1 /*
2 * Copyright (c) 2005, 2024, 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->as_Allocate()->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
162 assert(alloc_mem != nullptr, "Allocation without a memory projection.");
163 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
164 while (true) {
165 if (mem == alloc_mem || mem == start_mem ) {
166 return mem; // hit one of our sentinels
167 } else if (mem->is_MergeMem()) {
168 mem = mem->as_MergeMem()->memory_at(alias_idx);
169 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
170 Node *in = mem->in(0);
171 // we can safely skip over safepoints, calls, locks and membars because we
172 // already know that the object is safe to eliminate.
173 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
174 return in;
175 } else if (in->is_Call()) {
176 CallNode *call = in->as_Call();
177 if (call->may_modify(tinst, phase)) {
178 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape");
179 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) {
180 return in;
181 }
182 }
183 mem = in->in(TypeFunc::Memory);
184 } else if (in->is_MemBar()) {
185 ArrayCopyNode* ac = nullptr;
186 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) {
187 if (ac != nullptr) {
188 assert(ac->is_clonebasic(), "Only basic clone is a non escaping clone");
189 return ac;
190 }
191 }
192 mem = in->in(TypeFunc::Memory);
193 } else {
194 #ifdef ASSERT
195 in->dump();
196 mem->dump();
197 assert(false, "unexpected projection");
198 #endif
199 }
200 } else if (mem->is_Store()) {
201 const TypePtr* atype = mem->as_Store()->adr_type();
202 int adr_idx = phase->C->get_alias_index(atype);
203 if (adr_idx == alias_idx) {
204 assert(atype->isa_oopptr(), "address type must be oopptr");
205 int adr_offset = atype->flat_offset();
206 uint adr_iid = atype->is_oopptr()->instance_id();
207 // Array elements references have the same alias_idx
208 // but different offset and different instance_id.
209 if (adr_offset == offset && adr_iid == alloc->_idx) {
210 return mem;
211 }
212 } else {
213 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
214 }
215 mem = mem->in(MemNode::Memory);
216 } else if (mem->is_ClearArray()) {
217 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
218 // Can not bypass initialization of the instance
219 // we are looking.
220 debug_only(intptr_t offset;)
221 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
222 InitializeNode* init = alloc->as_Allocate()->initialization();
223 // We are looking for stored value, return Initialize node
224 // or memory edge from Allocate node.
225 if (init != nullptr) {
226 return init;
227 } else {
228 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
229 }
230 }
231 // Otherwise skip it (the call updated 'mem' value).
232 } else if (mem->Opcode() == Op_SCMemProj) {
233 mem = mem->in(0);
234 Node* adr = nullptr;
235 if (mem->is_LoadStore()) {
236 adr = mem->in(MemNode::Address);
237 } else {
238 assert(mem->Opcode() == Op_EncodeISOArray ||
239 mem->Opcode() == Op_StrCompressedCopy, "sanity");
240 adr = mem->in(3); // Destination array
241 }
242 const TypePtr* atype = adr->bottom_type()->is_ptr();
243 int adr_idx = phase->C->get_alias_index(atype);
244 if (adr_idx == alias_idx) {
245 DEBUG_ONLY(mem->dump();)
246 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
247 return nullptr;
248 }
249 mem = mem->in(MemNode::Memory);
250 } else if (mem->Opcode() == Op_StrInflatedCopy) {
251 Node* adr = mem->in(3); // Destination array
252 const TypePtr* atype = adr->bottom_type()->is_ptr();
253 int adr_idx = phase->C->get_alias_index(atype);
254 if (adr_idx == alias_idx) {
255 DEBUG_ONLY(mem->dump();)
256 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
257 return nullptr;
258 }
259 mem = mem->in(MemNode::Memory);
260 } else {
261 return mem;
262 }
263 assert(mem != orig_mem, "dead memory loop");
264 }
265 }
266
267 // Generate loads from source of the arraycopy for fields of
268 // destination needed at a deoptimization point
269 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) {
270 BasicType bt = ft;
271 const Type *type = ftype;
272 if (ft == T_NARROWOOP) {
273 bt = T_OBJECT;
274 type = ftype->make_oopptr();
275 }
276 Node* res = nullptr;
277 if (ac->is_clonebasic()) {
278 assert(ac->in(ArrayCopyNode::Src) != ac->in(ArrayCopyNode::Dest), "clone source equals destination");
279 Node* base = ac->in(ArrayCopyNode::Src);
280 Node* adr = _igvn.transform(new AddPNode(base, base, _igvn.MakeConX(offset)));
281 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
282 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
283 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
284 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
285 } else {
286 if (ac->modifies(offset, offset, &_igvn, true)) {
287 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
288 uint shift = exact_log2(type2aelembytes(bt));
289 Node* src_pos = ac->in(ArrayCopyNode::SrcPos);
290 Node* dest_pos = ac->in(ArrayCopyNode::DestPos);
291 const TypeInt* src_pos_t = _igvn.type(src_pos)->is_int();
292 const TypeInt* dest_pos_t = _igvn.type(dest_pos)->is_int();
293
294 Node* adr = nullptr;
295 Node* base = ac->in(ArrayCopyNode::Src);
296 const TypeAryPtr* adr_type = _igvn.type(base)->is_aryptr();
297 if (adr_type->is_flat()) {
298 shift = adr_type->flat_log_elem_size();
299 }
300 if (src_pos_t->is_con() && dest_pos_t->is_con()) {
301 intptr_t off = ((src_pos_t->get_con() - dest_pos_t->get_con()) << shift) + offset;
302 adr = _igvn.transform(new AddPNode(base, base, _igvn.MakeConX(off)));
303 adr_type = _igvn.type(adr)->is_aryptr();
304 assert(adr_type == _igvn.type(base)->is_aryptr()->add_field_offset_and_offset(off), "incorrect address type");
305 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
306 // Don't emit a new load from src if src == dst but try to get the value from memory instead
307 return value_from_mem(ac->in(TypeFunc::Memory), ctl, ft, ftype, adr_type, alloc);
308 }
309 } else {
310 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
311 // Non constant offset in the array: we can't statically
312 // determine the value
313 return nullptr;
314 }
315 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
316 #ifdef _LP64
317 diff = _igvn.transform(new ConvI2LNode(diff));
318 #endif
319 diff = _igvn.transform(new LShiftXNode(diff, _igvn.intcon(shift)));
320
321 Node* off = _igvn.transform(new AddXNode(_igvn.MakeConX(offset), diff));
322 adr = _igvn.transform(new AddPNode(base, base, off));
323 // In the case of a flat inline type array, each field has its
324 // own slice so we need to extract the field being accessed from
325 // the address computation
326 adr_type = adr_type->add_field_offset_and_offset(offset)->add_offset(Type::OffsetBot)->is_aryptr();
327 adr = _igvn.transform(new CastPPNode(ctl, adr, adr_type));
328 }
329 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
330 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
331 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
332 }
333 }
334 if (res != nullptr) {
335 if (ftype->isa_narrowoop()) {
336 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
337 assert(res->isa_DecodeN(), "should be narrow oop");
338 res = _igvn.transform(new EncodePNode(res, ftype));
339 }
340 return res;
341 }
342 return nullptr;
343 }
344
345 //
346 // Given a Memory Phi, compute a value Phi containing the values from stores
347 // on the input paths.
348 // Note: this function is recursive, its depth is limited by the "level" argument
349 // Returns the computed Phi, or null if it cannot compute it.
350 Node *PhaseMacroExpand::value_from_mem_phi(Node *mem, BasicType ft, const Type *phi_type, const TypeOopPtr *adr_t, AllocateNode *alloc, Node_Stack *value_phis, int level) {
351 assert(mem->is_Phi(), "sanity");
352 int alias_idx = C->get_alias_index(adr_t);
353 int offset = adr_t->flat_offset();
354 int instance_id = adr_t->instance_id();
355
356 // Check if an appropriate value phi already exists.
357 Node* region = mem->in(0);
358 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
359 Node* phi = region->fast_out(k);
360 if (phi->is_Phi() && phi != mem &&
361 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
362 return phi;
363 }
364 }
365 // Check if an appropriate new value phi already exists.
366 Node* new_phi = value_phis->find(mem->_idx);
367 if (new_phi != nullptr)
368 return new_phi;
369
370 if (level <= 0) {
371 return nullptr; // Give up: phi tree too deep
372 }
373 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
374 Node *alloc_mem = alloc->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
375 assert(alloc_mem != nullptr, "Allocation without a memory projection.");
376
377 uint length = mem->req();
378 GrowableArray <Node *> values(length, length, nullptr);
379
380 // create a new Phi for the value
381 PhiNode *phi = new PhiNode(mem->in(0), phi_type, nullptr, mem->_idx, instance_id, alias_idx, offset);
382 transform_later(phi);
383 value_phis->push(phi, mem->_idx);
384
385 for (uint j = 1; j < length; j++) {
386 Node *in = mem->in(j);
387 if (in == nullptr || in->is_top()) {
388 values.at_put(j, in);
389 } else {
390 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
391 if (val == start_mem || val == alloc_mem) {
392 // hit a sentinel, return appropriate 0 value
393 Node* default_value = alloc->in(AllocateNode::DefaultValue);
394 if (default_value != nullptr) {
395 values.at_put(j, default_value);
396 } else {
397 assert(alloc->in(AllocateNode::RawDefaultValue) == nullptr, "default value may not be null");
398 values.at_put(j, _igvn.zerocon(ft));
399 }
400 continue;
401 }
402 if (val->is_Initialize()) {
403 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
404 }
405 if (val == nullptr) {
406 return nullptr; // can't find a value on this path
407 }
408 if (val == mem) {
409 values.at_put(j, mem);
410 } else if (val->is_Store()) {
411 Node* n = val->in(MemNode::ValueIn);
412 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
413 n = bs->step_over_gc_barrier(n);
414 if (is_subword_type(ft)) {
415 n = Compile::narrow_value(ft, n, phi_type, &_igvn, true);
416 }
417 values.at_put(j, n);
418 } else if(val->is_Proj() && val->in(0) == alloc) {
419 Node* default_value = alloc->in(AllocateNode::DefaultValue);
420 if (default_value != nullptr) {
421 values.at_put(j, default_value);
422 } else {
423 assert(alloc->in(AllocateNode::RawDefaultValue) == nullptr, "default value may not be null");
424 values.at_put(j, _igvn.zerocon(ft));
425 }
426 } else if (val->is_Phi()) {
427 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
428 if (val == nullptr) {
429 return nullptr;
430 }
431 values.at_put(j, val);
432 } else if (val->Opcode() == Op_SCMemProj) {
433 assert(val->in(0)->is_LoadStore() ||
434 val->in(0)->Opcode() == Op_EncodeISOArray ||
435 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
436 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
437 return nullptr;
438 } else if (val->is_ArrayCopy()) {
439 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc);
440 if (res == nullptr) {
441 return nullptr;
442 }
443 values.at_put(j, res);
444 } else {
445 DEBUG_ONLY( val->dump(); )
446 assert(false, "unknown node on this path");
447 return nullptr; // unknown node on this path
448 }
449 }
450 }
451 // Set Phi's inputs
452 for (uint j = 1; j < length; j++) {
453 if (values.at(j) == mem) {
454 phi->init_req(j, phi);
455 } else {
456 phi->init_req(j, values.at(j));
457 }
458 }
459 return phi;
460 }
461
462 // Search the last value stored into the object's field.
463 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
464 assert(adr_t->is_known_instance_field(), "instance required");
465 int instance_id = adr_t->instance_id();
466 assert((uint)instance_id == alloc->_idx, "wrong allocation");
467
468 int alias_idx = C->get_alias_index(adr_t);
469 int offset = adr_t->flat_offset();
470 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
471 Node *alloc_mem = alloc->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
472 assert(alloc_mem != nullptr, "Allocation without a memory projection.");
473 VectorSet visited;
474
475 bool done = sfpt_mem == alloc_mem;
476 Node *mem = sfpt_mem;
477 while (!done) {
478 if (visited.test_set(mem->_idx)) {
479 return nullptr; // found a loop, give up
480 }
481 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
482 if (mem == start_mem || mem == alloc_mem) {
483 done = true; // hit a sentinel, return appropriate 0 value
484 } else if (mem->is_Initialize()) {
485 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
486 if (mem == nullptr) {
487 done = true; // Something went wrong.
488 } else if (mem->is_Store()) {
489 const TypePtr* atype = mem->as_Store()->adr_type();
490 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
491 done = true;
492 }
493 } else if (mem->is_Store()) {
494 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
495 assert(atype != nullptr, "address type must be oopptr");
496 assert(C->get_alias_index(atype) == alias_idx &&
497 atype->is_known_instance_field() && atype->flat_offset() == offset &&
498 atype->instance_id() == instance_id, "store is correct memory slice");
499 done = true;
500 } else if (mem->is_Phi()) {
501 // try to find a phi's unique input
502 Node *unique_input = nullptr;
503 Node *top = C->top();
504 for (uint i = 1; i < mem->req(); i++) {
505 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
506 if (n == nullptr || n == top || n == mem) {
507 continue;
508 } else if (unique_input == nullptr) {
509 unique_input = n;
510 } else if (unique_input != n) {
511 unique_input = top;
512 break;
513 }
514 }
515 if (unique_input != nullptr && unique_input != top) {
516 mem = unique_input;
517 } else {
518 done = true;
519 }
520 } else if (mem->is_ArrayCopy()) {
521 done = true;
522 } else {
523 DEBUG_ONLY( mem->dump(); )
524 assert(false, "unexpected node");
525 }
526 }
527 if (mem != nullptr) {
528 if (mem == start_mem || mem == alloc_mem) {
529 // hit a sentinel, return appropriate 0 value
530 Node* default_value = alloc->in(AllocateNode::DefaultValue);
531 if (default_value != nullptr) {
532 return default_value;
533 }
534 assert(alloc->in(AllocateNode::RawDefaultValue) == nullptr, "default value may not be null");
535 return _igvn.zerocon(ft);
536 } else if (mem->is_Store()) {
537 Node* n = mem->in(MemNode::ValueIn);
538 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
539 n = bs->step_over_gc_barrier(n);
540 return n;
541 } else if (mem->is_Phi()) {
542 // attempt to produce a Phi reflecting the values on the input paths of the Phi
543 Node_Stack value_phis(8);
544 Node* phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
545 if (phi != nullptr) {
546 return phi;
547 } else {
548 // Kill all new Phis
549 while(value_phis.is_nonempty()) {
550 Node* n = value_phis.node();
551 _igvn.replace_node(n, C->top());
552 value_phis.pop();
553 }
554 }
555 } else if (mem->is_ArrayCopy()) {
556 Node* ctl = mem->in(0);
557 Node* m = mem->in(TypeFunc::Memory);
558 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj()) {
559 // pin the loads in the uncommon trap path
560 ctl = sfpt_ctl;
561 m = sfpt_mem;
562 }
563 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc);
564 }
565 }
566 // Something went wrong.
567 return nullptr;
568 }
569
570 // Search the last value stored into the inline type's fields.
571 Node* PhaseMacroExpand::inline_type_from_mem(Node* mem, Node* ctl, ciInlineKlass* vk, const TypeAryPtr* adr_type, int offset, AllocateNode* alloc) {
572 // Subtract the offset of the first field to account for the missing oop header
573 offset -= vk->first_field_offset();
574 // Create a new InlineTypeNode and retrieve the field values from memory
575 InlineTypeNode* vt = InlineTypeNode::make_uninitialized(_igvn, vk);
576 transform_later(vt);
577 for (int i = 0; i < vk->nof_declared_nonstatic_fields(); ++i) {
578 ciType* field_type = vt->field_type(i);
579 int field_offset = offset + vt->field_offset(i);
580 Node* value = nullptr;
581 if (vt->field_is_flat(i)) {
582 value = inline_type_from_mem(mem, ctl, field_type->as_inline_klass(), adr_type, field_offset, alloc);
583 } else {
584 const Type* ft = Type::get_const_type(field_type);
585 BasicType bt = type2field[field_type->basic_type()];
586 if (UseCompressedOops && !is_java_primitive(bt)) {
587 ft = ft->make_narrowoop();
588 bt = T_NARROWOOP;
589 }
590 // Each inline type field has its own memory slice
591 adr_type = adr_type->with_field_offset(field_offset);
592 value = value_from_mem(mem, ctl, bt, ft, adr_type, alloc);
593 if (value != nullptr && ft->isa_narrowoop()) {
594 assert(UseCompressedOops, "unexpected narrow oop");
595 if (value->is_EncodeP()) {
596 value = value->in(1);
597 } else {
598 value = transform_later(new DecodeNNode(value, value->get_ptr_type()));
599 }
600 }
601 }
602 if (value != nullptr) {
603 vt->set_field_value(i, value);
604 } else {
605 // We might have reached the TrackedInitializationLimit
606 return nullptr;
607 }
608 }
609 return vt;
610 }
611
612 // Check the possibility of scalar replacement.
