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 "compiler/compileLog.hpp"
27 #include "gc/shared/collectedHeap.inline.hpp"
28 #include "gc/shared/tlab_globals.hpp"
29 #include "libadt/vectset.hpp"
30 #include "memory/universe.hpp"
31 #include "opto/addnode.hpp"
32 #include "opto/arraycopynode.hpp"
33 #include "opto/callnode.hpp"
34 #include "opto/castnode.hpp"
35 #include "opto/cfgnode.hpp"
36 #include "opto/compile.hpp"
37 #include "opto/convertnode.hpp"
38 #include "opto/graphKit.hpp"
39 #include "opto/inlinetypenode.hpp"
40 #include "opto/intrinsicnode.hpp"
41 #include "opto/locknode.hpp"
42 #include "opto/loopnode.hpp"
43 #include "opto/macro.hpp"
44 #include "opto/memnode.hpp"
45 #include "opto/narrowptrnode.hpp"
46 #include "opto/node.hpp"
47 #include "opto/opaquenode.hpp"
48 #include "opto/phaseX.hpp"
49 #include "opto/rootnode.hpp"
50 #include "opto/runtime.hpp"
51 #include "opto/subnode.hpp"
52 #include "opto/subtypenode.hpp"
53 #include "opto/type.hpp"
54 #include "prims/jvmtiExport.hpp"
55 #include "runtime/continuation.hpp"
56 #include "runtime/sharedRuntime.hpp"
57 #include "runtime/stubRoutines.hpp"
58 #include "utilities/macros.hpp"
59 #include "utilities/powerOfTwo.hpp"
60 #if INCLUDE_G1GC
61 #include "gc/g1/g1ThreadLocalData.hpp"
62 #endif // INCLUDE_G1GC
63
64
65 //
66 // Replace any references to "oldref" in inputs to "use" with "newref".
67 // Returns the number of replacements made.
68 //
69 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
70 int nreplacements = 0;
71 uint req = use->req();
72 for (uint j = 0; j < use->len(); j++) {
73 Node *uin = use->in(j);
74 if (uin == oldref) {
75 if (j < req)
76 use->set_req(j, newref);
77 else
78 use->set_prec(j, newref);
79 nreplacements++;
80 } else if (j >= req && uin == nullptr) {
81 break;
82 }
83 }
84 return nreplacements;
85 }
86
87 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
88 Node* cmp;
89 if (mask != 0) {
90 Node* and_node = transform_later(new AndXNode(word, MakeConX(mask)));
91 cmp = transform_later(new CmpXNode(and_node, MakeConX(bits)));
92 } else {
93 cmp = word;
94 }
95 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
96 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
97 transform_later(iff);
98
99 // Fast path taken.
100 Node *fast_taken = transform_later(new IfFalseNode(iff));
101
102 // Fast path not-taken, i.e. slow path
103 Node *slow_taken = transform_later(new IfTrueNode(iff));
104
105 if (return_fast_path) {
106 region->init_req(edge, slow_taken); // Capture slow-control
107 return fast_taken;
108 } else {
109 region->init_req(edge, fast_taken); // Capture fast-control
110 return slow_taken;
111 }
112 }
113
114 //--------------------copy_predefined_input_for_runtime_call--------------------
115 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
116 // Set fixed predefined input arguments
117 call->init_req( TypeFunc::Control, ctrl );
118 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
119 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
120 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
121 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
122 }
123
124 //------------------------------make_slow_call---------------------------------
125 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type,
126 address slow_call, const char* leaf_name, Node* slow_path,
127 Node* parm0, Node* parm1, Node* parm2) {
128
129 // Slow-path call
130 CallNode *call = leaf_name
131 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
132 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), TypeRawPtr::BOTTOM );
133
134 // Slow path call has no side-effects, uses few values
135 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
136 if (parm0 != nullptr) call->init_req(TypeFunc::Parms+0, parm0);
137 if (parm1 != nullptr) call->init_req(TypeFunc::Parms+1, parm1);
138 if (parm2 != nullptr) call->init_req(TypeFunc::Parms+2, parm2);
139 call->copy_call_debug_info(&_igvn, oldcall);
140 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
141 _igvn.replace_node(oldcall, call);
142 transform_later(call);
143
144 return call;
145 }
146
147 void PhaseMacroExpand::eliminate_gc_barrier(Node* p2x) {
148 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
149 bs->eliminate_gc_barrier(&_igvn, p2x);
150 #ifndef PRODUCT
151 if (PrintOptoStatistics) {
152 Atomic::inc(&PhaseMacroExpand::_GC_barriers_removed_counter);
153 }
154 #endif
155 }
156
157 // Search for a memory operation for the specified memory slice.
158 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
159 Node *orig_mem = mem;
160 Node *alloc_mem = alloc->as_Allocate()->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
161 assert(alloc_mem != nullptr, "Allocation without a memory projection.");
162 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
163 while (true) {
164 if (mem == alloc_mem || mem == start_mem ) {
165 return mem; // hit one of our sentinels
166 } else if (mem->is_MergeMem()) {
167 mem = mem->as_MergeMem()->memory_at(alias_idx);
168 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
169 Node *in = mem->in(0);
170 // we can safely skip over safepoints, calls, locks and membars because we
171 // already know that the object is safe to eliminate.
172 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
173 return in;
174 } else if (in->is_Call()) {
175 CallNode *call = in->as_Call();
176 if (call->may_modify(tinst, phase)) {
177 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape");
178 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) {
179 return in;
180 }
181 }
182 mem = in->in(TypeFunc::Memory);
183 } else if (in->is_MemBar()) {
184 ArrayCopyNode* ac = nullptr;
185 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) {
186 if (ac != nullptr) {
187 assert(ac->is_clonebasic(), "Only basic clone is a non escaping clone");
188 return ac;
189 }
190 }
191 mem = in->in(TypeFunc::Memory);
192 } else {
193 #ifdef ASSERT
194 in->dump();
195 mem->dump();
196 assert(false, "unexpected projection");
197 #endif
198 }
199 } else if (mem->is_Store()) {
200 const TypePtr* atype = mem->as_Store()->adr_type();
201 int adr_idx = phase->C->get_alias_index(atype);
202 if (adr_idx == alias_idx) {
203 assert(atype->isa_oopptr(), "address type must be oopptr");
204 int adr_offset = atype->flat_offset();
205 uint adr_iid = atype->is_oopptr()->instance_id();
206 // Array elements references have the same alias_idx
207 // but different offset and different instance_id.
208 if (adr_offset == offset && adr_iid == alloc->_idx) {
209 return mem;
210 }
211 } else {
212 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
213 }
214 mem = mem->in(MemNode::Memory);
215 } else if (mem->is_ClearArray()) {
216 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
217 // Can not bypass initialization of the instance
218 // we are looking.
219 debug_only(intptr_t offset;)
220 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
221 InitializeNode* init = alloc->as_Allocate()->initialization();
222 // We are looking for stored value, return Initialize node
223 // or memory edge from Allocate node.
224 if (init != nullptr) {
225 return init;
226 } else {
227 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
228 }
229 }
230 // Otherwise skip it (the call updated 'mem' value).
231 } else if (mem->Opcode() == Op_SCMemProj) {
232 mem = mem->in(0);
233 Node* adr = nullptr;
234 if (mem->is_LoadStore()) {
235 adr = mem->in(MemNode::Address);
236 } else {
237 assert(mem->Opcode() == Op_EncodeISOArray ||
238 mem->Opcode() == Op_StrCompressedCopy, "sanity");
239 adr = mem->in(3); // Destination array
240 }
241 const TypePtr* atype = adr->bottom_type()->is_ptr();
242 int adr_idx = phase->C->get_alias_index(atype);
243 if (adr_idx == alias_idx) {
244 DEBUG_ONLY(mem->dump();)
245 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
246 return nullptr;
247 }
248 mem = mem->in(MemNode::Memory);
249 } else if (mem->Opcode() == Op_StrInflatedCopy) {
250 Node* adr = mem->in(3); // Destination array
251 const TypePtr* atype = adr->bottom_type()->is_ptr();
252 int adr_idx = phase->C->get_alias_index(atype);
253 if (adr_idx == alias_idx) {
254 DEBUG_ONLY(mem->dump();)
255 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
256 return nullptr;
257 }
258 mem = mem->in(MemNode::Memory);
259 } else {
260 return mem;
261 }
262 assert(mem != orig_mem, "dead memory loop");
263 }
264 }
265
266 // Generate loads from source of the arraycopy for fields of
267 // destination needed at a deoptimization point
268 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) {
269 BasicType bt = ft;
270 const Type *type = ftype;
271 if (ft == T_NARROWOOP) {
272 bt = T_OBJECT;
273 type = ftype->make_oopptr();
274 }
275 Node* res = nullptr;
276 if (ac->is_clonebasic()) {
277 assert(ac->in(ArrayCopyNode::Src) != ac->in(ArrayCopyNode::Dest), "clone source equals destination");
278 Node* base = ac->in(ArrayCopyNode::Src);
279 Node* adr = _igvn.transform(new AddPNode(base, base, _igvn.MakeConX(offset)));
280 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
281 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
282 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
283 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
284 } else {
285 if (ac->modifies(offset, offset, &_igvn, true)) {
286 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
287 uint shift = exact_log2(type2aelembytes(bt));
288 Node* src_pos = ac->in(ArrayCopyNode::SrcPos);
289 Node* dest_pos = ac->in(ArrayCopyNode::DestPos);
290 const TypeInt* src_pos_t = _igvn.type(src_pos)->is_int();
291 const TypeInt* dest_pos_t = _igvn.type(dest_pos)->is_int();
292
293 Node* adr = nullptr;
294 Node* base = ac->in(ArrayCopyNode::Src);
295 const TypeAryPtr* adr_type = _igvn.type(base)->is_aryptr();
296 if (adr_type->is_flat()) {
297 shift = adr_type->flat_log_elem_size();
298 }
299 if (src_pos_t->is_con() && dest_pos_t->is_con()) {
300 intptr_t off = ((src_pos_t->get_con() - dest_pos_t->get_con()) << shift) + offset;
301 adr = _igvn.transform(new AddPNode(base, base, _igvn.MakeConX(off)));
302 adr_type = _igvn.type(adr)->is_aryptr();
303 assert(adr_type == _igvn.type(base)->is_aryptr()->add_field_offset_and_offset(off), "incorrect address type");
304 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
305 // Don't emit a new load from src if src == dst but try to get the value from memory instead
306 return value_from_mem(ac->in(TypeFunc::Memory), ctl, ft, ftype, adr_type, alloc);
307 }
308 } else {
309 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
310 // Non constant offset in the array: we can't statically
311 // determine the value
312 return nullptr;
313 }
314 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
315 #ifdef _LP64
316 diff = _igvn.transform(new ConvI2LNode(diff));
317 #endif
318 diff = _igvn.transform(new LShiftXNode(diff, _igvn.intcon(shift)));
319
320 Node* off = _igvn.transform(new AddXNode(_igvn.MakeConX(offset), diff));
321 adr = _igvn.transform(new AddPNode(base, base, off));
322 // In the case of a flat inline type array, each field has its
323 // own slice so we need to extract the field being accessed from
324 // the address computation
325 adr_type = adr_type->add_field_offset_and_offset(offset)->add_offset(Type::OffsetBot)->is_aryptr();
326 adr = _igvn.transform(new CastPPNode(ctl, adr, adr_type));
327 }
328 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
329 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
330 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
331 }
332 }
333 if (res != nullptr) {
334 if (ftype->isa_narrowoop()) {
335 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
336 assert(res->isa_DecodeN(), "should be narrow oop");
337 res = _igvn.transform(new EncodePNode(res, ftype));
338 }
339 return res;
340 }
341 return nullptr;
342 }
343
344 //
345 // Given a Memory Phi, compute a value Phi containing the values from stores
346 // on the input paths.
347 // Note: this function is recursive, its depth is limited by the "level" argument
348 // Returns the computed Phi, or null if it cannot compute it.
349 Node *PhaseMacroExpand::value_from_mem_phi(Node *mem, BasicType ft, const Type *phi_type, const TypeOopPtr *adr_t, AllocateNode *alloc, Node_Stack *value_phis, int level) {
350 assert(mem->is_Phi(), "sanity");
351 int alias_idx = C->get_alias_index(adr_t);
352 int offset = adr_t->flat_offset();
353 int instance_id = adr_t->instance_id();
354
355 // Check if an appropriate value phi already exists.
356 Node* region = mem->in(0);
357 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
358 Node* phi = region->fast_out(k);
359 if (phi->is_Phi() && phi != mem &&
360 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
361 return phi;
362 }
363 }
364 // Check if an appropriate new value phi already exists.
365 Node* new_phi = value_phis->find(mem->_idx);
366 if (new_phi != nullptr)
367 return new_phi;
368
369 if (level <= 0) {
370 return nullptr; // Give up: phi tree too deep
371 }
372 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
373 Node *alloc_mem = alloc->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
374 assert(alloc_mem != nullptr, "Allocation without a memory projection.");
375
376 uint length = mem->req();
377 GrowableArray <Node *> values(length, length, nullptr);
378
379 // create a new Phi for the value
380 PhiNode *phi = new PhiNode(mem->in(0), phi_type, nullptr, mem->_idx, instance_id, alias_idx, offset);
381 transform_later(phi);
382 value_phis->push(phi, mem->_idx);
383
384 for (uint j = 1; j < length; j++) {
385 Node *in = mem->in(j);
386 if (in == nullptr || in->is_top()) {
387 values.at_put(j, in);
388 } else {
389 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
390 if (val == start_mem || val == alloc_mem) {
391 // hit a sentinel, return appropriate value
392 Node* init_value = alloc->in(AllocateNode::InitValue);
393 if (init_value != nullptr) {
394 if (val == start_mem) {
395 // TODO 8350865 Somehow we ended up with root mem and therefore walked past the alloc. Fix this. Triggered by TestGenerated::test15
396 // Don't we need field_value_by_offset?
397 return nullptr;
398 }
399 values.at_put(j, init_value);
400 } else {
401 assert(alloc->in(AllocateNode::RawInitValue) == nullptr, "init value may not be null");
402 values.at_put(j, _igvn.zerocon(ft));
403 }
404 continue;
405 }
406 if (val->is_Initialize()) {
407 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
408 }
409 if (val == nullptr) {
410 return nullptr; // can't find a value on this path
411 }
412 if (val == mem) {
413 values.at_put(j, mem);
414 } else if (val->is_Store()) {
415 Node* n = val->in(MemNode::ValueIn);
416 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
417 n = bs->step_over_gc_barrier(n);
418 if (is_subword_type(ft)) {
419 n = Compile::narrow_value(ft, n, phi_type, &_igvn, true);
420 }
421 values.at_put(j, n);
422 } else if (val->is_Proj() && val->in(0) == alloc) {
423 Node* init_value = alloc->in(AllocateNode::InitValue);
424 if (init_value != nullptr) {
425 // TODO 8350865 Is this correct for non-all-zero init values? Don't we need field_value_by_offset?
426 values.at_put(j, init_value);
427 } else {
428 assert(alloc->in(AllocateNode::RawInitValue) == nullptr, "init value may not be null");
429 values.at_put(j, _igvn.zerocon(ft));
430 }
431 } else if (val->is_Phi()) {
432 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
433 if (val == nullptr) {
434 return nullptr;
435 }
436 values.at_put(j, val);
437 } else if (val->Opcode() == Op_SCMemProj) {
438 assert(val->in(0)->is_LoadStore() ||
439 val->in(0)->Opcode() == Op_EncodeISOArray ||
440 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
441 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
442 return nullptr;
443 } else if (val->is_ArrayCopy()) {
444 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc);
445 if (res == nullptr) {
446 return nullptr;
447 }
448 values.at_put(j, res);
449 } else if (val->is_top()) {
450 // This indicates that this path into the phi is dead. Top will eventually also propagate into the Region.
451 // IGVN will clean this up later.
452 values.at_put(j, val);
453 } else {
454 DEBUG_ONLY( val->dump(); )
455 assert(false, "unknown node on this path");
456 return nullptr; // unknown node on this path
457 }
458 }
459 }
460 // Set Phi's inputs
461 for (uint j = 1; j < length; j++) {
462 if (values.at(j) == mem) {
463 phi->init_req(j, phi);
464 } else {
465 phi->init_req(j, values.at(j));
466 }
467 }
468 return phi;
469 }
470
471 // Search the last value stored into the object's field.
472 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
473 assert(adr_t->is_known_instance_field(), "instance required");
474 int instance_id = adr_t->instance_id();
475 assert((uint)instance_id == alloc->_idx, "wrong allocation");
476
477 int alias_idx = C->get_alias_index(adr_t);
478 int offset = adr_t->flat_offset();
479 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
480 Node *alloc_mem = alloc->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
481 assert(alloc_mem != nullptr, "Allocation without a memory projection.");
482 VectorSet visited;
483
484 bool done = sfpt_mem == alloc_mem;
485 Node *mem = sfpt_mem;
486 while (!done) {
487 if (visited.test_set(mem->_idx)) {
488 return nullptr; // found a loop, give up
489 }
490 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
491 if (mem == start_mem || mem == alloc_mem) {
492 done = true; // hit a sentinel, return appropriate 0 value
493 } else if (mem->is_Initialize()) {
494 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
495 if (mem == nullptr) {
496 done = true; // Something went wrong.
