1 /*
2 * Copyright (c) 2005, 2025, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "ci/ciFlatArrayKlass.hpp"
26 #include "ci/ciInstanceKlass.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "gc/shared/collectedHeap.inline.hpp"
29 #include "gc/shared/tlab_globals.hpp"
30 #include "libadt/vectset.hpp"
31 #include "memory/universe.hpp"
32 #include "opto/addnode.hpp"
33 #include "opto/arraycopynode.hpp"
34 #include "opto/callnode.hpp"
35 #include "opto/castnode.hpp"
36 #include "opto/cfgnode.hpp"
37 #include "opto/compile.hpp"
38 #include "opto/convertnode.hpp"
39 #include "opto/graphKit.hpp"
40 #include "opto/inlinetypenode.hpp"
41 #include "opto/intrinsicnode.hpp"
42 #include "opto/locknode.hpp"
43 #include "opto/loopnode.hpp"
44 #include "opto/macro.hpp"
45 #include "opto/memnode.hpp"
46 #include "opto/narrowptrnode.hpp"
47 #include "opto/node.hpp"
48 #include "opto/opaquenode.hpp"
49 #include "opto/phaseX.hpp"
50 #include "opto/rootnode.hpp"
51 #include "opto/runtime.hpp"
52 #include "opto/subnode.hpp"
53 #include "opto/subtypenode.hpp"
54 #include "opto/type.hpp"
55 #include "prims/jvmtiExport.hpp"
56 #include "runtime/continuation.hpp"
57 #include "runtime/sharedRuntime.hpp"
58 #include "runtime/stubRoutines.hpp"
59 #include "utilities/globalDefinitions.hpp"
60 #include "utilities/macros.hpp"
61 #include "utilities/powerOfTwo.hpp"
62 #if INCLUDE_G1GC
63 #include "gc/g1/g1ThreadLocalData.hpp"
64 #endif // INCLUDE_G1GC
65
66
67 //
68 // Replace any references to "oldref" in inputs to "use" with "newref".
69 // Returns the number of replacements made.
70 //
71 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
72 int nreplacements = 0;
73 uint req = use->req();
74 for (uint j = 0; j < use->len(); j++) {
75 Node *uin = use->in(j);
76 if (uin == oldref) {
77 if (j < req)
78 use->set_req(j, newref);
79 else
80 use->set_prec(j, newref);
81 nreplacements++;
82 } else if (j >= req && uin == nullptr) {
83 break;
84 }
85 }
86 return nreplacements;
87 }
88
89
90 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word) {
91 Node* cmp = word;
92 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
93 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
94 transform_later(iff);
95
96 // Fast path taken.
97 Node *fast_taken = transform_later(new IfFalseNode(iff));
98
99 // Fast path not-taken, i.e. slow path
100 Node *slow_taken = transform_later(new IfTrueNode(iff));
101
102 region->init_req(edge, fast_taken); // Capture fast-control
103 return slow_taken;
104 }
105
106 //--------------------copy_predefined_input_for_runtime_call--------------------
107 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
108 // Set fixed predefined input arguments
109 call->init_req( TypeFunc::Control, ctrl );
110 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
111 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
112 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
113 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
114 }
115
116 //------------------------------make_slow_call---------------------------------
117 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type,
118 address slow_call, const char* leaf_name, Node* slow_path,
119 Node* parm0, Node* parm1, Node* parm2) {
120
121 // Slow-path call
122 CallNode *call = leaf_name
123 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
124 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), TypeRawPtr::BOTTOM );
125
126 // Slow path call has no side-effects, uses few values
127 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
128 if (parm0 != nullptr) call->init_req(TypeFunc::Parms+0, parm0);
129 if (parm1 != nullptr) call->init_req(TypeFunc::Parms+1, parm1);
130 if (parm2 != nullptr) call->init_req(TypeFunc::Parms+2, parm2);
131 call->copy_call_debug_info(&_igvn, oldcall);
132 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
133 _igvn.replace_node(oldcall, call);
134 transform_later(call);
135
136 return call;
137 }
138
139 void PhaseMacroExpand::eliminate_gc_barrier(Node* p2x) {
140 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
141 bs->eliminate_gc_barrier(&_igvn, p2x);
142 #ifndef PRODUCT
143 if (PrintOptoStatistics) {
144 Atomic::inc(&PhaseMacroExpand::_GC_barriers_removed_counter);
145 }
146 #endif
147 }
148
149 // Search for a memory operation for the specified memory slice.
150 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
151 Node *orig_mem = mem;
152 Node *alloc_mem = alloc->as_Allocate()->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
153 assert(alloc_mem != nullptr, "Allocation without a memory projection.");
154 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
155 while (true) {
156 if (mem == alloc_mem || mem == start_mem ) {
157 return mem; // hit one of our sentinels
158 } else if (mem->is_MergeMem()) {
159 mem = mem->as_MergeMem()->memory_at(alias_idx);
160 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
161 Node *in = mem->in(0);
162 // we can safely skip over safepoints, calls, locks and membars because we
163 // already know that the object is safe to eliminate.
164 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
165 return in;
166 } else if (in->is_Call()) {
167 CallNode *call = in->as_Call();
168 if (call->may_modify(tinst, phase)) {
169 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape");
170 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) {
171 return in;
172 }
173 }
174 mem = in->in(TypeFunc::Memory);
175 } else if (in->is_MemBar()) {
176 ArrayCopyNode* ac = nullptr;
177 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) {
178 if (ac != nullptr) {
179 assert(ac->is_clonebasic(), "Only basic clone is a non escaping clone");
180 return ac;
181 }
182 }
183 mem = in->in(TypeFunc::Memory);
184 } else {
185 #ifdef ASSERT
186 in->dump();
187 mem->dump();
188 assert(false, "unexpected projection");
189 #endif
190 }
191 } else if (mem->is_Store()) {
192 const TypePtr* atype = mem->as_Store()->adr_type();
193 int adr_idx = phase->C->get_alias_index(atype);
194 if (adr_idx == alias_idx) {
195 assert(atype->isa_oopptr(), "address type must be oopptr");
196 int adr_offset = atype->flat_offset();
197 uint adr_iid = atype->is_oopptr()->instance_id();
198 // Array elements references have the same alias_idx
199 // but different offset and different instance_id.
200 if (adr_offset == offset && adr_iid == alloc->_idx) {
201 return mem;
202 }
203 } else {
204 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
205 }
206 mem = mem->in(MemNode::Memory);
207 } else if (mem->is_ClearArray()) {
208 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
209 // Can not bypass initialization of the instance
210 // we are looking.
211 DEBUG_ONLY(intptr_t offset;)
212 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
213 InitializeNode* init = alloc->as_Allocate()->initialization();
214 // We are looking for stored value, return Initialize node
215 // or memory edge from Allocate node.
216 if (init != nullptr) {
217 return init;
218 } else {
219 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
220 }
221 }
222 // Otherwise skip it (the call updated 'mem' value).
223 } else if (mem->Opcode() == Op_SCMemProj) {
224 mem = mem->in(0);
225 Node* adr = nullptr;
226 if (mem->is_LoadStore()) {
227 adr = mem->in(MemNode::Address);
228 } else {
229 assert(mem->Opcode() == Op_EncodeISOArray ||
230 mem->Opcode() == Op_StrCompressedCopy, "sanity");
231 adr = mem->in(3); // Destination array
232 }
233 const TypePtr* atype = adr->bottom_type()->is_ptr();
234 int adr_idx = phase->C->get_alias_index(atype);
235 if (adr_idx == alias_idx) {
236 DEBUG_ONLY(mem->dump();)
237 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
238 return nullptr;
239 }
240 mem = mem->in(MemNode::Memory);
241 } else if (mem->Opcode() == Op_StrInflatedCopy) {
242 Node* adr = mem->in(3); // Destination array
243 const TypePtr* atype = adr->bottom_type()->is_ptr();
244 int adr_idx = phase->C->get_alias_index(atype);
245 if (adr_idx == alias_idx) {
246 DEBUG_ONLY(mem->dump();)
247 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
248 return nullptr;
249 }
250 mem = mem->in(MemNode::Memory);
251 } else {
252 return mem;
253 }
254 assert(mem != orig_mem, "dead memory loop");
255 }
256 }
257
258 // Generate loads from source of the arraycopy for fields of
259 // destination needed at a deoptimization point
260 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) {
261 BasicType bt = ft;
262 const Type *type = ftype;
263 if (ft == T_NARROWOOP) {
264 bt = T_OBJECT;
265 type = ftype->make_oopptr();
266 }
267 Node* res = nullptr;
268 if (ac->is_clonebasic()) {
269 assert(ac->in(ArrayCopyNode::Src) != ac->in(ArrayCopyNode::Dest), "clone source equals destination");
270 Node* base = ac->in(ArrayCopyNode::Src);
271 Node* adr = _igvn.transform(new AddPNode(base, base, _igvn.MakeConX(offset)));
272 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
273 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
274 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
275 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
276 } else {
277 if (ac->modifies(offset, offset, &_igvn, true)) {
278 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
279 uint shift = exact_log2(type2aelembytes(bt));
280 Node* src_pos = ac->in(ArrayCopyNode::SrcPos);
281 Node* dest_pos = ac->in(ArrayCopyNode::DestPos);
282 const TypeInt* src_pos_t = _igvn.type(src_pos)->is_int();
283 const TypeInt* dest_pos_t = _igvn.type(dest_pos)->is_int();
284
285 Node* adr = nullptr;
286 Node* base = ac->in(ArrayCopyNode::Src);
287 const TypeAryPtr* adr_type = _igvn.type(base)->is_aryptr();
288 if (adr_type->is_flat()) {
289 shift = adr_type->flat_log_elem_size();
290 }
291 if (src_pos_t->is_con() && dest_pos_t->is_con()) {
292 intptr_t off = ((src_pos_t->get_con() - dest_pos_t->get_con()) << shift) + offset;
293 adr = _igvn.transform(new AddPNode(base, base, _igvn.MakeConX(off)));
294 adr_type = _igvn.type(adr)->is_aryptr();
295 assert(adr_type == _igvn.type(base)->is_aryptr()->add_field_offset_and_offset(off), "incorrect address type");
296 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
297 // Don't emit a new load from src if src == dst but try to get the value from memory instead
298 return value_from_mem(ac->in(TypeFunc::Memory), ctl, ft, ftype, adr_type, alloc);
299 }
300 } else {
301 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
302 // Non constant offset in the array: we can't statically
303 // determine the value
304 return nullptr;
305 }
306 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
307 #ifdef _LP64
308 diff = _igvn.transform(new ConvI2LNode(diff));
309 #endif
310 diff = _igvn.transform(new LShiftXNode(diff, _igvn.intcon(shift)));
311
312 Node* off = _igvn.transform(new AddXNode(_igvn.MakeConX(offset), diff));
313 adr = _igvn.transform(new AddPNode(base, base, off));
314 // In the case of a flat inline type array, each field has its
315 // own slice so we need to extract the field being accessed from
316 // the address computation
317 adr_type = adr_type->add_field_offset_and_offset(offset)->add_offset(Type::OffsetBot)->is_aryptr();
318 adr = _igvn.transform(new CastPPNode(ctl, adr, adr_type));
319 }
320 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
321 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
322 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
323 }
324 }
325 if (res != nullptr) {
326 if (ftype->isa_narrowoop()) {
327 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
328 assert(res->isa_DecodeN(), "should be narrow oop");
329 res = _igvn.transform(new EncodePNode(res, ftype));
330 }
331 return res;
332 }
333 return nullptr;
334 }
335
336 //
337 // Given a Memory Phi, compute a value Phi containing the values from stores
338 // on the input paths.
339 // Note: this function is recursive, its depth is limited by the "level" argument
340 // Returns the computed Phi, or null if it cannot compute it.
341 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) {
342 assert(mem->is_Phi(), "sanity");
343 int alias_idx = C->get_alias_index(adr_t);
344 int offset = adr_t->flat_offset();
345 int instance_id = adr_t->instance_id();
346
347 // Check if an appropriate value phi already exists.
348 Node* region = mem->in(0);
349 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
350 Node* phi = region->fast_out(k);
351 if (phi->is_Phi() && phi != mem &&
352 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
353 return phi;
354 }
355 }
356 // Check if an appropriate new value phi already exists.
357 Node* new_phi = value_phis->find(mem->_idx);
358 if (new_phi != nullptr)
359 return new_phi;
360
361 if (level <= 0) {
362 return nullptr; // Give up: phi tree too deep
363 }
364 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
365 Node *alloc_mem = alloc->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
366 assert(alloc_mem != nullptr, "Allocation without a memory projection.");
367
368 uint length = mem->req();
369 GrowableArray <Node *> values(length, length, nullptr);
370
371 // create a new Phi for the value
372 PhiNode *phi = new PhiNode(mem->in(0), phi_type, nullptr, mem->_idx, instance_id, alias_idx, offset);
373 transform_later(phi);
374 value_phis->push(phi, mem->_idx);
375
376 for (uint j = 1; j < length; j++) {
377 Node *in = mem->in(j);
378 if (in == nullptr || in->is_top()) {
379 values.at_put(j, in);
380 } else {
381 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
382 if (val == start_mem || val == alloc_mem) {
383 // hit a sentinel, return appropriate value
384 Node* init_value = alloc->in(AllocateNode::InitValue);
385 if (init_value != nullptr) {
386 if (val == start_mem) {
387 // TODO 8350865 Scalar replacement does not work well for flat arrays.
388 // Somehow we ended up with root mem and therefore walked past the alloc. Fix this. Triggered by TestGenerated::test15
389 // Don't we need field_value_by_offset?
390 return nullptr;
391 }
392 values.at_put(j, init_value);
393 } else {
394 assert(alloc->in(AllocateNode::RawInitValue) == nullptr, "init value may not be null");
395 values.at_put(j, _igvn.zerocon(ft));
396 }
397 continue;
398 }
399 if (val->is_Initialize()) {
400 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
401 }
402 if (val == nullptr) {
403 return nullptr; // can't find a value on this path
404 }
405 if (val == mem) {
406 values.at_put(j, mem);
407 } else if (val->is_Store()) {
408 Node* n = val->in(MemNode::ValueIn);
409 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
410 n = bs->step_over_gc_barrier(n);
411 if (is_subword_type(ft)) {
412 n = Compile::narrow_value(ft, n, phi_type, &_igvn, true);
413 }
414 values.at_put(j, n);
415 } else if (val->is_Proj() && val->in(0) == alloc) {
416 Node* init_value = alloc->in(AllocateNode::InitValue);
417 if (init_value != nullptr) {
418 // TODO 8350865 Scalar replacement does not work well for flat arrays.
419 // Is this correct for non-all-zero init values? Don't we need field_value_by_offset?
420 values.at_put(j, init_value);
421 } else {
422 assert(alloc->in(AllocateNode::RawInitValue) == nullptr, "init value may not be null");
423 values.at_put(j, _igvn.zerocon(ft));
424 }
425 } else if (val->is_Phi()) {
426 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
427 if (val == nullptr) {
428 return nullptr;
429 }
430 values.at_put(j, val);
431 } else if (val->Opcode() == Op_SCMemProj) {
432 assert(val->in(0)->is_LoadStore() ||
433 val->in(0)->Opcode() == Op_EncodeISOArray ||
434 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
435 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
436 return nullptr;
437 } else if (val->is_ArrayCopy()) {
438 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc);
439 if (res == nullptr) {
440 return nullptr;
441 }
442 values.at_put(j, res);
443 } else if (val->is_top()) {
444 // This indicates that this path into the phi is dead. Top will eventually also propagate into the Region.
445 // IGVN will clean this up later.
446 values.at_put(j, val);
447 } else {
448 DEBUG_ONLY( val->dump(); )
449 assert(false, "unknown node on this path");
450 return nullptr; // unknown node on this path
451 }
452 }
453 }
454 // Set Phi's inputs
455 for (uint j = 1; j < length; j++) {
456 if (values.at(j) == mem) {
457 phi->init_req(j, phi);
458 } else {
459 phi->init_req(j, values.at(j));
460 }
461 }
462 return phi;
463 }
464
465 // Search the last value stored into the object's field.
466 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
467 assert(adr_t->is_known_instance_field(), "instance required");
468 int instance_id = adr_t->instance_id();
469 assert((uint)instance_id == alloc->_idx, "wrong allocation");
470
471 int alias_idx = C->get_alias_index(adr_t);
472 int offset = adr_t->flat_offset();
473 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
474 Node *alloc_mem = alloc->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
475 assert(alloc_mem != nullptr, "Allocation without a memory projection.");
476 VectorSet visited;
477
478 bool done = sfpt_mem == alloc_mem;
479 Node *mem = sfpt_mem;
480 while (!done) {
481 if (visited.test_set(mem->_idx)) {
482 return nullptr; // found a loop, give up
483 }
484 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
485 if (mem == start_mem || mem == alloc_mem) {
486 done = true; // hit a sentinel, return appropriate 0 value
487 } else if (mem->is_Initialize()) {
488 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
489 if (mem == nullptr) {
490 done = true; // Something went wrong.