613 bool PhaseMacroExpand::can_eliminate_allocation(PhaseIterGVN* igvn, AllocateNode *alloc, GrowableArray <SafePointNode *>* safepoints) {
614 // Scan the uses of the allocation to check for anything that would
615 // prevent us from eliminating it.
616 NOT_PRODUCT( const char* fail_eliminate = nullptr; )
617 DEBUG_ONLY( Node* disq_node = nullptr; )
618 bool can_eliminate = true;
619 bool reduce_merge_precheck = (safepoints == nullptr);
620
621 Unique_Node_List worklist;
622 Node* res = alloc->result_cast();
623 const TypeOopPtr* res_type = nullptr;
624 if (res == nullptr) {
625 // All users were eliminated.
626 } else if (!res->is_CheckCastPP()) {
627 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
628 can_eliminate = false;
629 } else {
630 worklist.push(res);
631 res_type = igvn->type(res)->isa_oopptr();
632 if (res_type == nullptr) {
633 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
634 can_eliminate = false;
635 } else if (!res_type->klass_is_exact()) {
636 NOT_PRODUCT(fail_eliminate = "Not an exact type.";)
637 can_eliminate = false;
638 } else if (res_type->isa_aryptr()) {
639 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
640 if (length < 0) {
641 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
642 can_eliminate = false;
643 }
644 }
645 }
646
647 while (can_eliminate && worklist.size() > 0) {
648 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
649 res = worklist.pop();
650 for (DUIterator_Fast jmax, j = res->fast_outs(jmax); j < jmax && can_eliminate; j++) {
651 Node* use = res->fast_out(j);
652
653 if (use->is_AddP()) {
654 const TypePtr* addp_type = igvn->type(use)->is_ptr();
655 int offset = addp_type->offset();
656
657 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
658 NOT_PRODUCT(fail_eliminate = "Undefined field reference";)
659 can_eliminate = false;
660 break;
661 }
662 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
663 k < kmax && can_eliminate; k++) {
664 Node* n = use->fast_out(k);
665 if (!n->is_Store() && n->Opcode() != Op_CastP2X && !bs->is_gc_pre_barrier_node(n) && !reduce_merge_precheck) {
666 DEBUG_ONLY(disq_node = n;)
667 if (n->is_Load() || n->is_LoadStore()) {
668 NOT_PRODUCT(fail_eliminate = "Field load";)
669 } else {
670 NOT_PRODUCT(fail_eliminate = "Not store field reference";)
671 }
672 can_eliminate = false;
673 }
674 }
675 } else if (use->is_ArrayCopy() &&
676 (use->as_ArrayCopy()->is_clonebasic() ||
677 use->as_ArrayCopy()->is_arraycopy_validated() ||
678 use->as_ArrayCopy()->is_copyof_validated() ||
679 use->as_ArrayCopy()->is_copyofrange_validated()) &&
680 use->in(ArrayCopyNode::Dest) == res) {
681 // ok to eliminate
682 } else if (use->is_SafePoint()) {
683 SafePointNode* sfpt = use->as_SafePoint();
684 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
685 // Object is passed as argument.
686 DEBUG_ONLY(disq_node = use;)
687 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
688 can_eliminate = false;
689 }
690 Node* sfptMem = sfpt->memory();
691 if (sfptMem == nullptr || sfptMem->is_top()) {
692 DEBUG_ONLY(disq_node = use;)
693 NOT_PRODUCT(fail_eliminate = "null or TOP memory";)
694 can_eliminate = false;
695 } else if (!reduce_merge_precheck) {
696 assert(!res->is_Phi() || !res->as_Phi()->can_be_inline_type(), "Inline type allocations should not have safepoint uses");
697 safepoints->append_if_missing(sfpt);
698 }
699 } else if (use->is_InlineType() && use->as_InlineType()->get_oop() == res) {
700 // Look at uses
701 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); k < kmax; k++) {
702 Node* u = use->fast_out(k);
703 if (u->is_InlineType()) {
704 // Use in flat field can be eliminated
705 InlineTypeNode* vt = u->as_InlineType();
706 for (uint i = 0; i < vt->field_count(); ++i) {
707 if (vt->field_value(i) == use && !vt->field_is_flat(i)) {
708 can_eliminate = false; // Use in non-flat field
709 break;
710 }
711 }
712 } else {
713 // Add other uses to the worklist to process individually
714 worklist.push(use);
715 }
716 }
717 } else if (use->Opcode() == Op_StoreX && use->in(MemNode::Address) == res) {
718 // Store to mark word of inline type larval buffer
719 assert(res_type->is_inlinetypeptr(), "Unexpected store to mark word");
720 } else if (res_type->is_inlinetypeptr() && (use->Opcode() == Op_MemBarRelease || use->Opcode() == Op_MemBarStoreStore)) {
721 // Inline type buffer allocations are followed by a membar
722 } else if (reduce_merge_precheck &&
723 (use->is_Phi() || use->is_EncodeP() ||
724 use->Opcode() == Op_MemBarRelease ||
725 (UseStoreStoreForCtor && use->Opcode() == Op_MemBarStoreStore))) {
726 // Nothing to do
727 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
728 if (use->is_Phi()) {
729 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
730 NOT_PRODUCT(fail_eliminate = "Object is return value";)
731 } else {
732 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
733 }
734 DEBUG_ONLY(disq_node = use;)
735 } else {
736 if (use->Opcode() == Op_Return) {
737 NOT_PRODUCT(fail_eliminate = "Object is return value";)
738 } else {
739 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
740 }
741 DEBUG_ONLY(disq_node = use;)
742 }
743 can_eliminate = false;
744 } else {
745 assert(use->Opcode() == Op_CastP2X, "should be");
746 assert(!use->has_out_with(Op_OrL), "should have been removed because oop is never null");
747 }
748 }
749 }
750
751 #ifndef PRODUCT
752 if (PrintEliminateAllocations && safepoints != nullptr) {
753 if (can_eliminate) {
754 tty->print("Scalar ");
755 if (res == nullptr)
756 alloc->dump();
757 else
758 res->dump();
759 } else {
760 tty->print("NotScalar (%s)", fail_eliminate);
761 if (res == nullptr)
762 alloc->dump();
763 else
764 res->dump();
765 #ifdef ASSERT
766 if (disq_node != nullptr) {
767 tty->print(" >>>> ");
768 disq_node->dump();
769 }
770 #endif /*ASSERT*/
771 }
772 }
773
774 if (TraceReduceAllocationMerges && !can_eliminate && reduce_merge_precheck) {
775 tty->print_cr("\tCan't eliminate allocation because '%s': ", fail_eliminate != nullptr ? fail_eliminate : "");
776 DEBUG_ONLY(if (disq_node != nullptr) disq_node->dump();)
777 }
778 #endif
779 return can_eliminate;
780 }
781
782 void PhaseMacroExpand::undo_previous_scalarizations(GrowableArray <SafePointNode *> safepoints_done, AllocateNode* alloc) {
783 Node* res = alloc->result_cast();
784 int nfields = 0;
785 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
786
787 if (res != nullptr) {
788 const TypeOopPtr* res_type = _igvn.type(res)->isa_oopptr();
789
790 if (res_type->isa_instptr()) {
791 // find the fields of the class which will be needed for safepoint debug information
792 ciInstanceKlass* iklass = res_type->is_instptr()->instance_klass();
793 nfields = iklass->nof_nonstatic_fields();
794 } else {
795 // find the array's elements which will be needed for safepoint debug information
796 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
797 assert(nfields >= 0, "must be an array klass.");
798 }
799 }
800
801 // rollback processed safepoints
802 while (safepoints_done.length() > 0) {
803 SafePointNode* sfpt_done = safepoints_done.pop();
804 // remove any extra entries we added to the safepoint
805 uint last = sfpt_done->req() - 1;
806 for (int k = 0; k < nfields; k++) {
807 sfpt_done->del_req(last--);
808 }
809 JVMState *jvms = sfpt_done->jvms();
810 jvms->set_endoff(sfpt_done->req());
811 // Now make a pass over the debug information replacing any references
812 // to SafePointScalarObjectNode with the allocated object.
813 int start = jvms->debug_start();
814 int end = jvms->debug_end();
815 for (int i = start; i < end; i++) {
816 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
817 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
818 if (scobj->first_index(jvms) == sfpt_done->req() &&
819 scobj->n_fields() == (uint)nfields) {
820 assert(scobj->alloc() == alloc, "sanity");
821 sfpt_done->set_req(i, res);
822 }
823 }
824 }
825 _igvn._worklist.push(sfpt_done);
826 }
827 }
828
829 SafePointScalarObjectNode* PhaseMacroExpand::create_scalarized_object_description(AllocateNode *alloc, SafePointNode* sfpt,
830 Unique_Node_List* value_worklist) {
831 // Fields of scalar objs are referenced only at the end
832 // of regular debuginfo at the last (youngest) JVMS.
833 // Record relative start index.
834 ciInstanceKlass* iklass = nullptr;
835 BasicType basic_elem_type = T_ILLEGAL;
836 const Type* field_type = nullptr;
837 const TypeOopPtr* res_type = nullptr;
838 int nfields = 0;
839 int array_base = 0;
840 int element_size = 0;
841 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
842 Node* res = alloc->result_cast();
843
844 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
845 assert(sfpt->jvms() != nullptr, "missed JVMS");
846
847 if (res != nullptr) { // Could be null when there are no users
848 res_type = _igvn.type(res)->isa_oopptr();
849
850 if (res_type->isa_instptr()) {
851 // find the fields of the class which will be needed for safepoint debug information
852 iklass = res_type->is_instptr()->instance_klass();
853 nfields = iklass->nof_nonstatic_fields();
854 } else {
855 // find the array's elements which will be needed for safepoint debug information
856 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
857 assert(nfields >= 0, "must be an array klass.");
858 basic_elem_type = res_type->is_aryptr()->elem()->array_element_basic_type();
859 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
860 element_size = type2aelembytes(basic_elem_type);
861 field_type = res_type->is_aryptr()->elem();
862 if (res_type->is_flat()) {
863 // Flat inline type array
864 element_size = res_type->is_aryptr()->flat_elem_size();
865 }
866 }
867
868 if (res->bottom_type()->is_inlinetypeptr()) {
869 // Nullable inline types have an IsInit field which is added to the safepoint when scalarizing them (see
870 // InlineTypeNode::make_scalar_in_safepoint()). When having circular inline types, we stop scalarizing at depth 1
871 // to avoid an endless recursion. Therefore, we do not have a SafePointScalarObjectNode node here, yet.
872 // We are about to create a SafePointScalarObjectNode as if this is a normal object. Add an additional int input
873 // with value 1 which sets IsInit to true to indicate that the object is always non-null. This input is checked
874 // later in PhaseOutput::filLocArray() for inline types.
875 sfpt->add_req(_igvn.intcon(1));
876 }
877 }
878
879 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type, alloc, first_ind, sfpt->jvms()->depth(), nfields);
880 sobj->init_req(0, C->root());
881 transform_later(sobj);
882
883 // Scan object's fields adding an input to the safepoint for each field.
884 for (int j = 0; j < nfields; j++) {
885 intptr_t offset;
886 ciField* field = nullptr;
887 if (iklass != nullptr) {
888 field = iklass->nonstatic_field_at(j);
889 offset = field->offset_in_bytes();
890 ciType* elem_type = field->type();
891 basic_elem_type = field->layout_type();
892 assert(!field->is_flat(), "flat inline type fields should not have safepoint uses");
893
894 // The next code is taken from Parse::do_get_xxx().
895 if (is_reference_type(basic_elem_type)) {
896 if (!elem_type->is_loaded()) {
897 field_type = TypeInstPtr::BOTTOM;
898 } else if (field != nullptr && field->is_static_constant()) {
899 ciObject* con = field->constant_value().as_object();
900 // Do not "join" in the previous type; it doesn't add value,
901 // and may yield a vacuous result if the field is of interface type.
902 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
903 assert(field_type != nullptr, "field singleton type must be consistent");
904 } else {
905 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
906 }
907 if (UseCompressedOops) {
908 field_type = field_type->make_narrowoop();
909 basic_elem_type = T_NARROWOOP;
910 }
911 } else {
912 field_type = Type::get_const_basic_type(basic_elem_type);
913 }
914 } else {
915 offset = array_base + j * (intptr_t)element_size;
916 }
917
918 Node* field_val = nullptr;
919 const TypeOopPtr* field_addr_type = res_type->add_offset(offset)->isa_oopptr();
920 if (res_type->is_flat()) {
921 ciInlineKlass* inline_klass = res_type->is_aryptr()->elem()->inline_klass();
922 assert(inline_klass->flat_in_array(), "must be flat in array");
923 field_val = inline_type_from_mem(sfpt->memory(), sfpt->control(), inline_klass, field_addr_type->isa_aryptr(), 0, alloc);
924 } else {
925 field_val = value_from_mem(sfpt->memory(), sfpt->control(), basic_elem_type, field_type, field_addr_type, alloc);
926 }
927
928 // We weren't able to find a value for this field,
929 // give up on eliminating this allocation.
930 if (field_val == nullptr) {
931 uint last = sfpt->req() - 1;
932 for (int k = 0; k < j; k++) {
933 sfpt->del_req(last--);
934 }
935 _igvn._worklist.push(sfpt);
936
937 #ifndef PRODUCT
938 if (PrintEliminateAllocations) {
939 if (field != nullptr) {
940 tty->print("=== At SafePoint node %d can't find value of field: ", sfpt->_idx);
941 field->print();
942 int field_idx = C->get_alias_index(field_addr_type);
943 tty->print(" (alias_idx=%d)", field_idx);
944 } else { // Array's element
945 tty->print("=== At SafePoint node %d can't find value of array element [%d]", sfpt->_idx, j);
946 }
947 tty->print(", which prevents elimination of: ");
948 if (res == nullptr)
949 alloc->dump();
950 else
951 res->dump();
952 }
953 #endif
954
955 return nullptr;
956 }
957
958 if (UseCompressedOops && field_type->isa_narrowoop()) {
959 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
960 // to be able scalar replace the allocation.
961 if (field_val->is_EncodeP()) {
962 field_val = field_val->in(1);
963 } else if (!field_val->is_InlineType()) {
964 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
965 }
966 }
967
968 // Keep track of inline types to scalarize them later
969 if (field_val->is_InlineType()) {
970 value_worklist->push(field_val);
971 } else if (field_val->is_Phi()) {
972 PhiNode* phi = field_val->as_Phi();
973 // Eagerly replace inline type phis now since we could be removing an inline type allocation where we must
974 // scalarize all its fields in safepoints.
975 field_val = phi->try_push_inline_types_down(&_igvn, true);
976 if (field_val->is_InlineType()) {
977 value_worklist->push(field_val);
978 }
979 }
980 sfpt->add_req(field_val);
981 }
982
983 sfpt->jvms()->set_endoff(sfpt->req());
984
985 return sobj;
986 }
987
988 // Do scalar replacement.
