1 /*
2 * Copyright (c) 2005, 2025, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "ci/ciFlatArrayKlass.hpp"
26 #include "ci/ciInstanceKlass.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/opcodes.hpp"
50 #include "opto/phaseX.hpp"
51 #include "opto/rootnode.hpp"
52 #include "opto/runtime.hpp"
53 #include "opto/subnode.hpp"
54 #include "opto/subtypenode.hpp"
55 #include "opto/type.hpp"
56 #include "prims/jvmtiExport.hpp"
57 #include "runtime/continuation.hpp"
58 #include "runtime/sharedRuntime.hpp"
59 #include "runtime/stubRoutines.hpp"
60 #include "utilities/globalDefinitions.hpp"
61 #include "utilities/macros.hpp"
62 #include "utilities/powerOfTwo.hpp"
63 #if INCLUDE_G1GC
64 #include "gc/g1/g1ThreadLocalData.hpp"
65 #endif // INCLUDE_G1GC
66
67
68 //
69 // Replace any references to "oldref" in inputs to "use" with "newref".
70 // Returns the number of replacements made.
71 //
72 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
73 int nreplacements = 0;
74 uint req = use->req();
75 for (uint j = 0; j < use->len(); j++) {
76 Node *uin = use->in(j);
77 if (uin == oldref) {
78 if (j < req)
79 use->set_req(j, newref);
80 else
81 use->set_prec(j, newref);
82 nreplacements++;
83 } else if (j >= req && uin == nullptr) {
84 break;
85 }
86 }
87 return nreplacements;
88 }
89
90
91 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word) {
92 Node* cmp = word;
93 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
94 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
95 transform_later(iff);
96
97 // Fast path taken.
98 Node *fast_taken = transform_later(new IfFalseNode(iff));
99
100 // Fast path not-taken, i.e. slow path
101 Node *slow_taken = transform_later(new IfTrueNode(iff));
102
103 region->init_req(edge, fast_taken); // Capture fast-control
104 return slow_taken;
105 }
106
107 //--------------------copy_predefined_input_for_runtime_call--------------------
108 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
109 // Set fixed predefined input arguments
110 call->init_req( TypeFunc::Control, ctrl );
111 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
112 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
113 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
114 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
115 }
116
117 //------------------------------make_slow_call---------------------------------
118 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type,
119 address slow_call, const char* leaf_name, Node* slow_path,
120 Node* parm0, Node* parm1, Node* parm2) {
121
122 // Slow-path call
123 CallNode *call = leaf_name
124 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
125 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), TypeRawPtr::BOTTOM );
126
127 // Slow path call has no side-effects, uses few values
128 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
129 if (parm0 != nullptr) call->init_req(TypeFunc::Parms+0, parm0);
130 if (parm1 != nullptr) call->init_req(TypeFunc::Parms+1, parm1);
131 if (parm2 != nullptr) call->init_req(TypeFunc::Parms+2, parm2);
132 call->copy_call_debug_info(&_igvn, oldcall);
133 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
134 _igvn.replace_node(oldcall, call);
135 transform_later(call);
136
137 return call;
138 }
139
140 void PhaseMacroExpand::eliminate_gc_barrier(Node* p2x) {
141 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
142 bs->eliminate_gc_barrier(&_igvn, p2x);
143 #ifndef PRODUCT
144 if (PrintOptoStatistics) {
145 AtomicAccess::inc(&PhaseMacroExpand::_GC_barriers_removed_counter);
146 }
147 #endif
148 }
149
150 // Search for a memory operation for the specified memory slice.
151 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
152 Node *orig_mem = mem;
153 Node *alloc_mem = alloc->as_Allocate()->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
154 assert(alloc_mem != nullptr, "Allocation without a memory projection.");
155 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
156 while (true) {
157 if (mem == alloc_mem || mem == start_mem ) {
158 return mem; // hit one of our sentinels
159 } else if (mem->is_MergeMem()) {
160 mem = mem->as_MergeMem()->memory_at(alias_idx);
161 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
162 Node *in = mem->in(0);
163 // we can safely skip over safepoints, calls, locks and membars because we
164 // already know that the object is safe to eliminate.
165 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
166 return in;
167 } else if (in->is_Call()) {
168 CallNode *call = in->as_Call();
169 if (call->may_modify(tinst, phase)) {
170 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape");
171 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) {
172 return in;
173 }
174 }
175 mem = in->in(TypeFunc::Memory);
176 } else if (in->is_MemBar()) {
177 ArrayCopyNode* ac = nullptr;
178 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) {
179 if (ac != nullptr) {
180 assert(ac->is_clonebasic(), "Only basic clone is a non escaping clone");
181 return ac;
182 }
183 }
184 mem = in->in(TypeFunc::Memory);
185 } else if (in->is_LoadFlat() || in->is_StoreFlat()) {
186 mem = in->in(TypeFunc::Memory);
187 } else {
188 #ifdef ASSERT
189 in->dump();
190 mem->dump();
191 assert(false, "unexpected projection");
192 #endif
193 }
194 } else if (mem->is_Store()) {
195 const TypePtr* atype = mem->as_Store()->adr_type();
196 int adr_idx = phase->C->get_alias_index(atype);
197 if (adr_idx == alias_idx) {
198 assert(atype->isa_oopptr(), "address type must be oopptr");
199 int adr_offset = atype->flat_offset();
200 uint adr_iid = atype->is_oopptr()->instance_id();
201 // Array elements references have the same alias_idx
202 // but different offset and different instance_id.
203 if (adr_offset == offset && adr_iid == alloc->_idx) {
204 return mem;
205 }
206 } else {
207 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
208 }
209 mem = mem->in(MemNode::Memory);
210 } else if (mem->is_ClearArray()) {
211 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
212 // Can not bypass initialization of the instance
213 // we are looking.
214 DEBUG_ONLY(intptr_t offset;)
215 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
216 InitializeNode* init = alloc->as_Allocate()->initialization();
217 // We are looking for stored value, return Initialize node
218 // or memory edge from Allocate node.
219 if (init != nullptr) {
220 return init;
221 } else {
222 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
223 }
224 }
225 // Otherwise skip it (the call updated 'mem' value).
226 } else if (mem->Opcode() == Op_SCMemProj) {
227 mem = mem->in(0);
228 Node* adr = nullptr;
229 if (mem->is_LoadStore()) {
230 adr = mem->in(MemNode::Address);
231 } else {
232 assert(mem->Opcode() == Op_EncodeISOArray ||
233 mem->Opcode() == Op_StrCompressedCopy, "sanity");
234 adr = mem->in(3); // Destination array
235 }
236 const TypePtr* atype = adr->bottom_type()->is_ptr();
237 int adr_idx = phase->C->get_alias_index(atype);
238 if (adr_idx == alias_idx) {
239 DEBUG_ONLY(mem->dump();)
240 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
241 return nullptr;
242 }
243 mem = mem->in(MemNode::Memory);
244 } else if (mem->Opcode() == Op_StrInflatedCopy) {
245 Node* adr = mem->in(3); // Destination array
246 const TypePtr* atype = adr->bottom_type()->is_ptr();
247 int adr_idx = phase->C->get_alias_index(atype);
248 if (adr_idx == alias_idx) {
249 DEBUG_ONLY(mem->dump();)
250 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
251 return nullptr;
252 }
253 mem = mem->in(MemNode::Memory);
254 } else {
255 return mem;
256 }
257 assert(mem != orig_mem, "dead memory loop");
258 }
259 }
260
261 // Generate loads from source of the arraycopy for fields of
262 // destination needed at a deoptimization point
263 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) {
264 BasicType bt = ft;
265 const Type *type = ftype;
266 if (ft == T_NARROWOOP) {
267 bt = T_OBJECT;
268 type = ftype->make_oopptr();
269 }
270 Node* res = nullptr;
271 if (ac->is_clonebasic()) {
272 assert(ac->in(ArrayCopyNode::Src) != ac->in(ArrayCopyNode::Dest), "clone source equals destination");
273 Node* base = ac->in(ArrayCopyNode::Src);
274 Node* adr = _igvn.transform(new AddPNode(base, base, _igvn.MakeConX(offset)));
275 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
276 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
277 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
278 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
279 } else {
280 if (ac->modifies(offset, offset, &_igvn, true)) {
281 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
282 uint shift = exact_log2(type2aelembytes(bt));
283 Node* src_pos = ac->in(ArrayCopyNode::SrcPos);
284 Node* dest_pos = ac->in(ArrayCopyNode::DestPos);
285 const TypeInt* src_pos_t = _igvn.type(src_pos)->is_int();
286 const TypeInt* dest_pos_t = _igvn.type(dest_pos)->is_int();
287
288 Node* adr = nullptr;
289 Node* base = ac->in(ArrayCopyNode::Src);
290 const TypeAryPtr* adr_type = _igvn.type(base)->is_aryptr();
291 if (adr_type->is_flat()) {
292 shift = adr_type->flat_log_elem_size();
293 }
294 if (src_pos_t->is_con() && dest_pos_t->is_con()) {
295 intptr_t off = ((src_pos_t->get_con() - dest_pos_t->get_con()) << shift) + offset;
296 adr = _igvn.transform(new AddPNode(base, base, _igvn.MakeConX(off)));
297 adr_type = _igvn.type(adr)->is_aryptr();
298 assert(adr_type == _igvn.type(base)->is_aryptr()->add_field_offset_and_offset(off), "incorrect address type");
299 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
300 // Don't emit a new load from src if src == dst but try to get the value from memory instead
301 return value_from_mem(ac->in(TypeFunc::Memory), ctl, ft, ftype, adr_type, alloc);
302 }
303 } else {
304 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
305 // Non constant offset in the array: we can't statically
306 // determine the value
307 return nullptr;
308 }
309 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
310 #ifdef _LP64
311 diff = _igvn.transform(new ConvI2LNode(diff));
312 #endif
313 diff = _igvn.transform(new LShiftXNode(diff, _igvn.intcon(shift)));
314
315 Node* off = _igvn.transform(new AddXNode(_igvn.MakeConX(offset), diff));
316 adr = _igvn.transform(new AddPNode(base, base, off));
317 // In the case of a flat inline type array, each field has its
318 // own slice so we need to extract the field being accessed from
319 // the address computation
320 adr_type = adr_type->add_field_offset_and_offset(offset)->add_offset(Type::OffsetBot)->is_aryptr();
321 adr = _igvn.transform(new CastPPNode(ctl, adr, adr_type));
322 }
323 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
324 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
325 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
326 }
327 }
328 if (res != nullptr) {
329 if (ftype->isa_narrowoop()) {
330 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
331 assert(res->isa_DecodeN(), "should be narrow oop");
332 res = _igvn.transform(new EncodePNode(res, ftype));
333 }
334 return res;
335 }
336 return nullptr;
337 }
338
339 //
340 // Given a Memory Phi, compute a value Phi containing the values from stores
341 // on the input paths.
342 // Note: this function is recursive, its depth is limited by the "level" argument
343 // Returns the computed Phi, or null if it cannot compute it.
344 Node *PhaseMacroExpand::value_from_mem_phi(Node *mem, BasicType ft, const Type *phi_type, const TypeOopPtr *adr_t, AllocateNode *alloc, Node_Stack *value_phis, int level) {
345 assert(mem->is_Phi(), "sanity");
346 int alias_idx = C->get_alias_index(adr_t);
347 int offset = adr_t->flat_offset();
348 int instance_id = adr_t->instance_id();
349
350 // Check if an appropriate value phi already exists.
351 Node* region = mem->in(0);
352 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
353 Node* phi = region->fast_out(k);
354 if (phi->is_Phi() && phi != mem &&
355 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
356 return phi;
357 }
358 }
359 // Check if an appropriate new value phi already exists.
360 Node* new_phi = value_phis->find(mem->_idx);
361 if (new_phi != nullptr)
362 return new_phi;
363
364 if (level <= 0) {
365 return nullptr; // Give up: phi tree too deep
366 }
367 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
368 Node *alloc_mem = alloc->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
369 assert(alloc_mem != nullptr, "Allocation without a memory projection.");
370
371 uint length = mem->req();
372 GrowableArray <Node *> values(length, length, nullptr);
373
374 // create a new Phi for the value
375 PhiNode *phi = new PhiNode(mem->in(0), phi_type, nullptr, mem->_idx, instance_id, alias_idx, offset);
376 transform_later(phi);
377 value_phis->push(phi, mem->_idx);
378
379 for (uint j = 1; j < length; j++) {
380 Node *in = mem->in(j);
381 if (in == nullptr || in->is_top()) {
382 values.at_put(j, in);
383 } else {
384 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
385 if (val == start_mem || val == alloc_mem) {
386 // hit a sentinel, return appropriate value
387 Node* init_value = alloc->in(AllocateNode::InitValue);
388 if (init_value != nullptr) {
389 if (val == start_mem) {
390 // TODO 8350865 Scalar replacement does not work well for flat arrays.
391 // Somehow we ended up with root mem and therefore walked past the alloc. Fix this. Triggered by TestGenerated::test15
392 // Don't we need field_value_by_offset?
393 return nullptr;
394 }
395 values.at_put(j, init_value);
396 } else {
397 assert(alloc->in(AllocateNode::RawInitValue) == nullptr, "init 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* init_value = alloc->in(AllocateNode::InitValue);
420 if (init_value != nullptr) {
421 // TODO 8350865 Scalar replacement does not work well for flat arrays.
422 // Is this correct for non-all-zero init values? Don't we need field_value_by_offset?
423 values.at_put(j, init_value);
424 } else {
425 assert(alloc->in(AllocateNode::RawInitValue) == nullptr, "init value may not be null");
426 values.at_put(j, _igvn.zerocon(ft));
427 }
428 } else if (val->is_Phi()) {
429 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
430 if (val == nullptr) {
431 return nullptr;
432 }
433 values.at_put(j, val);
434 } else if (val->Opcode() == Op_SCMemProj) {
435 assert(val->in(0)->is_LoadStore() ||
436 val->in(0)->Opcode() == Op_EncodeISOArray ||
437 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
438 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
439 return nullptr;
440 } else if (val->is_ArrayCopy()) {
441 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc);
442 if (res == nullptr) {
443 return nullptr;
444 }
445 values.at_put(j, res);
446 } else if (val->is_top()) {
447 // This indicates that this path into the phi is dead. Top will eventually also propagate into the Region.
448 // IGVN will clean this up later.
449 values.at_put(j, val);
450 } else {
451 DEBUG_ONLY( val->dump(); )
452 assert(false, "unknown node on this path");
453 return nullptr; // unknown node on this path
454 }
455 }
456 }
457 // Set Phi's inputs
458 for (uint j = 1; j < length; j++) {
459 if (values.at(j) == mem) {
460 phi->init_req(j, phi);
461 } else {
462 phi->init_req(j, values.at(j));
463 }
464 }
465 return phi;
466 }
467
468 // Search the last value stored into the object's field.
469 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
470 assert(adr_t->is_known_instance_field(), "instance required");
471 int instance_id = adr_t->instance_id();
472 assert((uint)instance_id == alloc->_idx, "wrong allocation");
473
474 int alias_idx = C->get_alias_index(adr_t);
475 int offset = adr_t->flat_offset();
476 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
477 Node *alloc_mem = alloc->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
478 assert(alloc_mem != nullptr, "Allocation without a memory projection.");
479 VectorSet visited;
480
481 bool done = sfpt_mem == alloc_mem;
482 Node *mem = sfpt_mem;
483 while (!done) {
484 if (visited.test_set(mem->_idx)) {
485 return nullptr; // found a loop, give up
486 }
487 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
488 if (mem == start_mem || mem == alloc_mem) {
489 done = true; // hit a sentinel, return appropriate 0 value
490 } else if (mem->is_Initialize()) {
491 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
492 if (mem == nullptr) {
493 done = true; // Something went wrong.
494 } else if (mem->is_Store()) {
495 const TypePtr* atype = mem->as_Store()->adr_type();
496 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
497 done = true;
498 }
499 } else if (mem->is_Store()) {
500 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
501 assert(atype != nullptr, "address type must be oopptr");
502 assert(C->get_alias_index(atype) == alias_idx &&
503 atype->is_known_instance_field() && atype->flat_offset() == offset &&
504 atype->instance_id() == instance_id, "store is correct memory slice");
505 done = true;
506 } else if (mem->is_Phi()) {
507 // try to find a phi's unique input
508 Node *unique_input = nullptr;
509 Node *top = C->top();
510 for (uint i = 1; i < mem->req(); i++) {
511 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
512 if (n == nullptr || n == top || n == mem) {
513 continue;
514 } else if (unique_input == nullptr) {
515 unique_input = n;
516 } else if (unique_input != n) {
517 unique_input = top;
518 break;
519 }
520 }
521 if (unique_input != nullptr && unique_input != top) {
522 mem = unique_input;
523 } else {
524 done = true;
525 }
526 } else if (mem->is_ArrayCopy()) {
527 done = true;
528 } else {
529 DEBUG_ONLY( mem->dump(); )
530 assert(false, "unexpected node");
531 }
532 }
533 if (mem != nullptr) {
534 if (mem == start_mem || mem == alloc_mem) {
535 // hit a sentinel, return appropriate value
536 Node* init_value = alloc->in(AllocateNode::InitValue);
537 if (init_value != nullptr) {
538 if (adr_t->is_flat()) {
539 if (init_value->is_EncodeP()) {
540 init_value = init_value->in(1);
541 }
542 assert(adr_t->is_aryptr()->field_offset().get() != Type::OffsetBot, "Unknown offset");
543 offset = adr_t->is_aryptr()->field_offset().get() + init_value->bottom_type()->inline_klass()->payload_offset();
544 init_value = init_value->as_InlineType()->field_value_by_offset(offset, true);
545 if (ft == T_NARROWOOP) {
546 init_value = transform_later(new EncodePNode(init_value, init_value->bottom_type()->make_ptr()));
547 }
548 }
549 return init_value;
550 }
551 assert(alloc->in(AllocateNode::RawInitValue) == nullptr, "init value may not be null");
552 return _igvn.zerocon(ft);
553 } else if (mem->is_Store()) {
554 Node* n = mem->in(MemNode::ValueIn);
555 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
556 n = bs->step_over_gc_barrier(n);
557 return n;
558 } else if (mem->is_Phi()) {
559 // attempt to produce a Phi reflecting the values on the input paths of the Phi
560 Node_Stack value_phis(8);
561 Node* phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
562 if (phi != nullptr) {
563 return phi;
564 } else {
565 // Kill all new Phis
566 while(value_phis.is_nonempty()) {
567 Node* n = value_phis.node();
568 _igvn.replace_node(n, C->top());
569 value_phis.pop();
570 }
571 }
572 } else if (mem->is_ArrayCopy()) {
573 Node* ctl = mem->in(0);
574 Node* m = mem->in(TypeFunc::Memory);
575 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj()) {
576 // pin the loads in the uncommon trap path
577 ctl = sfpt_ctl;
578 m = sfpt_mem;
579 }
580 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc);
581 }
582 }
583 // Something went wrong.