491 } else if (mem->is_Store()) {
492 const TypePtr* atype = mem->as_Store()->adr_type();
493 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
494 done = true;
495 }
496 } else if (mem->is_Store()) {
497 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
498 assert(atype != nullptr, "address type must be oopptr");
499 assert(C->get_alias_index(atype) == alias_idx &&
500 atype->is_known_instance_field() && atype->flat_offset() == offset &&
501 atype->instance_id() == instance_id, "store is correct memory slice");
502 done = true;
503 } else if (mem->is_Phi()) {
504 // try to find a phi's unique input
505 Node *unique_input = nullptr;
506 Node *top = C->top();
507 for (uint i = 1; i < mem->req(); i++) {
508 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
509 if (n == nullptr || n == top || n == mem) {
510 continue;
511 } else if (unique_input == nullptr) {
512 unique_input = n;
513 } else if (unique_input != n) {
514 unique_input = top;
515 break;
516 }
517 }
518 if (unique_input != nullptr && unique_input != top) {
519 mem = unique_input;
520 } else {
521 done = true;
522 }
523 } else if (mem->is_ArrayCopy()) {
524 done = true;
525 } else {
526 DEBUG_ONLY( mem->dump(); )
527 assert(false, "unexpected node");
528 }
529 }
530 if (mem != nullptr) {
531 if (mem == start_mem || mem == alloc_mem) {
532 // hit a sentinel, return appropriate value
533 Node* init_value = alloc->in(AllocateNode::InitValue);
534 if (init_value != nullptr) {
535 if (adr_t->is_flat()) {
536 if (init_value->is_EncodeP()) {
537 init_value = init_value->in(1);
538 }
539 assert(adr_t->is_aryptr()->field_offset().get() != Type::OffsetBot, "Unknown offset");
540 offset = adr_t->is_aryptr()->field_offset().get() + init_value->bottom_type()->inline_klass()->payload_offset();
541 init_value = init_value->as_InlineType()->field_value_by_offset(offset, true);
542 if (ft == T_NARROWOOP) {
543 init_value = transform_later(new EncodePNode(init_value, init_value->bottom_type()->make_ptr()));
544 }
545 }
546 return init_value;
547 }
548 assert(alloc->in(AllocateNode::RawInitValue) == nullptr, "init value may not be null");
549 return _igvn.zerocon(ft);
550 } else if (mem->is_Store()) {
551 Node* n = mem->in(MemNode::ValueIn);
552 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
553 n = bs->step_over_gc_barrier(n);
554 return n;
555 } else if (mem->is_Phi()) {
556 // attempt to produce a Phi reflecting the values on the input paths of the Phi
557 Node_Stack value_phis(8);
558 Node* phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
559 if (phi != nullptr) {
560 return phi;
561 } else {
562 // Kill all new Phis
563 while(value_phis.is_nonempty()) {
564 Node* n = value_phis.node();
565 _igvn.replace_node(n, C->top());
566 value_phis.pop();
567 }
568 }
569 } else if (mem->is_ArrayCopy()) {
570 Node* ctl = mem->in(0);
571 Node* m = mem->in(TypeFunc::Memory);
572 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj()) {
573 // pin the loads in the uncommon trap path
574 ctl = sfpt_ctl;
575 m = sfpt_mem;
576 }
577 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc);
578 }
579 }
580 // Something went wrong.
581 return nullptr;
582 }
583
584 // Search the last value stored into the inline type's fields (for flat arrays).
585 Node* PhaseMacroExpand::inline_type_from_mem(ciInlineKlass* vk, const TypeAryPtr* elem_adr_type, int elem_idx, int offset_in_element, bool null_free, AllocateNode* alloc, SafePointNode* sfpt) {
586 auto report_failure = [&](int field_offset_in_element) {
587 #ifndef PRODUCT
588 if (PrintEliminateAllocations) {
589 ciInlineKlass* elem_klass = elem_adr_type->elem()->inline_klass();
590 int offset = field_offset_in_element + elem_klass->payload_offset();
591 ciField* flattened_field = elem_klass->get_field_by_offset(offset, false);
592 assert(flattened_field != nullptr, "must have a field of type %s at offset %d", elem_klass->name()->as_utf8(), offset);
593 tty->print("=== At SafePoint node %d can't find value of field [%s] of array element [%d]", sfpt->_idx, flattened_field->name()->as_utf8(), elem_idx);
594 tty->print(", which prevents elimination of: ");
595 alloc->dump();
596 }
597 #endif // PRODUCT
598 };
599
600 // Create a new InlineTypeNode and retrieve the field values from memory
601 InlineTypeNode* vt = InlineTypeNode::make_uninitialized(_igvn, vk, false);
602 transform_later(vt);
603 if (null_free) {
604 vt->set_null_marker(_igvn);
605 } else {
606 int nm_offset_in_element = offset_in_element + vk->null_marker_offset_in_payload();
607 const TypeAryPtr* nm_adr_type = elem_adr_type->with_field_offset(nm_offset_in_element);
608 Node* nm_value = value_from_mem(sfpt->memory(), sfpt->control(), T_BOOLEAN, TypeInt::BOOL, nm_adr_type, alloc);
609 if (nm_value != nullptr) {
610 vt->set_null_marker(_igvn, nm_value);
611 } else {
612 report_failure(nm_offset_in_element);
613 return nullptr;
614 }
615 }
616
617 for (int i = 0; i < vk->nof_declared_nonstatic_fields(); ++i) {
618 ciType* field_type = vt->field_type(i);
619 int field_offset_in_element = offset_in_element + vt->field_offset(i) - vk->payload_offset();
620 Node* field_value = nullptr;
621 if (vt->field_is_flat(i)) {
622 field_value = inline_type_from_mem(field_type->as_inline_klass(), elem_adr_type, elem_idx, field_offset_in_element, vt->field_is_null_free(i), alloc, sfpt);
623 } else {
624 const Type* ft = Type::get_const_type(field_type);
625 BasicType bt = type2field[field_type->basic_type()];
626 if (UseCompressedOops && !is_java_primitive(bt)) {
627 ft = ft->make_narrowoop();
628 bt = T_NARROWOOP;
629 }
630 // Each inline type field has its own memory slice
631 const TypeAryPtr* field_adr_type = elem_adr_type->with_field_offset(field_offset_in_element);
632 field_value = value_from_mem(sfpt->memory(), sfpt->control(), bt, ft, field_adr_type, alloc);
633 if (field_value == nullptr) {
634 report_failure(field_offset_in_element);
635 } else if (ft->isa_narrowoop()) {
636 assert(UseCompressedOops, "unexpected narrow oop");
637 if (field_value->is_EncodeP()) {
638 field_value = field_value->in(1);
639 } else if (!field_value->is_InlineType()) {
640 field_value = transform_later(new DecodeNNode(field_value, field_value->get_ptr_type()));
641 }
642 }
643 }
644 if (field_value != nullptr) {
645 vt->set_field_value(i, field_value);
646 } else {
647 return nullptr;
648 }
649 }
650 return vt;
651 }
652
653 // Check the possibility of scalar replacement.
654 bool PhaseMacroExpand::can_eliminate_allocation(PhaseIterGVN* igvn, AllocateNode *alloc, GrowableArray <SafePointNode *>* safepoints) {
655 // Scan the uses of the allocation to check for anything that would
656 // prevent us from eliminating it.
657 NOT_PRODUCT( const char* fail_eliminate = nullptr; )
658 DEBUG_ONLY( Node* disq_node = nullptr; )
659 bool can_eliminate = true;
660 bool reduce_merge_precheck = (safepoints == nullptr);
661
662 Unique_Node_List worklist;
663 Node* res = alloc->result_cast();
664 const TypeOopPtr* res_type = nullptr;
665 if (res == nullptr) {
666 // All users were eliminated.
667 } else if (!res->is_CheckCastPP()) {
668 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
669 can_eliminate = false;
670 } else {
671 worklist.push(res);
672 res_type = igvn->type(res)->isa_oopptr();
673 if (res_type == nullptr) {
674 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
675 can_eliminate = false;
676 } else if (!res_type->klass_is_exact()) {
677 NOT_PRODUCT(fail_eliminate = "Not an exact type.";)
678 can_eliminate = false;
679 } else if (res_type->isa_aryptr()) {
680 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
681 if (length < 0) {
682 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
683 can_eliminate = false;
684 }
685 }
686 }
687
688 while (can_eliminate && worklist.size() > 0) {
689 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
690 res = worklist.pop();
691 for (DUIterator_Fast jmax, j = res->fast_outs(jmax); j < jmax && can_eliminate; j++) {
692 Node* use = res->fast_out(j);
693
694 if (use->is_AddP()) {
695 const TypePtr* addp_type = igvn->type(use)->is_ptr();
696 int offset = addp_type->offset();
697
698 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
699 NOT_PRODUCT(fail_eliminate = "Undefined field reference";)
700 can_eliminate = false;
701 break;
702 }
703 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
704 k < kmax && can_eliminate; k++) {
705 Node* n = use->fast_out(k);
706 if (!n->is_Store() && n->Opcode() != Op_CastP2X && !bs->is_gc_pre_barrier_node(n) && !reduce_merge_precheck) {
707 DEBUG_ONLY(disq_node = n;)
708 if (n->is_Load() || n->is_LoadStore()) {
709 NOT_PRODUCT(fail_eliminate = "Field load";)
710 } else {
711 NOT_PRODUCT(fail_eliminate = "Not store field reference";)
712 }
713 can_eliminate = false;
714 }
715 }
716 } else if (use->is_ArrayCopy() &&
717 (use->as_ArrayCopy()->is_clonebasic() ||
718 use->as_ArrayCopy()->is_arraycopy_validated() ||
719 use->as_ArrayCopy()->is_copyof_validated() ||
720 use->as_ArrayCopy()->is_copyofrange_validated()) &&
721 use->in(ArrayCopyNode::Dest) == res) {
722 // ok to eliminate
723 } else if (use->is_SafePoint()) {
724 SafePointNode* sfpt = use->as_SafePoint();
725 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
726 // Object is passed as argument.
727 DEBUG_ONLY(disq_node = use;)
728 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
729 can_eliminate = false;
730 }
731 Node* sfptMem = sfpt->memory();
732 if (sfptMem == nullptr || sfptMem->is_top()) {
733 DEBUG_ONLY(disq_node = use;)
734 NOT_PRODUCT(fail_eliminate = "null or TOP memory";)
735 can_eliminate = false;
736 } else if (!reduce_merge_precheck) {
737 assert(!res->is_Phi() || !res->as_Phi()->can_be_inline_type(), "Inline type allocations should not have safepoint uses");
738 safepoints->append_if_missing(sfpt);
739 }
740 } else if (use->is_InlineType() && use->as_InlineType()->get_oop() == res) {
741 // Look at uses
742 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); k < kmax; k++) {
743 Node* u = use->fast_out(k);
744 if (u->is_InlineType()) {
745 // Use in flat field can be eliminated
746 InlineTypeNode* vt = u->as_InlineType();
747 for (uint i = 0; i < vt->field_count(); ++i) {
748 if (vt->field_value(i) == use && !vt->field_is_flat(i)) {
749 can_eliminate = false; // Use in non-flat field
750 break;
751 }
752 }
753 } else {
754 // Add other uses to the worklist to process individually
755 worklist.push(use);
756 }
757 }
758 } else if (use->Opcode() == Op_StoreX && use->in(MemNode::Address) == res) {
759 // Store to mark word of inline type larval buffer
760 assert(res_type->is_inlinetypeptr(), "Unexpected store to mark word");
761 } else if (res_type->is_inlinetypeptr() && (use->Opcode() == Op_MemBarRelease || use->Opcode() == Op_MemBarStoreStore)) {
762 // Inline type buffer allocations are followed by a membar
763 } else if (reduce_merge_precheck &&
764 (use->is_Phi() || use->is_EncodeP() ||
765 use->Opcode() == Op_MemBarRelease ||
766 (UseStoreStoreForCtor && use->Opcode() == Op_MemBarStoreStore))) {
767 // Nothing to do
768 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
769 if (use->is_Phi()) {
770 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
771 NOT_PRODUCT(fail_eliminate = "Object is return value";)
772 } else {
773 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
774 }
775 DEBUG_ONLY(disq_node = use;)
776 } else {
777 if (use->Opcode() == Op_Return) {
778 NOT_PRODUCT(fail_eliminate = "Object is return value";)
779 } else {
780 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
781 }
782 DEBUG_ONLY(disq_node = use;)
783 }
784 can_eliminate = false;
785 } else {
786 assert(use->Opcode() == Op_CastP2X, "should be");
787 assert(!use->has_out_with(Op_OrL), "should have been removed because oop is never null");
788 }
789 }
790 }
791
792 #ifndef PRODUCT
793 if (PrintEliminateAllocations && safepoints != nullptr) {
794 if (can_eliminate) {
795 tty->print("Scalar ");
796 if (res == nullptr)
797 alloc->dump();
798 else
799 res->dump();
800 } else {
801 tty->print("NotScalar (%s)", fail_eliminate);
802 if (res == nullptr)
803 alloc->dump();
804 else
805 res->dump();
806 #ifdef ASSERT
807 if (disq_node != nullptr) {
808 tty->print(" >>>> ");
809 disq_node->dump();
810 }
811 #endif /*ASSERT*/
812 }
813 }
814
815 if (TraceReduceAllocationMerges && !can_eliminate && reduce_merge_precheck) {
816 tty->print_cr("\tCan't eliminate allocation because '%s': ", fail_eliminate != nullptr ? fail_eliminate : "");
817 DEBUG_ONLY(if (disq_node != nullptr) disq_node->dump();)
818 }
819 #endif
820 return can_eliminate;
821 }
822
823 void PhaseMacroExpand::undo_previous_scalarizations(GrowableArray <SafePointNode *> safepoints_done, AllocateNode* alloc) {
824 Node* res = alloc->result_cast();
825 int nfields = 0;
826 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
827
828 if (res != nullptr) {
829 const TypeOopPtr* res_type = _igvn.type(res)->isa_oopptr();
830
831 if (res_type->isa_instptr()) {
832 // find the fields of the class which will be needed for safepoint debug information
833 ciInstanceKlass* iklass = res_type->is_instptr()->instance_klass();
834 nfields = iklass->nof_nonstatic_fields();
835 } else {
836 // find the array's elements which will be needed for safepoint debug information
837 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
838 assert(nfields >= 0, "must be an array klass.");
839 }
840 }
841
842 // rollback processed safepoints
843 while (safepoints_done.length() > 0) {
844 SafePointNode* sfpt_done = safepoints_done.pop();
845 // remove any extra entries we added to the safepoint
846 uint last = sfpt_done->req() - 1;
847 for (int k = 0; k < nfields; k++) {
848 sfpt_done->del_req(last--);
849 }
850 JVMState *jvms = sfpt_done->jvms();
851 jvms->set_endoff(sfpt_done->req());
852 // Now make a pass over the debug information replacing any references
853 // to SafePointScalarObjectNode with the allocated object.
854 int start = jvms->debug_start();
855 int end = jvms->debug_end();
856 for (int i = start; i < end; i++) {
857 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
858 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
859 if (scobj->first_index(jvms) == sfpt_done->req() &&
860 scobj->n_fields() == (uint)nfields) {
861 assert(scobj->alloc() == alloc, "sanity");
862 sfpt_done->set_req(i, res);
863 }
864 }
865 }
866 _igvn._worklist.push(sfpt_done);
867 }
868 }
869
870 void PhaseMacroExpand::process_field_value_at_safepoint(const Type* field_type, Node* field_val, SafePointNode* sfpt, Unique_Node_List* value_worklist) {
871 if (UseCompressedOops && field_type->isa_narrowoop()) {
872 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
873 // to be able scalar replace the allocation.
874 if (field_val->is_EncodeP()) {
875 field_val = field_val->in(1);
876 } else if (!field_val->is_InlineType()) {
877 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
878 }
879 }
880
881 // Keep track of inline types to scalarize them later
882 if (field_val->is_InlineType()) {
883 value_worklist->push(field_val);
884 } else if (field_val->is_Phi()) {
885 PhiNode* phi = field_val->as_Phi();
886 // Eagerly replace inline type phis now since we could be removing an inline type allocation where we must
887 // scalarize all its fields in safepoints.
888 field_val = phi->try_push_inline_types_down(&_igvn, true);
889 if (field_val->is_InlineType()) {
890 value_worklist->push(field_val);
891 }
892 }
893 sfpt->add_req(field_val);
894 }
895
896 bool PhaseMacroExpand::add_array_elems_to_safepoint(AllocateNode* alloc, const TypeAryPtr* array_type, SafePointNode* sfpt, Unique_Node_List* value_worklist) {
897 const Type* elem_type = array_type->elem();
898 BasicType basic_elem_type = elem_type->array_element_basic_type();
899
900 intptr_t elem_size;
901 uint header_size;
902 if (array_type->is_flat()) {
903 elem_size = array_type->flat_elem_size();
904 header_size = arrayOopDesc::base_offset_in_bytes(T_FLAT_ELEMENT);
905 } else {
906 elem_size = type2aelembytes(basic_elem_type);
907 header_size = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
908 }
909
910 int n_elems = alloc->in(AllocateNode::ALength)->get_int();
911 for (int elem_idx = 0; elem_idx < n_elems; elem_idx++) {
912 intptr_t elem_offset = header_size + elem_idx * elem_size;
913 const TypeAryPtr* elem_adr_type = array_type->with_offset(elem_offset);
914 Node* elem_val;
915 if (array_type->is_flat()) {
916 ciInlineKlass* elem_klass = elem_type->inline_klass();
917 assert(elem_klass->maybe_flat_in_array(), "must be flat in array");
918 elem_val = inline_type_from_mem(elem_klass, elem_adr_type, elem_idx, 0, array_type->is_null_free(), alloc, sfpt);
919 } else {
920 elem_val = value_from_mem(sfpt->memory(), sfpt->control(), basic_elem_type, elem_type, elem_adr_type, alloc);
921 #ifndef PRODUCT
922 if (PrintEliminateAllocations && elem_val == nullptr) {
923 tty->print("=== At SafePoint node %d can't find value of array element [%d]", sfpt->_idx, elem_idx);
924 tty->print(", which prevents elimination of: ");
925 alloc->dump();
926 }
927 #endif // PRODUCT
928 }
929 if (elem_val == nullptr) {
930 return false;
931 }
932
933 process_field_value_at_safepoint(elem_type, elem_val, sfpt, value_worklist);
934 }
935
936 return true;
937 }
938
939 // Recursively adds all flattened fields of a type 'iklass' inside 'base' to 'sfpt'.
