1 /*
2 * Copyright (c) 2005, 2021, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "gc/shared/collectedHeap.inline.hpp"
28 #include "gc/shared/tlab_globals.hpp"
29 #include "libadt/vectset.hpp"
30 #include "memory/universe.hpp"
31 #include "opto/addnode.hpp"
32 #include "opto/arraycopynode.hpp"
33 #include "opto/callnode.hpp"
34 #include "opto/castnode.hpp"
35 #include "opto/cfgnode.hpp"
36 #include "opto/compile.hpp"
37 #include "opto/convertnode.hpp"
38 #include "opto/graphKit.hpp"
39 #include "opto/intrinsicnode.hpp"
40 #include "opto/locknode.hpp"
41 #include "opto/loopnode.hpp"
42 #include "opto/macro.hpp"
43 #include "opto/memnode.hpp"
44 #include "opto/narrowptrnode.hpp"
45 #include "opto/node.hpp"
46 #include "opto/opaquenode.hpp"
47 #include "opto/phaseX.hpp"
48 #include "opto/rootnode.hpp"
49 #include "opto/runtime.hpp"
50 #include "opto/subnode.hpp"
51 #include "opto/subtypenode.hpp"
52 #include "opto/type.hpp"
53 #include "prims/jvmtiExport.hpp"
54 #include "runtime/sharedRuntime.hpp"
55 #include "utilities/macros.hpp"
56 #include "utilities/powerOfTwo.hpp"
57 #if INCLUDE_G1GC
58 #include "gc/g1/g1ThreadLocalData.hpp"
59 #endif // INCLUDE_G1GC
60 #if INCLUDE_SHENANDOAHGC
61 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
62 #endif
63
64
65 //
66 // Replace any references to "oldref" in inputs to "use" with "newref".
67 // Returns the number of replacements made.
68 //
replace_input(Node * use,Node * oldref,Node * newref)69 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
70 int nreplacements = 0;
71 uint req = use->req();
72 for (uint j = 0; j < use->len(); j++) {
73 Node *uin = use->in(j);
74 if (uin == oldref) {
75 if (j < req)
76 use->set_req(j, newref);
77 else
78 use->set_prec(j, newref);
79 nreplacements++;
80 } else if (j >= req && uin == NULL) {
81 break;
82 }
83 }
84 return nreplacements;
85 }
86
migrate_outs(Node * old,Node * target)87 void PhaseMacroExpand::migrate_outs(Node *old, Node *target) {
88 assert(old != NULL, "sanity");
89 for (DUIterator_Fast imax, i = old->fast_outs(imax); i < imax; i++) {
90 Node* use = old->fast_out(i);
91 _igvn.rehash_node_delayed(use);
92 imax -= replace_input(use, old, target);
93 // back up iterator
94 --i;
95 }
96 assert(old->outcnt() == 0, "all uses must be deleted");
97 }
98
opt_bits_test(Node * ctrl,Node * region,int edge,Node * word,int mask,int bits,bool return_fast_path)99 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
100 Node* cmp;
101 if (mask != 0) {
102 Node* and_node = transform_later(new AndXNode(word, MakeConX(mask)));
103 cmp = transform_later(new CmpXNode(and_node, MakeConX(bits)));
104 } else {
105 cmp = word;
106 }
107 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
108 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
109 transform_later(iff);
110
111 // Fast path taken.
112 Node *fast_taken = transform_later(new IfFalseNode(iff));
113
114 // Fast path not-taken, i.e. slow path
115 Node *slow_taken = transform_later(new IfTrueNode(iff));
116
117 if (return_fast_path) {
118 region->init_req(edge, slow_taken); // Capture slow-control
119 return fast_taken;
120 } else {
121 region->init_req(edge, fast_taken); // Capture fast-control
122 return slow_taken;
123 }
124 }
125
126 //--------------------copy_predefined_input_for_runtime_call--------------------
copy_predefined_input_for_runtime_call(Node * ctrl,CallNode * oldcall,CallNode * call)127 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
128 // Set fixed predefined input arguments
129 call->init_req( TypeFunc::Control, ctrl );
130 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
131 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
132 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
133 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
134 }
135
136 //------------------------------make_slow_call---------------------------------
make_slow_call(CallNode * oldcall,const TypeFunc * slow_call_type,address slow_call,const char * leaf_name,Node * slow_path,Node * parm0,Node * parm1,Node * parm2)137 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type,
138 address slow_call, const char* leaf_name, Node* slow_path,
139 Node* parm0, Node* parm1, Node* parm2) {
140
141 // Slow-path call
142 CallNode *call = leaf_name
143 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
144 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), TypeRawPtr::BOTTOM );
145
146 // Slow path call has no side-effects, uses few values
147 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
148 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0);
149 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1);
150 if (parm2 != NULL) call->init_req(TypeFunc::Parms+2, parm2);
151 call->copy_call_debug_info(&_igvn, oldcall);
152 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
153 _igvn.replace_node(oldcall, call);
154 transform_later(call);
155
156 return call;
157 }
158
eliminate_gc_barrier(Node * p2x)159 void PhaseMacroExpand::eliminate_gc_barrier(Node* p2x) {
160 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
161 bs->eliminate_gc_barrier(this, p2x);
162 }
163
164 // Search for a memory operation for the specified memory slice.
scan_mem_chain(Node * mem,int alias_idx,int offset,Node * start_mem,Node * alloc,PhaseGVN * phase)165 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
166 Node *orig_mem = mem;
167 Node *alloc_mem = alloc->in(TypeFunc::Memory);
168 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
169 while (true) {
170 if (mem == alloc_mem || mem == start_mem ) {
171 return mem; // hit one of our sentinels
172 } else if (mem->is_MergeMem()) {
173 mem = mem->as_MergeMem()->memory_at(alias_idx);
174 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
175 Node *in = mem->in(0);
176 // we can safely skip over safepoints, calls, locks and membars because we
177 // already know that the object is safe to eliminate.
178 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
179 return in;
180 } else if (in->is_Call()) {
181 CallNode *call = in->as_Call();
182 if (call->may_modify(tinst, phase)) {
183 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape");
184 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) {
185 return in;
186 }
187 }
188 mem = in->in(TypeFunc::Memory);
189 } else if (in->is_MemBar()) {
190 ArrayCopyNode* ac = NULL;
191 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) {
192 if (ac != NULL) {
193 assert(ac->is_clonebasic(), "Only basic clone is a non escaping clone");
194 return ac;
195 }
196 }
197 mem = in->in(TypeFunc::Memory);
198 } else {
199 #ifdef ASSERT
200 in->dump();
201 mem->dump();
202 assert(false, "unexpected projection");
203 #endif
204 }
205 } else if (mem->is_Store()) {
206 const TypePtr* atype = mem->as_Store()->adr_type();
207 int adr_idx = phase->C->get_alias_index(atype);
208 if (adr_idx == alias_idx) {
209 assert(atype->isa_oopptr(), "address type must be oopptr");
210 int adr_offset = atype->offset();
211 uint adr_iid = atype->is_oopptr()->instance_id();
212 // Array elements references have the same alias_idx
213 // but different offset and different instance_id.
214 if (adr_offset == offset && adr_iid == alloc->_idx) {
215 return mem;
216 }
217 } else {
218 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
219 }
220 mem = mem->in(MemNode::Memory);
221 } else if (mem->is_ClearArray()) {
222 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
223 // Can not bypass initialization of the instance
224 // we are looking.
225 debug_only(intptr_t offset;)
226 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
227 InitializeNode* init = alloc->as_Allocate()->initialization();
228 // We are looking for stored value, return Initialize node
229 // or memory edge from Allocate node.
230 if (init != NULL) {
231 return init;
232 } else {
233 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
234 }
235 }
236 // Otherwise skip it (the call updated 'mem' value).
237 } else if (mem->Opcode() == Op_SCMemProj) {
238 mem = mem->in(0);
239 Node* adr = NULL;
240 if (mem->is_LoadStore()) {
241 adr = mem->in(MemNode::Address);
242 } else {
243 assert(mem->Opcode() == Op_EncodeISOArray ||
244 mem->Opcode() == Op_StrCompressedCopy, "sanity");
245 adr = mem->in(3); // Destination array
246 }
247 const TypePtr* atype = adr->bottom_type()->is_ptr();
248 int adr_idx = phase->C->get_alias_index(atype);
249 if (adr_idx == alias_idx) {
250 DEBUG_ONLY(mem->dump();)
251 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
252 return NULL;
253 }
254 mem = mem->in(MemNode::Memory);
255 } else if (mem->Opcode() == Op_StrInflatedCopy) {
256 Node* adr = mem->in(3); // Destination array
257 const TypePtr* atype = adr->bottom_type()->is_ptr();
258 int adr_idx = phase->C->get_alias_index(atype);
259 if (adr_idx == alias_idx) {
260 DEBUG_ONLY(mem->dump();)
261 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
262 return NULL;
263 }
264 mem = mem->in(MemNode::Memory);
265 } else {
266 return mem;
267 }
268 assert(mem != orig_mem, "dead memory loop");
269 }
270 }
271
272 // Generate loads from source of the arraycopy for fields of
273 // destination needed at a deoptimization point
make_arraycopy_load(ArrayCopyNode * ac,intptr_t offset,Node * ctl,Node * mem,BasicType ft,const Type * ftype,AllocateNode * alloc)274 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) {
275 BasicType bt = ft;
276 const Type *type = ftype;
277 if (ft == T_NARROWOOP) {
278 bt = T_OBJECT;
279 type = ftype->make_oopptr();
280 }
281 Node* res = NULL;
282 if (ac->is_clonebasic()) {
283 assert(ac->in(ArrayCopyNode::Src) != ac->in(ArrayCopyNode::Dest), "clone source equals destination");
284 Node* base = ac->in(ArrayCopyNode::Src);
285 Node* adr = _igvn.transform(new AddPNode(base, base, MakeConX(offset)));
286 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
287 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
288 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
289 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
290 } else {
291 if (ac->modifies(offset, offset, &_igvn, true)) {
292 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
293 uint shift = exact_log2(type2aelembytes(bt));
294 Node* src_pos = ac->in(ArrayCopyNode::SrcPos);
295 Node* dest_pos = ac->in(ArrayCopyNode::DestPos);
296 const TypeInt* src_pos_t = _igvn.type(src_pos)->is_int();
297 const TypeInt* dest_pos_t = _igvn.type(dest_pos)->is_int();
298
299 Node* adr = NULL;
300 const TypePtr* adr_type = NULL;
301 if (src_pos_t->is_con() && dest_pos_t->is_con()) {
302 intptr_t off = ((src_pos_t->get_con() - dest_pos_t->get_con()) << shift) + offset;
303 Node* base = ac->in(ArrayCopyNode::Src);
304 adr = _igvn.transform(new AddPNode(base, base, MakeConX(off)));
305 adr_type = _igvn.type(base)->is_ptr()->add_offset(off);
306 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
307 // Don't emit a new load from src if src == dst but try to get the value from memory instead
308 return value_from_mem(ac->in(TypeFunc::Memory), ctl, ft, ftype, adr_type->isa_oopptr(), alloc);
309 }
310 } else {
311 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
312 #ifdef _LP64
313 diff = _igvn.transform(new ConvI2LNode(diff));
314 #endif
315 diff = _igvn.transform(new LShiftXNode(diff, intcon(shift)));
316
317 Node* off = _igvn.transform(new AddXNode(MakeConX(offset), diff));
318 Node* base = ac->in(ArrayCopyNode::Src);
319 adr = _igvn.transform(new AddPNode(base, base, off));
320 adr_type = _igvn.type(base)->is_ptr()->add_offset(Type::OffsetBot);
321 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
322 // Non constant offset in the array: we can't statically
323 // determine the value
324 return NULL;
325 }
326 }
327 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
328 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
329 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
330 }
331 }
332 if (res != NULL) {
333 if (ftype->isa_narrowoop()) {
334 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
335 res = _igvn.transform(new EncodePNode(res, ftype));
336 }
337 return res;
338 }
339 return NULL;
340 }
341
342 //
343 // Given a Memory Phi, compute a value Phi containing the values from stores
344 // on the input paths.
345 // Note: this function is recursive, its depth is limited by the "level" argument
346 // Returns the computed Phi, or NULL if it cannot compute it.
value_from_mem_phi(Node * mem,BasicType ft,const Type * phi_type,const TypeOopPtr * adr_t,AllocateNode * alloc,Node_Stack * value_phis,int level)347 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) {
348 assert(mem->is_Phi(), "sanity");
349 int alias_idx = C->get_alias_index(adr_t);
350 int offset = adr_t->offset();
351 int instance_id = adr_t->instance_id();
352
353 // Check if an appropriate value phi already exists.
354 Node* region = mem->in(0);
355 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
356 Node* phi = region->fast_out(k);
357 if (phi->is_Phi() && phi != mem &&
358 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
359 return phi;
360 }
361 }
362 // Check if an appropriate new value phi already exists.