584 return nullptr;
585 }
586
587 // Search the last value stored into the inline type's fields (for flat arrays).
588 Node* PhaseMacroExpand::inline_type_from_mem(ciInlineKlass* vk, const TypeAryPtr* elem_adr_type, int elem_idx, int offset_in_element, bool null_free, AllocateNode* alloc, SafePointNode* sfpt) {
589 auto report_failure = [&](int field_offset_in_element) {
590 #ifndef PRODUCT
591 if (PrintEliminateAllocations) {
592 ciInlineKlass* elem_klass = elem_adr_type->elem()->inline_klass();
593 int offset = field_offset_in_element + elem_klass->payload_offset();
594 ciField* flattened_field = elem_klass->get_field_by_offset(offset, false);
595 assert(flattened_field != nullptr, "must have a field of type %s at offset %d", elem_klass->name()->as_utf8(), offset);
596 tty->print("=== At SafePoint node %d can't find value of field [%s] of array element [%d]", sfpt->_idx, flattened_field->name()->as_utf8(), elem_idx);
597 tty->print(", which prevents elimination of: ");
598 alloc->dump();
599 }
600 #endif // PRODUCT
601 };
602
603 // Create a new InlineTypeNode and retrieve the field values from memory
604 InlineTypeNode* vt = InlineTypeNode::make_uninitialized(_igvn, vk, false);
605 transform_later(vt);
606 if (null_free) {
607 vt->set_null_marker(_igvn);
608 } else {
609 int nm_offset_in_element = offset_in_element + vk->null_marker_offset_in_payload();
610 const TypeAryPtr* nm_adr_type = elem_adr_type->with_field_offset(nm_offset_in_element);
611 Node* nm_value = value_from_mem(sfpt->memory(), sfpt->control(), T_BOOLEAN, TypeInt::BOOL, nm_adr_type, alloc);
612 if (nm_value != nullptr) {
613 vt->set_null_marker(_igvn, nm_value);
614 } else {
615 report_failure(nm_offset_in_element);
616 return nullptr;
617 }
618 }
619
620 for (int i = 0; i < vk->nof_declared_nonstatic_fields(); ++i) {
621 ciType* field_type = vt->field_type(i);
622 int field_offset_in_element = offset_in_element + vt->field_offset(i) - vk->payload_offset();
623 Node* field_value = nullptr;
624 if (vt->field_is_flat(i)) {
625 field_value = inline_type_from_mem(field_type->as_inline_klass(), elem_adr_type, elem_idx, field_offset_in_element, vt->field_is_null_free(i), alloc, sfpt);
626 } else {
627 const Type* ft = Type::get_const_type(field_type);
628 BasicType bt = type2field[field_type->basic_type()];
629 if (UseCompressedOops && !is_java_primitive(bt)) {
630 ft = ft->make_narrowoop();
631 bt = T_NARROWOOP;
632 }
633 // Each inline type field has its own memory slice
634 const TypeAryPtr* field_adr_type = elem_adr_type->with_field_offset(field_offset_in_element);
635 field_value = value_from_mem(sfpt->memory(), sfpt->control(), bt, ft, field_adr_type, alloc);
636 if (field_value == nullptr) {
637 report_failure(field_offset_in_element);
638 } else if (ft->isa_narrowoop()) {
639 assert(UseCompressedOops, "unexpected narrow oop");
640 if (field_value->is_EncodeP()) {
641 field_value = field_value->in(1);
642 } else if (!field_value->is_InlineType()) {
643 field_value = transform_later(new DecodeNNode(field_value, field_value->get_ptr_type()));
644 }
645 }
646 }
647 if (field_value != nullptr) {
648 vt->set_field_value(i, field_value);
649 } else {
650 return nullptr;
651 }
652 }
653 return vt;
654 }
655
656 // Check the possibility of scalar replacement.
657 bool PhaseMacroExpand::can_eliminate_allocation(PhaseIterGVN* igvn, AllocateNode *alloc, GrowableArray <SafePointNode *>* safepoints) {
658 // Scan the uses of the allocation to check for anything that would
659 // prevent us from eliminating it.
660 NOT_PRODUCT( const char* fail_eliminate = nullptr; )
661 DEBUG_ONLY( Node* disq_node = nullptr; )
662 bool can_eliminate = true;
663 bool reduce_merge_precheck = (safepoints == nullptr);
664
665 Unique_Node_List worklist;
666 Node* res = alloc->result_cast();
667 const TypeOopPtr* res_type = nullptr;
668 if (res == nullptr) {
669 // All users were eliminated.
670 } else if (!res->is_CheckCastPP()) {
671 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
672 can_eliminate = false;
673 } else {
674 worklist.push(res);
675 res_type = igvn->type(res)->isa_oopptr();
676 if (res_type == nullptr) {
677 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
678 can_eliminate = false;
679 } else if (!res_type->klass_is_exact()) {
680 NOT_PRODUCT(fail_eliminate = "Not an exact type.";)
681 can_eliminate = false;
682 } else if (res_type->isa_aryptr()) {
683 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
684 if (length < 0) {
685 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
686 can_eliminate = false;
687 }
688 }
689 }
690
691 while (can_eliminate && worklist.size() > 0) {
692 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
693 res = worklist.pop();
694 for (DUIterator_Fast jmax, j = res->fast_outs(jmax); j < jmax && can_eliminate; j++) {
695 Node* use = res->fast_out(j);
696
697 if (use->is_AddP()) {
698 const TypePtr* addp_type = igvn->type(use)->is_ptr();
699 int offset = addp_type->offset();
700
701 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
702 NOT_PRODUCT(fail_eliminate = "Undefined field reference";)
703 can_eliminate = false;
704 break;
705 }
706 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
707 k < kmax && can_eliminate; k++) {
708 Node* n = use->fast_out(k);
709 if ((n->is_Mem() && n->as_Mem()->is_mismatched_access()) || n->is_LoadFlat() || n->is_StoreFlat()) {
710 DEBUG_ONLY(disq_node = n);
711 NOT_PRODUCT(fail_eliminate = "Mismatched access");
712 can_eliminate = false;
713 }
714 if (!n->is_Store() && n->Opcode() != Op_CastP2X && !bs->is_gc_pre_barrier_node(n) && !reduce_merge_precheck) {
715 DEBUG_ONLY(disq_node = n;)
716 if (n->is_Load() || n->is_LoadStore()) {
717 NOT_PRODUCT(fail_eliminate = "Field load";)
718 } else {
719 NOT_PRODUCT(fail_eliminate = "Not store field reference";)
720 }
721 can_eliminate = false;
722 }
723 }
724 } else if (use->is_ArrayCopy() &&
725 (use->as_ArrayCopy()->is_clonebasic() ||
726 use->as_ArrayCopy()->is_arraycopy_validated() ||
727 use->as_ArrayCopy()->is_copyof_validated() ||
728 use->as_ArrayCopy()->is_copyofrange_validated()) &&
729 use->in(ArrayCopyNode::Dest) == res) {
730 // ok to eliminate
731 } else if (use->is_SafePoint()) {
732 SafePointNode* sfpt = use->as_SafePoint();
733 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
734 // Object is passed as argument.
735 DEBUG_ONLY(disq_node = use;)
736 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
737 can_eliminate = false;
738 }
739 Node* sfptMem = sfpt->memory();
740 if (sfptMem == nullptr || sfptMem->is_top()) {
741 DEBUG_ONLY(disq_node = use;)
742 NOT_PRODUCT(fail_eliminate = "null or TOP memory";)
743 can_eliminate = false;
744 } else if (!reduce_merge_precheck) {
745 assert(!res->is_Phi() || !res->as_Phi()->can_be_inline_type(), "Inline type allocations should not have safepoint uses");
746 safepoints->append_if_missing(sfpt);
747 }
748 } else if (use->is_InlineType() && use->as_InlineType()->get_oop() == res) {
749 // Look at uses
750 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); k < kmax; k++) {
751 Node* u = use->fast_out(k);
752 if (u->is_InlineType()) {
753 // Use in flat field can be eliminated
754 InlineTypeNode* vt = u->as_InlineType();
755 for (uint i = 0; i < vt->field_count(); ++i) {
756 if (vt->field_value(i) == use && !vt->field_is_flat(i)) {
757 can_eliminate = false; // Use in non-flat field
758 break;
759 }
760 }
761 } else {
762 // Add other uses to the worklist to process individually
763 worklist.push(use);
764 }
765 }
766 } else if (use->Opcode() == Op_StoreX && use->in(MemNode::Address) == res) {
767 // Store to mark word of inline type larval buffer
768 assert(res_type->is_inlinetypeptr(), "Unexpected store to mark word");
769 } else if (res_type->is_inlinetypeptr() && (use->Opcode() == Op_MemBarRelease || use->Opcode() == Op_MemBarStoreStore)) {
770 // Inline type buffer allocations are followed by a membar
771 } else if (reduce_merge_precheck &&
772 (use->is_Phi() || use->is_EncodeP() ||
773 use->Opcode() == Op_MemBarRelease ||
774 (UseStoreStoreForCtor && use->Opcode() == Op_MemBarStoreStore))) {
775 // Nothing to do
776 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
777 if (use->is_Phi()) {
778 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
779 NOT_PRODUCT(fail_eliminate = "Object is return value";)
780 } else {
781 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
782 }
783 DEBUG_ONLY(disq_node = use;)
784 } else {
785 if (use->Opcode() == Op_Return) {
786 NOT_PRODUCT(fail_eliminate = "Object is return value";)
787 } else {
788 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
789 }
790 DEBUG_ONLY(disq_node = use;)
791 }
792 can_eliminate = false;
793 } else {
794 assert(use->Opcode() == Op_CastP2X, "should be");
795 assert(!use->has_out_with(Op_OrL), "should have been removed because oop is never null");
796 }
797 }
798 }
799
800 #ifndef PRODUCT
801 if (PrintEliminateAllocations && safepoints != nullptr) {
802 if (can_eliminate) {
803 tty->print("Scalar ");
804 if (res == nullptr)
805 alloc->dump();
806 else
807 res->dump();
808 } else {
809 tty->print("NotScalar (%s)", fail_eliminate);
810 if (res == nullptr)
811 alloc->dump();
812 else
813 res->dump();
814 #ifdef ASSERT
815 if (disq_node != nullptr) {
816 tty->print(" >>>> ");
817 disq_node->dump();
818 }
819 #endif /*ASSERT*/
820 }
821 }
822
823 if (TraceReduceAllocationMerges && !can_eliminate && reduce_merge_precheck) {
824 tty->print_cr("\tCan't eliminate allocation because '%s': ", fail_eliminate != nullptr ? fail_eliminate : "");
825 DEBUG_ONLY(if (disq_node != nullptr) disq_node->dump();)
826 }
827 #endif
828 return can_eliminate;
829 }
830
831 void PhaseMacroExpand::undo_previous_scalarizations(GrowableArray <SafePointNode *> safepoints_done, AllocateNode* alloc) {
832 Node* res = alloc->result_cast();
833 int nfields = 0;
834 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
835
836 if (res != nullptr) {
837 const TypeOopPtr* res_type = _igvn.type(res)->isa_oopptr();
838
839 if (res_type->isa_instptr()) {
840 // find the fields of the class which will be needed for safepoint debug information
841 ciInstanceKlass* iklass = res_type->is_instptr()->instance_klass();
842 nfields = iklass->nof_nonstatic_fields();
843 } else {
844 // find the array's elements which will be needed for safepoint debug information
845 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
846 assert(nfields >= 0, "must be an array klass.");
847 }
848 }
849
850 // rollback processed safepoints
851 while (safepoints_done.length() > 0) {
852 SafePointNode* sfpt_done = safepoints_done.pop();
853 // remove any extra entries we added to the safepoint
854 uint last = sfpt_done->req() - 1;
855 for (int k = 0; k < nfields; k++) {
856 sfpt_done->del_req(last--);
857 }
858 JVMState *jvms = sfpt_done->jvms();
859 jvms->set_endoff(sfpt_done->req());
860 // Now make a pass over the debug information replacing any references
861 // to SafePointScalarObjectNode with the allocated object.
862 int start = jvms->debug_start();
863 int end = jvms->debug_end();
864 for (int i = start; i < end; i++) {
865 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
866 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
867 if (scobj->first_index(jvms) == sfpt_done->req() &&
868 scobj->n_fields() == (uint)nfields) {
869 assert(scobj->alloc() == alloc, "sanity");
870 sfpt_done->set_req(i, res);
871 }
872 }
873 }
874 _igvn._worklist.push(sfpt_done);
875 }
876 }
877
878 void PhaseMacroExpand::process_field_value_at_safepoint(const Type* field_type, Node* field_val, SafePointNode* sfpt, Unique_Node_List* value_worklist) {
879 if (UseCompressedOops && field_type->isa_narrowoop()) {
880 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
881 // to be able scalar replace the allocation.
882 if (field_val->is_EncodeP()) {
883 field_val = field_val->in(1);
884 } else if (!field_val->is_InlineType()) {
885 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
886 }
887 }
888
889 // Keep track of inline types to scalarize them later
890 if (field_val->is_InlineType()) {
891 value_worklist->push(field_val);
892 } else if (field_val->is_Phi()) {
893 PhiNode* phi = field_val->as_Phi();
894 // Eagerly replace inline type phis now since we could be removing an inline type allocation where we must
895 // scalarize all its fields in safepoints.
896 field_val = phi->try_push_inline_types_down(&_igvn, true);
897 if (field_val->is_InlineType()) {
898 value_worklist->push(field_val);
899 }
900 }
901 sfpt->add_req(field_val);
902 }
903
904 bool PhaseMacroExpand::add_array_elems_to_safepoint(AllocateNode* alloc, const TypeAryPtr* array_type, SafePointNode* sfpt, Unique_Node_List* value_worklist) {
905 const Type* elem_type = array_type->elem();
906 BasicType basic_elem_type = elem_type->array_element_basic_type();
907
908 intptr_t elem_size;
909 uint header_size;
910 if (array_type->is_flat()) {
911 elem_size = array_type->flat_elem_size();
912 header_size = arrayOopDesc::base_offset_in_bytes(T_FLAT_ELEMENT);
913 } else {
914 elem_size = type2aelembytes(basic_elem_type);
915 header_size = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
916 }
917
918 int n_elems = alloc->in(AllocateNode::ALength)->get_int();
919 for (int elem_idx = 0; elem_idx < n_elems; elem_idx++) {
920 intptr_t elem_offset = header_size + elem_idx * elem_size;
921 const TypeAryPtr* elem_adr_type = array_type->with_offset(elem_offset);
922 Node* elem_val;
923 if (array_type->is_flat()) {
924 ciInlineKlass* elem_klass = elem_type->inline_klass();
925 assert(elem_klass->maybe_flat_in_array(), "must be flat in array");
926 elem_val = inline_type_from_mem(elem_klass, elem_adr_type, elem_idx, 0, array_type->is_null_free(), alloc, sfpt);
927 } else {
928 elem_val = value_from_mem(sfpt->memory(), sfpt->control(), basic_elem_type, elem_type, elem_adr_type, alloc);
929 #ifndef PRODUCT
930 if (PrintEliminateAllocations && elem_val == nullptr) {
931 tty->print("=== At SafePoint node %d can't find value of array element [%d]", sfpt->_idx, elem_idx);
932 tty->print(", which prevents elimination of: ");
933 alloc->dump();
934 }
935 #endif // PRODUCT
936 }
937 if (elem_val == nullptr) {
938 return false;
939 }
940
941 process_field_value_at_safepoint(elem_type, elem_val, sfpt, value_worklist);
942 }
943
944 return true;
945 }
946
947 // Recursively adds all flattened fields of a type 'iklass' inside 'base' to 'sfpt'.
948 // 'offset_minus_header' refers to the offset of the payload of 'iklass' inside 'base' minus the
949 // payload offset of 'iklass'. If 'base' is of type 'iklass' then 'offset_minus_header' == 0.
