptions, which are always
2407 // de-opted. So the compiler thinks the slow-call can never throw an
2408 // exception. If it DOES throw an exception we would need the debug
2409 // info removed first (since if it throws there is no monitor).
2410 assert(_callprojs->fallthrough_ioproj == nullptr && _callprojs->catchall_ioproj == nullptr &&
2411 _callprojs->catchall_memproj == nullptr && _callprojs->catchall_catchproj == nullptr, "Unexpected projection from Lock");
2412
2413 // Capture slow path
2414 // disconnect fall-through projection from call and create a new one
2415 // hook up users of fall-through projection to region
2416 Node *slow_ctrl = _callprojs->fallthrough_proj->clone();
2417 transform_later(slow_ctrl);
2418 _igvn.hash_delete(_callprojs->fallthrough_proj);
2419 _callprojs->fallthrough_proj->disconnect_inputs(C);
2420 region->init_req(1, slow_ctrl);
2421 // region inputs are now complete
2422 transform_later(region);
2423 _igvn.replace_node(_callprojs->fallthrough_proj, region);
2424
2425 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2426
2427 mem_phi->init_req(1, memproj);
2428
2429 transform_later(mem_phi);
2430
2431 _igvn.replace_node(_callprojs->fallthrough_memproj, mem_phi);
2432 }
2433
2434 //------------------------------expand_unlock_node----------------------
2435 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2436
2437 Node* ctrl = unlock->in(TypeFunc::Control);
2438 Node* mem = unlock->in(TypeFunc::Memory);
2439 Node* obj = unlock->obj_node();
2440 Node* box = unlock->box_node();
2441
2442 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2443
2444 // No need for a null check on unlock
2445
2446 // Make the merge point
2447 Node *region;
2448 Node *mem_phi;
2449
2450 region = new RegionNode(3);
2451 // create a Phi for the memory state
2452 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2453
2454 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2455 funlock = transform_later( funlock )->as_FastUnlock();
2456 // Optimize test; set region slot 2
2457 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2458 Node *thread = transform_later(new ThreadLocalNode());
2459
2460 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2461 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2462 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2463
2464 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2465 assert(_callprojs->fallthrough_ioproj == nullptr && _callprojs->catchall_ioproj == nullptr &&
2466 _callprojs->catchall_memproj == nullptr && _callprojs->catchall_catchproj == nullptr, "Unexpected projection from Lock");
2467
2468 // No exceptions for unlocking
2469 // Capture slow path
2470 // disconnect fall-through projection from call and create a new one
2471 // hook up users of fall-through projection to region
2472 Node *slow_ctrl = _callprojs->fallthrough_proj->clone();
2473 transform_later(slow_ctrl);
2474 _igvn.hash_delete(_callprojs->fallthrough_proj);
2475 _callprojs->fallthrough_proj->disconnect_inputs(C);
2476 region->init_req(1, slow_ctrl);
2477 // region inputs are now complete
2478 transform_later(region);
2479 _igvn.replace_node(_callprojs->fallthrough_proj, region);
2480
2481 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2482 mem_phi->init_req(1, memproj );
2483 mem_phi->init_req(2, mem);
2484 transform_later(mem_phi);
2485
2486 _igvn.replace_node(_callprojs->fallthrough_memproj, mem_phi);
2487 }
2488
2489 // An inline type might be returned from the call but we don't know its
2490 // type. Either we get a buffered inline type (and nothing needs to be done)
2491 // or one of the values being returned is the klass of the inline type
2492 // and we need to allocate an inline type instance of that type and
2493 // initialize it with other values being returned. In that case, we
2494 // first try a fast path allocation and initialize the value with the
2495 // inline klass's pack handler or we fall back to a runtime call.
2496 void PhaseMacroExpand::expand_mh_intrinsic_return(CallStaticJavaNode* call) {
2497 assert(call->method()->is_method_handle_intrinsic(), "must be a method handle intrinsic call");
2498 Node* ret = call->proj_out_or_null(TypeFunc::Parms);
2499 if (ret == nullptr) {
2500 return;
2501 }
2502 const TypeFunc* tf = call->_tf;
2503 const TypeTuple* domain = OptoRuntime::store_inline_type_fields_Type()->domain_cc();
2504 const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain);
2505 call->_tf = new_tf;
2506 // Make sure the change of type is applied before projections are processed by igvn
2507 _igvn.set_type(call, call->Value(&_igvn));
2508 _igvn.set_type(ret, ret->Value(&_igvn));
2509
2510 // Before any new projection is added:
2511 CallProjections* projs = call->extract_projections(true, true);
2512
2513 // Create temporary hook nodes that will be replaced below.
2514 // Add an input to prevent hook nodes from being dead.