989 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
990 GrowableArray <SafePointNode *> safepoints_done;
991 Node* res = alloc->result_cast();
992 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
993 const TypeOopPtr* res_type = nullptr;
994 if (res != nullptr) { // Could be null when there are no users
995 res_type = _igvn.type(res)->isa_oopptr();
996 }
997
998 // Process the safepoint uses
999 assert(safepoints.length() == 0 || !res_type->is_inlinetypeptr() || C->has_circular_inline_type(),
1000 "Inline type allocations should have been scalarized earlier");
1001 Unique_Node_List value_worklist;
1002 while (safepoints.length() > 0) {
1003 SafePointNode* sfpt = safepoints.pop();
1004 SafePointScalarObjectNode* sobj = create_scalarized_object_description(alloc, sfpt, &value_worklist);
1005
1006 if (sobj == nullptr) {
1007 undo_previous_scalarizations(safepoints_done, alloc);
1008 return false;
1009 }
1010
1011 // Now make a pass over the debug information replacing any references
1012 // to the allocated object with "sobj"
1013 JVMState *jvms = sfpt->jvms();
1014 sfpt->replace_edges_in_range(res, sobj, jvms->debug_start(), jvms->debug_end(), &_igvn);
1015 _igvn._worklist.push(sfpt);
1016
1017 // keep it for rollback
1018 safepoints_done.append_if_missing(sfpt);
1019 }
1020 // Scalarize inline types that were added to the safepoint.
1021 // Don't allow linking a constant oop (if available) for flat array elements
1022 // because Deoptimization::reassign_flat_array_elements needs field values.
1023 bool allow_oop = (res_type != nullptr) && !res_type->is_flat();
1024 for (uint i = 0; i < value_worklist.size(); ++i) {
1025 InlineTypeNode* vt = value_worklist.at(i)->as_InlineType();
1026 vt->make_scalar_in_safepoints(&_igvn, allow_oop);
1027 }
1028 return true;
1029 }
1030
1031 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
1032 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
1033 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
1034 if (ctl_proj != nullptr) {
1035 igvn.replace_node(ctl_proj, n->in(0));
1036 }
1037 if (mem_proj != nullptr) {
1038 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
1039 }
1040 }
1041
1042 // Process users of eliminated allocation.
1043 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc, bool inline_alloc) {
1044 Unique_Node_List worklist;
1045 Node* res = alloc->result_cast();
1046 if (res != nullptr) {
1047 worklist.push(res);
1048 }
1049 while (worklist.size() > 0) {
1050 res = worklist.pop();
1051 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
1052 Node *use = res->last_out(j);
1053 uint oc1 = res->outcnt();
1054
1055 if (use->is_AddP()) {
1056 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
1057 Node *n = use->last_out(k);
1058 uint oc2 = use->outcnt();
1059 if (n->is_Store()) {
1060 for (DUIterator_Fast pmax, p = n->fast_outs(pmax); p < pmax; p++) {
1061 MemBarNode* mb = n->fast_out(p)->isa_MemBar();
1062 if (mb != nullptr && mb->req() <= MemBarNode::Precedent && mb->in(MemBarNode::Precedent) == n) {
1063 // MemBarVolatiles should have been removed by MemBarNode::Ideal() for non-inline allocations
1064 assert(inline_alloc, "MemBarVolatile should be eliminated for non-escaping object");
1065 mb->remove(&_igvn);
1066 }
1067 }
1068 _igvn.replace_node(n, n->in(MemNode::Memory));
1069 } else {
1070 eliminate_gc_barrier(n);
1071 }
1072 k -= (oc2 - use->outcnt());
1073 }
1074 _igvn.remove_dead_node(use);
1075 } else if (use->is_ArrayCopy()) {
1076 // Disconnect ArrayCopy node
1077 ArrayCopyNode* ac = use->as_ArrayCopy();
1078 if (ac->is_clonebasic()) {
1079 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
1080 disconnect_projections(ac, _igvn);
1081 assert(alloc->in(TypeFunc::Memory)->is_Proj() && alloc->in(TypeFunc::Memory)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
1082 Node* membar_before = alloc->in(TypeFunc::Memory)->in(0);
1083 disconnect_projections(membar_before->as_MemBar(), _igvn);
1084 if (membar_after->is_MemBar()) {
1085 disconnect_projections(membar_after->as_MemBar(), _igvn);
1086 }
1087 } else {
1088 assert(ac->is_arraycopy_validated() ||
1089 ac->is_copyof_validated() ||
1090 ac->is_copyofrange_validated(), "unsupported");
1091 CallProjections* callprojs = ac->extract_projections(true);
1092
1093 _igvn.replace_node(callprojs->fallthrough_ioproj, ac->in(TypeFunc::I_O));
1094 _igvn.replace_node(callprojs->fallthrough_memproj, ac->in(TypeFunc::Memory));
1095 _igvn.replace_node(callprojs->fallthrough_catchproj, ac->in(TypeFunc::Control));
1096
1097 // Set control to top. IGVN will remove the remaining projections
1098 ac->set_req(0, top());
1099 ac->replace_edge(res, top(), &_igvn);
1100
1101 // Disconnect src right away: it can help find new
1102 // opportunities for allocation elimination
1103 Node* src = ac->in(ArrayCopyNode::Src);
1104 ac->replace_edge(src, top(), &_igvn);
1105 // src can be top at this point if src and dest of the
1106 // arraycopy were the same
1107 if (src->outcnt() == 0 && !src->is_top()) {
1108 _igvn.remove_dead_node(src);
1109 }
1110 }
1111 _igvn._worklist.push(ac);
1112 } else if (use->is_InlineType()) {
1113 assert(use->as_InlineType()->get_oop() == res, "unexpected inline type ptr use");
1114 // Cut off oop input and remove known instance id from type
1115 _igvn.rehash_node_delayed(use);
1116 use->as_InlineType()->set_oop(_igvn, _igvn.zerocon(T_OBJECT));
1117 const TypeOopPtr* toop = _igvn.type(use)->is_oopptr()->cast_to_instance_id(TypeOopPtr::InstanceBot);
1118 _igvn.set_type(use, toop);
1119 use->as_InlineType()->set_type(toop);
1120 // Process users
1121 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); k < kmax; k++) {
1122 Node* u = use->fast_out(k);
1123 if (!u->is_InlineType()) {
1124 worklist.push(u);
1125 }
1126 }
1127 } else if (use->Opcode() == Op_StoreX && use->in(MemNode::Address) == res) {
1128 // Store to mark word of inline type larval buffer
1129 assert(inline_alloc, "Unexpected store to mark word");
1130 _igvn.replace_node(use, use->in(MemNode::Memory));
1131 } else if (use->Opcode() == Op_MemBarRelease || use->Opcode() == Op_MemBarStoreStore) {
1132 // Inline type buffer allocations are followed by a membar
1133 assert(inline_alloc, "Unexpected MemBarRelease");
1134 use->as_MemBar()->remove(&_igvn);
1135 } else {
1136 eliminate_gc_barrier(use);
1137 }
1138 j -= (oc1 - res->outcnt());
1139 }
1140 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
1141 _igvn.remove_dead_node(res);
1142 }
1143
1144 //
1145 // Process other users of allocation's projections
1146 //
1147 if (_callprojs->resproj[0] != nullptr && _callprojs->resproj[0]->outcnt() != 0) {
1148 // First disconnect stores captured by Initialize node.
1149 // If Initialize node is eliminated first in the following code,
1150 // it will kill such stores and DUIterator_Last will assert.
1151 for (DUIterator_Fast jmax, j = _callprojs->resproj[0]->fast_outs(jmax); j < jmax; j++) {
1152 Node* use = _callprojs->resproj[0]->fast_out(j);
1153 if (use->is_AddP()) {
1154 // raw memory addresses used only by the initialization
1155 _igvn.replace_node(use, C->top());
1156 --j; --jmax;
1157 }
1158 }
1159 for (DUIterator_Last jmin, j = _callprojs->resproj[0]->last_outs(jmin); j >= jmin; ) {
1160 Node* use = _callprojs->resproj[0]->last_out(j);
1161 uint oc1 = _callprojs->resproj[0]->outcnt();
1162 if (use->is_Initialize()) {
1163 // Eliminate Initialize node.
1164 InitializeNode *init = use->as_Initialize();
1165 assert(init->outcnt() <= 2, "only a control and memory projection expected");
1166 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
1167 if (ctrl_proj != nullptr) {
1168 _igvn.replace_node(ctrl_proj, init->in(TypeFunc::Control));
1169 #ifdef ASSERT
1170 // If the InitializeNode has no memory out, it will die, and tmp will become null
1171 Node* tmp = init->in(TypeFunc::Control);
1172 assert(tmp == nullptr || tmp == _callprojs->fallthrough_catchproj, "allocation control projection");
1173 #endif
1174 }
1175 Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory);
1176 if (mem_proj != nullptr) {
1177 Node *mem = init->in(TypeFunc::Memory);
1178 #ifdef ASSERT
1179 if (mem->is_MergeMem()) {
1180 assert(mem->in(TypeFunc::Memory) == _callprojs->fallthrough_memproj, "allocation memory projection");
1181 } else {
1182 assert(mem == _callprojs->fallthrough_memproj, "allocation memory projection");
1183 }
1184 #endif
1185 _igvn.replace_node(mem_proj, mem);
1186 }
1187 } else if (use->Opcode() == Op_MemBarStoreStore) {
1188 // Inline type buffer allocations are followed by a membar
1189 assert(inline_alloc, "Unexpected MemBarStoreStore");
1190 use->as_MemBar()->remove(&_igvn);
1191 } else {
1192 assert(false, "only Initialize or AddP expected");
1193 }
1194 j -= (oc1 - _callprojs->resproj[0]->outcnt());
1195 }
1196 }
1197 if (_callprojs->fallthrough_catchproj != nullptr) {
1198 _igvn.replace_node(_callprojs->fallthrough_catchproj, alloc->in(TypeFunc::Control));
1199 }
1200 if (_callprojs->fallthrough_memproj != nullptr) {
1201 _igvn.replace_node(_callprojs->fallthrough_memproj, alloc->in(TypeFunc::Memory));
1202 }
1203 if (_callprojs->catchall_memproj != nullptr) {
1204 _igvn.replace_node(_callprojs->catchall_memproj, C->top());
1205 }
1206 if (_callprojs->fallthrough_ioproj != nullptr) {
1207 _igvn.replace_node(_callprojs->fallthrough_ioproj, alloc->in(TypeFunc::I_O));
1208 }
1209 if (_callprojs->catchall_ioproj != nullptr) {
1210 _igvn.replace_node(_callprojs->catchall_ioproj, C->top());
1211 }
1212 if (_callprojs->catchall_catchproj != nullptr) {
1213 _igvn.replace_node(_callprojs->catchall_catchproj, C->top());
1214 }
1215 }
1216
1217 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1218 // If reallocation fails during deoptimization we'll pop all
1219 // interpreter frames for this compiled frame and that won't play
1220 // nice with JVMTI popframe.
1221 // We avoid this issue by eager reallocation when the popframe request
1222 // is received.
1223 if (!EliminateAllocations) {
1224 return false;
1225 }
1226 Node* klass = alloc->in(AllocateNode::KlassNode);
1227 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1228
1229 // Attempt to eliminate inline type buffer allocations
1230 // regardless of usage and escape/replaceable status.
1231 bool inline_alloc = tklass->isa_instklassptr() &&
1232 tklass->is_instklassptr()->instance_klass()->is_inlinetype();
1233 if (!alloc->_is_non_escaping && !inline_alloc) {
1234 return false;
1235 }
1236 // Eliminate boxing allocations which are not used
1237 // regardless scalar replaceable status.
1238 Node* res = alloc->result_cast();
1239 bool boxing_alloc = (res == nullptr) && C->eliminate_boxing() &&
1240 tklass->isa_instklassptr() &&
1241 tklass->is_instklassptr()->instance_klass()->is_box_klass();
1242 if (!alloc->_is_scalar_replaceable && !boxing_alloc && !inline_alloc) {
1243 return false;
1244 }
1245
1246 _callprojs = alloc->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1247
1248 GrowableArray <SafePointNode *> safepoints;
1249 if (!can_eliminate_allocation(&_igvn, alloc, &safepoints)) {
1250 return false;
1251 }
1252
1253 if (!alloc->_is_scalar_replaceable) {
1254 assert(res == nullptr || inline_alloc, "sanity");
1255 // We can only eliminate allocation if all debug info references
1256 // are already replaced with SafePointScalarObject because
1257 // we can't search for a fields value without instance_id.
1258 if (safepoints.length() > 0) {
1259 assert(!inline_alloc || C->has_circular_inline_type(),
1260 "Inline type allocations should have been scalarized earlier");
1261 return false;
1262 }
1263 }
1264
1265 if (!scalar_replacement(alloc, safepoints)) {
1266 return false;
1267 }
1268
1269 CompileLog* log = C->log();
1270 if (log != nullptr) {
1271 log->head("eliminate_allocation type='%d'",
1272 log->identify(tklass->exact_klass()));
1273 JVMState* p = alloc->jvms();
1274 while (p != nullptr) {
1275 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1276 p = p->caller();
1277 }
1278 log->tail("eliminate_allocation");
1279 }
1280
1281 process_users_of_allocation(alloc, inline_alloc);
1282
1283 #ifndef PRODUCT
1284 if (PrintEliminateAllocations) {
1285 if (alloc->is_AllocateArray())
1286 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1287 else
1288 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1289 }
1290 #endif
1291
1292 return true;
1293 }
1294
1295 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1296 // EA should remove all uses of non-escaping boxing node.
1297 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != nullptr) {
1298 return false;
1299 }
1300
1301 assert(boxing->result_cast() == nullptr, "unexpected boxing node result");
1302
1303 _callprojs = boxing->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1304
1305 const TypeTuple* r = boxing->tf()->range_sig();
1306 assert(r->cnt() > TypeFunc::Parms, "sanity");
1307 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1308 assert(t != nullptr, "sanity");
1309
1310 CompileLog* log = C->log();
1311 if (log != nullptr) {
1312 log->head("eliminate_boxing type='%d'",
1313 log->identify(t->instance_klass()));
1314 JVMState* p = boxing->jvms();
1315 while (p != nullptr) {
1316 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1317 p = p->caller();
1318 }
1319 log->tail("eliminate_boxing");
1320 }
1321
1322 process_users_of_allocation(boxing);
1323
1324 #ifndef PRODUCT
1325 if (PrintEliminateAllocations) {
1326 tty->print("++++ Eliminated: %d ", boxing->_idx);
1327 boxing->method()->print_short_name(tty);
1328 tty->cr();
1329 }
1330 #endif
1331
1332 return true;
1333 }
1334
1335
1336 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1337 Node* adr = basic_plus_adr(base, offset);
1338 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1339 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered);
1340 transform_later(value);
1341 return value;
1342 }
1343
1344
1345 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1346 Node* adr = basic_plus_adr(base, offset);
1347 mem = StoreNode::make(_igvn, ctl, mem, adr, nullptr, value, bt, MemNode::unordered);
1348 transform_later(mem);
1349 return mem;
1350 }
1351
1352 //=============================================================================
1353 //
1354 // A L L O C A T I O N
1355 //
1356 // Allocation attempts to be fast in the case of frequent small objects.
1357 // It breaks down like this:
1358 //
1359 // 1) Size in doublewords is computed. This is a constant for objects and
1360 // variable for most arrays. Doubleword units are used to avoid size
1361 // overflow of huge doubleword arrays. We need doublewords in the end for
1362 // rounding.
1363 //
1364 // 2) Size is checked for being 'too large'. Too-large allocations will go
1365 // the slow path into the VM. The slow path can throw any required
1366 // exceptions, and does all the special checks for very large arrays. The
1367 // size test can constant-fold away for objects. For objects with
1368 // finalizers it constant-folds the otherway: you always go slow with
1369 // finalizers.
1370 //
1371 // 3) If NOT using TLABs, this is the contended loop-back point.
1372 // Load-Locked the heap top. If using TLABs normal-load the heap top.