940 // 'offset_minus_header' refers to the offset of the payload of 'iklass' inside 'base' minus the
941 // payload offset of 'iklass'. If 'base' is of type 'iklass' then 'offset_minus_header' == 0.
942 bool PhaseMacroExpand::add_inst_fields_to_safepoint(ciInstanceKlass* iklass, AllocateNode* alloc, Node* base, int offset_minus_header, SafePointNode* sfpt, Unique_Node_List* value_worklist) {
943 const TypeInstPtr* base_type = _igvn.type(base)->is_instptr();
944 auto report_failure = [&](int offset) {
945 #ifndef PRODUCT
946 if (PrintEliminateAllocations) {
947 ciInstanceKlass* base_klass = base_type->instance_klass();
948 ciField* flattened_field = base_klass->get_field_by_offset(offset, false);
949 assert(flattened_field != nullptr, "must have a field of type %s at offset %d", base_klass->name()->as_utf8(), offset);
950 tty->print("=== At SafePoint node %d can't find value of field: ", sfpt->_idx);
951 flattened_field->print();
952 int field_idx = C->alias_type(flattened_field)->index();
953 tty->print(" (alias_idx=%d)", field_idx);
954 tty->print(", which prevents elimination of: ");
955 base->dump();
956 }
957 #endif // PRODUCT
958 };
959
960 for (int i = 0; i < iklass->nof_declared_nonstatic_fields(); i++) {
961 ciField* field = iklass->declared_nonstatic_field_at(i);
962 if (field->is_flat()) {
963 ciInlineKlass* fvk = field->type()->as_inline_klass();
964 int field_offset_minus_header = offset_minus_header + field->offset_in_bytes() - fvk->payload_offset();
965 bool success = add_inst_fields_to_safepoint(fvk, alloc, base, field_offset_minus_header, sfpt, value_worklist);
966 if (!success) {
967 return false;
968 }
969
970 // The null marker of a field is added right after we scalarize that field
971 if (!field->is_null_free()) {
972 int nm_offset = offset_minus_header + field->null_marker_offset();
973 Node* null_marker = value_from_mem(sfpt->memory(), sfpt->control(), T_BOOLEAN, TypeInt::BOOL, base_type->with_offset(nm_offset), alloc);
974 if (null_marker == nullptr) {
975 report_failure(nm_offset);
976 return false;
977 }
978 process_field_value_at_safepoint(TypeInt::BOOL, null_marker, sfpt, value_worklist);
979 }
980
981 continue;
982 }
983
984 int offset = offset_minus_header + field->offset_in_bytes();
985 ciType* elem_type = field->type();
986 BasicType basic_elem_type = field->layout_type();
987
988 const Type* field_type;
989 if (is_reference_type(basic_elem_type)) {
990 if (!elem_type->is_loaded()) {
991 field_type = TypeInstPtr::BOTTOM;
992 } else {
993 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
994 }
995 if (UseCompressedOops) {
996 field_type = field_type->make_narrowoop();
997 basic_elem_type = T_NARROWOOP;
998 }
999 } else {
1000 field_type = Type::get_const_basic_type(basic_elem_type);
1001 }
1002
1003 const TypeInstPtr* field_addr_type = base_type->add_offset(offset)->isa_instptr();
1004 Node* field_val = value_from_mem(sfpt->memory(), sfpt->control(), basic_elem_type, field_type, field_addr_type, alloc);
1005 if (field_val == nullptr) {
1006 report_failure(offset);
1007 return false;
1008 }
1009 process_field_value_at_safepoint(field_type, field_val, sfpt, value_worklist);
1010 }
1011
1012 return true;
1013 }
1014
1015 SafePointScalarObjectNode* PhaseMacroExpand::create_scalarized_object_description(AllocateNode* alloc, SafePointNode* sfpt,
1016 Unique_Node_List* value_worklist) {
1017 // Fields of scalar objs are referenced only at the end
1018 // of regular debuginfo at the last (youngest) JVMS.
1019 // Record relative start index.
1020 ciInstanceKlass* iklass = nullptr;
1021 const TypeOopPtr* res_type = nullptr;
1022 int nfields = 0;
1023 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
1024 Node* res = alloc->result_cast();
1025
1026 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
1027 assert(sfpt->jvms() != nullptr, "missed JVMS");
1028 uint before_sfpt_req = sfpt->req();
1029
1030 if (res != nullptr) { // Could be null when there are no users
1031 res_type = _igvn.type(res)->isa_oopptr();
1032
1033 if (res_type->isa_instptr()) {
1034 // find the fields of the class which will be needed for safepoint debug information
1035 iklass = res_type->is_instptr()->instance_klass();
1036 nfields = iklass->nof_nonstatic_fields();
1037 } else {
1038 // find the array's elements which will be needed for safepoint debug information
1039 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
1040 assert(nfields >= 0, "must be an array klass.");
1041 }
1042
1043 if (res->bottom_type()->is_inlinetypeptr()) {
1044 // Nullable inline types have a null marker field which is added to the safepoint when scalarizing them (see
1045 // InlineTypeNode::make_scalar_in_safepoint()). When having circular inline types, we stop scalarizing at depth 1
1046 // to avoid an endless recursion. Therefore, we do not have a SafePointScalarObjectNode node here, yet.
1047 // We are about to create a SafePointScalarObjectNode as if this is a normal object. Add an additional int input
1048 // with value 1 which sets the null marker to true to indicate that the object is always non-null. This input is checked
1049 // later in PhaseOutput::filLocArray() for inline types.
1050 sfpt->add_req(_igvn.intcon(1));
1051 }
1052 }
1053
1054 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type, alloc, first_ind, sfpt->jvms()->depth(), nfields);
1055 sobj->init_req(0, C->root());
1056 transform_later(sobj);
1057
1058 if (res == nullptr) {
1059 sfpt->jvms()->set_endoff(sfpt->req());
1060 return sobj;
1061 }
1062
1063 bool success;
1064 if (iklass == nullptr) {
1065 success = add_array_elems_to_safepoint(alloc, res_type->is_aryptr(), sfpt, value_worklist);
1066 } else {
1067 success = add_inst_fields_to_safepoint(iklass, alloc, res, 0, sfpt, value_worklist);
1068 }
1069
1070 // We weren't able to find a value for this field, remove all the fields added to the safepoint
1071 if (!success) {
1072 for (uint i = sfpt->req() - 1; i >= before_sfpt_req; i--) {
1073 sfpt->del_req(i);
1074 }
1075 _igvn._worklist.push(sfpt);
1076 return nullptr;
1077 }
1078
1079 sfpt->jvms()->set_endoff(sfpt->req());
1080 return sobj;
1081 }
1082
1083 // Do scalar replacement.
1084 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
1085 GrowableArray <SafePointNode *> safepoints_done;
1086 Node* res = alloc->result_cast();
1087 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
1088 const TypeOopPtr* res_type = nullptr;
1089 if (res != nullptr) { // Could be null when there are no users
1090 res_type = _igvn.type(res)->isa_oopptr();
1091 }
1092
1093 // Process the safepoint uses
1094 Unique_Node_List value_worklist;
1095 while (safepoints.length() > 0) {
1096 SafePointNode* sfpt = safepoints.pop();
1097 SafePointScalarObjectNode* sobj = create_scalarized_object_description(alloc, sfpt, &value_worklist);
1098
1099 if (sobj == nullptr) {
1100 undo_previous_scalarizations(safepoints_done, alloc);
1101 return false;
1102 }
1103
1104 // Now make a pass over the debug information replacing any references
1105 // to the allocated object with "sobj"
1106 JVMState *jvms = sfpt->jvms();
1107 sfpt->replace_edges_in_range(res, sobj, jvms->debug_start(), jvms->debug_end(), &_igvn);
1108 _igvn._worklist.push(sfpt);
1109
1110 // keep it for rollback
1111 safepoints_done.append_if_missing(sfpt);
1112 }
1113 // Scalarize inline types that were added to the safepoint.
1114 // Don't allow linking a constant oop (if available) for flat array elements
1115 // because Deoptimization::reassign_flat_array_elements needs field values.
1116 bool allow_oop = (res_type != nullptr) && !res_type->is_flat();
1117 for (uint i = 0; i < value_worklist.size(); ++i) {
1118 InlineTypeNode* vt = value_worklist.at(i)->as_InlineType();
1119 vt->make_scalar_in_safepoints(&_igvn, allow_oop);
1120 }
1121 return true;
1122 }
1123
1124 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
1125 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
1126 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
1127 if (ctl_proj != nullptr) {
1128 igvn.replace_node(ctl_proj, n->in(0));
1129 }
1130 if (mem_proj != nullptr) {
1131 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
1132 }
1133 }
1134
1135 // Process users of eliminated allocation.
1136 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc, bool inline_alloc) {
1137 Unique_Node_List worklist;
1138 Node* res = alloc->result_cast();
1139 if (res != nullptr) {
1140 worklist.push(res);
1141 }
1142 while (worklist.size() > 0) {
1143 res = worklist.pop();
1144 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
1145 Node *use = res->last_out(j);
1146 uint oc1 = res->outcnt();
1147
1148 if (use->is_AddP()) {
1149 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
1150 Node *n = use->last_out(k);
1151 uint oc2 = use->outcnt();
1152 if (n->is_Store()) {
1153 for (DUIterator_Fast pmax, p = n->fast_outs(pmax); p < pmax; p++) {
1154 MemBarNode* mb = n->fast_out(p)->isa_MemBar();
1155 if (mb != nullptr && mb->req() <= MemBarNode::Precedent && mb->in(MemBarNode::Precedent) == n) {
1156 // MemBarVolatiles should have been removed by MemBarNode::Ideal() for non-inline allocations
1157 assert(inline_alloc, "MemBarVolatile should be eliminated for non-escaping object");
1158 mb->remove(&_igvn);
1159 }
1160 }
1161 _igvn.replace_node(n, n->in(MemNode::Memory));
1162 } else {
1163 eliminate_gc_barrier(n);
1164 }
1165 k -= (oc2 - use->outcnt());
1166 }
1167 _igvn.remove_dead_node(use);
1168 } else if (use->is_ArrayCopy()) {
1169 // Disconnect ArrayCopy node
1170 ArrayCopyNode* ac = use->as_ArrayCopy();
1171 if (ac->is_clonebasic()) {
1172 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
1173 disconnect_projections(ac, _igvn);
1174 assert(alloc->in(TypeFunc::Memory)->is_Proj() && alloc->in(TypeFunc::Memory)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
1175 Node* membar_before = alloc->in(TypeFunc::Memory)->in(0);
1176 disconnect_projections(membar_before->as_MemBar(), _igvn);
1177 if (membar_after->is_MemBar()) {
1178 disconnect_projections(membar_after->as_MemBar(), _igvn);
1179 }
1180 } else {
1181 assert(ac->is_arraycopy_validated() ||
1182 ac->is_copyof_validated() ||
1183 ac->is_copyofrange_validated(), "unsupported");
1184 CallProjections* callprojs = ac->extract_projections(true);
1185
1186 _igvn.replace_node(callprojs->fallthrough_ioproj, ac->in(TypeFunc::I_O));
1187 _igvn.replace_node(callprojs->fallthrough_memproj, ac->in(TypeFunc::Memory));
1188 _igvn.replace_node(callprojs->fallthrough_catchproj, ac->in(TypeFunc::Control));
1189
1190 // Set control to top. IGVN will remove the remaining projections
1191 ac->set_req(0, top());
1192 ac->replace_edge(res, top(), &_igvn);
1193
1194 // Disconnect src right away: it can help find new
1195 // opportunities for allocation elimination
1196 Node* src = ac->in(ArrayCopyNode::Src);
1197 ac->replace_edge(src, top(), &_igvn);
1198 // src can be top at this point if src and dest of the
1199 // arraycopy were the same
1200 if (src->outcnt() == 0 && !src->is_top()) {
1201 _igvn.remove_dead_node(src);
1202 }
1203 }
1204 _igvn._worklist.push(ac);
1205 } else if (use->is_InlineType()) {
1206 assert(use->as_InlineType()->get_oop() == res, "unexpected inline type ptr use");
1207 // Cut off oop input and remove known instance id from type
1208 _igvn.rehash_node_delayed(use);
1209 use->as_InlineType()->set_oop(_igvn, _igvn.zerocon(T_OBJECT));
1210 const TypeOopPtr* toop = _igvn.type(use)->is_oopptr()->cast_to_instance_id(TypeOopPtr::InstanceBot);
1211 _igvn.set_type(use, toop);
1212 use->as_InlineType()->set_type(toop);
1213 // Process users
1214 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); k < kmax; k++) {
1215 Node* u = use->fast_out(k);
1216 if (!u->is_InlineType()) {
1217 worklist.push(u);
1218 }
1219 }
1220 } else if (use->Opcode() == Op_StoreX && use->in(MemNode::Address) == res) {
1221 // Store to mark word of inline type larval buffer
1222 assert(inline_alloc, "Unexpected store to mark word");
1223 _igvn.replace_node(use, use->in(MemNode::Memory));
1224 } else if (use->Opcode() == Op_MemBarRelease || use->Opcode() == Op_MemBarStoreStore) {
1225 // Inline type buffer allocations are followed by a membar
1226 assert(inline_alloc, "Unexpected MemBarRelease");
1227 use->as_MemBar()->remove(&_igvn);
1228 } else {
1229 eliminate_gc_barrier(use);
1230 }
1231 j -= (oc1 - res->outcnt());
1232 }
1233 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
1234 _igvn.remove_dead_node(res);
1235 }
1236
1237 //
1238 // Process other users of allocation's projections
1239 //
1240 if (_callprojs->resproj[0] != nullptr && _callprojs->resproj[0]->outcnt() != 0) {
1241 // First disconnect stores captured by Initialize node.
1242 // If Initialize node is eliminated first in the following code,
1243 // it will kill such stores and DUIterator_Last will assert.
1244 for (DUIterator_Fast jmax, j = _callprojs->resproj[0]->fast_outs(jmax); j < jmax; j++) {
1245 Node* use = _callprojs->resproj[0]->fast_out(j);
1246 if (use->is_AddP()) {
1247 // raw memory addresses used only by the initialization
1248 _igvn.replace_node(use, C->top());
1249 --j; --jmax;
1250 }
1251 }
1252 for (DUIterator_Last jmin, j = _callprojs->resproj[0]->last_outs(jmin); j >= jmin; ) {
1253 Node* use = _callprojs->resproj[0]->last_out(j);
1254 uint oc1 = _callprojs->resproj[0]->outcnt();
1255 if (use->is_Initialize()) {
1256 // Eliminate Initialize node.
1257 InitializeNode *init = use->as_Initialize();
1258 assert(init->outcnt() <= 2, "only a control and memory projection expected");
1259 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
1260 if (ctrl_proj != nullptr) {
1261 _igvn.replace_node(ctrl_proj, init->in(TypeFunc::Control));
1262 #ifdef ASSERT
1263 // If the InitializeNode has no memory out, it will die, and tmp will become null
1264 Node* tmp = init->in(TypeFunc::Control);
1265 assert(tmp == nullptr || tmp == _callprojs->fallthrough_catchproj, "allocation control projection");
1266 #endif
1267 }
1268 Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory);
1269 if (mem_proj != nullptr) {
1270 Node *mem = init->in(TypeFunc::Memory);
1271 #ifdef ASSERT
1272 if (mem->is_MergeMem()) {
1273 assert(mem->in(TypeFunc::Memory) == _callprojs->fallthrough_memproj, "allocation memory projection");
1274 } else {
1275 assert(mem == _callprojs->fallthrough_memproj, "allocation memory projection");
1276 }
1277 #endif
1278 _igvn.replace_node(mem_proj, mem);
1279 }
1280 } else if (use->Opcode() == Op_MemBarStoreStore) {
1281 // Inline type buffer allocations are followed by a membar
1282 assert(inline_alloc, "Unexpected MemBarStoreStore");
1283 use->as_MemBar()->remove(&_igvn);
1284 } else {
1285 assert(false, "only Initialize or AddP expected");
1286 }
1287 j -= (oc1 - _callprojs->resproj[0]->outcnt());
1288 }
1289 }
1290 if (_callprojs->fallthrough_catchproj != nullptr) {
1291 _igvn.replace_node(_callprojs->fallthrough_catchproj, alloc->in(TypeFunc::Control));
1292 }
1293 if (_callprojs->fallthrough_memproj != nullptr) {
1294 _igvn.replace_node(_callprojs->fallthrough_memproj, alloc->in(TypeFunc::Memory));
1295 }
1296 if (_callprojs->catchall_memproj != nullptr) {
1297 _igvn.replace_node(_callprojs->catchall_memproj, C->top());
1298 }
1299 if (_callprojs->fallthrough_ioproj != nullptr) {
1300 _igvn.replace_node(_callprojs->fallthrough_ioproj, alloc->in(TypeFunc::I_O));
1301 }
1302 if (_callprojs->catchall_ioproj != nullptr) {
1303 _igvn.replace_node(_callprojs->catchall_ioproj, C->top());
1304 }
1305 if (_callprojs->catchall_catchproj != nullptr) {
1306 _igvn.replace_node(_callprojs->catchall_catchproj, C->top());
1307 }
1308 }
1309
1310 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1311 // If reallocation fails during deoptimization we'll pop all
1312 // interpreter frames for this compiled frame and that won't play
1313 // nice with JVMTI popframe.
1314 // We avoid this issue by eager reallocation when the popframe request
1315 // is received.
1316 if (!EliminateAllocations) {
1317 return false;
1318 }
1319 Node* klass = alloc->in(AllocateNode::KlassNode);
1320 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1321
1322 // Attempt to eliminate inline type buffer allocations
1323 // regardless of usage and escape/replaceable status.