2515 Node* ctl = new Node(call);
2516 Node* mem = new Node(ctl);
2517 Node* io = new Node(ctl);
2518 Node* ex_ctl = new Node(ctl);
2519 Node* ex_mem = new Node(ctl);
2520 Node* ex_io = new Node(ctl);
2521 Node* res = new Node(ctl);
2522
2523 // Allocate a new buffered inline type only if a new one is not returned
2524 Node* cast = transform_later(new CastP2XNode(ctl, res));
2525 Node* mask = MakeConX(0x1);
2526 Node* masked = transform_later(new AndXNode(cast, mask));
2527 Node* cmp = transform_later(new CmpXNode(masked, mask));
2528 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
2529 IfNode* allocation_iff = new IfNode(ctl, bol, PROB_MAX, COUNT_UNKNOWN);
2530 transform_later(allocation_iff);
2531 Node* allocation_ctl = transform_later(new IfTrueNode(allocation_iff));
2532 Node* no_allocation_ctl = transform_later(new IfFalseNode(allocation_iff));
2533 Node* no_allocation_res = transform_later(new CheckCastPPNode(no_allocation_ctl, res, TypeInstPtr::BOTTOM));
2534
2535 // Try to allocate a new buffered inline instance either from TLAB or eden space
2536 Node* needgc_ctrl = nullptr; // needgc means slowcase, i.e. allocation failed
2537 CallLeafNoFPNode* handler_call;
2538 const bool alloc_in_place = UseTLAB;
2539 if (alloc_in_place) {
2540 Node* fast_oop_ctrl = nullptr;
2541 Node* fast_oop_rawmem = nullptr;
2542 Node* mask2 = MakeConX(-2);
2543 Node* masked2 = transform_later(new AndXNode(cast, mask2));
2544 Node* rawklassptr = transform_later(new CastX2PNode(masked2));
2545 Node* klass_node = transform_later(new CheckCastPPNode(allocation_ctl, rawklassptr, TypeInstKlassPtr::OBJECT_OR_NULL));
2546 Node* layout_val = make_load(nullptr, mem, klass_node, in_bytes(Klass::layout_helper_offset()), TypeInt::INT, T_INT);
2547 Node* size_in_bytes = ConvI2X(layout_val);
2548 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2549 Node* fast_oop = bs->obj_allocate(this, mem, allocation_ctl, size_in_bytes, io, needgc_ctrl,
2550 fast_oop_ctrl, fast_oop_rawmem,
2551 AllocateInstancePrefetchLines);
2552 // Allocation succeed, initialize buffered inline instance header firstly,
2553 // and then initialize its fields with an inline class specific handler
2554 Node* mark_node = makecon(TypeRawPtr::make((address)markWord::inline_type_prototype().value()));
2555 fast_oop_rawmem = make_store(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
2556 fast_oop_rawmem = make_store(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
2557 if (UseCompressedClassPointers) {
2558 fast_oop_rawmem = make_store(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::klass_gap_offset_in_bytes(), intcon(0), T_INT);
2559 }
2560 Node* fixed_block = make_load(fast_oop_ctrl, fast_oop_rawmem, klass_node, in_bytes(InstanceKlass::adr_inlineklass_fixed_block_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2561 Node* pack_handler = make_load(fast_oop_ctrl, fast_oop_rawmem, fixed_block, in_bytes(InlineKlass::pack_handler_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2562 handler_call = new CallLeafNoFPNode(OptoRuntime::pack_inline_type_Type(),
2563 nullptr,
2564 "pack handler",
2565 TypeRawPtr::BOTTOM);
2566 handler_call->init_req(TypeFunc::Control, fast_oop_ctrl);
2567 handler_call->init_req(TypeFunc::Memory, fast_oop_rawmem);
2568 handler_call->init_req(TypeFunc::I_O, top());
2569 handler_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2570 handler_call->init_req(TypeFunc::ReturnAdr, top());
2571 handler_call->init_req(TypeFunc::Parms, pack_handler);
2572 handler_call->init_req(TypeFunc::Parms+1, fast_oop);
2573 } else {
2574 needgc_ctrl = allocation_ctl;
2575 }
2576
2577 // Allocation failed, fall back to a runtime call
2578 CallStaticJavaNode* slow_call = new CallStaticJavaNode(OptoRuntime::store_inline_type_fields_Type(),
2579 StubRoutines::store_inline_type_fields_to_buf(),
2580 "store_inline_type_fields",
2581 TypePtr::BOTTOM);
2582 slow_call->init_req(TypeFunc::Control, needgc_ctrl);
2583 slow_call->init_req(TypeFunc::Memory, mem);
2584 slow_call->init_req(TypeFunc::I_O, io);
2585 slow_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2586 slow_call->init_req(TypeFunc::ReturnAdr, call->in(TypeFunc::ReturnAdr));
2587 slow_call->init_req(TypeFunc::Parms, res);
2588