1373 //
1374 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
1375 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
1376 // "size*8" we always enter the VM, where "largish" is a constant picked small
1377 // enough that there's always space between the eden max and 4Gig (old space is
1378 // there so it's quite large) and large enough that the cost of entering the VM
1379 // is dwarfed by the cost to initialize the space.
1380 //
1381 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1382 // down. If contended, repeat at step 3. If using TLABs normal-store
1383 // adjusted heap top back down; there is no contention.
1384 //
1385 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
1386 // fields.
1387 //
1388 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1389 // oop flavor.
1390 //
1391 //=============================================================================
1392 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1393 // Allocations bigger than this always go the slow route.
1394 // This value must be small enough that allocation attempts that need to
1395 // trigger exceptions go the slow route. Also, it must be small enough so
1396 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1397 //=============================================================================j//
1398 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1399 // The allocator will coalesce int->oop copies away. See comment in
1400 // coalesce.cpp about how this works. It depends critically on the exact
1401 // code shape produced here, so if you are changing this code shape
1402 // make sure the GC info for the heap-top is correct in and around the
1403 // slow-path call.
1404 //
1405
1406 void PhaseMacroExpand::expand_allocate_common(
1407 AllocateNode* alloc, // allocation node to be expanded
1408 Node* length, // array length for an array allocation
1409 const TypeFunc* slow_call_type, // Type of slow call
1410 address slow_call_address, // Address of slow call
1411 Node* valid_length_test // whether length is valid or not
1412 )
1413 {
1414 Node* ctrl = alloc->in(TypeFunc::Control);
1415 Node* mem = alloc->in(TypeFunc::Memory);
1416 Node* i_o = alloc->in(TypeFunc::I_O);
1417 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1418 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1419 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1420 assert(ctrl != nullptr, "must have control");
1421
1422 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1423 // they will not be used if "always_slow" is set
1424 enum { slow_result_path = 1, fast_result_path = 2 };
1425 Node *result_region = nullptr;
1426 Node *result_phi_rawmem = nullptr;
1427 Node *result_phi_rawoop = nullptr;
1428 Node *result_phi_i_o = nullptr;
1429
1430 // The initial slow comparison is a size check, the comparison
1431 // we want to do is a BoolTest::gt
1432 bool expand_fast_path = true;
1433 int tv = _igvn.find_int_con(initial_slow_test, -1);
1434 if (tv >= 0) {
1435 // InitialTest has constant result
1436 // 0 - can fit in TLAB
1437 // 1 - always too big or negative
1438 assert(tv <= 1, "0 or 1 if a constant");
1439 expand_fast_path = (tv == 0);
1440 initial_slow_test = nullptr;
1441 } else {
1442 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1443 }
1444
1445 if (!UseTLAB) {
1446 // Force slow-path allocation
1447 expand_fast_path = false;
1448 initial_slow_test = nullptr;
1449 }
1450
1451 bool allocation_has_use = (alloc->result_cast() != nullptr);
1452 if (!allocation_has_use) {
1453 InitializeNode* init = alloc->initialization();
1454 if (init != nullptr) {
1455 init->remove(&_igvn);
1456 }
1457 if (expand_fast_path && (initial_slow_test == nullptr)) {
1458 // Remove allocation node and return.
1459 // Size is a non-negative constant -> no initial check needed -> directly to fast path.
1460 // Also, no usages -> empty fast path -> no fall out to slow path -> nothing left.
1461 #ifndef PRODUCT
1462 if (PrintEliminateAllocations) {
1463 tty->print("NotUsed ");
1464 Node* res = alloc->proj_out_or_null(TypeFunc::Parms);
1465 if (res != nullptr) {
1466 res->dump();
1467 } else {
1468 alloc->dump();
1469 }
1470 }
1471 #endif
1472 yank_alloc_node(alloc);
1473 return;
1474 }
1475 }
1476
1477 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1478 Node *slow_region = nullptr;
1479 Node *toobig_false = ctrl;
1480
1481 // generate the initial test if necessary
1482 if (initial_slow_test != nullptr ) {
1483 assert (expand_fast_path, "Only need test if there is a fast path");
1484 slow_region = new RegionNode(3);
1485
1486 // Now make the initial failure test. Usually a too-big test but
1487 // might be a TRUE for finalizers.
1488 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1489 transform_later(toobig_iff);
1490 // Plug the failing-too-big test into the slow-path region
1491 Node* toobig_true = new IfTrueNode(toobig_iff);
1492 transform_later(toobig_true);
1493 slow_region ->init_req( too_big_or_final_path, toobig_true );
1494 toobig_false = new IfFalseNode(toobig_iff);
1495 transform_later(toobig_false);
1496 } else {
1497 // No initial test, just fall into next case
1498 assert(allocation_has_use || !expand_fast_path, "Should already have been handled");
1499 toobig_false = ctrl;
1500 debug_only(slow_region = NodeSentinel);
1501 }
1502
1503 // If we are here there are several possibilities
1504 // - expand_fast_path is false - then only a slow path is expanded. That's it.
1505 // no_initial_check means a constant allocation.
1506 // - If check always evaluates to false -> expand_fast_path is false (see above)
1507 // - If check always evaluates to true -> directly into fast path (but may bailout to slowpath)
1508 // if !allocation_has_use the fast path is empty
1509 // if !allocation_has_use && no_initial_check
1510 // - Then there are no fastpath that can fall out to slowpath -> no allocation code at all.
1511 // removed by yank_alloc_node above.
1512
1513 Node *slow_mem = mem; // save the current memory state for slow path
1514 // generate the fast allocation code unless we know that the initial test will always go slow
1515 if (expand_fast_path) {
1516 // Fast path modifies only raw memory.
1517 if (mem->is_MergeMem()) {
1518 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1519 }
1520
1521 // allocate the Region and Phi nodes for the result
1522 result_region = new RegionNode(3);
1523 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1524 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1525
1526 // Grab regular I/O before optional prefetch may change it.
1527 // Slow-path does no I/O so just set it to the original I/O.
1528 result_phi_i_o->init_req(slow_result_path, i_o);
1529
1530 // Name successful fast-path variables
1531 Node* fast_oop_ctrl;
1532 Node* fast_oop_rawmem;
1533
1534 if (allocation_has_use) {
1535 Node* needgc_ctrl = nullptr;
1536 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1537
1538 intx prefetch_lines = length != nullptr ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1539 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1540 Node* fast_oop = bs->obj_allocate(this, mem, toobig_false, size_in_bytes, i_o, needgc_ctrl,
1541 fast_oop_ctrl, fast_oop_rawmem,
1542 prefetch_lines);
1543
1544 if (initial_slow_test != nullptr) {
1545 // This completes all paths into the slow merge point
1546 slow_region->init_req(need_gc_path, needgc_ctrl);
1547 transform_later(slow_region);
1548 } else {
1549 // No initial slow path needed!
1550 // Just fall from the need-GC path straight into the VM call.
1551 slow_region = needgc_ctrl;
1552 }
1553
1554 InitializeNode* init = alloc->initialization();
1555 fast_oop_rawmem = initialize_object(alloc,
1556 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1557 klass_node, length, size_in_bytes);
1558 expand_initialize_membar(alloc, init, fast_oop_ctrl, fast_oop_rawmem);
1559 expand_dtrace_alloc_probe(alloc, fast_oop, fast_oop_ctrl, fast_oop_rawmem);
1560
1561 result_phi_rawoop->init_req(fast_result_path, fast_oop);
1562 } else {
1563 assert (initial_slow_test != nullptr, "sanity");
1564 fast_oop_ctrl = toobig_false;
1565 fast_oop_rawmem = mem;
1566 transform_later(slow_region);
1567 }
1568
1569 // Plug in the successful fast-path into the result merge point
1570 result_region ->init_req(fast_result_path, fast_oop_ctrl);
1571 result_phi_i_o ->init_req(fast_result_path, i_o);
1572 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1573 } else {
1574 slow_region = ctrl;
1575 result_phi_i_o = i_o; // Rename it to use in the following code.
1576 }
1577
1578 // Generate slow-path call
1579 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1580 OptoRuntime::stub_name(slow_call_address),
1581 TypePtr::BOTTOM);
1582 call->init_req(TypeFunc::Control, slow_region);
1583 call->init_req(TypeFunc::I_O, top()); // does no i/o
1584 call->init_req(TypeFunc::Memory, slow_mem); // may gc ptrs
1585 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1586 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1587
1588 call->init_req(TypeFunc::Parms+0, klass_node);
1589 if (length != nullptr) {
1590 call->init_req(TypeFunc::Parms+1, length);
1591 } else {
1592 // Let the runtime know if this is a larval allocation
1593 call->init_req(TypeFunc::Parms+1, _igvn.intcon(alloc->_larval));
1594 }
1595
1596 // Copy debug information and adjust JVMState information, then replace
1597 // allocate node with the call
1598 call->copy_call_debug_info(&_igvn, alloc);
1599 // For array allocations, copy the valid length check to the call node so Compile::final_graph_reshaping() can verify
1600 // that the call has the expected number of CatchProj nodes (in case the allocation always fails and the fallthrough
1601 // path dies).
1602 if (valid_length_test != nullptr) {
1603 call->add_req(valid_length_test);
1604 }
1605 if (expand_fast_path) {
1606 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1607 } else {
1608 // Hook i_o projection to avoid its elimination during allocation
1609 // replacement (when only a slow call is generated).
1610 call->set_req(TypeFunc::I_O, result_phi_i_o);
1611 }
1612 _igvn.replace_node(alloc, call);
1613 transform_later(call);
1614
1615 // Identify the output projections from the allocate node and
1616 // adjust any references to them.
1617 // The control and io projections look like:
1618 //
1619 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1620 // Allocate Catch
1621 // ^---Proj(io) <-------+ ^---CatchProj(io)
1622 //
1623 // We are interested in the CatchProj nodes.
1624 //
1625 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1626
1627 // An allocate node has separate memory projections for the uses on
1628 // the control and i_o paths. Replace the control memory projection with
1629 // result_phi_rawmem (unless we are only generating a slow call when
1630 // both memory projections are combined)
1631 if (expand_fast_path && _callprojs->fallthrough_memproj != nullptr) {
1632 _igvn.replace_in_uses(_callprojs->fallthrough_memproj, result_phi_rawmem);
1633 }
1634 // Now change uses of catchall_memproj to use fallthrough_memproj and delete
1635 // catchall_memproj so we end up with a call that has only 1 memory projection.
1636 if (_callprojs->catchall_memproj != nullptr) {
1637 if (_callprojs->fallthrough_memproj == nullptr) {
1638 _callprojs->fallthrough_memproj = new ProjNode(call, TypeFunc::Memory);
1639 transform_later(_callprojs->fallthrough_memproj);
1640 }
1641 _igvn.replace_in_uses(_callprojs->catchall_memproj, _callprojs->fallthrough_memproj);
1642 _igvn.remove_dead_node(_callprojs->catchall_memproj);
1643 }
1644
1645 // An allocate node has separate i_o projections for the uses on the control
1646 // and i_o paths. Always replace the control i_o projection with result i_o
1647 // otherwise incoming i_o become dead when only a slow call is generated
1648 // (it is different from memory projections where both projections are
1649 // combined in such case).
1650 if (_callprojs->fallthrough_ioproj != nullptr) {
1651 _igvn.replace_in_uses(_callprojs->fallthrough_ioproj, result_phi_i_o);
1652 }
1653 // Now change uses of catchall_ioproj to use fallthrough_ioproj and delete
1654 // catchall_ioproj so we end up with a call that has only 1 i_o projection.
1655 if (_callprojs->catchall_ioproj != nullptr) {
1656 if (_callprojs->fallthrough_ioproj == nullptr) {
1657 _callprojs->fallthrough_ioproj = new ProjNode(call, TypeFunc::I_O);
1658 transform_later(_callprojs->fallthrough_ioproj);
1659 }
1660 _igvn.replace_in_uses(_callprojs->catchall_ioproj, _callprojs->fallthrough_ioproj);
1661 _igvn.remove_dead_node(_callprojs->catchall_ioproj);
1662 }
1663
1664 // if we generated only a slow call, we are done
1665 if (!expand_fast_path) {
1666 // Now we can unhook i_o.
1667 if (result_phi_i_o->outcnt() > 1) {
1668 call->set_req(TypeFunc::I_O, top());
1669 } else {
1670 assert(result_phi_i_o->unique_ctrl_out() == call, "sanity");
1671 // Case of new array with negative size known during compilation.
1672 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1673 // following code since call to runtime will throw exception.
1674 // As result there will be no users of i_o after the call.
1675 // Leave i_o attached to this call to avoid problems in preceding graph.
1676 }
1677 return;
1678 }
1679
1680 if (_callprojs->fallthrough_catchproj != nullptr) {
1681 ctrl = _callprojs->fallthrough_catchproj->clone();
1682 transform_later(ctrl);
1683 _igvn.replace_node(_callprojs->fallthrough_catchproj, result_region);
1684 } else {
1685 ctrl = top();
1686 }
1687 Node *slow_result;
1688 if (_callprojs->resproj[0] == nullptr) {
1689 // no uses of the allocation result
1690 slow_result = top();
1691 } else {
1692 slow_result = _callprojs->resproj[0]->clone();
1693 transform_later(slow_result);
1694 _igvn.replace_node(_callprojs->resproj[0], result_phi_rawoop);
1695 }
1696
1697 // Plug slow-path into result merge point
1698 result_region->init_req( slow_result_path, ctrl);
1699 transform_later(result_region);
1700 if (allocation_has_use) {
1701 result_phi_rawoop->init_req(slow_result_path, slow_result);
1702 transform_later(result_phi_rawoop);
1703 }
1704 result_phi_rawmem->init_req(slow_result_path, _callprojs->fallthrough_memproj);
1705 transform_later(result_phi_rawmem);
1706 transform_later(result_phi_i_o);
1707 // This completes all paths into the result merge point
1708 }
1709
1710 // Remove alloc node that has no uses.
1711 void PhaseMacroExpand::yank_alloc_node(AllocateNode* alloc) {
1712 Node* ctrl = alloc->in(TypeFunc::Control);
1713 Node* mem = alloc->in(TypeFunc::Memory);
1714 Node* i_o = alloc->in(TypeFunc::I_O);
1715
1716 _callprojs = alloc->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1717 if (_callprojs->resproj[0] != nullptr) {
1718 for (DUIterator_Fast imax, i = _callprojs->resproj[0]->fast_outs(imax); i < imax; i++) {
1719 Node* use = _callprojs->resproj[0]->fast_out(i);
1720 use->isa_MemBar()->remove(&_igvn);
1721 --imax;
1722 --i; // back up iterator
1723 }
1724 assert(_callprojs->resproj[0]->outcnt() == 0, "all uses must be deleted");
1725 _igvn.remove_dead_node(_callprojs->resproj[0]);
1726 }
1727 if (_callprojs->fallthrough_catchproj != nullptr) {
1728 _igvn.replace_in_uses(_callprojs->fallthrough_catchproj, ctrl);
1729 _igvn.remove_dead_node(_callprojs->fallthrough_catchproj);
1730 }
1731 if (_callprojs->catchall_catchproj != nullptr) {
1732 _igvn.rehash_node_delayed(_callprojs->catchall_catchproj);
1733 _callprojs->catchall_catchproj->set_req(0, top());
1734 }
1735 if (_callprojs->fallthrough_proj != nullptr) {
1736 Node* catchnode = _callprojs->fallthrough_proj->unique_ctrl_out();
1737 _igvn.remove_dead_node(catchnode);
1738 _igvn.remove_dead_node(_callprojs->fallthrough_proj);
1739 }
1740 if (_callprojs->fallthrough_memproj != nullptr) {
1741 _igvn.replace_in_uses(_callprojs->fallthrough_memproj, mem);
1742 _igvn.remove_dead_node(_callprojs->fallthrough_memproj);
1743 }
1744 if (_callprojs->fallthrough_ioproj != nullptr) {
1745 _igvn.replace_in_uses(_callprojs->fallthrough_ioproj, i_o);
1746 _igvn.remove_dead_node(_callprojs->fallthrough_ioproj);
1747 }
1748 if (_callprojs->catchall_memproj != nullptr) {
1749 _igvn.rehash_node_delayed(_callprojs->catchall_memproj);
1750 _callprojs->catchall_memproj->set_req(0, top());
1751 }
1752 if (_callprojs->catchall_ioproj != nullptr) {
1753 _igvn.rehash_node_delayed(_callprojs->catchall_ioproj);
1754 _callprojs->catchall_ioproj->set_req(0, top());
1755 }
1756 #ifndef PRODUCT
1757 if (PrintEliminateAllocations) {
1758 if (alloc->is_AllocateArray()) {
1759 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1760 } else {
1761 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1762 }
1763 }
1764 #endif
1765 _igvn.remove_dead_node(alloc);
1766 }
1767
1768 void PhaseMacroExpand::expand_initialize_membar(AllocateNode* alloc, InitializeNode* init,
1769 Node*& fast_oop_ctrl, Node*& fast_oop_rawmem) {
1770 // If initialization is performed by an array copy, any required
1771 // MemBarStoreStore was already added. If the object does not
1772 // escape no need for a MemBarStoreStore. If the object does not
1773 // escape in its initializer and memory barrier (MemBarStoreStore or
1774 // stronger) is already added at exit of initializer, also no need
1775 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore
1776 // so that stores that initialize this object can't be reordered
1777 // with a subsequent store that makes this object accessible by
1778 // other threads.