950 bool PhaseMacroExpand::add_inst_fields_to_safepoint(ciInstanceKlass* iklass, AllocateNode* alloc, Node* base, int offset_minus_header, SafePointNode* sfpt, Unique_Node_List* value_worklist) {
951 const TypeInstPtr* base_type = _igvn.type(base)->is_instptr();
952 auto report_failure = [&](int offset) {
953 #ifndef PRODUCT
954 if (PrintEliminateAllocations) {
955 ciInstanceKlass* base_klass = base_type->instance_klass();
956 ciField* flattened_field = base_klass->get_field_by_offset(offset, false);
957 assert(flattened_field != nullptr, "must have a field of type %s at offset %d", base_klass->name()->as_utf8(), offset);
958 tty->print("=== At SafePoint node %d can't find value of field: ", sfpt->_idx);
959 flattened_field->print();
960 int field_idx = C->alias_type(flattened_field)->index();
961 tty->print(" (alias_idx=%d)", field_idx);
962 tty->print(", which prevents elimination of: ");
963 base->dump();
964 }
965 #endif // PRODUCT
966 };
967
968 for (int i = 0; i < iklass->nof_declared_nonstatic_fields(); i++) {
969 ciField* field = iklass->declared_nonstatic_field_at(i);
970 if (field->is_flat()) {
971 ciInlineKlass* fvk = field->type()->as_inline_klass();
972 int field_offset_minus_header = offset_minus_header + field->offset_in_bytes() - fvk->payload_offset();
973 bool success = add_inst_fields_to_safepoint(fvk, alloc, base, field_offset_minus_header, sfpt, value_worklist);
974 if (!success) {
975 return false;
976 }
977
978 // The null marker of a field is added right after we scalarize that field
979 if (!field->is_null_free()) {
980 int nm_offset = offset_minus_header + field->null_marker_offset();
981 Node* null_marker = value_from_mem(sfpt->memory(), sfpt->control(), T_BOOLEAN, TypeInt::BOOL, base_type->with_offset(nm_offset), alloc);
982 if (null_marker == nullptr) {
983 report_failure(nm_offset);
984 return false;
985 }
986 process_field_value_at_safepoint(TypeInt::BOOL, null_marker, sfpt, value_worklist);
987 }
988
989 continue;
990 }
991
992 int offset = offset_minus_header + field->offset_in_bytes();
993 ciType* elem_type = field->type();
994 BasicType basic_elem_type = field->layout_type();
995
996 const Type* field_type;
997 if (is_reference_type(basic_elem_type)) {
998 if (!elem_type->is_loaded()) {
999 field_type = TypeInstPtr::BOTTOM;
1000 } else {
1001 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
1002 }
1003 if (UseCompressedOops) {
1004 field_type = field_type->make_narrowoop();
1005 basic_elem_type = T_NARROWOOP;
1006 }
1007 } else {
1008 field_type = Type::get_const_basic_type(basic_elem_type);
1009 }
1010
1011 const TypeInstPtr* field_addr_type = base_type->add_offset(offset)->isa_instptr();
1012 Node* field_val = value_from_mem(sfpt->memory(), sfpt->control(), basic_elem_type, field_type, field_addr_type, alloc);
1013 if (field_val == nullptr) {
1014 report_failure(offset);
1015 return false;
1016 }
1017 process_field_value_at_safepoint(field_type, field_val, sfpt, value_worklist);
1018 }
1019
1020 return true;
1021 }
1022
1023 SafePointScalarObjectNode* PhaseMacroExpand::create_scalarized_object_description(AllocateNode* alloc, SafePointNode* sfpt,
1024 Unique_Node_List* value_worklist) {
1025 // Fields of scalar objs are referenced only at the end
1026 // of regular debuginfo at the last (youngest) JVMS.
1027 // Record relative start index.
1028 ciInstanceKlass* iklass = nullptr;
1029 const TypeOopPtr* res_type = nullptr;
1030 int nfields = 0;
1031 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
1032 Node* res = alloc->result_cast();
1033
1034 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
1035 assert(sfpt->jvms() != nullptr, "missed JVMS");
1036 uint before_sfpt_req = sfpt->req();
1037
1038 if (res != nullptr) { // Could be null when there are no users
1039 res_type = _igvn.type(res)->isa_oopptr();
1040
1041 if (res_type->isa_instptr()) {
1042 // find the fields of the class which will be needed for safepoint debug information
1043 iklass = res_type->is_instptr()->instance_klass();
1044 nfields = iklass->nof_nonstatic_fields();
1045 } else {
1046 // find the array's elements which will be needed for safepoint debug information
1047 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
1048 assert(nfields >= 0, "must be an array klass.");
1049 }
1050
1051 if (res->bottom_type()->is_inlinetypeptr()) {
1052 // Nullable inline types have a null marker field which is added to the safepoint when scalarizing them (see
1053 // InlineTypeNode::make_scalar_in_safepoint()). When having circular inline types, we stop scalarizing at depth 1
1054 // to avoid an endless recursion. Therefore, we do not have a SafePointScalarObjectNode node here, yet.
1055 // We are about to create a SafePointScalarObjectNode as if this is a normal object. Add an additional int input
1056 // with value 1 which sets the null marker to true to indicate that the object is always non-null. This input is checked
1057 // later in PhaseOutput::filLocArray() for inline types.
1058 sfpt->add_req(_igvn.intcon(1));
1059 }
1060 }
1061
1062 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type, alloc, first_ind, sfpt->jvms()->depth(), nfields);
1063 sobj->init_req(0, C->root());
1064 transform_later(sobj);
1065
1066 if (res == nullptr) {
1067 sfpt->jvms()->set_endoff(sfpt->req());
1068 return sobj;
1069 }
1070
1071 bool success;
1072 if (iklass == nullptr) {
1073 success = add_array_elems_to_safepoint(alloc, res_type->is_aryptr(), sfpt, value_worklist);
1074 } else {
1075 success = add_inst_fields_to_safepoint(iklass, alloc, res, 0, sfpt, value_worklist);
1076 }
1077
1078 // We weren't able to find a value for this field, remove all the fields added to the safepoint
1079 if (!success) {
1080 for (uint i = sfpt->req() - 1; i >= before_sfpt_req; i--) {
1081 sfpt->del_req(i);
1082 }
1083 _igvn._worklist.push(sfpt);
1084 return nullptr;
1085 }
1086
1087 sfpt->jvms()->set_endoff(sfpt->req());
1088 return sobj;
1089 }
1090
1091 // Do scalar replacement.
1092 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
1093 GrowableArray <SafePointNode *> safepoints_done;
1094 Node* res = alloc->result_cast();
1095 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
1096 const TypeOopPtr* res_type = nullptr;
1097 if (res != nullptr) { // Could be null when there are no users
1098 res_type = _igvn.type(res)->isa_oopptr();
1099 }
1100
1101 // Process the safepoint uses
1102 Unique_Node_List value_worklist;
1103 while (safepoints.length() > 0) {
1104 SafePointNode* sfpt = safepoints.pop();
1105 SafePointScalarObjectNode* sobj = create_scalarized_object_description(alloc, sfpt, &value_worklist);
1106
1107 if (sobj == nullptr) {
1108 undo_previous_scalarizations(safepoints_done, alloc);
1109 return false;
1110 }
1111
1112 // Now make a pass over the debug information replacing any references
1113 // to the allocated object with "sobj"
1114 JVMState *jvms = sfpt->jvms();
1115 sfpt->replace_edges_in_range(res, sobj, jvms->debug_start(), jvms->debug_end(), &_igvn);
1116 _igvn._worklist.push(sfpt);
1117
1118 // keep it for rollback
1119 safepoints_done.append_if_missing(sfpt);
1120 }
1121 // Scalarize inline types that were added to the safepoint.
1122 // Don't allow linking a constant oop (if available) for flat array elements
1123 // because Deoptimization::reassign_flat_array_elements needs field values.
1124 bool allow_oop = (res_type != nullptr) && !res_type->is_flat();
1125 for (uint i = 0; i < value_worklist.size(); ++i) {
1126 InlineTypeNode* vt = value_worklist.at(i)->as_InlineType();
1127 vt->make_scalar_in_safepoints(&_igvn, allow_oop);
1128 }
1129 return true;
1130 }
1131
1132 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
1133 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
1134 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
1135 if (ctl_proj != nullptr) {
1136 igvn.replace_node(ctl_proj, n->in(0));
1137 }
1138 if (mem_proj != nullptr) {
1139 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
1140 }
1141 }
1142
1143 // Process users of eliminated allocation.
1144 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc, bool inline_alloc) {
1145 Unique_Node_List worklist;
1146 Node* res = alloc->result_cast();
1147 if (res != nullptr) {
1148 worklist.push(res);
1149 }
1150 while (worklist.size() > 0) {
1151 res = worklist.pop();
1152 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
1153 Node *use = res->last_out(j);
1154 uint oc1 = res->outcnt();
1155
1156 if (use->is_AddP()) {
1157 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
1158 Node *n = use->last_out(k);
1159 uint oc2 = use->outcnt();
1160 if (n->is_Store()) {
1161 for (DUIterator_Fast pmax, p = n->fast_outs(pmax); p < pmax; p++) {
1162 MemBarNode* mb = n->fast_out(p)->isa_MemBar();
1163 if (mb != nullptr && mb->req() <= MemBarNode::Precedent && mb->in(MemBarNode::Precedent) == n) {
1164 // MemBarVolatiles should have been removed by MemBarNode::Ideal() for non-inline allocations
1165 assert(inline_alloc, "MemBarVolatile should be eliminated for non-escaping object");
1166 mb->remove(&_igvn);
1167 }
1168 }
1169 _igvn.replace_node(n, n->in(MemNode::Memory));
1170 } else {
1171 eliminate_gc_barrier(n);
1172 }
1173 k -= (oc2 - use->outcnt());
1174 }
1175 _igvn.remove_dead_node(use);
1176 } else if (use->is_ArrayCopy()) {
1177 // Disconnect ArrayCopy node
1178 ArrayCopyNode* ac = use->as_ArrayCopy();
1179 if (ac->is_clonebasic()) {
1180 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
1181 disconnect_projections(ac, _igvn);
1182 assert(alloc->in(TypeFunc::Memory)->is_Proj() && alloc->in(TypeFunc::Memory)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
1183 Node* membar_before = alloc->in(TypeFunc::Memory)->in(0);
1184 disconnect_projections(membar_before->as_MemBar(), _igvn);
1185 if (membar_after->is_MemBar()) {
1186 disconnect_projections(membar_after->as_MemBar(), _igvn);
1187 }
1188 } else {
1189 assert(ac->is_arraycopy_validated() ||
1190 ac->is_copyof_validated() ||
1191 ac->is_copyofrange_validated(), "unsupported");
1192 CallProjections* callprojs = ac->extract_projections(true);
1193
1194 _igvn.replace_node(callprojs->fallthrough_ioproj, ac->in(TypeFunc::I_O));
1195 _igvn.replace_node(callprojs->fallthrough_memproj, ac->in(TypeFunc::Memory));
1196 _igvn.replace_node(callprojs->fallthrough_catchproj, ac->in(TypeFunc::Control));
1197
1198 // Set control to top. IGVN will remove the remaining projections
1199 ac->set_req(0, top());
1200 ac->replace_edge(res, top(), &_igvn);
1201
1202 // Disconnect src right away: it can help find new
1203 // opportunities for allocation elimination
1204 Node* src = ac->in(ArrayCopyNode::Src);
1205 ac->replace_edge(src, top(), &_igvn);
1206 // src can be top at this point if src and dest of the
1207 // arraycopy were the same
1208 if (src->outcnt() == 0 && !src->is_top()) {
1209 _igvn.remove_dead_node(src);
1210 }
1211 }
1212 _igvn._worklist.push(ac);
1213 } else if (use->is_InlineType()) {
1214 assert(use->as_InlineType()->get_oop() == res, "unexpected inline type ptr use");
1215 // Cut off oop input and remove known instance id from type
1216 _igvn.rehash_node_delayed(use);
1217 use->as_InlineType()->set_oop(_igvn, _igvn.zerocon(T_OBJECT));
1218 use->as_InlineType()->set_is_buffered(_igvn, false);
1219 const TypeOopPtr* toop = _igvn.type(use)->is_oopptr()->cast_to_instance_id(TypeOopPtr::InstanceBot);
1220 _igvn.set_type(use, toop);
1221 use->as_InlineType()->set_type(toop);
1222 // Process users
1223 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); k < kmax; k++) {
1224 Node* u = use->fast_out(k);
1225 if (!u->is_InlineType() && !u->is_StoreFlat()) {
1226 worklist.push(u);
1227 }
1228 }
1229 } else if (use->Opcode() == Op_StoreX && use->in(MemNode::Address) == res) {
1230 // Store to mark word of inline type larval buffer
1231 assert(inline_alloc, "Unexpected store to mark word");
1232 _igvn.replace_node(use, use->in(MemNode::Memory));
1233 } else if (use->Opcode() == Op_MemBarRelease || use->Opcode() == Op_MemBarStoreStore) {
1234 // Inline type buffer allocations are followed by a membar
1235 assert(inline_alloc, "Unexpected MemBarRelease");
1236 use->as_MemBar()->remove(&_igvn);
1237 } else {
1238 eliminate_gc_barrier(use);
1239 }
1240 j -= (oc1 - res->outcnt());
1241 }
1242 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
1243 _igvn.remove_dead_node(res);
1244 }
1245
1246 //
1247 // Process other users of allocation's projections
1248 //
1249 if (_callprojs->resproj[0] != nullptr && _callprojs->resproj[0]->outcnt() != 0) {
1250 // First disconnect stores captured by Initialize node.
1251 // If Initialize node is eliminated first in the following code,
1252 // it will kill such stores and DUIterator_Last will assert.
1253 for (DUIterator_Fast jmax, j = _callprojs->resproj[0]->fast_outs(jmax); j < jmax; j++) {
1254 Node* use = _callprojs->resproj[0]->fast_out(j);
1255 if (use->is_AddP()) {
1256 // raw memory addresses used only by the initialization
1257 _igvn.replace_node(use, C->top());
1258 --j; --jmax;
1259 }
1260 }
1261 for (DUIterator_Last jmin, j = _callprojs->resproj[0]->last_outs(jmin); j >= jmin; ) {
1262 Node* use = _callprojs->resproj[0]->last_out(j);
1263 uint oc1 = _callprojs->resproj[0]->outcnt();
1264 if (use->is_Initialize()) {
1265 // Eliminate Initialize node.
1266 InitializeNode *init = use->as_Initialize();
1267 assert(init->outcnt() <= 2, "only a control and memory projection expected");
1268 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
1269 if (ctrl_proj != nullptr) {
1270 _igvn.replace_node(ctrl_proj, init->in(TypeFunc::Control));
1271 #ifdef ASSERT
1272 // If the InitializeNode has no memory out, it will die, and tmp will become null
1273 Node* tmp = init->in(TypeFunc::Control);
1274 assert(tmp == nullptr || tmp == _callprojs->fallthrough_catchproj, "allocation control projection");
1275 #endif
1276 }
1277 Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory);
1278 if (mem_proj != nullptr) {
1279 Node *mem = init->in(TypeFunc::Memory);
1280 #ifdef ASSERT
1281 if (mem->is_MergeMem()) {
1282 assert(mem->in(TypeFunc::Memory) == _callprojs->fallthrough_memproj, "allocation memory projection");
1283 } else {
1284 assert(mem == _callprojs->fallthrough_memproj, "allocation memory projection");
1285 }
1286 #endif
1287 _igvn.replace_node(mem_proj, mem);
1288 }
1289 } else if (use->Opcode() == Op_MemBarStoreStore) {
1290 // Inline type buffer allocations are followed by a membar
1291 assert(inline_alloc, "Unexpected MemBarStoreStore");
1292 use->as_MemBar()->remove(&_igvn);
1293 } else {
1294 assert(false, "only Initialize or AddP expected");
1295 }
1296 j -= (oc1 - _callprojs->resproj[0]->outcnt());
1297 }
1298 }
1299 if (_callprojs->fallthrough_catchproj != nullptr) {
1300 _igvn.replace_node(_callprojs->fallthrough_catchproj, alloc->in(TypeFunc::Control));
1301 }
1302 if (_callprojs->fallthrough_memproj != nullptr) {
1303 _igvn.replace_node(_callprojs->fallthrough_memproj, alloc->in(TypeFunc::Memory));
1304 }
1305 if (_callprojs->catchall_memproj != nullptr) {
1306 _igvn.replace_node(_callprojs->catchall_memproj, C->top());
1307 }
1308 if (_callprojs->fallthrough_ioproj != nullptr) {
1309 _igvn.replace_node(_callprojs->fallthrough_ioproj, alloc->in(TypeFunc::I_O));
1310 }
1311 if (_callprojs->catchall_ioproj != nullptr) {
1312 _igvn.replace_node(_callprojs->catchall_ioproj, C->top());
1313 }
1314 if (_callprojs->catchall_catchproj != nullptr) {
1315 _igvn.replace_node(_callprojs->catchall_catchproj, C->top());
1316 }
1317 }
1318
1319 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1320 // If reallocation fails during deoptimization we'll pop all
1321 // interpreter frames for this compiled frame and that won't play
1322 // nice with JVMTI popframe.
1323 // We avoid this issue by eager reallocation when the popframe request
1324 // is received.
1325 if (!EliminateAllocations) {
1326 return false;
1327 }
1328 Node* klass = alloc->in(AllocateNode::KlassNode);
1329 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1330
1331 // Attempt to eliminate inline type buffer allocations
1332 // regardless of usage and escape/replaceable status.
1333 bool inline_alloc = tklass->isa_instklassptr() &&
1334 tklass->is_instklassptr()->instance_klass()->is_inlinetype();
1335 if (!alloc->_is_non_escaping && !inline_alloc) {
1336 return false;
1337 }
1338 // Eliminate boxing allocations which are not used
1339 // regardless scalar replaceable status.