1324 bool inline_alloc = tklass->isa_instklassptr() &&
1325 tklass->is_instklassptr()->instance_klass()->is_inlinetype();
1326 if (!alloc->_is_non_escaping && !inline_alloc) {
1327 return false;
1328 }
1329 // Eliminate boxing allocations which are not used
1330 // regardless scalar replaceable status.
1331 Node* res = alloc->result_cast();
1332 bool boxing_alloc = (res == nullptr) && C->eliminate_boxing() &&
1333 tklass->isa_instklassptr() &&
1334 tklass->is_instklassptr()->instance_klass()->is_box_klass();
1335 if (!alloc->_is_scalar_replaceable && !boxing_alloc && !inline_alloc) {
1336 return false;
1337 }
1338
1339 _callprojs = alloc->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1340
1341 GrowableArray <SafePointNode *> safepoints;
1342 if (!can_eliminate_allocation(&_igvn, alloc, &safepoints)) {
1343 return false;
1344 }
1345
1346 if (!alloc->_is_scalar_replaceable) {
1347 assert(res == nullptr || inline_alloc, "sanity");
1348 // We can only eliminate allocation if all debug info references
1349 // are already replaced with SafePointScalarObject because
1350 // we can't search for a fields value without instance_id.
1351 if (safepoints.length() > 0) {
1352 return false;
1353 }
1354 }
1355
1356 if (!scalar_replacement(alloc, safepoints)) {
1357 return false;
1358 }
1359
1360 CompileLog* log = C->log();
1361 if (log != nullptr) {
1362 log->head("eliminate_allocation type='%d'",
1363 log->identify(tklass->exact_klass()));
1364 JVMState* p = alloc->jvms();
1365 while (p != nullptr) {
1366 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1367 p = p->caller();
1368 }
1369 log->tail("eliminate_allocation");
1370 }
1371
1372 process_users_of_allocation(alloc, inline_alloc);
1373
1374 #ifndef PRODUCT
1375 if (PrintEliminateAllocations) {
1376 if (alloc->is_AllocateArray())
1377 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1378 else
1379 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1380 }
1381 #endif
1382
1383 return true;
1384 }
1385
1386 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1387 // EA should remove all uses of non-escaping boxing node.
1388 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != nullptr) {
1389 return false;
1390 }
1391
1392 assert(boxing->result_cast() == nullptr, "unexpected boxing node result");
1393
1394 _callprojs = boxing->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1395
1396 const TypeTuple* r = boxing->tf()->range_sig();
1397 assert(r->cnt() > TypeFunc::Parms, "sanity");
1398 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1399 assert(t != nullptr, "sanity");
1400
1401 CompileLog* log = C->log();
1402 if (log != nullptr) {
1403 log->head("eliminate_boxing type='%d'",
1404 log->identify(t->instance_klass()));
1405 JVMState* p = boxing->jvms();
1406 while (p != nullptr) {
1407 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1408 p = p->caller();
1409 }
1410 log->tail("eliminate_boxing");
1411 }
1412
1413 process_users_of_allocation(boxing);
1414
1415 #ifndef PRODUCT
1416 if (PrintEliminateAllocations) {
1417 tty->print("++++ Eliminated: %d ", boxing->_idx);
1418 boxing->method()->print_short_name(tty);
1419 tty->cr();
1420 }
1421 #endif
1422
1423 return true;
1424 }
1425
1426
1427 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1428 Node* adr = basic_plus_adr(base, offset);
1429 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1430 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered);
1431 transform_later(value);
1432 return value;
1433 }
1434
1435
1436 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1437 Node* adr = basic_plus_adr(base, offset);
1438 mem = StoreNode::make(_igvn, ctl, mem, adr, nullptr, value, bt, MemNode::unordered);
1439 transform_later(mem);
1440 return mem;
1441 }
1442
1443 //=============================================================================
1444 //
1445 // A L L O C A T I O N
1446 //
1447 // Allocation attempts to be fast in the case of frequent small objects.
1448 // It breaks down like this:
1449 //
1450 // 1) Size in doublewords is computed. This is a constant for objects and
1451 // variable for most arrays. Doubleword units are used to avoid size
1452 // overflow of huge doubleword arrays. We need doublewords in the end for
1453 // rounding.
1454 //
1455 // 2) Size is checked for being 'too large'. Too-large allocations will go
1456 // the slow path into the VM. The slow path can throw any required
1457 // exceptions, and does all the special checks for very large arrays. The
1458 // size test can constant-fold away for objects. For objects with
1459 // finalizers it constant-folds the otherway: you always go slow with
1460 // finalizers.
1461 //
1462 // 3) If NOT using TLABs, this is the contended loop-back point.
1463 // Load-Locked the heap top. If using TLABs normal-load the heap top.
1464 //
1465 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
1466 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
1467 // "size*8" we always enter the VM, where "largish" is a constant picked small
1468 // enough that there's always space between the eden max and 4Gig (old space is
1469 // there so it's quite large) and large enough that the cost of entering the VM
1470 // is dwarfed by the cost to initialize the space.
1471 //
1472 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1473 // down. If contended, repeat at step 3. If using TLABs normal-store
1474 // adjusted heap top back down; there is no contention.
1475 //
1476 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
1477 // fields.
1478 //
1479 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1480 // oop flavor.
1481 //
1482 //=============================================================================
1483 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1484 // Allocations bigger than this always go the slow route.
1485 // This value must be small enough that allocation attempts that need to
1486 // trigger exceptions go the slow route. Also, it must be small enough so
1487 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1488 //=============================================================================j//
1489 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1490 // The allocator will coalesce int->oop copies away. See comment in
1491 // coalesce.cpp about how this works. It depends critically on the exact
1492 // code shape produced here, so if you are changing this code shape
1493 // make sure the GC info for the heap-top is correct in and around the
1494 // slow-path call.
1495 //
1496
1497 void PhaseMacroExpand::expand_allocate_common(
1498 AllocateNode* alloc, // allocation node to be expanded
1499 Node* length, // array length for an array allocation
1500 Node* init_val, // value to initialize the array with
1501 const TypeFunc* slow_call_type, // Type of slow call
1502 address slow_call_address, // Address of slow call
1503 Node* valid_length_test // whether length is valid or not
1504 )
1505 {
1506 Node* ctrl = alloc->in(TypeFunc::Control);
1507 Node* mem = alloc->in(TypeFunc::Memory);
1508 Node* i_o = alloc->in(TypeFunc::I_O);
1509 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1510 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1511 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1512 assert(ctrl != nullptr, "must have control");
1513
1514 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1515 // they will not be used if "always_slow" is set
1516 enum { slow_result_path = 1, fast_result_path = 2 };
1517 Node *result_region = nullptr;
1518 Node *result_phi_rawmem = nullptr;
1519 Node *result_phi_rawoop = nullptr;
1520 Node *result_phi_i_o = nullptr;
1521
1522 // The initial slow comparison is a size check, the comparison
1523 // we want to do is a BoolTest::gt
1524 bool expand_fast_path = true;
1525 int tv = _igvn.find_int_con(initial_slow_test, -1);
1526 if (tv >= 0) {
1527 // InitialTest has constant result
1528 // 0 - can fit in TLAB
1529 // 1 - always too big or negative
1530 assert(tv <= 1, "0 or 1 if a constant");
1531 expand_fast_path = (tv == 0);
1532 initial_slow_test = nullptr;
1533 } else {
1534 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1535 }
1536
1537 if (!UseTLAB) {
1538 // Force slow-path allocation
1539 expand_fast_path = false;
1540 initial_slow_test = nullptr;
1541 }
1542
1543 bool allocation_has_use = (alloc->result_cast() != nullptr);
1544 if (!allocation_has_use) {
1545 InitializeNode* init = alloc->initialization();
1546 if (init != nullptr) {
1547 init->remove(&_igvn);
1548 }
1549 if (expand_fast_path && (initial_slow_test == nullptr)) {
1550 // Remove allocation node and return.
1551 // Size is a non-negative constant -> no initial check needed -> directly to fast path.
1552 // Also, no usages -> empty fast path -> no fall out to slow path -> nothing left.
1553 #ifndef PRODUCT
1554 if (PrintEliminateAllocations) {
1555 tty->print("NotUsed ");
1556 Node* res = alloc->proj_out_or_null(TypeFunc::Parms);
1557 if (res != nullptr) {
1558 res->dump();
1559 } else {
1560 alloc->dump();
1561 }
1562 }
1563 #endif
1564 yank_alloc_node(alloc);
1565 return;
1566 }
1567 }
1568
1569 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1570 Node *slow_region = nullptr;
1571 Node *toobig_false = ctrl;
1572
1573 // generate the initial test if necessary
1574 if (initial_slow_test != nullptr ) {
1575 assert (expand_fast_path, "Only need test if there is a fast path");
1576 slow_region = new RegionNode(3);
1577
1578 // Now make the initial failure test. Usually a too-big test but
1579 // might be a TRUE for finalizers.
1580 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1581 transform_later(toobig_iff);
1582 // Plug the failing-too-big test into the slow-path region
1583 Node* toobig_true = new IfTrueNode(toobig_iff);
1584 transform_later(toobig_true);
1585 slow_region ->init_req( too_big_or_final_path, toobig_true );
1586 toobig_false = new IfFalseNode(toobig_iff);
1587 transform_later(toobig_false);
1588 } else {
1589 // No initial test, just fall into next case
1590 assert(allocation_has_use || !expand_fast_path, "Should already have been handled");
1591 toobig_false = ctrl;
1592 DEBUG_ONLY(slow_region = NodeSentinel);
1593 }
1594
1595 // If we are here there are several possibilities
1596 // - expand_fast_path is false - then only a slow path is expanded. That's it.
1597 // no_initial_check means a constant allocation.
1598 // - If check always evaluates to false -> expand_fast_path is false (see above)
1599 // - If check always evaluates to true -> directly into fast path (but may bailout to slowpath)
1600 // if !allocation_has_use the fast path is empty
1601 // if !allocation_has_use && no_initial_check
1602 // - Then there are no fastpath that can fall out to slowpath -> no allocation code at all.
1603 // removed by yank_alloc_node above.
1604
1605 Node *slow_mem = mem; // save the current memory state for slow path
1606 // generate the fast allocation code unless we know that the initial test will always go slow
1607 if (expand_fast_path) {
1608 // Fast path modifies only raw memory.
1609 if (mem->is_MergeMem()) {
1610 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1611 }
1612
1613 // allocate the Region and Phi nodes for the result
1614 result_region = new RegionNode(3);
1615 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1616 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1617
1618 // Grab regular I/O before optional prefetch may change it.
1619 // Slow-path does no I/O so just set it to the original I/O.
1620 result_phi_i_o->init_req(slow_result_path, i_o);
1621
1622 // Name successful fast-path variables
1623 Node* fast_oop_ctrl;
1624 Node* fast_oop_rawmem;
1625
1626 if (allocation_has_use) {
1627 Node* needgc_ctrl = nullptr;
1628 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1629
1630 intx prefetch_lines = length != nullptr ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1631 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1632 Node* fast_oop = bs->obj_allocate(this, mem, toobig_false, size_in_bytes, i_o, needgc_ctrl,
1633 fast_oop_ctrl, fast_oop_rawmem,
1634 prefetch_lines);
1635
1636 if (initial_slow_test != nullptr) {
1637 // This completes all paths into the slow merge point
1638 slow_region->init_req(need_gc_path, needgc_ctrl);
1639 transform_later(slow_region);
1640 } else {
1641 // No initial slow path needed!
1642 // Just fall from the need-GC path straight into the VM call.
1643 slow_region = needgc_ctrl;
1644 }
1645
1646 InitializeNode* init = alloc->initialization();
1647 fast_oop_rawmem = initialize_object(alloc,
1648 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1649 klass_node, length, size_in_bytes);
1650 expand_initialize_membar(alloc, init, fast_oop_ctrl, fast_oop_rawmem);
1651 expand_dtrace_alloc_probe(alloc, fast_oop, fast_oop_ctrl, fast_oop_rawmem);
1652
1653 result_phi_rawoop->init_req(fast_result_path, fast_oop);
1654 } else {
1655 assert (initial_slow_test != nullptr, "sanity");
1656 fast_oop_ctrl = toobig_false;
1657 fast_oop_rawmem = mem;
1658 transform_later(slow_region);
1659 }
1660
1661 // Plug in the successful fast-path into the result merge point
1662 result_region ->init_req(fast_result_path, fast_oop_ctrl);
1663 result_phi_i_o ->init_req(fast_result_path, i_o);
1664 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1665 } else {
1666 slow_region = ctrl;
1667 result_phi_i_o = i_o; // Rename it to use in the following code.
1668 }
1669
1670 // Generate slow-path call
1671 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1672 OptoRuntime::stub_name(slow_call_address),
1673 TypePtr::BOTTOM);
1674 call->init_req(TypeFunc::Control, slow_region);
1675 call->init_req(TypeFunc::I_O, top()); // does no i/o
1676 call->init_req(TypeFunc::Memory, slow_mem); // may gc ptrs
1677 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1678 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1679
1680 call->init_req(TypeFunc::Parms+0, klass_node);
1681 if (length != nullptr) {
1682 call->init_req(TypeFunc::Parms+1, length);
1683 if (init_val != nullptr) {
1684 call->init_req(TypeFunc::Parms+2, init_val);
1685 }
1686 } else {
1687 // Let the runtime know if this is a larval allocation
1688 call->init_req(TypeFunc::Parms+1, _igvn.intcon(alloc->_larval));
1689 }
1690
1691 // Copy debug information and adjust JVMState information, then replace
1692 // allocate node with the call
1693 call->copy_call_debug_info(&_igvn, alloc);
1694 // For array allocations, copy the valid length check to the call node so Compile::final_graph_reshaping() can verify
1695 // that the call has the expected number of CatchProj nodes (in case the allocation always fails and the fallthrough
1696 // path dies).
1697 if (valid_length_test != nullptr) {
1698 call->add_req(valid_length_test);
1699 }
1700 if (expand_fast_path) {
1701 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1702 } else {
1703 // Hook i_o projection to avoid its elimination during allocation
1704 // replacement (when only a slow call is generated).
1705 call->set_req(TypeFunc::I_O, result_phi_i_o);
1706 }
1707 _igvn.replace_node(alloc, call);
1708 transform_later(call);
1709
1710 // Identify the output projections from the allocate node and
1711 // adjust any references to them.
1712 // The control and io projections look like:
1713 //
1714 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1715 // Allocate Catch
1716 // ^---Proj(io) <-------+ ^---CatchProj(io)
1717 //
1718 // We are interested in the CatchProj nodes.
1719 //
1720 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1721
1722 // An allocate node has separate memory projections for the uses on
1723 // the control and i_o paths. Replace the control memory projection with
1724 // result_phi_rawmem (unless we are only generating a slow call when
1725 // both memory projections are combined)
1726 if (expand_fast_path && _callprojs->fallthrough_memproj != nullptr) {
1727 _igvn.replace_in_uses(_callprojs->fallthrough_memproj, result_phi_rawmem);
1728 }
1729 // Now change uses of catchall_memproj to use fallthrough_memproj and delete
1730 // catchall_memproj so we end up with a call that has only 1 memory projection.
1731 if (_callprojs->catchall_memproj != nullptr) {
1732 if (_callprojs->fallthrough_memproj == nullptr) {
1733 _callprojs->fallthrough_memproj = new ProjNode(call, TypeFunc::Memory);
1734 transform_later(_callprojs->fallthrough_memproj);
1735 }
1736 _igvn.replace_in_uses(_callprojs->catchall_memproj, _callprojs->fallthrough_memproj);
1737 _igvn.remove_dead_node(_callprojs->catchall_memproj);
1738 }
1739
1740 // An allocate node has separate i_o projections for the uses on the control
1741 // and i_o paths. Always replace the control i_o projection with result i_o
1742 // otherwise incoming i_o become dead when only a slow call is generated
1743 // (it is different from memory projections where both projections are
1744 // combined in such case).
1745 if (_callprojs->fallthrough_ioproj != nullptr) {
1746 _igvn.replace_in_uses(_callprojs->fallthrough_ioproj, result_phi_i_o);
1747 }
1748 // Now change uses of catchall_ioproj to use fallthrough_ioproj and delete
1749 // catchall_ioproj so we end up with a call that has only 1 i_o projection.
1750 if (_callprojs->catchall_ioproj != nullptr) {
1751 if (_callprojs->fallthrough_ioproj == nullptr) {
1752 _callprojs->fallthrough_ioproj = new ProjNode(call, TypeFunc::I_O);
1753 transform_later(_callprojs->fallthrough_ioproj);
1754 }
1755 _igvn.replace_in_uses(_callprojs->catchall_ioproj, _callprojs->fallthrough_ioproj);
1756 _igvn.remove_dead_node(_callprojs->catchall_ioproj);
1757 }
1758
1759 // if we generated only a slow call, we are done
1760 if (!expand_fast_path) {
1761 // Now we can unhook i_o.
1762 if (result_phi_i_o->outcnt() > 1) {
1763 call->set_req(TypeFunc::I_O, top());
1764 } else {
1765 assert(result_phi_i_o->unique_ctrl_out() == call, "sanity");
1766 // Case of new array with negative size known during compilation.
1767 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1768 // following code since call to runtime will throw exception.
1769 // As result there will be no users of i_o after the call.
1770 // Leave i_o attached to this call to avoid problems in preceding graph.