1779 // Other threads include java threads and JVM internal threads
1780 // (for example concurrent GC threads). Current concurrent GC
1781 // implementation: G1 will not scan newly created object,
1782 // so it's safe to skip storestore barrier when allocation does
1783 // not escape.
1784 if (!alloc->does_not_escape_thread() &&
1785 !alloc->is_allocation_MemBar_redundant() &&
1786 (init == nullptr || !init->is_complete_with_arraycopy())) {
1787 if (init == nullptr || init->req() < InitializeNode::RawStores) {
1788 // No InitializeNode or no stores captured by zeroing
1789 // elimination. Simply add the MemBarStoreStore after object
1790 // initialization.
1791 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1792 transform_later(mb);
1793
1794 mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
1795 mb->init_req(TypeFunc::Control, fast_oop_ctrl);
1796 fast_oop_ctrl = new ProjNode(mb, TypeFunc::Control);
1797 transform_later(fast_oop_ctrl);
1798 fast_oop_rawmem = new ProjNode(mb, TypeFunc::Memory);
1799 transform_later(fast_oop_rawmem);
1800 } else {
1801 // Add the MemBarStoreStore after the InitializeNode so that
1802 // all stores performing the initialization that were moved
1803 // before the InitializeNode happen before the storestore
1804 // barrier.
1805
1806 Node* init_ctrl = init->proj_out_or_null(TypeFunc::Control);
1807 Node* init_mem = init->proj_out_or_null(TypeFunc::Memory);
1808
1809 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1810 transform_later(mb);
1811
1812 Node* ctrl = new ProjNode(init, TypeFunc::Control);
1813 transform_later(ctrl);
1814 Node* mem = new ProjNode(init, TypeFunc::Memory);
1815 transform_later(mem);
1816
1817 // The MemBarStoreStore depends on control and memory coming
1818 // from the InitializeNode
1819 mb->init_req(TypeFunc::Memory, mem);
1820 mb->init_req(TypeFunc::Control, ctrl);
1821
1822 ctrl = new ProjNode(mb, TypeFunc::Control);
1823 transform_later(ctrl);
1824 mem = new ProjNode(mb, TypeFunc::Memory);
1825 transform_later(mem);
1826
1827 // All nodes that depended on the InitializeNode for control
1828 // and memory must now depend on the MemBarNode that itself
1829 // depends on the InitializeNode
1830 if (init_ctrl != nullptr) {
1831 _igvn.replace_node(init_ctrl, ctrl);
1832 }
1833 if (init_mem != nullptr) {
1834 _igvn.replace_node(init_mem, mem);
1835 }
1836 }
1837 }
1838 }
1839
1840 void PhaseMacroExpand::expand_dtrace_alloc_probe(AllocateNode* alloc, Node* oop,
1841 Node*& ctrl, Node*& rawmem) {
1842 if (C->env()->dtrace_alloc_probes()) {
1843 // Slow-path call
1844 int size = TypeFunc::Parms + 2;
1845 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1846 CAST_FROM_FN_PTR(address,
1847 static_cast<int (*)(JavaThread*, oopDesc*)>(SharedRuntime::dtrace_object_alloc)),
1848 "dtrace_object_alloc",
1849 TypeRawPtr::BOTTOM);
1850
1851 // Get base of thread-local storage area
1852 Node* thread = new ThreadLocalNode();
1853 transform_later(thread);
1854
1855 call->init_req(TypeFunc::Parms + 0, thread);
1856 call->init_req(TypeFunc::Parms + 1, oop);
1857 call->init_req(TypeFunc::Control, ctrl);
1858 call->init_req(TypeFunc::I_O , top()); // does no i/o
1859 call->init_req(TypeFunc::Memory , rawmem);
1860 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1861 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1862 transform_later(call);
1863 ctrl = new ProjNode(call, TypeFunc::Control);
1864 transform_later(ctrl);
1865 rawmem = new ProjNode(call, TypeFunc::Memory);
1866 transform_later(rawmem);
1867 }
1868 }
1869
1870 // Helper for PhaseMacroExpand::expand_allocate_common.
1871 // Initializes the newly-allocated storage.
1872 Node* PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1873 Node* control, Node* rawmem, Node* object,
1874 Node* klass_node, Node* length,
1875 Node* size_in_bytes) {
1876 InitializeNode* init = alloc->initialization();
1877 // Store the klass & mark bits
1878 Node* mark_node = alloc->make_ideal_mark(&_igvn, control, rawmem);
1879 if (!mark_node->is_Con()) {
1880 transform_later(mark_node);
1881 }
1882 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, TypeX_X->basic_type());
1883
1884 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1885 int header_size = alloc->minimum_header_size(); // conservatively small
1886
1887 // Array length
1888 if (length != nullptr) { // Arrays need length field
1889 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1890 // conservatively small header size:
1891 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1892 if (_igvn.type(klass_node)->isa_aryklassptr()) { // we know the exact header size in most cases:
1893 BasicType elem = _igvn.type(klass_node)->is_klassptr()->as_instance_type()->isa_aryptr()->elem()->array_element_basic_type();
1894 if (is_reference_type(elem, true)) {
1895 elem = T_OBJECT;
1896 }
1897 header_size = Klass::layout_helper_header_size(Klass::array_layout_helper(elem));
1898 }
1899 }
1900
1901 // Clear the object body, if necessary.
1902 if (init == nullptr) {
1903 // The init has somehow disappeared; be cautious and clear everything.
1904 //
1905 // This can happen if a node is allocated but an uncommon trap occurs
1906 // immediately. In this case, the Initialize gets associated with the
1907 // trap, and may be placed in a different (outer) loop, if the Allocate
1908 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1909 // there can be two Allocates to one Initialize. The answer in all these
1910 // edge cases is safety first. It is always safe to clear immediately
1911 // within an Allocate, and then (maybe or maybe not) clear some more later.
1912 if (!(UseTLAB && ZeroTLAB)) {
1913 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1914 alloc->in(AllocateNode::DefaultValue),
1915 alloc->in(AllocateNode::RawDefaultValue),
1916 header_size, size_in_bytes,
1917 &_igvn);
1918 }
1919 } else {
1920 if (!init->is_complete()) {
1921 // Try to win by zeroing only what the init does not store.
1922 // We can also try to do some peephole optimizations,
1923 // such as combining some adjacent subword stores.
1924 rawmem = init->complete_stores(control, rawmem, object,
1925 header_size, size_in_bytes, &_igvn);
1926 }
1927 // We have no more use for this link, since the AllocateNode goes away:
1928 init->set_req(InitializeNode::RawAddress, top());
1929 // (If we keep the link, it just confuses the register allocator,
1930 // who thinks he sees a real use of the address by the membar.)
1931 }
1932
1933 return rawmem;
1934 }
1935
1936 // Generate prefetch instructions for next allocations.
1937 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1938 Node*& contended_phi_rawmem,
1939 Node* old_eden_top, Node* new_eden_top,
1940 intx lines) {
1941 enum { fall_in_path = 1, pf_path = 2 };
1942 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1943 // Generate prefetch allocation with watermark check.
1944 // As an allocation hits the watermark, we will prefetch starting
1945 // at a "distance" away from watermark.
1946
1947 Node *pf_region = new RegionNode(3);
1948 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY,
1949 TypeRawPtr::BOTTOM );
1950 // I/O is used for Prefetch
1951 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO );
1952
1953 Node *thread = new ThreadLocalNode();
1954 transform_later(thread);
1955
1956 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread,
1957 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1958 transform_later(eden_pf_adr);
1959
1960 Node *old_pf_wm = new LoadPNode(needgc_false,
1961 contended_phi_rawmem, eden_pf_adr,
1962 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM,
1963 MemNode::unordered);
1964 transform_later(old_pf_wm);
1965
1966 // check against new_eden_top
1967 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm );
1968 transform_later(need_pf_cmp);
1969 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge );
1970 transform_later(need_pf_bol);
1971 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol,
1972 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1973 transform_later(need_pf_iff);
1974
1975 // true node, add prefetchdistance
1976 Node *need_pf_true = new IfTrueNode( need_pf_iff );
1977 transform_later(need_pf_true);
1978
1979 Node *need_pf_false = new IfFalseNode( need_pf_iff );
1980 transform_later(need_pf_false);
1981
1982 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm,
1983 _igvn.MakeConX(AllocatePrefetchDistance) );
1984 transform_later(new_pf_wmt );
1985 new_pf_wmt->set_req(0, need_pf_true);
1986
1987 Node *store_new_wmt = new StorePNode(need_pf_true,
1988 contended_phi_rawmem, eden_pf_adr,
1989 TypeRawPtr::BOTTOM, new_pf_wmt,
1990 MemNode::unordered);
1991 transform_later(store_new_wmt);
1992
1993 // adding prefetches
1994 pf_phi_abio->init_req( fall_in_path, i_o );
1995
1996 Node *prefetch_adr;
1997 Node *prefetch;
1998 uint step_size = AllocatePrefetchStepSize;
1999 uint distance = 0;
2000
2001 for ( intx i = 0; i < lines; i++ ) {
2002 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt,
2003 _igvn.MakeConX(distance) );
2004 transform_later(prefetch_adr);
2005 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
2006 transform_later(prefetch);
2007 distance += step_size;
2008 i_o = prefetch;
2009 }
2010 pf_phi_abio->set_req( pf_path, i_o );
2011
2012 pf_region->init_req( fall_in_path, need_pf_false );
2013 pf_region->init_req( pf_path, need_pf_true );
2014
2015 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
2016 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
2017
2018 transform_later(pf_region);
2019 transform_later(pf_phi_rawmem);
2020 transform_later(pf_phi_abio);
2021
2022 needgc_false = pf_region;
2023 contended_phi_rawmem = pf_phi_rawmem;
2024 i_o = pf_phi_abio;
2025 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) {
2026 // Insert a prefetch instruction for each allocation.
2027 // This code is used to generate 1 prefetch instruction per cache line.
2028
2029 // Generate several prefetch instructions.
2030 uint step_size = AllocatePrefetchStepSize;
2031 uint distance = AllocatePrefetchDistance;
2032
2033 // Next cache address.
2034 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top,
2035 _igvn.MakeConX(step_size + distance));
2036 transform_later(cache_adr);
2037 cache_adr = new CastP2XNode(needgc_false, cache_adr);
2038 transform_later(cache_adr);
2039 // Address is aligned to execute prefetch to the beginning of cache line size
2040 // (it is important when BIS instruction is used on SPARC as prefetch).
2041 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
2042 cache_adr = new AndXNode(cache_adr, mask);
2043 transform_later(cache_adr);
2044 cache_adr = new CastX2PNode(cache_adr);
2045 transform_later(cache_adr);
2046
2047 // Prefetch
2048 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr );
2049 prefetch->set_req(0, needgc_false);
2050 transform_later(prefetch);
2051 contended_phi_rawmem = prefetch;
2052 Node *prefetch_adr;
2053 distance = step_size;
2054 for ( intx i = 1; i < lines; i++ ) {
2055 prefetch_adr = new AddPNode( cache_adr, cache_adr,
2056 _igvn.MakeConX(distance) );
2057 transform_later(prefetch_adr);
2058 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr );
2059 transform_later(prefetch);
2060 distance += step_size;
2061 contended_phi_rawmem = prefetch;
2062 }
2063 } else if( AllocatePrefetchStyle > 0 ) {
2064 // Insert a prefetch for each allocation only on the fast-path
2065 Node *prefetch_adr;
2066 Node *prefetch;
2067 // Generate several prefetch instructions.
2068 uint step_size = AllocatePrefetchStepSize;
2069 uint distance = AllocatePrefetchDistance;
2070 for ( intx i = 0; i < lines; i++ ) {
2071 prefetch_adr = new AddPNode( old_eden_top, new_eden_top,
2072 _igvn.MakeConX(distance) );
2073 transform_later(prefetch_adr);
2074 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
2075 // Do not let it float too high, since if eden_top == eden_end,
2076 // both might be null.
2077 if( i == 0 ) { // Set control for first prefetch, next follows it
2078 prefetch->init_req(0, needgc_false);
2079 }
2080 transform_later(prefetch);
2081 distance += step_size;
2082 i_o = prefetch;
2083 }
2084 }
2085 return i_o;
2086 }
2087
2088
2089 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
2090 expand_allocate_common(alloc, nullptr,
2091 OptoRuntime::new_instance_Type(),
2092 OptoRuntime::new_instance_Java(), nullptr);
2093 }
2094
2095 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
2096 Node* length = alloc->in(AllocateNode::ALength);
2097 Node* valid_length_test = alloc->in(AllocateNode::ValidLengthTest);
2098 InitializeNode* init = alloc->initialization();
2099 Node* klass_node = alloc->in(AllocateNode::KlassNode);
2100 const TypeAryKlassPtr* ary_klass_t = _igvn.type(klass_node)->isa_aryklassptr();
2101 address slow_call_address; // Address of slow call
2102 if (init != nullptr && init->is_complete_with_arraycopy() &&
2103 ary_klass_t && ary_klass_t->elem()->isa_klassptr() == nullptr) {
2104 // Don't zero type array during slow allocation in VM since
2105 // it will be initialized later by arraycopy in compiled code.
2106 slow_call_address = OptoRuntime::new_array_nozero_Java();
2107 } else {
2108 slow_call_address = OptoRuntime::new_array_Java();
2109 }
2110 expand_allocate_common(alloc, length,
2111 OptoRuntime::new_array_Type(),
2112 slow_call_address, valid_length_test);
2113 }
2114
2115 //-------------------mark_eliminated_box----------------------------------
2116 //
2117 // During EA obj may point to several objects but after few ideal graph
2118 // transformations (CCP) it may point to only one non escaping object
2119 // (but still using phi), corresponding locks and unlocks will be marked
2120 // for elimination. Later obj could be replaced with a new node (new phi)
2121 // and which does not have escape information. And later after some graph
2122 // reshape other locks and unlocks (which were not marked for elimination
2123 // before) are connected to this new obj (phi) but they still will not be
2124 // marked for elimination since new obj has no escape information.
2125 // Mark all associated (same box and obj) lock and unlock nodes for
2126 // elimination if some of them marked already.
2127 void PhaseMacroExpand::mark_eliminated_box(Node* box, Node* obj) {
2128 BoxLockNode* oldbox = box->as_BoxLock();
2129 if (oldbox->is_eliminated()) {
2130 return; // This BoxLock node was processed already.
2131 }
2132 assert(!oldbox->is_unbalanced(), "this should not be called for unbalanced region");
2133 // New implementation (EliminateNestedLocks) has separate BoxLock
2134 // node for each locked region so mark all associated locks/unlocks as
2135 // eliminated even if different objects are referenced in one locked region
2136 // (for example, OSR compilation of nested loop inside locked scope).