363 Node* new_phi = value_phis->find(mem->_idx);
364 if (new_phi != NULL)
365 return new_phi;
366
367 if (level <= 0) {
368 return NULL; // Give up: phi tree too deep
369 }
370 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
371 Node *alloc_mem = alloc->in(TypeFunc::Memory);
372
373 uint length = mem->req();
374 GrowableArray <Node *> values(length, length, NULL);
375
376 // create a new Phi for the value
377 PhiNode *phi = new PhiNode(mem->in(0), phi_type, NULL, mem->_idx, instance_id, alias_idx, offset);
378 transform_later(phi);
379 value_phis->push(phi, mem->_idx);
380
381 for (uint j = 1; j < length; j++) {
382 Node *in = mem->in(j);
383 if (in == NULL || in->is_top()) {
384 values.at_put(j, in);
385 } else {
386 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
387 if (val == start_mem || val == alloc_mem) {
388 // hit a sentinel, return appropriate 0 value
389 values.at_put(j, _igvn.zerocon(ft));
390 continue;
391 }
392 if (val->is_Initialize()) {
393 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
394 }
395 if (val == NULL) {
396 return NULL; // can't find a value on this path
397 }
398 if (val == mem) {
399 values.at_put(j, mem);
400 } else if (val->is_Store()) {
401 Node* n = val->in(MemNode::ValueIn);
402 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
403 n = bs->step_over_gc_barrier(n);
404 if (is_subword_type(ft)) {
405 n = Compile::narrow_value(ft, n, phi_type, &_igvn, true);
406 }
407 values.at_put(j, n);
408 } else if(val->is_Proj() && val->in(0) == alloc) {
409 values.at_put(j, _igvn.zerocon(ft));
410 } else if (val->is_Phi()) {
411 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
412 if (val == NULL) {
413 return NULL;
414 }
415 values.at_put(j, val);
416 } else if (val->Opcode() == Op_SCMemProj) {
417 assert(val->in(0)->is_LoadStore() ||
418 val->in(0)->Opcode() == Op_EncodeISOArray ||
419 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
420 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
421 return NULL;
422 } else if (val->is_ArrayCopy()) {
423 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc);
424 if (res == NULL) {
425 return NULL;
426 }
427 values.at_put(j, res);
428 } else {
429 DEBUG_ONLY( val->dump(); )
430 assert(false, "unknown node on this path");
431 return NULL; // unknown node on this path
432 }
433 }
434 }
435 // Set Phi's inputs
436 for (uint j = 1; j < length; j++) {
437 if (values.at(j) == mem) {
438 phi->init_req(j, phi);
439 } else {
440 phi->init_req(j, values.at(j));
441 }
442 }
443 return phi;
444 }
445
446 // Search the last value stored into the object's field.
value_from_mem(Node * sfpt_mem,Node * sfpt_ctl,BasicType ft,const Type * ftype,const TypeOopPtr * adr_t,AllocateNode * alloc)447 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
448 assert(adr_t->is_known_instance_field(), "instance required");
449 int instance_id = adr_t->instance_id();
450 assert((uint)instance_id == alloc->_idx, "wrong allocation");
451
452 int alias_idx = C->get_alias_index(adr_t);
453 int offset = adr_t->offset();
454 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
455 Node *alloc_ctrl = alloc->in(TypeFunc::Control);
456 Node *alloc_mem = alloc->in(TypeFunc::Memory);
457 VectorSet visited;
458
459 bool done = sfpt_mem == alloc_mem;
460 Node *mem = sfpt_mem;
461 while (!done) {
462 if (visited.test_set(mem->_idx)) {
463 return NULL; // found a loop, give up
464 }
465 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
466 if (mem == start_mem || mem == alloc_mem) {
467 done = true; // hit a sentinel, return appropriate 0 value
468 } else if (mem->is_Initialize()) {
469 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
470 if (mem == NULL) {
471 done = true; // Something go wrong.
472 } else if (mem->is_Store()) {
473 const TypePtr* atype = mem->as_Store()->adr_type();
474 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
475 done = true;
476 }
477 } else if (mem->is_Store()) {
478 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
479 assert(atype != NULL, "address type must be oopptr");
480 assert(C->get_alias_index(atype) == alias_idx &&
481 atype->is_known_instance_field() && atype->offset() == offset &&
482 atype->instance_id() == instance_id, "store is correct memory slice");
483 done = true;
484 } else if (mem->is_Phi()) {
485 // try to find a phi's unique input
486 Node *unique_input = NULL;
487 Node *top = C->top();
488 for (uint i = 1; i < mem->req(); i++) {
489 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
490 if (n == NULL || n == top || n == mem) {
491 continue;
492 } else if (unique_input == NULL) {
493 unique_input = n;
494 } else if (unique_input != n) {
495 unique_input = top;
496 break;
497 }
498 }
499 if (unique_input != NULL && unique_input != top) {
500 mem = unique_input;
501 } else {
502 done = true;
503 }
504 } else if (mem->is_ArrayCopy()) {
505 done = true;
506 } else {
507 DEBUG_ONLY( mem->dump(); )
508 assert(false, "unexpected node");
509 }
510 }
511 if (mem != NULL) {
512 if (mem == start_mem || mem == alloc_mem) {
513 // hit a sentinel, return appropriate 0 value
514 return _igvn.zerocon(ft);
515 } else if (mem->is_Store()) {
516 Node* n = mem->in(MemNode::ValueIn);
517 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
518 n = bs->step_over_gc_barrier(n);
519 return n;
520 } else if (mem->is_Phi()) {
521 // attempt to produce a Phi reflecting the values on the input paths of the Phi
522 Node_Stack value_phis(8);
523 Node* phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
524 if (phi != NULL) {
525 return phi;
526 } else {
527 // Kill all new Phis
528 while(value_phis.is_nonempty()) {
529 Node* n = value_phis.node();
530 _igvn.replace_node(n, C->top());
531 value_phis.pop();
532 }
533 }
534 } else if (mem->is_ArrayCopy()) {
535 Node* ctl = mem->in(0);
536 Node* m = mem->in(TypeFunc::Memory);
537 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
538 // pin the loads in the uncommon trap path
539 ctl = sfpt_ctl;
540 m = sfpt_mem;
541 }
542 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc);
543 }
544 }
545 // Something go wrong.
546 return NULL;
547 }
548
549 // Check the possibility of scalar replacement.
can_eliminate_allocation(AllocateNode * alloc,GrowableArray<SafePointNode * > & safepoints)550 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
551 // Scan the uses of the allocation to check for anything that would
552 // prevent us from eliminating it.
553 NOT_PRODUCT( const char* fail_eliminate = NULL; )
554 DEBUG_ONLY( Node* disq_node = NULL; )
555 bool can_eliminate = true;
556
557 Node* res = alloc->result_cast();
558 const TypeOopPtr* res_type = NULL;
559 if (res == NULL) {
560 // All users were eliminated.
561 } else if (!res->is_CheckCastPP()) {
562 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
563 can_eliminate = false;
564 } else {
565 res_type = _igvn.type(res)->isa_oopptr();
566 if (res_type == NULL) {
567 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
568 can_eliminate = false;
569 } else if (res_type->isa_aryptr()) {
570 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
571 if (length < 0) {
572 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
573 can_eliminate = false;
574 }
575 }
576 }
577
578 if (can_eliminate && res != NULL) {
579 for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
580 j < jmax && can_eliminate; j++) {
581 Node* use = res->fast_out(j);
582
583 if (use->is_AddP()) {
584 const TypePtr* addp_type = _igvn.type(use)->is_ptr();
585 int offset = addp_type->offset();
586
587 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
588 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";)
589 can_eliminate = false;
590 break;
591 }
592 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
593 k < kmax && can_eliminate; k++) {
594 Node* n = use->fast_out(k);
595 if (!n->is_Store() && n->Opcode() != Op_CastP2X
596 SHENANDOAHGC_ONLY(&& (!UseShenandoahGC || !ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(n))) ) {
597 DEBUG_ONLY(disq_node = n;)
598 if (n->is_Load() || n->is_LoadStore()) {
599 NOT_PRODUCT(fail_eliminate = "Field load";)
600 } else {
601 NOT_PRODUCT(fail_eliminate = "Not store field referrence";)
602 }
603 can_eliminate = false;
604 }
605 }
606 } else if (use->is_ArrayCopy() &&
607 (use->as_ArrayCopy()->is_clonebasic() ||
608 use->as_ArrayCopy()->is_arraycopy_validated() ||
609 use->as_ArrayCopy()->is_copyof_validated() ||
610 use->as_ArrayCopy()->is_copyofrange_validated()) &&
611 use->in(ArrayCopyNode::Dest) == res) {
612 // ok to eliminate
613 } else if (use->is_SafePoint()) {
614 SafePointNode* sfpt = use->as_SafePoint();
615 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
616 // Object is passed as argument.
617 DEBUG_ONLY(disq_node = use;)
618 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
619 can_eliminate = false;
620 }
621 Node* sfptMem = sfpt->memory();
622 if (sfptMem == NULL || sfptMem->is_top()) {
623 DEBUG_ONLY(disq_node = use;)
624 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
625 can_eliminate = false;
626 } else {
627 safepoints.append_if_missing(sfpt);
628 }
629 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
630 if (use->is_Phi()) {
631 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
632 NOT_PRODUCT(fail_eliminate = "Object is return value";)
633 } else {
634 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
635 }
636 DEBUG_ONLY(disq_node = use;)
637 } else {
638 if (use->Opcode() == Op_Return) {
639 NOT_PRODUCT(fail_eliminate = "Object is return value";)
640 }else {
641 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
642 }
643 DEBUG_ONLY(disq_node = use;)
644 }
645 can_eliminate = false;
646 }
647 }
648 }
649
650 #ifndef PRODUCT
651 if (PrintEliminateAllocations) {
652 if (can_eliminate) {
653 tty->print("Scalar ");
654 if (res == NULL)
655 alloc->dump();
656 else
657 res->dump();
658 } else if (alloc->_is_scalar_replaceable) {
659 tty->print("NotScalar (%s)", fail_eliminate);
660 if (res == NULL)
661 alloc->dump();
662 else
663 res->dump();
664 #ifdef ASSERT
665 if (disq_node != NULL) {
666 tty->print(" >>>> ");
667 disq_node->dump();
668 }
669 #endif /*ASSERT*/
670 }
671 }
672 #endif
673 return can_eliminate;
674 }
675
676 // Do scalar replacement.
scalar_replacement(AllocateNode * alloc,GrowableArray<SafePointNode * > & safepoints)677 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
678 GrowableArray <SafePointNode *> safepoints_done;
679
680 ciKlass* klass = NULL;
681 ciInstanceKlass* iklass = NULL;
682 int nfields = 0;
683 int array_base = 0;
684 int element_size = 0;
685 BasicType basic_elem_type = T_ILLEGAL;
686 ciType* elem_type = NULL;
687
688 Node* res = alloc->result_cast();
689 assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result");
690 const TypeOopPtr* res_type = NULL;
691 if (res != NULL) { // Could be NULL when there are no users
692 res_type = _igvn.type(res)->isa_oopptr();
693 }
694
695 if (res != NULL) {
696 klass = res_type->klass();
697 if (res_type->isa_instptr()) {
698 // find the fields of the class which will be needed for safepoint debug information
699 assert(klass->is_instance_klass(), "must be an instance klass.");
700 iklass = klass->as_instance_klass();
701 nfields = iklass->nof_nonstatic_fields();
702 } else {
703 // find the array's elements which will be needed for safepoint debug information
704 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
705 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass.");
706 elem_type = klass->as_array_klass()->element_type();
707 basic_elem_type = elem_type->basic_type();
708 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
709 element_size = type2aelembytes(basic_elem_type);
710 }
711 }
712 //
713 // Process the safepoint uses
714 //
715 while (safepoints.length() > 0) {
716 SafePointNode* sfpt = safepoints.pop();
717 Node* mem = sfpt->memory();
718 Node* ctl = sfpt->control();
719 assert(sfpt->jvms() != NULL, "missed JVMS");
720 // Fields of scalar objs are referenced only at the end
721 // of regular debuginfo at the last (youngest) JVMS.
722 // Record relative start index.
723 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
724 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type,
725 #ifdef ASSERT
726 alloc,
727 #endif
728 first_ind, nfields);
729 sobj->init_req(0, C->root());
730 transform_later(sobj);
731
732 // Scan object's fields adding an input to the safepoint for each field.
733 for (int j = 0; j < nfields; j++) {
734 intptr_t offset;
735 ciField* field = NULL;
736 if (iklass != NULL) {
737 field = iklass->nonstatic_field_at(j);
738 offset = field->offset();
739 elem_type = field->type();
740 basic_elem_type = field->layout_type();
741 } else {
742 offset = array_base + j * (intptr_t)element_size;
743 }
744
745 const Type *field_type;
746 // The next code is taken from Parse::do_get_xxx().
747 if (is_reference_type(basic_elem_type)) {
748 if (!elem_type->is_loaded()) {
749 field_type = TypeInstPtr::BOTTOM;
750 } else if (field != NULL && field->is_static_constant()) {
751 // This can happen if the constant oop is non-perm.
752 ciObject* con = field->constant_value().as_object();
753 // Do not "join" in the previous type; it doesn't add value,
754 // and may yield a vacuous result if the field is of interface type.