1340 Node* res = alloc->result_cast();
1341 bool boxing_alloc = (res == nullptr) && C->eliminate_boxing() &&
1342 tklass->isa_instklassptr() &&
1343 tklass->is_instklassptr()->instance_klass()->is_box_klass();
1344 if (!alloc->_is_scalar_replaceable && !boxing_alloc && !inline_alloc) {
1345 return false;
1346 }
1347
1348 _callprojs = alloc->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1349
1350 GrowableArray <SafePointNode *> safepoints;
1351 if (!can_eliminate_allocation(&_igvn, alloc, &safepoints)) {
1352 return false;
1353 }
1354
1355 if (!alloc->_is_scalar_replaceable) {
1356 assert(res == nullptr || inline_alloc, "sanity");
1357 // We can only eliminate allocation if all debug info references
1358 // are already replaced with SafePointScalarObject because
1359 // we can't search for a fields value without instance_id.
1360 if (safepoints.length() > 0) {
1361 return false;
1362 }
1363 }
1364
1365 if (!scalar_replacement(alloc, safepoints)) {
1366 return false;
1367 }
1368
1369 CompileLog* log = C->log();
1370 if (log != nullptr) {
1371 log->head("eliminate_allocation type='%d'",
1372 log->identify(tklass->exact_klass()));
1373 JVMState* p = alloc->jvms();
1374 while (p != nullptr) {
1375 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1376 p = p->caller();
1377 }
1378 log->tail("eliminate_allocation");
1379 }
1380
1381 process_users_of_allocation(alloc, inline_alloc);
1382
1383 #ifndef PRODUCT
1384 if (PrintEliminateAllocations) {
1385 if (alloc->is_AllocateArray())
1386 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1387 else
1388 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1389 }
1390 #endif
1391
1392 return true;
1393 }
1394
1395 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1396 // EA should remove all uses of non-escaping boxing node.
1397 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != nullptr) {
1398 return false;
1399 }
1400
1401 assert(boxing->result_cast() == nullptr, "unexpected boxing node result");
1402
1403 _callprojs = boxing->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1404
1405 const TypeTuple* r = boxing->tf()->range_sig();
1406 assert(r->cnt() > TypeFunc::Parms, "sanity");
1407 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1408 assert(t != nullptr, "sanity");
1409
1410 CompileLog* log = C->log();
1411 if (log != nullptr) {
1412 log->head("eliminate_boxing type='%d'",
1413 log->identify(t->instance_klass()));
1414 JVMState* p = boxing->jvms();
1415 while (p != nullptr) {
1416 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1417 p = p->caller();
1418 }
1419 log->tail("eliminate_boxing");
1420 }
1421
1422 process_users_of_allocation(boxing);
1423
1424 #ifndef PRODUCT
1425 if (PrintEliminateAllocations) {
1426 tty->print("++++ Eliminated: %d ", boxing->_idx);
1427 boxing->method()->print_short_name(tty);
1428 tty->cr();
1429 }
1430 #endif
1431
1432 return true;
1433 }
1434
1435
1436 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1437 Node* adr = basic_plus_adr(base, offset);
1438 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1439 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered);
1440 transform_later(value);
1441 return value;
1442 }
1443
1444
1445 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1446 Node* adr = basic_plus_adr(base, offset);
1447 mem = StoreNode::make(_igvn, ctl, mem, adr, nullptr, value, bt, MemNode::unordered);
1448 transform_later(mem);
1449 return mem;
1450 }
1451
1452 //=============================================================================
1453 //
1454 // A L L O C A T I O N
1455 //
1456 // Allocation attempts to be fast in the case of frequent small objects.
1457 // It breaks down like this:
1458 //
1459 // 1) Size in doublewords is computed. This is a constant for objects and
1460 // variable for most arrays. Doubleword units are used to avoid size
1461 // overflow of huge doubleword arrays. We need doublewords in the end for
1462 // rounding.
1463 //
1464 // 2) Size is checked for being 'too large'. Too-large allocations will go
1465 // the slow path into the VM. The slow path can throw any required
1466 // exceptions, and does all the special checks for very large arrays. The
1467 // size test can constant-fold away for objects. For objects with
1468 // finalizers it constant-folds the otherway: you always go slow with
1469 // finalizers.
1470 //
1471 // 3) If NOT using TLABs, this is the contended loop-back point.
1472 // Load-Locked the heap top. If using TLABs normal-load the heap top.
1473 //
1474 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
1475 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
1476 // "size*8" we always enter the VM, where "largish" is a constant picked small
1477 // enough that there's always space between the eden max and 4Gig (old space is
1478 // there so it's quite large) and large enough that the cost of entering the VM
1479 // is dwarfed by the cost to initialize the space.
1480 //
1481 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1482 // down. If contended, repeat at step 3. If using TLABs normal-store
1483 // adjusted heap top back down; there is no contention.
1484 //
1485 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
1486 // fields.
1487 //
1488 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1489 // oop flavor.
1490 //
1491 //=============================================================================
1492 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1493 // Allocations bigger than this always go the slow route.
1494 // This value must be small enough that allocation attempts that need to
1495 // trigger exceptions go the slow route. Also, it must be small enough so
1496 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1497 //=============================================================================j//
1498 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1499 // The allocator will coalesce int->oop copies away. See comment in
1500 // coalesce.cpp about how this works. It depends critically on the exact
1501 // code shape produced here, so if you are changing this code shape
1502 // make sure the GC info for the heap-top is correct in and around the
1503 // slow-path call.
1504 //
1505
1506 void PhaseMacroExpand::expand_allocate_common(
1507 AllocateNode* alloc, // allocation node to be expanded
1508 Node* length, // array length for an array allocation
1509 Node* init_val, // value to initialize the array with
1510 const TypeFunc* slow_call_type, // Type of slow call
1511 address slow_call_address, // Address of slow call
1512 Node* valid_length_test // whether length is valid or not
1513 )
1514 {
1515 Node* ctrl = alloc->in(TypeFunc::Control);
1516 Node* mem = alloc->in(TypeFunc::Memory);
1517 Node* i_o = alloc->in(TypeFunc::I_O);
1518 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1519 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1520 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1521 assert(ctrl != nullptr, "must have control");
1522
1523 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1524 // they will not be used if "always_slow" is set
1525 enum { slow_result_path = 1, fast_result_path = 2 };
1526 Node *result_region = nullptr;
1527 Node *result_phi_rawmem = nullptr;
1528 Node *result_phi_rawoop = nullptr;
1529 Node *result_phi_i_o = nullptr;
1530
1531 // The initial slow comparison is a size check, the comparison
1532 // we want to do is a BoolTest::gt
1533 bool expand_fast_path = true;
1534 int tv = _igvn.find_int_con(initial_slow_test, -1);
1535 if (tv >= 0) {
1536 // InitialTest has constant result
1537 // 0 - can fit in TLAB
1538 // 1 - always too big or negative
1539 assert(tv <= 1, "0 or 1 if a constant");
1540 expand_fast_path = (tv == 0);
1541 initial_slow_test = nullptr;
1542 } else {
1543 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1544 }
1545
1546 if (!UseTLAB) {
1547 // Force slow-path allocation
1548 expand_fast_path = false;
1549 initial_slow_test = nullptr;
1550 }
1551
1552 bool allocation_has_use = (alloc->result_cast() != nullptr);
1553 if (!allocation_has_use) {
1554 InitializeNode* init = alloc->initialization();
1555 if (init != nullptr) {
1556 init->remove(&_igvn);
1557 }
1558 if (expand_fast_path && (initial_slow_test == nullptr)) {
1559 // Remove allocation node and return.
1560 // Size is a non-negative constant -> no initial check needed -> directly to fast path.
1561 // Also, no usages -> empty fast path -> no fall out to slow path -> nothing left.
1562 #ifndef PRODUCT
1563 if (PrintEliminateAllocations) {
1564 tty->print("NotUsed ");
1565 Node* res = alloc->proj_out_or_null(TypeFunc::Parms);
1566 if (res != nullptr) {
1567 res->dump();
1568 } else {
1569 alloc->dump();
1570 }
1571 }
1572 #endif
1573 yank_alloc_node(alloc);
1574 return;
1575 }
1576 }
1577
1578 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1579 Node *slow_region = nullptr;
1580 Node *toobig_false = ctrl;
1581
1582 // generate the initial test if necessary
1583 if (initial_slow_test != nullptr ) {
1584 assert (expand_fast_path, "Only need test if there is a fast path");
1585 slow_region = new RegionNode(3);
1586
1587 // Now make the initial failure test. Usually a too-big test but
1588 // might be a TRUE for finalizers.
1589 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1590 transform_later(toobig_iff);
1591 // Plug the failing-too-big test into the slow-path region
1592 Node* toobig_true = new IfTrueNode(toobig_iff);
1593 transform_later(toobig_true);
1594 slow_region ->init_req( too_big_or_final_path, toobig_true );
1595 toobig_false = new IfFalseNode(toobig_iff);
1596 transform_later(toobig_false);
1597 } else {
1598 // No initial test, just fall into next case
1599 assert(allocation_has_use || !expand_fast_path, "Should already have been handled");
1600 toobig_false = ctrl;
1601 DEBUG_ONLY(slow_region = NodeSentinel);
1602 }
1603
1604 // If we are here there are several possibilities
1605 // - expand_fast_path is false - then only a slow path is expanded. That's it.
1606 // no_initial_check means a constant allocation.
1607 // - If check always evaluates to false -> expand_fast_path is false (see above)
1608 // - If check always evaluates to true -> directly into fast path (but may bailout to slowpath)
1609 // if !allocation_has_use the fast path is empty
1610 // if !allocation_has_use && no_initial_check
1611 // - Then there are no fastpath that can fall out to slowpath -> no allocation code at all.
1612 // removed by yank_alloc_node above.
1613
1614 Node *slow_mem = mem; // save the current memory state for slow path
1615 // generate the fast allocation code unless we know that the initial test will always go slow
1616 if (expand_fast_path) {
1617 // Fast path modifies only raw memory.
1618 if (mem->is_MergeMem()) {
1619 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1620 }
1621
1622 // allocate the Region and Phi nodes for the result
1623 result_region = new RegionNode(3);
1624 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1625 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1626
1627 // Grab regular I/O before optional prefetch may change it.
1628 // Slow-path does no I/O so just set it to the original I/O.
1629 result_phi_i_o->init_req(slow_result_path, i_o);
1630
1631 // Name successful fast-path variables
1632 Node* fast_oop_ctrl;
1633 Node* fast_oop_rawmem;
1634
1635 if (allocation_has_use) {
1636 Node* needgc_ctrl = nullptr;
1637 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1638
1639 intx prefetch_lines = length != nullptr ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1640 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1641 Node* fast_oop = bs->obj_allocate(this, mem, toobig_false, size_in_bytes, i_o, needgc_ctrl,
1642 fast_oop_ctrl, fast_oop_rawmem,
1643 prefetch_lines);
1644
1645 if (initial_slow_test != nullptr) {
1646 // This completes all paths into the slow merge point
1647 slow_region->init_req(need_gc_path, needgc_ctrl);
1648 transform_later(slow_region);
1649 } else {
1650 // No initial slow path needed!
1651 // Just fall from the need-GC path straight into the VM call.
1652 slow_region = needgc_ctrl;
1653 }
1654
1655 InitializeNode* init = alloc->initialization();
1656 fast_oop_rawmem = initialize_object(alloc,
1657 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1658 klass_node, length, size_in_bytes);
1659 expand_initialize_membar(alloc, init, fast_oop_ctrl, fast_oop_rawmem);
1660 expand_dtrace_alloc_probe(alloc, fast_oop, fast_oop_ctrl, fast_oop_rawmem);
1661
1662 result_phi_rawoop->init_req(fast_result_path, fast_oop);
1663 } else {
1664 assert (initial_slow_test != nullptr, "sanity");
1665 fast_oop_ctrl = toobig_false;
1666 fast_oop_rawmem = mem;
1667 transform_later(slow_region);
1668 }
1669
1670 // Plug in the successful fast-path into the result merge point
1671 result_region ->init_req(fast_result_path, fast_oop_ctrl);
1672 result_phi_i_o ->init_req(fast_result_path, i_o);
1673 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1674 } else {
1675 slow_region = ctrl;
1676 result_phi_i_o = i_o; // Rename it to use in the following code.
1677 }
1678
1679 // Generate slow-path call
1680 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1681 OptoRuntime::stub_name(slow_call_address),
1682 TypePtr::BOTTOM);
1683 call->init_req(TypeFunc::Control, slow_region);
1684 call->init_req(TypeFunc::I_O, top()); // does no i/o
1685 call->init_req(TypeFunc::Memory, slow_mem); // may gc ptrs
1686 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1687 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1688
1689 call->init_req(TypeFunc::Parms+0, klass_node);
1690 if (length != nullptr) {
1691 call->init_req(TypeFunc::Parms+1, length);
1692 if (init_val != nullptr) {
1693 call->init_req(TypeFunc::Parms+2, init_val);
1694 }
1695 } else {
1696 // Let the runtime know if this is a larval allocation
1697 call->init_req(TypeFunc::Parms+1, _igvn.intcon(alloc->_larval));
1698 }
1699
1700 // Copy debug information and adjust JVMState information, then replace
1701 // allocate node with the call
1702 call->copy_call_debug_info(&_igvn, alloc);
1703 // For array allocations, copy the valid length check to the call node so Compile::final_graph_reshaping() can verify
1704 // that the call has the expected number of CatchProj nodes (in case the allocation always fails and the fallthrough
1705 // path dies).
1706 if (valid_length_test != nullptr) {
1707 call->add_req(valid_length_test);
1708 }
1709 if (expand_fast_path) {
1710 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1711 } else {
1712 // Hook i_o projection to avoid its elimination during allocation
1713 // replacement (when only a slow call is generated).
1714 call->set_req(TypeFunc::I_O, result_phi_i_o);
1715 }
1716 _igvn.replace_node(alloc, call);
1717 transform_later(call);
1718
1719 // Identify the output projections from the allocate node and
1720 // adjust any references to them.
1721 // The control and io projections look like:
1722 //
1723 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1724 // Allocate Catch
1725 // ^---Proj(io) <-------+ ^---CatchProj(io)
1726 //
1727 // We are interested in the CatchProj nodes.
1728 //
1729 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1730
1731 // An allocate node has separate memory projections for the uses on
1732 // the control and i_o paths. Replace the control memory projection with
1733 // result_phi_rawmem (unless we are only generating a slow call when
1734 // both memory projections are combined)
1735 if (expand_fast_path && _callprojs->fallthrough_memproj != nullptr) {
1736 _igvn.replace_in_uses(_callprojs->fallthrough_memproj, result_phi_rawmem);
1737 }
1738 // Now change uses of catchall_memproj to use fallthrough_memproj and delete
1739 // catchall_memproj so we end up with a call that has only 1 memory projection.
1740 if (_callprojs->catchall_memproj != nullptr) {
1741 if (_callprojs->fallthrough_memproj == nullptr) {
1742 _callprojs->fallthrough_memproj = new ProjNode(call, TypeFunc::Memory);
1743 transform_later(_callprojs->fallthrough_memproj);
1744 }
1745 _igvn.replace_in_uses(_callprojs->catchall_memproj, _callprojs->fallthrough_memproj);
1746 _igvn.remove_dead_node(_callprojs->catchall_memproj);
1747 }
1748
1749 // An allocate node has separate i_o projections for the uses on the control
1750 // and i_o paths. Always replace the control i_o projection with result i_o
1751 // otherwise incoming i_o become dead when only a slow call is generated
1752 // (it is different from memory projections where both projections are
1753 // combined in such case).
1754 if (_callprojs->fallthrough_ioproj != nullptr) {
1755 _igvn.replace_in_uses(_callprojs->fallthrough_ioproj, result_phi_i_o);
1756 }
1757 // Now change uses of catchall_ioproj to use fallthrough_ioproj and delete
1758 // catchall_ioproj so we end up with a call that has only 1 i_o projection.
1759 if (_callprojs->catchall_ioproj != nullptr) {
1760 if (_callprojs->fallthrough_ioproj == nullptr) {
1761 _callprojs->fallthrough_ioproj = new ProjNode(call, TypeFunc::I_O);
1762 transform_later(_callprojs->fallthrough_ioproj);
1763 }
1764 _igvn.replace_in_uses(_callprojs->catchall_ioproj, _callprojs->fallthrough_ioproj);
1765 _igvn.remove_dead_node(_callprojs->catchall_ioproj);
1766 }
1767
1768 // if we generated only a slow call, we are done
1769 if (!expand_fast_path) {
1770 // Now we can unhook i_o.
1771 if (result_phi_i_o->outcnt() > 1) {
1772 call->set_req(TypeFunc::I_O, top());
1773 } else {
1774 assert(result_phi_i_o->unique_ctrl_out() == call, "sanity");
1775 // Case of new array with negative size known during compilation.
1776 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1777 // following code since call to runtime will throw exception.
1778 // As result there will be no users of i_o after the call.
1779 // Leave i_o attached to this call to avoid problems in preceding graph.
1780 }
1781 return;
1782 }
1783
1784 if (_callprojs->fallthrough_catchproj != nullptr) {
1785 ctrl = _callprojs->fallthrough_catchproj->clone();
1786 transform_later(ctrl);
1787 _igvn.replace_node(_callprojs->fallthrough_catchproj, result_region);
1788 } else {
1789 ctrl = top();
1790 }
1791 Node *slow_result;
1792 if (_callprojs->resproj[0] == nullptr) {
1793 // no uses of the allocation result
1794 slow_result = top();
1795 } else {
1796 slow_result = _callprojs->resproj[0]->clone();
1797 transform_later(slow_result);
1798 _igvn.replace_node(_callprojs->resproj[0], result_phi_rawoop);
1799 }
1800
1801 // Plug slow-path into result merge point
1802 result_region->init_req( slow_result_path, ctrl);
1803 transform_later(result_region);
1804 if (allocation_has_use) {
1805 result_phi_rawoop->init_req(slow_result_path, slow_result);
1806 transform_later(result_phi_rawoop);
1807 }
1808 result_phi_rawmem->init_req(slow_result_path, _callprojs->fallthrough_memproj);
1809 transform_later(result_phi_rawmem);
1810 transform_later(result_phi_i_o);
1811 // This completes all paths into the result merge point
1812 }
1813
1814 // Remove alloc node that has no uses.