2137 if (EliminateNestedLocks ||
2138 oldbox->as_BoxLock()->is_simple_lock_region(nullptr, obj, nullptr)) {
2139 // Box is used only in one lock region. Mark this box as eliminated.
2140 oldbox->set_local(); // This verifies correct state of BoxLock
2141 _igvn.hash_delete(oldbox);
2142 oldbox->set_eliminated(); // This changes box's hash value
2143 _igvn.hash_insert(oldbox);
2144
2145 for (uint i = 0; i < oldbox->outcnt(); i++) {
2146 Node* u = oldbox->raw_out(i);
2147 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
2148 AbstractLockNode* alock = u->as_AbstractLock();
2149 // Check lock's box since box could be referenced by Lock's debug info.
2150 if (alock->box_node() == oldbox) {
2151 // Mark eliminated all related locks and unlocks.
2152 #ifdef ASSERT
2153 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4");
2154 #endif
2155 alock->set_non_esc_obj();
2156 }
2157 }
2158 }
2159 return;
2160 }
2161
2162 // Create new "eliminated" BoxLock node and use it in monitor debug info
2163 // instead of oldbox for the same object.
2164 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
2165
2166 // Note: BoxLock node is marked eliminated only here and it is used
2167 // to indicate that all associated lock and unlock nodes are marked
2168 // for elimination.
2169 newbox->set_local(); // This verifies correct state of BoxLock
2170 newbox->set_eliminated();
2171 transform_later(newbox);
2172
2173 // Replace old box node with new box for all users of the same object.
2174 for (uint i = 0; i < oldbox->outcnt();) {
2175 bool next_edge = true;
2176
2177 Node* u = oldbox->raw_out(i);
2178 if (u->is_AbstractLock()) {
2179 AbstractLockNode* alock = u->as_AbstractLock();
2180 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
2181 // Replace Box and mark eliminated all related locks and unlocks.
2182 #ifdef ASSERT
2183 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5");
2184 #endif
2185 alock->set_non_esc_obj();
2186 _igvn.rehash_node_delayed(alock);
2187 alock->set_box_node(newbox);
2188 next_edge = false;
2189 }
2190 }
2191 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
2192 FastLockNode* flock = u->as_FastLock();
2193 assert(flock->box_node() == oldbox, "sanity");
2194 _igvn.rehash_node_delayed(flock);
2195 flock->set_box_node(newbox);
2196 next_edge = false;
2197 }
2198
2199 // Replace old box in monitor debug info.
2200 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
2201 SafePointNode* sfn = u->as_SafePoint();
2202 JVMState* youngest_jvms = sfn->jvms();
2203 int max_depth = youngest_jvms->depth();
2204 for (int depth = 1; depth <= max_depth; depth++) {
2205 JVMState* jvms = youngest_jvms->of_depth(depth);
2206 int num_mon = jvms->nof_monitors();
2207 // Loop over monitors
2208 for (int idx = 0; idx < num_mon; idx++) {
2209 Node* obj_node = sfn->monitor_obj(jvms, idx);
2210 Node* box_node = sfn->monitor_box(jvms, idx);
2211 if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
2212 int j = jvms->monitor_box_offset(idx);
2213 _igvn.replace_input_of(u, j, newbox);
2214 next_edge = false;
2215 }
2216 }
2217 }
2218 }
2219 if (next_edge) i++;
2220 }
2221 }
2222
2223 //-----------------------mark_eliminated_locking_nodes-----------------------
2224 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
2225 if (!alock->is_balanced()) {
2226 return; // Can't do any more elimination for this locking region
2227 }
2228 if (EliminateNestedLocks) {
2229 if (alock->is_nested()) {
2230 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
2231 return;
2232 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
2233 // Only Lock node has JVMState needed here.
2234 // Not that preceding claim is documented anywhere else.
2235 if (alock->jvms() != nullptr) {
2236 if (alock->as_Lock()->is_nested_lock_region()) {
2237 // Mark eliminated related nested locks and unlocks.
2238 Node* obj = alock->obj_node();
2239 BoxLockNode* box_node = alock->box_node()->as_BoxLock();
2240 assert(!box_node->is_eliminated(), "should not be marked yet");
2241 // Note: BoxLock node is marked eliminated only here
2242 // and it is used to indicate that all associated lock
2243 // and unlock nodes are marked for elimination.
2244 box_node->set_eliminated(); // Box's hash is always NO_HASH here
2245 for (uint i = 0; i < box_node->outcnt(); i++) {
2246 Node* u = box_node->raw_out(i);
2247 if (u->is_AbstractLock()) {
2248 alock = u->as_AbstractLock();
2249 if (alock->box_node() == box_node) {
2250 // Verify that this Box is referenced only by related locks.
2251 assert(alock->obj_node()->eqv_uncast(obj), "");
2252 // Mark all related locks and unlocks.
2253 #ifdef ASSERT
2254 alock->log_lock_optimization(C, "eliminate_lock_set_nested");
2255 #endif
2256 alock->set_nested();
2257 }
2258 }
2259 }
2260 } else {
2261 #ifdef ASSERT
2262 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region");
2263 if (C->log() != nullptr)
2264 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output
2265 #endif
2266 }
2267 }
2268 return;
2269 }
2270 // Process locks for non escaping object
2271 assert(alock->is_non_esc_obj(), "");
2272 } // EliminateNestedLocks
2273
2274 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2275 // Look for all locks of this object and mark them and
2276 // corresponding BoxLock nodes as eliminated.
2277 Node* obj = alock->obj_node();
2278 for (uint j = 0; j < obj->outcnt(); j++) {
2279 Node* o = obj->raw_out(j);
2280 if (o->is_AbstractLock() &&
2281 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2282 alock = o->as_AbstractLock();
2283 Node* box = alock->box_node();
2284 // Replace old box node with new eliminated box for all users
2285 // of the same object and mark related locks as eliminated.
2286 mark_eliminated_box(box, obj);
2287 }
2288 }
2289 }
2290 }
2291
2292 // we have determined that this lock/unlock can be eliminated, we simply
2293 // eliminate the node without expanding it.
2294 //
2295 // Note: The membar's associated with the lock/unlock are currently not
2296 // eliminated. This should be investigated as a future enhancement.
2297 //
2298 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2299
2300 if (!alock->is_eliminated()) {
2301 return false;
2302 }
2303 #ifdef ASSERT
2304 if (!alock->is_coarsened()) {
2305 // Check that new "eliminated" BoxLock node is created.
2306 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2307 assert(oldbox->is_eliminated(), "should be done already");
2308 }
2309 #endif
2310
2311 alock->log_lock_optimization(C, "eliminate_lock");
2312
2313 #ifndef PRODUCT
2314 if (PrintEliminateLocks) {
2315 tty->print_cr("++++ Eliminated: %d %s '%s'", alock->_idx, (alock->is_Lock() ? "Lock" : "Unlock"), alock->kind_as_string());
2316 }
2317 #endif
2318
2319 Node* mem = alock->in(TypeFunc::Memory);
2320 Node* ctrl = alock->in(TypeFunc::Control);
2321 guarantee(ctrl != nullptr, "missing control projection, cannot replace_node() with null");
2322
2323 _callprojs = alock->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2324 // There are 2 projections from the lock. The lock node will
2325 // be deleted when its last use is subsumed below.
2326 assert(alock->outcnt() == 2 &&
2327 _callprojs->fallthrough_proj != nullptr &&
2328 _callprojs->fallthrough_memproj != nullptr,
2329 "Unexpected projections from Lock/Unlock");
2330
2331 Node* fallthroughproj = _callprojs->fallthrough_proj;
2332 Node* memproj_fallthrough = _callprojs->fallthrough_memproj;
2333
2334 // The memory projection from a lock/unlock is RawMem
2335 // The input to a Lock is merged memory, so extract its RawMem input
2336 // (unless the MergeMem has been optimized away.)
2337 if (alock->is_Lock()) {
2338 // Search for MemBarAcquireLock node and delete it also.
2339 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2340 assert(membar != nullptr && membar->Opcode() == Op_MemBarAcquireLock, "");
2341 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2342 Node* memproj = membar->proj_out(TypeFunc::Memory);
2343 _igvn.replace_node(ctrlproj, fallthroughproj);
2344 _igvn.replace_node(memproj, memproj_fallthrough);
2345
2346 // Delete FastLock node also if this Lock node is unique user
2347 // (a loop peeling may clone a Lock node).
2348 Node* flock = alock->as_Lock()->fastlock_node();
2349 if (flock->outcnt() == 1) {
2350 assert(flock->unique_out() == alock, "sanity");
2351 _igvn.replace_node(flock, top());
2352 }
2353 }
2354
2355 // Search for MemBarReleaseLock node and delete it also.
2356 if (alock->is_Unlock() && ctrl->is_Proj() && ctrl->in(0)->is_MemBar()) {
2357 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2358 assert(membar->Opcode() == Op_MemBarReleaseLock &&
2359 mem->is_Proj() && membar == mem->in(0), "");
2360 _igvn.replace_node(fallthroughproj, ctrl);
2361 _igvn.replace_node(memproj_fallthrough, mem);
2362 fallthroughproj = ctrl;
2363 memproj_fallthrough = mem;
2364 ctrl = membar->in(TypeFunc::Control);
2365 mem = membar->in(TypeFunc::Memory);
2366 }
2367
2368 _igvn.replace_node(fallthroughproj, ctrl);
2369 _igvn.replace_node(memproj_fallthrough, mem);
2370 return true;
2371 }
2372
2373
2374 //------------------------------expand_lock_node----------------------
2375 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2376
2377 Node* ctrl = lock->in(TypeFunc::Control);
2378 Node* mem = lock->in(TypeFunc::Memory);
2379 Node* obj = lock->obj_node();
2380 Node* box = lock->box_node();
2381 Node* flock = lock->fastlock_node();
2382
2383 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2384
2385 // Make the merge point
2386 Node *region;
2387 Node *mem_phi;
2388 Node *slow_path;
2389
2390 region = new RegionNode(3);
2391 // create a Phi for the memory state
2392 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2393
2394 // Optimize test; set region slot 2
2395 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2396 mem_phi->init_req(2, mem);
2397
2398 // Make slow path call
2399 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(),
2400 OptoRuntime::complete_monitor_locking_Java(), nullptr, slow_path,
2401 obj, box, nullptr);
2402
2403 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2404
2405 // Slow path can only throw asynchronous exceptions, which are always
2406 // de-opted. So the compiler thinks the slow-call can never throw an
2407 // exception. If it DOES throw an exception we would need the debug
2408 // info removed first (since if it throws there is no monitor).
2409 assert(_callprojs->fallthrough_ioproj == nullptr && _callprojs->catchall_ioproj == nullptr &&
2410 _callprojs->catchall_memproj == nullptr && _callprojs->catchall_catchproj == nullptr, "Unexpected projection from Lock");
2411
2412 // Capture slow path
2413 // disconnect fall-through projection from call and create a new one
2414 // hook up users of fall-through projection to region
2415 Node *slow_ctrl = _callprojs->fallthrough_proj->clone();
2416 transform_later(slow_ctrl);
2417 _igvn.hash_delete(_callprojs->fallthrough_proj);
2418 _callprojs->fallthrough_proj->disconnect_inputs(C);
2419 region->init_req(1, slow_ctrl);
2420 // region inputs are now complete
2421 transform_later(region);
2422 _igvn.replace_node(_callprojs->fallthrough_proj, region);
2423
2424 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2425
2426 mem_phi->init_req(1, memproj);
2427
2428 transform_later(mem_phi);
2429
2430 _igvn.replace_node(_callprojs->fallthrough_memproj, mem_phi);
2431 }
2432
2433 //------------------------------expand_unlock_node----------------------
2434 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2435
2436 Node* ctrl = unlock->in(TypeFunc::Control);
2437 Node* mem = unlock->in(TypeFunc::Memory);
2438 Node* obj = unlock->obj_node();
2439 Node* box = unlock->box_node();
2440
2441 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2442
2443 // No need for a null check on unlock
2444
2445 // Make the merge point
2446 Node *region;
2447 Node *mem_phi;
2448
2449 region = new RegionNode(3);
2450 // create a Phi for the memory state
2451 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2452
2453 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2454 funlock = transform_later( funlock )->as_FastUnlock();
2455 // Optimize test; set region slot 2
2456 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2457 Node *thread = transform_later(new ThreadLocalNode());
2458
2459 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2460 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2461 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2462
2463 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2464 assert(_callprojs->fallthrough_ioproj == nullptr && _callprojs->catchall_ioproj == nullptr &&
2465 _callprojs->catchall_memproj == nullptr && _callprojs->catchall_catchproj == nullptr, "Unexpected projection from Lock");
2466
2467 // No exceptions for unlocking
2468 // Capture slow path
2469 // disconnect fall-through projection from call and create a new one
2470 // hook up users of fall-through projection to region
2471 Node *slow_ctrl = _callprojs->fallthrough_proj->clone();
2472 transform_later(slow_ctrl);
2473 _igvn.hash_delete(_callprojs->fallthrough_proj);
2474 _callprojs->fallthrough_proj->disconnect_inputs(C);
2475 region->init_req(1, slow_ctrl);
2476 // region inputs are now complete
2477 transform_later(region);
2478 _igvn.replace_node(_callprojs->fallthrough_proj, region);
2479
2480 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2481 mem_phi->init_req(1, memproj );
2482 mem_phi->init_req(2, mem);
2483 transform_later(mem_phi);
2484
2485 _igvn.replace_node(_callprojs->fallthrough_memproj, mem_phi);
2486 }
2487
2488 // An inline type might be returned from the call but we don't know its
2489 // type. Either we get a buffered inline type (and nothing needs to be done)
2490 // or one of the values being returned is the klass of the inline type
2491 // and we need to allocate an inline type instance of that type and
2492 // initialize it with other values being returned. In that case, we
2493 // first try a fast path allocation and initialize the value with the
2494 // inline klass's pack handler or we fall back to a runtime call.
2495 void PhaseMacroExpand::expand_mh_intrinsic_return(CallStaticJavaNode* call) {
2496 assert(call->method()->is_method_handle_intrinsic(), "must be a method handle intrinsic call");
2497 Node* ret = call->proj_out_or_null(TypeFunc::Parms);
2498 if (ret == nullptr) {
2499 return;
2500 }
2501 const TypeFunc* tf = call->_tf;
2502 const TypeTuple* domain = OptoRuntime::store_inline_type_fields_Type()->domain_cc();
2503 const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain);
2504 call->_tf = new_tf;
2505 // Make sure the change of type is applied before projections are processed by igvn
2506 _igvn.set_type(call, call->Value(&_igvn));
2507 _igvn.set_type(ret, ret->Value(&_igvn));
2508
2509 // Before any new projection is added:
2510 CallProjections* projs = call->extract_projections(true, true);
2511
2512 // Create temporary hook nodes that will be replaced below.
2513 // Add an input to prevent hook nodes from being dead.