2589 Node* slow_ctl = transform_later(new ProjNode(slow_call, TypeFunc::Control));
2590 Node* slow_mem = transform_later(new ProjNode(slow_call, TypeFunc::Memory));
2591 Node* slow_io = transform_later(new ProjNode(slow_call, TypeFunc::I_O));
2592 Node* slow_res = transform_later(new ProjNode(slow_call, TypeFunc::Parms));
2593 Node* slow_catc = transform_later(new CatchNode(slow_ctl, slow_io, 2));
2594 Node* slow_norm = transform_later(new CatchProjNode(slow_catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci));
2595 Node* slow_excp = transform_later(new CatchProjNode(slow_catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci));
2596
2597 Node* ex_r = new RegionNode(3);
2598 Node* ex_mem_phi = new PhiNode(ex_r, Type::MEMORY, TypePtr::BOTTOM);
2599 Node* ex_io_phi = new PhiNode(ex_r, Type::ABIO);
2600 ex_r->init_req(1, slow_excp);
2601 ex_mem_phi->init_req(1, slow_mem);
2602 ex_io_phi->init_req(1, slow_io);
2603 ex_r->init_req(2, ex_ctl);
2604 ex_mem_phi->init_req(2, ex_mem);
2605 ex_io_phi->init_req(2, ex_io);
2606 transform_later(ex_r);
2607 transform_later(ex_mem_phi);
2608 transform_later(ex_io_phi);
2609
2610 // We don't know how many values are returned. This assumes the
2611 // worst case, that all available registers are used.
2612 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2613 if (domain->field_at(i) == Type::HALF) {
2614 slow_call->init_req(i, top());
2615 if (alloc_in_place) {
2616 handler_call->init_req(i+1, top());
2617 }
2618 continue;
2619 }
2620 Node* proj = transform_later(new ProjNode(call, i));
2621 slow_call->init_req(i, proj);
2622 if (alloc_in_place) {
2623 handler_call->init_req(i+1, proj);
2624 }
2625 }
2626 // We can safepoint at that new call
2627 slow_call->copy_call_debug_info(&_igvn, call);
2628 transform_later(slow_call);
2629 if (alloc_in_place) {
2630 transform_later(handler_call);
2631 }
2632
2633 Node* fast_ctl = nullptr;
2634 Node* fast_res = nullptr;
2635 MergeMemNode* fast_mem = nullptr;
2636 if (alloc_in_place) {
2637 fast_ctl = transform_later(new ProjNode(handler_call, TypeFunc::Control));
2638 Node* rawmem = transform_later(new ProjNode(handler_call, TypeFunc::Memory));
2639 fast_res = transform_later(new ProjNode(handler_call, TypeFunc::Parms));
2640 fast_mem = MergeMemNode::make(mem);
2641 fast_mem->set_memory_at(Compile::AliasIdxRaw, rawmem);
2642 transform_later(fast_mem);
2643 }
2644
2645 Node* r = new RegionNode(alloc_in_place ? 4 : 3);
2646 Node* mem_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2647 Node* io_phi = new PhiNode(r, Type::ABIO);
2648 Node* res_phi = new PhiNode(r, TypeInstPtr::BOTTOM);
2649 r->init_req(1, no_allocation_ctl);
2650 mem_phi->init_req(1, mem);
2651 io_phi->init_req(1, io);
2652 res_phi->init_req(1, no_allocation_res);
2653 r->init_req(2, slow_norm);
2654 mem_phi->init_req(2, slow_mem);
2655 io_phi->init_req(2, slow_io);
2656 res_phi->init_req(2, slow_res);
2657 if (alloc_in_place) {
2658 r->init_req(3, fast_ctl);
2659 mem_phi->init_req(3, fast_mem);
2660 io_phi->init_req(3, io);
2661 res_phi->init_req(3, fast_res);
2662 }
2663 transform_later(r);
2664 transform_later(mem_phi);
2665 transform_later(io_phi);
2666 transform_later(res_phi);
2667
2668 // Do not let stores that initialize this buffer be reordered with a subsequent
2669 // store that would make this buffer accessible by other threads.
2670 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
2671 transform_later(mb);
2672 mb->init_req(TypeFunc::Memory, mem_phi);
2673 mb->init_req(TypeFunc::Control, r);
2674 r = new ProjNode(mb, TypeFunc::Control);
2675 transform_later(r);
2676 mem_phi = new ProjNode(mb, TypeFunc::Memory);
2677 transform_later(mem_phi);
2678
2679 assert(projs->nb_resproj == 1, "unexpected number of results");
2680 _igvn.replace_in_uses(projs->fallthrough_catchproj, r);
2681 _igvn.replace_in_uses(projs->fallthrough_memproj, mem_phi);
2682 _igvn.replace_in_uses(projs->fallthrough_ioproj, io_phi);
2683 _igvn.replace_in_uses(projs->resproj[0], res_phi);
2684 _igvn.replace_in_uses(projs->catchall_catchproj, ex_r);
2685 _igvn.replace_in_uses(projs->catchall_memproj, ex_mem_phi);
2686 _igvn.replace_in_uses(projs->catchall_ioproj, ex_io_phi);
2687 // The CatchNode should not use the ex_io_phi. Re-connect it to the catchall_ioproj.