755 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
756 assert(field_type != NULL, "field singleton type must be consistent");
757 } else {
758 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
759 }
760 if (UseCompressedOops) {
761 field_type = field_type->make_narrowoop();
762 basic_elem_type = T_NARROWOOP;
763 }
764 } else {
765 field_type = Type::get_const_basic_type(basic_elem_type);
766 }
767
768 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr();
769
770 Node *field_val = value_from_mem(mem, ctl, basic_elem_type, field_type, field_addr_type, alloc);
771 if (field_val == NULL) {
772 // We weren't able to find a value for this field,
773 // give up on eliminating this allocation.
774
775 // Remove any extra entries we added to the safepoint.
776 uint last = sfpt->req() - 1;
777 for (int k = 0; k < j; k++) {
778 sfpt->del_req(last--);
779 }
780 _igvn._worklist.push(sfpt);
781 // rollback processed safepoints
782 while (safepoints_done.length() > 0) {
783 SafePointNode* sfpt_done = safepoints_done.pop();
784 // remove any extra entries we added to the safepoint
785 last = sfpt_done->req() - 1;
786 for (int k = 0; k < nfields; k++) {
787 sfpt_done->del_req(last--);
788 }
789 JVMState *jvms = sfpt_done->jvms();
790 jvms->set_endoff(sfpt_done->req());
791 // Now make a pass over the debug information replacing any references
792 // to SafePointScalarObjectNode with the allocated object.
793 int start = jvms->debug_start();
794 int end = jvms->debug_end();
795 for (int i = start; i < end; i++) {
796 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
797 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
798 if (scobj->first_index(jvms) == sfpt_done->req() &&
799 scobj->n_fields() == (uint)nfields) {
800 assert(scobj->alloc() == alloc, "sanity");
801 sfpt_done->set_req(i, res);
802 }
803 }
804 }
805 _igvn._worklist.push(sfpt_done);
806 }
807 #ifndef PRODUCT
808 if (PrintEliminateAllocations) {
809 if (field != NULL) {
810 tty->print("=== At SafePoint node %d can't find value of Field: ",
811 sfpt->_idx);
812 field->print();
813 int field_idx = C->get_alias_index(field_addr_type);
814 tty->print(" (alias_idx=%d)", field_idx);
815 } else { // Array's element
816 tty->print("=== At SafePoint node %d can't find value of array element [%d]",
817 sfpt->_idx, j);
818 }
819 tty->print(", which prevents elimination of: ");
820 if (res == NULL)
821 alloc->dump();
822 else
823 res->dump();
824 }
825 #endif
826 return false;
827 }
828 if (UseCompressedOops && field_type->isa_narrowoop()) {
829 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
830 // to be able scalar replace the allocation.
831 if (field_val->is_EncodeP()) {
832 field_val = field_val->in(1);
833 } else {
834 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
835 }
836 }
837 sfpt->add_req(field_val);
838 }
839 JVMState *jvms = sfpt->jvms();
840 jvms->set_endoff(sfpt->req());
841 // Now make a pass over the debug information replacing any references
842 // to the allocated object with "sobj"
843 int start = jvms->debug_start();
844 int end = jvms->debug_end();
845 sfpt->replace_edges_in_range(res, sobj, start, end, &_igvn);
846 _igvn._worklist.push(sfpt);
847 safepoints_done.append_if_missing(sfpt); // keep it for rollback
848 }
849 return true;
850 }
851
disconnect_projections(MultiNode * n,PhaseIterGVN & igvn)852 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
853 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
854 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
855 if (ctl_proj != NULL) {
856 igvn.replace_node(ctl_proj, n->in(0));
857 }
858 if (mem_proj != NULL) {
859 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
860 }
861 }
862
863 // Process users of eliminated allocation.
process_users_of_allocation(CallNode * alloc)864 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) {
865 Node* res = alloc->result_cast();
866 if (res != NULL) {
867 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
868 Node *use = res->last_out(j);
869 uint oc1 = res->outcnt();
870
871 if (use->is_AddP()) {
872 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
873 Node *n = use->last_out(k);
874 uint oc2 = use->outcnt();
875 if (n->is_Store()) {
876 #ifdef ASSERT
877 // Verify that there is no dependent MemBarVolatile nodes,
878 // they should be removed during IGVN, see MemBarNode::Ideal().
879 for (DUIterator_Fast pmax, p = n->fast_outs(pmax);
880 p < pmax; p++) {
881 Node* mb = n->fast_out(p);
882 assert(mb->is_Initialize() || !mb->is_MemBar() ||
883 mb->req() <= MemBarNode::Precedent ||
884 mb->in(MemBarNode::Precedent) != n,
885 "MemBarVolatile should be eliminated for non-escaping object");
886 }
887 #endif
888 _igvn.replace_node(n, n->in(MemNode::Memory));
889 } else {
890 eliminate_gc_barrier(n);
891 }
892 k -= (oc2 - use->outcnt());
893 }
894 _igvn.remove_dead_node(use);
895 } else if (use->is_ArrayCopy()) {
896 // Disconnect ArrayCopy node
897 ArrayCopyNode* ac = use->as_ArrayCopy();
898 if (ac->is_clonebasic()) {
899 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
900 disconnect_projections(ac, _igvn);
901 assert(alloc->in(TypeFunc::Memory)->is_Proj() && alloc->in(TypeFunc::Memory)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
902 Node* membar_before = alloc->in(TypeFunc::Memory)->in(0);
903 disconnect_projections(membar_before->as_MemBar(), _igvn);
904 if (membar_after->is_MemBar()) {
905 disconnect_projections(membar_after->as_MemBar(), _igvn);
906 }
907 } else {
908 assert(ac->is_arraycopy_validated() ||
909 ac->is_copyof_validated() ||
910 ac->is_copyofrange_validated(), "unsupported");
911 CallProjections callprojs;
912 ac->extract_projections(&callprojs, true);
913
914 _igvn.replace_node(callprojs.fallthrough_ioproj, ac->in(TypeFunc::I_O));
915 _igvn.replace_node(callprojs.fallthrough_memproj, ac->in(TypeFunc::Memory));
916 _igvn.replace_node(callprojs.fallthrough_catchproj, ac->in(TypeFunc::Control));
917
918 // Set control to top. IGVN will remove the remaining projections
919 ac->set_req(0, top());
920 ac->replace_edge(res, top(), &_igvn);
921
922 // Disconnect src right away: it can help find new
923 // opportunities for allocation elimination
924 Node* src = ac->in(ArrayCopyNode::Src);
925 ac->replace_edge(src, top(), &_igvn);
926 // src can be top at this point if src and dest of the
927 // arraycopy were the same
928 if (src->outcnt() == 0 && !src->is_top()) {
929 _igvn.remove_dead_node(src);
930 }
931 }
932 _igvn._worklist.push(ac);
933 } else {
934 eliminate_gc_barrier(use);
935 }
936 j -= (oc1 - res->outcnt());
937 }
938 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
939 _igvn.remove_dead_node(res);
940 }
941
942 //
943 // Process other users of allocation's projections
944 //
945 if (_callprojs.resproj != NULL && _callprojs.resproj->outcnt() != 0) {
946 // First disconnect stores captured by Initialize node.
947 // If Initialize node is eliminated first in the following code,
948 // it will kill such stores and DUIterator_Last will assert.
949 for (DUIterator_Fast jmax, j = _callprojs.resproj->fast_outs(jmax); j < jmax; j++) {
950 Node* use = _callprojs.resproj->fast_out(j);
951 if (use->is_AddP()) {
952 // raw memory addresses used only by the initialization
953 _igvn.replace_node(use, C->top());
954 --j; --jmax;
955 }
956 }
957 for (DUIterator_Last jmin, j = _callprojs.resproj->last_outs(jmin); j >= jmin; ) {
958 Node* use = _callprojs.resproj->last_out(j);
959 uint oc1 = _callprojs.resproj->outcnt();
960 if (use->is_Initialize()) {
961 // Eliminate Initialize node.
962 InitializeNode *init = use->as_Initialize();
963 assert(init->outcnt() <= 2, "only a control and memory projection expected");
964 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
965 if (ctrl_proj != NULL) {
966 _igvn.replace_node(ctrl_proj, init->in(TypeFunc::Control));
967 #ifdef ASSERT
968 // If the InitializeNode has no memory out, it will die, and tmp will become NULL
969 Node* tmp = init->in(TypeFunc::Control);
970 assert(tmp == NULL || tmp == _callprojs.fallthrough_catchproj, "allocation control projection");
971 #endif
972 }
973 Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory);
974 if (mem_proj != NULL) {
975 Node *mem = init->in(TypeFunc::Memory);
976 #ifdef ASSERT
977 if (mem->is_MergeMem()) {
978 assert(mem->in(TypeFunc::Memory) == _callprojs.fallthrough_memproj, "allocation memory projection");
979 } else {
980 assert(mem == _callprojs.fallthrough_memproj, "allocation memory projection");
981 }
982 #endif
983 _igvn.replace_node(mem_proj, mem);
984 }
985 } else {
986 assert(false, "only Initialize or AddP expected");
987 }
988 j -= (oc1 - _callprojs.resproj->outcnt());
989 }
990 }
991 if (_callprojs.fallthrough_catchproj != NULL) {
992 _igvn.replace_node(_callprojs.fallthrough_catchproj, alloc->in(TypeFunc::Control));
993 }
994 if (_callprojs.fallthrough_memproj != NULL) {
995 _igvn.replace_node(_callprojs.fallthrough_memproj, alloc->in(TypeFunc::Memory));
996 }
997 if (_callprojs.catchall_memproj != NULL) {
998 _igvn.replace_node(_callprojs.catchall_memproj, C->top());
999 }
1000 if (_callprojs.fallthrough_ioproj != NULL) {
1001 _igvn.replace_node(_callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
1002 }
1003 if (_callprojs.catchall_ioproj != NULL) {
1004 _igvn.replace_node(_callprojs.catchall_ioproj, C->top());
1005 }
1006 if (_callprojs.catchall_catchproj != NULL) {
1007 _igvn.replace_node(_callprojs.catchall_catchproj, C->top());
1008 }
1009 }
1010
eliminate_allocate_node(AllocateNode * alloc)1011 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1012 // If reallocation fails during deoptimization we'll pop all
1013 // interpreter frames for this compiled frame and that won't play
1014 // nice with JVMTI popframe.
1015 // We avoid this issue by eager reallocation when the popframe request
1016 // is received.
1017 if (!EliminateAllocations || !alloc->_is_non_escaping) {
1018 return false;
1019 }
1020 Node* klass = alloc->in(AllocateNode::KlassNode);
1021 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1022 Node* res = alloc->result_cast();
1023 // Eliminate boxing allocations which are not used
1024 // regardless scalar replacable status.
1025 bool boxing_alloc = C->eliminate_boxing() &&
1026 tklass->klass()->is_instance_klass() &&
1027 tklass->klass()->as_instance_klass()->is_box_klass();
1028 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) {
1029 return false;
1030 }
1031
1032 alloc->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1033
1034 GrowableArray <SafePointNode *> safepoints;
1035 if (!can_eliminate_allocation(alloc, safepoints)) {
1036 return false;
1037 }
1038
1039 if (!alloc->_is_scalar_replaceable) {
1040 assert(res == NULL, "sanity");
1041 // We can only eliminate allocation if all debug info references
1042 // are already replaced with SafePointScalarObject because
1043 // we can't search for a fields value without instance_id.
1044 if (safepoints.length() > 0) {
1045 return false;
1046 }
1047 }
1048
1049 if (!scalar_replacement(alloc, safepoints)) {
1050 return false;
1051 }
1052
1053 CompileLog* log = C->log();
1054 if (log != NULL) {
1055 log->head("eliminate_allocation type='%d'",
1056 log->identify(tklass->klass()));
1057 JVMState* p = alloc->jvms();
1058 while (p != NULL) {
1059 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1060 p = p->caller();
1061 }
1062 log->tail("eliminate_allocation");
1063 }
1064
1065 process_users_of_allocation(alloc);
1066
1067 #ifndef PRODUCT
1068 if (PrintEliminateAllocations) {
1069 if (alloc->is_AllocateArray())
1070 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1071 else
1072 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1073 }
1074 #endif
1075
1076 return true;
1077 }
1078
eliminate_boxing_node(CallStaticJavaNode * boxing)1079 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1080 // EA should remove all uses of non-escaping boxing node.