1771 }
1772 return;
1773 }
1774
1775 if (_callprojs->fallthrough_catchproj != nullptr) {
1776 ctrl = _callprojs->fallthrough_catchproj->clone();
1777 transform_later(ctrl);
1778 _igvn.replace_node(_callprojs->fallthrough_catchproj, result_region);
1779 } else {
1780 ctrl = top();
1781 }
1782 Node *slow_result;
1783 if (_callprojs->resproj[0] == nullptr) {
1784 // no uses of the allocation result
1785 slow_result = top();
1786 } else {
1787 slow_result = _callprojs->resproj[0]->clone();
1788 transform_later(slow_result);
1789 _igvn.replace_node(_callprojs->resproj[0], result_phi_rawoop);
1790 }
1791
1792 // Plug slow-path into result merge point
1793 result_region->init_req( slow_result_path, ctrl);
1794 transform_later(result_region);
1795 if (allocation_has_use) {
1796 result_phi_rawoop->init_req(slow_result_path, slow_result);
1797 transform_later(result_phi_rawoop);
1798 }
1799 result_phi_rawmem->init_req(slow_result_path, _callprojs->fallthrough_memproj);
1800 transform_later(result_phi_rawmem);
1801 transform_later(result_phi_i_o);
1802 // This completes all paths into the result merge point
1803 }
1804
1805 // Remove alloc node that has no uses.
1806 void PhaseMacroExpand::yank_alloc_node(AllocateNode* alloc) {
1807 Node* ctrl = alloc->in(TypeFunc::Control);
1808 Node* mem = alloc->in(TypeFunc::Memory);
1809 Node* i_o = alloc->in(TypeFunc::I_O);
1810
1811 _callprojs = alloc->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1812 if (_callprojs->resproj[0] != nullptr) {
1813 for (DUIterator_Fast imax, i = _callprojs->resproj[0]->fast_outs(imax); i < imax; i++) {
1814 Node* use = _callprojs->resproj[0]->fast_out(i);
1815 use->isa_MemBar()->remove(&_igvn);
1816 --imax;
1817 --i; // back up iterator
1818 }
1819 assert(_callprojs->resproj[0]->outcnt() == 0, "all uses must be deleted");
1820 _igvn.remove_dead_node(_callprojs->resproj[0]);
1821 }
1822 if (_callprojs->fallthrough_catchproj != nullptr) {
1823 _igvn.replace_in_uses(_callprojs->fallthrough_catchproj, ctrl);
1824 _igvn.remove_dead_node(_callprojs->fallthrough_catchproj);
1825 }
1826 if (_callprojs->catchall_catchproj != nullptr) {
1827 _igvn.rehash_node_delayed(_callprojs->catchall_catchproj);
1828 _callprojs->catchall_catchproj->set_req(0, top());
1829 }
1830 if (_callprojs->fallthrough_proj != nullptr) {
1831 Node* catchnode = _callprojs->fallthrough_proj->unique_ctrl_out();
1832 _igvn.remove_dead_node(catchnode);
1833 _igvn.remove_dead_node(_callprojs->fallthrough_proj);
1834 }
1835 if (_callprojs->fallthrough_memproj != nullptr) {
1836 _igvn.replace_in_uses(_callprojs->fallthrough_memproj, mem);
1837 _igvn.remove_dead_node(_callprojs->fallthrough_memproj);
1838 }
1839 if (_callprojs->fallthrough_ioproj != nullptr) {
1840 _igvn.replace_in_uses(_callprojs->fallthrough_ioproj, i_o);
1841 _igvn.remove_dead_node(_callprojs->fallthrough_ioproj);
1842 }
1843 if (_callprojs->catchall_memproj != nullptr) {
1844 _igvn.rehash_node_delayed(_callprojs->catchall_memproj);
1845 _callprojs->catchall_memproj->set_req(0, top());
1846 }
1847 if (_callprojs->catchall_ioproj != nullptr) {
1848 _igvn.rehash_node_delayed(_callprojs->catchall_ioproj);
1849 _callprojs->catchall_ioproj->set_req(0, top());
1850 }
1851 #ifndef PRODUCT
1852 if (PrintEliminateAllocations) {
1853 if (alloc->is_AllocateArray()) {
1854 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1855 } else {
1856 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1857 }
1858 }
1859 #endif
1860 _igvn.remove_dead_node(alloc);
1861 }
1862
1863 void PhaseMacroExpand::expand_initialize_membar(AllocateNode* alloc, InitializeNode* init,
1864 Node*& fast_oop_ctrl, Node*& fast_oop_rawmem) {
1865 // If initialization is performed by an array copy, any required
1866 // MemBarStoreStore was already added. If the object does not
1867 // escape no need for a MemBarStoreStore. If the object does not
1868 // escape in its initializer and memory barrier (MemBarStoreStore or
1869 // stronger) is already added at exit of initializer, also no need
1870 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore
1871 // so that stores that initialize this object can't be reordered
1872 // with a subsequent store that makes this object accessible by
1873 // other threads.
1874 // Other threads include java threads and JVM internal threads
1875 // (for example concurrent GC threads). Current concurrent GC
1876 // implementation: G1 will not scan newly created object,
1877 // so it's safe to skip storestore barrier when allocation does
1878 // not escape.
1879 if (!alloc->does_not_escape_thread() &&
1880 !alloc->is_allocation_MemBar_redundant() &&
1881 (init == nullptr || !init->is_complete_with_arraycopy())) {
1882 if (init == nullptr || init->req() < InitializeNode::RawStores) {
1883 // No InitializeNode or no stores captured by zeroing
1884 // elimination. Simply add the MemBarStoreStore after object
1885 // initialization.
1886 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1887 transform_later(mb);
1888
1889 mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
1890 mb->init_req(TypeFunc::Control, fast_oop_ctrl);
1891 fast_oop_ctrl = new ProjNode(mb, TypeFunc::Control);
1892 transform_later(fast_oop_ctrl);
1893 fast_oop_rawmem = new ProjNode(mb, TypeFunc::Memory);
1894 transform_later(fast_oop_rawmem);
1895 } else {
1896 // Add the MemBarStoreStore after the InitializeNode so that
1897 // all stores performing the initialization that were moved
1898 // before the InitializeNode happen before the storestore
1899 // barrier.
1900
1901 Node* init_ctrl = init->proj_out_or_null(TypeFunc::Control);
1902 Node* init_mem = init->proj_out_or_null(TypeFunc::Memory);
1903
1904 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1905 transform_later(mb);
1906
1907 Node* ctrl = new ProjNode(init, TypeFunc::Control);
1908 transform_later(ctrl);
1909 Node* mem = new ProjNode(init, TypeFunc::Memory);
1910 transform_later(mem);
1911
1912 // The MemBarStoreStore depends on control and memory coming
1913 // from the InitializeNode
1914 mb->init_req(TypeFunc::Memory, mem);
1915 mb->init_req(TypeFunc::Control, ctrl);
1916
1917 ctrl = new ProjNode(mb, TypeFunc::Control);
1918 transform_later(ctrl);
1919 mem = new ProjNode(mb, TypeFunc::Memory);
1920 transform_later(mem);
1921
1922 // All nodes that depended on the InitializeNode for control
1923 // and memory must now depend on the MemBarNode that itself
1924 // depends on the InitializeNode
1925 if (init_ctrl != nullptr) {
1926 _igvn.replace_node(init_ctrl, ctrl);
1927 }
1928 if (init_mem != nullptr) {
1929 _igvn.replace_node(init_mem, mem);
1930 }
1931 }
1932 }
1933 }
1934
1935 void PhaseMacroExpand::expand_dtrace_alloc_probe(AllocateNode* alloc, Node* oop,
1936 Node*& ctrl, Node*& rawmem) {
1937 if (C->env()->dtrace_alloc_probes()) {
1938 // Slow-path call
1939 int size = TypeFunc::Parms + 2;
1940 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1941 CAST_FROM_FN_PTR(address,
1942 static_cast<int (*)(JavaThread*, oopDesc*)>(SharedRuntime::dtrace_object_alloc)),
1943 "dtrace_object_alloc",
1944 TypeRawPtr::BOTTOM);
1945
1946 // Get base of thread-local storage area
1947 Node* thread = new ThreadLocalNode();
1948 transform_later(thread);
1949
1950 call->init_req(TypeFunc::Parms + 0, thread);
1951 call->init_req(TypeFunc::Parms + 1, oop);
1952 call->init_req(TypeFunc::Control, ctrl);
1953 call->init_req(TypeFunc::I_O , top()); // does no i/o
1954 call->init_req(TypeFunc::Memory , rawmem);
1955 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1956 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1957 transform_later(call);
1958 ctrl = new ProjNode(call, TypeFunc::Control);
1959 transform_later(ctrl);
1960 rawmem = new ProjNode(call, TypeFunc::Memory);
1961 transform_later(rawmem);
1962 }
1963 }
1964
1965 // Helper for PhaseMacroExpand::expand_allocate_common.
1966 // Initializes the newly-allocated storage.
1967 Node* PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1968 Node* control, Node* rawmem, Node* object,
1969 Node* klass_node, Node* length,
1970 Node* size_in_bytes) {
1971 InitializeNode* init = alloc->initialization();
1972 // Store the klass & mark bits
1973 Node* mark_node = alloc->make_ideal_mark(&_igvn, control, rawmem);
1974 if (!mark_node->is_Con()) {
1975 transform_later(mark_node);
1976 }
1977 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, TypeX_X->basic_type());
1978
1979 if (!UseCompactObjectHeaders) {
1980 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1981 }
1982 int header_size = alloc->minimum_header_size(); // conservatively small
1983
1984 // Array length
1985 if (length != nullptr) { // Arrays need length field
1986 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1987 // conservatively small header size:
1988 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1989 if (_igvn.type(klass_node)->isa_aryklassptr()) { // we know the exact header size in most cases:
1990 BasicType elem = _igvn.type(klass_node)->is_klassptr()->as_instance_type()->isa_aryptr()->elem()->array_element_basic_type();
1991 if (is_reference_type(elem, true)) {
1992 elem = T_OBJECT;
1993 }
1994 header_size = Klass::layout_helper_header_size(Klass::array_layout_helper(elem));
1995 }
1996 }
1997
1998 // Clear the object body, if necessary.
1999 if (init == nullptr) {
2000 // The init has somehow disappeared; be cautious and clear everything.
2001 //
2002 // This can happen if a node is allocated but an uncommon trap occurs
2003 // immediately. In this case, the Initialize gets associated with the
2004 // trap, and may be placed in a different (outer) loop, if the Allocate
2005 // is in a loop. If (this is rare) the inner loop gets unrolled, then
2006 // there can be two Allocates to one Initialize. The answer in all these
2007 // edge cases is safety first. It is always safe to clear immediately
2008 // within an Allocate, and then (maybe or maybe not) clear some more later.
2009 if (!(UseTLAB && ZeroTLAB)) {
2010 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
2011 alloc->in(AllocateNode::InitValue),
2012 alloc->in(AllocateNode::RawInitValue),
2013 header_size, size_in_bytes,
2014 &_igvn);
2015 }
2016 } else {
2017 if (!init->is_complete()) {
2018 // Try to win by zeroing only what the init does not store.
2019 // We can also try to do some peephole optimizations,
2020 // such as combining some adjacent subword stores.
2021 rawmem = init->complete_stores(control, rawmem, object,
2022 header_size, size_in_bytes, &_igvn);
2023 }
2024 // We have no more use for this link, since the AllocateNode goes away:
2025 init->set_req(InitializeNode::RawAddress, top());
2026 // (If we keep the link, it just confuses the register allocator,
2027 // who thinks he sees a real use of the address by the membar.)
2028 }
2029
2030 return rawmem;
2031 }
2032
2033 // Generate prefetch instructions for next allocations.
2034 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
2035 Node*& contended_phi_rawmem,
2036 Node* old_eden_top, Node* new_eden_top,
2037 intx lines) {
2038 enum { fall_in_path = 1, pf_path = 2 };
2039 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
2040 // Generate prefetch allocation with watermark check.
2041 // As an allocation hits the watermark, we will prefetch starting
2042 // at a "distance" away from watermark.
2043
2044 Node *pf_region = new RegionNode(3);
2045 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY,
2046 TypeRawPtr::BOTTOM );
2047 // I/O is used for Prefetch
2048 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO );
2049
2050 Node *thread = new ThreadLocalNode();
2051 transform_later(thread);
2052
2053 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread,
2054 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
2055 transform_later(eden_pf_adr);
2056
2057 Node *old_pf_wm = new LoadPNode(needgc_false,
2058 contended_phi_rawmem, eden_pf_adr,
2059 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM,
2060 MemNode::unordered);
2061 transform_later(old_pf_wm);
2062
2063 // check against new_eden_top
2064 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm );
2065 transform_later(need_pf_cmp);
2066 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge );
2067 transform_later(need_pf_bol);
2068 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol,
2069 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
2070 transform_later(need_pf_iff);
2071
2072 // true node, add prefetchdistance
2073 Node *need_pf_true = new IfTrueNode( need_pf_iff );
2074 transform_later(need_pf_true);
2075
2076 Node *need_pf_false = new IfFalseNode( need_pf_iff );
2077 transform_later(need_pf_false);
2078
2079 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm,
2080 _igvn.MakeConX(AllocatePrefetchDistance) );
2081 transform_later(new_pf_wmt );
2082 new_pf_wmt->set_req(0, need_pf_true);
2083
2084 Node *store_new_wmt = new StorePNode(need_pf_true,
2085 contended_phi_rawmem, eden_pf_adr,
2086 TypeRawPtr::BOTTOM, new_pf_wmt,
2087 MemNode::unordered);
2088 transform_later(store_new_wmt);
2089
2090 // adding prefetches
2091 pf_phi_abio->init_req( fall_in_path, i_o );
2092
2093 Node *prefetch_adr;
2094 Node *prefetch;
2095 uint step_size = AllocatePrefetchStepSize;
2096 uint distance = 0;
2097
2098 for ( intx i = 0; i < lines; i++ ) {
2099 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt,
2100 _igvn.MakeConX(distance) );
2101 transform_later(prefetch_adr);
2102 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
2103 transform_later(prefetch);
2104 distance += step_size;
2105 i_o = prefetch;
2106 }
2107 pf_phi_abio->set_req( pf_path, i_o );
2108
2109 pf_region->init_req( fall_in_path, need_pf_false );
2110 pf_region->init_req( pf_path, need_pf_true );
2111
2112 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
2113 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
2114
2115 transform_later(pf_region);
2116 transform_later(pf_phi_rawmem);
2117 transform_later(pf_phi_abio);
2118
2119 needgc_false = pf_region;
2120 contended_phi_rawmem = pf_phi_rawmem;
2121 i_o = pf_phi_abio;
2122 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) {
2123 // Insert a prefetch instruction for each allocation.
2124 // This code is used to generate 1 prefetch instruction per cache line.
2125
2126 // Generate several prefetch instructions.
2127 uint step_size = AllocatePrefetchStepSize;
2128 uint distance = AllocatePrefetchDistance;
2129
2130 // Next cache address.
2131 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top,
2132 _igvn.MakeConX(step_size + distance));
2133 transform_later(cache_adr);
2134 cache_adr = new CastP2XNode(needgc_false, cache_adr);
2135 transform_later(cache_adr);
2136 // Address is aligned to execute prefetch to the beginning of cache line size
2137 // (it is important when BIS instruction is used on SPARC as prefetch).
2138 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
2139 cache_adr = new AndXNode(cache_adr, mask);
2140 transform_later(cache_adr);
2141 cache_adr = new CastX2PNode(cache_adr);
2142 transform_later(cache_adr);
2143
2144 // Prefetch
2145 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr );
2146 prefetch->set_req(0, needgc_false);
2147 transform_later(prefetch);
2148 contended_phi_rawmem = prefetch;
2149 Node *prefetch_adr;
2150 distance = step_size;
2151 for ( intx i = 1; i < lines; i++ ) {
2152 prefetch_adr = new AddPNode( cache_adr, cache_adr,
2153 _igvn.MakeConX(distance) );
2154 transform_later(prefetch_adr);
2155 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr );
2156 transform_later(prefetch);
2157 distance += step_size;
2158 contended_phi_rawmem = prefetch;
2159 }
2160 } else if( AllocatePrefetchStyle > 0 ) {
2161 // Insert a prefetch for each allocation only on the fast-path
2162 Node *prefetch_adr;
2163 Node *prefetch;
2164 // Generate several prefetch instructions.
2165 uint step_size = AllocatePrefetchStepSize;
2166 uint distance = AllocatePrefetchDistance;
2167 for ( intx i = 0; i < lines; i++ ) {
2168 prefetch_adr = new AddPNode( old_eden_top, new_eden_top,
2169 _igvn.MakeConX(distance) );
2170 transform_later(prefetch_adr);
2171 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
2172 // Do not let it float too high, since if eden_top == eden_end,
2173 // both might be null.
2174 if( i == 0 ) { // Set control for first prefetch, next follows it
2175 prefetch->init_req(0, needgc_false);
2176 }
2177 transform_later(prefetch);
2178 distance += step_size;
2179 i_o = prefetch;
2180 }
2181 }
2182 return i_o;
2183 }
2184
2185
2186 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
2187 expand_allocate_common(alloc, nullptr, nullptr,
2188 OptoRuntime::new_instance_Type(),
2189 OptoRuntime::new_instance_Java(), nullptr);
2190 }
2191
2192 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
2193 Node* length = alloc->in(AllocateNode::ALength);
2194 Node* valid_length_test = alloc->in(AllocateNode::ValidLengthTest);
2195 InitializeNode* init = alloc->initialization();
2196 Node* klass_node = alloc->in(AllocateNode::KlassNode);
2197 Node* init_value = alloc->in(AllocateNode::InitValue);
2198 const TypeAryKlassPtr* ary_klass_t = _igvn.type(klass_node)->isa_aryklassptr();
2199 // TODO 8366668 Compute the VM type, is this even needed now that we set it earlier? Should we assert instead?