1815 void PhaseMacroExpand::yank_alloc_node(AllocateNode* alloc) {
1816 Node* ctrl = alloc->in(TypeFunc::Control);
1817 Node* mem = alloc->in(TypeFunc::Memory);
1818 Node* i_o = alloc->in(TypeFunc::I_O);
1819
1820 _callprojs = alloc->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1821 if (_callprojs->resproj[0] != nullptr) {
1822 for (DUIterator_Fast imax, i = _callprojs->resproj[0]->fast_outs(imax); i < imax; i++) {
1823 Node* use = _callprojs->resproj[0]->fast_out(i);
1824 use->isa_MemBar()->remove(&_igvn);
1825 --imax;
1826 --i; // back up iterator
1827 }
1828 assert(_callprojs->resproj[0]->outcnt() == 0, "all uses must be deleted");
1829 _igvn.remove_dead_node(_callprojs->resproj[0]);
1830 }
1831 if (_callprojs->fallthrough_catchproj != nullptr) {
1832 _igvn.replace_in_uses(_callprojs->fallthrough_catchproj, ctrl);
1833 _igvn.remove_dead_node(_callprojs->fallthrough_catchproj);
1834 }
1835 if (_callprojs->catchall_catchproj != nullptr) {
1836 _igvn.rehash_node_delayed(_callprojs->catchall_catchproj);
1837 _callprojs->catchall_catchproj->set_req(0, top());
1838 }
1839 if (_callprojs->fallthrough_proj != nullptr) {
1840 Node* catchnode = _callprojs->fallthrough_proj->unique_ctrl_out();
1841 _igvn.remove_dead_node(catchnode);
1842 _igvn.remove_dead_node(_callprojs->fallthrough_proj);
1843 }
1844 if (_callprojs->fallthrough_memproj != nullptr) {
1845 _igvn.replace_in_uses(_callprojs->fallthrough_memproj, mem);
1846 _igvn.remove_dead_node(_callprojs->fallthrough_memproj);
1847 }
1848 if (_callprojs->fallthrough_ioproj != nullptr) {
1849 _igvn.replace_in_uses(_callprojs->fallthrough_ioproj, i_o);
1850 _igvn.remove_dead_node(_callprojs->fallthrough_ioproj);
1851 }
1852 if (_callprojs->catchall_memproj != nullptr) {
1853 _igvn.rehash_node_delayed(_callprojs->catchall_memproj);
1854 _callprojs->catchall_memproj->set_req(0, top());
1855 }
1856 if (_callprojs->catchall_ioproj != nullptr) {
1857 _igvn.rehash_node_delayed(_callprojs->catchall_ioproj);
1858 _callprojs->catchall_ioproj->set_req(0, top());
1859 }
1860 #ifndef PRODUCT
1861 if (PrintEliminateAllocations) {
1862 if (alloc->is_AllocateArray()) {
1863 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1864 } else {
1865 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1866 }
1867 }
1868 #endif
1869 _igvn.remove_dead_node(alloc);
1870 }
1871
1872 void PhaseMacroExpand::expand_initialize_membar(AllocateNode* alloc, InitializeNode* init,
1873 Node*& fast_oop_ctrl, Node*& fast_oop_rawmem) {
1874 // If initialization is performed by an array copy, any required
1875 // MemBarStoreStore was already added. If the object does not
1876 // escape no need for a MemBarStoreStore. If the object does not
1877 // escape in its initializer and memory barrier (MemBarStoreStore or
1878 // stronger) is already added at exit of initializer, also no need
1879 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore
1880 // so that stores that initialize this object can't be reordered
1881 // with a subsequent store that makes this object accessible by
1882 // other threads.
1883 // Other threads include java threads and JVM internal threads
1884 // (for example concurrent GC threads). Current concurrent GC
1885 // implementation: G1 will not scan newly created object,
1886 // so it's safe to skip storestore barrier when allocation does
1887 // not escape.
1888 if (!alloc->does_not_escape_thread() &&
1889 !alloc->is_allocation_MemBar_redundant() &&
1890 (init == nullptr || !init->is_complete_with_arraycopy())) {
1891 if (init == nullptr || init->req() < InitializeNode::RawStores) {
1892 // No InitializeNode or no stores captured by zeroing
1893 // elimination. Simply add the MemBarStoreStore after object
1894 // initialization.
1895 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1896 transform_later(mb);
1897
1898 mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
1899 mb->init_req(TypeFunc::Control, fast_oop_ctrl);
1900 fast_oop_ctrl = new ProjNode(mb, TypeFunc::Control);
1901 transform_later(fast_oop_ctrl);
1902 fast_oop_rawmem = new ProjNode(mb, TypeFunc::Memory);
1903 transform_later(fast_oop_rawmem);
1904 } else {
1905 // Add the MemBarStoreStore after the InitializeNode so that
1906 // all stores performing the initialization that were moved
1907 // before the InitializeNode happen before the storestore
1908 // barrier.
1909
1910 Node* init_ctrl = init->proj_out_or_null(TypeFunc::Control);
1911 Node* init_mem = init->proj_out_or_null(TypeFunc::Memory);
1912
1913 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1914 transform_later(mb);
1915
1916 Node* ctrl = new ProjNode(init, TypeFunc::Control);
1917 transform_later(ctrl);
1918 Node* mem = new ProjNode(init, TypeFunc::Memory);
1919 transform_later(mem);
1920
1921 // The MemBarStoreStore depends on control and memory coming
1922 // from the InitializeNode
1923 mb->init_req(TypeFunc::Memory, mem);
1924 mb->init_req(TypeFunc::Control, ctrl);
1925
1926 ctrl = new ProjNode(mb, TypeFunc::Control);
1927 transform_later(ctrl);
1928 mem = new ProjNode(mb, TypeFunc::Memory);
1929 transform_later(mem);
1930
1931 // All nodes that depended on the InitializeNode for control
1932 // and memory must now depend on the MemBarNode that itself
1933 // depends on the InitializeNode
1934 if (init_ctrl != nullptr) {
1935 _igvn.replace_node(init_ctrl, ctrl);
1936 }
1937 if (init_mem != nullptr) {
1938 _igvn.replace_node(init_mem, mem);
1939 }
1940 }
1941 }
1942 }
1943
1944 void PhaseMacroExpand::expand_dtrace_alloc_probe(AllocateNode* alloc, Node* oop,
1945 Node*& ctrl, Node*& rawmem) {
1946 if (C->env()->dtrace_alloc_probes()) {
1947 // Slow-path call
1948 int size = TypeFunc::Parms + 2;
1949 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1950 CAST_FROM_FN_PTR(address,
1951 static_cast<int (*)(JavaThread*, oopDesc*)>(SharedRuntime::dtrace_object_alloc)),
1952 "dtrace_object_alloc",
1953 TypeRawPtr::BOTTOM);
1954
1955 // Get base of thread-local storage area
1956 Node* thread = new ThreadLocalNode();
1957 transform_later(thread);
1958
1959 call->init_req(TypeFunc::Parms + 0, thread);
1960 call->init_req(TypeFunc::Parms + 1, oop);
1961 call->init_req(TypeFunc::Control, ctrl);
1962 call->init_req(TypeFunc::I_O , top()); // does no i/o
1963 call->init_req(TypeFunc::Memory , rawmem);
1964 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1965 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1966 transform_later(call);
1967 ctrl = new ProjNode(call, TypeFunc::Control);
1968 transform_later(ctrl);
1969 rawmem = new ProjNode(call, TypeFunc::Memory);
1970 transform_later(rawmem);
1971 }
1972 }
1973
1974 // Helper for PhaseMacroExpand::expand_allocate_common.
1975 // Initializes the newly-allocated storage.
1976 Node* PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1977 Node* control, Node* rawmem, Node* object,
1978 Node* klass_node, Node* length,
1979 Node* size_in_bytes) {
1980 InitializeNode* init = alloc->initialization();
1981 // Store the klass & mark bits
1982 Node* mark_node = alloc->make_ideal_mark(&_igvn, control, rawmem);
1983 if (!mark_node->is_Con()) {
1984 transform_later(mark_node);
1985 }
1986 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, TypeX_X->basic_type());
1987
1988 if (!UseCompactObjectHeaders) {
1989 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1990 }
1991 int header_size = alloc->minimum_header_size(); // conservatively small
1992
1993 // Array length
1994 if (length != nullptr) { // Arrays need length field
1995 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1996 // conservatively small header size:
1997 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1998 if (_igvn.type(klass_node)->isa_aryklassptr()) { // we know the exact header size in most cases:
1999 BasicType elem = _igvn.type(klass_node)->is_klassptr()->as_instance_type()->isa_aryptr()->elem()->array_element_basic_type();
2000 if (is_reference_type(elem, true)) {
2001 elem = T_OBJECT;
2002 }
2003 header_size = Klass::layout_helper_header_size(Klass::array_layout_helper(elem));
2004 }
2005 }
2006
2007 // Clear the object body, if necessary.
2008 if (init == nullptr) {
2009 // The init has somehow disappeared; be cautious and clear everything.
2010 //
2011 // This can happen if a node is allocated but an uncommon trap occurs
2012 // immediately. In this case, the Initialize gets associated with the
2013 // trap, and may be placed in a different (outer) loop, if the Allocate
2014 // is in a loop. If (this is rare) the inner loop gets unrolled, then
2015 // there can be two Allocates to one Initialize. The answer in all these
2016 // edge cases is safety first. It is always safe to clear immediately
2017 // within an Allocate, and then (maybe or maybe not) clear some more later.
2018 if (!(UseTLAB && ZeroTLAB)) {
2019 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
2020 alloc->in(AllocateNode::InitValue),
2021 alloc->in(AllocateNode::RawInitValue),
2022 header_size, size_in_bytes,
2023 &_igvn);
2024 }
2025 } else {
2026 if (!init->is_complete()) {
2027 // Try to win by zeroing only what the init does not store.
2028 // We can also try to do some peephole optimizations,
2029 // such as combining some adjacent subword stores.
2030 rawmem = init->complete_stores(control, rawmem, object,
2031 header_size, size_in_bytes, &_igvn);
2032 }
2033 // We have no more use for this link, since the AllocateNode goes away:
2034 init->set_req(InitializeNode::RawAddress, top());
2035 // (If we keep the link, it just confuses the register allocator,
2036 // who thinks he sees a real use of the address by the membar.)
2037 }
2038
2039 return rawmem;
2040 }
2041
2042 // Generate prefetch instructions for next allocations.
2043 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
2044 Node*& contended_phi_rawmem,
2045 Node* old_eden_top, Node* new_eden_top,
2046 intx lines) {
2047 enum { fall_in_path = 1, pf_path = 2 };
2048 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
2049 // Generate prefetch allocation with watermark check.
2050 // As an allocation hits the watermark, we will prefetch starting
2051 // at a "distance" away from watermark.
2052
2053 Node *pf_region = new RegionNode(3);
2054 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY,
2055 TypeRawPtr::BOTTOM );
2056 // I/O is used for Prefetch
2057 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO );
2058
2059 Node *thread = new ThreadLocalNode();
2060 transform_later(thread);
2061
2062 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread,
2063 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
2064 transform_later(eden_pf_adr);
2065
2066 Node *old_pf_wm = new LoadPNode(needgc_false,
2067 contended_phi_rawmem, eden_pf_adr,
2068 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM,
2069 MemNode::unordered);
2070 transform_later(old_pf_wm);
2071
2072 // check against new_eden_top
2073 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm );
2074 transform_later(need_pf_cmp);
2075 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge );
2076 transform_later(need_pf_bol);
2077 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol,
2078 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
2079 transform_later(need_pf_iff);
2080
2081 // true node, add prefetchdistance
2082 Node *need_pf_true = new IfTrueNode( need_pf_iff );
2083 transform_later(need_pf_true);
2084
2085 Node *need_pf_false = new IfFalseNode( need_pf_iff );
2086 transform_later(need_pf_false);
2087
2088 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm,
2089 _igvn.MakeConX(AllocatePrefetchDistance) );
2090 transform_later(new_pf_wmt );
2091 new_pf_wmt->set_req(0, need_pf_true);
2092
2093 Node *store_new_wmt = new StorePNode(need_pf_true,
2094 contended_phi_rawmem, eden_pf_adr,
2095 TypeRawPtr::BOTTOM, new_pf_wmt,
2096 MemNode::unordered);
2097 transform_later(store_new_wmt);
2098
2099 // adding prefetches
2100 pf_phi_abio->init_req( fall_in_path, i_o );
2101
2102 Node *prefetch_adr;
2103 Node *prefetch;
2104 uint step_size = AllocatePrefetchStepSize;
2105 uint distance = 0;
2106
2107 for ( intx i = 0; i < lines; i++ ) {
2108 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt,
2109 _igvn.MakeConX(distance) );
2110 transform_later(prefetch_adr);
2111 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
2112 transform_later(prefetch);
2113 distance += step_size;
2114 i_o = prefetch;
2115 }
2116 pf_phi_abio->set_req( pf_path, i_o );
2117
2118 pf_region->init_req( fall_in_path, need_pf_false );
2119 pf_region->init_req( pf_path, need_pf_true );
2120
2121 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
2122 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
2123
2124 transform_later(pf_region);
2125 transform_later(pf_phi_rawmem);
2126 transform_later(pf_phi_abio);
2127
2128 needgc_false = pf_region;
2129 contended_phi_rawmem = pf_phi_rawmem;
2130 i_o = pf_phi_abio;
2131 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) {
2132 // Insert a prefetch instruction for each allocation.
2133 // This code is used to generate 1 prefetch instruction per cache line.
2134
2135 // Generate several prefetch instructions.
2136 uint step_size = AllocatePrefetchStepSize;
2137 uint distance = AllocatePrefetchDistance;
2138
2139 // Next cache address.
2140 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top,
2141 _igvn.MakeConX(step_size + distance));
2142 transform_later(cache_adr);
2143 cache_adr = new CastP2XNode(needgc_false, cache_adr);
2144 transform_later(cache_adr);
2145 // Address is aligned to execute prefetch to the beginning of cache line size
2146 // (it is important when BIS instruction is used on SPARC as prefetch).
2147 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
2148 cache_adr = new AndXNode(cache_adr, mask);
2149 transform_later(cache_adr);
2150 cache_adr = new CastX2PNode(cache_adr);
2151 transform_later(cache_adr);
2152
2153 // Prefetch
2154 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr );
2155 prefetch->set_req(0, needgc_false);
2156 transform_later(prefetch);
2157 contended_phi_rawmem = prefetch;
2158 Node *prefetch_adr;
2159 distance = step_size;
2160 for ( intx i = 1; i < lines; i++ ) {
2161 prefetch_adr = new AddPNode( cache_adr, cache_adr,
2162 _igvn.MakeConX(distance) );
2163 transform_later(prefetch_adr);
2164 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr );
2165 transform_later(prefetch);
2166 distance += step_size;
2167 contended_phi_rawmem = prefetch;
2168 }
2169 } else if( AllocatePrefetchStyle > 0 ) {
2170 // Insert a prefetch for each allocation only on the fast-path
2171 Node *prefetch_adr;
2172 Node *prefetch;
2173 // Generate several prefetch instructions.
2174 uint step_size = AllocatePrefetchStepSize;
2175 uint distance = AllocatePrefetchDistance;
2176 for ( intx i = 0; i < lines; i++ ) {
2177 prefetch_adr = new AddPNode( old_eden_top, new_eden_top,
2178 _igvn.MakeConX(distance) );
2179 transform_later(prefetch_adr);
2180 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
2181 // Do not let it float too high, since if eden_top == eden_end,
2182 // both might be null.
2183 if( i == 0 ) { // Set control for first prefetch, next follows it
2184 prefetch->init_req(0, needgc_false);
2185 }
2186 transform_later(prefetch);
2187 distance += step_size;
2188 i_o = prefetch;
2189 }
2190 }
2191 return i_o;
2192 }
2193
2194
2195 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
2196 expand_allocate_common(alloc, nullptr, nullptr,
2197 OptoRuntime::new_instance_Type(),
2198 OptoRuntime::new_instance_Java(), nullptr);
2199 }
2200
2201 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
2202 Node* length = alloc->in(AllocateNode::ALength);
2203 Node* valid_length_test = alloc->in(AllocateNode::ValidLengthTest);
2204 InitializeNode* init = alloc->initialization();
2205 Node* klass_node = alloc->in(AllocateNode::KlassNode);
2206 Node* init_value = alloc->in(AllocateNode::InitValue);
2207 const TypeAryKlassPtr* ary_klass_t = _igvn.type(klass_node)->isa_aryklassptr();
2208 assert(!ary_klass_t || !ary_klass_t->klass_is_exact() || !ary_klass_t->exact_klass()->is_obj_array_klass() ||
2209 ary_klass_t->is_vm_type(), "Must be a refined array klass");
2210 const TypeFunc* slow_call_type;
2211 address slow_call_address; // Address of slow call
2212 if (init != nullptr && init->is_complete_with_arraycopy() &&
2213 ary_klass_t && ary_klass_t->elem()->isa_klassptr() == nullptr) {
2214 // Don't zero type array during slow allocation in VM since
2215 // it will be initialized later by arraycopy in compiled code.