1081 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != NULL) {
1082 return false;
1083 }
1084
1085 assert(boxing->result_cast() == NULL, "unexpected boxing node result");
1086
1087 boxing->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1088
1089 const TypeTuple* r = boxing->tf()->range();
1090 assert(r->cnt() > TypeFunc::Parms, "sanity");
1091 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1092 assert(t != NULL, "sanity");
1093
1094 CompileLog* log = C->log();
1095 if (log != NULL) {
1096 log->head("eliminate_boxing type='%d'",
1097 log->identify(t->klass()));
1098 JVMState* p = boxing->jvms();
1099 while (p != NULL) {
1100 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1101 p = p->caller();
1102 }
1103 log->tail("eliminate_boxing");
1104 }
1105
1106 process_users_of_allocation(boxing);
1107
1108 #ifndef PRODUCT
1109 if (PrintEliminateAllocations) {
1110 tty->print("++++ Eliminated: %d ", boxing->_idx);
1111 boxing->method()->print_short_name(tty);
1112 tty->cr();
1113 }
1114 #endif
1115
1116 return true;
1117 }
1118
1119 //---------------------------set_eden_pointers-------------------------
set_eden_pointers(Node * & eden_top_adr,Node * & eden_end_adr)1120 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) {
1121 if (UseTLAB) { // Private allocation: load from TLS
1122 Node* thread = transform_later(new ThreadLocalNode());
1123 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset());
1124 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset());
1125 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset);
1126 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset);
1127 } else { // Shared allocation: load from globals
1128 CollectedHeap* ch = Universe::heap();
1129 address top_adr = (address)ch->top_addr();
1130 address end_adr = (address)ch->end_addr();
1131 eden_top_adr = makecon(TypeRawPtr::make(top_adr));
1132 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr);
1133 }
1134 }
1135
1136
make_load(Node * ctl,Node * mem,Node * base,int offset,const Type * value_type,BasicType bt)1137 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1138 Node* adr = basic_plus_adr(base, offset);
1139 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1140 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered);
1141 transform_later(value);
1142 return value;
1143 }
1144
1145
make_store(Node * ctl,Node * mem,Node * base,int offset,Node * value,BasicType bt)1146 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1147 Node* adr = basic_plus_adr(base, offset);
1148 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered);
1149 transform_later(mem);
1150 return mem;
1151 }
1152
1153 //=============================================================================
1154 //
1155 // A L L O C A T I O N
1156 //
1157 // Allocation attempts to be fast in the case of frequent small objects.
1158 // It breaks down like this:
1159 //
1160 // 1) Size in doublewords is computed. This is a constant for objects and
1161 // variable for most arrays. Doubleword units are used to avoid size
1162 // overflow of huge doubleword arrays. We need doublewords in the end for
1163 // rounding.
1164 //
1165 // 2) Size is checked for being 'too large'. Too-large allocations will go
1166 // the slow path into the VM. The slow path can throw any required
1167 // exceptions, and does all the special checks for very large arrays. The
1168 // size test can constant-fold away for objects. For objects with
1169 // finalizers it constant-folds the otherway: you always go slow with
1170 // finalizers.
1171 //
1172 // 3) If NOT using TLABs, this is the contended loop-back point.
1173 // Load-Locked the heap top. If using TLABs normal-load the heap top.
1174 //
1175 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
1176 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
1177 // "size*8" we always enter the VM, where "largish" is a constant picked small
1178 // enough that there's always space between the eden max and 4Gig (old space is
1179 // there so it's quite large) and large enough that the cost of entering the VM
1180 // is dwarfed by the cost to initialize the space.
1181 //
1182 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1183 // down. If contended, repeat at step 3. If using TLABs normal-store
1184 // adjusted heap top back down; there is no contention.
1185 //
1186 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
1187 // fields.
1188 //
1189 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1190 // oop flavor.
1191 //
1192 //=============================================================================
1193 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1194 // Allocations bigger than this always go the slow route.
1195 // This value must be small enough that allocation attempts that need to
1196 // trigger exceptions go the slow route. Also, it must be small enough so
1197 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1198 //=============================================================================j//
1199 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1200 // The allocator will coalesce int->oop copies away. See comment in
1201 // coalesce.cpp about how this works. It depends critically on the exact
1202 // code shape produced here, so if you are changing this code shape
1203 // make sure the GC info for the heap-top is correct in and around the
1204 // slow-path call.
1205 //
1206
expand_allocate_common(AllocateNode * alloc,Node * length,const TypeFunc * slow_call_type,address slow_call_address)1207 void PhaseMacroExpand::expand_allocate_common(
1208 AllocateNode* alloc, // allocation node to be expanded
1209 Node* length, // array length for an array allocation
1210 const TypeFunc* slow_call_type, // Type of slow call
1211 address slow_call_address // Address of slow call
1212 )
1213 {
1214 Node* ctrl = alloc->in(TypeFunc::Control);
1215 Node* mem = alloc->in(TypeFunc::Memory);
1216 Node* i_o = alloc->in(TypeFunc::I_O);
1217 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1218 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1219 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1220 assert(ctrl != NULL, "must have control");
1221
1222 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1223 // they will not be used if "always_slow" is set
1224 enum { slow_result_path = 1, fast_result_path = 2 };
1225 Node *result_region = NULL;
1226 Node *result_phi_rawmem = NULL;
1227 Node *result_phi_rawoop = NULL;
1228 Node *result_phi_i_o = NULL;
1229
1230 // The initial slow comparison is a size check, the comparison
1231 // we want to do is a BoolTest::gt
1232 bool expand_fast_path = true;
1233 int tv = _igvn.find_int_con(initial_slow_test, -1);
1234 if (tv >= 0) {
1235 // InitialTest has constant result
1236 // 0 - can fit in TLAB
1237 // 1 - always too big or negative
1238 assert(tv <= 1, "0 or 1 if a constant");
1239 expand_fast_path = (tv == 0);
1240 initial_slow_test = NULL;
1241 } else {
1242 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1243 }
1244
1245 if (C->env()->dtrace_alloc_probes() ||
1246 (!UseTLAB && !Universe::heap()->supports_inline_contig_alloc())) {
1247 // Force slow-path allocation
1248 expand_fast_path = false;
1249 initial_slow_test = NULL;
1250 }
1251
1252 bool allocation_has_use = (alloc->result_cast() != NULL);
1253 if (!allocation_has_use) {
1254 InitializeNode* init = alloc->initialization();
1255 if (init != NULL) {
1256 init->remove(&_igvn);
1257 }
1258 if (expand_fast_path && (initial_slow_test == NULL)) {
1259 // Remove allocation node and return.
1260 // Size is a non-negative constant -> no initial check needed -> directly to fast path.
1261 // Also, no usages -> empty fast path -> no fall out to slow path -> nothing left.
1262 #ifndef PRODUCT
1263 if (PrintEliminateAllocations) {
1264 tty->print("NotUsed ");
1265 Node* res = alloc->proj_out_or_null(TypeFunc::Parms);
1266 if (res != NULL) {
1267 res->dump();
1268 } else {
1269 alloc->dump();
1270 }
1271 }
1272 #endif
1273 yank_alloc_node(alloc);
1274 return;
1275 }
1276 }
1277
1278 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1279 Node *slow_region = NULL;
1280 Node *toobig_false = ctrl;
1281
1282 // generate the initial test if necessary
1283 if (initial_slow_test != NULL ) {
1284 assert (expand_fast_path, "Only need test if there is a fast path");
1285 slow_region = new RegionNode(3);
1286
1287 // Now make the initial failure test. Usually a too-big test but
1288 // might be a TRUE for finalizers or a fancy class check for
1289 // newInstance0.
1290 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1291 transform_later(toobig_iff);
1292 // Plug the failing-too-big test into the slow-path region
1293 Node *toobig_true = new IfTrueNode( toobig_iff );
1294 transform_later(toobig_true);
1295 slow_region ->init_req( too_big_or_final_path, toobig_true );
1296 toobig_false = new IfFalseNode( toobig_iff );
1297 transform_later(toobig_false);
1298 } else {
1299 // No initial test, just fall into next case
1300 assert(allocation_has_use || !expand_fast_path, "Should already have been handled");
1301 toobig_false = ctrl;
1302 debug_only(slow_region = NodeSentinel);
1303 }
1304
1305 // If we are here there are several possibilities
1306 // - expand_fast_path is false - then only a slow path is expanded. That's it.
1307 // no_initial_check means a constant allocation.
1308 // - If check always evaluates to false -> expand_fast_path is false (see above)
1309 // - If check always evaluates to true -> directly into fast path (but may bailout to slowpath)
1310 // if !allocation_has_use the fast path is empty
1311 // if !allocation_has_use && no_initial_check
1312 // - Then there are no fastpath that can fall out to slowpath -> no allocation code at all.
1313 // removed by yank_alloc_node above.
1314
1315 Node *slow_mem = mem; // save the current memory state for slow path
1316 // generate the fast allocation code unless we know that the initial test will always go slow
1317 if (expand_fast_path) {
1318 // Fast path modifies only raw memory.
1319 if (mem->is_MergeMem()) {
1320 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1321 }
1322
1323 // allocate the Region and Phi nodes for the result
1324 result_region = new RegionNode(3);
1325 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1326 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1327
1328 // Grab regular I/O before optional prefetch may change it.
1329 // Slow-path does no I/O so just set it to the original I/O.
1330 result_phi_i_o->init_req(slow_result_path, i_o);
1331
1332 // Name successful fast-path variables
1333 Node* fast_oop_ctrl;
1334 Node* fast_oop_rawmem;
1335 if (allocation_has_use) {
1336 Node* needgc_ctrl = NULL;
1337 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1338
1339 intx prefetch_lines = length != NULL ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1340 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1341 Node* fast_oop = bs->obj_allocate(this, mem, toobig_false, size_in_bytes, i_o, needgc_ctrl,
1342 fast_oop_ctrl, fast_oop_rawmem,
1343 prefetch_lines);
1344
1345 if (initial_slow_test != NULL) {
1346 // This completes all paths into the slow merge point
1347 slow_region->init_req(need_gc_path, needgc_ctrl);
1348 transform_later(slow_region);
1349 } else {
1350 // No initial slow path needed!
1351 // Just fall from the need-GC path straight into the VM call.
1352 slow_region = needgc_ctrl;
1353 }
1354
1355 InitializeNode* init = alloc->initialization();
1356 fast_oop_rawmem = initialize_object(alloc,
1357 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1358 klass_node, length, size_in_bytes);
1359 expand_initialize_membar(alloc, init, fast_oop_ctrl, fast_oop_rawmem);
1360 expand_dtrace_alloc_probe(alloc, fast_oop, fast_oop_ctrl, fast_oop_rawmem);
1361
1362 result_phi_rawoop->init_req(fast_result_path, fast_oop);
1363 } else {
1364 assert (initial_slow_test != NULL, "sanity");
1365 fast_oop_ctrl = toobig_false;
1366 fast_oop_rawmem = mem;
1367 transform_later(slow_region);
1368 }
1369
1370 // Plug in the successful fast-path into the result merge point
1371 result_region ->init_req(fast_result_path, fast_oop_ctrl);
1372 result_phi_i_o ->init_req(fast_result_path, i_o);
1373 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1374 } else {
1375 slow_region = ctrl;
1376 result_phi_i_o = i_o; // Rename it to use in the following code.
1377 }
1378
1379 // Generate slow-path call
1380 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1381 OptoRuntime::stub_name(slow_call_address),
1382 TypePtr::BOTTOM);
1383 call->init_req(TypeFunc::Control, slow_region);
1384 call->init_req(TypeFunc::I_O, top()); // does no i/o
1385 call->init_req(TypeFunc::Memory, slow_mem); // may gc ptrs
1386 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1387 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1388
1389 call->init_req(TypeFunc::Parms+0, klass_node);
1390 if (length != NULL) {
1391 call->init_req(TypeFunc::Parms+1, length);
1392 }
1393
1394 // Copy debug information and adjust JVMState information, then replace
1395 // allocate node with the call
1396 call->copy_call_debug_info(&_igvn, alloc);
1397 if (expand_fast_path) {
1398 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1399 } else {
1400 // Hook i_o projection to avoid its elimination during allocation
1401 // replacement (when only a slow call is generated).
1402 call->set_req(TypeFunc::I_O, result_phi_i_o);
1403 }
1404 _igvn.replace_node(alloc, call);
1405 transform_later(call);
1406
1407 // Identify the output projections from the allocate node and
1408 // adjust any references to them.
1409 // The control and io projections look like:
1410 //
1411 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1412 // Allocate Catch
1413 // ^---Proj(io) <-------+ ^---CatchProj(io)
1414 //
1415 // We are interested in the CatchProj nodes.
1416 //
1417 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1418
1419 // An allocate node has separate memory projections for the uses on
1420 // the control and i_o paths. Replace the control memory projection with
1421 // result_phi_rawmem (unless we are only generating a slow call when
1422 // both memory projections are combined)
1423 if (expand_fast_path && _callprojs.fallthrough_memproj != NULL) {
1424 migrate_outs(_callprojs.fallthrough_memproj, result_phi_rawmem);
1425 }
1426 // Now change uses of catchall_memproj to use fallthrough_memproj and delete
1427 // catchall_memproj so we end up with a call that has only 1 memory projection.
1428 if (_callprojs.catchall_memproj != NULL ) {
1429 if (_callprojs.fallthrough_memproj == NULL) {
1430 _callprojs.fallthrough_memproj = new ProjNode(call, TypeFunc::Memory);
1431 transform_later(_callprojs.fallthrough_memproj);
1432 }
1433 migrate_outs(_callprojs.catchall_memproj, _callprojs.fallthrough_memproj);
1434 _igvn.remove_dead_node(_callprojs.catchall_memproj);
1435 }
1436
1437 // An allocate node has separate i_o projections for the uses on the control
1438 // and i_o paths. Always replace the control i_o projection with result i_o
1439 // otherwise incoming i_o become dead when only a slow call is generated
1440 // (it is different from memory projections where both projections are
1441 // combined in such case).