497 } else if (mem->is_Store()) {
498 const TypePtr* atype = mem->as_Store()->adr_type();
499 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
500 done = true;
501 }
502 } else if (mem->is_Store()) {
503 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
504 assert(atype != nullptr, "address type must be oopptr");
505 assert(C->get_alias_index(atype) == alias_idx &&
506 atype->is_known_instance_field() && atype->flat_offset() == offset &&
507 atype->instance_id() == instance_id, "store is correct memory slice");
508 done = true;
509 } else if (mem->is_Phi()) {
510 // try to find a phi's unique input
511 Node *unique_input = nullptr;
512 Node *top = C->top();
513 for (uint i = 1; i < mem->req(); i++) {
514 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
515 if (n == nullptr || n == top || n == mem) {
516 continue;
517 } else if (unique_input == nullptr) {
518 unique_input = n;
519 } else if (unique_input != n) {
520 unique_input = top;
521 break;
522 }
523 }
524 if (unique_input != nullptr && unique_input != top) {
525 mem = unique_input;
526 } else {
527 done = true;
528 }
529 } else if (mem->is_ArrayCopy()) {
530 done = true;
531 } else {
532 DEBUG_ONLY( mem->dump(); )
533 assert(false, "unexpected node");
534 }
535 }
536 if (mem != nullptr) {
537 if (mem == start_mem || mem == alloc_mem) {
538 // hit a sentinel, return appropriate value
539 Node* init_value = alloc->in(AllocateNode::InitValue);
540 if (init_value != nullptr) {
541 if (adr_t->is_flat()) {
542 if (init_value->is_EncodeP()) {
543 init_value = init_value->in(1);
544 }
545 assert(adr_t->is_aryptr()->field_offset().get() != Type::OffsetBot, "Unknown offset");
546 offset = adr_t->is_aryptr()->field_offset().get() + init_value->bottom_type()->inline_klass()->payload_offset();
547 init_value = init_value->as_InlineType()->field_value_by_offset(offset, true);
548 if (ft == T_NARROWOOP) {
549 init_value = transform_later(new EncodePNode(init_value, init_value->bottom_type()->make_ptr()));
550 }
551 }
552 return init_value;
553 }
554 assert(alloc->in(AllocateNode::RawInitValue) == nullptr, "init value may not be null");
555 return _igvn.zerocon(ft);
556 } else if (mem->is_Store()) {
557 Node* n = mem->in(MemNode::ValueIn);
558 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
559 n = bs->step_over_gc_barrier(n);
560 return n;
561 } else if (mem->is_Phi()) {
562 // attempt to produce a Phi reflecting the values on the input paths of the Phi
563 Node_Stack value_phis(8);
564 Node* phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
565 if (phi != nullptr) {
566 return phi;
567 } else {
568 // Kill all new Phis
569 while(value_phis.is_nonempty()) {
570 Node* n = value_phis.node();
571 _igvn.replace_node(n, C->top());
572 value_phis.pop();
573 }
574 }
575 } else if (mem->is_ArrayCopy()) {
576 Node* ctl = mem->in(0);
577 Node* m = mem->in(TypeFunc::Memory);
578 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj()) {
579 // pin the loads in the uncommon trap path
580 ctl = sfpt_ctl;
581 m = sfpt_mem;
582 }
583 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc);
584 }
585 }
586 // Something went wrong.
587 return nullptr;
588 }
589
590 // Search the last value stored into the inline type's fields (for flat arrays).
591 Node* PhaseMacroExpand::inline_type_from_mem(Node* mem, Node* ctl, ciInlineKlass* vk, const TypeAryPtr* adr_type, int offset, AllocateNode* alloc) {
592 // Subtract the offset of the first field to account for the missing oop header
593 offset -= vk->payload_offset();
594 // Create a new InlineTypeNode and retrieve the field values from memory
595 InlineTypeNode* vt = InlineTypeNode::make_uninitialized(_igvn, vk);
596 transform_later(vt);
597 for (int i = 0; i < vk->nof_declared_nonstatic_fields(); ++i) {
598 ciType* field_type = vt->field_type(i);
599 int field_offset = offset + vt->field_offset(i);
600 Node* value = nullptr;
601 if (vt->field_is_flat(i)) {
602 // TODO 8350865 Fix this
603 // assert(vt->field_is_null_free(i), "Unexpected nullable flat field");
604 if (!vt->field_is_null_free(i)) {
605 return nullptr;
606 }
607 value = inline_type_from_mem(mem, ctl, field_type->as_inline_klass(), adr_type, field_offset, alloc);
608 } else {
609 const Type* ft = Type::get_const_type(field_type);
610 BasicType bt = type2field[field_type->basic_type()];
611 if (UseCompressedOops && !is_java_primitive(bt)) {
612 ft = ft->make_narrowoop();
613 bt = T_NARROWOOP;
614 }
615 // Each inline type field has its own memory slice
616 adr_type = adr_type->with_field_offset(field_offset);
617 value = value_from_mem(mem, ctl, bt, ft, adr_type, alloc);
618 if (value != nullptr && ft->isa_narrowoop()) {
619 assert(UseCompressedOops, "unexpected narrow oop");
620 if (value->is_EncodeP()) {
621 value = value->in(1);
622 } else if (!value->is_InlineType()) {
623 value = transform_later(new DecodeNNode(value, value->get_ptr_type()));
624 }
625 }
626 }
627 if (value != nullptr) {
628 vt->set_field_value(i, value);
629 } else {
630 // We might have reached the TrackedInitializationLimit
631 return nullptr;
632 }
633 }
634 return vt;
635 }
636
637 // Check the possibility of scalar replacement.
638 bool PhaseMacroExpand::can_eliminate_allocation(PhaseIterGVN* igvn, AllocateNode *alloc, GrowableArray <SafePointNode *>* safepoints) {
639 // Scan the uses of the allocation to check for anything that would
640 // prevent us from eliminating it.
641 NOT_PRODUCT( const char* fail_eliminate = nullptr; )
642 DEBUG_ONLY( Node* disq_node = nullptr; )
643 bool can_eliminate = true;
644 bool reduce_merge_precheck = (safepoints == nullptr);
645
646 Unique_Node_List worklist;
647 Node* res = alloc->result_cast();
648 const TypeOopPtr* res_type = nullptr;
649 if (res == nullptr) {
650 // All users were eliminated.
651 } else if (!res->is_CheckCastPP()) {
652 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
653 can_eliminate = false;
654 } else {
655 worklist.push(res);
656 res_type = igvn->type(res)->isa_oopptr();
657 if (res_type == nullptr) {
658 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
659 can_eliminate = false;
660 } else if (!res_type->klass_is_exact()) {
661 NOT_PRODUCT(fail_eliminate = "Not an exact type.";)
662 can_eliminate = false;
663 } else if (res_type->isa_aryptr()) {
664 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
665 if (length < 0) {
666 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
667 can_eliminate = false;
668 }
669 }
670 }
671
672 while (can_eliminate && worklist.size() > 0) {
673 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
674 res = worklist.pop();
675 for (DUIterator_Fast jmax, j = res->fast_outs(jmax); j < jmax && can_eliminate; j++) {
676 Node* use = res->fast_out(j);
677
678 if (use->is_AddP()) {
679 const TypePtr* addp_type = igvn->type(use)->is_ptr();
680 int offset = addp_type->offset();
681
682 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
683 NOT_PRODUCT(fail_eliminate = "Undefined field reference";)
684 can_eliminate = false;
685 break;
686 }
687 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
688 k < kmax && can_eliminate; k++) {
689 Node* n = use->fast_out(k);
690 if (!n->is_Store() && n->Opcode() != Op_CastP2X && !bs->is_gc_pre_barrier_node(n) && !reduce_merge_precheck) {
691 DEBUG_ONLY(disq_node = n;)
692 if (n->is_Load() || n->is_LoadStore()) {
693 NOT_PRODUCT(fail_eliminate = "Field load";)
694 } else {
695 NOT_PRODUCT(fail_eliminate = "Not store field reference";)
696 }
697 can_eliminate = false;
698 }
699 }
700 } else if (use->is_ArrayCopy() &&
701 (use->as_ArrayCopy()->is_clonebasic() ||
702 use->as_ArrayCopy()->is_arraycopy_validated() ||
703 use->as_ArrayCopy()->is_copyof_validated() ||
704 use->as_ArrayCopy()->is_copyofrange_validated()) &&
705 use->in(ArrayCopyNode::Dest) == res) {
706 // ok to eliminate
707 } else if (use->is_SafePoint()) {
708 SafePointNode* sfpt = use->as_SafePoint();
709 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
710 // Object is passed as argument.
711 DEBUG_ONLY(disq_node = use;)
712 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
713 can_eliminate = false;
714 }
715 Node* sfptMem = sfpt->memory();
716 if (sfptMem == nullptr || sfptMem->is_top()) {
717 DEBUG_ONLY(disq_node = use;)
718 NOT_PRODUCT(fail_eliminate = "null or TOP memory";)
719 can_eliminate = false;
720 } else if (!reduce_merge_precheck) {
721 assert(!res->is_Phi() || !res->as_Phi()->can_be_inline_type(), "Inline type allocations should not have safepoint uses");
722 safepoints->append_if_missing(sfpt);
723 }
724 } else if (use->is_InlineType() && use->as_InlineType()->get_oop() == res) {
725 // Look at uses
726 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); k < kmax; k++) {
727 Node* u = use->fast_out(k);
728 if (u->is_InlineType()) {
729 // Use in flat field can be eliminated
730 InlineTypeNode* vt = u->as_InlineType();
731 for (uint i = 0; i < vt->field_count(); ++i) {
732 if (vt->field_value(i) == use && !vt->field_is_flat(i)) {
733 can_eliminate = false; // Use in non-flat field
734 break;
735 }
736 }
737 } else {
738 // Add other uses to the worklist to process individually
739 worklist.push(use);
740 }
741 }
742 } else if (use->Opcode() == Op_StoreX && use->in(MemNode::Address) == res) {
743 // Store to mark word of inline type larval buffer
744 assert(res_type->is_inlinetypeptr(), "Unexpected store to mark word");
745 } else if (res_type->is_inlinetypeptr() && (use->Opcode() == Op_MemBarRelease || use->Opcode() == Op_MemBarStoreStore)) {
746 // Inline type buffer allocations are followed by a membar
747 } else if (reduce_merge_precheck &&
748 (use->is_Phi() || use->is_EncodeP() ||
749 use->Opcode() == Op_MemBarRelease ||
750 (UseStoreStoreForCtor && use->Opcode() == Op_MemBarStoreStore))) {
751 // Nothing to do
752 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
753 if (use->is_Phi()) {
754 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
755 NOT_PRODUCT(fail_eliminate = "Object is return value";)
756 } else {
757 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
758 }
759 DEBUG_ONLY(disq_node = use;)
760 } else {
761 if (use->Opcode() == Op_Return) {
762 NOT_PRODUCT(fail_eliminate = "Object is return value";)
763 } else {
764 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
765 }
766 DEBUG_ONLY(disq_node = use;)
767 }
768 can_eliminate = false;
769 } else {
770 assert(use->Opcode() == Op_CastP2X, "should be");
771 assert(!use->has_out_with(Op_OrL), "should have been removed because oop is never null");
772 }
773 }
774 }
775
776 #ifndef PRODUCT
777 if (PrintEliminateAllocations && safepoints != nullptr) {
778 if (can_eliminate) {
779 tty->print("Scalar ");
780 if (res == nullptr)
781 alloc->dump();
782 else
783 res->dump();
784 } else {
785 tty->print("NotScalar (%s)", fail_eliminate);
786 if (res == nullptr)
787 alloc->dump();
788 else
789 res->dump();
790 #ifdef ASSERT
791 if (disq_node != nullptr) {
792 tty->print(" >>>> ");
793 disq_node->dump();
794 }
795 #endif /*ASSERT*/
796 }
797 }
798
799 if (TraceReduceAllocationMerges && !can_eliminate && reduce_merge_precheck) {
800 tty->print_cr("\tCan't eliminate allocation because '%s': ", fail_eliminate != nullptr ? fail_eliminate : "");
801 DEBUG_ONLY(if (disq_node != nullptr) disq_node->dump();)
802 }
803 #endif
804 return can_eliminate;
805 }
806
807 void PhaseMacroExpand::undo_previous_scalarizations(GrowableArray <SafePointNode *> safepoints_done, AllocateNode* alloc) {
808 Node* res = alloc->result_cast();
809 int nfields = 0;
810 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
811
812 if (res != nullptr) {
813 const TypeOopPtr* res_type = _igvn.type(res)->isa_oopptr();
814
815 if (res_type->isa_instptr()) {
816 // find the fields of the class which will be needed for safepoint debug information
817 ciInstanceKlass* iklass = res_type->is_instptr()->instance_klass();
818 nfields = iklass->nof_nonstatic_fields();
819 } else {
820 // find the array's elements which will be needed for safepoint debug information
821 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
822 assert(nfields >= 0, "must be an array klass.");
823 }
824 }
825
826 // rollback processed safepoints
827 while (safepoints_done.length() > 0) {
828 SafePointNode* sfpt_done = safepoints_done.pop();
829 // remove any extra entries we added to the safepoint
830 uint last = sfpt_done->req() - 1;
831 for (int k = 0; k < nfields; k++) {
832 sfpt_done->del_req(last--);
833 }
834 JVMState *jvms = sfpt_done->jvms();
835 jvms->set_endoff(sfpt_done->req());
836 // Now make a pass over the debug information replacing any references
837 // to SafePointScalarObjectNode with the allocated object.
838 int start = jvms->debug_start();
839 int end = jvms->debug_end();
840 for (int i = start; i < end; i++) {
841 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
842 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
843 if (scobj->first_index(jvms) == sfpt_done->req() &&
844 scobj->n_fields() == (uint)nfields) {
845 assert(scobj->alloc() == alloc, "sanity");
846 sfpt_done->set_req(i, res);
847 }
848 }
849 }
850 _igvn._worklist.push(sfpt_done);
851 }
852 }
853
854 SafePointScalarObjectNode* PhaseMacroExpand::create_scalarized_object_description(AllocateNode *alloc, SafePointNode* sfpt,
855 Unique_Node_List* value_worklist) {
856 // Fields of scalar objs are referenced only at the end
857 // of regular debuginfo at the last (youngest) JVMS.
858 // Record relative start index.
859 ciInstanceKlass* iklass = nullptr;
860 BasicType basic_elem_type = T_ILLEGAL;
861 const Type* field_type = nullptr;
862 const TypeOopPtr* res_type = nullptr;
863 int nfields = 0;
864 int array_base = 0;
865 int element_size = 0;
866 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
867 Node* res = alloc->result_cast();
868
869 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
870 assert(sfpt->jvms() != nullptr, "missed JVMS");
871
872 if (res != nullptr) { // Could be null when there are no users
873 res_type = _igvn.type(res)->isa_oopptr();
874
875 if (res_type->isa_instptr()) {
876 // find the fields of the class which will be needed for safepoint debug information
877 iklass = res_type->is_instptr()->instance_klass();
878 nfields = iklass->nof_nonstatic_fields();
879 } else {
880 // find the array's elements which will be needed for safepoint debug information
881 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
882 assert(nfields >= 0, "must be an array klass.");
883 basic_elem_type = res_type->is_aryptr()->elem()->array_element_basic_type();
884 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
885 element_size = type2aelembytes(basic_elem_type);
886 field_type = res_type->is_aryptr()->elem();
887 if (res_type->is_flat()) {
888 // Flat inline type array
889 element_size = res_type->is_aryptr()->flat_elem_size();
890 }
891 }
892
893 if (res->bottom_type()->is_inlinetypeptr()) {
894 // Nullable inline types have an IsInit field which is added to the safepoint when scalarizing them (see
895 // InlineTypeNode::make_scalar_in_safepoint()). When having circular inline types, we stop scalarizing at depth 1
896 // to avoid an endless recursion. Therefore, we do not have a SafePointScalarObjectNode node here, yet.
897 // We are about to create a SafePointScalarObjectNode as if this is a normal object. Add an additional int input
898 // with value 1 which sets IsInit to true to indicate that the object is always non-null. This input is checked
899 // later in PhaseOutput::filLocArray() for inline types.
900 sfpt->add_req(_igvn.intcon(1));
901 }
902 }
903
904 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type, alloc, first_ind, sfpt->jvms()->depth(), nfields);
905 sobj->init_req(0, C->root());
906 transform_later(sobj);
907
908 // Scan object's fields adding an input to the safepoint for each field.