2200 if (ary_klass_t && ary_klass_t->klass_is_exact() && ary_klass_t->exact_klass()->is_obj_array_klass()) {
2201 ary_klass_t = ary_klass_t->get_vm_type();
2202 klass_node = makecon(ary_klass_t);
2203 _igvn.replace_input_of(alloc, AllocateNode::KlassNode, klass_node);
2204 }
2205 const TypeFunc* slow_call_type;
2206 address slow_call_address; // Address of slow call
2207 if (init != nullptr && init->is_complete_with_arraycopy() &&
2208 ary_klass_t && ary_klass_t->elem()->isa_klassptr() == nullptr) {
2209 // Don't zero type array during slow allocation in VM since
2210 // it will be initialized later by arraycopy in compiled code.
2211 slow_call_address = OptoRuntime::new_array_nozero_Java();
2212 slow_call_type = OptoRuntime::new_array_nozero_Type();
2213 } else {
2214 slow_call_address = OptoRuntime::new_array_Java();
2215 slow_call_type = OptoRuntime::new_array_Type();
2216
2217 if (init_value == nullptr) {
2218 init_value = _igvn.zerocon(T_OBJECT);
2219 } else if (UseCompressedOops) {
2220 init_value = transform_later(new DecodeNNode(init_value, init_value->bottom_type()->make_ptr()));
2221 }
2222 }
2223 expand_allocate_common(alloc, length, init_value,
2224 slow_call_type,
2225 slow_call_address, valid_length_test);
2226 }
2227
2228 //-------------------mark_eliminated_box----------------------------------
2229 //
2230 // During EA obj may point to several objects but after few ideal graph
2231 // transformations (CCP) it may point to only one non escaping object
2232 // (but still using phi), corresponding locks and unlocks will be marked
2233 // for elimination. Later obj could be replaced with a new node (new phi)
2234 // and which does not have escape information. And later after some graph
2235 // reshape other locks and unlocks (which were not marked for elimination
2236 // before) are connected to this new obj (phi) but they still will not be
2237 // marked for elimination since new obj has no escape information.
2238 // Mark all associated (same box and obj) lock and unlock nodes for
2239 // elimination if some of them marked already.
2240 void PhaseMacroExpand::mark_eliminated_box(Node* box, Node* obj) {
2241 BoxLockNode* oldbox = box->as_BoxLock();
2242 if (oldbox->is_eliminated()) {
2243 return; // This BoxLock node was processed already.
2244 }
2245 assert(!oldbox->is_unbalanced(), "this should not be called for unbalanced region");
2246 // New implementation (EliminateNestedLocks) has separate BoxLock
2247 // node for each locked region so mark all associated locks/unlocks as
2248 // eliminated even if different objects are referenced in one locked region
2249 // (for example, OSR compilation of nested loop inside locked scope).
2250 if (EliminateNestedLocks ||
2251 oldbox->as_BoxLock()->is_simple_lock_region(nullptr, obj, nullptr)) {
2252 // Box is used only in one lock region. Mark this box as eliminated.
2253 oldbox->set_local(); // This verifies correct state of BoxLock
2254 _igvn.hash_delete(oldbox);
2255 oldbox->set_eliminated(); // This changes box's hash value
2256 _igvn.hash_insert(oldbox);
2257
2258 for (uint i = 0; i < oldbox->outcnt(); i++) {
2259 Node* u = oldbox->raw_out(i);
2260 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
2261 AbstractLockNode* alock = u->as_AbstractLock();
2262 // Check lock's box since box could be referenced by Lock's debug info.
2263 if (alock->box_node() == oldbox) {
2264 // Mark eliminated all related locks and unlocks.
2265 #ifdef ASSERT
2266 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4");
2267 #endif
2268 alock->set_non_esc_obj();
2269 }
2270 }
2271 }
2272 return;
2273 }
2274
2275 // Create new "eliminated" BoxLock node and use it in monitor debug info
2276 // instead of oldbox for the same object.
2277 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
2278
2279 // Note: BoxLock node is marked eliminated only here and it is used
2280 // to indicate that all associated lock and unlock nodes are marked
2281 // for elimination.
2282 newbox->set_local(); // This verifies correct state of BoxLock
2283 newbox->set_eliminated();
2284 transform_later(newbox);
2285
2286 // Replace old box node with new box for all users of the same object.
2287 for (uint i = 0; i < oldbox->outcnt();) {
2288 bool next_edge = true;
2289
2290 Node* u = oldbox->raw_out(i);
2291 if (u->is_AbstractLock()) {
2292 AbstractLockNode* alock = u->as_AbstractLock();
2293 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
2294 // Replace Box and mark eliminated all related locks and unlocks.
2295 #ifdef ASSERT
2296 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5");
2297 #endif
2298 alock->set_non_esc_obj();
2299 _igvn.rehash_node_delayed(alock);
2300 alock->set_box_node(newbox);
2301 next_edge = false;
2302 }
2303 }
2304 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
2305 FastLockNode* flock = u->as_FastLock();
2306 assert(flock->box_node() == oldbox, "sanity");
2307 _igvn.rehash_node_delayed(flock);
2308 flock->set_box_node(newbox);
2309 next_edge = false;
2310 }
2311
2312 // Replace old box in monitor debug info.
2313 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
2314 SafePointNode* sfn = u->as_SafePoint();
2315 JVMState* youngest_jvms = sfn->jvms();
2316 int max_depth = youngest_jvms->depth();
2317 for (int depth = 1; depth <= max_depth; depth++) {
2318 JVMState* jvms = youngest_jvms->of_depth(depth);
2319 int num_mon = jvms->nof_monitors();
2320 // Loop over monitors
2321 for (int idx = 0; idx < num_mon; idx++) {
2322 Node* obj_node = sfn->monitor_obj(jvms, idx);
2323 Node* box_node = sfn->monitor_box(jvms, idx);
2324 if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
2325 int j = jvms->monitor_box_offset(idx);
2326 _igvn.replace_input_of(u, j, newbox);
2327 next_edge = false;
2328 }
2329 }
2330 }
2331 }
2332 if (next_edge) i++;
2333 }
2334 }
2335
2336 //-----------------------mark_eliminated_locking_nodes-----------------------
2337 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
2338 if (!alock->is_balanced()) {
2339 return; // Can't do any more elimination for this locking region
2340 }
2341 if (EliminateNestedLocks) {
2342 if (alock->is_nested()) {
2343 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
2344 return;
2345 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
2346 // Only Lock node has JVMState needed here.
2347 // Not that preceding claim is documented anywhere else.
2348 if (alock->jvms() != nullptr) {
2349 if (alock->as_Lock()->is_nested_lock_region()) {
2350 // Mark eliminated related nested locks and unlocks.
2351 Node* obj = alock->obj_node();
2352 BoxLockNode* box_node = alock->box_node()->as_BoxLock();
2353 assert(!box_node->is_eliminated(), "should not be marked yet");
2354 // Note: BoxLock node is marked eliminated only here
2355 // and it is used to indicate that all associated lock
2356 // and unlock nodes are marked for elimination.
2357 box_node->set_eliminated(); // Box's hash is always NO_HASH here
2358 for (uint i = 0; i < box_node->outcnt(); i++) {
2359 Node* u = box_node->raw_out(i);
2360 if (u->is_AbstractLock()) {
2361 alock = u->as_AbstractLock();
2362 if (alock->box_node() == box_node) {
2363 // Verify that this Box is referenced only by related locks.
2364 assert(alock->obj_node()->eqv_uncast(obj), "");
2365 // Mark all related locks and unlocks.
2366 #ifdef ASSERT
2367 alock->log_lock_optimization(C, "eliminate_lock_set_nested");
2368 #endif
2369 alock->set_nested();
2370 }
2371 }
2372 }
2373 } else {
2374 #ifdef ASSERT
2375 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region");
2376 if (C->log() != nullptr)
2377 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output
2378 #endif
2379 }
2380 }
2381 return;
2382 }
2383 // Process locks for non escaping object
2384 assert(alock->is_non_esc_obj(), "");
2385 } // EliminateNestedLocks
2386
2387 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2388 // Look for all locks of this object and mark them and
2389 // corresponding BoxLock nodes as eliminated.
2390 Node* obj = alock->obj_node();
2391 for (uint j = 0; j < obj->outcnt(); j++) {
2392 Node* o = obj->raw_out(j);
2393 if (o->is_AbstractLock() &&
2394 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2395 alock = o->as_AbstractLock();
2396 Node* box = alock->box_node();
2397 // Replace old box node with new eliminated box for all users
2398 // of the same object and mark related locks as eliminated.
2399 mark_eliminated_box(box, obj);
2400 }
2401 }
2402 }
2403 }
2404
2405 // we have determined that this lock/unlock can be eliminated, we simply
2406 // eliminate the node without expanding it.
2407 //
2408 // Note: The membar's associated with the lock/unlock are currently not
2409 // eliminated. This should be investigated as a future enhancement.
2410 //
2411 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2412
2413 if (!alock->is_eliminated()) {
2414 return false;
2415 }
2416 #ifdef ASSERT
2417 if (!alock->is_coarsened()) {
2418 // Check that new "eliminated" BoxLock node is created.
2419 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2420 assert(oldbox->is_eliminated(), "should be done already");
2421 }
2422 #endif
2423
2424 alock->log_lock_optimization(C, "eliminate_lock");
2425
2426 #ifndef PRODUCT
2427 if (PrintEliminateLocks) {
2428 tty->print_cr("++++ Eliminated: %d %s '%s'", alock->_idx, (alock->is_Lock() ? "Lock" : "Unlock"), alock->kind_as_string());
2429 }
2430 #endif
2431
2432 Node* mem = alock->in(TypeFunc::Memory);
2433 Node* ctrl = alock->in(TypeFunc::Control);
2434 guarantee(ctrl != nullptr, "missing control projection, cannot replace_node() with null");
2435
2436 _callprojs = alock->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2437 // There are 2 projections from the lock. The lock node will
2438 // be deleted when its last use is subsumed below.
2439 assert(alock->outcnt() == 2 &&
2440 _callprojs->fallthrough_proj != nullptr &&
2441 _callprojs->fallthrough_memproj != nullptr,
2442 "Unexpected projections from Lock/Unlock");
2443
2444 Node* fallthroughproj = _callprojs->fallthrough_proj;
2445 Node* memproj_fallthrough = _callprojs->fallthrough_memproj;
2446
2447 // The memory projection from a lock/unlock is RawMem
2448 // The input to a Lock is merged memory, so extract its RawMem input
2449 // (unless the MergeMem has been optimized away.)
2450 if (alock->is_Lock()) {
2451 // Search for MemBarAcquireLock node and delete it also.
2452 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2453 assert(membar != nullptr && membar->Opcode() == Op_MemBarAcquireLock, "");
2454 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2455 Node* memproj = membar->proj_out(TypeFunc::Memory);
2456 _igvn.replace_node(ctrlproj, fallthroughproj);
2457 _igvn.replace_node(memproj, memproj_fallthrough);
2458
2459 // Delete FastLock node also if this Lock node is unique user
2460 // (a loop peeling may clone a Lock node).
2461 Node* flock = alock->as_Lock()->fastlock_node();
2462 if (flock->outcnt() == 1) {
2463 assert(flock->unique_out() == alock, "sanity");
2464 _igvn.replace_node(flock, top());
2465 }
2466 }
2467
2468 // Search for MemBarReleaseLock node and delete it also.
2469 if (alock->is_Unlock() && ctrl->is_Proj() && ctrl->in(0)->is_MemBar()) {
2470 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2471 assert(membar->Opcode() == Op_MemBarReleaseLock &&
2472 mem->is_Proj() && membar == mem->in(0), "");
2473 _igvn.replace_node(fallthroughproj, ctrl);
2474 _igvn.replace_node(memproj_fallthrough, mem);
2475 fallthroughproj = ctrl;
2476 memproj_fallthrough = mem;
2477 ctrl = membar->in(TypeFunc::Control);
2478 mem = membar->in(TypeFunc::Memory);
2479 }
2480
2481 _igvn.replace_node(fallthroughproj, ctrl);
2482 _igvn.replace_node(memproj_fallthrough, mem);
2483 return true;
2484 }
2485
2486
2487 //------------------------------expand_lock_node----------------------
2488 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2489
2490 Node* ctrl = lock->in(TypeFunc::Control);
2491 Node* mem = lock->in(TypeFunc::Memory);
2492 Node* obj = lock->obj_node();
2493 Node* box = lock->box_node();
2494 Node* flock = lock->fastlock_node();
2495
2496 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2497
2498 // Make the merge point
2499 Node *region;
2500 Node *mem_phi;
2501 Node *slow_path;
2502
2503 region = new RegionNode(3);
2504 // create a Phi for the memory state
2505 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2506
2507 // Optimize test; set region slot 2
2508 slow_path = opt_bits_test(ctrl, region, 2, flock);
2509 mem_phi->init_req(2, mem);
2510
2511 // Make slow path call
2512 CallNode* call = make_slow_call(lock, OptoRuntime::complete_monitor_enter_Type(),
2513 OptoRuntime::complete_monitor_locking_Java(), nullptr, slow_path,
2514 obj, box, nullptr);
2515
2516 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2517
2518 // Slow path can only throw asynchronous exceptions, which are always
2519 // de-opted. So the compiler thinks the slow-call can never throw an
2520 // exception. If it DOES throw an exception we would need the debug
2521 // info removed first (since if it throws there is no monitor).
2522 assert(_callprojs->fallthrough_ioproj == nullptr && _callprojs->catchall_ioproj == nullptr &&
2523 _callprojs->catchall_memproj == nullptr && _callprojs->catchall_catchproj == nullptr, "Unexpected projection from Lock");
2524
2525 // Capture slow path
2526 // disconnect fall-through projection from call and create a new one
2527 // hook up users of fall-through projection to region
2528 Node *slow_ctrl = _callprojs->fallthrough_proj->clone();
2529 transform_later(slow_ctrl);
2530 _igvn.hash_delete(_callprojs->fallthrough_proj);
2531 _callprojs->fallthrough_proj->disconnect_inputs(C);
2532 region->init_req(1, slow_ctrl);
2533 // region inputs are now complete
2534 transform_later(region);
2535 _igvn.replace_node(_callprojs->fallthrough_proj, region);
2536
2537 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2538
2539 mem_phi->init_req(1, memproj);
2540
2541 transform_later(mem_phi);
2542
2543 _igvn.replace_node(_callprojs->fallthrough_memproj, mem_phi);
2544 }
2545
2546 //------------------------------expand_unlock_node----------------------
2547 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2548
2549 Node* ctrl = unlock->in(TypeFunc::Control);
2550 Node* mem = unlock->in(TypeFunc::Memory);
2551 Node* obj = unlock->obj_node();
2552 Node* box = unlock->box_node();
2553
2554 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2555
2556 // No need for a null check on unlock
2557
2558 // Make the merge point
2559 Node *region;
2560 Node *mem_phi;
2561
2562 region = new RegionNode(3);
2563 // create a Phi for the memory state
2564 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2565
2566 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2567 funlock = transform_later( funlock )->as_FastUnlock();
2568 // Optimize test; set region slot 2
2569 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock);
2570 Node *thread = transform_later(new ThreadLocalNode());
2571
2572 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2573 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2574 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2575
2576 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2577 assert(_callprojs->fallthrough_ioproj == nullptr && _callprojs->catchall_ioproj == nullptr &&
2578 _callprojs->catchall_memproj == nullptr && _callprojs->catchall_catchproj == nullptr, "Unexpected projection from Lock");
2579
2580 // No exceptions for unlocking
2581 // Capture slow path
2582 // disconnect fall-through projection from call and create a new one
2583 // hook up users of fall-through projection to region
2584 Node *slow_ctrl = _callprojs->fallthrough_proj->clone();
2585 transform_later(slow_ctrl);
2586 _igvn.hash_delete(_callprojs->fallthrough_proj);
2587 _callprojs->fallthrough_proj->disconnect_inputs(C);
2588 region->init_req(1, slow_ctrl);
2589 // region inputs are now complete
2590 transform_later(region);
2591 _igvn.replace_node(_callprojs->fallthrough_proj, region);
2592
2593 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2594 mem_phi->init_req(1, memproj );
2595 mem_phi->init_req(2, mem);
2596 transform_later(mem_phi);
2597
2598 _igvn.replace_node(_callprojs->fallthrough_memproj, mem_phi);
2599 }
2600
2601 // An inline type might be returned from the call but we don't know its
2602 // type. Either we get a buffered inline type (and nothing needs to be done)
2603 // or one of the values being returned is the klass of the inline type
2604 // and we need to allocate an inline type instance of that type and
2605 // initialize it with other values being returned. In that case, we
2606 // first try a fast path allocation and initialize the value with the
2607 // inline klass's pack handler or we fall back to a runtime call.
2608 void PhaseMacroExpand::expand_mh_intrinsic_return(CallStaticJavaNode* call) {
2609 assert(call->method()->is_method_handle_intrinsic(), "must be a method handle intrinsic call");
2610 Node* ret = call->proj_out_or_null(TypeFunc::Parms);
2611 if (ret == nullptr) {
2612 return;
2613 }
2614 const TypeFunc* tf = call->_tf;
2615 const TypeTuple* domain = OptoRuntime::store_inline_type_fields_Type()->domain_cc();
2616 const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain);
2617 call->_tf = new_tf;
2618 // Make sure the change of type is applied before projections are processed by igvn
2619 _igvn.set_type(call, call->Value(&_igvn));
2620 _igvn.set_type(ret, ret->Value(&_igvn));
2621
2622 // Before any new projection is added:
2623 CallProjections* projs = call->extract_projections(true, true);
2624
2625 // Create temporary hook nodes that will be replaced below.
2626 // Add an input to prevent hook nodes from being dead.