1442 if (_callprojs.fallthrough_ioproj != NULL) {
1443 migrate_outs(_callprojs.fallthrough_ioproj, result_phi_i_o);
1444 }
1445 // Now change uses of catchall_ioproj to use fallthrough_ioproj and delete
1446 // catchall_ioproj so we end up with a call that has only 1 i_o projection.
1447 if (_callprojs.catchall_ioproj != NULL ) {
1448 if (_callprojs.fallthrough_ioproj == NULL) {
1449 _callprojs.fallthrough_ioproj = new ProjNode(call, TypeFunc::I_O);
1450 transform_later(_callprojs.fallthrough_ioproj);
1451 }
1452 migrate_outs(_callprojs.catchall_ioproj, _callprojs.fallthrough_ioproj);
1453 _igvn.remove_dead_node(_callprojs.catchall_ioproj);
1454 }
1455
1456 // if we generated only a slow call, we are done
1457 if (!expand_fast_path) {
1458 // Now we can unhook i_o.
1459 if (result_phi_i_o->outcnt() > 1) {
1460 call->set_req(TypeFunc::I_O, top());
1461 } else {
1462 assert(result_phi_i_o->unique_ctrl_out() == call, "sanity");
1463 // Case of new array with negative size known during compilation.
1464 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1465 // following code since call to runtime will throw exception.
1466 // As result there will be no users of i_o after the call.
1467 // Leave i_o attached to this call to avoid problems in preceding graph.
1468 }
1469 return;
1470 }
1471
1472 if (_callprojs.fallthrough_catchproj != NULL) {
1473 ctrl = _callprojs.fallthrough_catchproj->clone();
1474 transform_later(ctrl);
1475 _igvn.replace_node(_callprojs.fallthrough_catchproj, result_region);
1476 } else {
1477 ctrl = top();
1478 }
1479 Node *slow_result;
1480 if (_callprojs.resproj == NULL) {
1481 // no uses of the allocation result
1482 slow_result = top();
1483 } else {
1484 slow_result = _callprojs.resproj->clone();
1485 transform_later(slow_result);
1486 _igvn.replace_node(_callprojs.resproj, result_phi_rawoop);
1487 }
1488
1489 // Plug slow-path into result merge point
1490 result_region->init_req( slow_result_path, ctrl);
1491 transform_later(result_region);
1492 if (allocation_has_use) {
1493 result_phi_rawoop->init_req(slow_result_path, slow_result);
1494 transform_later(result_phi_rawoop);
1495 }
1496 result_phi_rawmem->init_req(slow_result_path, _callprojs.fallthrough_memproj);
1497 transform_later(result_phi_rawmem);
1498 transform_later(result_phi_i_o);
1499 // This completes all paths into the result merge point
1500 }
1501
1502 // Remove alloc node that has no uses.
yank_alloc_node(AllocateNode * alloc)1503 void PhaseMacroExpand::yank_alloc_node(AllocateNode* alloc) {
1504 Node* ctrl = alloc->in(TypeFunc::Control);
1505 Node* mem = alloc->in(TypeFunc::Memory);
1506 Node* i_o = alloc->in(TypeFunc::I_O);
1507
1508 alloc->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1509 if (_callprojs.resproj != NULL) {
1510 for (DUIterator_Fast imax, i = _callprojs.resproj->fast_outs(imax); i < imax; i++) {
1511 Node* use = _callprojs.resproj->fast_out(i);
1512 use->isa_MemBar()->remove(&_igvn);
1513 --imax;
1514 --i; // back up iterator
1515 }
1516 assert(_callprojs.resproj->outcnt() == 0, "all uses must be deleted");
1517 _igvn.remove_dead_node(_callprojs.resproj);
1518 }
1519 if (_callprojs.fallthrough_catchproj != NULL) {
1520 migrate_outs(_callprojs.fallthrough_catchproj, ctrl);
1521 _igvn.remove_dead_node(_callprojs.fallthrough_catchproj);
1522 }
1523 if (_callprojs.catchall_catchproj != NULL) {
1524 _igvn.rehash_node_delayed(_callprojs.catchall_catchproj);
1525 _callprojs.catchall_catchproj->set_req(0, top());
1526 }
1527 if (_callprojs.fallthrough_proj != NULL) {
1528 Node* catchnode = _callprojs.fallthrough_proj->unique_ctrl_out();
1529 _igvn.remove_dead_node(catchnode);
1530 _igvn.remove_dead_node(_callprojs.fallthrough_proj);
1531 }
1532 if (_callprojs.fallthrough_memproj != NULL) {
1533 migrate_outs(_callprojs.fallthrough_memproj, mem);
1534 _igvn.remove_dead_node(_callprojs.fallthrough_memproj);
1535 }
1536 if (_callprojs.fallthrough_ioproj != NULL) {
1537 migrate_outs(_callprojs.fallthrough_ioproj, i_o);
1538 _igvn.remove_dead_node(_callprojs.fallthrough_ioproj);
1539 }
1540 if (_callprojs.catchall_memproj != NULL) {
1541 _igvn.rehash_node_delayed(_callprojs.catchall_memproj);
1542 _callprojs.catchall_memproj->set_req(0, top());
1543 }
1544 if (_callprojs.catchall_ioproj != NULL) {
1545 _igvn.rehash_node_delayed(_callprojs.catchall_ioproj);
1546 _callprojs.catchall_ioproj->set_req(0, top());
1547 }
1548 #ifndef PRODUCT
1549 if (PrintEliminateAllocations) {
1550 if (alloc->is_AllocateArray()) {
1551 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1552 } else {
1553 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1554 }
1555 }
1556 #endif
1557 _igvn.remove_dead_node(alloc);
1558 }
1559
expand_initialize_membar(AllocateNode * alloc,InitializeNode * init,Node * & fast_oop_ctrl,Node * & fast_oop_rawmem)1560 void PhaseMacroExpand::expand_initialize_membar(AllocateNode* alloc, InitializeNode* init,
1561 Node*& fast_oop_ctrl, Node*& fast_oop_rawmem) {
1562 // If initialization is performed by an array copy, any required
1563 // MemBarStoreStore was already added. If the object does not
1564 // escape no need for a MemBarStoreStore. If the object does not
1565 // escape in its initializer and memory barrier (MemBarStoreStore or
1566 // stronger) is already added at exit of initializer, also no need
1567 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore
1568 // so that stores that initialize this object can't be reordered
1569 // with a subsequent store that makes this object accessible by
1570 // other threads.
1571 // Other threads include java threads and JVM internal threads
1572 // (for example concurrent GC threads). Current concurrent GC
1573 // implementation: G1 will not scan newly created object,
1574 // so it's safe to skip storestore barrier when allocation does
1575 // not escape.
1576 if (!alloc->does_not_escape_thread() &&
1577 !alloc->is_allocation_MemBar_redundant() &&
1578 (init == NULL || !init->is_complete_with_arraycopy())) {
1579 if (init == NULL || init->req() < InitializeNode::RawStores) {
1580 // No InitializeNode or no stores captured by zeroing
1581 // elimination. Simply add the MemBarStoreStore after object
1582 // initialization.
1583 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1584 transform_later(mb);
1585
1586 mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
1587 mb->init_req(TypeFunc::Control, fast_oop_ctrl);
1588 fast_oop_ctrl = new ProjNode(mb, TypeFunc::Control);
1589 transform_later(fast_oop_ctrl);
1590 fast_oop_rawmem = new ProjNode(mb, TypeFunc::Memory);
1591 transform_later(fast_oop_rawmem);
1592 } else {
1593 // Add the MemBarStoreStore after the InitializeNode so that
1594 // all stores performing the initialization that were moved
1595 // before the InitializeNode happen before the storestore
1596 // barrier.
1597
1598 Node* init_ctrl = init->proj_out_or_null(TypeFunc::Control);
1599 Node* init_mem = init->proj_out_or_null(TypeFunc::Memory);
1600
1601 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1602 transform_later(mb);
1603
1604 Node* ctrl = new ProjNode(init, TypeFunc::Control);
1605 transform_later(ctrl);
1606 Node* mem = new ProjNode(init, TypeFunc::Memory);
1607 transform_later(mem);
1608
1609 // The MemBarStoreStore depends on control and memory coming
1610 // from the InitializeNode
1611 mb->init_req(TypeFunc::Memory, mem);
1612 mb->init_req(TypeFunc::Control, ctrl);
1613
1614 ctrl = new ProjNode(mb, TypeFunc::Control);
1615 transform_later(ctrl);
1616 mem = new ProjNode(mb, TypeFunc::Memory);
1617 transform_later(mem);
1618
1619 // All nodes that depended on the InitializeNode for control
1620 // and memory must now depend on the MemBarNode that itself
1621 // depends on the InitializeNode
1622 if (init_ctrl != NULL) {
1623 _igvn.replace_node(init_ctrl, ctrl);
1624 }
1625 if (init_mem != NULL) {
1626 _igvn.replace_node(init_mem, mem);
1627 }
1628 }
1629 }
1630 }
1631
expand_dtrace_alloc_probe(AllocateNode * alloc,Node * oop,Node * & ctrl,Node * & rawmem)1632 void PhaseMacroExpand::expand_dtrace_alloc_probe(AllocateNode* alloc, Node* oop,
1633 Node*& ctrl, Node*& rawmem) {
1634 if (C->env()->dtrace_extended_probes()) {
1635 // Slow-path call
1636 int size = TypeFunc::Parms + 2;
1637 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1638 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base),
1639 "dtrace_object_alloc",
1640 TypeRawPtr::BOTTOM);
1641
1642 // Get base of thread-local storage area
1643 Node* thread = new ThreadLocalNode();
1644 transform_later(thread);
1645
1646 call->init_req(TypeFunc::Parms + 0, thread);
1647 call->init_req(TypeFunc::Parms + 1, oop);
1648 call->init_req(TypeFunc::Control, ctrl);
1649 call->init_req(TypeFunc::I_O , top()); // does no i/o
1650 call->init_req(TypeFunc::Memory , ctrl);
1651 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1652 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1653 transform_later(call);
1654 ctrl = new ProjNode(call, TypeFunc::Control);
1655 transform_later(ctrl);
1656 rawmem = new ProjNode(call, TypeFunc::Memory);
1657 transform_later(rawmem);
1658 }
1659 }
1660
1661 // Helper for PhaseMacroExpand::expand_allocate_common.
1662 // Initializes the newly-allocated storage.
1663 Node*
initialize_object(AllocateNode * alloc,Node * control,Node * rawmem,Node * object,Node * klass_node,Node * length,Node * size_in_bytes)1664 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1665 Node* control, Node* rawmem, Node* object,
1666 Node* klass_node, Node* length,
1667 Node* size_in_bytes) {
1668 InitializeNode* init = alloc->initialization();
1669 // Store the klass & mark bits
1670 Node* mark_node = alloc->make_ideal_mark(&_igvn, object, control, rawmem);
1671 if (!mark_node->is_Con()) {
1672 transform_later(mark_node);
1673 }
1674 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, TypeX_X->basic_type());
1675
1676 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1677 int header_size = alloc->minimum_header_size(); // conservatively small
1678
1679 // Array length
1680 if (length != NULL) { // Arrays need length field
1681 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1682 // conservatively small header size:
1683 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1684 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1685 if (k->is_array_klass()) // we know the exact header size in most cases:
1686 header_size = Klass::layout_helper_header_size(k->layout_helper());
1687 }
1688
1689 // Clear the object body, if necessary.
1690 if (init == NULL) {
1691 // The init has somehow disappeared; be cautious and clear everything.
1692 //
1693 // This can happen if a node is allocated but an uncommon trap occurs
1694 // immediately. In this case, the Initialize gets associated with the
1695 // trap, and may be placed in a different (outer) loop, if the Allocate
1696 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1697 // there can be two Allocates to one Initialize. The answer in all these
1698 // edge cases is safety first. It is always safe to clear immediately
1699 // within an Allocate, and then (maybe or maybe not) clear some more later.
1700 if (!(UseTLAB && ZeroTLAB)) {
1701 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1702 header_size, size_in_bytes,
1703 &_igvn);
1704 }
1705 } else {
1706 if (!init->is_complete()) {
1707 // Try to win by zeroing only what the init does not store.
1708 // We can also try to do some peephole optimizations,
1709 // such as combining some adjacent subword stores.
1710 rawmem = init->complete_stores(control, rawmem, object,
1711 header_size, size_in_bytes, &_igvn);
1712 }
1713 // We have no more use for this link, since the AllocateNode goes away:
1714 init->set_req(InitializeNode::RawAddress, top());
1715 // (If we keep the link, it just confuses the register allocator,
1716 // who thinks he sees a real use of the address by the membar.)
1717 }
1718
1719 return rawmem;
1720 }
1721
1722 // Generate prefetch instructions for next allocations.
prefetch_allocation(Node * i_o,Node * & needgc_false,Node * & contended_phi_rawmem,Node * old_eden_top,Node * new_eden_top,intx lines)1723 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1724 Node*& contended_phi_rawmem,
1725 Node* old_eden_top, Node* new_eden_top,
1726 intx lines) {
1727 enum { fall_in_path = 1, pf_path = 2 };
1728 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1729 // Generate prefetch allocation with watermark check.