909 for (int j = 0; j < nfields; j++) {
910 intptr_t offset;
911 ciField* field = nullptr;
912 if (iklass != nullptr) {
913 field = iklass->nonstatic_field_at(j);
914 offset = field->offset_in_bytes();
915 ciType* elem_type = field->type();
916 basic_elem_type = field->layout_type();
917 assert(!field->is_flat(), "flat inline type fields should not have safepoint uses");
918
919 ciField* flat_field = iklass->get_non_flat_field_by_offset(offset);
920 if (flat_field != nullptr && flat_field->is_flat() && !flat_field->is_null_free()) {
921 // TODO 8353432 Add support for nullable, flat fields in non-value class holders
922 // Below code only iterates over the flat representation and therefore misses to
923 // add null markers like we do in InlineTypeNode::add_fields_to_safepoint for value
924 // class holders.
925 _igvn._worklist.push(sfpt);
926 return nullptr;
927 }
928
929 // The next code is taken from Parse::do_get_xxx().
930 if (is_reference_type(basic_elem_type)) {
931 if (!elem_type->is_loaded()) {
932 field_type = TypeInstPtr::BOTTOM;
933 } else if (field != nullptr && field->is_static_constant()) {
934 ciObject* con = field->constant_value().as_object();
935 // Do not "join" in the previous type; it doesn't add value,
936 // and may yield a vacuous result if the field is of interface type.
937 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
938 assert(field_type != nullptr, "field singleton type must be consistent");
939 } else {
940 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
941 }
942 if (UseCompressedOops) {
943 field_type = field_type->make_narrowoop();
944 basic_elem_type = T_NARROWOOP;
945 }
946 } else {
947 field_type = Type::get_const_basic_type(basic_elem_type);
948 }
949 } else {
950 offset = array_base + j * (intptr_t)element_size;
951 }
952
953 Node* field_val = nullptr;
954 const TypeOopPtr* field_addr_type = res_type->add_offset(offset)->isa_oopptr();
955 if (res_type->is_flat()) {
956 ciInlineKlass* inline_klass = res_type->is_aryptr()->elem()->inline_klass();
957 assert(inline_klass->maybe_flat_in_array(), "must be flat in array");
958 field_val = inline_type_from_mem(sfpt->memory(), sfpt->control(), inline_klass, field_addr_type->isa_aryptr(), 0, alloc);
959 } else {
960 field_val = value_from_mem(sfpt->memory(), sfpt->control(), basic_elem_type, field_type, field_addr_type, alloc);
961 }
962
963 // We weren't able to find a value for this field,
964 // give up on eliminating this allocation.
965 if (field_val == nullptr) {
966 uint last = sfpt->req() - 1;
967 for (int k = 0; k < j; k++) {
968 sfpt->del_req(last--);
969 }
970 _igvn._worklist.push(sfpt);
971
972 #ifndef PRODUCT
973 if (PrintEliminateAllocations) {
974 if (field != nullptr) {
975 tty->print("=== At SafePoint node %d can't find value of field: ", sfpt->_idx);
976 field->print();
977 int field_idx = C->get_alias_index(field_addr_type);
978 tty->print(" (alias_idx=%d)", field_idx);
979 } else { // Array's element
980 tty->print("=== At SafePoint node %d can't find value of array element [%d]", sfpt->_idx, j);
981 }
982 tty->print(", which prevents elimination of: ");
983 if (res == nullptr)
984 alloc->dump();
985 else
986 res->dump();
987 }
988 #endif
989
990 return nullptr;
991 }
992
993 if (UseCompressedOops && field_type->isa_narrowoop()) {
994 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
995 // to be able scalar replace the allocation.
996 if (field_val->is_EncodeP()) {
997 field_val = field_val->in(1);
998 } else if (!field_val->is_InlineType()) {
999 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
1000 }
1001 }
1002
1003 // Keep track of inline types to scalarize them later
1004 if (field_val->is_InlineType()) {
1005 value_worklist->push(field_val);
1006 } else if (field_val->is_Phi()) {
1007 PhiNode* phi = field_val->as_Phi();
1008 // Eagerly replace inline type phis now since we could be removing an inline type allocation where we must
1009 // scalarize all its fields in safepoints.
1010 field_val = phi->try_push_inline_types_down(&_igvn, true);
1011 if (field_val->is_InlineType()) {
1012 value_worklist->push(field_val);
1013 }
1014 }
1015 sfpt->add_req(field_val);
1016 }
1017
1018 sfpt->jvms()->set_endoff(sfpt->req());
1019
1020 return sobj;
1021 }
1022
1023 // Do scalar replacement.
1024 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
1025 GrowableArray <SafePointNode *> safepoints_done;
1026 Node* res = alloc->result_cast();
1027 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
1028 const TypeOopPtr* res_type = nullptr;
1029 if (res != nullptr) { // Could be null when there are no users
1030 res_type = _igvn.type(res)->isa_oopptr();
1031 }
1032
1033 // Process the safepoint uses
1034 assert(safepoints.length() == 0 || !res_type->is_inlinetypeptr() || C->has_circular_inline_type(),
1035 "Inline type allocations should have been scalarized earlier");
1036 Unique_Node_List value_worklist;
1037 while (safepoints.length() > 0) {
1038 SafePointNode* sfpt = safepoints.pop();
1039 SafePointScalarObjectNode* sobj = create_scalarized_object_description(alloc, sfpt, &value_worklist);
1040
1041 if (sobj == nullptr) {
1042 undo_previous_scalarizations(safepoints_done, alloc);
1043 return false;
1044 }
1045
1046 // Now make a pass over the debug information replacing any references
1047 // to the allocated object with "sobj"
1048 JVMState *jvms = sfpt->jvms();
1049 sfpt->replace_edges_in_range(res, sobj, jvms->debug_start(), jvms->debug_end(), &_igvn);
1050 _igvn._worklist.push(sfpt);
1051
1052 // keep it for rollback
1053 safepoints_done.append_if_missing(sfpt);
1054 }
1055 // Scalarize inline types that were added to the safepoint.
1056 // Don't allow linking a constant oop (if available) for flat array elements
1057 // because Deoptimization::reassign_flat_array_elements needs field values.
1058 bool allow_oop = (res_type != nullptr) && !res_type->is_flat();
1059 for (uint i = 0; i < value_worklist.size(); ++i) {
1060 InlineTypeNode* vt = value_worklist.at(i)->as_InlineType();
1061 vt->make_scalar_in_safepoints(&_igvn, allow_oop);
1062 }
1063 return true;
1064 }
1065
1066 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
1067 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
1068 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
1069 if (ctl_proj != nullptr) {
1070 igvn.replace_node(ctl_proj, n->in(0));
1071 }
1072 if (mem_proj != nullptr) {
1073 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
1074 }
1075 }
1076
1077 // Process users of eliminated allocation.
1078 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc, bool inline_alloc) {
1079 Unique_Node_List worklist;
1080 Node* res = alloc->result_cast();
1081 if (res != nullptr) {
1082 worklist.push(res);
1083 }
1084 while (worklist.size() > 0) {
1085 res = worklist.pop();
1086 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
1087 Node *use = res->last_out(j);
1088 uint oc1 = res->outcnt();
1089
1090 if (use->is_AddP()) {
1091 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
1092 Node *n = use->last_out(k);
1093 uint oc2 = use->outcnt();
1094 if (n->is_Store()) {
1095 for (DUIterator_Fast pmax, p = n->fast_outs(pmax); p < pmax; p++) {
1096 MemBarNode* mb = n->fast_out(p)->isa_MemBar();
1097 if (mb != nullptr && mb->req() <= MemBarNode::Precedent && mb->in(MemBarNode::Precedent) == n) {
1098 // MemBarVolatiles should have been removed by MemBarNode::Ideal() for non-inline allocations
1099 assert(inline_alloc, "MemBarVolatile should be eliminated for non-escaping object");
1100 mb->remove(&_igvn);
1101 }
1102 }
1103 _igvn.replace_node(n, n->in(MemNode::Memory));
1104 } else {
1105 eliminate_gc_barrier(n);
1106 }
1107 k -= (oc2 - use->outcnt());
1108 }
1109 _igvn.remove_dead_node(use);
1110 } else if (use->is_ArrayCopy()) {
1111 // Disconnect ArrayCopy node
1112 ArrayCopyNode* ac = use->as_ArrayCopy();
1113 if (ac->is_clonebasic()) {
1114 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
1115 disconnect_projections(ac, _igvn);
1116 assert(alloc->in(TypeFunc::Memory)->is_Proj() && alloc->in(TypeFunc::Memory)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
1117 Node* membar_before = alloc->in(TypeFunc::Memory)->in(0);
1118 disconnect_projections(membar_before->as_MemBar(), _igvn);
1119 if (membar_after->is_MemBar()) {
1120 disconnect_projections(membar_after->as_MemBar(), _igvn);
1121 }
1122 } else {
1123 assert(ac->is_arraycopy_validated() ||
1124 ac->is_copyof_validated() ||
1125 ac->is_copyofrange_validated(), "unsupported");
1126 CallProjections* callprojs = ac->extract_projections(true);
1127
1128 _igvn.replace_node(callprojs->fallthrough_ioproj, ac->in(TypeFunc::I_O));
1129 _igvn.replace_node(callprojs->fallthrough_memproj, ac->in(TypeFunc::Memory));
1130 _igvn.replace_node(callprojs->fallthrough_catchproj, ac->in(TypeFunc::Control));
1131
1132 // Set control to top. IGVN will remove the remaining projections
1133 ac->set_req(0, top());
1134 ac->replace_edge(res, top(), &_igvn);
1135
1136 // Disconnect src right away: it can help find new
1137 // opportunities for allocation elimination
1138 Node* src = ac->in(ArrayCopyNode::Src);
1139 ac->replace_edge(src, top(), &_igvn);
1140 // src can be top at this point if src and dest of the
1141 // arraycopy were the same
1142 if (src->outcnt() == 0 && !src->is_top()) {
1143 _igvn.remove_dead_node(src);
1144 }
1145 }
1146 _igvn._worklist.push(ac);
1147 } else if (use->is_InlineType()) {
1148 assert(use->as_InlineType()->get_oop() == res, "unexpected inline type ptr use");
1149 // Cut off oop input and remove known instance id from type
1150 _igvn.rehash_node_delayed(use);
1151 use->as_InlineType()->set_oop(_igvn, _igvn.zerocon(T_OBJECT));
1152 const TypeOopPtr* toop = _igvn.type(use)->is_oopptr()->cast_to_instance_id(TypeOopPtr::InstanceBot);
1153 _igvn.set_type(use, toop);
1154 use->as_InlineType()->set_type(toop);
1155 // Process users
1156 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); k < kmax; k++) {
1157 Node* u = use->fast_out(k);
1158 if (!u->is_InlineType()) {
1159 worklist.push(u);
1160 }
1161 }
1162 } else if (use->Opcode() == Op_StoreX && use->in(MemNode::Address) == res) {
1163 // Store to mark word of inline type larval buffer
1164 assert(inline_alloc, "Unexpected store to mark word");
1165 _igvn.replace_node(use, use->in(MemNode::Memory));
1166 } else if (use->Opcode() == Op_MemBarRelease || use->Opcode() == Op_MemBarStoreStore) {
1167 // Inline type buffer allocations are followed by a membar
1168 assert(inline_alloc, "Unexpected MemBarRelease");
1169 use->as_MemBar()->remove(&_igvn);
1170 } else {
1171 eliminate_gc_barrier(use);
1172 }
1173 j -= (oc1 - res->outcnt());
1174 }
1175 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
1176 _igvn.remove_dead_node(res);
1177 }
1178
1179 //
1180 // Process other users of allocation's projections
1181 //
1182 if (_callprojs->resproj[0] != nullptr && _callprojs->resproj[0]->outcnt() != 0) {
1183 // First disconnect stores captured by Initialize node.
1184 // If Initialize node is eliminated first in the following code,
1185 // it will kill such stores and DUIterator_Last will assert.
1186 for (DUIterator_Fast jmax, j = _callprojs->resproj[0]->fast_outs(jmax); j < jmax; j++) {
1187 Node* use = _callprojs->resproj[0]->fast_out(j);
1188 if (use->is_AddP()) {
1189 // raw memory addresses used only by the initialization
1190 _igvn.replace_node(use, C->top());
1191 --j; --jmax;
1192 }
1193 }
1194 for (DUIterator_Last jmin, j = _callprojs->resproj[0]->last_outs(jmin); j >= jmin; ) {
1195 Node* use = _callprojs->resproj[0]->last_out(j);
1196 uint oc1 = _callprojs->resproj[0]->outcnt();
1197 if (use->is_Initialize()) {
1198 // Eliminate Initialize node.
1199 InitializeNode *init = use->as_Initialize();
1200 assert(init->outcnt() <= 2, "only a control and memory projection expected");
1201 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
1202 if (ctrl_proj != nullptr) {
1203 _igvn.replace_node(ctrl_proj, init->in(TypeFunc::Control));
1204 #ifdef ASSERT
1205 // If the InitializeNode has no memory out, it will die, and tmp will become null
1206 Node* tmp = init->in(TypeFunc::Control);
1207 assert(tmp == nullptr || tmp == _callprojs->fallthrough_catchproj, "allocation control projection");
1208 #endif
1209 }
1210 Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory);
1211 if (mem_proj != nullptr) {
1212 Node *mem = init->in(TypeFunc::Memory);
1213 #ifdef ASSERT
1214 if (mem->is_MergeMem()) {
1215 assert(mem->in(TypeFunc::Memory) == _callprojs->fallthrough_memproj, "allocation memory projection");
1216 } else {
1217 assert(mem == _callprojs->fallthrough_memproj, "allocation memory projection");
1218 }
1219 #endif
1220 _igvn.replace_node(mem_proj, mem);
1221 }
1222 } else if (use->Opcode() == Op_MemBarStoreStore) {
1223 // Inline type buffer allocations are followed by a membar
1224 assert(inline_alloc, "Unexpected MemBarStoreStore");
1225 use->as_MemBar()->remove(&_igvn);
1226 } else {
1227 assert(false, "only Initialize or AddP expected");
1228 }
1229 j -= (oc1 - _callprojs->resproj[0]->outcnt());
1230 }
1231 }
1232 if (_callprojs->fallthrough_catchproj != nullptr) {
1233 _igvn.replace_node(_callprojs->fallthrough_catchproj, alloc->in(TypeFunc::Control));
1234 }
1235 if (_callprojs->fallthrough_memproj != nullptr) {
1236 _igvn.replace_node(_callprojs->fallthrough_memproj, alloc->in(TypeFunc::Memory));
1237 }
1238 if (_callprojs->catchall_memproj != nullptr) {
1239 _igvn.replace_node(_callprojs->catchall_memproj, C->top());
1240 }
1241 if (_callprojs->fallthrough_ioproj != nullptr) {
1242 _igvn.replace_node(_callprojs->fallthrough_ioproj, alloc->in(TypeFunc::I_O));
1243 }
1244 if (_callprojs->catchall_ioproj != nullptr) {
1245 _igvn.replace_node(_callprojs->catchall_ioproj, C->top());
1246 }
1247 if (_callprojs->catchall_catchproj != nullptr) {
1248 _igvn.replace_node(_callprojs->catchall_catchproj, C->top());
1249 }
1250 }
1251
1252 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1253 // If reallocation fails during deoptimization we'll pop all
1254 // interpreter frames for this compiled frame and that won't play
1255 // nice with JVMTI popframe.
1256 // We avoid this issue by eager reallocation when the popframe request
1257 // is received.
1258 if (!EliminateAllocations) {
1259 return false;
1260 }
1261 Node* klass = alloc->in(AllocateNode::KlassNode);
1262 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1263
1264 // Attempt to eliminate inline type buffer allocations
1265 // regardless of usage and escape/replaceable status.
1266 bool inline_alloc = tklass->isa_instklassptr() &&
1267 tklass->is_instklassptr()->instance_klass()->is_inlinetype();
1268 if (!alloc->_is_non_escaping && !inline_alloc) {
1269 return false;
1270 }
1271 // Eliminate boxing allocations which are not used
1272 // regardless scalar replaceable status.
1273 Node* res = alloc->result_cast();
1274 bool boxing_alloc = (res == nullptr) && C->eliminate_boxing() &&
1275 tklass->isa_instklassptr() &&
1276 tklass->is_instklassptr()->instance_klass()->is_box_klass();
1277 if (!alloc->_is_scalar_replaceable && !boxing_alloc && !inline_alloc) {
1278 return false;
1279 }
1280
1281 _callprojs = alloc->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1282
1283 GrowableArray <SafePointNode *> safepoints;
1284 if (!can_eliminate_allocation(&_igvn, alloc, &safepoints)) {
1285 return false;
1286 }
1287
1288 if (!alloc->_is_scalar_replaceable) {
1289 assert(res == nullptr || inline_alloc, "sanity");
1290 // We can only eliminate allocation if all debug info references
1291 // are already replaced with SafePointScalarObject because
1292 // we can't search for a fields value without instance_id.