2688 Node* cn = projs->fallthrough_catchproj->in(0);
2689 _igvn.replace_input_of(cn, 1, projs->catchall_ioproj);
2690
2691 _igvn.replace_node(ctl, projs->fallthrough_catchproj);
2692 _igvn.replace_node(mem, projs->fallthrough_memproj);
2693 _igvn.replace_node(io, projs->fallthrough_ioproj);
2694 _igvn.replace_node(res, projs->resproj[0]);
2695 _igvn.replace_node(ex_ctl, projs->catchall_catchproj);
2696 _igvn.replace_node(ex_mem, projs->catchall_memproj);
2697 _igvn.replace_node(ex_io, projs->catchall_ioproj);
2698 }
2699
2700 void PhaseMacroExpand::expand_subtypecheck_node(SubTypeCheckNode *check) {
2701 assert(check->in(SubTypeCheckNode::Control) == nullptr, "should be pinned");
2702 Node* bol = check->unique_out();
2703 Node* obj_or_subklass = check->in(SubTypeCheckNode::ObjOrSubKlass);
2704 Node* superklass = check->in(SubTypeCheckNode::SuperKlass);
2705 assert(bol->is_Bool() && bol->as_Bool()->_test._test == BoolTest::ne, "unexpected bool node");
2706
2707 for (DUIterator_Last imin, i = bol->last_outs(imin); i >= imin; --i) {
2708 Node* iff = bol->last_out(i);
2709 assert(iff->is_If(), "where's the if?");
2710
2711 if (iff->in(0)->is_top()) {
2712 _igvn.replace_input_of(iff, 1, C->top());
2713 continue;
2714 }
2715
2716 Node* iftrue = iff->as_If()->proj_out(1);
2717 Node* iffalse = iff->as_If()->proj_out(0);
2718 Node* ctrl = iff->in(0);
2719
2720 Node* subklass = nullptr;
2721 if (_igvn.type(obj_or_subklass)->isa_klassptr()) {
2722 subklass = obj_or_subklass;
2723 } else {
2724 Node* k_adr = basic_plus_adr(obj_or_subklass, oopDesc::klass_offset_in_bytes());
2725 subklass = _igvn.transform(LoadKlassNode::make(_igvn, nullptr, C->immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
2726 }
2727
2728 Node* not_subtype_ctrl = Phase::gen_subtype_check(subklass, superklass, &ctrl, nullptr, _igvn);
2729
2730 _igvn.replace_input_of(iff, 0, C->top());
2731 _igvn.replace_node(iftrue, not_subtype_ctrl);
2732 _igvn.replace_node(iffalse, ctrl);
2733 }
2734 _igvn.replace_node(check, C->top());
2735 }
2736
2737 // FlatArrayCheckNode (array1 array2 ...) is expanded into:
2738 //
2739 // long mark = array1.mark | array2.mark | ...;
2740 // long locked_bit = markWord::unlocked_value & array1.mark & array2.mark & ...;
2741 // if (locked_bit == 0) {
2742 // // One array is locked, load prototype header from the klass
2743 // mark = array1.klass.proto | array2.klass.proto | ...
2744 // }
2745 // if ((mark & markWord::flat_array_bit_in_place) == 0) {
2746 // ...