2216 slow_call_address = OptoRuntime::new_array_nozero_Java();
2217 slow_call_type = OptoRuntime::new_array_nozero_Type();
2218 } else {
2219 slow_call_address = OptoRuntime::new_array_Java();
2220 slow_call_type = OptoRuntime::new_array_Type();
2221
2222 if (init_value == nullptr) {
2223 init_value = _igvn.zerocon(T_OBJECT);
2224 } else if (UseCompressedOops) {
2225 init_value = transform_later(new DecodeNNode(init_value, init_value->bottom_type()->make_ptr()));
2226 }
2227 }
2228 expand_allocate_common(alloc, length, init_value,
2229 slow_call_type,
2230 slow_call_address, valid_length_test);
2231 }
2232
2233 //-------------------mark_eliminated_box----------------------------------
2234 //
2235 // During EA obj may point to several objects but after few ideal graph
2236 // transformations (CCP) it may point to only one non escaping object
2237 // (but still using phi), corresponding locks and unlocks will be marked
2238 // for elimination. Later obj could be replaced with a new node (new phi)
2239 // and which does not have escape information. And later after some graph
2240 // reshape other locks and unlocks (which were not marked for elimination
2241 // before) are connected to this new obj (phi) but they still will not be
2242 // marked for elimination since new obj has no escape information.
2243 // Mark all associated (same box and obj) lock and unlock nodes for
2244 // elimination if some of them marked already.
2245 void PhaseMacroExpand::mark_eliminated_box(Node* box, Node* obj) {
2246 BoxLockNode* oldbox = box->as_BoxLock();
2247 if (oldbox->is_eliminated()) {
2248 return; // This BoxLock node was processed already.
2249 }
2250 assert(!oldbox->is_unbalanced(), "this should not be called for unbalanced region");
2251 // New implementation (EliminateNestedLocks) has separate BoxLock
2252 // node for each locked region so mark all associated locks/unlocks as
2253 // eliminated even if different objects are referenced in one locked region
2254 // (for example, OSR compilation of nested loop inside locked scope).
2255 if (EliminateNestedLocks ||
2256 oldbox->as_BoxLock()->is_simple_lock_region(nullptr, obj, nullptr)) {
2257 // Box is used only in one lock region. Mark this box as eliminated.
2258 oldbox->set_local(); // This verifies correct state of BoxLock
2259 _igvn.hash_delete(oldbox);
2260 oldbox->set_eliminated(); // This changes box's hash value
2261 _igvn.hash_insert(oldbox);
2262
2263 for (uint i = 0; i < oldbox->outcnt(); i++) {
2264 Node* u = oldbox->raw_out(i);
2265 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
2266 AbstractLockNode* alock = u->as_AbstractLock();
2267 // Check lock's box since box could be referenced by Lock's debug info.
2268 if (alock->box_node() == oldbox) {
2269 // Mark eliminated all related locks and unlocks.
2270 #ifdef ASSERT
2271 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4");
2272 #endif
2273 alock->set_non_esc_obj();
2274 }
2275 }
2276 }
2277 return;
2278 }
2279
2280 // Create new "eliminated" BoxLock node and use it in monitor debug info
2281 // instead of oldbox for the same object.
2282 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
2283
2284 // Note: BoxLock node is marked eliminated only here and it is used
2285 // to indicate that all associated lock and unlock nodes are marked
2286 // for elimination.
2287 newbox->set_local(); // This verifies correct state of BoxLock
2288 newbox->set_eliminated();
2289 transform_later(newbox);
2290
2291 // Replace old box node with new box for all users of the same object.
2292 for (uint i = 0; i < oldbox->outcnt();) {
2293 bool next_edge = true;
2294
2295 Node* u = oldbox->raw_out(i);
2296 if (u->is_AbstractLock()) {
2297 AbstractLockNode* alock = u->as_AbstractLock();
2298 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
2299 // Replace Box and mark eliminated all related locks and unlocks.
2300 #ifdef ASSERT
2301 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5");
2302 #endif
2303 alock->set_non_esc_obj();
2304 _igvn.rehash_node_delayed(alock);
2305 alock->set_box_node(newbox);
2306 next_edge = false;
2307 }
2308 }
2309 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
2310 FastLockNode* flock = u->as_FastLock();
2311 assert(flock->box_node() == oldbox, "sanity");
2312 _igvn.rehash_node_delayed(flock);
2313 flock->set_box_node(newbox);
2314 next_edge = false;
2315 }
2316
2317 // Replace old box in monitor debug info.
2318 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
2319 SafePointNode* sfn = u->as_SafePoint();
2320 JVMState* youngest_jvms = sfn->jvms();
2321 int max_depth = youngest_jvms->depth();
2322 for (int depth = 1; depth <= max_depth; depth++) {
2323 JVMState* jvms = youngest_jvms->of_depth(depth);
2324 int num_mon = jvms->nof_monitors();
2325 // Loop over monitors
2326 for (int idx = 0; idx < num_mon; idx++) {
2327 Node* obj_node = sfn->monitor_obj(jvms, idx);
2328 Node* box_node = sfn->monitor_box(jvms, idx);
2329 if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
2330 int j = jvms->monitor_box_offset(idx);
2331 _igvn.replace_input_of(u, j, newbox);
2332 next_edge = false;
2333 }
2334 }
2335 }
2336 }
2337 if (next_edge) i++;
2338 }
2339 }
2340
2341 //-----------------------mark_eliminated_locking_nodes-----------------------
2342 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
2343 if (!alock->is_balanced()) {
2344 return; // Can't do any more elimination for this locking region
2345 }
2346 if (EliminateNestedLocks) {
2347 if (alock->is_nested()) {
2348 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
2349 return;
2350 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
2351 // Only Lock node has JVMState needed here.
2352 // Not that preceding claim is documented anywhere else.
2353 if (alock->jvms() != nullptr) {
2354 if (alock->as_Lock()->is_nested_lock_region()) {
2355 // Mark eliminated related nested locks and unlocks.
2356 Node* obj = alock->obj_node();
2357 BoxLockNode* box_node = alock->box_node()->as_BoxLock();
2358 assert(!box_node->is_eliminated(), "should not be marked yet");
2359 // Note: BoxLock node is marked eliminated only here
2360 // and it is used to indicate that all associated lock
2361 // and unlock nodes are marked for elimination.
2362 box_node->set_eliminated(); // Box's hash is always NO_HASH here
2363 for (uint i = 0; i < box_node->outcnt(); i++) {
2364 Node* u = box_node->raw_out(i);
2365 if (u->is_AbstractLock()) {
2366 alock = u->as_AbstractLock();
2367 if (alock->box_node() == box_node) {
2368 // Verify that this Box is referenced only by related locks.
2369 assert(alock->obj_node()->eqv_uncast(obj), "");
2370 // Mark all related locks and unlocks.
2371 #ifdef ASSERT
2372 alock->log_lock_optimization(C, "eliminate_lock_set_nested");
2373 #endif
2374 alock->set_nested();
2375 }
2376 }
2377 }
2378 } else {
2379 #ifdef ASSERT
2380 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region");
2381 if (C->log() != nullptr)
2382 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output
2383 #endif
2384 }
2385 }
2386 return;
2387 }
2388 // Process locks for non escaping object
2389 assert(alock->is_non_esc_obj(), "");
2390 } // EliminateNestedLocks
2391
2392 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2393 // Look for all locks of this object and mark them and
2394 // corresponding BoxLock nodes as eliminated.
2395 Node* obj = alock->obj_node();
2396 for (uint j = 0; j < obj->outcnt(); j++) {
2397 Node* o = obj->raw_out(j);
2398 if (o->is_AbstractLock() &&
2399 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2400 alock = o->as_AbstractLock();
2401 Node* box = alock->box_node();
2402 // Replace old box node with new eliminated box for all users
2403 // of the same object and mark related locks as eliminated.
2404 mark_eliminated_box(box, obj);
2405 }
2406 }
2407 }
2408 }
2409
2410 // we have determined that this lock/unlock can be eliminated, we simply
2411 // eliminate the node without expanding it.
2412 //
2413 // Note: The membar's associated with the lock/unlock are currently not
2414 // eliminated. This should be investigated as a future enhancement.
2415 //
2416 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2417
2418 if (!alock->is_eliminated()) {
2419 return false;
2420 }
2421 #ifdef ASSERT
2422 if (!alock->is_coarsened()) {
2423 // Check that new "eliminated" BoxLock node is created.
2424 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2425 assert(oldbox->is_eliminated(), "should be done already");
2426 }
2427 #endif
2428
2429 alock->log_lock_optimization(C, "eliminate_lock");
2430
2431 #ifndef PRODUCT
2432 if (PrintEliminateLocks) {
2433 tty->print_cr("++++ Eliminated: %d %s '%s'", alock->_idx, (alock->is_Lock() ? "Lock" : "Unlock"), alock->kind_as_string());
2434 }
2435 #endif
2436
2437 Node* mem = alock->in(TypeFunc::Memory);
2438 Node* ctrl = alock->in(TypeFunc::Control);
2439 guarantee(ctrl != nullptr, "missing control projection, cannot replace_node() with null");
2440
2441 _callprojs = alock->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2442 // There are 2 projections from the lock. The lock node will
2443 // be deleted when its last use is subsumed below.
2444 assert(alock->outcnt() == 2 &&
2445 _callprojs->fallthrough_proj != nullptr &&
2446 _callprojs->fallthrough_memproj != nullptr,
2447 "Unexpected projections from Lock/Unlock");
2448
2449 Node* fallthroughproj = _callprojs->fallthrough_proj;
2450 Node* memproj_fallthrough = _callprojs->fallthrough_memproj;
2451
2452 // The memory projection from a lock/unlock is RawMem
2453 // The input to a Lock is merged memory, so extract its RawMem input
2454 // (unless the MergeMem has been optimized away.)
2455 if (alock->is_Lock()) {
2456 // Search for MemBarAcquireLock node and delete it also.
2457 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2458 assert(membar != nullptr && membar->Opcode() == Op_MemBarAcquireLock, "");
2459 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2460 Node* memproj = membar->proj_out(TypeFunc::Memory);
2461 _igvn.replace_node(ctrlproj, fallthroughproj);
2462 _igvn.replace_node(memproj, memproj_fallthrough);
2463
2464 // Delete FastLock node also if this Lock node is unique user
2465 // (a loop peeling may clone a Lock node).
2466 Node* flock = alock->as_Lock()->fastlock_node();
2467 if (flock->outcnt() == 1) {
2468 assert(flock->unique_out() == alock, "sanity");
2469 _igvn.replace_node(flock, top());
2470 }
2471 }
2472
2473 // Search for MemBarReleaseLock node and delete it also.
2474 if (alock->is_Unlock() && ctrl->is_Proj() && ctrl->in(0)->is_MemBar()) {
2475 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2476 assert(membar->Opcode() == Op_MemBarReleaseLock &&
2477 mem->is_Proj() && membar == mem->in(0), "");
2478 _igvn.replace_node(fallthroughproj, ctrl);
2479 _igvn.replace_node(memproj_fallthrough, mem);
2480 fallthroughproj = ctrl;
2481 memproj_fallthrough = mem;
2482 ctrl = membar->in(TypeFunc::Control);
2483 mem = membar->in(TypeFunc::Memory);
2484 }
2485
2486 _igvn.replace_node(fallthroughproj, ctrl);
2487 _igvn.replace_node(memproj_fallthrough, mem);
2488 return true;
2489 }
2490
2491
2492 //------------------------------expand_lock_node----------------------
2493 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2494
2495 Node* ctrl = lock->in(TypeFunc::Control);
2496 Node* mem = lock->in(TypeFunc::Memory);
2497 Node* obj = lock->obj_node();
2498 Node* box = lock->box_node();
2499 Node* flock = lock->fastlock_node();
2500
2501 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2502
2503 // Make the merge point
2504 Node *region;
2505 Node *mem_phi;
2506 Node *slow_path;
2507
2508 region = new RegionNode(3);
2509 // create a Phi for the memory state
2510 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2511
2512 // Optimize test; set region slot 2
2513 slow_path = opt_bits_test(ctrl, region, 2, flock);
2514 mem_phi->init_req(2, mem);
2515
2516 // Make slow path call
2517 CallNode* call = make_slow_call(lock, OptoRuntime::complete_monitor_enter_Type(),
2518 OptoRuntime::complete_monitor_locking_Java(), nullptr, slow_path,
2519 obj, box, nullptr);
2520
2521 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2522
2523 // Slow path can only throw asynchronous exceptions, which are always
2524 // de-opted. So the compiler thinks the slow-call can never throw an
2525 // exception. If it DOES throw an exception we would need the debug
2526 // info removed first (since if it throws there is no monitor).
2527 assert(_callprojs->fallthrough_ioproj == nullptr && _callprojs->catchall_ioproj == nullptr &&
2528 _callprojs->catchall_memproj == nullptr && _callprojs->catchall_catchproj == nullptr, "Unexpected projection from Lock");
2529
2530 // Capture slow path
2531 // disconnect fall-through projection from call and create a new one
2532 // hook up users of fall-through projection to region
2533 Node *slow_ctrl = _callprojs->fallthrough_proj->clone();
2534 transform_later(slow_ctrl);
2535 _igvn.hash_delete(_callprojs->fallthrough_proj);
2536 _callprojs->fallthrough_proj->disconnect_inputs(C);
2537 region->init_req(1, slow_ctrl);
2538 // region inputs are now complete
2539 transform_later(region);
2540 _igvn.replace_node(_callprojs->fallthrough_proj, region);
2541
2542 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2543
2544 mem_phi->init_req(1, memproj);
2545
2546 transform_later(mem_phi);
2547
2548 _igvn.replace_node(_callprojs->fallthrough_memproj, mem_phi);
2549 }
2550
2551 //------------------------------expand_unlock_node----------------------
2552 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2553
2554 Node* ctrl = unlock->in(TypeFunc::Control);
2555 Node* mem = unlock->in(TypeFunc::Memory);
2556 Node* obj = unlock->obj_node();
2557 Node* box = unlock->box_node();
2558
2559 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2560
2561 // No need for a null check on unlock
2562
2563 // Make the merge point
2564 Node *region;
2565 Node *mem_phi;
2566
2567 region = new RegionNode(3);
2568 // create a Phi for the memory state
2569 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2570
2571 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2572 funlock = transform_later( funlock )->as_FastUnlock();
2573 // Optimize test; set region slot 2
2574 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock);
2575 Node *thread = transform_later(new ThreadLocalNode());
2576
2577 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2578 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2579 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2580
2581 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2582 assert(_callprojs->fallthrough_ioproj == nullptr && _callprojs->catchall_ioproj == nullptr &&
2583 _callprojs->catchall_memproj == nullptr && _callprojs->catchall_catchproj == nullptr, "Unexpected projection from Lock");
2584
2585 // No exceptions for unlocking
2586 // Capture slow path
2587 // disconnect fall-through projection from call and create a new one
2588 // hook up users of fall-through projection to region
2589 Node *slow_ctrl = _callprojs->fallthrough_proj->clone();
2590 transform_later(slow_ctrl);
2591 _igvn.hash_delete(_callprojs->fallthrough_proj);
2592 _callprojs->fallthrough_proj->disconnect_inputs(C);
2593 region->init_req(1, slow_ctrl);
2594 // region inputs are now complete
2595 transform_later(region);
2596 _igvn.replace_node(_callprojs->fallthrough_proj, region);
2597
2598 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2599 mem_phi->init_req(1, memproj );
2600 mem_phi->init_req(2, mem);
2601 transform_later(mem_phi);
2602
2603 _igvn.replace_node(_callprojs->fallthrough_memproj, mem_phi);
2604 }
2605
2606 // An inline type might be returned from the call but we don't know its
2607 // type. Either we get a buffered inline type (and nothing needs to be done)
2608 // or one of the values being returned is the klass of the inline type
2609 // and we need to allocate an inline type instance of that type and
2610 // initialize it with other values being returned. In that case, we
2611 // first try a fast path allocation and initialize the value with the
2612 // inline klass's pack handler or we fall back to a runtime call.
2613 void PhaseMacroExpand::expand_mh_intrinsic_return(CallStaticJavaNode* call) {
2614 assert(call->method()->is_method_handle_intrinsic(), "must be a method handle intrinsic call");
2615 Node* ret = call->proj_out_or_null(TypeFunc::Parms);
2616 if (ret == nullptr) {
2617 return;
2618 }
2619 const TypeFunc* tf = call->_tf;
2620 const TypeTuple* domain = OptoRuntime::store_inline_type_fields_Type()->domain_cc();
2621 const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain);
2622 call->_tf = new_tf;
2623 // Make sure the change of type is applied before projections are processed by igvn
2624 _igvn.set_type(call, call->Value(&_igvn));
2625 _igvn.set_type(ret, ret->Value(&_igvn));
2626
2627 // Before any new projection is added:
2628 CallProjections* projs = call->extract_projections(true, true);
2629
2630 // Create temporary hook nodes that will be replaced below.
2631 // Add an input to prevent hook nodes from being dead.