1293 if (safepoints.length() > 0) {
1294 assert(!inline_alloc || C->has_circular_inline_type(),
1295 "Inline type allocations should have been scalarized earlier");
1296 return false;
1297 }
1298 }
1299
1300 if (!scalar_replacement(alloc, safepoints)) {
1301 return false;
1302 }
1303
1304 CompileLog* log = C->log();
1305 if (log != nullptr) {
1306 log->head("eliminate_allocation type='%d'",
1307 log->identify(tklass->exact_klass()));
1308 JVMState* p = alloc->jvms();
1309 while (p != nullptr) {
1310 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1311 p = p->caller();
1312 }
1313 log->tail("eliminate_allocation");
1314 }
1315
1316 process_users_of_allocation(alloc, inline_alloc);
1317
1318 #ifndef PRODUCT
1319 if (PrintEliminateAllocations) {
1320 if (alloc->is_AllocateArray())
1321 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1322 else
1323 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1324 }
1325 #endif
1326
1327 return true;
1328 }
1329
1330 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1331 // EA should remove all uses of non-escaping boxing node.
1332 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != nullptr) {
1333 return false;
1334 }
1335
1336 assert(boxing->result_cast() == nullptr, "unexpected boxing node result");
1337
1338 _callprojs = boxing->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1339
1340 const TypeTuple* r = boxing->tf()->range_sig();
1341 assert(r->cnt() > TypeFunc::Parms, "sanity");
1342 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1343 assert(t != nullptr, "sanity");
1344
1345 CompileLog* log = C->log();
1346 if (log != nullptr) {
1347 log->head("eliminate_boxing type='%d'",
1348 log->identify(t->instance_klass()));
1349 JVMState* p = boxing->jvms();
1350 while (p != nullptr) {
1351 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1352 p = p->caller();
1353 }
1354 log->tail("eliminate_boxing");
1355 }
1356
1357 process_users_of_allocation(boxing);
1358
1359 #ifndef PRODUCT
1360 if (PrintEliminateAllocations) {
1361 tty->print("++++ Eliminated: %d ", boxing->_idx);
1362 boxing->method()->print_short_name(tty);
1363 tty->cr();
1364 }
1365 #endif
1366
1367 return true;
1368 }
1369
1370
1371 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1372 Node* adr = basic_plus_adr(base, offset);
1373 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1374 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered);
1375 transform_later(value);
1376 return value;
1377 }
1378
1379
1380 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1381 Node* adr = basic_plus_adr(base, offset);
1382 mem = StoreNode::make(_igvn, ctl, mem, adr, nullptr, value, bt, MemNode::unordered);
1383 transform_later(mem);
1384 return mem;
1385 }
1386
1387 //=============================================================================
1388 //
1389 // A L L O C A T I O N
1390 //
1391 // Allocation attempts to be fast in the case of frequent small objects.
1392 // It breaks down like this:
1393 //
1394 // 1) Size in doublewords is computed. This is a constant for objects and
1395 // variable for most arrays. Doubleword units are used to avoid size
1396 // overflow of huge doubleword arrays. We need doublewords in the end for
1397 // rounding.
1398 //
1399 // 2) Size is checked for being 'too large'. Too-large allocations will go
1400 // the slow path into the VM. The slow path can throw any required
1401 // exceptions, and does all the special checks for very large arrays. The
1402 // size test can constant-fold away for objects. For objects with
1403 // finalizers it constant-folds the otherway: you always go slow with
1404 // finalizers.
1405 //
1406 // 3) If NOT using TLABs, this is the contended loop-back point.
1407 // Load-Locked the heap top. If using TLABs normal-load the heap top.
1408 //
1409 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
1410 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
1411 // "size*8" we always enter the VM, where "largish" is a constant picked small
1412 // enough that there's always space between the eden max and 4Gig (old space is
1413 // there so it's quite large) and large enough that the cost of entering the VM
1414 // is dwarfed by the cost to initialize the space.
1415 //
1416 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1417 // down. If contended, repeat at step 3. If using TLABs normal-store
1418 // adjusted heap top back down; there is no contention.
1419 //
1420 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
1421 // fields.
1422 //
1423 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1424 // oop flavor.
1425 //
1426 //=============================================================================
1427 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1428 // Allocations bigger than this always go the slow route.
1429 // This value must be small enough that allocation attempts that need to
1430 // trigger exceptions go the slow route. Also, it must be small enough so
1431 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1432 //=============================================================================j//
1433 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1434 // The allocator will coalesce int->oop copies away. See comment in
1435 // coalesce.cpp about how this works. It depends critically on the exact
1436 // code shape produced here, so if you are changing this code shape
1437 // make sure the GC info for the heap-top is correct in and around the
1438 // slow-path call.
1439 //
1440
1441 void PhaseMacroExpand::expand_allocate_common(
1442 AllocateNode* alloc, // allocation node to be expanded
1443 Node* length, // array length for an array allocation
1444 Node* init_val, // value to initialize the array with
1445 const TypeFunc* slow_call_type, // Type of slow call
1446 address slow_call_address, // Address of slow call
1447 Node* valid_length_test // whether length is valid or not
1448 )
1449 {
1450 Node* ctrl = alloc->in(TypeFunc::Control);
1451 Node* mem = alloc->in(TypeFunc::Memory);
1452 Node* i_o = alloc->in(TypeFunc::I_O);
1453 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1454 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1455 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1456 assert(ctrl != nullptr, "must have control");
1457
1458 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1459 // they will not be used if "always_slow" is set
1460 enum { slow_result_path = 1, fast_result_path = 2 };
1461 Node *result_region = nullptr;
1462 Node *result_phi_rawmem = nullptr;
1463 Node *result_phi_rawoop = nullptr;
1464 Node *result_phi_i_o = nullptr;
1465
1466 // The initial slow comparison is a size check, the comparison
1467 // we want to do is a BoolTest::gt
1468 bool expand_fast_path = true;
1469 int tv = _igvn.find_int_con(initial_slow_test, -1);
1470 if (tv >= 0) {
1471 // InitialTest has constant result
1472 // 0 - can fit in TLAB
1473 // 1 - always too big or negative
1474 assert(tv <= 1, "0 or 1 if a constant");
1475 expand_fast_path = (tv == 0);
1476 initial_slow_test = nullptr;
1477 } else {
1478 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1479 }
1480
1481 if (!UseTLAB) {
1482 // Force slow-path allocation
1483 expand_fast_path = false;
1484 initial_slow_test = nullptr;
1485 }
1486
1487 bool allocation_has_use = (alloc->result_cast() != nullptr);
1488 if (!allocation_has_use) {
1489 InitializeNode* init = alloc->initialization();
1490 if (init != nullptr) {
1491 init->remove(&_igvn);
1492 }
1493 if (expand_fast_path && (initial_slow_test == nullptr)) {
1494 // Remove allocation node and return.
1495 // Size is a non-negative constant -> no initial check needed -> directly to fast path.
1496 // Also, no usages -> empty fast path -> no fall out to slow path -> nothing left.
1497 #ifndef PRODUCT
1498 if (PrintEliminateAllocations) {
1499 tty->print("NotUsed ");
1500 Node* res = alloc->proj_out_or_null(TypeFunc::Parms);
1501 if (res != nullptr) {
1502 res->dump();
1503 } else {
1504 alloc->dump();
1505 }
1506 }
1507 #endif
1508 yank_alloc_node(alloc);
1509 return;
1510 }
1511 }
1512
1513 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1514 Node *slow_region = nullptr;
1515 Node *toobig_false = ctrl;
1516
1517 // generate the initial test if necessary
1518 if (initial_slow_test != nullptr ) {
1519 assert (expand_fast_path, "Only need test if there is a fast path");
1520 slow_region = new RegionNode(3);
1521
1522 // Now make the initial failure test. Usually a too-big test but
1523 // might be a TRUE for finalizers.
1524 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1525 transform_later(toobig_iff);
1526 // Plug the failing-too-big test into the slow-path region
1527 Node* toobig_true = new IfTrueNode(toobig_iff);
1528 transform_later(toobig_true);
1529 slow_region ->init_req( too_big_or_final_path, toobig_true );
1530 toobig_false = new IfFalseNode(toobig_iff);
1531 transform_later(toobig_false);
1532 } else {
1533 // No initial test, just fall into next case
1534 assert(allocation_has_use || !expand_fast_path, "Should already have been handled");
1535 toobig_false = ctrl;
1536 debug_only(slow_region = NodeSentinel);
1537 }
1538
1539 // If we are here there are several possibilities
1540 // - expand_fast_path is false - then only a slow path is expanded. That's it.
1541 // no_initial_check means a constant allocation.
1542 // - If check always evaluates to false -> expand_fast_path is false (see above)
1543 // - If check always evaluates to true -> directly into fast path (but may bailout to slowpath)
1544 // if !allocation_has_use the fast path is empty
1545 // if !allocation_has_use && no_initial_check
1546 // - Then there are no fastpath that can fall out to slowpath -> no allocation code at all.
1547 // removed by yank_alloc_node above.
1548
1549 Node *slow_mem = mem; // save the current memory state for slow path
1550 // generate the fast allocation code unless we know that the initial test will always go slow
1551 if (expand_fast_path) {
1552 // Fast path modifies only raw memory.
1553 if (mem->is_MergeMem()) {
1554 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1555 }
1556
1557 // allocate the Region and Phi nodes for the result
1558 result_region = new RegionNode(3);
1559 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1560 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1561
1562 // Grab regular I/O before optional prefetch may change it.
1563 // Slow-path does no I/O so just set it to the original I/O.
1564 result_phi_i_o->init_req(slow_result_path, i_o);
1565
1566 // Name successful fast-path variables
1567 Node* fast_oop_ctrl;
1568 Node* fast_oop_rawmem;
1569
1570 if (allocation_has_use) {
1571 Node* needgc_ctrl = nullptr;
1572 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1573
1574 intx prefetch_lines = length != nullptr ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1575 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1576 Node* fast_oop = bs->obj_allocate(this, mem, toobig_false, size_in_bytes, i_o, needgc_ctrl,
1577 fast_oop_ctrl, fast_oop_rawmem,
1578 prefetch_lines);
1579
1580 if (initial_slow_test != nullptr) {
1581 // This completes all paths into the slow merge point
1582 slow_region->init_req(need_gc_path, needgc_ctrl);
1583 transform_later(slow_region);
1584 } else {
1585 // No initial slow path needed!
1586 // Just fall from the need-GC path straight into the VM call.
1587 slow_region = needgc_ctrl;
1588 }
1589
1590 InitializeNode* init = alloc->initialization();
1591 fast_oop_rawmem = initialize_object(alloc,
1592 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1593 klass_node, length, size_in_bytes);
1594 expand_initialize_membar(alloc, init, fast_oop_ctrl, fast_oop_rawmem);
1595 expand_dtrace_alloc_probe(alloc, fast_oop, fast_oop_ctrl, fast_oop_rawmem);
1596
1597 result_phi_rawoop->init_req(fast_result_path, fast_oop);
1598 } else {
1599 assert (initial_slow_test != nullptr, "sanity");
1600 fast_oop_ctrl = toobig_false;
1601 fast_oop_rawmem = mem;
1602 transform_later(slow_region);
1603 }
1604
1605 // Plug in the successful fast-path into the result merge point
1606 result_region ->init_req(fast_result_path, fast_oop_ctrl);
1607 result_phi_i_o ->init_req(fast_result_path, i_o);
1608 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1609 } else {
1610 slow_region = ctrl;
1611 result_phi_i_o = i_o; // Rename it to use in the following code.
1612 }
1613
1614 // Generate slow-path call
1615 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1616 OptoRuntime::stub_name(slow_call_address),
1617 TypePtr::BOTTOM);
1618 call->init_req(TypeFunc::Control, slow_region);
1619 call->init_req(TypeFunc::I_O, top()); // does no i/o
1620 call->init_req(TypeFunc::Memory, slow_mem); // may gc ptrs
1621 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1622 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1623
1624 call->init_req(TypeFunc::Parms+0, klass_node);
1625 if (length != nullptr) {
1626 call->init_req(TypeFunc::Parms+1, length);
1627 if (init_val != nullptr) {
1628 call->init_req(TypeFunc::Parms+2, init_val);
1629 }
1630 } else {
1631 // Let the runtime know if this is a larval allocation
1632 call->init_req(TypeFunc::Parms+1, _igvn.intcon(alloc->_larval));
1633 }
1634
1635 // Copy debug information and adjust JVMState information, then replace
1636 // allocate node with the call
1637 call->copy_call_debug_info(&_igvn, alloc);
1638 // For array allocations, copy the valid length check to the call node so Compile::final_graph_reshaping() can verify
1639 // that the call has the expected number of CatchProj nodes (in case the allocation always fails and the fallthrough
1640 // path dies).
1641 if (valid_length_test != nullptr) {
1642 call->add_req(valid_length_test);
1643 }
1644 if (expand_fast_path) {
1645 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1646 } else {
1647 // Hook i_o projection to avoid its elimination during allocation
1648 // replacement (when only a slow call is generated).
1649 call->set_req(TypeFunc::I_O, result_phi_i_o);
1650 }
1651 _igvn.replace_node(alloc, call);
1652 transform_later(call);
1653
1654 // Identify the output projections from the allocate node and
1655 // adjust any references to them.
1656 // The control and io projections look like:
1657 //
1658 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1659 // Allocate Catch
1660 // ^---Proj(io) <-------+ ^---CatchProj(io)
1661 //
1662 // We are interested in the CatchProj nodes.
1663 //
1664 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1665
1666 // An allocate node has separate memory projections for the uses on
1667 // the control and i_o paths. Replace the control memory projection with
1668 // result_phi_rawmem (unless we are only generating a slow call when
1669 // both memory projections are combined)
1670 if (expand_fast_path && _callprojs->fallthrough_memproj != nullptr) {
1671 _igvn.replace_in_uses(_callprojs->fallthrough_memproj, result_phi_rawmem);
1672 }
1673 // Now change uses of catchall_memproj to use fallthrough_memproj and delete
1674 // catchall_memproj so we end up with a call that has only 1 memory projection.
1675 if (_callprojs->catchall_memproj != nullptr) {
1676 if (_callprojs->fallthrough_memproj == nullptr) {
1677 _callprojs->fallthrough_memproj = new ProjNode(call, TypeFunc::Memory);
1678 transform_later(_callprojs->fallthrough_memproj);
1679 }
1680 _igvn.replace_in_uses(_callprojs->catchall_memproj, _callprojs->fallthrough_memproj);
1681 _igvn.remove_dead_node(_callprojs->catchall_memproj);
1682 }
1683
1684 // An allocate node has separate i_o projections for the uses on the control
1685 // and i_o paths. Always replace the control i_o projection with result i_o
1686 // otherwise incoming i_o become dead when only a slow call is generated
1687 // (it is different from memory projections where both projections are
1688 // combined in such case).
1689 if (_callprojs->fallthrough_ioproj != nullptr) {
1690 _igvn.replace_in_uses(_callprojs->fallthrough_ioproj, result_phi_i_o);
1691 }
1692 // Now change uses of catchall_ioproj to use fallthrough_ioproj and delete
1693 // catchall_ioproj so we end up with a call that has only 1 i_o projection.
1694 if (_callprojs->catchall_ioproj != nullptr) {
1695 if (_callprojs->fallthrough_ioproj == nullptr) {
1696 _callprojs->fallthrough_ioproj = new ProjNode(call, TypeFunc::I_O);
1697 transform_later(_callprojs->fallthrough_ioproj);
1698 }
1699 _igvn.replace_in_uses(_callprojs->catchall_ioproj, _callprojs->fallthrough_ioproj);
1700 _igvn.remove_dead_node(_callprojs->catchall_ioproj);
1701 }
1702
1703 // if we generated only a slow call, we are done
1704 if (!expand_fast_path) {
1705 // Now we can unhook i_o.
1706 if (result_phi_i_o->outcnt() > 1) {
1707 call->set_req(TypeFunc::I_O, top());
1708 } else {
1709 assert(result_phi_i_o->unique_ctrl_out() == call, "sanity");
1710 // Case of new array with negative size known during compilation.
1711 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1712 // following code since call to runtime will throw exception.
1713 // As result there will be no users of i_o after the call.
1714 // Leave i_o attached to this call to avoid problems in preceding graph.
1715 }
1716 return;
1717 }
1718
1719 if (_callprojs->fallthrough_catchproj != nullptr) {
1720 ctrl = _callprojs->fallthrough_catchproj->clone();
1721 transform_later(ctrl);
1722 _igvn.replace_node(_callprojs->fallthrough_catchproj, result_region);
1723 } else {
1724 ctrl = top();
1725 }
1726 Node *slow_result;
1727 if (_callprojs->resproj[0] == nullptr) {
1728 // no uses of the allocation result
1729 slow_result = top();
1730 } else {
1731 slow_result = _callprojs->resproj[0]->clone();
1732 transform_later(slow_result);
1733 _igvn.replace_node(_callprojs->resproj[0], result_phi_rawoop);
1734 }
1735
1736 // Plug slow-path into result merge point
1737 result_region->init_req( slow_result_path, ctrl);
1738 transform_later(result_region);
1739 if (allocation_has_use) {
1740 result_phi_rawoop->init_req(slow_result_path, slow_result);
1741 transform_later(result_phi_rawoop);
1742 }
1743 result_phi_rawmem->init_req(slow_result_path, _callprojs->fallthrough_memproj);
1744 transform_later(result_phi_rawmem);
1745 transform_later(result_phi_i_o);
1746 // This completes all paths into the result merge point
1747 }
1748
1749 // Remove alloc node that has no uses.