2627 Node* ctl = new Node(call);
2628 Node* mem = new Node(ctl);
2629 Node* io = new Node(ctl);
2630 Node* ex_ctl = new Node(ctl);
2631 Node* ex_mem = new Node(ctl);
2632 Node* ex_io = new Node(ctl);
2633 Node* res = new Node(ctl);
2634
2635 // Allocate a new buffered inline type only if a new one is not returned
2636 Node* cast = transform_later(new CastP2XNode(ctl, res));
2637 Node* mask = MakeConX(0x1);
2638 Node* masked = transform_later(new AndXNode(cast, mask));
2639 Node* cmp = transform_later(new CmpXNode(masked, mask));
2640 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
2641 IfNode* allocation_iff = new IfNode(ctl, bol, PROB_MAX, COUNT_UNKNOWN);
2642 transform_later(allocation_iff);
2643 Node* allocation_ctl = transform_later(new IfTrueNode(allocation_iff));
2644 Node* no_allocation_ctl = transform_later(new IfFalseNode(allocation_iff));
2645 Node* no_allocation_res = transform_later(new CheckCastPPNode(no_allocation_ctl, res, TypeInstPtr::BOTTOM));
2646
2647 // Try to allocate a new buffered inline instance either from TLAB or eden space
2648 Node* needgc_ctrl = nullptr; // needgc means slowcase, i.e. allocation failed
2649 CallLeafNoFPNode* handler_call;
2650 const bool alloc_in_place = UseTLAB;
2651 if (alloc_in_place) {
2652 Node* fast_oop_ctrl = nullptr;
2653 Node* fast_oop_rawmem = nullptr;
2654 Node* mask2 = MakeConX(-2);
2655 Node* masked2 = transform_later(new AndXNode(cast, mask2));
2656 Node* rawklassptr = transform_later(new CastX2PNode(masked2));
2657 Node* klass_node = transform_later(new CheckCastPPNode(allocation_ctl, rawklassptr, TypeInstKlassPtr::OBJECT_OR_NULL));
2658 Node* layout_val = make_load(nullptr, mem, klass_node, in_bytes(Klass::layout_helper_offset()), TypeInt::INT, T_INT);
2659 Node* size_in_bytes = ConvI2X(layout_val);
2660 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2661 Node* fast_oop = bs->obj_allocate(this, mem, allocation_ctl, size_in_bytes, io, needgc_ctrl,
2662 fast_oop_ctrl, fast_oop_rawmem,
2663 AllocateInstancePrefetchLines);
2664 // Allocation succeed, initialize buffered inline instance header firstly,
2665 // and then initialize its fields with an inline class specific handler
2666 Node* mark_word_node;
2667 if (UseCompactObjectHeaders) {
2668 // COH: We need to load the prototype from the klass at runtime since it encodes the klass pointer already.
2669 mark_word_node = make_load(fast_oop_ctrl, fast_oop_rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2670 } else {
2671 // Otherwise, use the static prototype.
2672 mark_word_node = makecon(TypeRawPtr::make((address)markWord::inline_type_prototype().value()));
2673 }
2674
2675 fast_oop_rawmem = make_store(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::mark_offset_in_bytes(), mark_word_node, T_ADDRESS);
2676 if (!UseCompactObjectHeaders) {
2677 // COH: Everything is encoded in the mark word, so nothing left to do.
2678 fast_oop_rawmem = make_store(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
2679 if (UseCompressedClassPointers) {
2680 fast_oop_rawmem = make_store(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::klass_gap_offset_in_bytes(), intcon(0), T_INT);
2681 }
2682 }
2683 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);
2684 Node* pack_handler = make_load(fast_oop_ctrl, fast_oop_rawmem, fixed_block, in_bytes(InlineKlass::pack_handler_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2685 handler_call = new CallLeafNoFPNode(OptoRuntime::pack_inline_type_Type(),
2686 nullptr,
2687 "pack handler",
2688 TypeRawPtr::BOTTOM);
2689 handler_call->init_req(TypeFunc::Control, fast_oop_ctrl);
2690 handler_call->init_req(TypeFunc::Memory, fast_oop_rawmem);
2691 handler_call->init_req(TypeFunc::I_O, top());
2692 handler_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2693 handler_call->init_req(TypeFunc::ReturnAdr, top());
2694 handler_call->init_req(TypeFunc::Parms, pack_handler);
2695 handler_call->init_req(TypeFunc::Parms+1, fast_oop);
2696 } else {
2697 needgc_ctrl = allocation_ctl;
2698 }
2699
2700 // Allocation failed, fall back to a runtime call
2701 CallStaticJavaNode* slow_call = new CallStaticJavaNode(OptoRuntime::store_inline_type_fields_Type(),
2702 StubRoutines::store_inline_type_fields_to_buf(),
2703 "store_inline_type_fields",
2704 TypePtr::BOTTOM);
2705 slow_call->init_req(TypeFunc::Control, needgc_ctrl);
2706 slow_call->init_req(TypeFunc::Memory, mem);
2707 slow_call->init_req(TypeFunc::I_O, io);
2708 slow_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2709 slow_call->init_req(TypeFunc::ReturnAdr, call->in(TypeFunc::ReturnAdr));
2710 slow_call->init_req(TypeFunc::Parms, res);
2711
2712 Node* slow_ctl = transform_later(new ProjNode(slow_call, TypeFunc::Control));
2713 Node* slow_mem = transform_later(new ProjNode(slow_call, TypeFunc::Memory));
2714 Node* slow_io = transform_later(new ProjNode(slow_call, TypeFunc::I_O));
2715 Node* slow_res = transform_later(new ProjNode(slow_call, TypeFunc::Parms));
2716 Node* slow_catc = transform_later(new CatchNode(slow_ctl, slow_io, 2));
2717 Node* slow_norm = transform_later(new CatchProjNode(slow_catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci));
2718 Node* slow_excp = transform_later(new CatchProjNode(slow_catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci));
2719
2720 Node* ex_r = new RegionNode(3);
2721 Node* ex_mem_phi = new PhiNode(ex_r, Type::MEMORY, TypePtr::BOTTOM);
2722 Node* ex_io_phi = new PhiNode(ex_r, Type::ABIO);
2723 ex_r->init_req(1, slow_excp);
2724 ex_mem_phi->init_req(1, slow_mem);
2725 ex_io_phi->init_req(1, slow_io);
2726 ex_r->init_req(2, ex_ctl);
2727 ex_mem_phi->init_req(2, ex_mem);
2728 ex_io_phi->init_req(2, ex_io);
2729 transform_later(ex_r);
2730 transform_later(ex_mem_phi);
2731 transform_later(ex_io_phi);
2732
2733 // We don't know how many values are returned. This assumes the
2734 // worst case, that all available registers are used.
2735 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2736 if (domain->field_at(i) == Type::HALF) {
2737 slow_call->init_req(i, top());
2738 if (alloc_in_place) {
2739 handler_call->init_req(i+1, top());
2740 }
2741 continue;
2742 }
2743 Node* proj = transform_later(new ProjNode(call, i));
2744 slow_call->init_req(i, proj);
2745 if (alloc_in_place) {
2746 handler_call->init_req(i+1, proj);
2747 }
2748 }
2749 // We can safepoint at that new call
2750 slow_call->copy_call_debug_info(&_igvn, call);
2751 transform_later(slow_call);
2752 if (alloc_in_place) {
2753 transform_later(handler_call);
2754 }
2755
2756 Node* fast_ctl = nullptr;
2757 Node* fast_res = nullptr;
2758 MergeMemNode* fast_mem = nullptr;
2759 if (alloc_in_place) {
2760 fast_ctl = transform_later(new ProjNode(handler_call, TypeFunc::Control));
2761 Node* rawmem = transform_later(new ProjNode(handler_call, TypeFunc::Memory));
2762 fast_res = transform_later(new ProjNode(handler_call, TypeFunc::Parms));
2763 fast_mem = MergeMemNode::make(mem);
2764 fast_mem->set_memory_at(Compile::AliasIdxRaw, rawmem);
2765 transform_later(fast_mem);
2766 }
2767
2768 Node* r = new RegionNode(alloc_in_place ? 4 : 3);
2769 Node* mem_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2770 Node* io_phi = new PhiNode(r, Type::ABIO);
2771 Node* res_phi = new PhiNode(r, TypeInstPtr::BOTTOM);
2772 r->init_req(1, no_allocation_ctl);
2773 mem_phi->init_req(1, mem);
2774 io_phi->init_req(1, io);
2775 res_phi->init_req(1, no_allocation_res);
2776 r->init_req(2, slow_norm);
2777 mem_phi->init_req(2, slow_mem);
2778 io_phi->init_req(2, slow_io);
2779 res_phi->init_req(2, slow_res);
2780 if (alloc_in_place) {
2781 r->init_req(3, fast_ctl);
2782 mem_phi->init_req(3, fast_mem);
2783 io_phi->init_req(3, io);
2784 res_phi->init_req(3, fast_res);
2785 }
2786 transform_later(r);
2787 transform_later(mem_phi);
2788 transform_later(io_phi);
2789 transform_later(res_phi);
2790
2791 // Do not let stores that initialize this buffer be reordered with a subsequent
2792 // store that would make this buffer accessible by other threads.
2793 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
2794 transform_later(mb);
2795 mb->init_req(TypeFunc::Memory, mem_phi);
2796 mb->init_req(TypeFunc::Control, r);
2797 r = new ProjNode(mb, TypeFunc::Control);
2798 transform_later(r);
2799 mem_phi = new ProjNode(mb, TypeFunc::Memory);
2800 transform_later(mem_phi);
2801
2802 assert(projs->nb_resproj == 1, "unexpected number of results");
2803 _igvn.replace_in_uses(projs->fallthrough_catchproj, r);
2804 _igvn.replace_in_uses(projs->fallthrough_memproj, mem_phi);
2805 _igvn.replace_in_uses(projs->fallthrough_ioproj, io_phi);
2806 _igvn.replace_in_uses(projs->resproj[0], res_phi);
2807 _igvn.replace_in_uses(projs->catchall_catchproj, ex_r);
2808 _igvn.replace_in_uses(projs->catchall_memproj, ex_mem_phi);
2809 _igvn.replace_in_uses(projs->catchall_ioproj, ex_io_phi);
2810 // The CatchNode should not use the ex_io_phi. Re-connect it to the catchall_ioproj.
2811 Node* cn = projs->fallthrough_catchproj->in(0);
2812 _igvn.replace_input_of(cn, 1, projs->catchall_ioproj);
2813
2814 _igvn.replace_node(ctl, projs->fallthrough_catchproj);
2815 _igvn.replace_node(mem, projs->fallthrough_memproj);
2816 _igvn.replace_node(io, projs->fallthrough_ioproj);
2817 _igvn.replace_node(res, projs->resproj[0]);
2818 _igvn.replace_node(ex_ctl, projs->catchall_catchproj);
2819 _igvn.replace_node(ex_mem, projs->catchall_memproj);
2820 _igvn.replace_node(ex_io, projs->catchall_ioproj);
2821 }
2822
2823 void PhaseMacroExpand::expand_subtypecheck_node(SubTypeCheckNode *check) {
2824 assert(check->in(SubTypeCheckNode::Control) == nullptr, "should be pinned");
2825 Node* bol = check->unique_out();
2826 Node* obj_or_subklass = check->in(SubTypeCheckNode::ObjOrSubKlass);
2827 Node* superklass = check->in(SubTypeCheckNode::SuperKlass);
2828 assert(bol->is_Bool() && bol->as_Bool()->_test._test == BoolTest::ne, "unexpected bool node");
2829
2830 for (DUIterator_Last imin, i = bol->last_outs(imin); i >= imin; --i) {
2831 Node* iff = bol->last_out(i);
2832 assert(iff->is_If(), "where's the if?");
2833
2834 if (iff->in(0)->is_top()) {
2835 _igvn.replace_input_of(iff, 1, C->top());
2836 continue;
2837 }
2838
2839 Node* iftrue = iff->as_If()->proj_out(1);
2840 Node* iffalse = iff->as_If()->proj_out(0);
2841 Node* ctrl = iff->in(0);
2842
2843 Node* subklass = nullptr;
2844 if (_igvn.type(obj_or_subklass)->isa_klassptr()) {
2845 subklass = obj_or_subklass;
2846 } else {
2847 Node* k_adr = basic_plus_adr(obj_or_subklass, oopDesc::klass_offset_in_bytes());
2848 subklass = _igvn.transform(LoadKlassNode::make(_igvn, C->immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
2849 }
2850
2851 Node* not_subtype_ctrl = Phase::gen_subtype_check(subklass, superklass, &ctrl, nullptr, _igvn, check->method(), check->bci());
2852
2853 _igvn.replace_input_of(iff, 0, C->top());
2854 _igvn.replace_node(iftrue, not_subtype_ctrl);
2855 _igvn.replace_node(iffalse, ctrl);
2856 }
2857 _igvn.replace_node(check, C->top());
2858 }
2859
2860 // FlatArrayCheckNode (array1 array2 ...) is expanded into:
2861 //
2862 // long mark = array1.mark | array2.mark | ...;
2863 // long locked_bit = markWord::unlocked_value & array1.mark & array2.mark & ...;
2864 // if (locked_bit == 0) {
2865 // // One array is locked, load prototype header from the klass
2866 // mark = array1.klass.proto | array2.klass.proto | ...
2867 // }
2868 // if ((mark & markWord::flat_array_bit_in_place) == 0) {
2869 // ...
2870 // }
2871 void PhaseMacroExpand::expand_flatarraycheck_node(FlatArrayCheckNode* check) {
2872 bool array_inputs = _igvn.type(check->in(FlatArrayCheckNode::ArrayOrKlass))->isa_oopptr() != nullptr;
2873 if (array_inputs) {
2874 Node* mark = MakeConX(0);
2875 Node* locked_bit = MakeConX(markWord::unlocked_value);
2876 Node* mem = check->in(FlatArrayCheckNode::Memory);
2877 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
2878 Node* ary = check->in(i);
2879 const TypeOopPtr* t = _igvn.type(ary)->isa_oopptr();
2880 assert(t != nullptr, "Mixing array and klass inputs");
2881 assert(!t->is_flat() && !t->is_not_flat(), "Should have been optimized out");
2882 Node* mark_adr = basic_plus_adr(ary, oopDesc::mark_offset_in_bytes());
2883 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));
2884 mark = _igvn.transform(new OrXNode(mark, mark_load));
2885 locked_bit = _igvn.transform(new AndXNode(locked_bit, mark_load));
2886 }
2887 assert(!mark->is_Con(), "Should have been optimized out");
2888 Node* cmp = _igvn.transform(new CmpXNode(locked_bit, MakeConX(0)));
2889 Node* is_unlocked = _igvn.transform(new BoolNode(cmp, BoolTest::ne));
2890
2891 // BoolNode might be shared, replace each if user
2892 Node* old_bol = check->unique_out();
2893 assert(old_bol->is_Bool() && old_bol->as_Bool()->_test._test == BoolTest::ne, "unexpected condition");
2894 for (DUIterator_Last imin, i = old_bol->last_outs(imin); i >= imin; --i) {
2895 IfNode* old_iff = old_bol->last_out(i)->as_If();
2896 Node* ctrl = old_iff->in(0);
2897 RegionNode* region = new RegionNode(3);
2898 Node* mark_phi = new PhiNode(region, TypeX_X);
2899
2900 // Check if array is unlocked
2901 IfNode* iff = _igvn.transform(new IfNode(ctrl, is_unlocked, PROB_MAX, COUNT_UNKNOWN))->as_If();
2902
2903 // Unlocked: Use bits from mark word
2904 region->init_req(1, _igvn.transform(new IfTrueNode(iff)));
2905 mark_phi->init_req(1, mark);
2906
2907 // Locked: Load prototype header from klass
2908 ctrl = _igvn.transform(new IfFalseNode(iff));
2909 Node* proto = MakeConX(0);
2910 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
2911 Node* ary = check->in(i);
2912 // Make loads control dependent to make sure they are only executed if array is locked
2913 Node* klass_adr = basic_plus_adr(ary, oopDesc::klass_offset_in_bytes());
2914 Node* klass = _igvn.transform(LoadKlassNode::make(_igvn, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
2915 Node* proto_adr = basic_plus_adr(klass, in_bytes(Klass::prototype_header_offset()));
2916 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));
2917 proto = _igvn.transform(new OrXNode(proto, proto_load));
2918 }
2919 region->init_req(2, ctrl);
2920 mark_phi->init_req(2, proto);
2921
2922 // Check if flat array bits are set
2923 Node* mask = MakeConX(markWord::flat_array_bit_in_place);
2924 Node* masked = _igvn.transform(new AndXNode(_igvn.transform(mark_phi), mask));
2925 cmp = _igvn.transform(new CmpXNode(masked, MakeConX(0)));
2926 Node* is_not_flat = _igvn.transform(new BoolNode(cmp, BoolTest::eq));
2927
2928 ctrl = _igvn.transform(region);
2929 iff = _igvn.transform(new IfNode(ctrl, is_not_flat, PROB_MAX, COUNT_UNKNOWN))->as_If();
2930 _igvn.replace_node(old_iff, iff);
2931 }
2932 _igvn.replace_node(check, C->top());
2933 } else {
2934 // Fall back to layout helper check
2935 Node* lhs = intcon(0);
2936 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
2937 Node* array_or_klass = check->in(i);
2938 Node* klass = nullptr;
2939 const TypePtr* t = _igvn.type(array_or_klass)->is_ptr();
2940 assert(!t->is_flat() && !t->is_not_flat(), "Should have been optimized out");
2941 if (t->isa_oopptr() != nullptr) {
2942 Node* klass_adr = basic_plus_adr(array_or_klass, oopDesc::klass_offset_in_bytes());
2943 klass = transform_later(LoadKlassNode::make(_igvn, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
2944 } else {
2945 assert(t->isa_klassptr(), "Unexpected input type");
2946 klass = array_or_klass;
2947 }
2948 Node* lh_addr = basic_plus_adr(klass, in_bytes(Klass::layout_helper_offset()));
2949 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));
2950 lhs = _igvn.transform(new OrINode(lhs, lh_val));
2951 }
2952 Node* masked = transform_later(new AndINode(lhs, intcon(Klass::_lh_array_tag_flat_value_bit_inplace)));
2953 Node* cmp = transform_later(new CmpINode(masked, intcon(0)));
2954 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
2955 Node* m2b = transform_later(new Conv2BNode(masked));
2956 // The matcher expects the input to If nodes to be produced by a Bool(CmpI..)