2747 // }
2748 void PhaseMacroExpand::expand_flatarraycheck_node(FlatArrayCheckNode* check) {
2749 bool array_inputs = _igvn.type(check->in(FlatArrayCheckNode::ArrayOrKlass))->isa_oopptr() != nullptr;
2750 if (UseArrayMarkWordCheck && array_inputs) {
2751 Node* mark = MakeConX(0);
2752 Node* locked_bit = MakeConX(markWord::unlocked_value);
2753 Node* mem = check->in(FlatArrayCheckNode::Memory);
2754 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
2755 Node* ary = check->in(i);
2756 const TypeOopPtr* t = _igvn.type(ary)->isa_oopptr();
2757 assert(t != nullptr, "Mixing array and klass inputs");
2758 assert(!t->is_flat() && !t->is_not_flat(), "Should have been optimized out");
2759 Node* mark_adr = basic_plus_adr(ary, oopDesc::mark_offset_in_bytes());
2760 Node* mark_load = _igvn.transform(LoadNode::make(_igvn, nullptr, mem, mark_adr, mark_adr->bottom_type()->is_ptr(), TypeX_X, TypeX_X->basic_type(), MemNode::unordered));
2761 mark = _igvn.transform(new OrXNode(mark, mark_load));
2762 locked_bit = _igvn.transform(new AndXNode(locked_bit, mark_load));
2763 }
2764 assert(!mark->is_Con(), "Should have been optimized out");
2765 Node* cmp = _igvn.transform(new CmpXNode(locked_bit, MakeConX(0)));
2766 Node* is_unlocked = _igvn.transform(new BoolNode(cmp, BoolTest::ne));
2767
2768 // BoolNode might be shared, replace each if user
2769 Node* old_bol = check->unique_out();
2770 assert(old_bol->is_Bool() && old_bol->as_Bool()->_test._test == BoolTest::ne, "unexpected condition");
2771 for (DUIterator_Last imin, i = old_bol->last_outs(imin); i >= imin; --i) {
2772 IfNode* old_iff = old_bol->last_out(i)->as_If();
2773 Node* ctrl = old_iff->in(0);
2774 RegionNode* region = new RegionNode(3);
2775 Node* mark_phi = new PhiNode(region, TypeX_X);
2776
2777 // Check if array is unlocked
2778 IfNode* iff = _igvn.transform(new IfNode(ctrl, is_unlocked, PROB_MAX, COUNT_UNKNOWN))->as_If();
2779
2780 // Unlocked: Use bits from mark word
2781 region->init_req(1, _igvn.transform(new IfTrueNode(iff)));
2782 mark_phi->init_req(1, mark);
2783
2784 // Locked: Load prototype header from klass
2785 ctrl = _igvn.transform(new IfFalseNode(iff));
2786 Node* proto = MakeConX(0);
2787 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
2788 Node* ary = check->in(i);
2789 // Make loads control dependent to make sure they are only executed if array is locked
2790 Node* klass_adr = basic_plus_adr(ary, oopDesc::klass_offset_in_bytes());
2791 Node* klass = _igvn.transform(LoadKlassNode::make(_igvn, ctrl, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
2792 Node* proto_adr = basic_plus_adr(klass, in_bytes(Klass::prototype_header_offset()));
2793 Node* proto_load = _igvn.transform(LoadNode::make(_igvn, ctrl, C->immutable_memory(), proto_adr, proto_adr->bottom_type()->is_ptr(), TypeX_X, TypeX_X->basic_type(), MemNode::unordered));
2794 proto = _igvn.transform(new OrXNode(proto, proto_load));
2795 }
2796 region->init_req(2, ctrl);
2797 mark_phi->init_req(2, proto);
2798
2799 // Check if flat array bits are set
2800 Node* mask = MakeConX(markWord::flat_array_bit_in_place);
2801 Node* masked = _igvn.transform(new AndXNode(_igvn.transform(mark_phi), mask));
2802 cmp = _igvn.transform(new CmpXNode(masked, MakeConX(0)));
2803 Node* is_not_flat = _igvn.transform(new BoolNode(cmp, BoolTest::eq));
2804
2805 ctrl = _igvn.transform(region);
2806 iff = _igvn.transform(new IfNode(ctrl, is_not_flat, PROB_MAX, COUNT_UNKNOWN))->as_If();
2807 _igvn.replace_node(old_iff, iff);
2808 }
2809 _igvn.replace_node(check, C->top());
2810 } else {
2811 // Fall back to layout helper check
2812 Node* lhs = intcon(0);
2813 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
2814 Node* array_or_klass = check->in(i);
2815 Node* klass = nullptr;
2816 const TypePtr* t = _igvn.type(array_or_klass)->is_ptr();
2817 assert(!t->is_flat() && !t->is_not_flat(), "Should have been optimized out");
2818 if (t->isa_oopptr() != nullptr) {
2819 Node* klass_adr = basic_plus_adr(array_or_klass, oopDesc::klass_offset_in_bytes());
2820 klass = transform_later(LoadKlassNode::make(_igvn, nullptr, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
2821 } else {
2822 assert(t->isa_aryklassptr(), "Unexpected input type");
2823 klass = array_or_klass;
2824 }
2825 Node* lh_addr = basic_plus_adr(klass, in_bytes(Klass::layout_helper_offset()));
2826 Node* lh_val = _igvn.transform(LoadNode::make(_igvn, nullptr, C->immutable_memory(), lh_addr, lh_addr->bottom_type()->is_ptr(), TypeInt::INT, T_INT, MemNode::unordered));
2827 lhs = _igvn.transform(new OrINode(lhs, lh_val));
2828 }
2829 Node* masked = transform_later(new AndINode(lhs, intcon(Klass::_lh_array_tag_flat_value_bit_inplace)));
2830 Node* cmp = transform_later(new CmpINode(masked, intcon(0)));
2831 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
2832 Node* old_bol = check->unique_out();
2833 _igvn.replace_node(old_bol, bol);
2834 _igvn.replace_node(check, C->top());
2835 }
2836 }
2837
2838 //---------------------------eliminate_macro_nodes----------------------
2839 // Eliminate scalar replaced allocations and associated locks.