1750 void PhaseMacroExpand::yank_alloc_node(AllocateNode* alloc) {
1751 Node* ctrl = alloc->in(TypeFunc::Control);
1752 Node* mem = alloc->in(TypeFunc::Memory);
1753 Node* i_o = alloc->in(TypeFunc::I_O);
1754
1755 _callprojs = alloc->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1756 if (_callprojs->resproj[0] != nullptr) {
1757 for (DUIterator_Fast imax, i = _callprojs->resproj[0]->fast_outs(imax); i < imax; i++) {
1758 Node* use = _callprojs->resproj[0]->fast_out(i);
1759 use->isa_MemBar()->remove(&_igvn);
1760 --imax;
1761 --i; // back up iterator
1762 }
1763 assert(_callprojs->resproj[0]->outcnt() == 0, "all uses must be deleted");
1764 _igvn.remove_dead_node(_callprojs->resproj[0]);
1765 }
1766 if (_callprojs->fallthrough_catchproj != nullptr) {
1767 _igvn.replace_in_uses(_callprojs->fallthrough_catchproj, ctrl);
1768 _igvn.remove_dead_node(_callprojs->fallthrough_catchproj);
1769 }
1770 if (_callprojs->catchall_catchproj != nullptr) {
1771 _igvn.rehash_node_delayed(_callprojs->catchall_catchproj);
1772 _callprojs->catchall_catchproj->set_req(0, top());
1773 }
1774 if (_callprojs->fallthrough_proj != nullptr) {
1775 Node* catchnode = _callprojs->fallthrough_proj->unique_ctrl_out();
1776 _igvn.remove_dead_node(catchnode);
1777 _igvn.remove_dead_node(_callprojs->fallthrough_proj);
1778 }
1779 if (_callprojs->fallthrough_memproj != nullptr) {
1780 _igvn.replace_in_uses(_callprojs->fallthrough_memproj, mem);
1781 _igvn.remove_dead_node(_callprojs->fallthrough_memproj);
1782 }
1783 if (_callprojs->fallthrough_ioproj != nullptr) {
1784 _igvn.replace_in_uses(_callprojs->fallthrough_ioproj, i_o);
1785 _igvn.remove_dead_node(_callprojs->fallthrough_ioproj);
1786 }
1787 if (_callprojs->catchall_memproj != nullptr) {
1788 _igvn.rehash_node_delayed(_callprojs->catchall_memproj);
1789 _callprojs->catchall_memproj->set_req(0, top());
1790 }
1791 if (_callprojs->catchall_ioproj != nullptr) {
1792 _igvn.rehash_node_delayed(_callprojs->catchall_ioproj);
1793 _callprojs->catchall_ioproj->set_req(0, top());
1794 }
1795 #ifndef PRODUCT
1796 if (PrintEliminateAllocations) {
1797 if (alloc->is_AllocateArray()) {
1798 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1799 } else {
1800 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1801 }
1802 }
1803 #endif
1804 _igvn.remove_dead_node(alloc);
1805 }
1806
1807 void PhaseMacroExpand::expand_initialize_membar(AllocateNode* alloc, InitializeNode* init,
1808 Node*& fast_oop_ctrl, Node*& fast_oop_rawmem) {
1809 // If initialization is performed by an array copy, any required
1810 // MemBarStoreStore was already added. If the object does not
1811 // escape no need for a MemBarStoreStore. If the object does not
1812 // escape in its initializer and memory barrier (MemBarStoreStore or
1813 // stronger) is already added at exit of initializer, also no need
1814 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore
1815 // so that stores that initialize this object can't be reordered
1816 // with a subsequent store that makes this object accessible by
1817 // other threads.
1818 // Other threads include java threads and JVM internal threads
1819 // (for example concurrent GC threads). Current concurrent GC
1820 // implementation: G1 will not scan newly created object,
1821 // so it's safe to skip storestore barrier when allocation does
1822 // not escape.
1823 if (!alloc->does_not_escape_thread() &&
1824 !alloc->is_allocation_MemBar_redundant() &&
1825 (init == nullptr || !init->is_complete_with_arraycopy())) {
1826 if (init == nullptr || init->req() < InitializeNode::RawStores) {
1827 // No InitializeNode or no stores captured by zeroing
1828 // elimination. Simply add the MemBarStoreStore after object
1829 // initialization.
1830 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1831 transform_later(mb);
1832
1833 mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
1834 mb->init_req(TypeFunc::Control, fast_oop_ctrl);
1835 fast_oop_ctrl = new ProjNode(mb, TypeFunc::Control);
1836 transform_later(fast_oop_ctrl);
1837 fast_oop_rawmem = new ProjNode(mb, TypeFunc::Memory);
1838 transform_later(fast_oop_rawmem);
1839 } else {
1840 // Add the MemBarStoreStore after the InitializeNode so that
1841 // all stores performing the initialization that were moved
1842 // before the InitializeNode happen before the storestore
1843 // barrier.
1844
1845 Node* init_ctrl = init->proj_out_or_null(TypeFunc::Control);
1846 Node* init_mem = init->proj_out_or_null(TypeFunc::Memory);
1847
1848 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1849 transform_later(mb);
1850
1851 Node* ctrl = new ProjNode(init, TypeFunc::Control);
1852 transform_later(ctrl);
1853 Node* mem = new ProjNode(init, TypeFunc::Memory);
1854 transform_later(mem);
1855
1856 // The MemBarStoreStore depends on control and memory coming
1857 // from the InitializeNode
1858 mb->init_req(TypeFunc::Memory, mem);
1859 mb->init_req(TypeFunc::Control, ctrl);
1860
1861 ctrl = new ProjNode(mb, TypeFunc::Control);
1862 transform_later(ctrl);
1863 mem = new ProjNode(mb, TypeFunc::Memory);
1864 transform_later(mem);
1865
1866 // All nodes that depended on the InitializeNode for control
1867 // and memory must now depend on the MemBarNode that itself
1868 // depends on the InitializeNode
1869 if (init_ctrl != nullptr) {
1870 _igvn.replace_node(init_ctrl, ctrl);
1871 }
1872 if (init_mem != nullptr) {
1873 _igvn.replace_node(init_mem, mem);
1874 }
1875 }
1876 }
1877 }
1878
1879 void PhaseMacroExpand::expand_dtrace_alloc_probe(AllocateNode* alloc, Node* oop,
1880 Node*& ctrl, Node*& rawmem) {
1881 if (C->env()->dtrace_alloc_probes()) {
1882 // Slow-path call
1883 int size = TypeFunc::Parms + 2;
1884 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1885 CAST_FROM_FN_PTR(address,
1886 static_cast<int (*)(JavaThread*, oopDesc*)>(SharedRuntime::dtrace_object_alloc)),
1887 "dtrace_object_alloc",
1888 TypeRawPtr::BOTTOM);
1889
1890 // Get base of thread-local storage area
1891 Node* thread = new ThreadLocalNode();
1892 transform_later(thread);
1893
1894 call->init_req(TypeFunc::Parms + 0, thread);
1895 call->init_req(TypeFunc::Parms + 1, oop);
1896 call->init_req(TypeFunc::Control, ctrl);
1897 call->init_req(TypeFunc::I_O , top()); // does no i/o
1898 call->init_req(TypeFunc::Memory , rawmem);
1899 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1900 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1901 transform_later(call);
1902 ctrl = new ProjNode(call, TypeFunc::Control);
1903 transform_later(ctrl);
1904 rawmem = new ProjNode(call, TypeFunc::Memory);
1905 transform_later(rawmem);
1906 }
1907 }
1908
1909 // Helper for PhaseMacroExpand::expand_allocate_common.
1910 // Initializes the newly-allocated storage.
1911 Node* PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1912 Node* control, Node* rawmem, Node* object,
1913 Node* klass_node, Node* length,
1914 Node* size_in_bytes) {
1915 InitializeNode* init = alloc->initialization();
1916 // Store the klass & mark bits
1917 Node* mark_node = alloc->make_ideal_mark(&_igvn, control, rawmem);
1918 if (!mark_node->is_Con()) {
1919 transform_later(mark_node);
1920 }
1921 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, TypeX_X->basic_type());
1922
1923 if (!UseCompactObjectHeaders) {
1924 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1925 }
1926 int header_size = alloc->minimum_header_size(); // conservatively small
1927
1928 // Array length
1929 if (length != nullptr) { // Arrays need length field
1930 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1931 // conservatively small header size:
1932 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1933 if (_igvn.type(klass_node)->isa_aryklassptr()) { // we know the exact header size in most cases:
1934 BasicType elem = _igvn.type(klass_node)->is_klassptr()->as_instance_type()->isa_aryptr()->elem()->array_element_basic_type();
1935 if (is_reference_type(elem, true)) {
1936 elem = T_OBJECT;
1937 }
1938 header_size = Klass::layout_helper_header_size(Klass::array_layout_helper(elem));
1939 }
1940 }
1941
1942 // Clear the object body, if necessary.
1943 if (init == nullptr) {
1944 // The init has somehow disappeared; be cautious and clear everything.
1945 //
1946 // This can happen if a node is allocated but an uncommon trap occurs
1947 // immediately. In this case, the Initialize gets associated with the
1948 // trap, and may be placed in a different (outer) loop, if the Allocate
1949 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1950 // there can be two Allocates to one Initialize. The answer in all these
1951 // edge cases is safety first. It is always safe to clear immediately
1952 // within an Allocate, and then (maybe or maybe not) clear some more later.
1953 if (!(UseTLAB && ZeroTLAB)) {
1954 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1955 alloc->in(AllocateNode::InitValue),
1956 alloc->in(AllocateNode::RawInitValue),
1957 header_size, size_in_bytes,
1958 &_igvn);
1959 }
1960 } else {
1961 if (!init->is_complete()) {
1962 // Try to win by zeroing only what the init does not store.
1963 // We can also try to do some peephole optimizations,
1964 // such as combining some adjacent subword stores.
1965 rawmem = init->complete_stores(control, rawmem, object,
1966 header_size, size_in_bytes, &_igvn);
1967 }
1968 // We have no more use for this link, since the AllocateNode goes away:
1969 init->set_req(InitializeNode::RawAddress, top());
1970 // (If we keep the link, it just confuses the register allocator,
1971 // who thinks he sees a real use of the address by the membar.)
1972 }
1973
1974 return rawmem;
1975 }
1976
1977 // Generate prefetch instructions for next allocations.
1978 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1979 Node*& contended_phi_rawmem,
1980 Node* old_eden_top, Node* new_eden_top,
1981 intx lines) {
1982 enum { fall_in_path = 1, pf_path = 2 };
1983 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1984 // Generate prefetch allocation with watermark check.
1985 // As an allocation hits the watermark, we will prefetch starting
1986 // at a "distance" away from watermark.
1987
1988 Node *pf_region = new RegionNode(3);
1989 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY,
1990 TypeRawPtr::BOTTOM );
1991 // I/O is used for Prefetch
1992 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO );
1993
1994 Node *thread = new ThreadLocalNode();
1995 transform_later(thread);
1996
1997 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread,
1998 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1999 transform_later(eden_pf_adr);
2000
2001 Node *old_pf_wm = new LoadPNode(needgc_false,
2002 contended_phi_rawmem, eden_pf_adr,
2003 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM,
2004 MemNode::unordered);
2005 transform_later(old_pf_wm);
2006
2007 // check against new_eden_top
2008 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm );
2009 transform_later(need_pf_cmp);
2010 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge );
2011 transform_later(need_pf_bol);
2012 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol,
2013 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
2014 transform_later(need_pf_iff);
2015
2016 // true node, add prefetchdistance
2017 Node *need_pf_true = new IfTrueNode( need_pf_iff );
2018 transform_later(need_pf_true);
2019
2020 Node *need_pf_false = new IfFalseNode( need_pf_iff );
2021 transform_later(need_pf_false);
2022
2023 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm,
2024 _igvn.MakeConX(AllocatePrefetchDistance) );
2025 transform_later(new_pf_wmt );
2026 new_pf_wmt->set_req(0, need_pf_true);
2027
2028 Node *store_new_wmt = new StorePNode(need_pf_true,
2029 contended_phi_rawmem, eden_pf_adr,
2030 TypeRawPtr::BOTTOM, new_pf_wmt,
2031 MemNode::unordered);
2032 transform_later(store_new_wmt);
2033
2034 // adding prefetches
2035 pf_phi_abio->init_req( fall_in_path, i_o );
2036
2037 Node *prefetch_adr;
2038 Node *prefetch;
2039 uint step_size = AllocatePrefetchStepSize;
2040 uint distance = 0;
2041
2042 for ( intx i = 0; i < lines; i++ ) {
2043 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt,
2044 _igvn.MakeConX(distance) );
2045 transform_later(prefetch_adr);
2046 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
2047 transform_later(prefetch);
2048 distance += step_size;
2049 i_o = prefetch;
2050 }
2051 pf_phi_abio->set_req( pf_path, i_o );
2052
2053 pf_region->init_req( fall_in_path, need_pf_false );
2054 pf_region->init_req( pf_path, need_pf_true );
2055
2056 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
2057 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
2058
2059 transform_later(pf_region);
2060 transform_later(pf_phi_rawmem);
2061 transform_later(pf_phi_abio);
2062
2063 needgc_false = pf_region;
2064 contended_phi_rawmem = pf_phi_rawmem;
2065 i_o = pf_phi_abio;
2066 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) {
2067 // Insert a prefetch instruction for each allocation.
2068 // This code is used to generate 1 prefetch instruction per cache line.
2069
2070 // Generate several prefetch instructions.
2071 uint step_size = AllocatePrefetchStepSize;
2072 uint distance = AllocatePrefetchDistance;
2073
2074 // Next cache address.
2075 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top,
2076 _igvn.MakeConX(step_size + distance));
2077 transform_later(cache_adr);
2078 cache_adr = new CastP2XNode(needgc_false, cache_adr);
2079 transform_later(cache_adr);
2080 // Address is aligned to execute prefetch to the beginning of cache line size
2081 // (it is important when BIS instruction is used on SPARC as prefetch).
2082 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
2083 cache_adr = new AndXNode(cache_adr, mask);
2084 transform_later(cache_adr);
2085 cache_adr = new CastX2PNode(cache_adr);
2086 transform_later(cache_adr);
2087
2088 // Prefetch
2089 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr );
2090 prefetch->set_req(0, needgc_false);
2091 transform_later(prefetch);
2092 contended_phi_rawmem = prefetch;
2093 Node *prefetch_adr;
2094 distance = step_size;
2095 for ( intx i = 1; i < lines; i++ ) {
2096 prefetch_adr = new AddPNode( cache_adr, cache_adr,
2097 _igvn.MakeConX(distance) );
2098 transform_later(prefetch_adr);
2099 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr );
2100 transform_later(prefetch);
2101 distance += step_size;
2102 contended_phi_rawmem = prefetch;
2103 }
2104 } else if( AllocatePrefetchStyle > 0 ) {
2105 // Insert a prefetch for each allocation only on the fast-path
2106 Node *prefetch_adr;
2107 Node *prefetch;
2108 // Generate several prefetch instructions.
2109 uint step_size = AllocatePrefetchStepSize;
2110 uint distance = AllocatePrefetchDistance;
2111 for ( intx i = 0; i < lines; i++ ) {
2112 prefetch_adr = new AddPNode( old_eden_top, new_eden_top,
2113 _igvn.MakeConX(distance) );
2114 transform_later(prefetch_adr);
2115 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
2116 // Do not let it float too high, since if eden_top == eden_end,
2117 // both might be null.
2118 if( i == 0 ) { // Set control for first prefetch, next follows it
2119 prefetch->init_req(0, needgc_false);
2120 }
2121 transform_later(prefetch);
2122 distance += step_size;
2123 i_o = prefetch;
2124 }
2125 }
2126 return i_o;
2127 }
2128
2129
2130 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
2131 expand_allocate_common(alloc, nullptr, nullptr,
2132 OptoRuntime::new_instance_Type(),
2133 OptoRuntime::new_instance_Java(), nullptr);
2134 }
2135
2136 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
2137 Node* length = alloc->in(AllocateNode::ALength);
2138 Node* valid_length_test = alloc->in(AllocateNode::ValidLengthTest);
2139 InitializeNode* init = alloc->initialization();
2140 Node* klass_node = alloc->in(AllocateNode::KlassNode);
2141 Node* init_value = alloc->in(AllocateNode::InitValue);
2142 const TypeAryKlassPtr* ary_klass_t = _igvn.type(klass_node)->isa_aryklassptr();
2143 const TypeFunc* slow_call_type;
2144 address slow_call_address; // Address of slow call
2145 if (init != nullptr && init->is_complete_with_arraycopy() &&
2146 ary_klass_t && ary_klass_t->elem()->isa_klassptr() == nullptr) {
2147 // Don't zero type array during slow allocation in VM since
2148 // it will be initialized later by arraycopy in compiled code.