2632 Node* ctl = new Node(call);
2633 Node* mem = new Node(ctl);
2634 Node* io = new Node(ctl);
2635 Node* ex_ctl = new Node(ctl);
2636 Node* ex_mem = new Node(ctl);
2637 Node* ex_io = new Node(ctl);
2638 Node* res = new Node(ctl);
2639
2640 // Allocate a new buffered inline type only if a new one is not returned
2641 Node* cast = transform_later(new CastP2XNode(ctl, res));
2642 Node* mask = MakeConX(0x1);
2643 Node* masked = transform_later(new AndXNode(cast, mask));
2644 Node* cmp = transform_later(new CmpXNode(masked, mask));
2645 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
2646 IfNode* allocation_iff = new IfNode(ctl, bol, PROB_MAX, COUNT_UNKNOWN);
2647 transform_later(allocation_iff);
2648 Node* allocation_ctl = transform_later(new IfTrueNode(allocation_iff));
2649 Node* no_allocation_ctl = transform_later(new IfFalseNode(allocation_iff));
2650 Node* no_allocation_res = transform_later(new CheckCastPPNode(no_allocation_ctl, res, TypeInstPtr::BOTTOM));
2651
2652 // Try to allocate a new buffered inline instance either from TLAB or eden space
2653 Node* needgc_ctrl = nullptr; // needgc means slowcase, i.e. allocation failed
2654 CallLeafNoFPNode* handler_call;
2655 const bool alloc_in_place = UseTLAB;
2656 if (alloc_in_place) {
2657 Node* fast_oop_ctrl = nullptr;
2658 Node* fast_oop_rawmem = nullptr;
2659 Node* mask2 = MakeConX(-2);
2660 Node* masked2 = transform_later(new AndXNode(cast, mask2));
2661 Node* rawklassptr = transform_later(new CastX2PNode(masked2));
2662 Node* klass_node = transform_later(new CheckCastPPNode(allocation_ctl, rawklassptr, TypeInstKlassPtr::OBJECT_OR_NULL));
2663 Node* layout_val = make_load(nullptr, mem, klass_node, in_bytes(Klass::layout_helper_offset()), TypeInt::INT, T_INT);
2664 Node* size_in_bytes = ConvI2X(layout_val);
2665 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2666 Node* fast_oop = bs->obj_allocate(this, mem, allocation_ctl, size_in_bytes, io, needgc_ctrl,
2667 fast_oop_ctrl, fast_oop_rawmem,
2668 AllocateInstancePrefetchLines);
2669 // Allocation succeed, initialize buffered inline instance header firstly,
2670 // and then initialize its fields with an inline class specific handler
2671 Node* mark_word_node;
2672 if (UseCompactObjectHeaders) {
2673 // COH: We need to load the prototype from the klass at runtime since it encodes the klass pointer already.
2674 mark_word_node = make_load(fast_oop_ctrl, fast_oop_rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2675 } else {
2676 // Otherwise, use the static prototype.
2677 mark_word_node = makecon(TypeRawPtr::make((address)markWord::inline_type_prototype().value()));
2678 }
2679
2680 fast_oop_rawmem = make_store(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::mark_offset_in_bytes(), mark_word_node, T_ADDRESS);
2681 if (!UseCompactObjectHeaders) {
2682 // COH: Everything is encoded in the mark word, so nothing left to do.
2683 fast_oop_rawmem = make_store(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
2684 if (UseCompressedClassPointers) {
2685 fast_oop_rawmem = make_store(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::klass_gap_offset_in_bytes(), intcon(0), T_INT);
2686 }
2687 }
2688 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);
2689 Node* pack_handler = make_load(fast_oop_ctrl, fast_oop_rawmem, fixed_block, in_bytes(InlineKlass::pack_handler_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2690 handler_call = new CallLeafNoFPNode(OptoRuntime::pack_inline_type_Type(),
2691 nullptr,
2692 "pack handler",
2693 TypeRawPtr::BOTTOM);
2694 handler_call->init_req(TypeFunc::Control, fast_oop_ctrl);
2695 handler_call->init_req(TypeFunc::Memory, fast_oop_rawmem);
2696 handler_call->init_req(TypeFunc::I_O, top());
2697 handler_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2698 handler_call->init_req(TypeFunc::ReturnAdr, top());
2699 handler_call->init_req(TypeFunc::Parms, pack_handler);
2700 handler_call->init_req(TypeFunc::Parms+1, fast_oop);
2701 } else {
2702 needgc_ctrl = allocation_ctl;
2703 }
2704
2705 // Allocation failed, fall back to a runtime call
2706 CallStaticJavaNode* slow_call = new CallStaticJavaNode(OptoRuntime::store_inline_type_fields_Type(),
2707 StubRoutines::store_inline_type_fields_to_buf(),
2708 "store_inline_type_fields",
2709 TypePtr::BOTTOM);
2710 slow_call->init_req(TypeFunc::Control, needgc_ctrl);
2711 slow_call->init_req(TypeFunc::Memory, mem);
2712 slow_call->init_req(TypeFunc::I_O, io);
2713 slow_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2714 slow_call->init_req(TypeFunc::ReturnAdr, call->in(TypeFunc::ReturnAdr));
2715 slow_call->init_req(TypeFunc::Parms, res);
2716
2717 Node* slow_ctl = transform_later(new ProjNode(slow_call, TypeFunc::Control));
2718 Node* slow_mem = transform_later(new ProjNode(slow_call, TypeFunc::Memory));
2719 Node* slow_io = transform_later(new ProjNode(slow_call, TypeFunc::I_O));
2720 Node* slow_res = transform_later(new ProjNode(slow_call, TypeFunc::Parms));
2721 Node* slow_catc = transform_later(new CatchNode(slow_ctl, slow_io, 2));
2722 Node* slow_norm = transform_later(new CatchProjNode(slow_catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci));
2723 Node* slow_excp = transform_later(new CatchProjNode(slow_catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci));
2724
2725 Node* ex_r = new RegionNode(3);
2726 Node* ex_mem_phi = new PhiNode(ex_r, Type::MEMORY, TypePtr::BOTTOM);
2727 Node* ex_io_phi = new PhiNode(ex_r, Type::ABIO);
2728 ex_r->init_req(1, slow_excp);
2729 ex_mem_phi->init_req(1, slow_mem);
2730 ex_io_phi->init_req(1, slow_io);
2731 ex_r->init_req(2, ex_ctl);
2732 ex_mem_phi->init_req(2, ex_mem);
2733 ex_io_phi->init_req(2, ex_io);
2734 transform_later(ex_r);
2735 transform_later(ex_mem_phi);
2736 transform_later(ex_io_phi);
2737
2738 // We don't know how many values are returned. This assumes the
2739 // worst case, that all available registers are used.
2740 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2741 if (domain->field_at(i) == Type::HALF) {
2742 slow_call->init_req(i, top());
2743 if (alloc_in_place) {
2744 handler_call->init_req(i+1, top());
2745 }
2746 continue;
2747 }
2748 Node* proj = transform_later(new ProjNode(call, i));
2749 slow_call->init_req(i, proj);
2750 if (alloc_in_place) {
2751 handler_call->init_req(i+1, proj);
2752 }
2753 }
2754 // We can safepoint at that new call
2755 slow_call->copy_call_debug_info(&_igvn, call);
2756 transform_later(slow_call);
2757 if (alloc_in_place) {
2758 transform_later(handler_call);
2759 }
2760
2761 Node* fast_ctl = nullptr;
2762 Node* fast_res = nullptr;
2763 MergeMemNode* fast_mem = nullptr;
2764 if (alloc_in_place) {
2765 fast_ctl = transform_later(new ProjNode(handler_call, TypeFunc::Control));
2766 Node* rawmem = transform_later(new ProjNode(handler_call, TypeFunc::Memory));
2767 fast_res = transform_later(new ProjNode(handler_call, TypeFunc::Parms));
2768 fast_mem = MergeMemNode::make(mem);
2769 fast_mem->set_memory_at(Compile::AliasIdxRaw, rawmem);
2770 transform_later(fast_mem);
2771 }
2772
2773 Node* r = new RegionNode(alloc_in_place ? 4 : 3);
2774 Node* mem_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2775 Node* io_phi = new PhiNode(r, Type::ABIO);
2776 Node* res_phi = new PhiNode(r, TypeInstPtr::BOTTOM);
2777 r->init_req(1, no_allocation_ctl);
2778 mem_phi->init_req(1, mem);
2779 io_phi->init_req(1, io);
2780 res_phi->init_req(1, no_allocation_res);
2781 r->init_req(2, slow_norm);
2782 mem_phi->init_req(2, slow_mem);
2783 io_phi->init_req(2, slow_io);
2784 res_phi->init_req(2, slow_res);
2785 if (alloc_in_place) {
2786 r->init_req(3, fast_ctl);
2787 mem_phi->init_req(3, fast_mem);
2788 io_phi->init_req(3, io);
2789 res_phi->init_req(3, fast_res);
2790 }
2791 transform_later(r);
2792 transform_later(mem_phi);
2793 transform_later(io_phi);
2794 transform_later(res_phi);
2795
2796 // Do not let stores that initialize this buffer be reordered with a subsequent
2797 // store that would make this buffer accessible by other threads.
2798 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
2799 transform_later(mb);
2800 mb->init_req(TypeFunc::Memory, mem_phi);
2801 mb->init_req(TypeFunc::Control, r);
2802 r = new ProjNode(mb, TypeFunc::Control);
2803 transform_later(r);
2804 mem_phi = new ProjNode(mb, TypeFunc::Memory);
2805 transform_later(mem_phi);
2806
2807 assert(projs->nb_resproj == 1, "unexpected number of results");
2808 _igvn.replace_in_uses(projs->fallthrough_catchproj, r);
2809 _igvn.replace_in_uses(projs->fallthrough_memproj, mem_phi);
2810 _igvn.replace_in_uses(projs->fallthrough_ioproj, io_phi);
2811 _igvn.replace_in_uses(projs->resproj[0], res_phi);
2812 _igvn.replace_in_uses(projs->catchall_catchproj, ex_r);
2813 _igvn.replace_in_uses(projs->catchall_memproj, ex_mem_phi);
2814 _igvn.replace_in_uses(projs->catchall_ioproj, ex_io_phi);
2815 // The CatchNode should not use the ex_io_phi. Re-connect it to the catchall_ioproj.
2816 Node* cn = projs->fallthrough_catchproj->in(0);
2817 _igvn.replace_input_of(cn, 1, projs->catchall_ioproj);
2818
2819 _igvn.replace_node(ctl, projs->fallthrough_catchproj);
2820 _igvn.replace_node(mem, projs->fallthrough_memproj);
2821 _igvn.replace_node(io, projs->fallthrough_ioproj);
2822 _igvn.replace_node(res, projs->resproj[0]);
2823 _igvn.replace_node(ex_ctl, projs->catchall_catchproj);
2824 _igvn.replace_node(ex_mem, projs->catchall_memproj);
2825 _igvn.replace_node(ex_io, projs->catchall_ioproj);
2826 }
2827
2828 void PhaseMacroExpand::expand_subtypecheck_node(SubTypeCheckNode *check) {
2829 assert(check->in(SubTypeCheckNode::Control) == nullptr, "should be pinned");
2830 Node* bol = check->unique_out();
2831 Node* obj_or_subklass = check->in(SubTypeCheckNode::ObjOrSubKlass);
2832 Node* superklass = check->in(SubTypeCheckNode::SuperKlass);
2833 assert(bol->is_Bool() && bol->as_Bool()->_test._test == BoolTest::ne, "unexpected bool node");
2834
2835 for (DUIterator_Last imin, i = bol->last_outs(imin); i >= imin; --i) {
2836 Node* iff = bol->last_out(i);
2837 assert(iff->is_If(), "where's the if?");
2838
2839 if (iff->in(0)->is_top()) {
2840 _igvn.replace_input_of(iff, 1, C->top());
2841 continue;
2842 }
2843
2844 Node* iftrue = iff->as_If()->proj_out(1);
2845 Node* iffalse = iff->as_If()->proj_out(0);
2846 Node* ctrl = iff->in(0);
2847
2848 Node* subklass = nullptr;
2849 if (_igvn.type(obj_or_subklass)->isa_klassptr()) {
2850 subklass = obj_or_subklass;
2851 } else {
2852 Node* k_adr = basic_plus_adr(obj_or_subklass, oopDesc::klass_offset_in_bytes());
2853 subklass = _igvn.transform(LoadKlassNode::make(_igvn, C->immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
2854 }
2855
2856 Node* not_subtype_ctrl = Phase::gen_subtype_check(subklass, superklass, &ctrl, nullptr, _igvn, check->method(), check->bci());
2857
2858 _igvn.replace_input_of(iff, 0, C->top());
2859 _igvn.replace_node(iftrue, not_subtype_ctrl);
2860 _igvn.replace_node(iffalse, ctrl);
2861 }
2862 _igvn.replace_node(check, C->top());
2863 }
2864
2865 // FlatArrayCheckNode (array1 array2 ...) is expanded into:
2866 //
2867 // long mark = array1.mark | array2.mark | ...;
2868 // long locked_bit = markWord::unlocked_value & array1.mark & array2.mark & ...;
2869 // if (locked_bit == 0) {
2870 // // One array is locked, load prototype header from the klass
2871 // mark = array1.klass.proto | array2.klass.proto | ...
2872 // }
2873 // if ((mark & markWord::flat_array_bit_in_place) == 0) {
2874 // ...
2875 // }
2876 void PhaseMacroExpand::expand_flatarraycheck_node(FlatArrayCheckNode* check) {
2877 bool array_inputs = _igvn.type(check->in(FlatArrayCheckNode::ArrayOrKlass))->isa_oopptr() != nullptr;
2878 if (array_inputs) {
2879 Node* mark = MakeConX(0);
2880 Node* locked_bit = MakeConX(markWord::unlocked_value);
2881 Node* mem = check->in(FlatArrayCheckNode::Memory);
2882 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
2883 Node* ary = check->in(i);
2884 const TypeOopPtr* t = _igvn.type(ary)->isa_oopptr();
2885 assert(t != nullptr, "Mixing array and klass inputs");
2886 assert(!t->is_flat() && !t->is_not_flat(), "Should have been optimized out");
2887 Node* mark_adr = basic_plus_adr(ary, oopDesc::mark_offset_in_bytes());
2888 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));
2889 mark = _igvn.transform(new OrXNode(mark, mark_load));
2890 locked_bit = _igvn.transform(new AndXNode(locked_bit, mark_load));
2891 }
2892 assert(!mark->is_Con(), "Should have been optimized out");
2893 Node* cmp = _igvn.transform(new CmpXNode(locked_bit, MakeConX(0)));
2894 Node* is_unlocked = _igvn.transform(new BoolNode(cmp, BoolTest::ne));
2895
2896 // BoolNode might be shared, replace each if user
2897 Node* old_bol = check->unique_out();
2898 assert(old_bol->is_Bool() && old_bol->as_Bool()->_test._test == BoolTest::ne, "unexpected condition");
2899 for (DUIterator_Last imin, i = old_bol->last_outs(imin); i >= imin; --i) {
2900 IfNode* old_iff = old_bol->last_out(i)->as_If();
2901 Node* ctrl = old_iff->in(0);
2902 RegionNode* region = new RegionNode(3);
2903 Node* mark_phi = new PhiNode(region, TypeX_X);
2904
2905 // Check if array is unlocked
2906 IfNode* iff = _igvn.transform(new IfNode(ctrl, is_unlocked, PROB_MAX, COUNT_UNKNOWN))->as_If();
2907
2908 // Unlocked: Use bits from mark word
2909 region->init_req(1, _igvn.transform(new IfTrueNode(iff)));
2910 mark_phi->init_req(1, mark);
2911
2912 // Locked: Load prototype header from klass
2913 ctrl = _igvn.transform(new IfFalseNode(iff));
2914 Node* proto = MakeConX(0);
2915 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
2916 Node* ary = check->in(i);
2917 // Make loads control dependent to make sure they are only executed if array is locked
2918 Node* klass_adr = basic_plus_adr(ary, oopDesc::klass_offset_in_bytes());
2919 Node* klass = _igvn.transform(LoadKlassNode::make(_igvn, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
2920 Node* proto_adr = basic_plus_adr(klass, in_bytes(Klass::prototype_header_offset()));
2921 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));
2922 proto = _igvn.transform(new OrXNode(proto, proto_load));
2923 }
2924 region->init_req(2, ctrl);
2925 mark_phi->init_req(2, proto);
2926
2927 // Check if flat array bits are set
2928 Node* mask = MakeConX(markWord::flat_array_bit_in_place);
2929 Node* masked = _igvn.transform(new AndXNode(_igvn.transform(mark_phi), mask));
2930 cmp = _igvn.transform(new CmpXNode(masked, MakeConX(0)));
2931 Node* is_not_flat = _igvn.transform(new BoolNode(cmp, BoolTest::eq));
2932
2933 ctrl = _igvn.transform(region);
2934 iff = _igvn.transform(new IfNode(ctrl, is_not_flat, PROB_MAX, COUNT_UNKNOWN))->as_If();
2935 _igvn.replace_node(old_iff, iff);
2936 }
2937 _igvn.replace_node(check, C->top());
2938 } else {
2939 // Fall back to layout helper check
2940 Node* lhs = intcon(0);
2941 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
2942 Node* array_or_klass = check->in(i);
2943 Node* klass = nullptr;
2944 const TypePtr* t = _igvn.type(array_or_klass)->is_ptr();
2945 assert(!t->is_flat() && !t->is_not_flat(), "Should have been optimized out");
2946 if (t->isa_oopptr() != nullptr) {
2947 Node* klass_adr = basic_plus_adr(array_or_klass, oopDesc::klass_offset_in_bytes());
2948 klass = transform_later(LoadKlassNode::make(_igvn, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
2949 } else {
2950 assert(t->isa_klassptr(), "Unexpected input type");
2951 klass = array_or_klass;
2952 }
2953 Node* lh_addr = basic_plus_adr(klass, in_bytes(Klass::layout_helper_offset()));
2954 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));
2955 lhs = _igvn.transform(new OrINode(lhs, lh_val));
2956 }
2957 Node* masked = transform_later(new AndINode(lhs, intcon(Klass::_lh_array_tag_flat_value_bit_inplace)));
2958 Node* cmp = transform_later(new CmpINode(masked, intcon(0)));
2959 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
2960 Node* m2b = transform_later(new Conv2BNode(masked));
2961 // The matcher expects the input to If/CMove nodes to be produced by a Bool(CmpI..)