1730 // As an allocation hits the watermark, we will prefetch starting
1731 // at a "distance" away from watermark.
1732
1733 Node *pf_region = new RegionNode(3);
1734 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY,
1735 TypeRawPtr::BOTTOM );
1736 // I/O is used for Prefetch
1737 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO );
1738
1739 Node *thread = new ThreadLocalNode();
1740 transform_later(thread);
1741
1742 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread,
1743 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1744 transform_later(eden_pf_adr);
1745
1746 Node *old_pf_wm = new LoadPNode(needgc_false,
1747 contended_phi_rawmem, eden_pf_adr,
1748 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM,
1749 MemNode::unordered);
1750 transform_later(old_pf_wm);
1751
1752 // check against new_eden_top
1753 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm );
1754 transform_later(need_pf_cmp);
1755 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge );
1756 transform_later(need_pf_bol);
1757 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol,
1758 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1759 transform_later(need_pf_iff);
1760
1761 // true node, add prefetchdistance
1762 Node *need_pf_true = new IfTrueNode( need_pf_iff );
1763 transform_later(need_pf_true);
1764
1765 Node *need_pf_false = new IfFalseNode( need_pf_iff );
1766 transform_later(need_pf_false);
1767
1768 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm,
1769 _igvn.MakeConX(AllocatePrefetchDistance) );
1770 transform_later(new_pf_wmt );
1771 new_pf_wmt->set_req(0, need_pf_true);
1772
1773 Node *store_new_wmt = new StorePNode(need_pf_true,
1774 contended_phi_rawmem, eden_pf_adr,
1775 TypeRawPtr::BOTTOM, new_pf_wmt,
1776 MemNode::unordered);
1777 transform_later(store_new_wmt);
1778
1779 // adding prefetches
1780 pf_phi_abio->init_req( fall_in_path, i_o );
1781
1782 Node *prefetch_adr;
1783 Node *prefetch;
1784 uint step_size = AllocatePrefetchStepSize;
1785 uint distance = 0;
1786
1787 for ( intx i = 0; i < lines; i++ ) {
1788 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt,
1789 _igvn.MakeConX(distance) );
1790 transform_later(prefetch_adr);
1791 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1792 transform_later(prefetch);
1793 distance += step_size;
1794 i_o = prefetch;
1795 }
1796 pf_phi_abio->set_req( pf_path, i_o );
1797
1798 pf_region->init_req( fall_in_path, need_pf_false );
1799 pf_region->init_req( pf_path, need_pf_true );
1800
1801 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
1802 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
1803
1804 transform_later(pf_region);
1805 transform_later(pf_phi_rawmem);
1806 transform_later(pf_phi_abio);
1807
1808 needgc_false = pf_region;
1809 contended_phi_rawmem = pf_phi_rawmem;
1810 i_o = pf_phi_abio;
1811 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) {
1812 // Insert a prefetch instruction for each allocation.
1813 // This code is used to generate 1 prefetch instruction per cache line.
1814
1815 // Generate several prefetch instructions.
1816 uint step_size = AllocatePrefetchStepSize;
1817 uint distance = AllocatePrefetchDistance;
1818
1819 // Next cache address.
1820 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top,
1821 _igvn.MakeConX(step_size + distance));
1822 transform_later(cache_adr);
1823 cache_adr = new CastP2XNode(needgc_false, cache_adr);
1824 transform_later(cache_adr);
1825 // Address is aligned to execute prefetch to the beginning of cache line size
1826 // (it is important when BIS instruction is used on SPARC as prefetch).
1827 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
1828 cache_adr = new AndXNode(cache_adr, mask);
1829 transform_later(cache_adr);
1830 cache_adr = new CastX2PNode(cache_adr);
1831 transform_later(cache_adr);
1832
1833 // Prefetch
1834 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr );
1835 prefetch->set_req(0, needgc_false);
1836 transform_later(prefetch);
1837 contended_phi_rawmem = prefetch;
1838 Node *prefetch_adr;
1839 distance = step_size;
1840 for ( intx i = 1; i < lines; i++ ) {
1841 prefetch_adr = new AddPNode( cache_adr, cache_adr,
1842 _igvn.MakeConX(distance) );
1843 transform_later(prefetch_adr);
1844 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr );
1845 transform_later(prefetch);
1846 distance += step_size;
1847 contended_phi_rawmem = prefetch;
1848 }
1849 } else if( AllocatePrefetchStyle > 0 ) {
1850 // Insert a prefetch for each allocation only on the fast-path
1851 Node *prefetch_adr;
1852 Node *prefetch;
1853 // Generate several prefetch instructions.
1854 uint step_size = AllocatePrefetchStepSize;
1855 uint distance = AllocatePrefetchDistance;
1856 for ( intx i = 0; i < lines; i++ ) {
1857 prefetch_adr = new AddPNode( old_eden_top, new_eden_top,
1858 _igvn.MakeConX(distance) );
1859 transform_later(prefetch_adr);
1860 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1861 // Do not let it float too high, since if eden_top == eden_end,
1862 // both might be null.
1863 if( i == 0 ) { // Set control for first prefetch, next follows it
1864 prefetch->init_req(0, needgc_false);
1865 }
1866 transform_later(prefetch);
1867 distance += step_size;
1868 i_o = prefetch;
1869 }
1870 }
1871 return i_o;
1872 }
1873
1874
expand_allocate(AllocateNode * alloc)1875 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
1876 expand_allocate_common(alloc, NULL,
1877 OptoRuntime::new_instance_Type(),
1878 OptoRuntime::new_instance_Java());
1879 }
1880
expand_allocate_array(AllocateArrayNode * alloc)1881 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
1882 Node* length = alloc->in(AllocateNode::ALength);
1883 InitializeNode* init = alloc->initialization();
1884 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1885 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1886 address slow_call_address; // Address of slow call
1887 if (init != NULL && init->is_complete_with_arraycopy() &&
1888 k->is_type_array_klass()) {
1889 // Don't zero type array during slow allocation in VM since
1890 // it will be initialized later by arraycopy in compiled code.
1891 slow_call_address = OptoRuntime::new_array_nozero_Java();
1892 } else {
1893 slow_call_address = OptoRuntime::new_array_Java();
1894 }
1895 expand_allocate_common(alloc, length,
1896 OptoRuntime::new_array_Type(),
1897 slow_call_address);
1898 }
1899
1900 //-------------------mark_eliminated_box----------------------------------
1901 //
1902 // During EA obj may point to several objects but after few ideal graph
1903 // transformations (CCP) it may point to only one non escaping object
1904 // (but still using phi), corresponding locks and unlocks will be marked
1905 // for elimination. Later obj could be replaced with a new node (new phi)
1906 // and which does not have escape information. And later after some graph
1907 // reshape other locks and unlocks (which were not marked for elimination
1908 // before) are connected to this new obj (phi) but they still will not be
1909 // marked for elimination since new obj has no escape information.
1910 // Mark all associated (same box and obj) lock and unlock nodes for
1911 // elimination if some of them marked already.
mark_eliminated_box(Node * oldbox,Node * obj)1912 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) {
1913 if (oldbox->as_BoxLock()->is_eliminated()) {
1914 return; // This BoxLock node was processed already.
1915 }
1916 // New implementation (EliminateNestedLocks) has separate BoxLock
1917 // node for each locked region so mark all associated locks/unlocks as
1918 // eliminated even if different objects are referenced in one locked region
1919 // (for example, OSR compilation of nested loop inside locked scope).
1920 if (EliminateNestedLocks ||
1921 oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj, NULL)) {
1922 // Box is used only in one lock region. Mark this box as eliminated.
1923 _igvn.hash_delete(oldbox);
1924 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value
1925 _igvn.hash_insert(oldbox);
1926
1927 for (uint i = 0; i < oldbox->outcnt(); i++) {
1928 Node* u = oldbox->raw_out(i);
1929 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
1930 AbstractLockNode* alock = u->as_AbstractLock();
1931 // Check lock's box since box could be referenced by Lock's debug info.
1932 if (alock->box_node() == oldbox) {
1933 // Mark eliminated all related locks and unlocks.
1934 #ifdef ASSERT
1935 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4");
1936 #endif
1937 alock->set_non_esc_obj();
1938 }
1939 }
1940 }
1941 return;
1942 }
1943
1944 // Create new "eliminated" BoxLock node and use it in monitor debug info
1945 // instead of oldbox for the same object.
1946 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
1947
1948 // Note: BoxLock node is marked eliminated only here and it is used
1949 // to indicate that all associated lock and unlock nodes are marked
1950 // for elimination.
1951 newbox->set_eliminated();
1952 transform_later(newbox);
1953
1954 // Replace old box node with new box for all users of the same object.
1955 for (uint i = 0; i < oldbox->outcnt();) {
1956 bool next_edge = true;
1957
1958 Node* u = oldbox->raw_out(i);
1959 if (u->is_AbstractLock()) {
1960 AbstractLockNode* alock = u->as_AbstractLock();
1961 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
1962 // Replace Box and mark eliminated all related locks and unlocks.
1963 #ifdef ASSERT
1964 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5");
1965 #endif
1966 alock->set_non_esc_obj();
1967 _igvn.rehash_node_delayed(alock);
1968 alock->set_box_node(newbox);
1969 next_edge = false;
1970 }
1971 }
1972 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
1973 FastLockNode* flock = u->as_FastLock();
1974 assert(flock->box_node() == oldbox, "sanity");
1975 _igvn.rehash_node_delayed(flock);
1976 flock->set_box_node(newbox);
1977 next_edge = false;
1978 }
1979
1980 // Replace old box in monitor debug info.
1981 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
1982 SafePointNode* sfn = u->as_SafePoint();
1983 JVMState* youngest_jvms = sfn->jvms();
1984 int max_depth = youngest_jvms->depth();
1985 for (int depth = 1; depth <= max_depth; depth++) {
1986 JVMState* jvms = youngest_jvms->of_depth(depth);
1987 int num_mon = jvms->nof_monitors();
1988 // Loop over monitors
1989 for (int idx = 0; idx < num_mon; idx++) {
1990 Node* obj_node = sfn->monitor_obj(jvms, idx);
1991 Node* box_node = sfn->monitor_box(jvms, idx);
1992 if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
1993 int j = jvms->monitor_box_offset(idx);
1994 _igvn.replace_input_of(u, j, newbox);
1995 next_edge = false;
1996 }
1997 }
1998 }
1999 }
2000 if (next_edge) i++;
2001 }
2002 }
2003
2004 //-----------------------mark_eliminated_locking_nodes-----------------------
mark_eliminated_locking_nodes(AbstractLockNode * alock)2005 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
2006 if (EliminateNestedLocks) {
2007 if (alock->is_nested()) {
2008 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
2009 return;
2010 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
2011 // Only Lock node has JVMState needed here.
2012 // Not that preceding claim is documented anywhere else.
2013 if (alock->jvms() != NULL) {
2014 if (alock->as_Lock()->is_nested_lock_region()) {
2015 // Mark eliminated related nested locks and unlocks.
2016 Node* obj = alock->obj_node();
2017 BoxLockNode* box_node = alock->box_node()->as_BoxLock();
2018 assert(!box_node->is_eliminated(), "should not be marked yet");
2019 // Note: BoxLock node is marked eliminated only here
2020 // and it is used to indicate that all associated lock
2021 // and unlock nodes are marked for elimination.
2022 box_node->set_eliminated(); // Box's hash is always NO_HASH here
2023 for (uint i = 0; i < box_node->outcnt(); i++) {
2024 Node* u = box_node->raw_out(i);
2025 if (u->is_AbstractLock()) {
2026 alock = u->as_AbstractLock();
2027 if (alock->box_node() == box_node) {
2028 // Verify that this Box is referenced only by related locks.
2029 assert(alock->obj_node()->eqv_uncast(obj), "");
2030 // Mark all related locks and unlocks.
2031 #ifdef ASSERT
2032 alock->log_lock_optimization(C, "eliminate_lock_set_nested");
2033 #endif
2034 alock->set_nested();
2035 }
2036 }
2037 }
2038 } else {
2039 #ifdef ASSERT
2040 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region");
2041 if (C->log() != NULL)
2042 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output
2043 #endif
2044 }
2045 }
2046 return;
2047 }
2048 // Process locks for non escaping object
2049 assert(alock->is_non_esc_obj(), "");
2050 } // EliminateNestedLocks
2051
2052 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2053 // Look for all locks of this object and mark them and
2054 // corresponding BoxLock nodes as eliminated.
2055 Node* obj = alock->obj_node();
2056 for (uint j = 0; j < obj->outcnt(); j++) {
2057 Node* o = obj->raw_out(j);
2058 if (o->is_AbstractLock() &&
2059 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2060 alock = o->as_AbstractLock();
2061 Node* box = alock->box_node();
2062 // Replace old box node with new eliminated box for all users
2063 // of the same object and mark related locks as eliminated.
2064 mark_eliminated_box(box, obj);
2065 }
2066 }
2067 }
2068 }
2069
2070 // we have determined that this lock/unlock can be eliminated, we simply
2071 // eliminate the node without expanding it.