2149 slow_call_address = OptoRuntime::new_array_nozero_Java();
2150 slow_call_type = OptoRuntime::new_array_nozero_Type();
2151 } else {
2152 slow_call_address = OptoRuntime::new_array_Java();
2153 slow_call_type = OptoRuntime::new_array_Type();
2154
2155 if (init_value == nullptr) {
2156 init_value = _igvn.zerocon(T_OBJECT);
2157 } else if (UseCompressedOops) {
2158 init_value = transform_later(new DecodeNNode(init_value, init_value->bottom_type()->make_ptr()));
2159 }
2160 }
2161 expand_allocate_common(alloc, length, init_value,
2162 slow_call_type,
2163 slow_call_address, valid_length_test);
2164 }
2165
2166 //-------------------mark_eliminated_box----------------------------------
2167 //
2168 // During EA obj may point to several objects but after few ideal graph
2169 // transformations (CCP) it may point to only one non escaping object
2170 // (but still using phi), corresponding locks and unlocks will be marked
2171 // for elimination. Later obj could be replaced with a new node (new phi)
2172 // and which does not have escape information. And later after some graph
2173 // reshape other locks and unlocks (which were not marked for elimination
2174 // before) are connected to this new obj (phi) but they still will not be
2175 // marked for elimination since new obj has no escape information.
2176 // Mark all associated (same box and obj) lock and unlock nodes for
2177 // elimination if some of them marked already.
2178 void PhaseMacroExpand::mark_eliminated_box(Node* box, Node* obj) {
2179 BoxLockNode* oldbox = box->as_BoxLock();
2180 if (oldbox->is_eliminated()) {
2181 return; // This BoxLock node was processed already.
2182 }
2183 assert(!oldbox->is_unbalanced(), "this should not be called for unbalanced region");
2184 // New implementation (EliminateNestedLocks) has separate BoxLock
2185 // node for each locked region so mark all associated locks/unlocks as
2186 // eliminated even if different objects are referenced in one locked region
2187 // (for example, OSR compilation of nested loop inside locked scope).
2188 if (EliminateNestedLocks ||
2189 oldbox->as_BoxLock()->is_simple_lock_region(nullptr, obj, nullptr)) {
2190 // Box is used only in one lock region. Mark this box as eliminated.
2191 oldbox->set_local(); // This verifies correct state of BoxLock
2192 _igvn.hash_delete(oldbox);
2193 oldbox->set_eliminated(); // This changes box's hash value
2194 _igvn.hash_insert(oldbox);
2195
2196 for (uint i = 0; i < oldbox->outcnt(); i++) {
2197 Node* u = oldbox->raw_out(i);
2198 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
2199 AbstractLockNode* alock = u->as_AbstractLock();
2200 // Check lock's box since box could be referenced by Lock's debug info.
2201 if (alock->box_node() == oldbox) {
2202 // Mark eliminated all related locks and unlocks.
2203 #ifdef ASSERT
2204 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4");
2205 #endif
2206 alock->set_non_esc_obj();
2207 }
2208 }
2209 }
2210 return;
2211 }
2212
2213 // Create new "eliminated" BoxLock node and use it in monitor debug info
2214 // instead of oldbox for the same object.
2215 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
2216
2217 // Note: BoxLock node is marked eliminated only here and it is used
2218 // to indicate that all associated lock and unlock nodes are marked
2219 // for elimination.
2220 newbox->set_local(); // This verifies correct state of BoxLock
2221 newbox->set_eliminated();
2222 transform_later(newbox);
2223
2224 // Replace old box node with new box for all users of the same object.
2225 for (uint i = 0; i < oldbox->outcnt();) {
2226 bool next_edge = true;
2227
2228 Node* u = oldbox->raw_out(i);
2229 if (u->is_AbstractLock()) {
2230 AbstractLockNode* alock = u->as_AbstractLock();
2231 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
2232 // Replace Box and mark eliminated all related locks and unlocks.
2233 #ifdef ASSERT
2234 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5");
2235 #endif
2236 alock->set_non_esc_obj();
2237 _igvn.rehash_node_delayed(alock);
2238 alock->set_box_node(newbox);
2239 next_edge = false;
2240 }
2241 }
2242 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
2243 FastLockNode* flock = u->as_FastLock();
2244 assert(flock->box_node() == oldbox, "sanity");
2245 _igvn.rehash_node_delayed(flock);
2246 flock->set_box_node(newbox);
2247 next_edge = false;
2248 }
2249
2250 // Replace old box in monitor debug info.
2251 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
2252 SafePointNode* sfn = u->as_SafePoint();
2253 JVMState* youngest_jvms = sfn->jvms();
2254 int max_depth = youngest_jvms->depth();
2255 for (int depth = 1; depth <= max_depth; depth++) {
2256 JVMState* jvms = youngest_jvms->of_depth(depth);
2257 int num_mon = jvms->nof_monitors();
2258 // Loop over monitors
2259 for (int idx = 0; idx < num_mon; idx++) {
2260 Node* obj_node = sfn->monitor_obj(jvms, idx);
2261 Node* box_node = sfn->monitor_box(jvms, idx);
2262 if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
2263 int j = jvms->monitor_box_offset(idx);
2264 _igvn.replace_input_of(u, j, newbox);
2265 next_edge = false;
2266 }
2267 }
2268 }
2269 }
2270 if (next_edge) i++;
2271 }
2272 }
2273
2274 //-----------------------mark_eliminated_locking_nodes-----------------------
2275 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
2276 if (!alock->is_balanced()) {
2277 return; // Can't do any more elimination for this locking region
2278 }
2279 if (EliminateNestedLocks) {
2280 if (alock->is_nested()) {
2281 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
2282 return;
2283 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
2284 // Only Lock node has JVMState needed here.
2285 // Not that preceding claim is documented anywhere else.
2286 if (alock->jvms() != nullptr) {
2287 if (alock->as_Lock()->is_nested_lock_region()) {
2288 // Mark eliminated related nested locks and unlocks.
2289 Node* obj = alock->obj_node();
2290 BoxLockNode* box_node = alock->box_node()->as_BoxLock();
2291 assert(!box_node->is_eliminated(), "should not be marked yet");
2292 // Note: BoxLock node is marked eliminated only here
2293 // and it is used to indicate that all associated lock
2294 // and unlock nodes are marked for elimination.
2295 box_node->set_eliminated(); // Box's hash is always NO_HASH here
2296 for (uint i = 0; i < box_node->outcnt(); i++) {
2297 Node* u = box_node->raw_out(i);
2298 if (u->is_AbstractLock()) {
2299 alock = u->as_AbstractLock();
2300 if (alock->box_node() == box_node) {
2301 // Verify that this Box is referenced only by related locks.
2302 assert(alock->obj_node()->eqv_uncast(obj), "");
2303 // Mark all related locks and unlocks.
2304 #ifdef ASSERT
2305 alock->log_lock_optimization(C, "eliminate_lock_set_nested");
2306 #endif
2307 alock->set_nested();
2308 }
2309 }
2310 }
2311 } else {
2312 #ifdef ASSERT
2313 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region");
2314 if (C->log() != nullptr)
2315 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output
2316 #endif
2317 }
2318 }
2319 return;
2320 }
2321 // Process locks for non escaping object
2322 assert(alock->is_non_esc_obj(), "");
2323 } // EliminateNestedLocks
2324
2325 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2326 // Look for all locks of this object and mark them and
2327 // corresponding BoxLock nodes as eliminated.
2328 Node* obj = alock->obj_node();
2329 for (uint j = 0; j < obj->outcnt(); j++) {
2330 Node* o = obj->raw_out(j);
2331 if (o->is_AbstractLock() &&
2332 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2333 alock = o->as_AbstractLock();
2334 Node* box = alock->box_node();
2335 // Replace old box node with new eliminated box for all users
2336 // of the same object and mark related locks as eliminated.
2337 mark_eliminated_box(box, obj);
2338 }
2339 }
2340 }
2341 }
2342
2343 // we have determined that this lock/unlock can be eliminated, we simply
2344 // eliminate the node without expanding it.
2345 //
2346 // Note: The membar's associated with the lock/unlock are currently not
2347 // eliminated. This should be investigated as a future enhancement.
2348 //
2349 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2350
2351 if (!alock->is_eliminated()) {
2352 return false;
2353 }
2354 #ifdef ASSERT
2355 if (!alock->is_coarsened()) {
2356 // Check that new "eliminated" BoxLock node is created.
2357 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2358 assert(oldbox->is_eliminated(), "should be done already");
2359 }
2360 #endif
2361
2362 alock->log_lock_optimization(C, "eliminate_lock");
2363
2364 #ifndef PRODUCT
2365 if (PrintEliminateLocks) {
2366 tty->print_cr("++++ Eliminated: %d %s '%s'", alock->_idx, (alock->is_Lock() ? "Lock" : "Unlock"), alock->kind_as_string());
2367 }
2368 #endif
2369
2370 Node* mem = alock->in(TypeFunc::Memory);
2371 Node* ctrl = alock->in(TypeFunc::Control);
2372 guarantee(ctrl != nullptr, "missing control projection, cannot replace_node() with null");
2373
2374 _callprojs = alock->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2375 // There are 2 projections from the lock. The lock node will
2376 // be deleted when its last use is subsumed below.
2377 assert(alock->outcnt() == 2 &&
2378 _callprojs->fallthrough_proj != nullptr &&
2379 _callprojs->fallthrough_memproj != nullptr,
2380 "Unexpected projections from Lock/Unlock");
2381
2382 Node* fallthroughproj = _callprojs->fallthrough_proj;
2383 Node* memproj_fallthrough = _callprojs->fallthrough_memproj;
2384
2385 // The memory projection from a lock/unlock is RawMem
2386 // The input to a Lock is merged memory, so extract its RawMem input
2387 // (unless the MergeMem has been optimized away.)
2388 if (alock->is_Lock()) {
2389 // Search for MemBarAcquireLock node and delete it also.
2390 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2391 assert(membar != nullptr && membar->Opcode() == Op_MemBarAcquireLock, "");
2392 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2393 Node* memproj = membar->proj_out(TypeFunc::Memory);
2394 _igvn.replace_node(ctrlproj, fallthroughproj);
2395 _igvn.replace_node(memproj, memproj_fallthrough);
2396
2397 // Delete FastLock node also if this Lock node is unique user
2398 // (a loop peeling may clone a Lock node).
2399 Node* flock = alock->as_Lock()->fastlock_node();
2400 if (flock->outcnt() == 1) {
2401 assert(flock->unique_out() == alock, "sanity");
2402 _igvn.replace_node(flock, top());
2403 }
2404 }
2405
2406 // Search for MemBarReleaseLock node and delete it also.
2407 if (alock->is_Unlock() && ctrl->is_Proj() && ctrl->in(0)->is_MemBar()) {
2408 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2409 assert(membar->Opcode() == Op_MemBarReleaseLock &&
2410 mem->is_Proj() && membar == mem->in(0), "");
2411 _igvn.replace_node(fallthroughproj, ctrl);
2412 _igvn.replace_node(memproj_fallthrough, mem);
2413 fallthroughproj = ctrl;
2414 memproj_fallthrough = mem;
2415 ctrl = membar->in(TypeFunc::Control);
2416 mem = membar->in(TypeFunc::Memory);
2417 }
2418
2419 _igvn.replace_node(fallthroughproj, ctrl);
2420 _igvn.replace_node(memproj_fallthrough, mem);
2421 return true;
2422 }
2423
2424
2425 //------------------------------expand_lock_node----------------------
2426 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2427
2428 Node* ctrl = lock->in(TypeFunc::Control);
2429 Node* mem = lock->in(TypeFunc::Memory);
2430 Node* obj = lock->obj_node();
2431 Node* box = lock->box_node();
2432 Node* flock = lock->fastlock_node();
2433
2434 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2435
2436 // Make the merge point
2437 Node *region;
2438 Node *mem_phi;
2439 Node *slow_path;
2440
2441 region = new RegionNode(3);
2442 // create a Phi for the memory state
2443 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2444
2445 // Optimize test; set region slot 2
2446 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2447 mem_phi->init_req(2, mem);
2448
2449 // Make slow path call
2450 CallNode* call = make_slow_call(lock, OptoRuntime::complete_monitor_enter_Type(),
2451 OptoRuntime::complete_monitor_locking_Java(), nullptr, slow_path,
2452 obj, box, nullptr);
2453
2454 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2455
2456 // Slow path can only throw asynchronous exceptions, which are always
2457 // de-opted. So the compiler thinks the slow-call can never throw an
2458 // exception. If it DOES throw an exception we would need the debug
2459 // info removed first (since if it throws there is no monitor).
2460 assert(_callprojs->fallthrough_ioproj == nullptr && _callprojs->catchall_ioproj == nullptr &&
2461 _callprojs->catchall_memproj == nullptr && _callprojs->catchall_catchproj == nullptr, "Unexpected projection from Lock");
2462
2463 // Capture slow path
2464 // disconnect fall-through projection from call and create a new one
2465 // hook up users of fall-through projection to region
2466 Node *slow_ctrl = _callprojs->fallthrough_proj->clone();
2467 transform_later(slow_ctrl);
2468 _igvn.hash_delete(_callprojs->fallthrough_proj);
2469 _callprojs->fallthrough_proj->disconnect_inputs(C);
2470 region->init_req(1, slow_ctrl);
2471 // region inputs are now complete
2472 transform_later(region);
2473 _igvn.replace_node(_callprojs->fallthrough_proj, region);
2474
2475 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2476
2477 mem_phi->init_req(1, memproj);
2478
2479 transform_later(mem_phi);
2480
2481 _igvn.replace_node(_callprojs->fallthrough_memproj, mem_phi);
2482 }
2483
2484 //------------------------------expand_unlock_node----------------------
2485 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2486
2487 Node* ctrl = unlock->in(TypeFunc::Control);
2488 Node* mem = unlock->in(TypeFunc::Memory);
2489 Node* obj = unlock->obj_node();
2490 Node* box = unlock->box_node();
2491
2492 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2493
2494 // No need for a null check on unlock
2495
2496 // Make the merge point
2497 Node *region;
2498 Node *mem_phi;
2499
2500 region = new RegionNode(3);
2501 // create a Phi for the memory state
2502 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2503
2504 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2505 funlock = transform_later( funlock )->as_FastUnlock();
2506 // Optimize test; set region slot 2
2507 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2508 Node *thread = transform_later(new ThreadLocalNode());
2509
2510 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2511 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2512 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2513
2514 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2515 assert(_callprojs->fallthrough_ioproj == nullptr && _callprojs->catchall_ioproj == nullptr &&
2516 _callprojs->catchall_memproj == nullptr && _callprojs->catchall_catchproj == nullptr, "Unexpected projection from Lock");
2517
2518 // No exceptions for unlocking
2519 // Capture slow path
2520 // disconnect fall-through projection from call and create a new one
2521 // hook up users of fall-through projection to region
2522 Node *slow_ctrl = _callprojs->fallthrough_proj->clone();
2523 transform_later(slow_ctrl);
2524 _igvn.hash_delete(_callprojs->fallthrough_proj);
2525 _callprojs->fallthrough_proj->disconnect_inputs(C);
2526 region->init_req(1, slow_ctrl);
2527 // region inputs are now complete
2528 transform_later(region);
2529 _igvn.replace_node(_callprojs->fallthrough_proj, region);
2530
2531 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2532 mem_phi->init_req(1, memproj );
2533 mem_phi->init_req(2, mem);
2534 transform_later(mem_phi);
2535
2536 _igvn.replace_node(_callprojs->fallthrough_memproj, mem_phi);
2537 }
2538
2539 // An inline type might be returned from the call but we don't know its
2540 // type. Either we get a buffered inline type (and nothing needs to be done)
2541 // or one of the values being returned is the klass of the inline type
2542 // and we need to allocate an inline type instance of that type and
2543 // initialize it with other values being returned. In that case, we
2544 // first try a fast path allocation and initialize the value with the
2545 // inline klass's pack handler or we fall back to a runtime call.
2546 void PhaseMacroExpand::expand_mh_intrinsic_return(CallStaticJavaNode* call) {
2547 assert(call->method()->is_method_handle_intrinsic(), "must be a method handle intrinsic call");
2548 Node* ret = call->proj_out_or_null(TypeFunc::Parms);
2549 if (ret == nullptr) {
2550 return;
2551 }
2552 const TypeFunc* tf = call->_tf;
2553 const TypeTuple* domain = OptoRuntime::store_inline_type_fields_Type()->domain_cc();
2554 const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain);
2555 call->_tf = new_tf;
2556 // Make sure the change of type is applied before projections are processed by igvn
2557 _igvn.set_type(call, call->Value(&_igvn));
2558 _igvn.set_type(ret, ret->Value(&_igvn));
2559
2560 // Before any new projection is added:
2561 CallProjections* projs = call->extract_projections(true, true);
2562
2563 // Create temporary hook nodes that will be replaced below.
2564 // Add an input to prevent hook nodes from being dead.