2840 void PhaseMacroExpand::eliminate_macro_nodes() {
2841 if (C->macro_count() == 0)
2842 return;
2843 NOT_PRODUCT(int membar_before = count_MemBar(C);)
2844
2845 // Before elimination may re-mark (change to Nested or NonEscObj)
2846 // all associated (same box and obj) lock and unlock nodes.
2847 int cnt = C->macro_count();
2848 for (int i=0; i < cnt; i++) {
2849 Node *n = C->macro_node(i);
2850 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2851 mark_eliminated_locking_nodes(n->as_AbstractLock());
2852 }
2853 }
2854 // Re-marking may break consistency of Coarsened locks.
2855 if (!C->coarsened_locks_consistent()) {
2856 return; // recompile without Coarsened locks if broken
2857 }
2858
2859 // First, attempt to eliminate locks
2860 bool progress = true;
2861 while (progress) {
2862 progress = false;
2863 for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once
2864 Node* n = C->macro_node(i - 1);
2865 bool success = false;
2866 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2867 if (n->is_AbstractLock()) {
2868 success = eliminate_locking_node(n->as_AbstractLock());
2869 #ifndef PRODUCT
2870 if (success && PrintOptoStatistics) {
2871 Atomic::inc(&PhaseMacroExpand::_monitor_objects_removed_counter);
2872 }
2873 #endif
2874 }
2875 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2876 progress = progress || success;
2877 }
2878 }
2879 // Next, attempt to eliminate allocations
2880 _has_locks = false;
2881 progress = true;
2882 while (progress) {
2883 progress = false;
2884 for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once
2885 Node* n = C->macro_node(i - 1);
2886 bool success = false;
2887 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2888 switch (n->class_id()) {
2889 case Node::Class_Allocate:
2890 case Node::Class_AllocateArray:
2891 success = eliminate_allocate_node(n->as_Allocate());
2892 #ifndef PRODUCT
2893 if (success && PrintOptoStatistics) {
2894 Atomic::inc(&PhaseMacroExpand::_objs_scalar_replaced_counter);
2895 }
2896 #endif
2897 break;
2898 case Node::Class_CallStaticJava: {
2899 CallStaticJavaNode* call = n->as_CallStaticJava();
2900 if (!call->method()->is_method_handle_intrinsic()) {
2901 success = eliminate_boxing_node(n->as_CallStaticJava());
2902 }
2903 break;
2904 }
2905 case Node::Class_Lock:
2906 case Node::Class_Unlock:
2907 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2908 _has_locks = true;
2909 break;
2910 case Node::Class_ArrayCopy:
2911 break;
2912 case Node::Class_OuterStripMinedLoop:
2913 break;
2914 case Node::Class_SubTypeCheck:
2915 break;
2916 case Node::Class_Opaque1:
2917 break;
2918 case Node::Class_FlatArrayCheck:
2919 break;
2920 default:
2921 assert(n->Opcode() == Op_LoopLimit ||
2922 n->Opcode() == Op_Opaque3 ||
2923 n->Opcode() == Op_Opaque4 ||
2924 n->Opcode() == Op_MaxL ||
2925 n->Opcode() == Op_MinL ||
2926 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n),
2927 "unknown node type in macro list");
2928 }
2929 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2930 progress = progress || success;
2931 }
2932 }
2933 #ifndef PRODUCT
2934 if (PrintOptoStatistics) {
2935 int membar_after = count_MemBar(C);
2936 Atomic::add(&PhaseMacroExpand::_memory_barriers_removed_counter, membar_before - membar_after);
2937 }
2938 #endif
2939 }
2940
2941 //------------------------------expand_macro_nodes----------------------
2942 // Returns true if a failure occurred.
2943 bool PhaseMacroExpand::expand_macro_nodes() {
2944 // Last attempt to eliminate macro nodes.
2945 eliminate_macro_nodes();
2946 if (C->failing()) return true;
2947
2948 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2949 bool progress = true;
2950 while (progress) {
2951 progress = false;
2952 for (int i = C->macro_count(); i > 0; i--) {
2953 Node* n = C->macro_node(i-1);
2954 bool success = false;
2955 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2956 if (n->Opcode() == Op_LoopLimit) {
2957 // Remove it from macro list and put on IGVN worklist to optimize.