2072 //
2073 // Note: The membar's associated with the lock/unlock are currently not
2074 // eliminated. This should be investigated as a future enhancement.
2075 //
eliminate_locking_node(AbstractLockNode * alock)2076 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2077
2078 if (!alock->is_eliminated()) {
2079 return false;
2080 }
2081 #ifdef ASSERT
2082 if (!alock->is_coarsened()) {
2083 // Check that new "eliminated" BoxLock node is created.
2084 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2085 assert(oldbox->is_eliminated(), "should be done already");
2086 }
2087 #endif
2088
2089 alock->log_lock_optimization(C, "eliminate_lock");
2090
2091 #ifndef PRODUCT
2092 if (PrintEliminateLocks) {
2093 tty->print_cr("++++ Eliminated: %d %s '%s'", alock->_idx, (alock->is_Lock() ? "Lock" : "Unlock"), alock->kind_as_string());
2094 }
2095 #endif
2096
2097 Node* mem = alock->in(TypeFunc::Memory);
2098 Node* ctrl = alock->in(TypeFunc::Control);
2099 guarantee(ctrl != NULL, "missing control projection, cannot replace_node() with NULL");
2100
2101 alock->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
2102 // There are 2 projections from the lock. The lock node will
2103 // be deleted when its last use is subsumed below.
2104 assert(alock->outcnt() == 2 &&
2105 _callprojs.fallthrough_proj != NULL &&
2106 _callprojs.fallthrough_memproj != NULL,
2107 "Unexpected projections from Lock/Unlock");
2108
2109 Node* fallthroughproj = _callprojs.fallthrough_proj;
2110 Node* memproj_fallthrough = _callprojs.fallthrough_memproj;
2111
2112 // The memory projection from a lock/unlock is RawMem
2113 // The input to a Lock is merged memory, so extract its RawMem input
2114 // (unless the MergeMem has been optimized away.)
2115 if (alock->is_Lock()) {
2116 // Seach for MemBarAcquireLock node and delete it also.
2117 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2118 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, "");
2119 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2120 Node* memproj = membar->proj_out(TypeFunc::Memory);
2121 _igvn.replace_node(ctrlproj, fallthroughproj);
2122 _igvn.replace_node(memproj, memproj_fallthrough);
2123
2124 // Delete FastLock node also if this Lock node is unique user
2125 // (a loop peeling may clone a Lock node).
2126 Node* flock = alock->as_Lock()->fastlock_node();
2127 if (flock->outcnt() == 1) {
2128 assert(flock->unique_out() == alock, "sanity");
2129 _igvn.replace_node(flock, top());
2130 }
2131 }
2132
2133 // Seach for MemBarReleaseLock node and delete it also.
2134 if (alock->is_Unlock() && ctrl->is_Proj() && ctrl->in(0)->is_MemBar()) {
2135 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2136 assert(membar->Opcode() == Op_MemBarReleaseLock &&
2137 mem->is_Proj() && membar == mem->in(0), "");
2138 _igvn.replace_node(fallthroughproj, ctrl);
2139 _igvn.replace_node(memproj_fallthrough, mem);
2140 fallthroughproj = ctrl;
2141 memproj_fallthrough = mem;
2142 ctrl = membar->in(TypeFunc::Control);
2143 mem = membar->in(TypeFunc::Memory);
2144 }
2145
2146 _igvn.replace_node(fallthroughproj, ctrl);
2147 _igvn.replace_node(memproj_fallthrough, mem);
2148 return true;
2149 }
2150
2151
2152 //------------------------------expand_lock_node----------------------
expand_lock_node(LockNode * lock)2153 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2154
2155 Node* ctrl = lock->in(TypeFunc::Control);
2156 Node* mem = lock->in(TypeFunc::Memory);
2157 Node* obj = lock->obj_node();
2158 Node* box = lock->box_node();
2159 Node* flock = lock->fastlock_node();
2160
2161 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2162
2163 // Make the merge point
2164 Node *region;
2165 Node *mem_phi;
2166 Node *slow_path;
2167
2168 if (UseOptoBiasInlining) {
2169 /*
2170 * See the full description in MacroAssembler::biased_locking_enter().
2171 *
2172 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) {
2173 * // The object is biased.
2174 * proto_node = klass->prototype_header;
2175 * o_node = thread | proto_node;
2176 * x_node = o_node ^ mark_word;
2177 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ?
2178 * // Done.
2179 * } else {
2180 * if( (x_node & biased_lock_mask) != 0 ) {
2181 * // The klass's prototype header is no longer biased.
2182 * cas(&mark_word, mark_word, proto_node)
2183 * goto cas_lock;
2184 * } else {
2185 * // The klass's prototype header is still biased.
2186 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch?
2187 * old = mark_word;
2188 * new = o_node;
2189 * } else {
2190 * // Different thread or anonymous biased.
2191 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask);
2192 * new = thread | old;
2193 * }
2194 * // Try to rebias.
2195 * if( cas(&mark_word, old, new) == 0 ) {
2196 * // Done.
2197 * } else {
2198 * goto slow_path; // Failed.
2199 * }
2200 * }
2201 * }
2202 * } else {
2203 * // The object is not biased.
2204 * cas_lock:
2205 * if( FastLock(obj) == 0 ) {
2206 * // Done.
2207 * } else {
2208 * slow_path:
2209 * OptoRuntime::complete_monitor_locking_Java(obj);
2210 * }
2211 * }
2212 */
2213
2214 region = new RegionNode(5);
2215 // create a Phi for the memory state
2216 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2217
2218 Node* fast_lock_region = new RegionNode(3);
2219 Node* fast_lock_mem_phi = new PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM);
2220
2221 // First, check mark word for the biased lock pattern.
2222 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2223
2224 // Get fast path - mark word has the biased lock pattern.
2225 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node,
2226 markWord::biased_lock_mask_in_place,
2227 markWord::biased_lock_pattern, true);
2228 // fast_lock_region->in(1) is set to slow path.
2229 fast_lock_mem_phi->init_req(1, mem);
2230
2231 // Now check that the lock is biased to the current thread and has
2232 // the same epoch and bias as Klass::_prototype_header.
2233
2234 // Special-case a fresh allocation to avoid building nodes:
2235 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn);
2236 if (klass_node == NULL) {
2237 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
2238 klass_node = transform_later(LoadKlassNode::make(_igvn, NULL, mem, k_adr, _igvn.type(k_adr)->is_ptr()));
2239 #ifdef _LP64
2240 if (UseCompressedClassPointers && klass_node->is_DecodeNKlass()) {
2241 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity");
2242 klass_node->in(1)->init_req(0, ctrl);
2243 } else
2244 #endif
2245 klass_node->init_req(0, ctrl);
2246 }
2247 Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type());
2248
2249 Node* thread = transform_later(new ThreadLocalNode());
2250 Node* cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2251 Node* o_node = transform_later(new OrXNode(cast_thread, proto_node));
2252 Node* x_node = transform_later(new XorXNode(o_node, mark_node));
2253
2254 // Get slow path - mark word does NOT match the value.
2255 STATIC_ASSERT(markWord::age_mask_in_place <= INT_MAX);
2256 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node,
2257 (~(int)markWord::age_mask_in_place), 0);
2258 // region->in(3) is set to fast path - the object is biased to the current thread.
2259 mem_phi->init_req(3, mem);
2260
2261
2262 // Mark word does NOT match the value (thread | Klass::_prototype_header).
2263
2264
2265 // First, check biased pattern.
2266 // Get fast path - _prototype_header has the same biased lock pattern.
2267 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node,
2268 markWord::biased_lock_mask_in_place, 0, true);
2269
2270 not_biased_ctrl = fast_lock_region->in(2); // Slow path
2271 // fast_lock_region->in(2) - the prototype header is no longer biased
2272 // and we have to revoke the bias on this object.
2273 // We are going to try to reset the mark of this object to the prototype
2274 // value and fall through to the CAS-based locking scheme.
2275 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
2276 Node* cas = new StoreXConditionalNode(not_biased_ctrl, mem, adr,
2277 proto_node, mark_node);
2278 transform_later(cas);
2279 Node* proj = transform_later(new SCMemProjNode(cas));
2280 fast_lock_mem_phi->init_req(2, proj);
2281
2282
2283 // Second, check epoch bits.
2284 Node* rebiased_region = new RegionNode(3);
2285 Node* old_phi = new PhiNode( rebiased_region, TypeX_X);
2286 Node* new_phi = new PhiNode( rebiased_region, TypeX_X);
2287
2288 // Get slow path - mark word does NOT match epoch bits.
2289 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node,
2290 markWord::epoch_mask_in_place, 0);
2291 // The epoch of the current bias is not valid, attempt to rebias the object
2292 // toward the current thread.
2293 rebiased_region->init_req(2, epoch_ctrl);
2294 old_phi->init_req(2, mark_node);
2295 new_phi->init_req(2, o_node);
2296
2297 // rebiased_region->in(1) is set to fast path.
2298 // The epoch of the current bias is still valid but we know
2299 // nothing about the owner; it might be set or it might be clear.
2300 Node* cmask = MakeConX(markWord::biased_lock_mask_in_place |
2301 markWord::age_mask_in_place |
2302 markWord::epoch_mask_in_place);
2303 Node* old = transform_later(new AndXNode(mark_node, cmask));
2304 cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2305 Node* new_mark = transform_later(new OrXNode(cast_thread, old));
2306 old_phi->init_req(1, old);
2307 new_phi->init_req(1, new_mark);
2308
2309 transform_later(rebiased_region);
2310 transform_later(old_phi);
2311 transform_later(new_phi);
2312
2313 // Try to acquire the bias of the object using an atomic operation.
2314 // If this fails we will go in to the runtime to revoke the object's bias.
2315 cas = new StoreXConditionalNode(rebiased_region, mem, adr, new_phi, old_phi);
2316 transform_later(cas);
2317 proj = transform_later(new SCMemProjNode(cas));
2318
2319 // Get slow path - Failed to CAS.
2320 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0);
2321 mem_phi->init_req(4, proj);
2322 // region->in(4) is set to fast path - the object is rebiased to the current thread.
2323
2324 // Failed to CAS.
2325 slow_path = new RegionNode(3);
2326 Node *slow_mem = new PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM);
2327
2328 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control
2329 slow_mem->init_req(1, proj);
2330
2331 // Call CAS-based locking scheme (FastLock node).
2332
2333 transform_later(fast_lock_region);
2334 transform_later(fast_lock_mem_phi);
2335
2336 // Get slow path - FastLock failed to lock the object.
2337 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0);
2338 mem_phi->init_req(2, fast_lock_mem_phi);
2339 // region->in(2) is set to fast path - the object is locked to the current thread.
2340
2341 slow_path->init_req(2, ctrl); // Capture slow-control
2342 slow_mem->init_req(2, fast_lock_mem_phi);
2343
2344 transform_later(slow_path);
2345 transform_later(slow_mem);
2346 // Reset lock's memory edge.
2347 lock->set_req(TypeFunc::Memory, slow_mem);
2348
2349 } else {
2350 region = new RegionNode(3);
2351 // create a Phi for the memory state
2352 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2353
2354 // Optimize test; set region slot 2
2355 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2356 mem_phi->init_req(2, mem);
2357 }
2358
2359 // Make slow path call
2360 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(),
2361 OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path,
2362 obj, box, NULL);
2363
2364 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
2365
2366 // Slow path can only throw asynchronous exceptions, which are always
2367 // de-opted. So the compiler thinks the slow-call can never throw an
2368 // exception. If it DOES throw an exception we would need the debug
2369 // info removed first (since if it throws there is no monitor).
2370 assert(_callprojs.fallthrough_ioproj == NULL && _callprojs.catchall_ioproj == NULL &&
2371 _callprojs.catchall_memproj == NULL && _callprojs.catchall_catchproj == NULL, "Unexpected projection from Lock");
2372
2373 // Capture slow path
2374 // disconnect fall-through projection from call and create a new one
2375 // hook up users of fall-through projection to region
2376 Node *slow_ctrl = _callprojs.fallthrough_proj->clone();
2377 transform_later(slow_ctrl);
2378 _igvn.hash_delete(_callprojs.fallthrough_proj);
2379 _callprojs.fallthrough_proj->disconnect_inputs(C);
2380 region->init_req(1, slow_ctrl);
2381 // region inputs are now complete
2382 transform_later(region);
2383 _igvn.replace_node(_callprojs.fallthrough_proj, region);
2384
2385 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2386 mem_phi->init_req(1, memproj );
2387 transform_later(mem_phi);
2388 _igvn.replace_node(_callprojs.fallthrough_memproj, mem_phi);
2389 }
2390
2391 //------------------------------expand_unlock_node----------------------
expand_unlock_node(UnlockNode * unlock)2392 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2393
2394 Node* ctrl = unlock->in(TypeFunc::Control);
2395 Node* mem = unlock->in(TypeFunc::Memory);
2396 Node* obj = unlock->obj_node();
2397 Node* box = unlock->box_node();
2398
2399 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2400
2401 // No need for a null check on unlock
2402
2403 // Make the merge point
2404 Node *region;
2405 Node *mem_phi;
2406
2407 if (UseOptoBiasInlining) {
2408 // Check for biased locking unlock case, which is a no-op.