2565 Node* ctl = new Node(call);
2566 Node* mem = new Node(ctl);
2567 Node* io = new Node(ctl);
2568 Node* ex_ctl = new Node(ctl);
2569 Node* ex_mem = new Node(ctl);
2570 Node* ex_io = new Node(ctl);
2571 Node* res = new Node(ctl);
2572
2573 // Allocate a new buffered inline type only if a new one is not returned
2574 Node* cast = transform_later(new CastP2XNode(ctl, res));
2575 Node* mask = MakeConX(0x1);
2576 Node* masked = transform_later(new AndXNode(cast, mask));
2577 Node* cmp = transform_later(new CmpXNode(masked, mask));
2578 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
2579 IfNode* allocation_iff = new IfNode(ctl, bol, PROB_MAX, COUNT_UNKNOWN);
2580 transform_later(allocation_iff);
2581 Node* allocation_ctl = transform_later(new IfTrueNode(allocation_iff));
2582 Node* no_allocation_ctl = transform_later(new IfFalseNode(allocation_iff));
2583 Node* no_allocation_res = transform_later(new CheckCastPPNode(no_allocation_ctl, res, TypeInstPtr::BOTTOM));
2584
2585 // Try to allocate a new buffered inline instance either from TLAB or eden space
2586 Node* needgc_ctrl = nullptr; // needgc means slowcase, i.e. allocation failed
2587 CallLeafNoFPNode* handler_call;
2588 const bool alloc_in_place = UseTLAB;
2589 if (alloc_in_place) {
2590 Node* fast_oop_ctrl = nullptr;
2591 Node* fast_oop_rawmem = nullptr;
2592 Node* mask2 = MakeConX(-2);
2593 Node* masked2 = transform_later(new AndXNode(cast, mask2));
2594 Node* rawklassptr = transform_later(new CastX2PNode(masked2));
2595 Node* klass_node = transform_later(new CheckCastPPNode(allocation_ctl, rawklassptr, TypeInstKlassPtr::OBJECT_OR_NULL));
2596 Node* layout_val = make_load(nullptr, mem, klass_node, in_bytes(Klass::layout_helper_offset()), TypeInt::INT, T_INT);
2597 Node* size_in_bytes = ConvI2X(layout_val);
2598 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2599 Node* fast_oop = bs->obj_allocate(this, mem, allocation_ctl, size_in_bytes, io, needgc_ctrl,
2600 fast_oop_ctrl, fast_oop_rawmem,
2601 AllocateInstancePrefetchLines);
2602 // Allocation succeed, initialize buffered inline instance header firstly,
2603 // and then initialize its fields with an inline class specific handler
2604 Node* mark_node = makecon(TypeRawPtr::make((address)markWord::inline_type_prototype().value()));
2605 fast_oop_rawmem = make_store(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
2606 fast_oop_rawmem = make_store(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
2607 if (UseCompressedClassPointers) {
2608 fast_oop_rawmem = make_store(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::klass_gap_offset_in_bytes(), intcon(0), T_INT);
2609 }
2610 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);
2611 Node* pack_handler = make_load(fast_oop_ctrl, fast_oop_rawmem, fixed_block, in_bytes(InlineKlass::pack_handler_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2612 handler_call = new CallLeafNoFPNode(OptoRuntime::pack_inline_type_Type(),
2613 nullptr,
2614 "pack handler",
2615 TypeRawPtr::BOTTOM);
2616 handler_call->init_req(TypeFunc::Control, fast_oop_ctrl);
2617 handler_call->init_req(TypeFunc::Memory, fast_oop_rawmem);
2618 handler_call->init_req(TypeFunc::I_O, top());
2619 handler_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2620 handler_call->init_req(TypeFunc::ReturnAdr, top());
2621 handler_call->init_req(TypeFunc::Parms, pack_handler);
2622 handler_call->init_req(TypeFunc::Parms+1, fast_oop);
2623 } else {
2624 needgc_ctrl = allocation_ctl;
2625 }
2626
2627 // Allocation failed, fall back to a runtime call
2628 CallStaticJavaNode* slow_call = new CallStaticJavaNode(OptoRuntime::store_inline_type_fields_Type(),
2629 StubRoutines::store_inline_type_fields_to_buf(),
2630 "store_inline_type_fields",
2631 TypePtr::BOTTOM);
2632 slow_call->init_req(TypeFunc::Control, needgc_ctrl);
2633 slow_call->init_req(TypeFunc::Memory, mem);
2634 slow_call->init_req(TypeFunc::I_O, io);
2635 slow_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2636 slow_call->init_req(TypeFunc::ReturnAdr, call->in(TypeFunc::ReturnAdr));
2637 slow_call->init_req(TypeFunc::Parms, res);
2638
2639 Node* slow_ctl = transform_later(new ProjNode(slow_call, TypeFunc::Control));
2640 Node* slow_mem = transform_later(new ProjNode(slow_call, TypeFunc::Memory));
2641 Node* slow_io = transform_later(new ProjNode(slow_call, TypeFunc::I_O));
2642 Node* slow_res = transform_later(new ProjNode(slow_call, TypeFunc::Parms));
2643 Node* slow_catc = transform_later(new CatchNode(slow_ctl, slow_io, 2));
2644 Node* slow_norm = transform_later(new CatchProjNode(slow_catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci));
2645 Node* slow_excp = transform_later(new CatchProjNode(slow_catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci));
2646
2647 Node* ex_r = new RegionNode(3);
2648 Node* ex_mem_phi = new PhiNode(ex_r, Type::MEMORY, TypePtr::BOTTOM);
2649 Node* ex_io_phi = new PhiNode(ex_r, Type::ABIO);
2650 ex_r->init_req(1, slow_excp);
2651 ex_mem_phi->init_req(1, slow_mem);
2652 ex_io_phi->init_req(1, slow_io);
2653 ex_r->init_req(2, ex_ctl);
2654 ex_mem_phi->init_req(2, ex_mem);
2655 ex_io_phi->init_req(2, ex_io);
2656 transform_later(ex_r);
2657 transform_later(ex_mem_phi);
2658 transform_later(ex_io_phi);
2659
2660 // We don't know how many values are returned. This assumes the
2661 // worst case, that all available registers are used.
2662 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2663 if (domain->field_at(i) == Type::HALF) {
2664 slow_call->init_req(i, top());
2665 if (alloc_in_place) {
2666 handler_call->init_req(i+1, top());
2667 }
2668 continue;
2669 }
2670 Node* proj = transform_later(new ProjNode(call, i));
2671 slow_call->init_req(i, proj);
2672 if (alloc_in_place) {
2673 handler_call->init_req(i+1, proj);
2674 }
2675 }
2676 // We can safepoint at that new call
2677 slow_call->copy_call_debug_info(&_igvn, call);
2678 transform_later(slow_call);
2679 if (alloc_in_place) {
2680 transform_later(handler_call);
2681 }
2682
2683 Node* fast_ctl = nullptr;
2684 Node* fast_res = nullptr;
2685 MergeMemNode* fast_mem = nullptr;
2686 if (alloc_in_place) {
2687 fast_ctl = transform_later(new ProjNode(handler_call, TypeFunc::Control));
2688 Node* rawmem = transform_later(new ProjNode(handler_call, TypeFunc::Memory));
2689 fast_res = transform_later(new ProjNode(handler_call, TypeFunc::Parms));
2690 fast_mem = MergeMemNode::make(mem);
2691 fast_mem->set_memory_at(Compile::AliasIdxRaw, rawmem);
2692 transform_later(fast_mem);
2693 }
2694
2695 Node* r = new RegionNode(alloc_in_place ? 4 : 3);
2696 Node* mem_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2697 Node* io_phi = new PhiNode(r, Type::ABIO);
2698 Node* res_phi = new PhiNode(r, TypeInstPtr::BOTTOM);
2699 r->init_req(1, no_allocation_ctl);
2700 mem_phi->init_req(1, mem);
2701 io_phi->init_req(1, io);
2702 res_phi->init_req(1, no_allocation_res);
2703 r->init_req(2, slow_norm);
2704 mem_phi->init_req(2, slow_mem);
2705 io_phi->init_req(2, slow_io);
2706 res_phi->init_req(2, slow_res);
2707 if (alloc_in_place) {
2708 r->init_req(3, fast_ctl);
2709 mem_phi->init_req(3, fast_mem);
2710 io_phi->init_req(3, io);
2711 res_phi->init_req(3, fast_res);
2712 }
2713 transform_later(r);
2714 transform_later(mem_phi);
2715 transform_later(io_phi);
2716 transform_later(res_phi);
2717
2718 // Do not let stores that initialize this buffer be reordered with a subsequent
2719 // store that would make this buffer accessible by other threads.
2720 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
2721 transform_later(mb);
2722 mb->init_req(TypeFunc::Memory, mem_phi);
2723 mb->init_req(TypeFunc::Control, r);
2724 r = new ProjNode(mb, TypeFunc::Control);
2725 transform_later(r);
2726 mem_phi = new ProjNode(mb, TypeFunc::Memory);
2727 transform_later(mem_phi);
2728
2729 assert(projs->nb_resproj == 1, "unexpected number of results");
2730 _igvn.replace_in_uses(projs->fallthrough_catchproj, r);
2731 _igvn.replace_in_uses(projs->fallthrough_memproj, mem_phi);
2732 _igvn.replace_in_uses(projs->fallthrough_ioproj, io_phi);
2733 _igvn.replace_in_uses(projs->resproj[0], res_phi);
2734 _igvn.replace_in_uses(projs->catchall_catchproj, ex_r);
2735 _igvn.replace_in_uses(projs->catchall_memproj, ex_mem_phi);
2736 _igvn.replace_in_uses(projs->catchall_ioproj, ex_io_phi);
2737 // The CatchNode should not use the ex_io_phi. Re-connect it to the catchall_ioproj.
2738 Node* cn = projs->fallthrough_catchproj->in(0);
2739 _igvn.replace_input_of(cn, 1, projs->catchall_ioproj);
2740
2741 _igvn.replace_node(ctl, projs->fallthrough_catchproj);
2742 _igvn.replace_node(mem, projs->fallthrough_memproj);
2743 _igvn.replace_node(io, projs->fallthrough_ioproj);
2744 _igvn.replace_node(res, projs->resproj[0]);
2745 _igvn.replace_node(ex_ctl, projs->catchall_catchproj);
2746 _igvn.replace_node(ex_mem, projs->catchall_memproj);
2747 _igvn.replace_node(ex_io, projs->catchall_ioproj);
2748 }
2749
2750 void PhaseMacroExpand::expand_subtypecheck_node(SubTypeCheckNode *check) {
2751 assert(check->in(SubTypeCheckNode::Control) == nullptr, "should be pinned");
2752 Node* bol = check->unique_out();
2753 Node* obj_or_subklass = check->in(SubTypeCheckNode::ObjOrSubKlass);
2754 Node* superklass = check->in(SubTypeCheckNode::SuperKlass);
2755 assert(bol->is_Bool() && bol->as_Bool()->_test._test == BoolTest::ne, "unexpected bool node");
2756
2757 for (DUIterator_Last imin, i = bol->last_outs(imin); i >= imin; --i) {
2758 Node* iff = bol->last_out(i);
2759 assert(iff->is_If(), "where's the if?");
2760
2761 if (iff->in(0)->is_top()) {
2762 _igvn.replace_input_of(iff, 1, C->top());
2763 continue;
2764 }
2765
2766 Node* iftrue = iff->as_If()->proj_out(1);
2767 Node* iffalse = iff->as_If()->proj_out(0);
2768 Node* ctrl = iff->in(0);
2769
2770 Node* subklass = nullptr;
2771 if (_igvn.type(obj_or_subklass)->isa_klassptr()) {
2772 subklass = obj_or_subklass;
2773 } else {
2774 Node* k_adr = basic_plus_adr(obj_or_subklass, oopDesc::klass_offset_in_bytes());
2775 subklass = _igvn.transform(LoadKlassNode::make(_igvn, C->immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
2776 }
2777
2778 Node* not_subtype_ctrl = Phase::gen_subtype_check(subklass, superklass, &ctrl, nullptr, _igvn, check->method(), check->bci());
2779
2780 _igvn.replace_input_of(iff, 0, C->top());
2781 _igvn.replace_node(iftrue, not_subtype_ctrl);
2782 _igvn.replace_node(iffalse, ctrl);
2783 }
2784 _igvn.replace_node(check, C->top());
2785 }
2786
2787 // FlatArrayCheckNode (array1 array2 ...) is expanded into:
2788 //
2789 // long mark = array1.mark | array2.mark | ...;
2790 // long locked_bit = markWord::unlocked_value & array1.mark & array2.mark & ...;
2791 // if (locked_bit == 0) {
2792 // // One array is locked, load prototype header from the klass
2793 // mark = array1.klass.proto | array2.klass.proto | ...
2794 // }
2795 // if ((mark & markWord::flat_array_bit_in_place) == 0) {
2796 // ...
2797 // }
2798 void PhaseMacroExpand::expand_flatarraycheck_node(FlatArrayCheckNode* check) {
2799 bool array_inputs = _igvn.type(check->in(FlatArrayCheckNode::ArrayOrKlass))->isa_oopptr() != nullptr;
2800 if (array_inputs) {
2801 Node* mark = MakeConX(0);
2802 Node* locked_bit = MakeConX(markWord::unlocked_value);
2803 Node* mem = check->in(FlatArrayCheckNode::Memory);
2804 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
2805 Node* ary = check->in(i);
2806 const TypeOopPtr* t = _igvn.type(ary)->isa_oopptr();
2807 assert(t != nullptr, "Mixing array and klass inputs");
2808 assert(!t->is_flat() && !t->is_not_flat(), "Should have been optimized out");
2809 Node* mark_adr = basic_plus_adr(ary, oopDesc::mark_offset_in_bytes());
2810 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));
2811 mark = _igvn.transform(new OrXNode(mark, mark_load));
2812 locked_bit = _igvn.transform(new AndXNode(locked_bit, mark_load));
2813 }
2814 assert(!mark->is_Con(), "Should have been optimized out");
2815 Node* cmp = _igvn.transform(new CmpXNode(locked_bit, MakeConX(0)));
2816 Node* is_unlocked = _igvn.transform(new BoolNode(cmp, BoolTest::ne));
2817
2818 // BoolNode might be shared, replace each if user
2819 Node* old_bol = check->unique_out();
2820 assert(old_bol->is_Bool() && old_bol->as_Bool()->_test._test == BoolTest::ne, "unexpected condition");
2821 for (DUIterator_Last imin, i = old_bol->last_outs(imin); i >= imin; --i) {
2822 IfNode* old_iff = old_bol->last_out(i)->as_If();
2823 Node* ctrl = old_iff->in(0);
2824 RegionNode* region = new RegionNode(3);
2825 Node* mark_phi = new PhiNode(region, TypeX_X);
2826
2827 // Check if array is unlocked
2828 IfNode* iff = _igvn.transform(new IfNode(ctrl, is_unlocked, PROB_MAX, COUNT_UNKNOWN))->as_If();
2829
2830 // Unlocked: Use bits from mark word
2831 region->init_req(1, _igvn.transform(new IfTrueNode(iff)));
2832 mark_phi->init_req(1, mark);
2833
2834 // Locked: Load prototype header from klass
2835 ctrl = _igvn.transform(new IfFalseNode(iff));
2836 Node* proto = MakeConX(0);
2837 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
2838 Node* ary = check->in(i);
2839 // Make loads control dependent to make sure they are only executed if array is locked
2840 Node* klass_adr = basic_plus_adr(ary, oopDesc::klass_offset_in_bytes());
2841 Node* klass = _igvn.transform(LoadKlassNode::make(_igvn, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
2842 Node* proto_adr = basic_plus_adr(klass, in_bytes(Klass::prototype_header_offset()));
2843 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));
2844 proto = _igvn.transform(new OrXNode(proto, proto_load));
2845 }
2846 region->init_req(2, ctrl);
2847 mark_phi->init_req(2, proto);
2848
2849 // Check if flat array bits are set
2850 Node* mask = MakeConX(markWord::flat_array_bit_in_place);
2851 Node* masked = _igvn.transform(new AndXNode(_igvn.transform(mark_phi), mask));
2852 cmp = _igvn.transform(new CmpXNode(masked, MakeConX(0)));
2853 Node* is_not_flat = _igvn.transform(new BoolNode(cmp, BoolTest::eq));
2854
2855 ctrl = _igvn.transform(region);
2856 iff = _igvn.transform(new IfNode(ctrl, is_not_flat, PROB_MAX, COUNT_UNKNOWN))->as_If();
2857 _igvn.replace_node(old_iff, iff);
2858 }
2859 _igvn.replace_node(check, C->top());
2860 } else {
2861 // Fall back to layout helper check
2862 Node* lhs = intcon(0);
2863 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
2864 Node* array_or_klass = check->in(i);
2865 Node* klass = nullptr;
2866 const TypePtr* t = _igvn.type(array_or_klass)->is_ptr();
2867 assert(!t->is_flat() && !t->is_not_flat(), "Should have been optimized out");
2868 if (t->isa_oopptr() != nullptr) {
2869 Node* klass_adr = basic_plus_adr(array_or_klass, oopDesc::klass_offset_in_bytes());
2870 klass = transform_later(LoadKlassNode::make(_igvn, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
2871 } else {
2872 assert(t->isa_klassptr(), "Unexpected input type");
2873 klass = array_or_klass;
2874 }
2875 Node* lh_addr = basic_plus_adr(klass, in_bytes(Klass::layout_helper_offset()));
2876 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));
2877 lhs = _igvn.transform(new OrINode(lhs, lh_val));
2878 }
2879 Node* masked = transform_later(new AndINode(lhs, intcon(Klass::_lh_array_tag_flat_value_bit_inplace)));
2880 Node* cmp = transform_later(new CmpINode(masked, intcon(0)));
2881 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
2882 Node* m2b = transform_later(new Conv2BNode(masked));
2883 // The matcher expects the input to If nodes to be produced by a Bool(CmpI..)