2958 C->remove_macro_node(n);
2959 _igvn._worklist.push(n);
2960 success = true;
2961 } else if (n->Opcode() == Op_CallStaticJava) {
2962 CallStaticJavaNode* call = n->as_CallStaticJava();
2963 if (!call->method()->is_method_handle_intrinsic()) {
2964 // Remove it from macro list and put on IGVN worklist to optimize.
2965 C->remove_macro_node(n);
2966 _igvn._worklist.push(n);
2967 success = true;
2968 }
2969 } else if (n->is_Opaque1()) {
2970 _igvn.replace_node(n, n->in(1));
2971 success = true;
2972 #if INCLUDE_RTM_OPT
2973 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) {
2974 assert(C->profile_rtm(), "should be used only in rtm deoptimization code");
2975 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), "");
2976 Node* cmp = n->unique_out();
2977 #ifdef ASSERT
2978 // Validate graph.
2979 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), "");
2980 BoolNode* bol = cmp->unique_out()->as_Bool();
2981 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() &&
2982 (bol->_test._test == BoolTest::ne), "");
2983 IfNode* ifn = bol->unique_out()->as_If();
2984 assert((ifn->outcnt() == 2) &&
2985 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != nullptr, "");
2986 #endif
2987 Node* repl = n->in(1);
2988 if (!_has_locks) {
2989 // Remove RTM state check if there are no locks in the code.
2990 // Replace input to compare the same value.
2991 repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1);
2992 }
2993 _igvn.replace_node(n, repl);
2994 success = true;
2995 #endif
2996 } else if (n->Opcode() == Op_Opaque4) {
2997 // With Opaque4 nodes, the expectation is that the test of input 1
2998 // is always equal to the constant value of input 2. So we can
2999 // remove the Opaque4 and replace it by input 2. In debug builds,
3000 // leave the non constant test in instead to sanity check that it
3001 // never fails (if it does, that subgraph was constructed so, at
3002 // runtime, a Halt node is executed).
3003 #ifdef ASSERT
3004 _igvn.replace_node(n, n->in(1));
3005 #else
3006 _igvn.replace_node(n, n->in(2));
3007 #endif
3008 success = true;
3009 } else if (n->Opcode() == Op_OuterStripMinedLoop) {
3010 n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn);
3011 C->remove_macro_node(n);
3012 success = true;
3013 } else if (n->Opcode() == Op_MaxL) {
3014 // Since MaxL and MinL are not implemented in the backend, we expand them to
3015 // a CMoveL construct now. At least until here, the type could be computed
3016 // precisely. CMoveL is not so smart, but we can give it at least the best
3017 // type we know abouot n now.
3018 Node* repl = MaxNode::signed_max(n->in(1), n->in(2), _igvn.type(n), _igvn);
3019 _igvn.replace_node(n, repl);
3020 success = true;
3021 } else if (n->Opcode() == Op_MinL) {
3022 Node* repl = MaxNode::signed_min(n->in(1), n->in(2), _igvn.type(n), _igvn);
3023 _igvn.replace_node(n, repl);
3024 success = true;
3025 }
3026 assert(!success || (C->macro_count() == (old_macro_count - 1)), "elimination must have deleted one node from macro list");
3027 progress = progress || success;
3028 }
3029 }
3030
3031 // Clean up the graph so we're less likely to hit the maximum node
3032 // limit
3033 _igvn.set_delay_transform(false);
3034 _igvn.optimize();
3035 if (C->failing()) return true;
3036 _igvn.set_delay_transform(true);
3037
3038
3039 // Because we run IGVN after each expansion, some macro nodes may go
3040 // dead and be removed from the list as we iterate over it. Move
3041 // Allocate nodes (processed in a second pass) at the beginning of
3042 // the list and then iterate from the last element of the list until
3043 // an Allocate node is seen. This is robust to random deletion in
3044 // the list due to nodes going dead.
3045 C->sort_macro_nodes();
3046
3047 // expand arraycopy "macro" nodes first
3048 // For ReduceBulkZeroing, we must first process all arraycopy nodes
3049 // before the allocate nodes are expanded.
3050 while (C->macro_count() > 0) {
3051 int macro_count = C->macro_count();
3052 Node * n = C->macro_node(macro_count-1);
3053 assert(n->is_macro(), "only macro nodes expected here");
3054 if (_igvn.type(n) == Type::TOP || (n->in(0) != nullptr && n->in(0)->is_top())) {
3055 // node is unreachable, so don't try to expand it
3056 C->remove_macro_node(n);
3057 continue;
3058 }
3059 if (n->is_Allocate()) {
3060 break;
3061 }
3062 // Make sure expansion will not cause node limit to be exceeded.
3063 // Worst case is a macro node gets expanded into about 200 nodes.
3064 // Allow 50% more for optimization.