2409 // See the full description in MacroAssembler::biased_locking_exit().
2410 region = new RegionNode(4);
2411 // create a Phi for the memory state
2412 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2413 mem_phi->init_req(3, mem);
2414
2415 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2416 ctrl = opt_bits_test(ctrl, region, 3, mark_node,
2417 markWord::biased_lock_mask_in_place,
2418 markWord::biased_lock_pattern);
2419 } else {
2420 region = new RegionNode(3);
2421 // create a Phi for the memory state
2422 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2423 }
2424
2425 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2426 funlock = transform_later( funlock )->as_FastUnlock();
2427 // Optimize test; set region slot 2
2428 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2429 Node *thread = transform_later(new ThreadLocalNode());
2430
2431 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2432 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2433 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2434
2435 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
2436 assert(_callprojs.fallthrough_ioproj == NULL && _callprojs.catchall_ioproj == NULL &&
2437 _callprojs.catchall_memproj == NULL && _callprojs.catchall_catchproj == NULL, "Unexpected projection from Lock");
2438
2439 // No exceptions for unlocking
2440 // Capture slow path
2441 // disconnect fall-through projection from call and create a new one
2442 // hook up users of fall-through projection to region
2443 Node *slow_ctrl = _callprojs.fallthrough_proj->clone();
2444 transform_later(slow_ctrl);
2445 _igvn.hash_delete(_callprojs.fallthrough_proj);
2446 _callprojs.fallthrough_proj->disconnect_inputs(C);
2447 region->init_req(1, slow_ctrl);
2448 // region inputs are now complete
2449 transform_later(region);
2450 _igvn.replace_node(_callprojs.fallthrough_proj, region);
2451
2452 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2453 mem_phi->init_req(1, memproj );
2454 mem_phi->init_req(2, mem);
2455 transform_later(mem_phi);
2456 _igvn.replace_node(_callprojs.fallthrough_memproj, mem_phi);
2457 }
2458
expand_subtypecheck_node(SubTypeCheckNode * check)2459 void PhaseMacroExpand::expand_subtypecheck_node(SubTypeCheckNode *check) {
2460 assert(check->in(SubTypeCheckNode::Control) == NULL, "should be pinned");
2461 Node* bol = check->unique_out();
2462 Node* obj_or_subklass = check->in(SubTypeCheckNode::ObjOrSubKlass);
2463 Node* superklass = check->in(SubTypeCheckNode::SuperKlass);
2464 assert(bol->is_Bool() && bol->as_Bool()->_test._test == BoolTest::ne, "unexpected bool node");
2465
2466 for (DUIterator_Last imin, i = bol->last_outs(imin); i >= imin; --i) {
2467 Node* iff = bol->last_out(i);
2468 assert(iff->is_If(), "where's the if?");
2469
2470 if (iff->in(0)->is_top()) {
2471 _igvn.replace_input_of(iff, 1, C->top());
2472 continue;
2473 }
2474
2475 Node* iftrue = iff->as_If()->proj_out(1);
2476 Node* iffalse = iff->as_If()->proj_out(0);
2477 Node* ctrl = iff->in(0);
2478
2479 Node* subklass = NULL;
2480 if (_igvn.type(obj_or_subklass)->isa_klassptr()) {
2481 subklass = obj_or_subklass;
2482 } else {
2483 Node* k_adr = basic_plus_adr(obj_or_subklass, oopDesc::klass_offset_in_bytes());
2484 subklass = _igvn.transform(LoadKlassNode::make(_igvn, NULL, C->immutable_memory(), k_adr, TypeInstPtr::KLASS));
2485 }
2486
2487 Node* not_subtype_ctrl = Phase::gen_subtype_check(subklass, superklass, &ctrl, NULL, _igvn);
2488
2489 _igvn.replace_input_of(iff, 0, C->top());
2490 _igvn.replace_node(iftrue, not_subtype_ctrl);
2491 _igvn.replace_node(iffalse, ctrl);
2492 }
2493 _igvn.replace_node(check, C->top());
2494 }
2495
2496 //---------------------------eliminate_macro_nodes----------------------
2497 // Eliminate scalar replaced allocations and associated locks.
eliminate_macro_nodes()2498 void PhaseMacroExpand::eliminate_macro_nodes() {
2499 if (C->macro_count() == 0)
2500 return;
2501
2502 // Before elimination may re-mark (change to Nested or NonEscObj)
2503 // all associated (same box and obj) lock and unlock nodes.
2504 int cnt = C->macro_count();
2505 for (int i=0; i < cnt; i++) {
2506 Node *n = C->macro_node(i);
2507 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2508 mark_eliminated_locking_nodes(n->as_AbstractLock());
2509 }
2510 }
2511 // Re-marking may break consistency of Coarsened locks.
2512 if (!C->coarsened_locks_consistent()) {
2513 return; // recompile without Coarsened locks if broken
2514 }
2515
2516 // First, attempt to eliminate locks
2517 bool progress = true;
2518 while (progress) {
2519 progress = false;
2520 for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once
2521 Node* n = C->macro_node(i - 1);
2522 bool success = false;
2523 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2524 if (n->is_AbstractLock()) {
2525 success = eliminate_locking_node(n->as_AbstractLock());
2526 }
2527 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2528 progress = progress || success;
2529 }
2530 }
2531 // Next, attempt to eliminate allocations
2532 _has_locks = false;
2533 progress = true;
2534 while (progress) {
2535 progress = false;
2536 for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once
2537 Node* n = C->macro_node(i - 1);
2538 bool success = false;
2539 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2540 switch (n->class_id()) {
2541 case Node::Class_Allocate:
2542 case Node::Class_AllocateArray:
2543 success = eliminate_allocate_node(n->as_Allocate());
2544 break;
2545 case Node::Class_CallStaticJava:
2546 success = eliminate_boxing_node(n->as_CallStaticJava());
2547 break;
2548 case Node::Class_Lock:
2549 case Node::Class_Unlock:
2550 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2551 _has_locks = true;
2552 break;
2553 case Node::Class_ArrayCopy:
2554 break;
2555 case Node::Class_OuterStripMinedLoop:
2556 break;
2557 case Node::Class_SubTypeCheck:
2558 break;
2559 case Node::Class_Opaque1:
2560 break;
2561 default:
2562 assert(n->Opcode() == Op_LoopLimit ||
2563 n->Opcode() == Op_Opaque2 ||
2564 n->Opcode() == Op_Opaque3 ||
2565 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n),
2566 "unknown node type in macro list");
2567 }
2568 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2569 progress = progress || success;
2570 }
2571 }
2572 }
2573
2574 //------------------------------expand_macro_nodes----------------------
2575 // Returns true if a failure occurred.
expand_macro_nodes()2576 bool PhaseMacroExpand::expand_macro_nodes() {
2577 // Last attempt to eliminate macro nodes.
2578 eliminate_macro_nodes();
2579 if (C->failing()) return true;
2580
2581 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2582 bool progress = true;
2583 while (progress) {
2584 progress = false;
2585 for (int i = C->macro_count(); i > 0; i--) {
2586 Node* n = C->macro_node(i-1);
2587 bool success = false;
2588 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2589 if (n->Opcode() == Op_LoopLimit) {
2590 // Remove it from macro list and put on IGVN worklist to optimize.
2591 C->remove_macro_node(n);
2592 _igvn._worklist.push(n);
2593 success = true;
2594 } else if (n->Opcode() == Op_CallStaticJava) {
2595 // Remove it from macro list and put on IGVN worklist to optimize.
2596 C->remove_macro_node(n);
2597 _igvn._worklist.push(n);
2598 success = true;
2599 } else if (n->is_Opaque1() || n->Opcode() == Op_Opaque2) {
2600 _igvn.replace_node(n, n->in(1));
2601 success = true;
2602 #if INCLUDE_RTM_OPT
2603 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) {
2604 assert(C->profile_rtm(), "should be used only in rtm deoptimization code");
2605 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), "");
2606 Node* cmp = n->unique_out();
2607 #ifdef ASSERT
2608 // Validate graph.
2609 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), "");
2610 BoolNode* bol = cmp->unique_out()->as_Bool();
2611 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() &&
2612 (bol->_test._test == BoolTest::ne), "");
2613 IfNode* ifn = bol->unique_out()->as_If();
2614 assert((ifn->outcnt() == 2) &&
2615 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != NULL, "");
2616 #endif
2617 Node* repl = n->in(1);
2618 if (!_has_locks) {
2619 // Remove RTM state check if there are no locks in the code.
2620 // Replace input to compare the same value.
2621 repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1);
2622 }
2623 _igvn.replace_node(n, repl);
2624 success = true;
2625 #endif
2626 } else if (n->Opcode() == Op_OuterStripMinedLoop) {
2627 n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn);
2628 C->remove_macro_node(n);
2629 success = true;
2630 }
2631 assert(!success || (C->macro_count() == (old_macro_count - 1)), "elimination must have deleted one node from macro list");
2632 progress = progress || success;
2633 }
2634 }
2635
2636 // Clean up the graph so we're less likely to hit the maximum node
2637 // limit
2638 _igvn.set_delay_transform(false);
2639 _igvn.optimize();
2640 if (C->failing()) return true;
2641 _igvn.set_delay_transform(true);
2642
2643
2644 // Because we run IGVN after each expansion, some macro nodes may go
2645 // dead and be removed from the list as we iterate over it. Move
2646 // Allocate nodes (processed in a second pass) at the beginning of
2647 // the list and then iterate from the last element of the list until
2648 // an Allocate node is seen. This is robust to random deletion in
2649 // the list due to nodes going dead.
2650 C->sort_macro_nodes();
2651
2652 // expand arraycopy "macro" nodes first
2653 // For ReduceBulkZeroing, we must first process all arraycopy nodes
2654 // before the allocate nodes are expanded.
2655 while (C->macro_count() > 0) {
2656 int macro_count = C->macro_count();
2657 Node * n = C->macro_node(macro_count-1);
2658 assert(n->is_macro(), "only macro nodes expected here");
2659 if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) {
2660 // node is unreachable, so don't try to expand it
2661 C->remove_macro_node(n);
2662 continue;
2663 }
2664 if (n->is_Allocate()) {
2665 break;
2666 }
2667 // Make sure expansion will not cause node limit to be exceeded.
2668 // Worst case is a macro node gets expanded into about 200 nodes.
2669 // Allow 50% more for optimization.
2670 if (C->check_node_count(300, "out of nodes before macro expansion")) {
2671 return true;
2672 }
2673
2674 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2675 switch (n->class_id()) {
2676 case Node::Class_Lock:
2677 expand_lock_node(n->as_Lock());
2678 break;
2679 case Node::Class_Unlock:
2680 expand_unlock_node(n->as_Unlock());
2681 break;
2682 case Node::Class_ArrayCopy:
2683 expand_arraycopy_node(n->as_ArrayCopy());
2684 break;
2685 case Node::Class_SubTypeCheck:
2686 expand_subtypecheck_node(n->as_SubTypeCheck());
2687 break;
2688 default:
2689 assert(false, "unknown node type in macro list");
2690 }
2691 assert(C->macro_count() == (old_macro_count - 1), "expansion must have deleted one node from macro list");
2692 if (C->failing()) return true;
2693
2694 // Clean up the graph so we're less likely to hit the maximum node
2695 // limit
2696 _igvn.set_delay_transform(false);
2697 _igvn.optimize();
2698 if (C->failing()) return true;
2699 _igvn.set_delay_transform(true);
2700 }
2701
2702 // All nodes except Allocate nodes are expanded now. There could be
2703 // new optimization opportunities (such as folding newly created
2704 // load from a just allocated object). Run IGVN.
2705
2706 // expand "macro" nodes
2707 // nodes are removed from the macro list as they are processed
2708 while (C->macro_count() > 0) {
2709 int macro_count = C->macro_count();
2710 Node * n = C->macro_node(macro_count-1);
2711 assert(n->is_macro(), "only macro nodes expected here");
2712 if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) {
2713 // node is unreachable, so don't try to expand it
2714 C->remove_macro_node(n);
2715 continue;
2716 }
2717 // Make sure expansion will not cause node limit to be exceeded.
2718 // Worst case is a macro node gets expanded into about 200 nodes.
2719 // Allow 50% more for optimization.
2720 if (C->check_node_count(300, "out of nodes before macro expansion")) {
2721 return true;
2722 }
2723 switch (n->class_id()) {
2724 case Node::Class_Allocate:
2725 expand_allocate(n->as_Allocate());
2726 break;
2727 case Node::Class_AllocateArray:
2728 expand_allocate_array(n->as_AllocateArray());
2729 break;
2730 default:
2731 assert(false, "unknown node type in macro list");
2732 }
2733 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2734 if (C->failing()) return true;
2735
2736 // Clean up the graph so we're less likely to hit the maximum node
2737 // limit
2738 _igvn.set_delay_transform(false);
2739 _igvn.optimize();
2740 if (C->failing()) return true;
2741 _igvn.set_delay_transform(true);
2742 }
2743
2744 _igvn.set_delay_transform(false);
2745 return false;
2746 }
2747