2884 // pattern, but the input to other potential users (e.g. Phi) to be some
2885 // other pattern (e.g. a Conv2B node, possibly idealized as a CMoveI).
2886 Node* old_bol = check->unique_out();
2887 for (DUIterator_Last imin, i = old_bol->last_outs(imin); i >= imin; --i) {
2888 Node* user = old_bol->last_out(i);
2889 for (uint j = 0; j < user->req(); j++) {
2890 Node* n = user->in(j);
2891 if (n == old_bol) {
2892 _igvn.replace_input_of(user, j, user->is_If() ? bol : m2b);
2893 }
2894 }
2895 }
2896 _igvn.replace_node(check, C->top());
2897 }
2898 }
2899
2900 //---------------------------eliminate_macro_nodes----------------------
2901 // Eliminate scalar replaced allocations and associated locks.
2902 void PhaseMacroExpand::eliminate_macro_nodes() {
2903 if (C->macro_count() == 0)
2904 return;
2905 NOT_PRODUCT(int membar_before = count_MemBar(C);)
2906
2907 // Before elimination may re-mark (change to Nested or NonEscObj)
2908 // all associated (same box and obj) lock and unlock nodes.
2909 int cnt = C->macro_count();
2910 for (int i=0; i < cnt; i++) {
2911 Node *n = C->macro_node(i);
2912 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2913 mark_eliminated_locking_nodes(n->as_AbstractLock());
2914 }
2915 }
2916 // Re-marking may break consistency of Coarsened locks.
2917 if (!C->coarsened_locks_consistent()) {
2918 return; // recompile without Coarsened locks if broken
2919 } else {
2920 // After coarsened locks are eliminated locking regions
2921 // become unbalanced. We should not execute any more
2922 // locks elimination optimizations on them.
2923 C->mark_unbalanced_boxes();
2924 }
2925
2926 // First, attempt to eliminate locks
2927 bool progress = true;
2928 while (progress) {
2929 progress = false;
2930 for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once
2931 Node* n = C->macro_node(i - 1);
2932 bool success = false;
2933 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2934 if (n->is_AbstractLock()) {
2935 success = eliminate_locking_node(n->as_AbstractLock());
2936 #ifndef PRODUCT
2937 if (success && PrintOptoStatistics) {
2938 Atomic::inc(&PhaseMacroExpand::_monitor_objects_removed_counter);
2939 }
2940 #endif
2941 }
2942 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2943 progress = progress || success;
2944 }
2945 }
2946 // Next, attempt to eliminate allocations
2947 _has_locks = false;
2948 progress = true;
2949 while (progress) {
2950 progress = false;
2951 for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once
2952 Node* n = C->macro_node(i - 1);
2953 bool success = false;
2954 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2955 switch (n->class_id()) {
2956 case Node::Class_Allocate:
2957 case Node::Class_AllocateArray:
2958 success = eliminate_allocate_node(n->as_Allocate());
2959 #ifndef PRODUCT
2960 if (success && PrintOptoStatistics) {
2961 Atomic::inc(&PhaseMacroExpand::_objs_scalar_replaced_counter);
2962 }
2963 #endif
2964 break;
2965 case Node::Class_CallStaticJava: {
2966 CallStaticJavaNode* call = n->as_CallStaticJava();
2967 if (!call->method()->is_method_handle_intrinsic()) {
2968 success = eliminate_boxing_node(n->as_CallStaticJava());
2969 }
2970 break;
2971 }
2972 case Node::Class_Lock:
2973 case Node::Class_Unlock:
2974 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2975 _has_locks = true;
2976 break;
2977 case Node::Class_ArrayCopy:
2978 break;
2979 case Node::Class_OuterStripMinedLoop:
2980 break;
2981 case Node::Class_SubTypeCheck:
2982 break;
2983 case Node::Class_Opaque1:
2984 break;
2985 case Node::Class_FlatArrayCheck:
2986 break;
2987 default:
2988 assert(n->Opcode() == Op_LoopLimit ||
2989 n->Opcode() == Op_ModD ||
2990 n->Opcode() == Op_ModF ||
2991 n->is_OpaqueNotNull() ||
2992 n->is_OpaqueInitializedAssertionPredicate() ||
2993 n->Opcode() == Op_MaxL ||
2994 n->Opcode() == Op_MinL ||
2995 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n),
2996 "unknown node type in macro list");
2997 }
2998 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2999 progress = progress || success;
3000 }
3001 }
3002 #ifndef PRODUCT
3003 if (PrintOptoStatistics) {
3004 int membar_after = count_MemBar(C);
3005 Atomic::add(&PhaseMacroExpand::_memory_barriers_removed_counter, membar_before - membar_after);
3006 }
3007 #endif
3008 }
3009
3010 //------------------------------expand_macro_nodes----------------------
3011 // Returns true if a failure occurred.
3012 bool PhaseMacroExpand::expand_macro_nodes() {
3013 // Do not allow new macro nodes once we started to expand
3014 C->reset_allow_macro_nodes();
3015 if (StressMacroExpansion) {
3016 C->shuffle_macro_nodes();
3017 }
3018 // Last attempt to eliminate macro nodes.
3019 eliminate_macro_nodes();
3020 if (C->failing()) return true;
3021
3022 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
3023 bool progress = true;
3024 while (progress) {
3025 progress = false;
3026 for (int i = C->macro_count(); i > 0; i--) {
3027 Node* n = C->macro_node(i-1);
3028 bool success = false;
3029 DEBUG_ONLY(int old_macro_count = C->macro_count();)
3030 if (n->Opcode() == Op_LoopLimit) {
3031 // Remove it from macro list and put on IGVN worklist to optimize.
3032 C->remove_macro_node(n);
3033 _igvn._worklist.push(n);
3034 success = true;
3035 } else if (n->Opcode() == Op_CallStaticJava) {
3036 CallStaticJavaNode* call = n->as_CallStaticJava();
3037 if (!call->method()->is_method_handle_intrinsic()) {
3038 // Remove it from macro list and put on IGVN worklist to optimize.
3039 C->remove_macro_node(n);
3040 _igvn._worklist.push(n);
3041 success = true;
3042 }
3043 } else if (n->is_Opaque1()) {
3044 _igvn.replace_node(n, n->in(1));
3045 success = true;
3046 } else if (n->is_OpaqueNotNull()) {
3047 // Tests with OpaqueNotNull nodes are implicitly known to be true. Replace the node with true. In debug builds,
3048 // we leave the test in the graph to have an additional sanity check at runtime. If the test fails (i.e. a bug),
3049 // we will execute a Halt node.
3050 #ifdef ASSERT
3051 _igvn.replace_node(n, n->in(1));
3052 #else
3053 _igvn.replace_node(n, _igvn.intcon(1));
3054 #endif
3055 success = true;
3056 } else if (n->is_OpaqueInitializedAssertionPredicate()) {
3057 // Initialized Assertion Predicates must always evaluate to true. Therefore, we get rid of them in product
3058 // builds as they are useless. In debug builds we keep them as additional verification code. Even though
3059 // loop opts are already over, we want to keep Initialized Assertion Predicates alive as long as possible to
3060 // enable folding of dead control paths within which cast nodes become top after due to impossible types -
3061 // even after loop opts are over. Therefore, we delay the removal of these opaque nodes until now.
3062 #ifdef ASSERT
3063 _igvn.replace_node(n, n->in(1));
3064 #else
3065 _igvn.replace_node(n, _igvn.intcon(1));
3066 #endif // ASSERT
3067 } else if (n->Opcode() == Op_OuterStripMinedLoop) {
3068 n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn);
3069 C->remove_macro_node(n);
3070 success = true;
3071 } else if (n->Opcode() == Op_MaxL) {
3072 // Since MaxL and MinL are not implemented in the backend, we expand them to
3073 // a CMoveL construct now. At least until here, the type could be computed
3074 // precisely. CMoveL is not so smart, but we can give it at least the best
3075 // type we know abouot n now.
3076 Node* repl = MaxNode::signed_max(n->in(1), n->in(2), _igvn.type(n), _igvn);
3077 _igvn.replace_node(n, repl);
3078 success = true;
3079 } else if (n->Opcode() == Op_MinL) {
3080 Node* repl = MaxNode::signed_min(n->in(1), n->in(2), _igvn.type(n), _igvn);
3081 _igvn.replace_node(n, repl);
3082 success = true;
3083 }
3084 assert(!success || (C->macro_count() == (old_macro_count - 1)), "elimination must have deleted one node from macro list");
3085 progress = progress || success;
3086 if (success) {
3087 C->print_method(PHASE_AFTER_MACRO_EXPANSION_STEP, 5, n);
3088 }
3089 }
3090 }
3091
3092 // Clean up the graph so we're less likely to hit the maximum node
3093 // limit
3094 _igvn.set_delay_transform(false);
3095 _igvn.optimize();
3096 if (C->failing()) return true;
3097 _igvn.set_delay_transform(true);
3098
3099
3100 // Because we run IGVN after each expansion, some macro nodes may go
3101 // dead and be removed from the list as we iterate over it. Move
3102 // Allocate nodes (processed in a second pass) at the beginning of
3103 // the list and then iterate from the last element of the list until
3104 // an Allocate node is seen. This is robust to random deletion in
3105 // the list due to nodes going dead.
3106 C->sort_macro_nodes();
3107
3108 // expand arraycopy "macro" nodes first
3109 // For ReduceBulkZeroing, we must first process all arraycopy nodes
3110 // before the allocate nodes are expanded.
3111 while (C->macro_count() > 0) {
3112 int macro_count = C->macro_count();
3113 Node * n = C->macro_node(macro_count-1);
3114 assert(n->is_macro(), "only macro nodes expected here");
3115 if (_igvn.type(n) == Type::TOP || (n->in(0) != nullptr && n->in(0)->is_top())) {
3116 // node is unreachable, so don't try to expand it
3117 C->remove_macro_node(n);
3118 continue;
3119 }
3120 if (n->is_Allocate()) {
3121 break;
3122 }
3123 // Make sure expansion will not cause node limit to be exceeded.
3124 // Worst case is a macro node gets expanded into about 200 nodes.
3125 // Allow 50% more for optimization.
3126 if (C->check_node_count(300, "out of nodes before macro expansion")) {
3127 return true;
3128 }
3129
3130 DEBUG_ONLY(int old_macro_count = C->macro_count();)
3131 switch (n->class_id()) {
3132 case Node::Class_Lock:
3133 expand_lock_node(n->as_Lock());
3134 break;
3135 case Node::Class_Unlock:
3136 expand_unlock_node(n->as_Unlock());
3137 break;
3138 case Node::Class_ArrayCopy:
3139 expand_arraycopy_node(n->as_ArrayCopy());
3140 break;
3141 case Node::Class_SubTypeCheck:
3142 expand_subtypecheck_node(n->as_SubTypeCheck());
3143 break;
3144 case Node::Class_CallStaticJava:
3145 expand_mh_intrinsic_return(n->as_CallStaticJava());
3146 C->remove_macro_node(n);
3147 break;
3148 case Node::Class_FlatArrayCheck:
3149 expand_flatarraycheck_node(n->as_FlatArrayCheck());
3150 break;
3151 default:
3152 switch (n->Opcode()) {
3153 case Op_ModD:
3154 case Op_ModF: {
3155 bool is_drem = n->Opcode() == Op_ModD;
3156 CallNode* mod_macro = n->as_Call();
3157 CallNode* call = new CallLeafNode(mod_macro->tf(),
3158 is_drem ? CAST_FROM_FN_PTR(address, SharedRuntime::drem)
3159 : CAST_FROM_FN_PTR(address, SharedRuntime::frem),
3160 is_drem ? "drem" : "frem", TypeRawPtr::BOTTOM);
3161 call->init_req(TypeFunc::Control, mod_macro->in(TypeFunc::Control));
3162 call->init_req(TypeFunc::I_O, mod_macro->in(TypeFunc::I_O));
3163 call->init_req(TypeFunc::Memory, mod_macro->in(TypeFunc::Memory));
3164 call->init_req(TypeFunc::ReturnAdr, mod_macro->in(TypeFunc::ReturnAdr));
3165 call->init_req(TypeFunc::FramePtr, mod_macro->in(TypeFunc::FramePtr));
3166 for (unsigned int i = 0; i < mod_macro->tf()->domain_cc()->cnt() - TypeFunc::Parms; i++) {
3167 call->init_req(TypeFunc::Parms + i, mod_macro->in(TypeFunc::Parms + i));
3168 }
3169 _igvn.replace_node(mod_macro, call);
3170 transform_later(call);
3171 break;
3172 }
3173 default:
3174 assert(false, "unknown node type in macro list");
3175 }
3176 }
3177 assert(C->macro_count() == (old_macro_count - 1), "expansion must have deleted one node from macro list");
3178 if (C->failing()) return true;
3179 C->print_method(PHASE_AFTER_MACRO_EXPANSION_STEP, 5, n);
3180
3181 // Clean up the graph so we're less likely to hit the maximum node
3182 // limit
3183 _igvn.set_delay_transform(false);
3184 _igvn.optimize();
3185 if (C->failing()) return true;
3186 _igvn.set_delay_transform(true);
3187 }
3188
3189 // All nodes except Allocate nodes are expanded now. There could be
3190 // new optimization opportunities (such as folding newly created
3191 // load from a just allocated object). Run IGVN.
3192
3193 // expand "macro" nodes
3194 // nodes are removed from the macro list as they are processed
3195 while (C->macro_count() > 0) {
3196 int macro_count = C->macro_count();
3197 Node * n = C->macro_node(macro_count-1);
3198 assert(n->is_macro(), "only macro nodes expected here");
3199 if (_igvn.type(n) == Type::TOP || (n->in(0) != nullptr && n->in(0)->is_top())) {
3200 // node is unreachable, so don't try to expand it
3201 C->remove_macro_node(n);
3202 continue;
3203 }
3204 // Make sure expansion will not cause node limit to be exceeded.
3205 // Worst case is a macro node gets expanded into about 200 nodes.
3206 // Allow 50% more for optimization.
3207 if (C->check_node_count(300, "out of nodes before macro expansion")) {
3208 return true;
3209 }
3210 switch (n->class_id()) {
3211 case Node::Class_Allocate:
3212 expand_allocate(n->as_Allocate());
3213 break;
3214 case Node::Class_AllocateArray:
3215 expand_allocate_array(n->as_AllocateArray());
3216 break;
3217 default:
3218 assert(false, "unknown node type in macro list");
3219 }
3220 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
3221 if (C->failing()) return true;
3222 C->print_method(PHASE_AFTER_MACRO_EXPANSION_STEP, 5, n);
3223
3224 // Clean up the graph so we're less likely to hit the maximum node
3225 // limit
3226 _igvn.set_delay_transform(false);
3227 _igvn.optimize();
3228 if (C->failing()) return true;
3229 _igvn.set_delay_transform(true);
3230 }
3231
3232 _igvn.set_delay_transform(false);
3233 return false;
3234 }
3235
3236 #ifndef PRODUCT
3237 int PhaseMacroExpand::_objs_scalar_replaced_counter = 0;
3238 int PhaseMacroExpand::_monitor_objects_removed_counter = 0;
3239 int PhaseMacroExpand::_GC_barriers_removed_counter = 0;
3240 int PhaseMacroExpand::_memory_barriers_removed_counter = 0;
3241
3242 void PhaseMacroExpand::print_statistics() {
3243 tty->print("Objects scalar replaced = %d, ", Atomic::load(&_objs_scalar_replaced_counter));
3244 tty->print("Monitor objects removed = %d, ", Atomic::load(&_monitor_objects_removed_counter));
3245 tty->print("GC barriers removed = %d, ", Atomic::load(&_GC_barriers_removed_counter));
3246 tty->print_cr("Memory barriers removed = %d", Atomic::load(&_memory_barriers_removed_counter));
3247 }
3248
3249 int PhaseMacroExpand::count_MemBar(Compile *C) {
3250 if (!PrintOptoStatistics) {
3251 return 0;
3252 }
3253 Unique_Node_List ideal_nodes;
3254 int total = 0;
3255 ideal_nodes.map(C->live_nodes(), nullptr);
3256 ideal_nodes.push(C->root());
3257 for (uint next = 0; next < ideal_nodes.size(); ++next) {
3258 Node* n = ideal_nodes.at(next);
3259 if (n->is_MemBar()) {
3260 total++;
3261 }
3262 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3263 Node* m = n->fast_out(i);
3264 ideal_nodes.push(m);
3265 }
3266 }
3267 return total;
3268 }
3269 #endif