2957 // pattern, but the input to other potential users (e.g. Phi) to be some
2958 // other pattern (e.g. a Conv2B node, possibly idealized as a CMoveI).
2959 Node* old_bol = check->unique_out();
2960 for (DUIterator_Last imin, i = old_bol->last_outs(imin); i >= imin; --i) {
2961 Node* user = old_bol->last_out(i);
2962 for (uint j = 0; j < user->req(); j++) {
2963 Node* n = user->in(j);
2964 if (n == old_bol) {
2965 _igvn.replace_input_of(user, j, user->is_If() ? bol : m2b);
2966 }
2967 }
2968 }
2969 _igvn.replace_node(check, C->top());
2970 }
2971 }
2972
2973 // Perform refining of strip mined loop nodes in the macro nodes list.
2974 void PhaseMacroExpand::refine_strip_mined_loop_macro_nodes() {
2975 for (int i = C->macro_count(); i > 0; i--) {
2976 Node* n = C->macro_node(i - 1);
2977 if (n->is_OuterStripMinedLoop()) {
2978 n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn);
2979 }
2980 }
2981 }
2982
2983 //---------------------------eliminate_macro_nodes----------------------
2984 // Eliminate scalar replaced allocations and associated locks.
2985 void PhaseMacroExpand::eliminate_macro_nodes(bool eliminate_locks) {
2986 if (C->macro_count() == 0) {
2987 return;
2988 }
2989
2990 if (StressMacroElimination) {
2991 C->shuffle_macro_nodes();
2992 }
2993 NOT_PRODUCT(int membar_before = count_MemBar(C);)
2994
2995 int iteration = 0;
2996 while (C->macro_count() > 0) {
2997 if (iteration++ > 100) {
2998 assert(false, "Too slow convergence of macro elimination");
2999 break;
3000 }
3001
3002 // Postpone lock elimination to after EA when most allocations are eliminated
3003 // because they might block lock elimination if their escape state isn't
3004 // determined yet and we only got one chance at eliminating the lock.
3005 if (eliminate_locks) {
3006 // Before elimination may re-mark (change to Nested or NonEscObj)
3007 // all associated (same box and obj) lock and unlock nodes.
3008 int cnt = C->macro_count();
3009 for (int i=0; i < cnt; i++) {
3010 Node *n = C->macro_node(i);
3011 if (n->is_AbstractLock()) { // Lock and Unlock nodes
3012 mark_eliminated_locking_nodes(n->as_AbstractLock());
3013 }
3014 }
3015 // Re-marking may break consistency of Coarsened locks.
3016 if (!C->coarsened_locks_consistent()) {
3017 return; // recompile without Coarsened locks if broken
3018 } else {
3019 // After coarsened locks are eliminated locking regions
3020 // become unbalanced. We should not execute any more
3021 // locks elimination optimizations on them.
3022 C->mark_unbalanced_boxes();
3023 }
3024 }
3025
3026 bool progress = false;
3027 for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once
3028 Node* n = C->macro_node(i - 1);
3029 bool success = false;
3030 DEBUG_ONLY(int old_macro_count = C->macro_count();)
3031 switch (n->class_id()) {
3032 case Node::Class_Allocate:
3033 case Node::Class_AllocateArray:
3034 success = eliminate_allocate_node(n->as_Allocate());
3035 #ifndef PRODUCT
3036 if (success && PrintOptoStatistics) {
3037 Atomic::inc(&PhaseMacroExpand::_objs_scalar_replaced_counter);
3038 }
3039 #endif
3040 break;
3041 case Node::Class_CallStaticJava: {
3042 CallStaticJavaNode* call = n->as_CallStaticJava();
3043 if (!call->method()->is_method_handle_intrinsic()) {
3044 success = eliminate_boxing_node(n->as_CallStaticJava());
3045 }
3046 break;
3047 }
3048 case Node::Class_Lock:
3049 case Node::Class_Unlock:
3050 if (eliminate_locks) {
3051 success = eliminate_locking_node(n->as_AbstractLock());
3052 #ifndef PRODUCT
3053 if (success && PrintOptoStatistics) {
3054 Atomic::inc(&PhaseMacroExpand::_monitor_objects_removed_counter);
3055 }
3056 #endif
3057 }
3058 break;
3059 case Node::Class_ArrayCopy:
3060 break;
3061 case Node::Class_OuterStripMinedLoop:
3062 break;
3063 case Node::Class_SubTypeCheck:
3064 break;
3065 case Node::Class_Opaque1:
3066 break;
3067 case Node::Class_FlatArrayCheck:
3068 break;
3069 default:
3070 assert(n->Opcode() == Op_LoopLimit ||
3071 n->Opcode() == Op_ModD ||
3072 n->Opcode() == Op_ModF ||
3073 n->is_OpaqueNotNull() ||
3074 n->is_OpaqueInitializedAssertionPredicate() ||
3075 n->Opcode() == Op_MaxL ||
3076 n->Opcode() == Op_MinL ||
3077 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n),
3078 "unknown node type in macro list");
3079 }
3080 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
3081 progress = progress || success;
3082 if (success) {
3083 C->print_method(PHASE_AFTER_MACRO_ELIMINATION_STEP, 5, n);
3084 }
3085 }
3086
3087 // Ensure the graph after PhaseMacroExpand::eliminate_macro_nodes is canonical (no igvn
3088 // transformation is pending). If an allocation is used only in safepoints, elimination of
3089 // other macro nodes can remove all these safepoints, allowing the allocation to be removed.
3090 // Hence after igvn we retry removing macro nodes if some progress that has been made in this
3091 // iteration.
3092 _igvn.set_delay_transform(false);
3093 _igvn.optimize();
3094 if (C->failing()) {
3095 return;
3096 }
3097 _igvn.set_delay_transform(true);
3098
3099 if (!progress) {
3100 break;
3101 }
3102 }
3103 #ifndef PRODUCT
3104 if (PrintOptoStatistics) {
3105 int membar_after = count_MemBar(C);
3106 Atomic::add(&PhaseMacroExpand::_memory_barriers_removed_counter, membar_before - membar_after);
3107 }
3108 #endif
3109 }
3110
3111 void PhaseMacroExpand::eliminate_opaque_looplimit_macro_nodes() {
3112 if (C->macro_count() == 0) {
3113 return;
3114 }
3115 refine_strip_mined_loop_macro_nodes();
3116 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
3117 bool progress = true;
3118 while (progress) {
3119 progress = false;
3120 for (int i = C->macro_count(); i > 0; i--) {
3121 Node* n = C->macro_node(i-1);
3122 bool success = false;
3123 DEBUG_ONLY(int old_macro_count = C->macro_count();)
3124 if (n->Opcode() == Op_LoopLimit) {
3125 // Remove it from macro list and put on IGVN worklist to optimize.
3126 C->remove_macro_node(n);
3127 _igvn._worklist.push(n);
3128 success = true;
3129 } else if (n->Opcode() == Op_CallStaticJava) {
3130 CallStaticJavaNode* call = n->as_CallStaticJava();
3131 if (!call->method()->is_method_handle_intrinsic()) {
3132 // Remove it from macro list and put on IGVN worklist to optimize.
3133 C->remove_macro_node(n);
3134 _igvn._worklist.push(n);
3135 success = true;
3136 }
3137 } else if (n->is_Opaque1()) {
3138 _igvn.replace_node(n, n->in(1));
3139 success = true;
3140 } else if (n->is_OpaqueNotNull()) {
3141 // Tests with OpaqueNotNull nodes are implicitly known to be true. Replace the node with true. In debug builds,
3142 // we leave the test in the graph to have an additional sanity check at runtime. If the test fails (i.e. a bug),
3143 // we will execute a Halt node.
3144 #ifdef ASSERT
3145 _igvn.replace_node(n, n->in(1));
3146 #else
3147 _igvn.replace_node(n, _igvn.intcon(1));
3148 #endif
3149 success = true;
3150 } else if (n->is_OpaqueInitializedAssertionPredicate()) {
3151 // Initialized Assertion Predicates must always evaluate to true. Therefore, we get rid of them in product
3152 // builds as they are useless. In debug builds we keep them as additional verification code. Even though
3153 // loop opts are already over, we want to keep Initialized Assertion Predicates alive as long as possible to
3154 // enable folding of dead control paths within which cast nodes become top after due to impossible types -
3155 // even after loop opts are over. Therefore, we delay the removal of these opaque nodes until now.
3156 #ifdef ASSERT
3157 _igvn.replace_node(n, n->in(1));
3158 #else
3159 _igvn.replace_node(n, _igvn.intcon(1));
3160 #endif // ASSERT
3161 } else if (n->Opcode() == Op_OuterStripMinedLoop) {
3162 C->remove_macro_node(n);
3163 success = true;
3164 } else if (n->Opcode() == Op_MaxL) {
3165 // Since MaxL and MinL are not implemented in the backend, we expand them to
3166 // a CMoveL construct now. At least until here, the type could be computed
3167 // precisely. CMoveL is not so smart, but we can give it at least the best
3168 // type we know abouot n now.
3169 Node* repl = MaxNode::signed_max(n->in(1), n->in(2), _igvn.type(n), _igvn);
3170 _igvn.replace_node(n, repl);
3171 success = true;
3172 } else if (n->Opcode() == Op_MinL) {
3173 Node* repl = MaxNode::signed_min(n->in(1), n->in(2), _igvn.type(n), _igvn);
3174 _igvn.replace_node(n, repl);
3175 success = true;
3176 }
3177 assert(!success || (C->macro_count() == (old_macro_count - 1)), "elimination must have deleted one node from macro list");
3178 progress = progress || success;
3179 if (success) {
3180 C->print_method(PHASE_AFTER_MACRO_ELIMINATION_STEP, 5, n);
3181 }
3182 }
3183 }
3184 }
3185
3186 //------------------------------expand_macro_nodes----------------------
3187 // Returns true if a failure occurred.
3188 bool PhaseMacroExpand::expand_macro_nodes() {
3189 if (StressMacroExpansion) {
3190 C->shuffle_macro_nodes();
3191 }
3192
3193 // Clean up the graph so we're less likely to hit the maximum node
3194 // limit
3195 _igvn.set_delay_transform(false);
3196 _igvn.optimize();
3197 if (C->failing()) return true;
3198 _igvn.set_delay_transform(true);
3199
3200
3201 // Because we run IGVN after each expansion, some macro nodes may go
3202 // dead and be removed from the list as we iterate over it. Move
3203 // Allocate nodes (processed in a second pass) at the beginning of
3204 // the list and then iterate from the last element of the list until
3205 // an Allocate node is seen. This is robust to random deletion in
3206 // the list due to nodes going dead.
3207 C->sort_macro_nodes();
3208
3209 // expand arraycopy "macro" nodes first
3210 // For ReduceBulkZeroing, we must first process all arraycopy nodes
3211 // before the allocate nodes are expanded.
3212 while (C->macro_count() > 0) {
3213 int macro_count = C->macro_count();
3214 Node * n = C->macro_node(macro_count-1);
3215 assert(n->is_macro(), "only macro nodes expected here");
3216 if (_igvn.type(n) == Type::TOP || (n->in(0) != nullptr && n->in(0)->is_top())) {
3217 // node is unreachable, so don't try to expand it
3218 C->remove_macro_node(n);
3219 continue;
3220 }
3221 if (n->is_Allocate()) {
3222 break;
3223 }
3224 // Make sure expansion will not cause node limit to be exceeded.
3225 // Worst case is a macro node gets expanded into about 200 nodes.
3226 // Allow 50% more for optimization.
3227 if (C->check_node_count(300, "out of nodes before macro expansion")) {
3228 return true;
3229 }
3230
3231 DEBUG_ONLY(int old_macro_count = C->macro_count();)
3232 switch (n->class_id()) {
3233 case Node::Class_Lock:
3234 expand_lock_node(n->as_Lock());
3235 break;
3236 case Node::Class_Unlock:
3237 expand_unlock_node(n->as_Unlock());
3238 break;
3239 case Node::Class_ArrayCopy:
3240 expand_arraycopy_node(n->as_ArrayCopy());
3241 break;
3242 case Node::Class_SubTypeCheck:
3243 expand_subtypecheck_node(n->as_SubTypeCheck());
3244 break;
3245 case Node::Class_CallStaticJava:
3246 expand_mh_intrinsic_return(n->as_CallStaticJava());
3247 C->remove_macro_node(n);
3248 break;
3249 case Node::Class_FlatArrayCheck:
3250 expand_flatarraycheck_node(n->as_FlatArrayCheck());
3251 break;
3252 default:
3253 switch (n->Opcode()) {
3254 case Op_ModD:
3255 case Op_ModF: {
3256 CallNode* mod_macro = n->as_Call();
3257 CallNode* call = new CallLeafPureNode(mod_macro->tf(), mod_macro->entry_point(),
3258 mod_macro->_name, TypeRawPtr::BOTTOM);
3259 call->init_req(TypeFunc::Control, mod_macro->in(TypeFunc::Control));
3260 call->init_req(TypeFunc::I_O, C->top());
3261 call->init_req(TypeFunc::Memory, C->top());
3262 call->init_req(TypeFunc::ReturnAdr, C->top());
3263 call->init_req(TypeFunc::FramePtr, C->top());
3264 for (unsigned int i = 0; i < mod_macro->tf()->domain_cc()->cnt() - TypeFunc::Parms; i++) {
3265 call->init_req(TypeFunc::Parms + i, mod_macro->in(TypeFunc::Parms + i));
3266 }
3267 _igvn.replace_node(mod_macro, call);
3268 transform_later(call);
3269 break;
3270 }
3271 default:
3272 assert(false, "unknown node type in macro list");
3273 }
3274 }
3275 assert(C->macro_count() == (old_macro_count - 1), "expansion must have deleted one node from macro list");
3276 if (C->failing()) return true;
3277 C->print_method(PHASE_AFTER_MACRO_EXPANSION_STEP, 5, n);
3278
3279 // Clean up the graph so we're less likely to hit the maximum node
3280 // limit
3281 _igvn.set_delay_transform(false);
3282 _igvn.optimize();
3283 if (C->failing()) return true;
3284 _igvn.set_delay_transform(true);
3285 }
3286
3287 // All nodes except Allocate nodes are expanded now. There could be
3288 // new optimization opportunities (such as folding newly created
3289 // load from a just allocated object). Run IGVN.
3290
3291 // expand "macro" nodes
3292 // nodes are removed from the macro list as they are processed
3293 while (C->macro_count() > 0) {
3294 int macro_count = C->macro_count();
3295 Node * n = C->macro_node(macro_count-1);
3296 assert(n->is_macro(), "only macro nodes expected here");
3297 if (_igvn.type(n) == Type::TOP || (n->in(0) != nullptr && n->in(0)->is_top())) {
3298 // node is unreachable, so don't try to expand it
3299 C->remove_macro_node(n);
3300 continue;
3301 }
3302 // Make sure expansion will not cause node limit to be exceeded.
3303 // Worst case is a macro node gets expanded into about 200 nodes.
3304 // Allow 50% more for optimization.
3305 if (C->check_node_count(300, "out of nodes before macro expansion")) {
3306 return true;
3307 }
3308 switch (n->class_id()) {
3309 case Node::Class_Allocate:
3310 expand_allocate(n->as_Allocate());
3311 break;
3312 case Node::Class_AllocateArray:
3313 expand_allocate_array(n->as_AllocateArray());
3314 break;
3315 default:
3316 assert(false, "unknown node type in macro list");
3317 }
3318 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
3319 if (C->failing()) return true;
3320 C->print_method(PHASE_AFTER_MACRO_EXPANSION_STEP, 5, n);
3321
3322 // Clean up the graph so we're less likely to hit the maximum node
3323 // limit
3324 _igvn.set_delay_transform(false);
3325 _igvn.optimize();
3326 if (C->failing()) return true;
3327 _igvn.set_delay_transform(true);
3328 }
3329
3330 _igvn.set_delay_transform(false);
3331 return false;
3332 }
3333
3334 #ifndef PRODUCT
3335 int PhaseMacroExpand::_objs_scalar_replaced_counter = 0;
3336 int PhaseMacroExpand::_monitor_objects_removed_counter = 0;
3337 int PhaseMacroExpand::_GC_barriers_removed_counter = 0;
3338 int PhaseMacroExpand::_memory_barriers_removed_counter = 0;
3339
3340 void PhaseMacroExpand::print_statistics() {
3341 tty->print("Objects scalar replaced = %d, ", Atomic::load(&_objs_scalar_replaced_counter));
3342 tty->print("Monitor objects removed = %d, ", Atomic::load(&_monitor_objects_removed_counter));
3343 tty->print("GC barriers removed = %d, ", Atomic::load(&_GC_barriers_removed_counter));
3344 tty->print_cr("Memory barriers removed = %d", Atomic::load(&_memory_barriers_removed_counter));
3345 }
3346
3347 int PhaseMacroExpand::count_MemBar(Compile *C) {
3348 if (!PrintOptoStatistics) {
3349 return 0;
3350 }
3351 Unique_Node_List ideal_nodes;
3352 int total = 0;
3353 ideal_nodes.map(C->live_nodes(), nullptr);
3354 ideal_nodes.push(C->root());
3355 for (uint next = 0; next < ideal_nodes.size(); ++next) {
3356 Node* n = ideal_nodes.at(next);
3357 if (n->is_MemBar()) {
3358 total++;
3359 }
3360 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3361 Node* m = n->fast_out(i);
3362 ideal_nodes.push(m);
3363 }
3364 }
3365 return total;
3366 }
3367 #endif