3065 if (C->check_node_count(300, "out of nodes before macro expansion")) {
3066 return true;
3067 }
3068
3069 DEBUG_ONLY(int old_macro_count = C->macro_count();)
3070 switch (n->class_id()) {
3071 case Node::Class_Lock:
3072 expand_lock_node(n->as_Lock());
3073 break;
3074 case Node::Class_Unlock:
3075 expand_unlock_node(n->as_Unlock());
3076 break;
3077 case Node::Class_ArrayCopy:
3078 expand_arraycopy_node(n->as_ArrayCopy());
3079 break;
3080 case Node::Class_SubTypeCheck:
3081 expand_subtypecheck_node(n->as_SubTypeCheck());
3082 break;
3083 case Node::Class_CallStaticJava:
3084 expand_mh_intrinsic_return(n->as_CallStaticJava());
3085 C->remove_macro_node(n);
3086 break;
3087 case Node::Class_FlatArrayCheck:
3088 expand_flatarraycheck_node(n->as_FlatArrayCheck());
3089 break;
3090 default:
3091 assert(false, "unknown node type in macro list");
3092 }
3093 assert(C->macro_count() == (old_macro_count - 1), "expansion must have deleted one node from macro list");
3094 if (C->failing()) return true;
3095
3096 // Clean up the graph so we're less likely to hit the maximum node
3097 // limit
3098 _igvn.set_delay_transform(false);
3099 _igvn.optimize();
3100 if (C->failing()) return true;
3101 _igvn.set_delay_transform(true);
3102 }
3103
3104 // All nodes except Allocate nodes are expanded now. There could be
3105 // new optimization opportunities (such as folding newly created
3106 // load from a just allocated object). Run IGVN.
3107
3108 // expand "macro" nodes
3109 // nodes are removed from the macro list as they are processed
3110 while (C->macro_count() > 0) {
3111 int macro_count = C->macro_count();
3112 Node * n = C->macro_node(macro_count-1);
3113 assert(n->is_macro(), "only macro nodes expected here");
3114 if (_igvn.type(n) == Type::TOP || (n->in(0) != nullptr && n->in(0)->is_top())) {
3115 // node is unreachable, so don't try to expand it
3116 C->remove_macro_node(n);
3117 continue;
3118 }
3119 // Make sure expansion will not cause node limit to be exceeded.
3120 // Worst case is a macro node gets expanded into about 200 nodes.
3121 // Allow 50% more for optimization.
3122 if (C->check_node_count(300, "out of nodes before macro expansion")) {
3123 return true;
3124 }
3125 switch (n->class_id()) {
3126 case Node::Class_Allocate:
3127 expand_allocate(n->as_Allocate());
3128 break;
3129 case Node::Class_AllocateArray:
3130 expand_allocate_array(n->as_AllocateArray());
3131 break;
3132 default:
3133 assert(false, "unknown node type in macro list");
3134 }
3135 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
3136 if (C->failing()) return true;
3137
3138 // Clean up the graph so we're less likely to hit the maximum node
3139 // limit
3140 _igvn.set_delay_transform(false);
3141 _igvn.optimize();
3142 if (C->failing()) return true;
3143 _igvn.set_delay_transform(true);
3144 }
3145
3146 _igvn.set_delay_transform(false);
3147 return false;
3148 }
3149
3150 #ifndef PRODUCT
3151 int PhaseMacroExpand::_objs_scalar_replaced_counter = 0;
3152 int PhaseMacroExpand::_monitor_objects_removed_counter = 0;
3153 int PhaseMacroExpand::_GC_barriers_removed_counter = 0;
3154 int PhaseMacroExpand::_memory_barriers_removed_counter = 0;
3155
3156 void PhaseMacroExpand::print_statistics() {
3157 tty->print("Objects scalar replaced = %d, ", Atomic::load(&_objs_scalar_replaced_counter));
3158 tty->print("Monitor objects removed = %d, ", Atomic::load(&_monitor_objects_removed_counter));
3159 tty->print("GC barriers removed = %d, ", Atomic::load(&_GC_barriers_removed_counter));
3160 tty->print_cr("Memory barriers removed = %d", Atomic::load(&_memory_barriers_removed_counter));
3161 }
3162
3163 int PhaseMacroExpand::count_MemBar(Compile *C) {
3164 if (!PrintOptoStatistics) {
3165 return 0;
3166 }
3167 Unique_Node_List ideal_nodes;
3168 int total = 0;
3169 ideal_nodes.map(C->live_nodes(), nullptr);
3170 ideal_nodes.push(C->root());
3171 for (uint next = 0; next < ideal_nodes.size(); ++next) {
3172 Node* n = ideal_nodes.at(next);
3173 if (n->is_MemBar()) {
3174 total++;
3175 }
3176 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3177 Node* m = n->fast_out(i);
3178 ideal_nodes.push(m);
3179 }
3180 }
3181 return total;
3182 }
3183 #endif