2514 Node* ctl = new Node(call);
2515 Node* mem = new Node(ctl);
2516 Node* io = new Node(ctl);
2517 Node* ex_ctl = new Node(ctl);
2518 Node* ex_mem = new Node(ctl);
2519 Node* ex_io = new Node(ctl);
2520 Node* res = new Node(ctl);
2521
2522 // Allocate a new buffered inline type only if a new one is not returned
2523 Node* cast = transform_later(new CastP2XNode(ctl, res));
2524 Node* mask = MakeConX(0x1);
2525 Node* masked = transform_later(new AndXNode(cast, mask));
2526 Node* cmp = transform_later(new CmpXNode(masked, mask));
2527 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
2528 IfNode* allocation_iff = new IfNode(ctl, bol, PROB_MAX, COUNT_UNKNOWN);
2529 transform_later(allocation_iff);
2530 Node* allocation_ctl = transform_later(new IfTrueNode(allocation_iff));
2531 Node* no_allocation_ctl = transform_later(new IfFalseNode(allocation_iff));
2532 Node* no_allocation_res = transform_later(new CheckCastPPNode(no_allocation_ctl, res, TypeInstPtr::BOTTOM));
2533
2534 // Try to allocate a new buffered inline instance either from TLAB or eden space
2535 Node* needgc_ctrl = nullptr; // needgc means slowcase, i.e. allocation failed
2536 CallLeafNoFPNode* handler_call;
2537 const bool alloc_in_place = UseTLAB;
2538 if (alloc_in_place) {
2539 Node* fast_oop_ctrl = nullptr;
2540 Node* fast_oop_rawmem = nullptr;
2541 Node* mask2 = MakeConX(-2);
2542 Node* masked2 = transform_later(new AndXNode(cast, mask2));
2543 Node* rawklassptr = transform_later(new CastX2PNode(masked2));
2544 Node* klass_node = transform_later(new CheckCastPPNode(allocation_ctl, rawklassptr, TypeInstKlassPtr::OBJECT_OR_NULL));
2545 Node* layout_val = make_load(nullptr, mem, klass_node, in_bytes(Klass::layout_helper_offset()), TypeInt::INT, T_INT);
2546 Node* size_in_bytes = ConvI2X(layout_val);
2547 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2548 Node* fast_oop = bs->obj_allocate(this, mem, allocation_ctl, size_in_bytes, io, needgc_ctrl,
2549 fast_oop_ctrl, fast_oop_rawmem,
2550 AllocateInstancePrefetchLines);
2551 // Allocation succeed, initialize buffered inline instance header firstly,
2552 // and then initialize its fields with an inline class specific handler
2553 Node* mark_node = makecon(TypeRawPtr::make((address)markWord::inline_type_prototype().value()));
2554 fast_oop_rawmem = make_store(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
2555 fast_oop_rawmem = make_store(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
2556 if (UseCompressedClassPointers) {
2557 fast_oop_rawmem = make_store(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::klass_gap_offset_in_bytes(), intcon(0), T_INT);
2558 }
2559 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);
2560 Node* pack_handler = make_load(fast_oop_ctrl, fast_oop_rawmem, fixed_block, in_bytes(InlineKlass::pack_handler_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2561 handler_call = new CallLeafNoFPNode(OptoRuntime::pack_inline_type_Type(),
2562 nullptr,
2563 "pack handler",
2564 TypeRawPtr::BOTTOM);
2565 handler_call->init_req(TypeFunc::Control, fast_oop_ctrl);
2566 handler_call->init_req(TypeFunc::Memory, fast_oop_rawmem);
2567 handler_call->init_req(TypeFunc::I_O, top());
2568 handler_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2569 handler_call->init_req(TypeFunc::ReturnAdr, top());
2570 handler_call->init_req(TypeFunc::Parms, pack_handler);
2571 handler_call->init_req(TypeFunc::Parms+1, fast_oop);
2572 } else {
2573 needgc_ctrl = allocation_ctl;
2574 }
2575
2576 // Allocation failed, fall back to a runtime call
2577 CallStaticJavaNode* slow_call = new CallStaticJavaNode(OptoRuntime::store_inline_type_fields_Type(),
2578 StubRoutines::store_inline_type_fields_to_buf(),
2579 "store_inline_type_fields",
2580 TypePtr::BOTTOM);
2581 slow_call->init_req(TypeFunc::Control, needgc_ctrl);
2582 slow_call->init_req(TypeFunc::Memory, mem);
2583 slow_call->init_req(TypeFunc::I_O, io);
2584 slow_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2585 slow_call->init_req(TypeFunc::ReturnAdr, call->in(TypeFunc::ReturnAdr));
2586 slow_call->init_req(TypeFunc::Parms, res);
2587
2588 Node* slow_ctl = transform_later(new ProjNode(slow_call, TypeFunc::Control));
2589 Node* slow_mem = transform_later(new ProjNode(slow_call, TypeFunc::Memory));
2590 Node* slow_io = transform_later(new ProjNode(slow_call, TypeFunc::I_O));
2591 Node* slow_res = transform_later(new ProjNode(slow_call, TypeFunc::Parms));
2592 Node* slow_catc = transform_later(new CatchNode(slow_ctl, slow_io, 2));
2593 Node* slow_norm = transform_later(new CatchProjNode(slow_catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci));
2594 Node* slow_excp = transform_later(new CatchProjNode(slow_catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci));
2595
2596 Node* ex_r = new RegionNode(3);
2597 Node* ex_mem_phi = new PhiNode(ex_r, Type::MEMORY, TypePtr::BOTTOM);
2598 Node* ex_io_phi = new PhiNode(ex_r, Type::ABIO);
2599 ex_r->init_req(1, slow_excp);
2600 ex_mem_phi->init_req(1, slow_mem);
2601 ex_io_phi->init_req(1, slow_io);
2602 ex_r->init_req(2, ex_ctl);
2603 ex_mem_phi->init_req(2, ex_mem);
2604 ex_io_phi->init_req(2, ex_io);
2605 transform_later(ex_r);
2606 transform_later(ex_mem_phi);
2607 transform_later(ex_io_phi);
2608
2609 // We don't know how many values are returned. This assumes the
2610 // worst case, that all available registers are used.
2611 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2612 if (domain->field_at(i) == Type::HALF) {
2613 slow_call->init_req(i, top());
2614 if (alloc_in_place) {
2615 handler_call->init_req(i+1, top());
2616 }
2617 continue;
2618 }
2619 Node* proj = transform_later(new ProjNode(call, i));
2620 slow_call->init_req(i, proj);
2621 if (alloc_in_place) {
2622 handler_call->init_req(i+1, proj);
2623 }
2624 }
2625 // We can safepoint at that new call
2626 slow_call->copy_call_debug_info(&_igvn, call);
2627 transform_later(slow_call);
2628 if (alloc_in_place) {
2629 transform_later(handler_call);
2630 }
2631
2632 Node* fast_ctl = nullptr;
2633 Node* fast_res = nullptr;
2634 MergeMemNode* fast_mem = nullptr;
2635 if (alloc_in_place) {
2636 fast_ctl = transform_later(new ProjNode(handler_call, TypeFunc::Control));
2637 Node* rawmem = transform_later(new ProjNode(handler_call, TypeFunc::Memory));
2638 fast_res = transform_later(new ProjNode(handler_call, TypeFunc::Parms));
2639 fast_mem = MergeMemNode::make(mem);
2640 fast_mem->set_memory_at(Compile::AliasIdxRaw, rawmem);
2641 transform_later(fast_mem);
2642 }
2643
2644 Node* r = new RegionNode(alloc_in_place ? 4 : 3);
2645 Node* mem_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2646 Node* io_phi = new PhiNode(r, Type::ABIO);
2647 Node* res_phi = new PhiNode(r, TypeInstPtr::BOTTOM);
2648 r->init_req(1, no_allocation_ctl);
2649 mem_phi->init_req(1, mem);
2650 io_phi->init_req(1, io);
2651 res_phi->init_req(1, no_allocation_res);
2652 r->init_req(2, slow_norm);
2653 mem_phi->init_req(2, slow_mem);
2654 io_phi->init_req(2, slow_io);
2655 res_phi->init_req(2, slow_res);
2656 if (alloc_in_place) {
2657 r->init_req(3, fast_ctl);
2658 mem_phi->init_req(3, fast_mem);
2659 io_phi->init_req(3, io);
2660 res_phi->init_req(3, fast_res);
2661 }
2662 transform_later(r);
2663 transform_later(mem_phi);
2664 transform_later(io_phi);
2665 transform_later(res_phi);
2666
2667 // Do not let stores that initialize this buffer be reordered with a subsequent
2668 // store that would make this buffer accessible by other threads.
2669 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
2670 transform_later(mb);
2671 mb->init_req(TypeFunc::Memory, mem_phi);
2672 mb->init_req(TypeFunc::Control, r);
2673 r = new ProjNode(mb, TypeFunc::Control);
2674 transform_later(r);
2675 mem_phi = new ProjNode(mb, TypeFunc::Memory);
2676 transform_later(mem_phi);
2677
2678 assert(projs->nb_resproj == 1, "unexpected number of results");
2679 _igvn.replace_in_uses(projs->fallthrough_catchproj, r);
2680 _igvn.replace_in_uses(projs->fallthrough_memproj, mem_phi);
2681 _igvn.replace_in_uses(projs->fallthrough_ioproj, io_phi);
2682 _igvn.replace_in_uses(projs->resproj[0], res_phi);
2683 _igvn.replace_in_uses(projs->catchall_catchproj, ex_r);
2684 _igvn.replace_in_uses(projs->catchall_memproj, ex_mem_phi);
2685 _igvn.replace_in_uses(projs->catchall_ioproj, ex_io_phi);
2686 // The CatchNode should not use the ex_io_phi. Re-connect it to the catchall_ioproj.
2687 Node* cn = projs->fallthrough_catchproj->in(0);
2688 _igvn.replace_input_of(cn, 1, projs->catchall_ioproj);
2689
2690 _igvn.replace_node(ctl, projs->fallthrough_catchproj);
2691 _igvn.replace_node(mem, projs->fallthrough_memproj);
2692 _igvn.replace_node(io, projs->fallthrough_ioproj);
2693 _igvn.replace_node(res, projs->resproj[0]);
2694 _igvn.replace_node(ex_ctl, projs->catchall_catchproj);
2695 _igvn.replace_node(ex_mem, projs->catchall_memproj);
2696 _igvn.replace_node(ex_io, projs->catchall_ioproj);
2697 }
2698
2699 void PhaseMacroExpand::expand_subtypecheck_node(SubTypeCheckNode *check) {
2700 assert(check->in(SubTypeCheckNode::Control) == nullptr, "should be pinned");
2701 Node* bol = check->unique_out();
2702 Node* obj_or_subklass = check->in(SubTypeCheckNode::ObjOrSubKlass);
2703 Node* superklass = check->in(SubTypeCheckNode::SuperKlass);
2704 assert(bol->is_Bool() && bol->as_Bool()->_test._test == BoolTest::ne, "unexpected bool node");
2705
2706 for (DUIterator_Last imin, i = bol->last_outs(imin); i >= imin; --i) {
2707 Node* iff = bol->last_out(i);
2708 assert(iff->is_If(), "where's the if?");
2709
2710 if (iff->in(0)->is_top()) {
2711 _igvn.replace_input_of(iff, 1, C->top());
2712 continue;
2713 }
2714
2715 Node* iftrue = iff->as_If()->proj_out(1);
2716 Node* iffalse = iff->as_If()->proj_out(0);
2717 Node* ctrl = iff->in(0);
2718
2719 Node* subklass = nullptr;
2720 if (_igvn.type(obj_or_subklass)->isa_klassptr()) {
2721 subklass = obj_or_subklass;
2722 } else {
2723 Node* k_adr = basic_plus_adr(obj_or_subklass, oopDesc::klass_offset_in_bytes());
2724 subklass = _igvn.transform(LoadKlassNode::make(_igvn, nullptr, C->immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
2725 }
2726
2727 Node* not_subtype_ctrl = Phase::gen_subtype_check(subklass, superklass, &ctrl, nullptr, _igvn, check->method(), check->bci());
2728
2729 _igvn.replace_input_of(iff, 0, C->top());
2730 _igvn.replace_node(iftrue, not_subtype_ctrl);
2731 _igvn.replace_node(iffalse, ctrl);
2732 }
2733 _igvn.replace_node(check, C->top());
2734 }
2735
2736 // FlatArrayCheckNode (array1 array2 ...) is expanded into:
2737 //
2738 // long mark = array1.mark | array2.mark | ...;
2739 // long locked_bit = markWord::unlocked_value & array1.mark & array2.mark & ...;
2740 // if (locked_bit == 0) {
2741 // // One array is locked, load prototype header from the klass
2742 // mark = array1.klass.proto | array2.klass.proto | ...
2743 // }
2744 // if ((mark & markWord::flat_array_bit_in_place) == 0) {
2745 // ...
2746 // }
2747 void PhaseMacroExpand::expand_flatarraycheck_node(FlatArrayCheckNode* check) {
2748 bool array_inputs = _igvn.type(check->in(FlatArrayCheckNode::ArrayOrKlass))->isa_oopptr() != nullptr;
2749 if (array_inputs) {
2750 Node* mark = MakeConX(0);
2751 Node* locked_bit = MakeConX(markWord::unlocked_value);
2752 Node* mem = check->in(FlatArrayCheckNode::Memory);
2753 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
2754 Node* ary = check->in(i);
2755 const TypeOopPtr* t = _igvn.type(ary)->isa_oopptr();
2756 assert(t != nullptr, "Mixing array and klass inputs");
2757 assert(!t->is_flat() && !t->is_not_flat(), "Should have been optimized out");
2758 Node* mark_adr = basic_plus_adr(ary, oopDesc::mark_offset_in_bytes());
2759 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));
2760 mark = _igvn.transform(new OrXNode(mark, mark_load));
2761 locked_bit = _igvn.transform(new AndXNode(locked_bit, mark_load));
2762 }
2763 assert(!mark->is_Con(), "Should have been optimized out");
2764 Node* cmp = _igvn.transform(new CmpXNode(locked_bit, MakeConX(0)));
2765 Node* is_unlocked = _igvn.transform(new BoolNode(cmp, BoolTest::ne));
2766
2767 // BoolNode might be shared, replace each if user
2768 Node* old_bol = check->unique_out();
2769 assert(old_bol->is_Bool() && old_bol->as_Bool()->_test._test == BoolTest::ne, "unexpected condition");
2770 for (DUIterator_Last imin, i = old_bol->last_outs(imin); i >= imin; --i) {
2771 IfNode* old_iff = old_bol->last_out(i)->as_If();
2772 Node* ctrl = old_iff->in(0);
2773 RegionNode* region = new RegionNode(3);
2774 Node* mark_phi = new PhiNode(region, TypeX_X);
2775
2776 // Check if array is unlocked
2777 IfNode* iff = _igvn.transform(new IfNode(ctrl, is_unlocked, PROB_MAX, COUNT_UNKNOWN))->as_If();
2778
2779 // Unlocked: Use bits from mark word
2780 region->init_req(1, _igvn.transform(new IfTrueNode(iff)));
2781 mark_phi->init_req(1, mark);
2782
2783 // Locked: Load prototype header from klass
2784 ctrl = _igvn.transform(new IfFalseNode(iff));
2785 Node* proto = MakeConX(0);
2786 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
2787 Node* ary = check->in(i);
2788 // Make loads control dependent to make sure they are only executed if array is locked
2789 Node* klass_adr = basic_plus_adr(ary, oopDesc::klass_offset_in_bytes());
2790 Node* klass = _igvn.transform(LoadKlassNode::make(_igvn, ctrl, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
2791 Node* proto_adr = basic_plus_adr(klass, in_bytes(Klass::prototype_header_offset()));
2792 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));
2793 proto = _igvn.transform(new OrXNode(proto, proto_load));
2794 }
2795 region->init_req(2, ctrl);
2796 mark_phi->init_req(2, proto);
2797
2798 // Check if flat array bits are set
2799 Node* mask = MakeConX(markWord::flat_array_bit_in_place);
2800 Node* masked = _igvn.transform(new AndXNode(_igvn.transform(mark_phi), mask));
2801 cmp = _igvn.transform(new CmpXNode(masked, MakeConX(0)));
2802 Node* is_not_flat = _igvn.transform(new BoolNode(cmp, BoolTest::eq));
2803
2804 ctrl = _igvn.transform(region);
2805 iff = _igvn.transform(new IfNode(ctrl, is_not_flat, PROB_MAX, COUNT_UNKNOWN))->as_If();
2806 _igvn.replace_node(old_iff, iff);
2807 }
2808 _igvn.replace_node(check, C->top());
2809 } else {
2810 // Fall back to layout helper check
2811 Node* lhs = intcon(0);
2812 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
2813 Node* array_or_klass = check->in(i);
2814 Node* klass = nullptr;
2815 const TypePtr* t = _igvn.type(array_or_klass)->is_ptr();
2816 assert(!t->is_flat() && !t->is_not_flat(), "Should have been optimized out");
2817 if (t->isa_oopptr() != nullptr) {
2818 Node* klass_adr = basic_plus_adr(array_or_klass, oopDesc::klass_offset_in_bytes());
2819 klass = transform_later(LoadKlassNode::make(_igvn, nullptr, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
2820 } else {
2821 assert(t->isa_klassptr(), "Unexpected input type");
2822 klass = array_or_klass;
2823 }
2824 Node* lh_addr = basic_plus_adr(klass, in_bytes(Klass::layout_helper_offset()));
2825 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));
2826 lhs = _igvn.transform(new OrINode(lhs, lh_val));
2827 }
2828 Node* masked = transform_later(new AndINode(lhs, intcon(Klass::_lh_array_tag_flat_value_bit_inplace)));
2829 Node* cmp = transform_later(new CmpINode(masked, intcon(0)));
2830 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
2831 Node* m2b = transform_later(new Conv2BNode(masked));
2832 // The matcher expects the input to If nodes to be produced by a Bool(CmpI..)