2962 // pattern, but the input to other potential users (e.g. Phi) to be some
2963 // other pattern (e.g. a Conv2B node, possibly idealized as a CMoveI).
2964 Node* old_bol = check->unique_out();
2965 for (DUIterator_Last imin, i = old_bol->last_outs(imin); i >= imin; --i) {
2966 Node* user = old_bol->last_out(i);
2967 for (uint j = 0; j < user->req(); j++) {
2968 Node* n = user->in(j);
2969 if (n == old_bol) {
2970 _igvn.replace_input_of(user, j, (user->is_If() || user->is_CMove()) ? bol : m2b);
2971 }
2972 }
2973 }
2974 _igvn.replace_node(check, C->top());
2975 }
2976 }
2977
2978 // Perform refining of strip mined loop nodes in the macro nodes list.
2979 void PhaseMacroExpand::refine_strip_mined_loop_macro_nodes() {
2980 for (int i = C->macro_count(); i > 0; i--) {
2981 Node* n = C->macro_node(i - 1);
2982 if (n->is_OuterStripMinedLoop()) {
2983 n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn);
2984 }
2985 }
2986 }
2987
2988 //---------------------------eliminate_macro_nodes----------------------
2989 // Eliminate scalar replaced allocations and associated locks.
2990 void PhaseMacroExpand::eliminate_macro_nodes(bool eliminate_locks) {
2991 if (C->macro_count() == 0) {
2992 return;
2993 }
2994
2995 if (StressMacroElimination) {
2996 C->shuffle_macro_nodes();
2997 }
2998 NOT_PRODUCT(int membar_before = count_MemBar(C);)
2999
3000 int iteration = 0;
3001 while (C->macro_count() > 0) {
3002 if (iteration++ > 100) {
3003 assert(false, "Too slow convergence of macro elimination");
3004 break;
3005 }
3006
3007 // Postpone lock elimination to after EA when most allocations are eliminated
3008 // because they might block lock elimination if their escape state isn't
3009 // determined yet and we only got one chance at eliminating the lock.
3010 if (eliminate_locks) {
3011 // Before elimination may re-mark (change to Nested or NonEscObj)
3012 // all associated (same box and obj) lock and unlock nodes.
3013 int cnt = C->macro_count();
3014 for (int i=0; i < cnt; i++) {
3015 Node *n = C->macro_node(i);
3016 if (n->is_AbstractLock()) { // Lock and Unlock nodes
3017 mark_eliminated_locking_nodes(n->as_AbstractLock());
3018 }
3019 }
3020 // Re-marking may break consistency of Coarsened locks.
3021 if (!C->coarsened_locks_consistent()) {
3022 return; // recompile without Coarsened locks if broken
3023 } else {
3024 // After coarsened locks are eliminated locking regions
3025 // become unbalanced. We should not execute any more
3026 // locks elimination optimizations on them.
3027 C->mark_unbalanced_boxes();
3028 }
3029 }
3030
3031 bool progress = false;
3032 for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once
3033 Node* n = C->macro_node(i - 1);
3034 bool success = false;
3035 DEBUG_ONLY(int old_macro_count = C->macro_count();)
3036 switch (n->class_id()) {
3037 case Node::Class_Allocate:
3038 case Node::Class_AllocateArray:
3039 success = eliminate_allocate_node(n->as_Allocate());
3040 #ifndef PRODUCT
3041 if (success && PrintOptoStatistics) {
3042 AtomicAccess::inc(&PhaseMacroExpand::_objs_scalar_replaced_counter);
3043 }
3044 #endif
3045 break;
3046 case Node::Class_CallStaticJava: {
3047 CallStaticJavaNode* call = n->as_CallStaticJava();
3048 if (!call->method()->is_method_handle_intrinsic()) {
3049 success = eliminate_boxing_node(n->as_CallStaticJava());
3050 }
3051 break;
3052 }
3053 case Node::Class_Lock:
3054 case Node::Class_Unlock:
3055 if (eliminate_locks) {
3056 success = eliminate_locking_node(n->as_AbstractLock());
3057 #ifndef PRODUCT
3058 if (success && PrintOptoStatistics) {
3059 AtomicAccess::inc(&PhaseMacroExpand::_monitor_objects_removed_counter);
3060 }
3061 #endif
3062 }
3063 break;
3064 case Node::Class_ArrayCopy:
3065 break;
3066 case Node::Class_OuterStripMinedLoop:
3067 break;
3068 case Node::Class_SubTypeCheck:
3069 break;
3070 case Node::Class_Opaque1:
3071 break;
3072 case Node::Class_FlatArrayCheck:
3073 break;
3074 default:
3075 assert(n->Opcode() == Op_LoopLimit ||
3076 n->Opcode() == Op_ModD ||
3077 n->Opcode() == Op_ModF ||
3078 n->is_OpaqueNotNull() ||
3079 n->is_OpaqueInitializedAssertionPredicate() ||
3080 n->Opcode() == Op_MaxL ||
3081 n->Opcode() == Op_MinL ||
3082 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n),
3083 "unknown node type in macro list");
3084 }
3085 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
3086 progress = progress || success;
3087 if (success) {
3088 C->print_method(PHASE_AFTER_MACRO_ELIMINATION_STEP, 5, n);
3089 }
3090 }
3091
3092 // Ensure the graph after PhaseMacroExpand::eliminate_macro_nodes is canonical (no igvn
3093 // transformation is pending). If an allocation is used only in safepoints, elimination of
3094 // other macro nodes can remove all these safepoints, allowing the allocation to be removed.
3095 // Hence after igvn we retry removing macro nodes if some progress that has been made in this
3096 // iteration.
3097 _igvn.set_delay_transform(false);
3098 _igvn.optimize();
3099 if (C->failing()) {
3100 return;
3101 }
3102 _igvn.set_delay_transform(true);
3103
3104 if (!progress) {
3105 break;
3106 }
3107 }
3108 #ifndef PRODUCT
3109 if (PrintOptoStatistics) {
3110 int membar_after = count_MemBar(C);
3111 AtomicAccess::add(&PhaseMacroExpand::_memory_barriers_removed_counter, membar_before - membar_after);
3112 }
3113 #endif
3114 }
3115
3116 void PhaseMacroExpand::eliminate_opaque_looplimit_macro_nodes() {
3117 if (C->macro_count() == 0) {
3118 return;
3119 }
3120 refine_strip_mined_loop_macro_nodes();
3121 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
3122 bool progress = true;
3123 while (progress) {
3124 progress = false;
3125 for (int i = C->macro_count(); i > 0; i--) {
3126 Node* n = C->macro_node(i-1);
3127 bool success = false;
3128 DEBUG_ONLY(int old_macro_count = C->macro_count();)
3129 if (n->Opcode() == Op_LoopLimit) {
3130 // Remove it from macro list and put on IGVN worklist to optimize.
3131 C->remove_macro_node(n);
3132 _igvn._worklist.push(n);
3133 success = true;
3134 } else if (n->Opcode() == Op_CallStaticJava) {
3135 CallStaticJavaNode* call = n->as_CallStaticJava();
3136 if (!call->method()->is_method_handle_intrinsic()) {
3137 // Remove it from macro list and put on IGVN worklist to optimize.
3138 C->remove_macro_node(n);
3139 _igvn._worklist.push(n);
3140 success = true;
3141 }
3142 } else if (n->is_Opaque1()) {
3143 _igvn.replace_node(n, n->in(1));
3144 success = true;
3145 } else if (n->is_OpaqueNotNull()) {
3146 // Tests with OpaqueNotNull nodes are implicitly known to be true. Replace the node with true. In debug builds,
3147 // we leave the test in the graph to have an additional sanity check at runtime. If the test fails (i.e. a bug),
3148 // we will execute a Halt node.
3149 #ifdef ASSERT
3150 _igvn.replace_node(n, n->in(1));
3151 #else
3152 _igvn.replace_node(n, _igvn.intcon(1));
3153 #endif
3154 success = true;
3155 } else if (n->is_OpaqueInitializedAssertionPredicate()) {
3156 // Initialized Assertion Predicates must always evaluate to true. Therefore, we get rid of them in product
3157 // builds as they are useless. In debug builds we keep them as additional verification code. Even though
3158 // loop opts are already over, we want to keep Initialized Assertion Predicates alive as long as possible to
3159 // enable folding of dead control paths within which cast nodes become top after due to impossible types -
3160 // even after loop opts are over. Therefore, we delay the removal of these opaque nodes until now.
3161 #ifdef ASSERT
3162 _igvn.replace_node(n, n->in(1));
3163 #else
3164 _igvn.replace_node(n, _igvn.intcon(1));
3165 #endif // ASSERT
3166 } else if (n->Opcode() == Op_OuterStripMinedLoop) {
3167 C->remove_macro_node(n);
3168 success = true;
3169 } else if (n->Opcode() == Op_MaxL) {
3170 // Since MaxL and MinL are not implemented in the backend, we expand them to
3171 // a CMoveL construct now. At least until here, the type could be computed
3172 // precisely. CMoveL is not so smart, but we can give it at least the best
3173 // type we know abouot n now.
3174 Node* repl = MaxNode::signed_max(n->in(1), n->in(2), _igvn.type(n), _igvn);
3175 _igvn.replace_node(n, repl);
3176 success = true;
3177 } else if (n->Opcode() == Op_MinL) {
3178 Node* repl = MaxNode::signed_min(n->in(1), n->in(2), _igvn.type(n), _igvn);
3179 _igvn.replace_node(n, repl);
3180 success = true;
3181 }
3182 assert(!success || (C->macro_count() == (old_macro_count - 1)), "elimination must have deleted one node from macro list");
3183 progress = progress || success;
3184 if (success) {
3185 C->print_method(PHASE_AFTER_MACRO_ELIMINATION_STEP, 5, n);
3186 }
3187 }
3188 }
3189 }
3190
3191 //------------------------------expand_macro_nodes----------------------
3192 // Returns true if a failure occurred.
3193 bool PhaseMacroExpand::expand_macro_nodes() {
3194 if (StressMacroExpansion) {
3195 C->shuffle_macro_nodes();
3196 }
3197
3198 // Clean up the graph so we're less likely to hit the maximum node
3199 // limit
3200 _igvn.set_delay_transform(false);
3201 _igvn.optimize();
3202 if (C->failing()) return true;
3203 _igvn.set_delay_transform(true);
3204
3205
3206 // Because we run IGVN after each expansion, some macro nodes may go
3207 // dead and be removed from the list as we iterate over it. Move
3208 // Allocate nodes (processed in a second pass) at the beginning of
3209 // the list and then iterate from the last element of the list until
3210 // an Allocate node is seen. This is robust to random deletion in
3211 // the list due to nodes going dead.
3212 C->sort_macro_nodes();
3213
3214 // expand arraycopy "macro" nodes first
3215 // For ReduceBulkZeroing, we must first process all arraycopy nodes
3216 // before the allocate nodes are expanded.
3217 while (C->macro_count() > 0) {
3218 int macro_count = C->macro_count();
3219 Node * n = C->macro_node(macro_count-1);
3220 assert(n->is_macro(), "only macro nodes expected here");
3221 if (_igvn.type(n) == Type::TOP || (n->in(0) != nullptr && n->in(0)->is_top())) {
3222 // node is unreachable, so don't try to expand it
3223 C->remove_macro_node(n);
3224 continue;
3225 }
3226 if (n->is_Allocate()) {
3227 break;
3228 }
3229 // Make sure expansion will not cause node limit to be exceeded.
3230 // Worst case is a macro node gets expanded into about 200 nodes.
3231 // Allow 50% more for optimization.
3232 if (C->check_node_count(300, "out of nodes before macro expansion")) {
3233 return true;
3234 }
3235
3236 DEBUG_ONLY(int old_macro_count = C->macro_count();)
3237 switch (n->class_id()) {
3238 case Node::Class_Lock:
3239 expand_lock_node(n->as_Lock());
3240 break;
3241 case Node::Class_Unlock:
3242 expand_unlock_node(n->as_Unlock());
3243 break;
3244 case Node::Class_ArrayCopy:
3245 expand_arraycopy_node(n->as_ArrayCopy());
3246 break;
3247 case Node::Class_SubTypeCheck:
3248 expand_subtypecheck_node(n->as_SubTypeCheck());
3249 break;
3250 case Node::Class_CallStaticJava:
3251 expand_mh_intrinsic_return(n->as_CallStaticJava());
3252 C->remove_macro_node(n);
3253 break;
3254 case Node::Class_FlatArrayCheck:
3255 expand_flatarraycheck_node(n->as_FlatArrayCheck());
3256 break;
3257 default:
3258 switch (n->Opcode()) {
3259 case Op_ModD:
3260 case Op_ModF: {
3261 CallNode* mod_macro = n->as_Call();
3262 CallNode* call = new CallLeafPureNode(mod_macro->tf(), mod_macro->entry_point(),
3263 mod_macro->_name, TypeRawPtr::BOTTOM);
3264 call->init_req(TypeFunc::Control, mod_macro->in(TypeFunc::Control));
3265 call->init_req(TypeFunc::I_O, C->top());
3266 call->init_req(TypeFunc::Memory, C->top());
3267 call->init_req(TypeFunc::ReturnAdr, C->top());
3268 call->init_req(TypeFunc::FramePtr, C->top());
3269 for (unsigned int i = 0; i < mod_macro->tf()->domain_cc()->cnt() - TypeFunc::Parms; i++) {
3270 call->init_req(TypeFunc::Parms + i, mod_macro->in(TypeFunc::Parms + i));
3271 }
3272 _igvn.replace_node(mod_macro, call);
3273 transform_later(call);
3274 break;
3275 }
3276 default:
3277 assert(false, "unknown node type in macro list");
3278 }
3279 }
3280 assert(C->macro_count() == (old_macro_count - 1), "expansion must have deleted one node from macro list");
3281 if (C->failing()) return true;
3282 C->print_method(PHASE_AFTER_MACRO_EXPANSION_STEP, 5, n);
3283
3284 // Clean up the graph so we're less likely to hit the maximum node
3285 // limit
3286 _igvn.set_delay_transform(false);
3287 _igvn.optimize();
3288 if (C->failing()) return true;
3289 _igvn.set_delay_transform(true);
3290 }
3291
3292 // All nodes except Allocate nodes are expanded now. There could be
3293 // new optimization opportunities (such as folding newly created
3294 // load from a just allocated object). Run IGVN.
3295
3296 // expand "macro" nodes
3297 // nodes are removed from the macro list as they are processed
3298 while (C->macro_count() > 0) {
3299 int macro_count = C->macro_count();
3300 Node * n = C->macro_node(macro_count-1);
3301 assert(n->is_macro(), "only macro nodes expected here");
3302 if (_igvn.type(n) == Type::TOP || (n->in(0) != nullptr && n->in(0)->is_top())) {
3303 // node is unreachable, so don't try to expand it
3304 C->remove_macro_node(n);
3305 continue;
3306 }
3307 // Make sure expansion will not cause node limit to be exceeded.
3308 // Worst case is a macro node gets expanded into about 200 nodes.
3309 // Allow 50% more for optimization.
3310 if (C->check_node_count(300, "out of nodes before macro expansion")) {
3311 return true;
3312 }
3313 switch (n->class_id()) {
3314 case Node::Class_Allocate:
3315 expand_allocate(n->as_Allocate());
3316 break;
3317 case Node::Class_AllocateArray:
3318 expand_allocate_array(n->as_AllocateArray());
3319 break;
3320 default:
3321 assert(false, "unknown node type in macro list");
3322 }
3323 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
3324 if (C->failing()) return true;
3325 C->print_method(PHASE_AFTER_MACRO_EXPANSION_STEP, 5, n);
3326
3327 // Clean up the graph so we're less likely to hit the maximum node
3328 // limit
3329 _igvn.set_delay_transform(false);
3330 _igvn.optimize();
3331 if (C->failing()) return true;
3332 _igvn.set_delay_transform(true);
3333 }
3334
3335 _igvn.set_delay_transform(false);
3336 return false;
3337 }
3338
3339 #ifndef PRODUCT
3340 int PhaseMacroExpand::_objs_scalar_replaced_counter = 0;
3341 int PhaseMacroExpand::_monitor_objects_removed_counter = 0;
3342 int PhaseMacroExpand::_GC_barriers_removed_counter = 0;
3343 int PhaseMacroExpand::_memory_barriers_removed_counter = 0;
3344
3345 void PhaseMacroExpand::print_statistics() {
3346 tty->print("Objects scalar replaced = %d, ", AtomicAccess::load(&_objs_scalar_replaced_counter));
3347 tty->print("Monitor objects removed = %d, ", AtomicAccess::load(&_monitor_objects_removed_counter));
3348 tty->print("GC barriers removed = %d, ", AtomicAccess::load(&_GC_barriers_removed_counter));
3349 tty->print_cr("Memory barriers removed = %d", AtomicAccess::load(&_memory_barriers_removed_counter));
3350 }
3351
3352 int PhaseMacroExpand::count_MemBar(Compile *C) {
3353 if (!PrintOptoStatistics) {
3354 return 0;
3355 }
3356 Unique_Node_List ideal_nodes;
3357 int total = 0;
3358 ideal_nodes.map(C->live_nodes(), nullptr);
3359 ideal_nodes.push(C->root());
3360 for (uint next = 0; next < ideal_nodes.size(); ++next) {
3361 Node* n = ideal_nodes.at(next);
3362 if (n->is_MemBar()) {
3363 total++;
3364 }
3365 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3366 Node* m = n->fast_out(i);
3367 ideal_nodes.push(m);
3368 }
3369 }
3370 return total;
3371 }
3372 #endif