2833 // pattern, but the input to other potential users (e.g. Phi) to be some
2834 // other pattern (e.g. a Conv2B node, possibly idealized as a CMoveI).
2835 Node* old_bol = check->unique_out();
2836 for (DUIterator_Last imin, i = old_bol->last_outs(imin); i >= imin; --i) {
2837 Node* user = old_bol->last_out(i);
2838 for (uint j = 0; j < user->req(); j++) {
2839 Node* n = user->in(j);
2840 if (n == old_bol) {
2841 _igvn.replace_input_of(user, j, user->is_If() ? bol : m2b);
2842 }
2843 }
2844 }
2845 _igvn.replace_node(check, C->top());
2846 }
2847 }
2848
2849 //---------------------------eliminate_macro_nodes----------------------
2850 // Eliminate scalar replaced allocations and associated locks.
2851 void PhaseMacroExpand::eliminate_macro_nodes() {
2852 if (C->macro_count() == 0)
2853 return;
2854 NOT_PRODUCT(int membar_before = count_MemBar(C);)
2855
2856 // Before elimination may re-mark (change to Nested or NonEscObj)
2857 // all associated (same box and obj) lock and unlock nodes.
2858 int cnt = C->macro_count();
2859 for (int i=0; i < cnt; i++) {
2860 Node *n = C->macro_node(i);
2861 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2862 mark_eliminated_locking_nodes(n->as_AbstractLock());
2863 }
2864 }
2865 // Re-marking may break consistency of Coarsened locks.
2866 if (!C->coarsened_locks_consistent()) {
2867 return; // recompile without Coarsened locks if broken
2868 } else {
2869 // After coarsened locks are eliminated locking regions
2870 // become unbalanced. We should not execute any more
2871 // locks elimination optimizations on them.
2872 C->mark_unbalanced_boxes();
2873 }
2874
2875 // First, attempt to eliminate locks
2876 bool progress = true;
2877 while (progress) {
2878 progress = false;
2879 for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once
2880 Node* n = C->macro_node(i - 1);
2881 bool success = false;
2882 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2883 if (n->is_AbstractLock()) {
2884 success = eliminate_locking_node(n->as_AbstractLock());
2885 #ifndef PRODUCT
2886 if (success && PrintOptoStatistics) {
2887 Atomic::inc(&PhaseMacroExpand::_monitor_objects_removed_counter);
2888 }
2889 #endif
2890 }
2891 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2892 progress = progress || success;
2893 }
2894 }
2895 // Next, attempt to eliminate allocations
2896 _has_locks = false;
2897 progress = true;
2898 while (progress) {
2899 progress = false;
2900 for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once
2901 Node* n = C->macro_node(i - 1);
2902 bool success = false;
2903 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2904 switch (n->class_id()) {
2905 case Node::Class_Allocate:
2906 case Node::Class_AllocateArray:
2907 success = eliminate_allocate_node(n->as_Allocate());
2908 #ifndef PRODUCT
2909 if (success && PrintOptoStatistics) {
2910 Atomic::inc(&PhaseMacroExpand::_objs_scalar_replaced_counter);
2911 }
2912 #endif
2913 break;
2914 case Node::Class_CallStaticJava: {
2915 CallStaticJavaNode* call = n->as_CallStaticJava();
2916 if (!call->method()->is_method_handle_intrinsic()) {
2917 success = eliminate_boxing_node(n->as_CallStaticJava());
2918 }
2919 break;
2920 }
2921 case Node::Class_Lock:
2922 case Node::Class_Unlock:
2923 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2924 _has_locks = true;
2925 break;
2926 case Node::Class_ArrayCopy:
2927 break;
2928 case Node::Class_OuterStripMinedLoop:
2929 break;
2930 case Node::Class_SubTypeCheck:
2931 break;
2932 case Node::Class_Opaque1:
2933 break;
2934 case Node::Class_FlatArrayCheck:
2935 break;
2936 default:
2937 assert(n->Opcode() == Op_LoopLimit ||
2938 n->is_Opaque4() ||
2939 n->is_OpaqueInitializedAssertionPredicate() ||
2940 n->Opcode() == Op_MaxL ||
2941 n->Opcode() == Op_MinL ||
2942 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n),
2943 "unknown node type in macro list");
2944 }
2945 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2946 progress = progress || success;
2947 }
2948 }
2949 #ifndef PRODUCT
2950 if (PrintOptoStatistics) {
2951 int membar_after = count_MemBar(C);
2952 Atomic::add(&PhaseMacroExpand::_memory_barriers_removed_counter, membar_before - membar_after);
2953 }
2954 #endif
2955 }
2956
2957 //------------------------------expand_macro_nodes----------------------
2958 // Returns true if a failure occurred.
2959 bool PhaseMacroExpand::expand_macro_nodes() {
2960 // Do not allow new macro nodes once we started to expand
2961 C->reset_allow_macro_nodes();
2962 if (StressMacroExpansion) {
2963 C->shuffle_macro_nodes();
2964 }
2965 // Last attempt to eliminate macro nodes.
2966 eliminate_macro_nodes();
2967 if (C->failing()) return true;
2968
2969 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2970 bool progress = true;
2971 while (progress) {
2972 progress = false;
2973 for (int i = C->macro_count(); i > 0; i--) {
2974 Node* n = C->macro_node(i-1);
2975 bool success = false;
2976 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2977 if (n->Opcode() == Op_LoopLimit) {
2978 // Remove it from macro list and put on IGVN worklist to optimize.
2979 C->remove_macro_node(n);
2980 _igvn._worklist.push(n);
2981 success = true;
2982 } else if (n->Opcode() == Op_CallStaticJava) {
2983 CallStaticJavaNode* call = n->as_CallStaticJava();
2984 if (!call->method()->is_method_handle_intrinsic()) {
2985 // Remove it from macro list and put on IGVN worklist to optimize.
2986 C->remove_macro_node(n);
2987 _igvn._worklist.push(n);
2988 success = true;
2989 }
2990 } else if (n->is_Opaque1()) {
2991 _igvn.replace_node(n, n->in(1));
2992 success = true;
2993 } else if (n->is_Opaque4()) {
2994 // With Opaque4 nodes, the expectation is that the test of input 1
2995 // is always equal to the constant value of input 2. So we can
2996 // remove the Opaque4 and replace it by input 2. In debug builds,
2997 // leave the non constant test in instead to sanity check that it
2998 // never fails (if it does, that subgraph was constructed so, at
2999 // runtime, a Halt node is executed).
3000 #ifdef ASSERT
3001 _igvn.replace_node(n, n->in(1));
3002 #else
3003 _igvn.replace_node(n, n->in(2));
3004 #endif
3005 success = true;
3006 } else if (n->is_OpaqueInitializedAssertionPredicate()) {
3007 // Initialized Assertion Predicates must always evaluate to true. Therefore, we get rid of them in product
3008 // builds as they are useless. In debug builds we keep them as additional verification code. Even though
3009 // loop opts are already over, we want to keep Initialized Assertion Predicates alive as long as possible to
3010 // enable folding of dead control paths within which cast nodes become top after due to impossible types -
3011 // even after loop opts are over. Therefore, we delay the removal of these opaque nodes until now.
3012 #ifdef ASSERT
3013 _igvn.replace_node(n, n->in(1));
3014 #else
3015 _igvn.replace_node(n, _igvn.intcon(1));
3016 #endif // ASSERT
3017 } else if (n->Opcode() == Op_OuterStripMinedLoop) {
3018 n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn);
3019 C->remove_macro_node(n);
3020 success = true;
3021 } else if (n->Opcode() == Op_MaxL) {
3022 // Since MaxL and MinL are not implemented in the backend, we expand them to
3023 // a CMoveL construct now. At least until here, the type could be computed
3024 // precisely. CMoveL is not so smart, but we can give it at least the best
3025 // type we know abouot n now.
3026 Node* repl = MaxNode::signed_max(n->in(1), n->in(2), _igvn.type(n), _igvn);
3027 _igvn.replace_node(n, repl);
3028 success = true;
3029 } else if (n->Opcode() == Op_MinL) {
3030 Node* repl = MaxNode::signed_min(n->in(1), n->in(2), _igvn.type(n), _igvn);
3031 _igvn.replace_node(n, repl);
3032 success = true;
3033 }
3034 assert(!success || (C->macro_count() == (old_macro_count - 1)), "elimination must have deleted one node from macro list");
3035 progress = progress || success;
3036 if (success) {
3037 C->print_method(PHASE_AFTER_MACRO_EXPANSION_STEP, 5, n);
3038 }
3039 }
3040 }
3041
3042 // Clean up the graph so we're less likely to hit the maximum node
3043 // limit
3044 _igvn.set_delay_transform(false);
3045 _igvn.optimize();
3046 if (C->failing()) return true;
3047 _igvn.set_delay_transform(true);
3048
3049
3050 // Because we run IGVN after each expansion, some macro nodes may go
3051 // dead and be removed from the list as we iterate over it. Move
3052 // Allocate nodes (processed in a second pass) at the beginning of
3053 // the list and then iterate from the last element of the list until
3054 // an Allocate node is seen. This is robust to random deletion in
3055 // the list due to nodes going dead.
3056 C->sort_macro_nodes();
3057
3058 // expand arraycopy "macro" nodes first
3059 // For ReduceBulkZeroing, we must first process all arraycopy nodes
3060 // before the allocate nodes are expanded.
3061 while (C->macro_count() > 0) {
3062 int macro_count = C->macro_count();
3063 Node * n = C->macro_node(macro_count-1);
3064 assert(n->is_macro(), "only macro nodes expected here");
3065 if (_igvn.type(n) == Type::TOP || (n->in(0) != nullptr && n->in(0)->is_top())) {
3066 // node is unreachable, so don't try to expand it
3067 C->remove_macro_node(n);
3068 continue;
3069 }
3070 if (n->is_Allocate()) {
3071 break;
3072 }
3073 // Make sure expansion will not cause node limit to be exceeded.
3074 // Worst case is a macro node gets expanded into about 200 nodes.
3075 // Allow 50% more for optimization.
3076 if (C->check_node_count(300, "out of nodes before macro expansion")) {
3077 return true;
3078 }
3079
3080 DEBUG_ONLY(int old_macro_count = C->macro_count();)
3081 switch (n->class_id()) {
3082 case Node::Class_Lock:
3083 expand_lock_node(n->as_Lock());
3084 break;
3085 case Node::Class_Unlock:
3086 expand_unlock_node(n->as_Unlock());
3087 break;
3088 case Node::Class_ArrayCopy:
3089 expand_arraycopy_node(n->as_ArrayCopy());
3090 break;
3091 case Node::Class_SubTypeCheck:
3092 expand_subtypecheck_node(n->as_SubTypeCheck());
3093 break;
3094 case Node::Class_CallStaticJava:
3095 expand_mh_intrinsic_return(n->as_CallStaticJava());
3096 C->remove_macro_node(n);
3097 break;
3098 case Node::Class_FlatArrayCheck:
3099 expand_flatarraycheck_node(n->as_FlatArrayCheck());
3100 break;
3101 default:
3102 assert(false, "unknown node type in macro list");
3103 }
3104 assert(C->macro_count() == (old_macro_count - 1), "expansion must have deleted one node from macro list");
3105 if (C->failing()) return true;
3106 C->print_method(PHASE_AFTER_MACRO_EXPANSION_STEP, 5, n);
3107
3108 // Clean up the graph so we're less likely to hit the maximum node
3109 // limit
3110 _igvn.set_delay_transform(false);
3111 _igvn.optimize();
3112 if (C->failing()) return true;
3113 _igvn.set_delay_transform(true);
3114 }
3115
3116 // All nodes except Allocate nodes are expanded now. There could be
3117 // new optimization opportunities (such as folding newly created
3118 // load from a just allocated object). Run IGVN.
3119
3120 // expand "macro" nodes
3121 // nodes are removed from the macro list as they are processed
3122 while (C->macro_count() > 0) {
3123 int macro_count = C->macro_count();
3124 Node * n = C->macro_node(macro_count-1);
3125 assert(n->is_macro(), "only macro nodes expected here");
3126 if (_igvn.type(n) == Type::TOP || (n->in(0) != nullptr && n->in(0)->is_top())) {
3127 // node is unreachable, so don't try to expand it
3128 C->remove_macro_node(n);
3129 continue;
3130 }
3131 // Make sure expansion will not cause node limit to be exceeded.
3132 // Worst case is a macro node gets expanded into about 200 nodes.
3133 // Allow 50% more for optimization.
3134 if (C->check_node_count(300, "out of nodes before macro expansion")) {
3135 return true;
3136 }
3137 switch (n->class_id()) {
3138 case Node::Class_Allocate:
3139 expand_allocate(n->as_Allocate());
3140 break;
3141 case Node::Class_AllocateArray:
3142 expand_allocate_array(n->as_AllocateArray());
3143 break;
3144 default:
3145 assert(false, "unknown node type in macro list");
3146 }
3147 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
3148 if (C->failing()) return true;
3149 C->print_method(PHASE_AFTER_MACRO_EXPANSION_STEP, 5, n);
3150
3151 // Clean up the graph so we're less likely to hit the maximum node
3152 // limit
3153 _igvn.set_delay_transform(false);
3154 _igvn.optimize();
3155 if (C->failing()) return true;
3156 _igvn.set_delay_transform(true);
3157 }
3158
3159 _igvn.set_delay_transform(false);
3160 return false;
3161 }
3162
3163 #ifndef PRODUCT
3164 int PhaseMacroExpand::_objs_scalar_replaced_counter = 0;
3165 int PhaseMacroExpand::_monitor_objects_removed_counter = 0;
3166 int PhaseMacroExpand::_GC_barriers_removed_counter = 0;
3167 int PhaseMacroExpand::_memory_barriers_removed_counter = 0;
3168
3169 void PhaseMacroExpand::print_statistics() {
3170 tty->print("Objects scalar replaced = %d, ", Atomic::load(&_objs_scalar_replaced_counter));
3171 tty->print("Monitor objects removed = %d, ", Atomic::load(&_monitor_objects_removed_counter));
3172 tty->print("GC barriers removed = %d, ", Atomic::load(&_GC_barriers_removed_counter));
3173 tty->print_cr("Memory barriers removed = %d", Atomic::load(&_memory_barriers_removed_counter));
3174 }
3175
3176 int PhaseMacroExpand::count_MemBar(Compile *C) {
3177 if (!PrintOptoStatistics) {
3178 return 0;
3179 }
3180 Unique_Node_List ideal_nodes;
3181 int total = 0;
3182 ideal_nodes.map(C->live_nodes(), nullptr);
3183 ideal_nodes.push(C->root());
3184 for (uint next = 0; next < ideal_nodes.size(); ++next) {
3185 Node* n = ideal_nodes.at(next);
3186 if (n->is_MemBar()) {
3187 total++;
3188 }
3189 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3190 Node* m = n->fast_out(i);
3191 ideal_nodes.push(m);
3192 }
3193 }
3194 return total;
3195 }
3196 #endif